Pharmacoepidemiology and economic evaluation of measures of potentially inappropriate prescribing

Transcription

1 Royal College of Surgeons in Ireland PhD theses Theses and Dissertations Pharmacoepidemiology and economic evaluation of measures of potentially inappropriate prescribing Frank Moriarty Royal College of Surgeons in Ireland, Citation Moriarty F. Pharmacoepidemiology and economic evaluation of measures of potentially inappropriate prescribing [PhD Thesis]. Dublin: Royal College of Surgeons in Ireland; This Thesis is brought to you for free and open access by the Theses and Dissertations at It has been accepted for inclusion in PhD theses by an authorized administrator of For more information, please contact

2 Use Licence Creative Commons Licence: This work is licensed under a Creative Commons Attribution-Noncommercial-Share Alike 4.0 License. This thesis is available at e-publications@rcsi:

3 Pharmacoepidemiology and economic evaluation of measures of potentially inappropriate prescribing Francis Moriarty BSc (Pharm) MPharm Department of General Practice RCSI A thesis submitted to the School of Postgraduate Studies, Faculty of Medicine and Health Sciences, Royal College of Surgeons in Ireland, in fulfilment of the degree of Doctor of Philosophy Supervisors Professor Tom Fahey Professor Kathleen Bennett Dr Caitriona Cahir October 2016

4 Candidate thesis declaration I declare that this thesis, which I submit to RCSI for examination in consideration of the award of a higher degree of Doctor of Philosophy is my own personal effort. Where any of the content presented is the result of input or data from a related collaborative research programme this is duly acknowledged in the text such that it is possible to ascertain how much of the work is my own. I have not already obtained a degree in RCSI or elsewhere on the basis of this work. Furthermore, I took reasonable care to ensure that the work is original, and, to the best of my knowledge, does not breach copyright law, and has not been taken from other sources except where such work has been cited and acknowledged within the text. Signed: Student number: Date: 19 th October

5 Table of contents Candidate thesis declaration... 2 Table of contents... 3 List of abbreviations... 9 Table of figures Table of tables Summary Acknowledgements Publications Publications arising from this thesis Related publications Conference presentations Oral presentations Poster presentations Other oral presentations Chapter 1 Introduction Introduction Background Demographic changes Challenges of prescribing in older adults Common prescribing challenges in middle-aged adults Potentially inappropriate prescribing Measurement of PIP Examples of explicit PIP measures Optimal characteristics of explicit measures Previous research on PIP in older adults Time trends in appropriate prescribing PIP and patient outcomes Economic impact of PIP Previous research on PIP in middle aged adults Aims and objectives of this thesis The health system in Ireland Justification of methods used Cohort study analysis

6 1.8.2 Economic modelling Thesis structure Chapter 2 Trends in polypharmacy and PIP in primary care Introduction Methods Study design and setting Ethical approval Analysis Results Regular medications and polypharmacy in the total population PIP in older individuals The relationship between PIP and polypharmacy in older individuals Discussion Principal findings Findings in the context of previous research Practice and policy implications Strengths and limitations Conclusions Chapter 3 Characterising potentially inappropriate prescribing of proton pump inhibitors over time Introduction Methods Study design and setting Ethical approval Analysis Results PPI use Factors associated with dosage of long-term PPIs Discussion Principal findings Findings in the context of previous research Strengths and limitations Practice and policy implications Conclusions Chapter 4 Prevalence of PIP and its association with adverse outcomes in a cohort of community-dwelling older people from TILDA

7 4.1 Introduction Methods Study population Data collection Ethical approval Prevalence of PIMs and PPOs Association with adverse outcomes Results Participants PIP exposure Association of STOPP and START with patient outcomes Discussion Principal findings Findings in the context of previous research Practice and policy implications Strengths and limitations Conclusions Chapter 5 Evaluating the economic impact of PIP and related adverse events in older people in Ireland using a Markov modelling approach Introduction Methods Markov models Description of model structures Transition probabilities Costing Utility values Analysis Results Economic impact of PIP relative to non-pip Cost-effectiveness of hypothetical interventions Discussion Principal findings Findings in the context of previous research Practice and policy implications Strengths and limitations

8 5.4.5 Conclusions Chapter 6 PIP in middle age: prevalence of the PROMPT criteria in the PCRS Introduction Methods Study design and setting Study population Ethical approval Data collection Outcomes Statistical analysis Results Descriptive statistics Prevalence of overall PIP and criterion-specific PIP Factors associated with PIP Discussion Principal findings in the context of previous research Comparison with analysis in Northern Ireland Practice and policy implications Strengths and limitations Conclusions Chapter 7 Prevalence of PIP and its association with adverse outcomes in a cohort of community-dwelling middle-aged people from TILDA Introduction Methods Study design and setting Data collection Ethical approval Prevalence of PIP Association with patient outcomes Results Participants Prevalence of PIP and change over time Association with patient outcomes Discussion Principal findings

9 7.4.2 Findings in the context of previous research Practice and policy implications Strengths and limitations Conclusions Chapter 8 Conclusions Introduction Empirical findings Implications and recommendations Trends and prescribing monitoring The challenges of optimising prescribing Efficiency of medicines optimisation The economics of medicines optimisation Role for pharmacists Limitations Proposals for future research Impact Research impact Policy impact Service impact Societal impact Conclusions References Appendices Appendix I: PIP criteria included in analysis of TILDA participants aged 65 years and reasons for exclusion Appendix II: Overlap between PIM and PPO criteria from different screening tools Appendix III: Prevalence of PIP in TILDA participants aged 65 years Appendix IV: Descriptions of Markov model inputs Transition probabilities Costs Utilities Appendix V: Approaches used to specify distributions used in probabilistic sensitivity analysis Probability parameters Relative risk parameters Cost parameters

10 Utility parameters Appendix VI: Structures for benzodiazepine and PPI models in TreeAge Pro Appendix VII: Detailed results of economic evaluation Appendix VIII: Probabilistic sensitivity analysis ICER plot using alternative NSAID scenario Appendix IX: Comparison of linked middle-aged TILDA participants

11 List of abbreviations A&E AC ACE ACOVE ADE ADR AIC AP ARB ATC BIC BNF CCB CDSS CE CES-D CI CNS COPD COX CPI CPRD CS CSO CV DDD DDI DRG ED EHB EHR EPD EU EVPI FDA GDP GEE GI GMS Accident and emergency Anticoagulant drug Angiotensin converting enzyme Assessing Care Of Vulnerable Elders Adverse drug event Adverse drug reaction Akaike information criterion Antiplatelet drug Angiotensin II receptor blocker Anatomical Therapeutic Chemical Bayesian information criterion British National Formulary Calcium channel blocker Computerised decision support system Cost-effectiveness Centre for Epidemiological Studies Depression scale Confidence interval Central nervous system Chronic obstructive pulmonary disease Cyclo-oxygenase Consumer Price Index Clinical Practice Research Datalink Corticosteroid Central Statistics Office Cardiovascular Defined daily dosage Drug-drug interactions Diagnosis-related group Emergency department Eastern Health Board Electronic health record Enhanced Prescribing Database European Union Expected value of perfect information Food & Drug Administration Gross Domestic Product Generalised estimating equations Gastrointestinal General Medical Services 9

12 GORD GP H2 H2RA HARM HEDIS HF HIQA HRT HSE HYT ICER IHD IHI IPET IPW IQR IRR LOS LR LTI LY MAI MAOI MI MMP MSM NCPE NHS NICE NMIC NORGEP NSAID OPD OR PCRS PIM PIP PMR PPI PPO Gastro-oesophageal reflux disease General practioner Histamine-2 receptor H2 receptor antagonists Hospital Admission Related to Medication Health plan Employer Data and Information Set Heart failure Health Information and Quality Authority Hormone replacement therapy Health Service Executive Hypertension Incremental cost-effectiveness ratio Ischaemic heart disease Individual health identifier Improving Prescribing in the Elderly Tool Inverse probability weight Interquartile range Incident rate ratio Length of stay Likelihood ratio Long Term Illness Life year Medication Appropriateness Index Monoamine oxidase inhibitors Myocardial infarction Medicines Management Programme Marginal structural model National Centre for Pharmacoeconomics National Health Service National Institute of Health and Care Excellence National Medicines Information Centre Norwegian General Practice Non-steroidal anti-inflammatory drug Outpatient department Odds ratios Primary Care Reimbursement Service Potentially inappropriate medication Potentially inappropriate prescribing Patient medical record Proton pump inhibitor Potential prescribing omission 10

13 PROMPT PSA PUD QALY QOF QoL RAND RCSI RCT RR SD SES SSRI START STOPP STROBE T2DM TCA TCD TIA TILDA TTO UK US VAS VE WHO WTP ZE PRescribing Optimally in Middle-aged People s Treatments Probabilistic sensitivity analysis Peptic ulcer disease Quality-adjusted life year Quality and Outcomes Framework Quality of life Research ANd Development Royal College of Surgeons in Ireland Randomised controlled trial Relative risks Standard deviation Socioeconomic status Selective serotonin reuptake inhibitor Screening Tool to Alert doctors to Right Treatment Screening Tool for Older Persons (potentially inappropriate) Prescriptions STrengthening the Reporting of Observational studies in Epidemiology Type 2 diabetes mellitus Tricyclic antidepressant Trinity College Dublin Transient ischaemic attack The Irish Longitudinal Study on Ageing Time trade off United Kingdom United States Visual analogue scale Vulnerable elder World Health Organisation Willingness to pay Zollinger-Ellison 11

14 Table of figures Figure 1-1 Proportion of total population by age group for EU-28 in 2014 (bordered) and projected values for 2080 (solid) Figure 1-2 Example of the impact of different approaches to confounder control, X representing pathways blocked by different approaches Figure 2-1 Percentage of eligible population by number of regular medications for the years Figure 2-2 Standardised rates of prescribing of most common regular medications in all individuals in Figure 2-3 Prevalence of most common types of PIP in individuals aged 65 years Figure 2-4 Overall prevalence of PIP using STOPP by number of criteria (left) and prevalence in sensitivity analysis excluding PPIs at maximal dosage for greater than eight weeks (right) Figure 3-1 Prescribing of PPIs and ulcerogenic medicines to GMS-eligible population of the EHB region of Ireland aged 65 years in 1997, 2002, 2007 and Figure 3-2 Adjusted odds ratios with 95% CIs for long-term PPI at maximal dose relative to maintenance dose for various factors stratified by study year Figure 4-1 Flow diagram of study participants from TILDA cohort aged 65 years Figure 4-2 Distribution of number of ED visits in previous 12 months reported by participants at follow-up interview Figure 4-3 Distribution of number of GP visits in previous 12 months reported by participants at follow-up interview Figure 4-4 Distribution of participant CASP-R12 scores at follow-up Figure 5-1 Illustrative example of a Markov model structure Figure 5-2 State transitions of NSAID Markov model Figure 5-3 State transitions of benzodiazepine Markov model Figure 5-4 State transitions of PPI Markov model Figure 5-5 Decision tree structure for NSAID Markov model in TreeAge Pro Figure 5-6 Cost-effectiveness plane Figure 5-7 Decision tree structure of hypothetical intervention analysis Figure 5-8 Incremental costs and utilities for PIP compared to non-pip from probabilistic sensitivity analysis for each model (northwest quadrant) Figure 5-9 Two-way sensitivity analyses of intervention cost and effectiveness at willingness-to-pay threshold of 45,000 per QALY for a) benzodiazepines, b) PPIs, and c) NSAID models Figure 5-10 Threshold effectiveness value for NSAID intervention at intervention cost of 500 and willingness-to-pay threshold of 45,000 per QALY

15 Figure 6-1 Prevalence of PROMPT criteria (dark blue) and relevant drug or disease prevalence (light blue, dashed border) as a proportion of all study participants Figure 7-1 Flow diagram of study participants from TILDA cohort aged years Figure 7-2 Distribution of number of ED visits in previous 12 months reported by participants at follow-up interview Figure 7-3 Distribution of number of GP visits in previous 12 months reported by participants at follow-up interview Figure 7-4 Distribution of participant CASP-R12 scores at follow-up Figure A-1 Decision tree structure for benzodiazepine Markov model in TreeAge Pro Figure A-2 Decision tree structure for PPI Markov model in TreeAge Pro Figure A-3 Incremental costs and utilities for PIP compared to non-pip from probabilistic sensitivity analysis using alternative NSAID scenario

16 Table of tables Table 1-1 Terms and definitions relating to potentially inappropriate prescribing Table 2-1 Prevalence of polypharmacy (being dispensed 5 regular medications) and excessive polypharmacy ( 10 regular medications) in each study year by age group Table 2-2 Adjusted negative binomial regression models for polypharmacy and excessive polypharmacy in all individuals Table 2-3 Overall prevalence of PIP in individuals aged 65 years in each study year Table 2-4 Number and prevalence (%) of individual STOPP criteria in individuals aged 65 years in each study year Table 2-5 Unadjusted and adjusted logistic regression models for having any potentially inappropriate prescribing (PIP) criteria in individuals aged 65 years Table 2-6 Adjusted odds ratios for interaction between level of polypharmacy and study year in individuals aged 65 years Table 3-1 Number and percentage (95% CI) of individuals prescribed a PPI across study years, categorised by duration of use, dosage and concurrent medications use Table 3-2 Unadjusted and adjusted odds ratios and 95% CIs for factors associated with maximal dose compared to maintenance dose in long-term PPI users Table 4-1 Description of covariates adjusted for in multivariate regression models Table 4-2 Descriptive statistics for participants at baseline (Wave 1) and follow-up (Wave 2) Table 4-3 Number and percentage of participants with PIMs and PPOs at baseline and two year follow-up Table 4-4 Prevalence of individual PIM criteria (prevalence 2%) at baseline and two year follow-up Table 4-5 Prevalence of individual PPO criteria (prevalence 2%) at baseline and two year follow-up Table 4-6 Population-averaged GEE models for change in sample prevalence of PIMs and PPOs Table 4-7 Number (percentage) with an ED visit and median (IQR) GP visits the 12 months preceding follow-up by subgroup and adjusted incident rate ratios (95% CI) for ED visits (n=1,748) and GP visits (n=1,741) Table 4-8 Number (percentage) with an increase in ADL difficulties (functional decline) between baseline and follow-up by subgroup and adjusted odds ratios (95% CI) for functional decline (n=1,753) Table 4-9 Mean (SD) of CASP-R12 quality of life score at follow-up by subgroup and adjusted β coefficient (95% CI) for CASP-R12 score (n=986) Table 4-10 Sensitivity analysis comparing parameter estimates with 95% CIs by outcome for multivariate regression and marginal structural models (MSMs) Table 5-1 Description of included PIP criteria from STOPP

17 Table 5-2 Point estimates for each parameter input and distributions used in probabilistic sensitivity analysis Table 5-3 Cost, effect, and ICER outputs for PIP compared to non-pip scenarios for each model Table 5-4 Threshold values across willingness-to-pay thresholds for (i) intervention effectiveness at an intervention cost of 500 per person and (ii) intervention cost at levels of effectiveness from published trials a Table 6-1 Descriptive statistics for included participants Table 6-2 Summary of PIP prevalence and most common criteria Table 6-3 Number of individuals with and prevalence (%) of all PROMPT criteria Table 6-4 Unadjusted and adjusted odds ratios and 95% CIs for factors associated with having any PROMPT criterion compared to none (adjusted only for variables shown) Table 7-1 Description of covariates adjusted for in multivariate regression models Table 7-2 Descriptive statistics for participants at baseline (Wave 1) and follow-up (Wave 2) Table 7-3 Number and percentage of participants with each PROMPT criterion and change in percentage prevalence between baseline and two year follow up Table 7-4 Population-averaged GEE models for change in prevalence of the PROMPT criteria Table 7-5 Number and percentage of participants with an ED visit and unadjusted and adjusted incident rate ratios and 95% CIs for rate of ED visits (n=806) Table 7-6 Median (IQR) GP visits and unadjusted and adjusted incident rate ratios and 95% CIs for rate of GP visits (n=806) Table 7-7 Mean (SD) CASP-R12 score at follow up and unadjusted and adjusted β coefficient with 95% CIs for CASP-R12 score at follow up (n=524) Table 7-8 Sensitivity analysis comparing parameter estimates with 95% CIs by outcome for multivariate regression and marginal structural models (MSMs) Table A-1 Prevalence of all PIM criteria and PPO criteria at baseline and two year followup Table A-2 Full cost, effect, and ICER results for each model for PIP scenarios relative to non-pip scenarios Table A-3 Number of adverse events for PIP and non-pip scenarios Table A-4 One way deterministic sensitivity analysis results Table A-5 Baseline characteristics of TILDA participants aged years who had linked medications data, those non-linked and those non-linked who were GMS eligible

18 Summary Several measures of potentially inappropriate prescribing (PIP) exist, however their validity has been under-researched. The aim of this thesis was to assess measures of PIP in older and middle-aged people in primary care in terms of their applicability and relevance in Ireland, effect on patient outcomes and economic impact. This thesis focussed on community-dwelling adults in Ireland, aged 65 years (older adults) or years (middle-aged adults). Measures of PIP, the Screening Tool for Older Persons Prescriptions (STOPP), the Screening Tool to Alter doctors to Right Treatment (START), and PRescribing Optimally in Middle-aged People s Treatments (PROMPT) criteria, were applied to two national data sources, the General Medical Services (GMS) scheme dispensing database and The Irish Longitudinal Study on Ageing. Economic analysis was also conducted by developing Markov models of PIP. The prevalence of PIP in older people rose from 1997 to 2012 (32.6%-37.3%), though the odds of having PIP decreased over time after accounting for the increase in medications prescribed. Long-term prescribing of maximal dose proton pump inhibitors grew sharply and was not consistently associated with expected risk factors for gastrointestinal bleeding. PIP was present in 41.6% of middle-aged GMS patients, with prevalent criteria similar to those in older people. For older adults, having 2 STOPP criteria was significantly associated with higher rates of emergency department and GP visits, while having 2 START omissions was also associated with increased healthcare utilisation, functional decline and reduced quality of life (QoL). In middle-aged people, there was no evidence of a relationship between PROMPT criteria and healthcare utilisation or QoL after controlling for confounders. Of the three PIP criteria evaluated relative to appropriate alternatives in Markov models, long-term benzodiazepine prescribing had the greatest cost and quality-adjusted life year impact, although long-term non-steroidal anti-inflammatory drug use was the most costeffective PIP to target. This thesis demonstrates that PIP is prevalent and can impact on patient and economic outcomes. Optimising prescribing to reduce PIP may provide benefits for patients and the wider health system. 16

19 Acknowledgements Firstly I wish to acknowledge the Health Research Board in Ireland (HRB) for supporting this PhD through the HRB PhD Scholars Programme in Health Services Research (grant no. PHD/2007/16). I would also like to acknowledge the Primary Care Reimbursement Service for providing the administrative pharmacy claims data used in this work and the staff and participants of The Irish Longitudinal Study on Ageing (TILDA) who kindly gave their time. Thank you to my supervisors, to Professor Tom Fahey for his mentorship, insightful clinical input and for making me feel at home at the HRB Centre for Primary Care Research; to Professor Kathleen Bennett for her always sound methodological advice and for making me a better scientist; and to Dr Caitriona Cahir for her patience and guidance throughout this research. I am indebted to Dr Helen Sheridan and Dr Catriona Bradley, for both their encouragement to pursue this course of study and their sage counsel along the way. Thank you to Dr Rose Galvin for her mentorship which yielded both my first peerreviewed publication and a valued friendship. I must thank my co-authors on publications arising from this thesis, Professor Susan Smith, Professor Rose Anne Kenny, Professor Carmel Hughes, and especially Colin Hardy and Dr Janine Cooper for their close collaboration. I am also grateful to all the staff at the HRB Centre for Primary Care Research and the SPHeRE programme for their assistance, and to Professor Rachel Elliott for hosting me on my international placement at the University of Nottingham. Thanks to my colleagues on the SPHeRE programme at RCSI for their friendship and camaraderie: Mary-Ann, Nora, Eithne, Sara, Lisa, Yvonne, Emma, Áine, Mark, Patrick, Amelia, and Mary. Particular thanks must go to Barbara, who having trod a similar doctoral path always provided a sympathetic ear and sound advice, especially when I wasn t braining good. Thanks also to Ronan, for being a friend and great company on the same road for the last 10 years chapeau! I am grateful to all my friends, in particular Michelle and Cliodhna, for listening and for the laughs. Lastly, thanks to my parents, Irene and Frank words cannot do justice to my appreciation for their constant love, support and encouragement. 17

20 Publications Publications arising from this thesis Moriarty F, Bennett K, Cahir C, Fahey T. Characterising potentially inappropriate prescribing of proton pump inhibitors in older people in primary care in Ireland from 1997 to Journal of the American Geriatrics Society. In press (Chapter 3) Moriarty F, Bennett K, Cahir C, Kenny RA, Fahey T. Potentially inappropriate prescribing according to STOPP and START and adverse outcomes in community-dwelling older people: a prospective cohort study. British Journal of Clinical Pharmacology. 2016; 82(3): (Chapter 4) Cooper JA*, Moriarty F*, Ryan C, Smith SM, Bennett K, Fahey T, Wallace E, Cahir C, Williams D, Teeling M, Hughes CM. Potentially inappropriate prescribing in two populations with differing socio-economic profiles: a cross-sectional database study using the PROMPT criteria. European Journal of Clinical Pharmacology. 2016; 72(5): (Chapter 6) Moriarty F*, Hardy C*, Bennett K, Smith SM, Fahey T. Trends and interaction of polypharmacy and potentially inappropriate prescribing in primary care over 15 years in Ireland: a repeated cross-sectional study. BMJ Open. 2015; 5(9): e (Chapter 2) Moriarty F, Bennett K, Fahey T, Kenny RA, Cahir C. Longitudinal prevalence of potentially inappropriate medicines and potential prescribing omissions in a cohort of communitydwelling older people. European Journal of Clinical Pharmacology. 2015; 71(4): (Chapter 4) *Denotes joint first authorship Related publications Barry E, O'Brien K, Moriarty F, Cooper J, Redmond P, Hughes CM, Bennett K, Fahey T, Smith SM; PIPc Project Steering group. PIPc study: development of indicators of potentially inappropriate prescribing in children (PIPc) in primary care using a modified Delphi technique. BMJ Open. 2016; 6(9): e Cahir C, Moriarty F, Teljeur C, Fahey T, Bennett K. Potentially Inappropriate Prescribing and Vulnerability and Hospitalization in Older Community-Dwelling Patients. Annals of Pharmacotherapy. 2014; 48(12): Galvin R, Moriarty F, Cousins G, Cahir C, Motterlini N, Bradley M, Hughes CM, Bennett K, Smith SM, Fahey T, Kenny RA. Prevalence of potentially inappropriate prescribing and prescribing omissions in older Irish adults: findings from The Irish LongituDinal Study on Ageing study (TILDA). European Journal of Clinical Pharmacology. 2014; 70(5):

21 Conference presentations Oral presentations Moriarty F, Bennett K, Cahir C, Kenny RA, Fahey T. How does potentially inappropriate prescribing measured by STOPP and START relate to healthcare utilisation in older people? A cohort study. Society for Academic Primary Care 45 th Annual Conference, Dublin Castle, 3 rd July Moriarty F, Bennett K, Cahir C, Kenny RA, Fahey T. Potentially inappropriate prescribing and healthcare utilisation in older people: a cohort study using marginal structural models. 21 st Health Services Research and Pharmacy Practice Conference, University of Reading, 7 th April Moriarty F, Cooper J, Bennett K, Cahir C, Hughes CM, Fahey T. Potentially inappropriate prescribing (PIP) in two populations with differing socio-economic profiles: a crosssectional database study using the PROMPT criteria. SPHeRE conference, RCSI, 29 th February Moriarty F, Hardy C, Bennett K, Smith SM, Fahey T. Trends and interaction of potentially inappropriate prescribing and polypharmacy over 15 years in Ireland: a repeated crosssectional study. Society for Social Medicine 59 th Annual Scientific Meeting, University College Dublin, 3 rd September 2015 Moriarty F, Hardy C, Bennett K, Smith SM, Fahey T. The rise and fall of potentially inappropriate prescribing: trends and interaction with polypharmacy over 15 years in Ireland. Society for Academic Primary Care 44 th Annual Conference, University of Oxford, 8 th July 2015 Moriarty F, Bennett K, Cahir C, Kenny RA, Fahey T. Determining the relationship between potentially inappropriate medications and quality of life in a cohort of older adults. International Association of Geriatrics and Gerontology-European Region Congress, Dublin Convention Centre, 24 th April Moriarty F, Hardy C, Bennett K, Smith SM, Fahey T. Trends in polypharmacy and prescribing appropriateness from 1997 to st Health Services Research and Pharmacy Practice Conference, Queen s University Belfast, 16 th April Received Best Oral Presentation award. Moriarty F, Bennett K, Cahir C, Kenny RA, Fahey T. Determining the prevalence between potentially inappropriate medications and quality of life in a cohort of older adults. 1st Annual SPHeRE Network Conference, RCSI, 9 th January Moriarty F, Cahir C, Fahey T, Bennett K. Potentially inappropriate medicines and potential prescribing omissions in older people and their association with health care utilization. North American Primary Care Research Group (NAPCRG) Annual Meeting, 23 rd November

22 Moriarty F. Student Research Showcase - Outcomes research having a high impact on new challenges for improving European health care. International Society for Pharmacoeconomics and Outcomes Research (ISPOR) 17 th Annual European Congress, Amsterdam, 10 th November Poster presentations Moriarty F, Bennett K, Cahir C, Kenny RA, Fahey T. Potentially inappropriate prescribing measured by STOPP and START and functional decline and quality of life in older people: a cohort study. 32 nd International Conference on Pharmacoepidemiology and Therapeutic Risk Management (ICPE), the Convention Centre Dublin, 27 th August Moriarty F, Bennett K, Fahey T, Kenny RA, Cahir C. Prevalence of potentially inappropriate medicines and potential prescribing omissions over time in cohort of a communitydwelling older people. All Ireland Pharmacy Healthcare Conference, Dundalk, th January Moriarty F, Cahir C, Fahey T, Bennett K. Potentially inappropriate medicines and potential prescribing omissions in older people and their association with health care utilization: a retrospective cohort study. ISPOR 17 th Annual European Congress, Amsterdam, th November Received Best Student Poster Research Presentation award Moriarty F, Cahir C, Fahey T, Bennett K. Potentially inappropriate prescribing and its association with Instrumental Activities of Daily Living (IADL) impairment in older people. AUDGPI SEM, University College Cork, 6-7 th March 2014 Moriarty F, Cahir C, Fahey T, Bennett K. Potentially inappropriate prescribing and its association with Instrumental Activities of Daily Living (IADL) impairment in older people. ISPOR 16th Annual European Congress, Dublin, 4-6 th November Received Best Student Poster Research Presentation award Other oral presentations Moriarty F. Potentially inappropriate prescribing: tools, trends and a target for medicines optimisation. 4 th National Medicines Forum, Royal College of Physicians in Ireland, Dublin, 16 th May Moriarty F. Potentially inappropriate prescribing a good barometer of medication safety? Reducing Medication Errors in Healthcare Services Conference, Royal Marine Hotel, Dun Laoghaire, 26 th June

23 Chapter 1 Introduction 21

24 1.1 Introduction This thesis is focussed on the evaluation of measures of potentially inappropriate prescribing (PIP) in older people (aged 65 years and over) and middle-aged people (aged between 45 and 64 years) in Ireland in terms of their applicability and relevance, effect on patient outcomes, and economic impact. This chapter provides background information on why demographic changes and complexity of medication use in older and middle-aged people are of importance in Section 1.2, and in Section 1.3, a discussion of what PIP is, how it can be measured, and the optimal characteristics of measures of PIP. Some of the published research that has been conducted to date in older people is reviewed in Section 1.4 followed by a discussion of the lack of prescribing research in the middle-aged population in Section 1.5. The aims and objectives of this thesis are then presented in Section 1.6, followed by a brief description of the health service context for this research (Section 1.7), rationale for the methods used (Section 1.8) and an outline of the structure of the rest of the thesis (Section 1.9). 1.2 Background Demographic changes One of the major challenges facing Ireland and Europe in general in the coming decades is the ageing population. The proportion of the population composed of people aged 65 years or older is growing substantially.[1] The European commission projects that by 2025 more than 20% of Europeans will be 65 or over, and by 2060, 28% of the population of the 28 European Union (EU) member states will be aged 65 years and over (up from 18% in 2014) and those aged 80 years and over will rise from 5% to 12%.[1] This is illustrated in Figure 1-1 below taken from Eurostat,[2] which shows the distribution of age groups in 2014 being pyramid shaped but this is expected to become column shaped by 2080, with slight narrowing in the middle age range of years and a large growth in those aged 85 years and over. This highlights the progressive ageing within the older population itself, as the very old group is growing at a faster rate than any other age segment of the EU s population, primarily due to extended life expectancy.[1] This poses a challenge with respect to expenditure on health due to increasing demand for healthcare resources. At an EU level, the increase in spending on healthcare due to the effect of an ageing population over the next fifty years is projected to be 1.1% of 22

25 gross domestic profit (GDP), i.e. from 6.9% to 8% of GDP from 2013 to 2060.[1] With respect to Ireland, the projected rise in spending attributable to demographic changes was 1.3% of GDP, or up to 1.9% of GDP if non-demographic cost drivers such as technological advancement and the development of new treatments are included. Medications are a particular area of focus given they are the most common healthcare intervention in the developed world,[3] and because older adults have the highest levels of prescribing.[4] Figure 1-1 Proportion of total population by age group for EU-28 in 2014 (bordered) and projected values for 2080 (solid) Spending on pharmaceuticals was identified as a specific area of focus in a 2009 report projecting the impact of the ageing population on future demand for healthcare in Ireland.[5] Public spending on pharmaceuticals in 2006 was 1.86 billion, approximately 15% of all health expenditure. Due to the higher number of prescribed items and average cost of medicines in older patients, demographic changes were projected to lead to an increase in prescription items from 54 million in 2006 to between 75 and 100 million in 2020, with total drug ingredient costs growing by between 40% and 100%.[5] Given the magnitude of medications use and spending, which in 2014 was still 1.8 billion or 14% of state health expenditure,[6] efforts to optimise prescribing to maximise the benefits and 23

26 minimise harms may not only improve patient safety but also help to control healthcare costs. With advancing age tends to come increased use of medications, and this along with other factors in ageing can make prescribing complex, and some of these challenges are outlined below Challenges of prescribing in older adults Underrepresentation in trials Older adults carry the majority of disease burden, however, it has been reported that only 32% of patients in phase II and III clinical trials are aged 65 years and over.[7] This implies that medicines being prescribed have not necessarily been effectively evaluated in this age group,[8] and, therefore, evidence for the effectiveness and/or safety of treatments in older people may be deficient. Poor representation of this group can be due to explicit exclusion by age at the level of eligibility criteria, or by excluding those with comorbidities or cognitive impairment.[8] Other factors which may contribute are the combination of obstacles faced by older adults, such as economic barriers, communication issues (i.e. impaired vision or hearing difficulties which can preclude engagement with written surveys and telephone interviews respectively) and physical limitations that may limit transport options.[7] There tends to be low participation rates in trials of a number of age-related disease areas, including Alzheimer's disease, arthritis, incontinence, and cardiovascular disease.[7] As a consequence, results of trials have limited generalisability to this age group and uncertainty regarding risks and benefits of treatments can delay the use of new treatments in older people.[7,8] Often, clinicians have to rely on evidence from observational research to inform medication treatment decisions in these patients or extrapolate from trials of younger participants which may not be applicable Age-related physiological changes Natural changes in physiology as people age can impact on prescribing. These can relate to either pharmacodynamics, what a drug does to the body, or pharmacokinetics, what the body does to a drug.[9] Pharmacodynamic changes arise due to alterations in drug concentrations at the receptor they act on, drug-receptor interactions (resulting from variations in receptor density or affinity, and cellular response), and the body s regulation of homeostatic equilibrium.[10] This most often gives rise to an increased sensitivity to medications, for example enhanced sedative effects of central nervous system medications due to higher drug concentrations in the brain,[11,12] or hypotension with 24

27 antihypertensive drugs owing to impaired homeostatic counterregulation.[12] However in some cases pharmacodynamic changes can reduce responsiveness, as occurs with β agonists and antagonists possibly due to downregulation of β adrenoreceptors in older age or cell signalling changes which decrease the sensitivity of the heart and lungs to these agents.[9,12] Pharmacodynamic changes in ageing are not well understood and are often difficult to separate from pharmacokinetic effects.[10] Pharmacokinetic changes are better characterised and primarily affect how a drug is absorbed, distributed, metabolised and excreted by the body. For example, absorption of calcium and iron may be reduced in older aged individuals, predisposing them to osteoporosis and anaemia secondary to deficiencies in these minerals.[9] Of greatest concern in older adults is that these changes can result in prolonged drug half-life (the time taken for the concentration of a medication to be halved due to metabolism or excretion), increased duration of action or accumulation of a medication, all of which increase the risk of adverse events. Ageing is associated with a decrease in total body water leading to higher plasma concentrations of water-soluble drugs such as lithium and digoxin.[9,13] In tandem with this fall in lean body mass is an increase in body fat which prolongs the action of lipophilic drugs including diazepam.[9] Conversely in the very old frail patient, loss of body fat can reduce the duration of action but increase the plasma concentration of such drugs.[12] Many drugs bind to plasma proteins such as albumin in the blood, and, although albumin production may be reduced due to ageing or specific comorbidities, increasing the amount of active drugs such as naproxen or theophylline circulating, this is likely to be of limited clinical significance.[9,12] Considering drug metabolism, this primarily occurs in the liver where hepatic mass and blood flow can decrease with age.[14] Drugs that are highly extracted by the liver are most likely to be affected by changes in liver blood flow, and the implications of this are twofold. Firstly, some drugs undergo first-pass metabolism after absorption from the intestine and before reaching the bloodstream, so if this is impaired, elevated plasma concentrations can result as occurs with propranolol and labetalol.[13] Secondly, drugs in the bloodstream are often metabolised by the liver to convert them into an inactive, less active, or more easily eliminated form. Again, if metabolism is reduced circulating levels of drugs such as amitriptyline and theophylline can rise and their duration of action may be prolonged.[12] 25

28 Lastly, elimination or excretion of drugs can be impaired as the kidney may be impacted by physiological changes in ageing or alterations due to diagnoses such as hypertension and diabetes.[11,12] Reduction in glomerular filtration rate, indicating the volume of blood filtered by the kidneys, can decrease the rate of elimination of drugs which are excreted in their active form such as metformin, digoxin, and angiotensin converting enzyme (ACE) inhibitors.[9,14] Renal impairment is one of the few pharmacokinetic changes which can be easily detected and quantified allowing for reduction in dose or dosing frequency to avoid potential adverse effects, however such dosing changes are not consistently carried out in older aged patients and other pharmacokinetics changes are more difficult to anticipate and avoid.[12] Burden of morbidity and medications use In addition to these intrinsic changes, older patients are more likely to have a high illness and medication burden. Advances in the treatment of many conditions in recent decades has increased survival following events and infections that previously had high mortality, thus extending life expectancy.[15] The net result of this is people are now living for longer with chronic disease, as illustrated in a Dutch study where the prevalence doubled from 361 chronic disease registrations per 1,000 person years in 1985 to 764 per 1,000 person years in 2005.[16] This has led to an increasing burden of multimorbidity, the coexistence of two or more chronic conditions.[16] Multimorbidity is strongly associated with age, with prevalence estimates ranging from 55% to 98% among adults aged 65 years and over.[15 17] It is also strongly linked to socioeconomic deprivation, so that in deprived areas the level of multimorbidity may be higher than expected given the age profile.[17,18] Despite multimorbidity being an increasingly common feature of healthcare for older individuals, typically clinical management still tends to focus on single diseases.[19] This is exemplified by the case of clinical guidelines which outline effective management of individual conditions but recommendations do not take account of other comorbidities, therefore conflicts may arise leading to drug-disease interactions.[20 22] For example, for an individual with both chronic obstructive pulmonary disease (COPD) and coronary artery disease following a myocardial infarction (MI), recommended therapy of a β adrenoreceptor blocker for the latter may increase the risk of bronchoconstriction due to the former.[23] An analysis of National Institute of Health and Care Excellence (NICE) clinical guidelines found that for treatments recommended for type 2 diabetes mellitus, there were 32 potentially serious drug-drug interactions with 26

29 medications included in the guidelines for 11 other conditions, including heart failure and chronic kidney disease.[24] People with multimorbidity are at higher risk of functional decline, increased healthcare utilisation, and reduced quality of life (QoL).[21] They are also likely to be prescribed multiple medications and such complex drug regimens carry their own risks.[4,21] The rising number of medications being prescribed in recent decades has been well characterised,[25 27] and old age is one of the most important predictors of multiple medication use or polypharmacy.[4,28] Polypharmacy is most often classified as the use of five or more medicines, although there is no universal definition.[29] Polypharmacy in older patients has been a source of concern for healthcare professionals and was thought of as a synonym for overprescribing and an indicator of suboptimal prescribing.[30] However, polypharmacy may be entirely appropriate, particularly in the presence of multimorbidity,[29,31] and should not be a reason to avoid prescribing additional beneficial treatments.[32,33] Despite the potential benefits, such complex drug regimens also pose a number of risks including drug-drug or drug-disease interactions,[15,34,35] unplanned hospital admissions,[31] adverse drug events (ADEs),[36,37] and poor adherence.[37] Older people are particularly vulnerable to these risks which may be propagated by existing physiological changes in pharmacodynamics and pharmacokinetics.[9,34] Conversely, drug-drug interactions (DDIs) can also impact on pharmacokinetics and pharmacodynamics. For example, one drug may inhibit or induce the metabolism of another (pharmacokinetic effects), or two drugs acting on the same receptor could lead to either a synergistic effect or competitive antagonism which reduces effectiveness (pharmacodynamic effects).[34] Common prescribing challenges in middle-aged adults Although the major demographic challenge in the coming years is the growth in the number of people aged 65 years and over, consideration should also be given to middleaged people (45-64 years). The numbers in this age group are projected to decrease slightly although will remain fairly constant and this will also be the source population for the older and very old adults in the coming decades. However, this demographic group have seldom been the specific population of interest in studies of medicines optimisation. There are clear reasons why the focus of such research has been primarily on patients at older ages as outlined in Section However, a number of these factors may also 27

30 apply to individuals aged between 45 and 64 years, as discussed below, and complexity of medication prescribing in this age group is often overlooked. Physiological changes in ageing are often cited as the main reason for the vulnerability of older people to medication-related adverse events. Although 65 years is used as an arbitrary cut-off to indicate entry into older age, the changes in pharmacokinetics and pharmacodynamics do not occur at a single point in time.[38] Rather, these develop on a continuum and the decline in some physiological functioning can begin at a relatively early age.[14] Regarding kidney function, renal blood flow reduces by 1% per year from the age of 50 years.[14] The number of functioning glomeruli decreases leading to a steady decline in glomerular filtration rate from the age of 30 years by between 0.4 and 1ml/minute per year,[11] and kidney mass declines by between one fifth and one quarter between the ages of 30 years and 80 years.[11] These changes can reduce the clearance of renallyeliminated drugs and their metabolites such as lithium, digoxin, hydrochlorothiazide, oxycodone, ACE inhibitors and metformin, increasing the risk of adverse effects.[14] Hepatic changes which may influence pharmacokinetics include a decrease in liver mass, and consequently the total amount of drug-metabolising enzymes, and reduced liver blood flow.[14] Enzyme activity itself does not appear to be age-related,[13] however, enzyme-responsiveness may vary, for example with younger people possibly being more susceptible to the enzyme inductive effect of cigarette smoking which may increase theophylline metabolism.[9] How these changes progress with age is less well characterised than in the case of renal function but they appear to be moderate from middle age onwards as routine tests of liver function do not show significant differences between those aged and years.[9] This may result in reduced first-pass metabolism of drugs such as β blockers and tricyclic antidepressants leading to increased bioavailability and more rapid onset of action.[14] Pharmacodynamic changes in ageing have received less attention, although cardiovascular and central nervous systems have been investigated to some extent.[14] For instance neuronal loss and a 20% reduction in brain weight occur from the age of 20 to 80 years,[12] and this coupled with changes in receptor density and neurotransmitter production lead to increased sensitivity to the effect of benzodiazepines and opioids.[10] 28

31 However, research on such pharmacodynamic changes does not indicate at what age these changes may become clinically significant. The number of chronic illnesses and multiple medications are contributors to the risk of adverse medication effects, but again, these are not exclusively the domain of older patients. Although the prevalence of multimorbidity is lower in middle-aged people, this population may contain the highest number of people with multimorbidity in absolute terms, as was found in Scottish primary care.[17] In addition, as number of medications prescribed is strongly associated with multimorbidity, polypharmacy is also a consideration in middle-aged people. A Swedish population study found that in 2008 the number of individuals on five or more medicines was only slightly lower in those aged years than in those 70 years or over.[27] A comparable number of people with polypharmacy appeared to fall into each of these age categories in an analysis in Scottish primary care.[4] The use of multiple medications can be associated with increased risk of high-risk prescribing, adverse events, and drug-drug interactions.[28,39,40] Such interactions can alter pharmacokinetics, for example due to receptor site competition or metabolic enzyme inhibition.[11] Although these may be amplified by age-related changes, pharmacokinetic interactions are principally linked with the number of pharmacotherapies.[34,40] 1.3 Potentially inappropriate prescribing Due to the limited generalisability of evidence from trials for prescribing in older age as discussed in Section , along with the complexity of physiological changes, multimorbidity and polypharmacy, concerns regarding quality of prescribing and medication safety in older patients have led to a focus on the concept of PIP. Despite these factors also being a consideration in middle-aged people, this age group has to date not been the subject of studies of PIP. PIP can be defined as suboptimal medicines use relating to overprescribing, misprescribing, or underprescribing.[41] Firstly, overprescribing refers to the use of a medication which is not clinically indicated. Despite a lack of clinical trial evidence in older age groups, evidence in younger populations and from observational research can be used to determine that there is a lack of efficacy for an indication, such as a loop diuretic as first line management for ankle oedema.[42] Secondly, misprescribing refers to the use of a medication that is clinically indicated but in a way in which the risks are likely to 29

32 outweigh the benefits. This may be due to prescription of too high a dose, such as digoxin at greater than 125 micrograms, which can lead to accumulation and adverse effects due to reduced renal function and increased susceptibility to the therapeutic effects.[43,44] Misprescribing also includes prescriptions of too long a duration, for example, long-acting benzodiazepines for greater than four weeks where dependence and tolerability are likely to develop beyond this and sedative effects increase the risk of falls.[45,46] Misprescribing can also be due to inappropriate combinations with another drug, such as a member of the same drug class leading to duplication or a different drug class like warfarin with an non-steroidal anti-inflammatory drug (NSAID). Combination of a drug with a diagnosis, such as non-cardioselective β blocker use in patients with COPD as mentioned earlier, can also lead to an interaction. Instances where a medication is used but an alternative with a superior risk-benefit ratio exists can also be considered misprescribing. Thirdly, underprescribing refers to the omission of a clinically indicated medication. This may be a primary treatment such as metformin in type 2 diabetes mellitus, preventive treatment such as bisphosphonate therapy in patients taking oral corticosteroids long-term, or symptomatic treatment for example a laxative in patients requiring strong opioids.[41] Overprescribing and misprescribing relate to the use of potentially inappropriate medications (PIMs) and underprescribing concerns potential prescribing omissions (PPOs). The relationship between these terms and definitions are summarised in Table 1-1. Of the three categories, overprescribing and misprescribing tend to receive the greatest attention, as these acts of commission align better with the concept that inappropriate prescribing only encompasses the use of too many medicines.[47] Underprescribing, an act of omission, may be neglected when focusing on medicines optimisation.[32] The evidence on what constitutes overprescribing and misprescribing may be more equivocal than that for underprescribing,[8] and due to omission bias, doctors may be more concerned regarding harm resulting from an act of commission than an act of omission.[48] Starting a new medicine is a more complex decision which may be affected by patient preferences and prescriber fears of adverse consequences to a greater extent than stopping a medication with an active risk of harm.[33] This is reflected in the maxim which is included in the ethical principles of healthcare, primum non nocere (first, do no harm). 30

33 Table 1-1 Terms and definitions relating to potentially inappropriate prescribing Potentially inappropriate prescribing Acts of commission Acts of omission Potentially inappropriate medications Potential prescribing omissions Overprescribing The use of a medication which lacks an indication Misprescribing The use of an indicated medication such that the risks outweigh the benefits or a better alternative exists Underprescribing The omission of a clinically indicated medication Measurement of PIP Process or outcome measures of quality of care PIP can be considered as an indicator of quality of prescribing.[41] According to the model proposed by Donabedian, quality of care can be assessed on the basis of structures, processes and outcomes of care.[49,50] Measurement of outcomes, such as mortality or adverse events, offers a number of advantages, including improved construct validity (i.e. they tend to measure things that are of importance), applicability across healthcare settings and countries, and the ability to measure many outcomes routinely and with precision.[51,52] There are also limitations, however, such as that outcomes may be affected by a number of factors outside of the control of healthcare professionals and providers, and so outcome measures may not specifically assess the quality of care being delivered.[51] Also, it may take considerable time for outcomes to manifest or for a sufficient number of rare outcomes to occur to allow measurement, and in the case of some outcomes such as patient attitudes and rehabilitation, they may be difficult to measure.[51] Assessing process of care, i.e. whether what is known to be good care has been applied, is an alternative which requires clear specification of what values and standards constitute good care that will be assessed.[49] Examples of process measures include use of recommended drug therapy or time to admission following emergency department presentation. Measuring processes rather than outcomes gets closer to ascertaining whether the actions undertaken and care provided are appropriate,[49] and as such, providers can feel more accountable for these and if integrated into routine practice can 31

34 provide useful feedback at the point of care.[53] Measurement can also occur at or closer to the point of care, eliminating the time lag between actions and effects which outcome measures can be subject to.[51] Relevant processes should, however, have a causal relationship with outcomes, that is a change in the process should result in a change in an important outcome independent of other factors.[53] This should be demonstrated empirically as although expert consensus is widely accepted, it is not a strong form of evidence to support a link between process and outcome.[51] Even in cases when studies have provided evidence, a true process-outcome effect may not have been demonstrated due to methodological flaws including lack of a temporal relationship between the two,[41] or failure to adequately control for confounding.[51] Although in some cases medicines use can be evaluated by outcome measures of quality of care, for example A patient admitted to hospital for a fall and taking a long-acting benzodiazepine indicates that the benzodiazepine prescription is inappropriate,[41] the act of prescribing is more commonly assessed by process measures of quality of care Implicit versus explicit measures For both process and outcomes measures, PIP can be determined either implicitly or explicitly.[41] Implicit or judgement-based assessment involves a prescriber using information about the patient alongside published research to evaluate the suitability and appropriateness of a medicine in the context of that individual, their circumstances and preferences. Alternatively, explicit or criteria-based approaches involve the use of sets of indicators, which are typically drug- or disease-orientated, and specify circumstances in which prescribing particular medications may be inappropriate. Although implicit judgement of appropriateness is by its very nature non-prescriptive, a structured measure using an implicit approach has been developed, the Medication Appropriateness Index (MAI).[54] It is not drug specific and lists 10 attributes of a treatment, for example, indication, dose, and interactions, used to appraise the overall appropriateness of a medication in that instance. While some of the attributes require a clinician s judgement, such as effectiveness and practicality of directions, operational definitions and explicit instructions are included.[55] This may help to standardise rating and improve the validity, reproducibility and generalisability of this implicit measure.[41] The rating of attributes produces an overall weighted score or index and the MAI has been shown to have good inter- and intra-rater reliability as well as content and face 32

35 validity.[54,56,57] However it can be time-consuming to apply, requiring an experienced clinician with access to detailed clinical information to consider current medications, disease states, and outcome data.[58] This is in contrast to explicit measures which require little or no clinical judgement and can be applied manually to patient medical records (PMRs) or used in conducting medication reviews with patients. An alternative approach and an advantage over implicit assessment is that application of explicit measures can be automated and incorporated into computerised decision support systems (CDSS). Prescribing or dispensing data alone can be sufficient to apply drug-orientated explicit criteria, whereas for disease- and patient-orientated criteria, validated proxies for diagnoses may be required in the absence of clinical information.[59] A disadvantage of explicit criteria is that they cannot incorporate all factors that define quality and so, may not comprehensively identify all medication issues.[41,60] Their development is usually based on reviews of published studies and expert consensus as the evidence base for treatment benefits and harms in older aged patients is often lacking.[8,41] Examples of explicit PIP measures A number of explicit measures of PIP have been developed, the earliest of which was the Beers criteria.[61] It was first published in 1991 and was originally a list of medications to avoid in older residential care patients in the United States (US).[61] A two-round Delphi survey method was used among 13 national experts in geriatrics and pharmacotherapy and consensus was reached on 30 explicit criteria. At the time it was envisioned that these could be updated in the future and since then a number of revisions have taken place. The first of these in 1997 widened the scope to older people in all care settings and added grading of the likely clinical severity of adverse events.[62] This update also went beyond simply drugs to avoid and included criteria relating to medications which are inappropriate in people with certain medical conditions. The 2003 update continued in this vein and the only change was adding, removing or modifying criteria.[63] The next revision in 2012 differed in a number of ways.[64] Firstly, preceding the consensus panel a systematic review of the literature was carried out in order to inform which criteria to include and to introduce a grading of the level of evidence and strength of recommendation as a replacement for the severity rating. For some criteria, exceptions for clinically appropriate cases were specified (for example based on dosage or duration) 33

36 and a new category of drugs to use with caution in older people was also introduced. The most recently published 2015 update adopted a similar methodological approach and added criteria for DDIs as well as medications to avoid or reduce the dose of at varying levels of renal impairment.[65] The Beers criteria are the most commonly used criteria in the published literature for identifying PIP,[58] however, a limitation is that the early consensus panels included only national experts and hence there was a focus on US prescribing. A number of the drugs included were not available or widely used in other countries, meaning some criteria were effectively redundant which was a barrier to the adoption of Beers into routine geriatric practice outside of the US.[55] Due to these limitations a number of other measures of PIP were developed, including several country-specific criteria. Possibly the most widely adopted of these are the Screening Tool for Older Persons (potentially inappropriate) Prescriptions (STOPP) and the Screening Tool to Alert doctors to Right Treatment (START).[42] They were developed with a focus on currently used and frequently prescribed medications in the European context to allow adoption by prescribers in day-to-day practice. The former includes criteria relating to acts of commission (i.e. overprescribing and misprescribing) while the latter relates to the other aspect of PIP, acts of omission (underprescribing). A Delphi consensus approach, similar to that employed in the development of the Beers criteria, was used and the panel comprised 18 experts from Ireland and the United Kingdom (UK), including consultant geriatricians, clinical pharmacologists, academic general practitioners (GPs) and hospital pharmacists. Consensus was reached after two rounds on 65 STOPP criteria and 22 START criteria, and as well as this content validation, good inter-rater reliability has been demonstrated among doctors and pharmacists.[42,66,67] Advantages of STOPP and START include their comprehensiveness, applicability in different countries, and organisation by physiological system to facilitate rapid application in routine practice.[68] STOPP and START were revised in 2015 by carrying out a Delphi consensus process on the original criteria along with a further 51 new indicators.[69] This resulted in fifteen criteria from the original version being removed and an overall increase from 87 to 114 criteria. Although three of the new criteria introduced are implicit, the tools remain predominantly explicit. The original STOPP and START criteria had been extensively used at the time of publication of the second version, with 74 publications having described 34

37 their use, including 45 original research articles, and these related to 24 different countries illustrating international applicability across not only Europe, but also Asia, Australia, and both North and South America.[69] Owing to the broad literature that exists on the original versions of STOPP and START and the timing of the publication of the updated version, this thesis will focus only on the original editions of these tools. A number of other country-specific measures of PIP have been developed, in some cases through modification of the Beers criteria,[70] or in other cases by using existing criteria as a starting point and adopting a Delphi consensus approach using experts from that geographical area. This approach was used for the Canadian McLeod criteria, the Laroche criteria in France, the Norwegian General Practice (NORGEP) criteria and the PRISCUS list from Germany.[71 74] A different method was used in developing prescribing indicators for older people in Australia.[75] Firstly, the most common medical diagnoses that older Australians seek healthcare for and the most commonly prescribed drug agents in Australia were identified.[75] Then, national drug information, clinical practice guidance, international literature and other evidence sources were used to identify optimal use or significant problems related to those common drugs in people with the common conditions identified. This process yielded 48 prescribing indicators with footnotes providing additional information on definitions included in the indicators. No consensus method was used as the authors judged that consensus had already been achieved in the evidence sources they used. The authors acknowledge that many of the criteria may also be applicable in patients aged less than 65 years but that the conditions and medications included are only those most common in the older population. The Assessing Care Of Vulnerable Elders (ACOVE) indicators were produced by the Research ANd Development (RAND) corporation in the US to assess the quality of medical care of older people,[76] and although not specifically developed to identify PIP, they have since been adopted for this purpose. The third and most recent version contains 392 indicators relating to 26 different medical conditions and is not merely focused on prescribing with the four domains included being screening and prevention, diagnosis, treatment, and follow-up and continuity. A number of indicators relate to medications to be avoided in certain circumstances (i.e. PIMs) and others to medications that should be prescribed (i.e. PPOs). Their suitability as a measure of PIP has been illustrated in a 35

38 number of studies which have also applied more conventional PIP screening tools,[77 79] and a validation study has demonstrated good inter-rater reliability.[80] Although PIP is one aspect of medications safety, other explicit indicators of prescribing have been developed and used in primary care. These include high-risk prescribing indicators from Scotland,[39] preventable drug-related morbidity indicators,[81] and the UK Royal College of General Practitioner indicators,[82] all of which have been used in primary care research.[83 85] As these are primarily focused on prescribing safety and harm rather than (in)appropriateness of prescribing, they were not a focus of this thesis Optimal characteristics of explicit measures Explicit measures of PIP have generally been developed using the same approach, being a review of the literature followed by formulation of criteria usually by Delphi technique. This consensus method does provide face and content validity, however, it is based on expert opinion, does not establish internal validity, and cannot replace evidence provided by scientific evaluation of criteria.[86] There have been few studies which have evaluated other aspects of validity of prescribing indicators.[59,86] Due to the large number of published indicators, it is important to consider what the characteristics an optimal measure of prescribing quality should have. Kaufmann and colleagues suggest that an ideal tool to assess the appropriateness of drug prescriptions should generally: cover all aspects of appropriateness (efficacy, safety, cost-effectiveness and patients preferences), be developed using evidence-based methods, show significant correlation between the degree of inappropriateness and clinical outcomes, and be applicable not only in research conditions but also in daily health care practice.[87] O Mahony and Gallagher in outlining the need for new criteria ahead of their development of STOPP and START made similar suggestions by proposing the following: organisation based on physiological systems and rapid applicability in practice, include the more common errors of commission and omission in prescribing for older adults, 36

39 generalisability to the global community of doctors and pharmacists, ease of interface with computer records of patients co-morbidities and drugs, ability to reduce the prevalence of PIP in older adults across clinical settings, and ability to reduce the incidence of adverse drug reactions and their effects e.g. hospitalisation and healthcare utilisation.[88,89] Establishing relevance and applicability of PIP screening tools is an important validation step, captured in the characteristics of including the most common errors and covering all aspects of appropriateness above.[86] One of the primary reasons for several countryspecific measures was to develop indicators more applicable to the setting than the Beers criteria which contained drugs not often used. The frequent revisions of the Beers criteria to add new indicators and remove others illustrate that even within the same setting relevance may change with time. Including irrelevant indicators can unnecessarily increase the time needed to apply a screening tool and so reduce its utility, therefore monitoring how trends in prescribing of potentially inappropriate drugs vary is important. Another factor included in the proposed ideal characteristics is an ability to reduce rates of ADEs and consequences, such as hospitalisation and healthcare utilisation,[88] and to have causal links with important clinical outcomes.[41,87,89] As these PIP indicators have primarily been developed by consensus methods, demonstrating such predictive validity is vital.[86] This is congruent with Donabedian s model of evaluation of quality of care,[49] which would suggest that PIP criteria, like other valid process measures, should be associated with meaningful patient outcomes.[41,53] This causal relationship is important so that changing the process (i.e. intervening to reduce PIP) results in a change in outcomes (i.e. a reduction in ADEs or hospitalisations).[53] An overview of some research which supports how measures of PIP, in particular STOPP and START, reflect the ideal characteristics is presented. 1.4 Previous research on PIP in older adults Time trends in appropriate prescribing Prescribing of medications has increased in recent years both in terms of costs and volume, and the use of multiple medicines by older patients has become commonplace.[28,90] In the US, visits to doctors in ambulatory care by older people on at least five medicines quadrupled over a 10 year period between 1990 and 2000.[91] 37

40 Number of medications is one of the strongest predictors of PIP,[92] but most often PIP prevalence has been assessed at one point in time or during one time period, despite the rising levels of medicines use. An overview of research which has assessed time trends in appropriate prescribing in community-dwelling older people is presented. Evaluating trends in PIP is important both to understand if indicators remain applicable and relevant but also to understand how the rising prevalence of polypharmacy is related to PIP, given this is the strongest predictor. Regarding numbers of medications, incident polypharmacy, defined as being prescribed five or more medications in a month for at least 6 months out of 12, almost doubled from 1.46% of an older people included in an Italian administrative health database in 2001 to 2.86% in 2009.[93] Incident polypharmacy was associated with hospitalisation, institutionalisation and death at both time points. The effect of incident polypharmacy appears to have changed with time - an interaction between year and polypharmacy was found for the outcome of hospitalisation, with a significantly higher adjusted hazard ratio in 2009 (1.17, 95% confidence interval (CI) 1.16, 1.19) relative to 2001 (1.12, 95% CI 1.10, 1.15).[93] A Swedish cohort study from 1992 to 2002 found concurrent use of five or more drugs doubled and was more common among those with lower educational attainment and potential DDIs increased from 17% to 25% and were also associated with level of education.[26] Similarly, Guthrie and colleagues identified significant increases in both number of medicines prescribed and DDIs between 1995 and 2010 among primary care patients in Tayside in Scotland.[40] The proportion of adults on five or more and 10 or more medications doubled (to 20.8%) and tripled (to 5.8%) respectively over this time and was strongly associated with age. Potentially serious DDI prevalence rose from 5.8% of adults in 1995 to 13.1% in 2010, mainly involving cardiovascular, central nervous system (CNS) or musculoskeletal drugs. The strongest predictor of a DDI was number of drugs but as with the Swedish cohort study mentioned previously, this analysis only considered predictors at one point in time. Specifically focusing on potentially inappropriate prescribing, some studies have only reported the change in prevalence with time and did not consider factors associated with this prescribing.[94] For example, amongst older Swedish people, the proportion prescribed a long-acting benzodiazepine or tramadol fell while zopiclone use rose over a 38

41 seven year period between 2006 and 2013.[95] During this time, a marginal reduction in anticholinergics was observed. This is in contrast to a Scottish study where there were increases in both the number of older people prescribed anticholinergics and the average anticholinergic burden between 1995 and 2010.[96] Interestingly, the proportion of individuals with high anticholinergic burden among those on five or more medications was reported and this was significantly lower in people with polypharmacy in 2010 compared to 1995, suggesting this cut-off has become less specific for suboptimal prescribing with time. A US study applied a subset of prescribing measures from the Health Plan Employer Data and Information Set (HEDIS), a tool used by health plans to monitor quality of care performance, to administrative data and found the proportion of older veterans receiving a high-risk medication declined from 24% in 2004 to 10% in 2009.[97] The only predictors assessed were age, sex, race, cognitive function and specific comorbidities. Trends in Beers criteria inappropriate drugs have been evaluated in several studies, including one applying a modified version of the 2003 Beers criteria to older patients in a UK primary care database from 1994 to 2003.[98] The prevalence remained fairly stable at about one third of patients over this period, with the most commonly implicated drugs being amitriptyline (although in many cases this was used at low doses) and dextropropoxyphene (as co-proxamol) while benzodiazepine use fell over this period. In a related paper evaluating prescribing between 1996 and 2005, although the number of repeat medicines prescribed were found to have increased, PIP actually declined, particularly in the case of dextropropoxyphene.[99] Predictors of PIP in 2005 were assessed and number of repeat medicines was the strongest, with few other factors being significant once this was accounted for. Again, a limitation of this study was that predictors were only evaluated for one time point and the relationship with PIP over time was not considered. Price and colleagues evaluated the prevalence of PIP according to the 2003 Beers criteria over a 13 year period up to 2005 among older Australians using an administrative claims database.[100] The prevalence fluctuated over this period, peaking at 46.9% in 1999 and reaching a low of 40.6% in 2005, and temazepam was consistently the most common PIP identified with others drugs such as digoxin and NSAIDs declining during the study. Number of medications was again the strongest predictor of PIP, and, after accounting for 39

42 the number of prescribed medications, a trend of reducing odds of having PIP over time was apparent PIP and patient outcomes It is important to demonstrate that PIP criteria have an effect on patient outcomes, not only because a causal link is a desirable characteristic of explicit criteria but also to establish the validity and justify their use as a process measure of quality of care. However, this important step in the development of explicit prescribing criteria is often not carried out.[86] Even when studies have been conducted, insufficient confounding control, failure to address criteria for causality (such as the Bradford Hill criteria which include temporal relationship and dose-response), and other methodological issues limit the strength of evidence.[41,51,101] While the majority of older people are community dwelling, many studies of PIP have been in hospital settings, which limits the generalisability to the wider older population. For example a prospective study by Hamilton and colleagues recruited 600 older adults admitted to hospital with acute illness.[102] PIP was assessed using both STOPP and the Beers criteria based on medications at the point of hospitalisation. For each additional STOPP criteria, there was a statistically significant 85% increase in the odds of a serious avoidable ADE occurring before hospitalisation, and in many cases this contributed to the admission. No significant association with the Beers criteria was found, and STOPP identified significantly more medications involved in ADEs, avoidable ADEs, and avoidable ADEs that contributed to hospital admission than Beers. A similar prospective study including 715 acutely ill older adults admitted to hospital which assessed medication on admission found PIP according to STOPP caused or contributed to 11.5% of hospitalisations due to an ADE while Beers criteria were implicated in 6%, a significant difference.[103] Although the prospective designs of these studies are a strength, the non-representative acutely ill population, and inability to assess the strength of the exposure-outcome temporal relationship are limitations. A study of PIP among older hospital inpatients partaking in a study of appropriate medication use in Italy applied the STOPP and Beers criteria to medications taken seven days before and during admission.[104] The two outcomes evaluated were occurrence of an adverse drug reaction (ADR) definitely or probably caused by a drug taken during the exposure period (as assessed by the Naranjo criteria), and a decline in functional status 40

43 between admission and discharge (defined as an increase in the number of Activities of Daily Living an individual is dependent for). Again, occurrence of an ADR and functional decline was significantly associated with exposure to STOPP but not with the Beers criteria. Although a period of pre-hospitalisation medication exposure was considered, the inclusion of exposure during hospitalisation as well complicates the temporal sequence in this study. A large retrospective cohort study was conducted in Sweden using data from 813 older patients who had a healthcare visit in primary care, or as an outpatient or inpatient in specialised care over a three month period.[105] Having any STOPP criterion was associated with almost 2.5 times higher odds of an ADR, and almost 30% of ADRs were considered to be caused by a PIP, although the same period of exposure and outcome appears to have been used. This study population is highly representative given it included individuals from all care settings and the finding that only 7% of participants had a hospitalisation during the study period challenges the generalisability of hospital-based studies. A cohort study evaluating the impact of PIP has been conducted in Irish general practice, including over 900 community-dwelling adults aged 70 years and over.[106] ADEs, accident and emergency (A&E) visits, and hospital visits were determined from GP medical records and patient self-report with review by two independent clinicians, and retrospective medication dispensing information from the six months previous to this was used to apply STOPP and the Beers criteria. Patients with multiple STOPP criteria were twice as likely to have an ADE and had double the rate of A&E visits. In the case of hospitalisation, the increased rate that was associated with STOPP was not found for Beers criteria.[107] STOPP was also associated with vulnerability and a reduction in health-related QoL. Although this cohort contained a large community-dwelling sample of older adults, as the temporal sequence of PIP and outcomes during the six month retrospective period could not be determined, this limits the inference of causality. Regarding underprescribing, the consequences of cardiovascular medication underuse included in the START criteria have been evaluated in a subsample of 1,454 older participants in a population-based cohort.[108] Medication underuse was assessed at each participant s study entry. Of the outcomes of cardiovascular mortality, non-fatal MI, stroke, or coronary interventions, and non-cardiovascular mortality, the only significant 41

44 association identified during a mean follow-up of 2.24 years was between START and noncardiovascular deaths. While only a subset of PIP criteria were considered here, the prospective nature of this study in community dwellers is a major strength. Few studies have assessed the impact of overprescribing and misprescribing as well as underprescribing in the same study. This is despite both acts of commission and omission contributing to PIP and evidence suggesting people on five or more regular medicines may be more likely to also be subjected to underprescribing.[109] Data from a randomised controlled trial (RCT) of a pharmacist intervention delivered to 368 patients aged 80 years and over was used to evaluate any relationship between measures of prescribing quality (including STOPP, START, and MAI) and clinical outcomes.[110,111] In the first of two papers reporting this secondary analysis, each measure was assessed in separate multivariate regression models and although no significant relationship with total number of hospital visits or readmissions was found, number of STOPP criteria and MAI score were associated with drug-related readmissions (rate ratios of 1.34 and 1.09 respectively).[110] The second paper examined the use of STOPP, START, and MAI as well as validated risk prediction tools to predict rehospitalisation or death in the 12 months following discharge.[111] All three of the prescribing measures had poor discrimination between patients at risk of the composite outcome, though the authors did acknowledge that STOPP, START and MAI were not devised as risk prediction scores. A further limitation is the difficulty in isolating the effect of PIP from the effect of the intervention delivered to patients in the RCT these two papers analysed data from (which was effective in reducing each PIP measure between admission and discharge). Gosch and colleagues conducted a retrospective cohort study of 457 older patients admitted to a geriatric hospital department following a previous hip fracture.[112] Appropriateness of prescribing was assessed at discharge and three-year mortality was the outcome of interest determined using a death register. The combined number of STOPP and START criteria was significantly associated with mortality, independent of demographics and comorbidities. Although STOPP and START were developed together as complementary measures, the direct combination of them is questionable given they are theoretically and practically different issues. These findings are also limited by this being single centre study, using a retrospective design and a select study sample of only hip fracture patients. 42

45 A nested case-control design including 338 older patients from the Hospital Admission Related to Medication (HARM) cohort study evaluated the effects of Beers, STOPP, START, and STOPP and START together.[113] Medication-related hospital admissions were significantly associated with STOPP criteria and the composite of STOPP and START in multivariate analysis, however it is not clear whether STOPP and START were combined as a single measure or if they were both included as independent predictors. The controls used here, planned surgery admissions in the same hospital as cases, matched for age and sex, may have introduced bias as these patients were likely to be less ill than patients with unplanned admissions, which may have related to the prevalence of PIP. All of the studies that assessed both STOPP and START have been conducted in the hospital setting with no primary care based research on the more general older population to date. None of these studies appear to have included both PIMs and PPOs as two independent predictors in the one analysis, having either conducted separate analyses for each or combined the total number of STOPP and START criteria. To provide robust evidence, a study should use a prospective design to account for temporality in the relationship between cause and effect and to adequately control for factors which may confound the relationship between PIP and patient outcomes. Even if confounders are adjusted for using standard methods, there is still potential for bias in longitudinal studies.[114] Economic impact of PIP As well as linking process measures to patient outcomes, the effect on the wider health system is also important to consider and healthcare costs are one of the most relevant outcomes at this level. A causal relationship with economic outcomes may also support addressing prescribing appropriateness as a cost-saving measure. The economic impact of PIP has not been extensively researched in the literature. A systematic review of the application of STOPP and START identified only three studies that assessed economic outcomes associated with PIP and in each case, only the direct cost of the medications from each criterion was included.[115] A national populationbased analysis of a subset of STOPP criteria amongst those over 70 years of age in Ireland estimated the cost of dispensed PIP medications at 45 million in 2007, approximately 9% of the total state pharmaceutical expenditure for this age group.[92] The equivalent costing from a similar study in Northern Ireland was also high at 6 million or 5.4% of 43

46 drug spending.[116] Costs of PIP in these two countries have also been evaluated in the nursing home setting.[117] Among a total of 630 nursing home residents in Ireland and Northern Ireland, the annual net ingredient costs of medicines defined by STOPP and Beers as potentially inappropriate were 165,500 and 45,500 respectively. A further systematic review of the economic burden of a variety of medication problems included several studies relating to inappropriate prescribing.[118] Among 500 patients in primary care in Ireland in 2006, the monthly net ingredient cost of prescribed medications from Beers criteria and the Improving Prescribing in the Elderly Tool (IPET) were 825 and 381 respectively.[119] In a US study of 946 patients in ambulatory care, 36% were prescribed a PPI with no documented appropriate indication at an annual cost of $234,000 or $1.57 million using over-the-counter or prescription prices respectively.[120] Inappropriate medicines use defined by Beers criteria has also been shown to be associated with increased healthcare utilisation and costs in a Medicare managed care population.[121] In only assessing the cost of inappropriate drugs these studies neglect to include the economic ramifications of appropriate alternative prescriptions that may be required. A recent study has assessed this with respect to the German PRISCUS list of inappropriate medications in those aged 65 years and over and estimated the cost to the health system in 2009 for these was million.[122] Considering the appropriate replacement prescription, substituting each PIM with an medication at the mean price of all the possible alternative would incur additional costs of 125 million, with a range between 20 and 504 million if the least or most expensive alternatives were used. A 2004 study of the monthly costs of drugs from START that were omitted in 600 older people who were hospitalised amounted to over 9,000, with omission of indicated statin therapy contributing the largest cost.[123] Trials have demonstrated that interventions which included STOPP and START can significantly reduce mean drug cost,[124] and the prescription of medications when a less expensive alternative exists.[125] Relating to specific drug classes, the economic cost of adverse events of NSAIDs has been evaluated in two studies.[126,127] Costs and healthcare use attributable to inappropriate benzodiazepine prescribing and the costs of hospital-treated falls associated with benzodiazepine use in Europe have also been evaluated using regression methods.[128,129] 44

47 Overall these studies have primarily focused on direct costs of inappropriate prescribing, usually have not accounted for the cost of the alternative, and even when adverse effects have also been considered, these studies have used simplistic methods. Economic evaluations of other areas of medications use have adopted more sophisticated modelbased approaches. Karnon and colleagues evaluated medication errors in hospitals by developing a decision tree model which mapped the medication processes that occur in secondary care, including pathways from potential error points to the outcomes of undetected errors.[130] Errors could occur at prescription, dispensing or administration and four types of error for each of these stages were described e.g. for prescribing errors, wrong drug, dose, route, or frequency. The outcomes of an undetected error were classified into four categories of harm: none, mild, moderate and severe/life-threatening. The model was populated with error rates, detection rates, probabilities that undetected errors cause harm, severity of errors causing harm and the effectiveness of interventions being investigated in reducing the impact of medication errors. Costs for patients experiencing an ADE were applied from the literature.[130] A similar approach has been described in a protocol to evaluate the cost-effectiveness of eprescribing using a health economic model.[131] A German study to evaluate the costs resulting from ADEs, which may be due to inappropriate prescribing, also adopted a decision tree approach, in which probabilities were assigned to consequences of different types of ADE and the costs of resource utilisation associated with these consequences or treatments were determined.[132] While in these cases there is little ability to examine specific drugs, the advantage of such approaches is that they provide a framework that is generalisable to all types of medication error or ADE. This is in contrast to the methods used by Elliott and colleagues to evaluate the cost-effectiveness of PINCER, an intervention to reduce medicines management errors. They evaluated six prescribing and monitoring errors using a series of Markov models, one for each error being targeted.[133] This approach provides improved face validity and allows for data relating to the likely clinical consequences of specific errors to be used to populate the models. 1.5 Previous research on PIP in middle aged adults As has been illustrated, most studies on PIP have been in older aged populations. Research on prescribing has generally not included middle-aged people as a specific 45

48 group of interest, although there are some exceptions to this. A population-based study evaluated potentially serious DDIs in adults of all ages and, although there was increasing risk of interactions across increasing age groups, controlling for patient and practice factors, this did not appear to be significantly different between those aged and years.[40] Fifteen indicators of high-risk prescribing relating mainly to NSAIDs were evaluated in a study which included middle-aged people among adults of all ages, and, after adjustment for patient and practice factors, this study found that the risk of highrisk prescribing in middle-aged groups was similar to older individuals with the highest odds ratio in those aged years.[39] These studies illustrate that middle-aged individuals are subject to similar prescribing issues as the older population. However, despite this and the shared risk factors for adverse effects of medicines, PIP has not been considered to date in those aged 45 to 64 years, either by applying existing screening tools or developing indicators specific for this age group. The lack of specific indicators has recently been addressed by the publication of the PROMPT (PRescribing Optimally in Middle-aged People s Treatments) criteria in 2014.[134] These were developed to identify PIP specifically in middle-aged people using prescribing or dispensing data in the absence of clinical information. The methods used were similar to those employed in the development of explicit PIP criteria for older people. An initial list was compiled by reviewing prescribing indicators used in older people for applicability to middle age. Criteria were excluded if they required clinical information to be applied or if they related to conditions which are rare in middle-aged people such as dementia, or included medications not used in the UK or Ireland. This list was then assessed by a Delphi consensus panel, composed of 17 members with relevant clinical, pharmacy and epidemiological expertise from both the UK and Ireland. A webbased questionnaire was used and members rated agreement to each proposed indicator and of 34 criteria included initially, consensus was achieved on 22 criteria following two rounds. The PROMPT criteria are divided by physiological system, with six criteria relating to the central nervous system, cardiovascular and respiratory systems contain four each and three gastrointestinal system criteria. STOPP and START were an original source of the majority of the final criteria (twelve) and nine were included that had appeared in the Beers criteria, hence most of the PROMPT criteria are recognisable given familiarity with PIP in older people. 46

49 As with other measures of PIP, the PROMPT criteria can be considered process measures of medication safety according to Donabedian s classification,[49] and should reflect the optimal characteristics of explicit measures described in Section such as encompassing common prescribing errors, generalisability for the global community of doctors and pharmacists, and being easily interfaced with computer records of patients details.[88] The relevance and applicability of the PROMPT criteria have not been demonstrated owing to their recent development, however, this is an important validation step.[86] As with any process measure of quality of care, a further important characteristic of prescribing criteria is an ability to reduce rates of adverse drug events and consequences such as hospitalisation and healthcare utilisation,[88] through causal links with important clinical outcomes.[41,87,89] Although the effect of PIP on ADEs has been shown in older people, it is still necessary to demonstrate that similar criteria have an impact in middle-age. One of the determinants of beneficial and adverse effects of drugs is baseline characteristics of individuals; hence middle-aged people may have improved physiological functioning and thus be more resistant to adverse effects of the same drugs than older aged people.[11] 1.6 Aims and objectives of this thesis The overall aim of this thesis was to assess measures of PIP in older and middle-aged people in primary care in terms of their (1) applicability and relevance in Ireland, (2) effect on patient outcomes and (3) economic impact. These aims were achieved through a number of specific objectives. Aim 1: To assess the applicability and relevance of measures of PIP. Objectives: i) To determine the prevalence of PIP using STOPP over a 15 year period from 1997 to 2012 in people aged 65 years using regional administrative pharmacy claims data. ii) iii) iv) To assess the relationship between PIP and polypharmacy over this period. To evaluate medications co-prescribed with the most common PIP criterion which may explain the change in this pattern of prescribing over this period. To establish the prevalence of PIP using the PROMPT criteria in middle-aged people (aged years) using national administrative pharmacy claims data covering approximately one third of the population in this age group. 47

50 v) To identify the relationship between PIP in middle-aged people and polypharmacy, age and sex. Aim 2: To assess the effect of measures of PIP on patient outcomes. Objectives: i) To determine the prevalence of PIP over time, changes in prevalence between baseline and two year follow up, and factors associated with any change, in a cohort of community-dwelling older people aged 65 years and over. ii) iii) iv) To evaluate the prospective relationship between PIP determined by STOPP and START and healthcare utilisation, functional decline, and QoL in this cohort. To determine the prevalence of PIP over time, changes in prevalence between baseline and two year follow up, and factors associated with any change, in a cohort of community-dwelling people in middle-age (between 45 and 64 years). To evaluate the prospective relationship between PIP determined by the PROMPT criteria and healthcare utilisation and QoL in this cohort. Aim 3: To assess the economic impact of measures of PIP. Objectives: i) To develop and populate a Markov model for the three most prevalent STOPP criteria using published and thesis-reported data. ii) iii) To use these models to compare each PIP to an appropriate non-pip alternative in terms of the effects on patient outcomes, namely quality and quantity of life, and costs to the health service of the primary adverse events. To apply these models to evaluate the cost-effectiveness of hypothetical interventions to reduce PIP. 1.7 The health system in Ireland The research presented in this thesis relates to adults in primary care in Ireland and therefore some contextual information on the Irish health service with a focus on medication prescribing is presented here. The health system in Ireland is primarily tax funded, with the remainder of funding coming from out-of-pocket payments and private insurance.[135] The Health Service Executive (HSE) is the main provider of health services, particularly in secondary care, 48

51 while in primary care, GPs and community pharmacists are private contractors who provide services for the HSE. Some patients can access medical services, including hospital inpatient and outpatient care, GPs and dental services, free at the point of use through the General Medical Services (GMS) or medical card scheme. Eligibility to this is based on household income, with a higher threshold applying to those aged 70 years so a greater proportion of this age group are covered. Approximately 40% of the general population are covered compared to 96.5% of those aged 70 years and over.[6] For people with income levels just above the GMS scheme threshold, a GP visit (GPV) card is granted which entitles them to only GP visits free of charge.[135] Eligibility to the GPV card was extended in 2015 to all children aged <6 years and all adults 70 years who do not qualify for a medical card. For patients not covered by the GMS/GPV scheme, they must pay out of pocket for GP visits. GPs are paid by the HSE for care provided to GMS/GPV patients primarily on the basis of capitation, but this may change in the coming years. A number of national clinical programmes have been introduced by the HSE to provide clinical leadership to improve care delivery in line with the Tackling Chronic Disease policy framework, including health failure, mental health, and care of older people.[136] One of these programmes, for diabetes, has recently resulted in a cycle of care being introduced for type 2 diabetes mellitus in primary care and from October 2016 this will link appropriate care for diabetes patients to GP payments from the HSE.[137] Medications are provided through a number of schemes.[5,135] GMS patients can receive most prescribed medications free of charge, although a co-payment of 0.50 per prescription item was introduced in 2010 and this has since been increased to 1.50 and 2.50 in January and December 2013 respectively. Secondly, the Long Term Illness (LTI) scheme entitles people with any of 16 specified conditions, including diabetes, epilepsy, and multiple sclerosis, to free medications to treat their illness.[135] All individuals with these conditions are automatically eligible with no test of means and no co-payment is required. Those not covered by either the GMS or LTI schemes pay out-of-pocket for their medications, though the Drugs Payment Scheme, which all individuals are automatically eligible for, caps household payments at a maximum level per month, currently 144, beyond which the state pays.[135] These schemes are administered by the Primary Care Reimbursement Service (PCRS) within the HSE. Pharmacies transmit claims of medications dispensed under these schemes to the PCRS at the end of each month for reimbursement. Pharmacies are paid on a fee per item basis for dispensing of prescribed 49

52 medications and the only other state-remunerated service is the administration of winter influenza vaccines, with no formal role in medication review outside of the dispensing process. The administrative pharmacy claims data transmitted to the PCRS can and has been used for research purposes however as different numerical patient identifiers are used for the different schemes, each scheme can generally only be examined in isolation. The recent Health Information Act 2014 has laid the legislative foundations for the introduction of an individual health identifier (IHI), which may address this limitation. Building on this is the ehealth Ireland strategy which outlines the planned development and introduction of electronic health records (EHRs) in the coming years among other improvements in IT infrastructure.[138] This move away from paper-based systems has already begun with the recent introduction of electronic transmission of discharge summaries from hospital to GPs and electronic referrals, with electronic prescribing also planned. A number of bodies related to the HSE have a role in rational prescribing. Most medications are covered by the PCRS drug schemes however agents of questionable costeffectiveness or major budget impact are referred to the National Centre for Pharmacoeconomics (NCPE) for economic evaluation before reimbursement is approved. The cost-effective use of medications also comes under the remit of the HSE s Medicines Management Programme (MMP) along with enhancing safe and effective prescribing. Among steps undertaken by the MMP since its establishment in 2013 is the implementation of a preferred drug initiative which recommends a preferred drug molecule to be prescribed within a number of drug classes such as statins and PPIs.[139] To date the MMP has mainly focussed on cost-effective prescribing while the National Medicines Information Centre (NMIC) has played a bigger role in the area of safety and effectiveness. Although its primary purpose is the provision of a medication query answering service, the NMIC also issues therapeutics bulletins and topic guides on safe and effective prescribing for information purposes and these are sent in print form to all GPs. 1.8 Justification of methods used The population of interest for this thesis was community-dwelling adults. A majority of studies of PIP have focussed on acutely-ill people in the hospital setting whereas the 50

53 majority of individuals residing in primary care have been less frequently studied. The methods used were: (i) (ii) Secondary analysis of two existing large-scale data sources, namely administrative pharmacy claims data from the PCRS GMS scheme and The Irish Longitudinal Study on Ageing (TILDA), a nationally-representative cohort study of ageing, and Synthesis of previously published data in an economic analysis using Markov models. The pharmacy claims database from the PCRS contains records of prescribed medications which were dispensed to individuals eligible for the GMS scheme as described in Section 1.7. Study designs adopted included cross-sectional and repeated cross-sectional studies for Aim 1, cohort studies including marginal structural models (MSMs) to achieve Aim 2, and economic modelling using a Markov model approach for Aim 3. Justification for the selection of the second and third of these methods is presented below. The procedure for accessing the two national databases used in this thesis was similar. A protocol outlining the planned research was submitted to the Department of Pharmacology & Therapeutics, Trinity College Dublin, which is the data controller for the PCRS database, and to the Study Management Committee of TILDA for approval. Once the approval had been obtained, all analysis was conducted on secure hotdesks located in the Department of Pharmacology & Therapeutics and TILDA, where only summarised data containing no individual identifiable information could be removed from the hotdesk computer Cohort study analysis Given the lack of evidence from RCTs regarding drugs in older people,[8] the use of observational data can provide real world evidence to clinicians regarding prescribing decisions in this age group. However, associations identified in such studies may be due to correlation rather than causation and can be confounded even if regression methods are used to control for these effects.[140] Longitudinal data from a cohort study allows for the prospective association between PIP and patient outcomes to be determined, addressing the temporal relationship required to demonstrate causal effects.[101] While this does eliminate the potential for bias due to reverse causality that can occur in crosssectional analysis, standard multivariate regression methods may still lead to biased 51

54 estimates if there is confounding by time-varying covariates or unmeasured confounders.[114] Considering an example of a study, represented in Figure 1-2, to estimate the causal effect of PIP on healthcare use in older people, the number of medications a patient is on can predict whether they have PIP or not and can also predict their healthcare use. Additionally, exposure to PIP may increase the number of medications prescribed (e.g. if a new drug is started due to an adverse effect of PIP). Therefore, the number of medications is a time-dependent confounder. A number of standard regression methods to estimate the effect of PIP on healthcare use may produce biased results. For example: a) Not controlling for confounding and using the crude estimate will bias estimates of any effect of PIP as those with PIP will tend to be on more medications and therefore have more healthcare utilisation. b) Controlling for the baseline value of confounders such as number of medications will produce biased results as it does not account for the fact that those who received a PIP after the start of the study may be those who had an increase in the number of medications. c) Controlling for time-updated values of time-varying confounders affected by previous PIP exposure, such as number of medications, will give biased results because some of the effects of PIP on healthcare use may be through increased number of medications. As shown in Figure 1-2, if number of medications at time t+1 lies on the causal pathway between PIP exposure and the outcome, controlling for this would bias the exposure effect estimation, as any effect of PIP acting via number of medications would be lost. Figure 1-2 Example of the impact of different approaches to confounder control, X representing pathways blocked by different approaches 52

55 MSMs can be used to overcome these challenges as they separate the control of confounding from the estimation of effects and so account for time-dependent confounding without biasing estimates of the exposure effect on outcomes.[141] MSMs have been used in a number of observational studies of medication exposures.[114,142,143] Unlike RCTs, individuals characteristics are not equally distributed and MSMs address this by upweighting individuals unlikely to have received the exposure they did given their covariate history.[141] By applying these weights this method attempts to replicate an experimental study by balancing individual characteristics between the exposure groups, as occurs with the randomisation process in trials. The focus of previous studies that have used MSMs has been predominantly on beneficial effects of medications (effectiveness), with few cases assessing medication safety, which may be a more important application given observational studies are necessary to evaluate drug harms which often cannot be investigated in trials.[144,145] Economic modelling The use of a modelling approach to evaluate the economic impact of PIP allows for existing evidence to be synthesised to address this research question. Among the modelling methods available,[146] Markov models are preferable in this case as time dependency can be included, which has been lacking in previous studies evaluating the adverse effects of some drug classes in older people.[ ] As described in Section 1.4.3, a number of studies using economic models to evaluate the economic impact of medication errors and adverse drug events have taken a generalised approach.[ ] In these cases specific instances of medication error and resulting adverse events were not being investigated but rather the totality of error types divided into categories of error (e.g. wrong drug prescribing error) and categories of adverse event (e.g. severity of harm or treatment required as a result of event). It would be possible to assign PIP criteria into various categories to use a generalised approach, for example inappropriate duration, inappropriate dose, inappropriate drugdrug combination, and inappropriate drug-disease combination. However, such groupings would combine prescribing issues with varying types and risk of adverse events with varying magnitudes of harm. For example, long-term use of benzodiazepines and longterm maximal dose PPIs would both be categorised as inappropriate due to duration, despite the former having a high propensity to cause harm due to falls risk while the 53

56 latter carries a lower risk of harm.[45,147] A generalised approach would result in an unnecessary reduction in face validity of the model given the finite number of PIP criteria that are of interest. Therefore, similar to the methods used to evaluate the costeffectiveness of addressing six prescribing and monitoring errors in the PINCER trial,[133] a separate drug-specific Markov model was developed for each of three specific PIP indicators. Markov model approaches generally involve mapping out the stages of progression of a disease or other clinical situation as health states, the progression of which may be altered by an intervention of interest.[148] In this case, the intervention of interest is a harmful one, PIP, and the consequences mapped out are the adverse events which each type of PIP may lead to. 1.9 Thesis structure Following this chapter, repeated cross-sectional analysis of regional pharmacy claims data is presented in Chapter 2, which assesses the volume of medicines dispensed, how this has changed over the last 15 years, and its relationship with PIP in older aged patients to address objectives i) and ii) of Aim 1. Further analysis characterising the prescription of proton pump inhibitors (PPIs) in this age group is presented in Chapter 3 to determine what factors may explain the change in its prevalence from 1997 to 2012 in line with objective iii) of Aim 1. Longitudinal analysis of a cohort of older patients with linked pharmacy claims data is reported in Chapter 4 to meet objectives i) and ii) of Aim 2. This work determined exposure to PIP over time using a number of different screening tools, and then evaluated the relationship of STOPP and START with patient outcomes, namely healthcare use, functional decline, and QoL. This is followed in Chapter 5 by an economic evaluation of the impact of the three most common STOPP criteria identified in Chapter 2 to address all objectives of Aim 3. A Markov modelling approach was used to quantify the cost and QoL effect of adverse events which may result from these patterns of prescribing to identify which PIP criteria to prioritise. From here, the focus shifts to the middle-aged population, and cross-sectional analysis of national prescribing data for one third of the middle-aged population of Ireland who are most socioeconomically deprived is presented in Chapter 6. This addresses objectives iv) and v) of Aim 1 and served to establish the burden of PIP and applicability of the PROMPT criteria in this age group for the first time. Then to assess the effect of PIP in middle age on patient outcomes in line with objectives iii) and iv) of Aim 2, Chapter 7 presents 54

57 longitudinal analysis of a similarly deprived cohort of patients aged between 45 and 64 years with linked pharmacy claims data. Finally, Chapter 8 discusses the overall findings, the implications of these, the overall limitations of the thesis as well as recommendations for further research and the impact of the findings. To expand on the terminology shown in Table 1-1, while PIP does incorporate both acts of commission (PIMs) and acts of omission (PPOs), most chapters focus only on STOPP and therefore PIP is for the most part used synonymously with PIMs. In Chapter 4 where information on diagnoses allows for the evaluation of PPOs from START, PIP is used when referring to both acts of commission and omission while PIMs and PPOs are used when discussing STOPP and START respectively. 55

58 Chapter 2 Trends in polypharmacy and PIP in primary care 56

59 2.1 Introduction As discussed in Chapter 1, the volume of medicines being prescribed has risen in recent years.[90] Despite this, concerns have been raised that patients who are eligible for evidence-based treatments are not receiving them, particularly those already prescribed multiple medications.[32,149] Set against the background of a rising burden of illness, some commentators have called for the use of more medicines to alleviate pain and disability, prolong life, and prevent avoidable disease.[149,150] In contrast, there is increasing concern about overdiagnosis and overtreatment, particularly in older patients.[30,151] There has been a proliferation of clinical guidelines focused on single conditions, which fail to account for the growing cohort of patients with multimorbidity.[152] There has been particular focus on polypharmacy in older people, commonly defined as the use of five or more regular medications. Although polypharmacy has been used as a crude marker of prescribing quality, it can in many cases, such as multimorbidity, be entirely appropriate.[29,31] A clearer indicator of medications safety and prescribing quality is PIP, the use of a medicine such that the harms outweigh the benefits,[107] and exposure to PIP medications is associated with adverse outcomes, including ADEs and hospitalisations.[106,107] These two issues are inter-related, as polypharmacy is the single biggest predictor of being prescribed a PIP medication.[107,153] Several studies have evaluated changes in the number of medications being prescribed over time,[25 27,154] however trends in PIP have been researched less extensively and the only PIP screening tools evaluated so far are the Beers criteria and HEDIS list.[94,97 99] STOPP has yet to be considered over time and so it is unclear if the PIP criteria included in this measure have remained applicable and relevant. Although it appears that both polypharmacy and PIP have remained prevalent or become more widespread in recent years, the relationship between these issues over time is not clearly understood. Predictors of PIP have generally only been assessed at one point in time, and where the relationship of polypharmacy with other indicators of appropriate medication use has been evaluated longitudinally, the strength of this association may have decreased in recent years.[93,96] The aims of this study were to analyse administrative pharmacy claims data over a 15 year time period from 1997 to 2012 in primary care in Ireland to examine: (1) the change in prescribing patterns and rates of polypharmacy in all individuals, (2) the change in 57

60 prevalence of PIP as determined by a subset of applicable criteria from STOPP in individuals aged 65 years and (3) the relationship between PIP and polypharmacy in these older individuals. 2.2 Methods Study design and setting A repeated cross-sectional study was conducted using patient-level dispensing data from an administrative pharmacy claims database in the years 1997, 2002, 2007, and 2012, a period of 15 years. Data were included for all people eligible for the General Medical Services (GMS) scheme in the study years in the former Eastern Health Board (EHB) region of Ireland, where 29.1% of the national population resided in 2012.[155] The GMS scheme is a means-tested form of public health cover in Ireland providing free health services, including most prescribed medicines, to people based on income and age, although a monthly co-payment per prescription item was introduced in As of 2012, 40.1% of the general Irish population and 96.5% of those over 70 years were covered by the scheme.[155,156] From 2002 to 2008, all people aged 70 years or over were automatically eligible for the scheme and since January 2009, a higher income threshold was applied to this age group compared to the general population. The Health Services Executive-Primary Care Reimbursement Service (HSE-PCRS) database used for this study contains dispensing records of medications prescribed to patients in primary care by their general practitioners. For this analysis, medications were classified into drug classes based on the first five characters of their World Health Organisation (WHO) Anatomical Therapeutic Chemical (ATC) code and only dispensing data relating to individual patients were included Ethical approval Permission was given by the data controller of the HSE-PCRS database for analysis if the data were anonymised and analysed so that no subgroup contained fewer than five individuals. This was adhered to and so, in accordance with the data controller s policy, it was not necessary to seek specific ethical approval for this analysis. 58

61 2.2.3 Analysis Regular medications and polypharmacy in the total population The number of regular medications (dispensed in at least three consecutive months) per person during each study year was analysed to determine the distribution of individuals by age group and number of regular medications (category aggregated at 15 or more medications). The most prevalent medications (grouped by five character ATC codes) were also examined by determining the number of individuals to whom each medication was regularly dispensed. For the years 1997 to 2007, this was directly standardised by age group and sex to the 2012 GMS population of the EHB region to account for changing demographics and to allow comparison across years. The prevalence of polypharmacy (being dispensed five or more regular medications in the study year) and excessive polypharmacy (conventionally defined as 10 or more regular medications) was calculated.[157] Negative binomial regression was used to quantify the change in the rate of these outcomes associated with study year (using 1997 as reference year), controlling for age group and sex. Interaction terms between these independent variables were included if they provided a statistically significant improvement to model fit. Incident rate ratios (IRR) with 95% confidence intervals (CI) are presented. Negative binomial models were used over Poisson models due to overdispersion in the rates of polypharmacy.[158] PIP in older individuals Prescribing inappropriateness was assessed in individuals aged 65 years using a subset of criteria from the Screening Tool for Older Persons Prescription (STOPP). This explicit measure of PIP was applied as it was developed for use in older European populations and includes more medications than the Beers criteria which are commonly prescribed in the study setting.[42,159] Analysis was restricted to older individuals, as explicit measures of PIP such as STOPP have not been validated in younger age groups. Thirty of sixty-five criteria were applicable to the information in the HSE-PCRS administrative dataset and lack of detailed clinical information precluded application of the remaining criteria, consistent with other studies applying STOPP to pharmacy claims data.[160] PIP prevalence was assessed by the percentage of individuals with any PIP criteria. Prevalence for each individual STOPP criteria was also determined to examine the most common forms of PIP over the study period. 59

62 The relationship between PIP and polypharmacy in older individuals Unadjusted logistic regression was performed to assess the change in odds of having any PIP over the study years and odds ratios with 95% CI are presented. The relationship between PIP prevalence (dependent variable) and rates of polypharmacy across years was explored using multivariate logistic regression. Covariates included in the model were year (1997 as the reference), level of polypharmacy (0-4 medications (reference), 5-9 medications, and 10 medications), and sex. A sensitivity analysis was performed posthoc to assess the impact of excluding the STOPP criterion with the largest contribution to PIP in 2012, long-term use of proton pump inhibitors (PPIs) at maximal dose, from the analysis. Interaction between level of polypharmacy and year was investigated and significance was evaluated using likelihood ratio (LR) tests. Statistical analyses were performed using SAS 9.2 (SAS Institute Inc., Cary, NC, USA). Significance at p < 0.01 was assumed. 2.3 Results Regular medications and polypharmacy in the total population The number of individuals included in this study in 1997, 2002, 2007, and 2012 were 338,025, 344,270, 373,007, and 539,752 respectively. Figure 2-1 shows the proportion of each age group by the number of regular medications prescribed for each study year. There was a clear increase in the proportion of individuals on higher numbers of medications, particularly in the two oldest age groups. Of those aged years, the percentage with polypharmacy (on five or more regular medications) increased from 8.3% to 30.2% over the study period, and for those aged 65 years it rose from 17.8% to 60.4% (shown below in Table 2-1). A similar trend was observed for excessive polypharmacy (on 10 or more regular medications), with the prevalence increasing during this time from 0.8% to 8.3% in those aged years and from 1.5% to 21.9% in people 65 years. 60

63 61 Figure 2-1 Percentage of eligible population by number of regular medications for the years

64 Table 2-1 Prevalence of polypharmacy (being dispensed 5 regular medications) and excessive polypharmacy ( 10 regular medications) in each study year by age group n (%) Polypharmacy <5 years 7 (< 0.1) 45 (0.2) 68 (0.2) 130 (0.3) 5-15 years 61 (0.1) 77 (0.2) 164 (0.3) 390 (0.4) years 1406 (1.3) 2495 (2.8) 4496 (4.6) 8390 (4.6) years 5172 (8.3) (18.6) (29.6) (30.2) 65 years (17.8) (30.9) (49.8) (60.4) Total (6.1) (14.7) (23.2) (21.9) Excessive polypharmacy <5 years 0 (0.0) 5 (< 0.1) 5 (< 0.1) 20 (0.1) 5-15 years 7 (< 0.1) 5 (< 0.1) 10 (< 0.1) 25 (< 0.1) years 84 (0.1) 257 (0.3) 615 (0.6) 1262 (0.7) years 468 (0.8) 1681 (3.0) 4172 (7.0) 7707 (8.3) 65 years 1205 (1.5) 6589 (5.4) (14.3) (21.9) Total 1764 (0.6) 8537 (2.5) (6.3) (7.1) The rates of prescription per 1,000 GMS-eligible patients of the 15 most common regular medications in 2012 compared to the age-standardised and sex-standardised rates in previous years are shown in Figure 2-2. Statins were prescribed to the highest number of individuals, and similar to anti-platelet drugs and PPIs, there were large increases in the numbers of people on these medications over the study period. Several other cardiovascular drugs were among the most commonly dispensed medicines. Benzodiazepine anxiolytics were one of the few medications not to show a year on year increase over this time. Prescribing of related medicines, such as selective serotonin reuptake inhibitors (SSRIs) and non-benzodiazepine (Z-drug) hypnotics, did show an upward trend during the study period. In the negative binomial regression analysis (Table 2-2), the adjusted IRR for polypharmacy in 2012 compared to 1997 was 4.16 (95% CI 3.23, 5.36). In the model for excessive polypharmacy, the adjusted IRR for 2012 compared to 1997 was (95% CI 8.58, 12.91). For both of the polypharmacy outcomes there was a trend across the study years of increasing adjusted IRR for polypharmacy, controlling for age and sex. 62

65 63 Figure 2-2 Standardised rates of prescribing of most common regular medications in all individuals in 2012 Abbreviations: ACE, angiotensin converting enzyme; Inh, inhaled; PPIs, proton pump inhibitor; SSRIs, selective serotonin reuptake inhibitors.

66 Table 2-2 Adjusted negative binomial regression models for polypharmacy and excessive polypharmacy in all individuals Polypharmacy a Excessive polypharmacy b Adjusted IRR c (95% CI) Adjusted IRR d (95% CI) Year 1997 (reference) (1.69, 2.82) 3.53 (2.86, 4.36) (2.71, 4.49) 8.07 (6.55, 9.93) (3.23, 5.36) (8.58, 12.91) Abbreviation: IRR, incident rate ratio. a Polypharmacy defined as 5 or more regular medications. b Excessive polypharmacy defined as 10 or more regular medications. c Adjusted for age group and sex (both significant, p < 0.01). d Adjusted for age group, sex and age group-sex interaction (all significant, p < 0.01) PIP in older individuals There were 78,489 individuals aged 65 years included in 1997, 121,726 in 2002, 129,162 in 2007 and 133,884 in The prevalence of PIP in these individuals in 1997 using 30 STOPP criteria was 32.6%. This fell to 28.6% in 2002; however the percentage with PIP increased in the more recent study years to 32.8% in 2007 and 37.3% in 2012 (Table 2-3). Table 2-3 Overall prevalence of PIP in individuals aged 65 years in each study year n (%) 1997 (n=78,489) 2002 (n=121,726) 2007 (n=129,162) 2012 (n=133,884) 1 PIP (24.1) (21.5) (24.8) (27.6) 2 PIP 5219 (6.7) 6746 (5.5) 8063 (6.2) 9801 (7.3) 3 PIP 1459 (1.9) 1918 (1.6) 2307 (1.8) 3118 (2.3) Any PIP criteria (32.6) (28.6) (32.8) (37.3) A number of PIP criteria decreased in prevalence across the study period, with the largest reductions in prescribing of high doses of aspirin and digoxin (Figure 2-3). Although the rates of long-term use of benzodiazepines and non-steroidal anti-inflammatory drugs (NSAIDs) have fluctuated across the study period, the prevalence of these criteria has remained above 3%. An increase in duplication of drug classes was observed, in particular duplicate opioids. The largest increase in prevalence of any criterion was PPIs at maximum dosage for >8 weeks, which rose from 0.8% in 1997 to 23.6% in This was the major contributor to the overall PIP prevalence and a sensitivity analysis excluding this criterion showed a consistent decrease in overall PIP prevalence (32.3%, 24.9%, 22.6%, and 20.8% from 1997 to 2012, see Figure 2-4). The prevalence of individual STOPP criteria in each study year is presented in Table

67 65 Figure 2-3 Prevalence of most common types of PIP in individuals aged 65 years Abbreviations: H2, histamine-2 receptor; TCA, tricyclic antidepressant; CCB, calcium channel blocker; PPI, proton pump inhibitor; COPD, chronic obstructive pulmonary disease; NSAID, non-steroidal anti-inflammatory drug; T2DM, type 2 diabetes mellitus.

68 Table 2-4 Number and prevalence (%) of individual STOPP criteria in individuals aged 65 years in each study year n (%) Criteria description Cardiovascular system Digoxin >125 μg/day 3135 (4.0) 3557 (2.9) 2136 (1.7) 1166 (0.9) Thiazide with gout 94 (0.1) 249 (0.2) 376 (0.3) 292 (0.2) Cardioselective β blocker with COPD 149 (0.2) 247 (0.2) 375 (0.3) 369 (0.3) β blocker with verapamil 92 (0.1) 226 (0.2) 281 (0.2) 211 (0.2) Aspirin and warfarin without H2 218 (0.3) 764 (0.6) 1381 (1.1) 1616 (1.2) antagonist/ppi Dipyridamole monotherapy 308 (0.4) 230 (0.2) 84 (0.1) 41 (< 0.1) Aspirin >150mg/day (13.1) 8186 (6.7) 2528 (2.0) 1933 (1.4) Central nervous system TCA with dementia <5 (< 0.1) a 93 (0.1) 200 (0.2) 312 (0.2) TCA with glaucoma 130 (0.2) 152 (0.1) 138 (0.1) 151 (0.1) TCA with opiate/ccb 1928 (2.5) 2427 (2.0) 2777 (2.2) 3842 (2.9) Long-term benzodiazepines >4 weeks 6373 (8.1) 3999 (3.3) 6608 (5.1) 5684 (4.3) Long-term neuroleptic 627 (0.8) 867 (0.7) 1306 (1.0) 2300 (1.7) Long-term neuroleptic with 228 (0.3) 248 (0.2) 324 (0.3) 504 (0.4) Parkinson s disease Phenothiazines with epilepsy 299 (0.4) 263 (0.2) 299 (0.2) 371 (0.3) Anticholinergic and neuroleptics 191 (0.2) 334 (0.3) 327 (0.3) 456 (0.3) Use of 1 st generation antihistamines 1292 (1.7) 2023 (1.7) 1985 (1.5) 2464 (1.8) for >1 week Gastrointestinal system Prochlorperazine or metoclopramide 146 (0.2) 160 (0.1) 225 (0.2) 245 (0.2) with Parkinson s disease PPI at max dosage for >8 weeks 662 (0.8) 7700 (6.3) (15.4) (23.6) Respiratory system Theophylline as monotherapy for 2036 (2.6) 1939 (1.6) 1218 (0.9) 623 (0.5) COPD Nebulised ipratropium with 21 (< 0.1) 12 (< 0.1) 14 (< 0.1) 16 (< 0.1) glaucoma Musculoskeletal system Long-term use (>3 months) of NSAID 3215 (4.1) 7075 (5.8) 6590 (5.1) 5491 (4.1) Warfarin and NSAID 254 (0.3) 823 (0.7) 635 (0.5) 343 (0.3) Urogenitary system Antimuscarinic with dementia 0 ( 0) 205 (0.2) 679 (0.5) 850 (0.6) Antimuscarinic with chronic 28 (< 0.1) 108 (0.1) 164 (0.1) 183 (0.1) glaucoma Endocrine system Glibenclamide or chlorpropamide for 1074 (1.4) 750 (0.6) 237 (0.2) 65 (0.1) type 2 diabetes Any duplicate drug class 1247 (1.6) 3168 (2.6) 4830 (3.7) 5566 (4.2) Duplicate antidepressants <5 (< 0.1) a 51 (< 0.1) 358 (0.3) 862 (0.6) Duplicate SSRIs 7 (< 0.1) 20 (< 0.1) 31 (< 0.1) 41 (< 0.1) Duplicate opiates 443 (0.6) 1483 (1.2) 2031 (1.6) 3030 (2.3) Duplicate NSAIDs 766 (1.0) 1052 (0.9) 624 (0.5) 415 (0.3) Duplicate loop diuretics 28 (< 0.1) 81 (0.1) 101 (0.1) 105 (0.1) Duplicate ACE inhibitor 31 (< 0.1) 547 (0.5) 1759 (1.4) 1238 (0.9) 66

69 Prevalence Abbreviations: ACE, angiotensin converting enzyme; CCB, calcium channel blocker; COPD, chronic obstructive pulmonary disease; NSAID, non-steroidal anti-inflammatory drug; PPI, proton pump inhibitor; TCA, tricyclic antidepressant. a Exact numbers not reported for cells with less than five observations to avoid potential identification. 40% 30% 1 PIP 2 PIPs 3 PIPs 20% 10% 0% Figure 2-4 Overall prevalence of PIP using STOPP by number of criteria (left) and prevalence in sensitivity analysis excluding PPIs at maximal dosage for greater than eight weeks (right) The relationship between PIP and polypharmacy in older individuals The trend of PIP prevalence across the study years was confirmed in the univariate logistic regression where the odds of having any PIP were lower in 2002 compared to 1997 and then rose in 2007 and 2012 compared to 1997 (Table 2-5). After adjusting for sex and level of polypharmacy in the multivariable logistic regression, a trend of reducing odds of having a PIP across the study years was observed. The adjusted OR for having a PIP for polypharmacy compared to no polypharmacy was 6.83 (95% CI 6.73, 6.93) and for excessive polypharmacy ( 10 medications) compared to no polypharmacy was (95% CI 21.57, 22.54). In the sensitivity analysis using prevalence of any PIP excluding long-term maximal dose PPI as the outcome, an even greater reduction in the odds of having a PIP was observed over the study period (adjusted OR for PIP in 2012 was 0.2 (95% CI 0.19, 0.2) compared to in 1997). 67

70 Table 2-5 Unadjusted and adjusted logistic regression models for having any potentially inappropriate prescribing (PIP) criteria in individuals aged 65 years Any PIP Crude odds ratio (95% CI) Adjusted odds ratio a (95% CI) Year 1997 (reference) (0.81, 0.84) 0.53 (0.52, 0.54) (0.99, 1.03) 0.38 (0.38, 0.39) (1.21, 1.25) 0.39 (0.39, 0.40) a Adjusted for level of polypharmacy (0-4 medications (reference), 5-9 medications, 10 medications) and sex (all significant, p < 0.01). There was evidence that the effect of polypharmacy and excessive polypharmacy on the risk of having any PIP varied across study years (LR tests p < 0.001). When interaction terms were included between level of polypharmacy and year (see Table 2-6), the odds of PIP among those on between five and nine regular medications (polypharmacy) decreased significantly between 1997 and 2012 (adjusted OR 0.88, 95% CI 0.84, 0.93), while for excessive polypharmacy (10 or more regular medications) the odds have increased significantly with time (for 2012 relative to 1997, adjusted OR 1.19, 95% CI 1.02, 1.39). Table 2-6 Adjusted odds ratios for interaction between level of polypharmacy and study year in individuals aged 65 years Year and polypharmacy interactions Adjusted odds ratio (95% CI) a 0-4 medications (reference) (reference) (0.51, 0.53)** (0.37, 0.39)** (0.41, 0.43)** 5-9 medications 7.16 (6.88, 7.46)** (0.97, 1.07) (0.90, 0.99)* (0.84, 0.93)** 10+ medications (15.14, 20.39)** (0.98, 1.35) (1.13, 1.54)** (1.02, 1.39)* * p < 0.05 ** p < 0.01 a Also adjusted for sex. 68

71 2.4 Discussion Principal findings Between 1997 and 2012 there was a substantial increase in the prescribing of regular medications, particularly in older adults, with a fourfold increase in polypharmacy and a 10-fold increase in excessive polypharmacy, independent of age and sex. PIP prevalence also rose, largely due to increasing maximal dose PPI use that masked the reduction across most of the other PIP medicines. After controlling for changes in polypharmacy over time, there has been a reduction in the odds of having any PIP in recent years. The effect of polypharmacy on PIP has decreased with time while for excessive polypharmacy this effect has increased Findings in the context of previous research Other studies have also reported an increase over time in drugs prescribed to individual patients, though they examined different time frames to the present study.[25 27] This has implications for healthcare provision, for example the number of office-based visits by elderly patients with polypharmacy quadrupled between 1990 and 2000 in the US.[91] Higher rates of excessive polypharmacy were observed in the present study, possibly due to the proportion of individuals with lower incomes being included, as lower socioeconomic status (SES) and deprivation can be associated with polypharmacy, multimorbidity, and lower quality prescribing.[17,26,161] In terms of commonly prescribed medicines, a study of the period noted the change in treatment patterns of cardiovascular disease over this time, with rises in prescribing of calcium channel antagonists, angiotensin converting enzyme inhibitors, beta blockers and lipidlowering drugs.[154] Much research on trends in PIP found decreasing prevalence over time,[95,99,100] and the number of regular medications or polypharmacy was consistently reported as being the strongest predictor of PIP.[99,100] The only study to assess how this relationship has changed with time found that among older people with polypharmacy, the proportion who had a high anticholinergic drug burden was lower in 2010 compared to fifteen years previously.[96] In relation to specific PIP medicines, significant quantities of maximal dose PPIs continue to be prescribed. This is despite the potential cost-savings of optimising use being raised in both the USA and Ireland ($47.1 billion and 40.5 million per 69

72 annum)[160,162] and concerns regarding the clinical implications of such PPI overprescribing.[163] Long-term NSAID and benzodiazepine use in older people, defined as inappropriate in both the STOPP and Beers criteria,[159] is also of concern as such prescribing remains prevalent across countries and is associated with high-risk adverse events in vulnerable elderly patients.[39,164] Practice and policy implications The growth in prescribing in recent years, particularly in middle and older age groups, means more individuals have polypharmacy than ever before (Table 2-1), suggesting that a threshold of five or more medications may no longer specifically identify higher risk patients.[28] This is also supported by the findings that the risk of PIP in those with polypharmacy has decreased with time, whereas being on 10 or more medications has become more strongly associated with PIP since 1997 as illustrated in Table 2-6. Polypharmacy was estimated to cost US health plans at least $50 billion annually in 2002 and continued growth since then is likely to have had a major impact on pharmaceutical expenditure.[165] A number of factors may be contributing to increasing polypharmacy. The growing prevalence of multimorbidity twinned with the use of single condition-focused treatment guidelines are likely to have contributed to the higher rates of polypharmacy.[166] More patient-centred care may help address this conflict between evidence-based medicine and effective multimorbidity management.[167] There may also be a growing acceptance that the medicalisation of older age is of benefit to patients.[150] Prescribing indications for common medicines such as statins and PPIs have expanded since they were first marketed and the use of such agents has become increasingly widespread.[168,169] It appears that the mass prescription of preventive medicines is becoming more acceptable, as illustrated by the increasing number of people on statins.[170] A prescription no longer signals treatment of a sick patient and this has implications for what polypharmacy indicates in modern health care.[170] One of the few drug classes not to increase following each time period was benzodiazepine anxiolytics. While this may be an encouraging sign of preferential use of other anxiety treatments such as SSRIs, it may also reflect switching from these benzodiazepines to Z-drug hypnotics in cases of insomnia, however evidence is unclear on whether Z-drugs are a superior alternative to benzodiazepines.[171] 70

73 The relationship between polypharmacy and use of PIP drugs is complex. While polypharmacy and excessive polypharmacy are strongly associated with PIP, the picture that emerges in this setting is that when the number of prescribed medicines is accounted for, the chance of being prescribed a PIP medicine has decreased over time (see Table 2-5). These analyses illustrate the complex, competing factors influencing prescribing. On the one hand, there is increasing medicines use being driven by clinical practice guidelines and other forms of external evidence that mandate prescribing. On the other hand, there is an awareness of iatrogenic harm and that the use PIP medicines need to be justified and, if possible, limited. Interventions to improve the appropriate use of polypharmacy have been effective in reducing inappropriate prescribing but the clinical significance of such improvements is unclear.[172] Deprescribing of medications among older people with multimorbidity and polypharmacy to reduce the drug burden may also yield patient benefits however the evidence base for this approach needs to be further developed.[173] Strengths and limitations The large study population and length of the follow-up period means these results are more likely to be generalisable and less likely to be due to short-term fluctuations in prescribing practice. The study utilises primary care dispensing data from a pharmacy claims database, which allows for medications use of individuals to be explored as opposed to population-level drugs consumption. Dispensing data sources, unlike prescribing databases, can account for primary non-adherence to prescriptions and so is more likely to reflect actual medications use by patients; however, it is not known whether patients actually take the medications dispensed or if they are taking over-thecounter medicines. No clinical or diagnostic information is available in this data source. Therefore, only a subset of STOPP criteria could be applied, possibly underestimating the actual prevalence of PIP in this population. There could also be clinical justification for some instances of PIP, which cannot be identified without a full patient medical record. For example, long-term PPIs at high doses may be appropriate in managing Barrett s oesophagus. A further limitation of this study is that only people eligible for the means-tested GMS scheme in the EHB region could be included. Deprived individuals may therefore be overrepresented, as are females and the younger and older populations who are possibly 71

74 more likely to have polypharmacy.[26] However for much of the study period, all over 70 s were eligible for the GMS scheme and so for the older population, the findings are likely to be largely representative and not biased by socioeconomic status. Some forms of prescribing included may not have been considered potentially inappropriate from evidence available before the publication of the STOPP criteria in The reduction in prevalence of such criteria may illustrate evidence being incorporated into practice as it became available Conclusions This study showed a marked increase in prescribed medications over a relatively short time and increasing polypharmacy is the main driver of exposure to PIP. Prescribing of certain PIP medicines has declined, which could indicate an increasing recognition amongst prescribers that these medicines may be potentially inappropriate. Reassuringly after controlling for increasing polypharmacy, the odds of older individuals being exposed to PIP have reduced with time; given the increased medications burden, clinicians seem to be prescribing more appropriately. The increasing prevalence of PIP over time was driven by one criterion, PPIs at maximal therapeutic dose for >8 weeks, which has risen substantially since 1997 and warrants further investigation. In the future, two related issues need to be addressed. Firstly, quality improvement strategies and interventions should be considered to improve prescribing appropriateness further, particularly for prevalent PIP medicines such as benzodiazepines, NSAIDs, and PPIs. Secondly, interventions for patients need to make clear the trade-off between taking more medicines to prevent disease or minimise disability and the potential for iatrogenic harm from PIP drugs that occur more commonly with polypharmacy. 72

75 Chapter 3 Characterising potentially inappropriate prescribing of proton pump inhibitors over time 73

76 3.1 Introduction Proton pump inhibitors (PPIs) were first marketed in the late 1980 s and since this time, the volume prescribed has surpassed other acid-suppressant agents such as H2 receptor antagonists (H2RA) due to superior pharmacokinetics and efficacy to make PPIs one of the most commonly prescribed drug classes in developed countries.[ ] They were initially marketed for the treatment and prevention of gastric acid-related symptoms and diseases, however, these indications do not fully account for the rising number of prescriptions observed in recent years.[177] Their widespread use has prompted concerns regarding overprescribing of PPIs and that their use by many patients may cause more harm than benefit, particularly in patients at low risk of gastrointestinal (GI) adverse events.[174,178] Use of PPIs for longer than indicated in older people is a particular source of concern, as illustrated by the addition of this prescribing to the recently updated Beers criteria.[65] The Screening Tool for Older Person s Prescriptions (STOPP) also defines use of PPIs at therapeutic dose or higher for longer than eight weeks as potentially inappropriate for people aged 65 and over.[69] Studies suggest this is the optimal duration of high-dose prescribing to maximise ulcer healing and symptom resolution and minimise rebound acid secretion.[178] This is one of the most prevalent types of potentially inappropriate prescribing defined by STOPP in several European countries.[179] The analysis in Chapter 2 showed that this type of PIP rose sharply over the last 15 years, with almost a quarter of older participants having this criterion in 2012, and that this increase was driving the overall rise in PIP prevalence.[176] As PPIs are regarded as having a good safety profile and high efficacy,[178] such prescribing is often seen as a conservative approach to reduce the likelihood of GI bleeds in patients who are at elevated risk of such adverse events. However, using high doses for longer than indicated is not cost-effective and, as well as increasing drug burden for patients, may also put patients at risk of other adverse events, including fractures and Clostridium difficile infection.[147,178,180] The aims of this study were: (1) to characterise how PPI prescribing and co-prescribed ulcerogenic medicines have changed in community-dwelling older people in Ireland (aged 65 years) from 1997 to 2012 and (2) for long-term PPI users, to determine what coprescribed medications and demographic factors were associated with prescription of 74

77 maximal dose (rather than maintenance dose) and which factors account for the increase in long-term use of maximal doses from 1997 to Methods Study design and setting This was a repeated cross-sectional study using administrative pharmacy claims data available from the Health Service Executive Primary Care Reimbursement Service (HSE- PCRS). The study population includes all individuals aged 65 years eligible for the General Medical Services (GMS) scheme in the former Eastern Health Board region of Ireland (representing one third of older people nationally) in the years 1997, 2002, 2007, and The GMS scheme is a form of public health cover in Ireland with means-tested eligibility, however, from 2002 to 2008 all people aged 70 years or over were automatically eligible for the scheme and since January 2009, a higher income threshold was applied to this age group compared to the general population still covering 96.5% of this population.[155] Ethical approval Permission was given by the data controller of the HSE-PCRS database for this analysis as a part of the protocol for the study presented in Chapter 2. Again, data were anonymised and analysed so that no subgroup contained fewer than five individuals and so specific ethical approval for this analysis was not required Analysis PPI use For each study year, the percentage of the study population prescribed a PPI (ATC code A02BC and combinations) was determined, categorised by duration ( and > 8 weeks for short- and long-term use respectively) and for those prescribed a long-term PPI, the dosage prescribed for this period (either maximal i.e. treatment dosage or higher, or maintenance i.e. less than treatment dose). Individuals with long-term periods at both maximal and maintenance doses were categorised as the former. Treatment doses were classified as 40mg daily for omeprazole, pantoprazole, and esomeprazole, 30mg daily for lansoprazole and 20mg daily for rabeprazole, consistent with other studies and United Kingdom treatment guidelines.[160,181] These doses are in some cases higher than those indicated on Food & Drug Administration (FDA) labels and so estimates of potential 75

78 overprescribing are likely to be more conservative than if FDA dose recommendations had been used. For individuals prescribed a long-term PPI at maximal dose their concomitant use of medicines that can increase risk of GI bleeding (ulcerogenic agents including non-steroidal anti-inflammatory drugs (NSAIDs), aspirin, other antiplatelet or anticoagulant drugs, and oral corticosteroids prescribed for three months or greater) were also examined. Concomitant use was classified as a dispensing of the ulcerogenic medicine within the period of long-term PPI use. Data were only available on doctor-prescribed drugs and therefore over the counter use of NSAIDs for example could not be considered. Prevalence estimates with 95% confidence intervals (CIs) were calculated for each of these scenarios as a percentage of all persons aged 65 years and over eligible for the GMS scheme in the final month of each study year Factors associated with dosage of long-term PPIs Among long-term PPI users across all years, logistic regression was used to assess medication and demographic factors associated with being prescribed a maximal dose (considered to be potentially inappropriate) compared to a lower, more appropriate maintenance dose after eight weeks. Factors included as explanatory variables were concomitant use of ulcerogenic medicines, age group (categorised as years (reference), years, or 75 years), sex, and level of polypharmacy (number of regular medication classes prescribed to an individual for three months or longer, categorised as 0-4 (reference), 5-9, or 10 or more regular medications). Study year was also included to assess if risk of long-term users being prescribed a maximal dose changed with time, controlling for demographic and medication factors. Post-hoc sensitivity analysis considered interaction between study year and other covariates and LR tests were used to assess significance. Adjusted odds ratios (OR) with 95% CI are presented. Some individuals may be included at multiple time points, so a sensitivity analysis was undertaken to account for clustering/non-independence of repeated observations of a patient in estimating standard errors. All analyses were performed using Stata version 12 (Stata Corporation, College Station, TX, USA) and significance was assumed at p <

79 3.3 Results PPI use The number of individuals aged 65 years and over included in this study was 78,489 in 1997, 121,726 in 2002, 129,162 in 2007, and 133,884 in Previous analysis found the percentage of these older adults on five or more regular medications (i.e. with polypharmacy) increased steadily from 17.8% in 1997 to 60.4% in 2012, while those on 10 or more regular medications rose from 1.5% to 21.9% over this period.[176] Approximately half of participants were prescribed a PPI in 2007 and 2012, up from 10.7% in 1997 (see Table 3-1). Long-term prescribing became more common, with the percentage of participants on a PPI for greater than eight weeks increasing from 4.1% in 1997 to 35.5% in The prevalence of potentially inappropriate prescription of a PPI (i.e. treatment dosage or higher for greater than eight weeks) also rose over this period, from 0.8% of individuals in 1997 to 23.6% of those in Long-term use at maximal dosages increased both as a proportion of those prescribed any PPI (from 7.9% in 1997 to 48.7% in 2012) and those on a long-term PPI (from 20.5% in 1997 to 66.4% in 2012), as illustrated in Figure 3-1, and in all cases the trend across years was statistically significant. Figure 3-1 is a stacked area graph where the total height refers to all PPI prescribing and the height of each coloured section shows the proportion of the population in each subgroup. Excluding individuals also receiving a concurrent prescription for an ulcerogenic agent, 4.7% of the 2012 study population were prescribed a potentially inappropriate PPI with no concurrent medications that increase the risk of GI bleeding. The corresponding percentages were 0.3% of individuals in 1997, 1.4% in 2002, and 2.7% in 2007, a statistically significant trend Factors associated with dosage of long-term PPIs Adjusted logistic regression analysis of individuals prescribed a long-term PPI found that concurrent anticoagulant, antiplatelet (excluding aspirin), or long-term corticosteroid use, and being on five or more regular medications were significantly associated with higher odds of being prescribed a potentially inappropriate maximal dose (Table 3-2). Concurrent NSAID (adjusted OR 0.87, 95% CI 0.85, 0.89) or aspirin use (OR 0.95, 95% CI 0.92, 0.97) as well as female sex and older age were associated with lower odds of being prescribed a maximal dose among long-term PPI users. 77

80 78 Table 3-1 Number and percentage (95% CI) of individuals prescribed a PPI across study years, categorised by duration of use, dosage and concurrent medications use 1997 (n=78,489) a 2002 (n=121,726) a 2007 (n=129,162) a 2012 (n=133,884) a n % prevalence (95% CI) n % prevalence (95% CI) n % prevalence (95% CI) n % prevalence (95% CI) Prescribed a PPI (10.5, 10.9) (27.6, 28.1) (50.1, 50.6) (48.2, 48.7) Short-term PPI ( 8 weeks) (6.4, 6.7) (14.3, 14.7) (24.2, 24.6) (12.8, 13.1) Long-term PPI (>8 weeks) (4.0, 4.3) (13.2, 13.5) (25.7, 26.2) (35.2, 35.8) Long-term PPI, maintenance dose b (3.2, 3.4) (6.9, 7.2) (10.3, 10.7) (11.8, 12.1) Long-term PPI, maximal dose b (0.8, 0.9) (6.2, 6.5) (15.2, 15.6) (23.3, 23.8) - with an NSAID c (0.3, 0.3) (2.5, 2.7) (6.3, 6.6) (6.3, 6.5) - with aspirin c (0.2, 0.3) (2.9, 3.1) (8.5, 8.8) (13.0, 13.4) - with an antiplatelet d or (0.1, 0.1) (1.2, 1.3) (4.2, 4.4) (5.9, 6.2) anticoagulant drug c - with a long-term ( 3 months) (0.1, 0.1) (0.6, 0.7) (1.3, 1.4) (1.8, 1.9) corticosteroid c Long-term PPI, maximal dose b, none of above medicines (0.3, 0.3) (1.3, 1.5) (2.7, 2.8) (4.6, 4.8) a Total number of people aged 65 years eligible for General Medical Services scheme in Eastern Health Board region for each study year b Dose classified as maximal i.e. greater than or equal to treatment dosage recommend by the UK National Institute for Clinical and Care Excellence (40mg daily for omeprazole, pantoprazole, and esomeprazole, 30mg daily for lansoprazole and 20mg daily for rabeprazole), or maintenance i.e. less than treatment dose.[181] c Categories not mutually exclusive, individuals may have been co-prescribed more than one type of ulcerogenic medicine d Antiplatelet drugs excluding aspirin

81 79 Percentage of study participants 60% Short-term PPI 50% Long-term PPI at maintenance dose 40% 30% Long-term PPI at maximal dose with an NSAID or with aspirin 20% or with AP/AC 10% or with CS long-term 0% Study year Long-term PPI at maximal dose, no concurrent GI risk meds Figure 3-1 Prescribing of PPIs and ulcerogenic medicines to GMS-eligible population of the EHB region of Ireland aged 65 years in 1997, 2002, 2007 and Abbreviations: AP, antiplatelet drug (excluding aspirin); AC, anticoagulant drug; CS, corticosteroid; GI, gastrointestinal; NSAID, non-steroidal anti-inflammatory drug; PPI, proton pump inhibitor.

82 The increased odds of a long-term PPI user in 2012 being on a maximal dose compared to in 1997 remained significantly higher after adjusting for concomitant ulcerogenic medicines, age group, sex and polypharmacy in the multivariate model (adjusted OR 6.30, 95% CI 5.76, 6.88, see Table 3-2). Regression analyses were repeated by study year and the associations for each variable are shown in Figure 3-2. When the addition of interaction terms between each covariate and year in the pooled analysis was assessed in post-hoc sensitivity analysis, the only statistically significant (LR test p < 0.01) interaction was between year and excessive polypharmacy. In the sensitivity analysis, accounting for patient clustering of repeated observations across years did not significantly alter the results. Table 3-2 Unadjusted and adjusted odds ratios and 95% CIs for factors associated with maximal dose compared to maintenance dose in long-term PPI users Unadjusted OR (95% CI) Adjusted OR (95% CI) Concurrent aspirin 1.13 (1.11, 1.16) 0.95 (0.92, 0.97) Concurrent NSAID 0.84 (0.82, 0.86) 0.87 (0.85, 0.89) Concurrent antiplatelet a /anticoagulant 1.64 (1.59, 1.69) 1.36 (1.31, 1.41) Concurrent corticosteroid for 3 months 1.19 (1.13, 1.24) 1.09 (1.04, 1.15) Age group years (reference) years 0.90 (0.86, 0.94) 0.90 (0.86, 0.93) 75+ years 0.94 (0.91, 0.97) 0.86 (0.83, 0.89) Female (vs male) 0.89 (0.87, 0.91) 0.90 (0.88, 0.92) Level of polypharmacy 0-4 medications (reference) medications 1.45 (1.40, 1.50) 1.28 (1.24, 1.34) 10+ medications 2.39 (2.30, 2.49) 1.91 (1.83, 2.00) Study year 1997 (reference) (3.20, 3.83) 3.28 (2.99, 3.59) (5.24, 6.25) 4.86 (4.45, 5.32) (7.04, 8.39) 6.30 (5.76, 6.88) a Antiplatelet drugs excluding aspirin 3.4 Discussion Principal findings Although the proportion of older people in this study prescribed PPIs was four times higher in 2012 compared to 1997, the prevalence of potentially inappropriate long-term PPI at maximal dose increased almost thirtyfold over this period. Discounting instances 80

83 where an ulcerogenic drug was co-prescribed (even though higher doses of PPIs would not necessarily be indicated for gastroprotection in such cases), there were still almost 5% of people on a potentially inappropriate PPI in It was hypothesised that long-term PPI users with GI bleeding risk factors would be more likely to be prescribed a maximal dose; however, maximal-dose use was not consistently associated with such risk factors, including aspirin and NSAID use and increasing age. The rise in potentially inappropriate PPI use from 1997 to 2012 was not fully explained by changes in demographics and medication co-prescribing over this time period. Concurrent aspirin Concurrent NSAID Concurrent AC/AP Concurrent CS, 3m yrs (vs 65-69) 75+ yrs (vs 65-69) Female 5-9 meds (vs 0-4) 10+ meds (vs 0-4) Adjusted OR (95% CI) Figure 3-2 Adjusted odds ratios with 95% CIs for long-term PPI at maximal dose relative to maintenance dose for various factors stratified by study year Abbreviations: AC, anticoagulant drug; AP, antiplatelet drug (excluding aspirin); CS, corticosteroid; NSAID, non-steroidal anti-inflammatory drug; OR, odds ratio; PPI, proton pump inhibitor. 81

84 Behavioural economic theories may provide some explanation of the decision-making processes leading to increased use of maximal-dose PPIs long-term. The availability heuristic, estimating the probability of events based on how easily one recalls similar events could play a role.[48] The threat of a GI bleed is well established and may be more widely recognised by prescribers than PPI-associated adverse effects such as Clostridium difficile infection and fracture, for which evidence has only emerged more recently and is mainly observational.[147,180] Omission bias, which leads to risks from acts of omission being underestimated relative to acts of commission, may also influence decision making.[48] The more passive approach of continuing a PPI at treatment dose may be favoured over the act of commission of dose reduction or discontinuation, for fear of adverse consequences of altering the status quo.[182] Patient resistance to change due to beliefs that continuation is beneficial to their health may also impede deprescribing.[183] The unexpected association of NSAID use with lower doses of PPI could be due to the presence of confounding not measured in this study, for example a factor associated with the outcome of maximal-dose PPI prescribing and inversely associated with NSAID use, such as history of peptic ulcer disease (PUD) or symptomatic gastro-oesophageal reflux disease (GORD). Given that GI mucosal protective mechanisms are impaired with age, the relationship between increasing age and lower odds of maximal-dose PPI is also surprising and unlikely to be explained by confounding due to PUD or GORD.[184] Findings in the context of previous research Research has shown a consistent upward trend in PPI utilisation, up to a 13-fold increase in the 12 years up to 2006 in Australia.[162, ] A Dutch study found a rise in PPI prescribing to primary care patients initiated on low-dose aspirin or an NSAID from 2000 to 2012, and this increase was observed in patients across levels of risk for upper GI bleeding.[188] Regarding trends in maximal dose prescribing, a growing prevalence of high-intensity PPI use was noted in a Canadian study.[189] Considering prescribing between 1997 and 2004 in this Canadian research, polypharmacy was associated with use of high doses, while increasing age was associated with reduced odds, consistent with the present study. Interestingly the authors concluded that severity of GORD symptoms was a weak predictor of maximal-dose use. An important finding was that demographics and concomitant ulcerogenic medicines dispensed did not account for much of the increase in maximal treatment dose 82

85 prescribing from 1997 to This suggests that prescribing practices have changed independent of these factors and that many patients may not require this level of PPI prescribing. This is consistent with other studies, where a small proportion of PPI prescribing (13.5%) was associated with regular indications of PUD or taking an NSAID in French nursing homes,[190] while non-specific morbidity accounted for 46% of new prescriptions in an English primary-care setting in 1995.[185] The former study concluded that much prescribing was probably inappropriately related to general health vulnerability, reflected by polypharmacy and comorbidities.[190] Strengths and limitations Strengths of this study are the large numbers included and their representativeness of older Irish adults. The long study duration provides better evidence of actual changes in prescribing practice rather than short-term fluctuations. Due to lack of clinical information, possible predictors of prolonged use of maximal-dose PPIs such as PUD, GORD, or GI bleeding history could not be included. However, with the exception of Barrett s oesophagus and Zollinger-Ellison (ZE) syndrome, long-term PPI use at higher treatment doses is no more effective than maintenance doses for these conditions.[189] The lack of data on the setting of initiation (i.e. hospital or ambulatory care) or the indication for PPI prescriptions precluded examination of potentially important contributors to PPI overuse such as stress ulcer prophylaxis.[178,191] Similar to concomitant ulcerogenic drugs, conditions such as PUD may offer some justification for maximal-dose use, however, this would still be contrary to treatment guidelines and unlikely to benefit patients.[181,189] The use of pharmacy claims data captured all prescribed medications and allowed for medications use of individuals to be analysed rather than population-level drug consumption. Although the primary focus was prescribing of these agents, it was assumed that dispensed medications were then consumed. While this is a limitation, long-term use indicated by repeated dispensing here is more likely to indicate actual consumption compared to once-off or intermittent dispensing. Information on nonprescription use of medicines, including NSAIDs, aspirin, PPIs or alternative gastroprotective medications (e.g. antacids, H2RA) was lacking, however, patients would have to purchase these themselves; given the extra cost, this is probably uncommon. Although this is not the first study to examine factors associated with maximal-dose PPI 83

86 use in older people, this does appear to be the first since the continued increase in longterm prescribing over the last decade Practice and policy implications Deprescribing of PPIs targeted at patients at low risk of GI bleeding, through dose reduction or discontinuation where tolerated, may help reduce the occurrence of PPIassociated adverse effects for patients and associated costs.[180] Sub-optimal use of PPIs has been identified as a significant source of drug expenditure and rationalising prescribing in this area could yield substantial cost savings.[160,162] A recent systematic review has shown interventions to improve PPI prescribing are feasible, with discontinuation rates without symptom control deterioration ranging from 14% to 64%.[192] Approaches to optimise prescribing more generally in older patients may also provide improvements. OPTI-SCRIPT, a multi-faceted intervention including academic detailing by a pharmacist and medication review with web-based treatment algorithms, targeted potentially inappropriate prescribing in Irish general practice and was effective in reducing the use of maximal-dose PPIs long-term (OR 0.3, 95% CI 0.14, 0.68, comparing the intervention group to control).[193] Reducing overuse needs involvement of both healthcare professionals, to discuss PPI use, offer alternative treatments and trial discontinuation, as well as patients, who if fully informed of the risks and benefits may be more likely to try other approaches such as lifestyle changes.[177] Strategies to optimise such prescribing should extend beyond primary care as many long-term prescriptions are hospital initiated.[194] PPIs commenced in secondary care without appropriate indications, for example for stress ulcer prophylaxis, or appropriate indications for long-term prescribing are often continued following discharge leading to inappropriate prolonged use, particularly if communication of indication at the transition of care is lacking.[174,191,195,196] Providing clear information to prescribers initiating or continuing PPIs regarding indicated dosage and duration of treatment may help reduce inappropriate long-term use in ambulatory care.[197] Conclusions This study contextualises the growing trend in potentially inappropriate PPI use in recent years and demonstrates that a significant portion of this maximal-dose PPI prescribing in older patients may not be justified. Further research should focus on determining the 84

87 most effective interventions to optimise prescribing in this area and ways to implement these in practice in order to reduce PPI overuse while enhancing patient care. This supports the content validity of this criterion by demonstrating that in many cases, this type of prescribing is potentially inappropriate. 85

88 Chapter 4 Prevalence of PIP and its association with adverse outcomes in a cohort of community-dwelling older people from TILDA 86

89 4.1 Introduction In this chapter, the applicability of a number of different screening tools is assessed and the validity of PIP criteria as process measures of care is evaluated by determining their relationship with outcomes in older people. As was outlined in Chapter 1, medications are the most common healthcare intervention in developed countries,[3] and older people are particularly vulnerable to adverse effects of medications due to physiological changes in ageing and the prevalence of multimorbidity and polypharmacy.[15,34,35] Concerns regarding appropriate medication use in this age group led to several screening tools being devised to define what constitutes PIP in older people. PIP can be classified as either: (i) potentially inappropriate medications (PIMs), the use of a medication where no clear clinical indication exists or the use of an indicated medication in circumstances where the risks outweigh the benefits or a better alternative exists; or (ii) potential prescribing omissions (PPOs), not prescribing a beneficial medication for which there is a clear clinical indication.[198] Although PIP can be determined implicitly on the basis of clinician s judgement, the majority of research has determined it explicitly using published criteria and a large number of these have been developed.[198] As described in 1.3.2, the earliest such tool was Beers criteria, first published in 1991 as a list of drugs to be avoided in older nursing home residents.[61] It contained many medicines not commonly prescribed outside of the United States and updates to the Beers criteria include drugs which are more widely used internationally and also drugs to avoid with certain conditions.[64] Although earlier iterations of the Beers criteria have been used extensively in the literature, the 2012 criteria have not been widely applied.[199] The Screening Tool for Older People s Prescriptions (STOPP) and Screening Tool to Alert doctors to Right Treatment (START) were developed as screening tools for PIMs and PPOs respectively, suitable for use in European countries and have been applied and validated in the literature.[42,66,67,115] Outside of specific PIP screening tools, the Assessing Care Of Vulnerable Elders (ACOVE) indicators were developed by the RAND corporation to assess the overall quality of care of older people.[76] Several ACOVE indicators relate to PIMs and PPOs and the use of these as a PIP screening tool has been assessed and good inter-rater reliability has been demonstrated.[80,200] Much published literature on this topic has only focussed on PIMs with few studies utilising PPO screening tools to assess underprescribing. Of those 87

90 studies that considered both PIMs and PPOs, few have considered how these two forms of PIP overlap (i.e. how commonly do patients have a prescribing omission and an inappropriate medication).[125,179, ] Additionally little is known about how the prevalence of PIMs and PPOs changes as older individuals are followed up over time and the determinants of change. While screening tools such as STOPP and START can be considered process measures of medication safety,[49] it is important to establish that such prescribing does have an effect on patient outcomes, such as adverse drug events (ADEs), hospitalisations or quality of life (QoL). Older individuals, particularly those with multiple chronic illnesses, can spend a significant portion of their time attending healthcare appointments to manage their conditions.[205] By identifying and addressing factors which result in increased healthcare utilisation, this can reduce the burden of doctor visits for patients and reduce demand on and costs of health services.[206] Another important outcome is the ability to carry out activities of daily living as these are critical for people to live independently and to facilitate healthy and active ageing. Functional decline may limit a person s independence and necessitate availing of caring or institutionalisation if they cannot be supported at home. Quality of life (QoL) is increasingly recognised as a vital component of successful ageing.[207] Satisfaction of people s needs is now the focus of QoL measurement in older age, whereas previously the only aspect that was considered in older people was health status.[207] A number of studies have assessed the association of STOPP PIMs using similar outcomes; however the impact of START PPOs on patients has received little attention.[115] Much of this research has been conducted in the secondary care setting and used cross-sectional designs with limited capacity to determine the prospective relationship of PIMs and PPOs with patient outcomes. It is difficult to establish whether an association is causal using such study designs due to potential bias and confounding. Longitudinal cohort studies provide a more robust method to assess the impact of medication exposure as they allow control for baseline differences, can account for confounding by time-varying factors using appropriate methods, and may allow for inference of causal effects.[208] This study had three aims, firstly to compare the prevalence of PIP (both PIMs and PPOs) in a cohort of community-dwelling people aged 65 years and older in Ireland according to a number of screening tools. Secondly, to assess if the prevalence of PIP in this cohort 88

91 changes over time and to determine the factors associated with any change. The third aim was to determine the association of PIP detected by STOPP and START with healthcare utilisation, functional decline, and QoL in this cohort over time. 4.2 Methods Study population This prospective cohort study included participants from The Irish Longitudinal Study on Ageing (TILDA), a nationally-representative cohort study charting the health, economic and social circumstances of community dwellers aged 50 years and over. Some TILDA participants also consented to use of their administrative pharmacy claims data from the Health Service Executive Primary Care Reimbursement Service (HSE-PCRS). Participants were included in the present study if they were aged 65 years or over at baseline TILDA interview, had been followed up after two years, were eligible for the General Medical Services (GMS) scheme and provided a GMS identifier which was successfully linked to their pharmacy claims data. The GMS scheme provides free health services and prescribed medicines to eligible persons in Ireland, although a small monthly co-payment of 0.50 per prescription item has applied since October Eligibility for the GMS scheme is based on means testing, although all people aged over 70 were eligible until December 2008 when a higher income threshold was introduced for this age group compared to the general population. However approximately 96% of this age group were still eligible in 2012.[155] The STROBE standardised reporting guidelines for cohort studies have been followed in the reporting of this research.[209] Data collection TILDA participant recruitment and Wave 1 baseline data collection was carried out between October 2009 and February 2011 when participants were interviewed face-toface, completed a questionnaire and underwent a health assessment. TILDA follow-up waves are scheduled every two years, the first of which (Wave 2) was carried out from February 2012 to March The HSE-PCRS database which was the source of the medication dispensing data contains a GMS identifier, patient sex and date of birth, drug information (World Health Organisation Anatomical Therapeutic Chemical (ATC) code, strength, defined daily dosage (DDD)), quantity dispensed and date of dispensing. Linkage of participants TILDA and 89

92 HSE-PCRS dispensing data was carried out using methods published previously.[210] In brief, for GMS-eligible participants who provided a valid GMS identifier, linkage was verified by matching participant birth month, birth year and sex from the HSE-PCRS database and TILDA dataset and linkage was considered to be accurate if at least two of these matched. Medication dispensing data were extracted from the HSE-PCRS pharmacy claims database on the basis of GMS identifier for each participant in the present study for the 15 months preceding the date of their TILDA baseline and follow-up interviews and all data were anonymised after linkage was performed Ethical approval Ethical approval for TILDA was provided by the Faculty of Health Sciences Ethics Committee, Trinity College Dublin (TCD), and included provision for the secondary analysis of collected data and provision for linkage to participant s GMS dispensing information. This approval was submitted to the Royal College of Surgeons in Ireland (RCSI) Research Ethics Committee who recognised it for the purposes of the present thesis Prevalence of PIMs and PPOs Explicit criteria A subset of criteria from STOPP, Beers criteria (2012) and the third iteration of the ACOVE indicators were applied to dispensing data and information from TILDA to measure PIMs. Forty-five of sixty-five (69%) STOPP criteria, forty-two of fifty-two (81%) Beers criteria and seventeen of twenty-two (77%) ACOVE indicators relating to inappropriate medicines were used. The availability of information on diagnoses from TILDA allowed underprescribing to also be assessed. To measure PPOs, a subset of criteria from START and the ACOVE indicators were used. Fifteen of twenty-two (68%) START criteria were applicable, while twenty-one of sixty-five (34%) ACOVE indicators relating to prescribing indicated medicines could be applied. All criteria for which the necessary participant information was available in the HSE-PCRS pharmacy claims database and TILDA were applied. The most common reasons that criteria could not be applied were lack of information on particular diagnoses (e.g. chronic constipation), lack of biochemical information (e.g. receptor subtype in breast cancer, left ventricular ejection fraction in heart failure) or that criteria related to hospital inpatient prescribing, which cannot be assessed in this study as dispensing information relates to primary care. Appendix I 90

93 contains a full list of included criteria and reasons for exclusion for those that could not be applied. For STOPP, Beers criteria and ACOVE indicators for PIMs, six criteria are directly duplicated and there are nine cases where criteria may overlap to some extent. For PPO screening tools, two criteria are directly duplicated and criteria may overlap in definition in eight cases (details of which are presented in Appendix II). Directly duplicated criteria were not counted when considering number of PIMs and PPOs, while for those criteria with some overlap, these were not counted if a participant had more than one of the overlapping criteria Application of criteria The prevalence of PIP was determined in this cohort during two time intervals: (i) the 12 months preceding each participant s baseline TILDA interview and (ii) the 12 months preceding their follow-up interview. A longer period of time was analysed for criteria dependent on duration of medication use of greater than one month to allow for 12 months of potential exposure. For example from STOPP, Long-term use of NSAID (>3 months) was assessed using 15 months of data. Dispensing data from the HSE-PCRS database and information from TILDA on diagnoses, medications not included in the HSE- PCRS database (self-reported) and other characteristics were used to assess if participants had received a PIM or had a PPO. Participants were classified as having a PIM if they were prescribed the potentially inappropriate medicine at any time during the study periods, while having a PPO was classified as not receiving the indicated medicine at any time during the study periods. The HSE-PCRS database records each individual dispensed item as a single observation and these are nested in a hierarchy, within prescriptions, within months and within patients. Therefore a stepwise approach was taken to the application of PIP criteria to the data. Firstly, criteria dependent on presence or strength of a medicine were applied (for example Aspirin at dose >150 mg/day from STOPP) as these could be determined by a single dispensing. Then, data were aggregated by prescription to apply criteria dependent on co-prescribing of medicines (for example Warfarin and NSAID from STOPP). Lastly, data were aggregated by month of dispensing and remaining criteria were applied, including those dependent on omission of a medicine, presence of co-morbidity (for example Theophylline as monotherapy for COPD from STOPP) and long-term use of a 91

94 medicine (for example Long-term use of NSAID (>3 months) from STOPP). Long-term use was classified using number of DDDs dispensed, which were calculated as follows: Number of DDDs = Quantity Strength DDD For example, long-term use for more than three months was defined as dispensing of greater than 90 DDDs across four consecutive months with a dispensing in each. DDDs are a validated statistical measure of drug consumption maintained by the WHO and may be defined as the assumed average maintenance dose per day for a drug used for its main indication in adults.[211] Statistical analysis The prevalence of PIP for each time period was calculated for each set of PIM and PPO criteria separately (i.e. STOPP, Beers, ACOVE PIMs, START and ACOVE PPOs) as the proportion of the cohort with at least one criterion from the screening tool. For participants with any PIM or PPO, the prevalence was assessed by number of criteria per individual (considering any number of duplicated drug classes from STOPP as one criterion). The proportion of the study cohort with any PIM and with any PPO was determined, as well as the prevalence of each individual criterion. Additionally, the prevalence of each criterion as a proportion of the number of participants with the condition/disease or prescribed the drug of interest was assessed, where applicable. For example, of those prescribed a benzodiazepine during the study period, the proportion who were prescribed it for >4 weeks. The absolute change in prevalence between the two periods with 95% confidence intervals (CI) was determined and McNemar s test for paired groups was used to assess if the overall prevalence of each screening tool or the prevalence of individual criteria changed significantly over time. Generalised estimating equations (GEE) with exchangeable correlations were used to investigate determinants of the change in prevalence of PIMs and PPOs over time.[212] Firstly unadjusted analysis estimating change in overall PIM and PPO prevalence from baseline to follow-up was conducted. This was followed by multivariate GEE analysis which adjusted for factors known to be associated with PIP at baseline and follow-up in the models. These included sex, age, number of medicines, number of diagnosed chronic conditions (reported at TILDA interview). Level of educational attainment as an indicator of socioeconomic status (SES) was also assessed for inclusion in the models. 92

95 Level of educational attainment has a number of strengths as a measure of SES, such as being fixed from early in life and thus is time stable, and avoiding reverse causality which can occur with other measures where for example, poor health may cause loss of income.[213,214] However in older people educational attainment can have less utility as many individuals may have left education after the minimum required time and so it may only be able to separate the most advantaged from the rest of the population.[214] SES has been associated with quality of prescribing in previous research and for example, the prevalence of drug interactions has been found to vary across educational levels.[26,215] A study which investigated the most appropriate measure to use in older adults found either occupationally derived social class or education in combination with a deprivation indicator were preferred.[214] As no deprivation indicator, such as the Townsend Index, was collected in TILDA and due to limitations of using occupational class (e.g. classes can comprise heterogeneous occupations with varying incomes and prestige, difficulty in classifying homemakers and retirees, and a higher degree of missing data for this variable in TILDA),[213] educational attainment was used in this analysis. Unadjusted and adjusted odds ratios with 95% CIs are presented and statistical significance was assumed at p < Analyses were performed using SAS 9.2 (SAS Institute Inc., Cary, NC, USA) and Stata version 12 (Stata Corporation, College Station, TX, USA) Association with adverse outcomes Primary outcome The primary outcome under investigation was healthcare utilisation. Two types of utilisation were considered, emergency department (ED) visits and general practitioner (GP) visits, to determine the relationship between PIP and both urgent secondary care and routine primary care use. Healthcare utilisation was assessed during TILDA interview by asking participants how many times in the previous 12 months did they visit a hospital emergency department as a patient, and in the last 12 months about how often did they visit their GP. Although there have been concerns over the accuracy of self-reported healthcare utilisation, most often underreporting due to long recall time period,[216] hospital ED visits and inpatient admissions are often recalled with more than 90% accuracy.[217,218] Self-reported healthcare utilisation is valid across socioeconomic groups,[219] however 93

96 younger people, males, those with higher education and healthier individuals may recall more accurately.[217] Sensitivity analysis was conducted to assess the impact of inaccuracy of reported utilisation and is outlined in Section below Secondary outcomes Two secondary outcomes were also considered. The first was decline in physical functioning. Physical functioning was assessed during TILDA interview by asking participants if they have difficulty doing any of a number of named activities due to a health or memory problem (excluding any difficulties expected to last less than three months). Six Activities of Daily Living (ADLs) are asked about, encompassing basic tasks of everyday life, important for self-care and hygiene.[220] The following ADLs were included: Dressing; Walking across a room; Bathing or showering; Eating; Getting in or out of bed; Using the toilet. Any increase in the number of ADLs a participant reported difficulty with between baseline and follow-up was classified as functional decline. Secondly quality of life (QoL) was investigated. This was assessed in TILDA using the CASP- 19, a measure designed for use in middle-aged and older people which captures the domains of control, autonomy, self-realisation and pleasure.[207] This was included in the self-completion questionnaire which TILDA participants complete and return postinterview. The measure is composed of 19 statements that describe feelings, for example My age prevents me from doing the things I would like to do and participants are asked to rate on a four-point scale how often a statement describes how they feel (Often, Sometimes, Rarely or Never). These correspond to scores of 0, 1, 2 and 3 for negative statements, with reverse coding for those worded positively, so a higher score corresponds to a higher quality of life. This gives a possible range from 0 (worst quality of life) to 57 (best quality of life). The CASP-19 scale has been widely used in longitudinal studies of ageing, however psychometric evaluation has questioned its reliability and validity.[221] In TILDA, confirmatory factor analysis did not support the validity of the original CASP-19 scale and found a single-factor model which includes only 12 of the original 19 statements (CASP-R12) had excellent fit to the data and improved psychometric validity, and so this was used as the secondary outcome measure in this analysis.[222] 94

97 Exposure The exposure of interest was potentially inappropriate prescribing (PIP), including both PIMs measured by STOPP and PPOs measured by START. STOPP and START were used to determine exposure as the most applicable measures in the cohort, having the fewest criteria with zero prevalence. The number of criteria that a participant was exposed to in the 12 months preceding outcome measurement was determined separately for both screening tools. Exposure to any STOPP PIM and to any START PPO was also considered Covariates Several covariates were considered which may confound the relationship between PIP and the outcomes of interest. These included participant demographics, such as age group, sex, level of educational attainment as a measure of SES, and living arrangements. As polypharmacy and multimorbidity increase the likelihood of having PIP, number of repeat medicines and the number of doctor-diagnosed chronic conditions reported by participants were also included. In addition, specific covariates were considered in each model relating to individual outcome of interest, for example, private health insurance status and number of ED/GP visits reported by the participants in the 12 months preceding baseline in the models for healthcare utilisation. Detailed descriptions of each covariate are included in Table Statistical analysis Descriptive statistics were generated for each outcome (number of participants and percentage of subgroup with an ED visit and with functional decline, median and interquartile range (IQR) for GP visits, or mean and standard deviation (SD) for CASP-R12 score) by exposure variables and covariates of interest. For healthcare utilisation, separate regression models were fitted for the outcomes of (i) the reported number of ED visits in the 12 months preceding follow-up interview and (ii) the number of GP visits in the same period. Suitability of count model types was investigated using the countfit function in Stata which compares candidate models on the basis of residuals and tests of model fit (AIC (Akaike information criterion), BIC (Bayesian information criterion) and the Vuong test).[223] Negative binomial models were found to provide superior fit to Poisson models and zero-inflated negative binomial and Poisson models.[158] 95

98 Exposure to PIP was assessed by including two independent variables for the number of STOPP criteria and number of START criteria, categorised as 0 (reference), 1 and 2+ criteria. The effect of any PIM exposure and any PPO exposure was also considered in models by using binary exposure variables for having any STOPP and any START versus none. Unadjusted analysis was conducted including only these independent variables for PIP, followed by multivariate analysis, which adjusted for the covariates discussed previously. For covariates that can vary with time, the value at the latest time point prior to outcome measurement was used. Results are presented as incident ratio ratios (IRR) with 95% confidence intervals (CI). Table 4-1 Description of covariates adjusted for in multivariate regression models Variable Format Description of categories Age group Binary years (reference) or 75 years Sex Binary Male (reference) or Female Number of repeat drug classes a Continuous N/A Number of reported doctor-diagnosed conditions b Categorical 0 (reference) or more Level of educational attainment Categorical None/primary (reference) Secondary Third/higher Living arrangements Categorical Living alone (reference) Living with spouse Living with others Private health insurance Binary No (reference) or Yes Number of ED visits at baseline Continuous N/A Number of GP visits at baseline Continuous N/A Diagnosed mental health condition c Binary No (reference) or Yes Any hospital admission in 12 months pre Binary No (reference) or Yes follow-up Moderate physical activity at baseline Binary No (reference) or Yes Depressive symptoms d Categorical None (reference) Sub-clinical Clinical Social participation e Binary No (reference) Yes CASP-R12 score at baseline Continuous N/A a Number of medicines (defined by level 3 ATC code) dispensed in at least 3 months to a participant during the 12 months of PIP exposure measurement in HSE-PCRS (with an upper bound of 10 or more medicines). b The number of doctor-diagnosed chronic conditions reported by the participants at the TILDA interview from the following list: cardiovascular disease (heart attack, heart failure or angina), cataracts, hypertension, high cholesterol, stroke, diabetes, lung disease, asthma, arthritis, osteoporosis, cancer, Parkinson's disease, peptic ulcer, and hip fracture. 96

99 c Reported a doctor-diagnosed emotional, nervous or psychiatric problem during TILDA interview d Level of symptoms screened by the Centre for Epidemiological Studies Depression scale (CES-D) in the self-completion questionnaire at follow-up. None corresponds to a CES-D score of 0-7, subclinical to a score of 8-15, and clinical to a score of >15. e Reported social participation in any groups such as sports or social groups A similar analytical approach was then used for the secondary outcomes. For functional decline, the outcome variable used in the analysis was binary, classified as decline (an increase in the number of ADLs a participant reported difficulty with) or no decline. Logistic regression models were fitted and results are presented as odds ratios (OR) and 95% CI. For QoL, participant CASP-R12 score measured at follow up (possible range from worst to best QoL of 0 to 36) was the continuous outcome variable. Linear regression was used as CASP-R12 score had an approximately normal distribution which was confirmed in post-estimation testing of model assumptions, and results are presented as β regression coefficients with 95% CI. Statistical significance was assumed at p < Analyses were performed using Stata version 13 (Stata Corporation, College Station, TX, USA) Sensitivity analysis and marginal structural models The possibility of an additional effect in individuals exposed to both PIMs and PPOs was assessed by the addition of an interaction term to each model and LR tests were used to evaluate if this improved model fit. To evaluate whether possible overreporting of healthcare utilisation due to recall bias affected the parameter estimates, sensitivity analysis was conducted after categorising extreme numbers of GP visits reported at a value of 52 (i.e. one visit per week). Sensitivity analysis was conducted using CASP-19 score at follow up as the dependent variable to assess if this impacted on results. The impact of time-dependent confounding was investigated using marginal structural models (MSMs). As outlined in Section 1.8.1, in the analysis of longitudinal studies, multivariate regression may lead to biased estimates if a time-dependent covariate predicts subsequent exposure and is an independent predictor of the outcome, and if past exposure predicts the covariate.[114] In such cases MSMs, a two-step estimation strategy which separates confounding control for covariates that vary with time from parameter estimation and avoids overadjustment of confounders, can produce unbiased estimates.[114,208] A number of potential time-dependent confounders exist in the current analysis such as number of regular medicines and number of diagnosed 97

100 conditions, although these may not vary substantially over the relatively short study length of follow up of two years. In the first stage of using MSM, an exposure model was fitted using logistic regression to derive weights for each participant (of value w i ) which is the inverse of the probability of having the PIP exposure they did conditional on past PIP exposure and covariate history (including measurements from both baseline and follow up).[114,224] If A i1 is an individual s follow-up PIP exposure, A i0 their baseline exposure, and L i1 and L i0 their covariate values at follow up and baseline, this can be given by: w i = 1 P(A i1 L i0, A i0, L i1 ) This inverse probability weight (IPW) was used in a weighted regression analysis in the second stage, where a pseudopopulation is created by replicating each participant w i times, which eliminates confounding by the covariates used to derive weights from the parameter estimation. This reweighting balances these covariates between the exposed to PIP and unexposed groups so that, for example, the number of regular medicines at baseline is not associated with subsequent PIP exposure in the pseudopopulation. To avoid extreme values, these weights were stabilised by replacing the numerator with the conditional probability of having the exposure they did given past exposure and only baseline covariates.[114,224] This is given as follows: sw i = P(A i1 L i0, A i0 ) P(A i1 L i0, A i0, L i1 ) Similarly to exposure, attrition in longitudinal studies can bias results if those lost to follow up differ from those who were followed up and this was accounted for using censoring weights, which are similar to IPW if censoring is considered as a timedependent exposure.[224] Again, a logistic regression model was fitted to estimate the probability of remaining uncensored given past PIP exposure and covariate history. The following factors, shown to be associated with attrition in previous cohort studies of ageing,[225,226] were included as covariates: age group, sex, level of educational attainment, living arrangements, number of repeat drug classes, number of diagnosed chronic conditions, employment status (employed, retired, or other), and area of residence (Dublin, another town or city, or rural). The inverse of this probability of remaining uncensored was used to weight individuals and eliminate confounding due to 98

101 censoring. For example, if older individuals have a lower probability of remaining uncensored (i.e. if they have a higher loss to follow-up), they will be weighted more heavily in the second stage regression to give an estimate of the effect of PIP not biased by age-related loss to follow-up. The final weights used in the MSMs were the product of the stabilised IPW and censoring weight and such weighted regression models produce parameter estimates for PIP exposure which are not confounded by the included time-varying covariates or censoring.[224] MSMs were fitted in Stata version 13 (Stata Corporation, College Station, TX, USA) for each outcome for STOPP PIM exposure and separately for START PPO exposure, with adjustment for baseline covariates only. 4.3 Results Participants This study included 2,051 TILDA participants from baseline of TILDA (Figure 4-1), of which 1,107 (54.0%) were female and mean participant age (SD) in this sample was 74.8 (6.2) years, with 1,087 (53%) participants aged between 65 and 74 and 967 (47%) aged greater than 75 years (see Table 4-2). Median number of repeat drug classes was 5 with an interquartile range (IQR) of 3 to 8 and 1,414 participants (69%) had multimorbidity as they reported two or more doctor-diagnosed medical conditions. A majority of participants (51.5%) had only primary level or no education. A total of 1,753 participants completed full follow up after two years at Wave 2 (Figure 4-1) PIP exposure Prevalence of potentially inappropriate medicine use and potential prescribing omissions When assessed using STOPP, 1,081 participants (52.7%) were prescribed a PIM during the baseline study period, while prevalence of Beers PIMs and ACOVE PIMs were significantly lower (30.5% and 19.8% respectively, p < 0.05). The prevalence of PPOs at baseline was 44.8% when assessed using ACOVE indicators and 38.2% using START (see Table 4-3). Overall, 61.4% of the sample had a PIM defined by any of the screening tools, 53.3% had any PPO and 753 (36.7%) participants had both a PIM and PPO. A total of 2,963 PIMs and 2,315 PPOs were identified during this study period. 99

102 Table 4-2 Descriptive statistics for participants at baseline (Wave 1) and follow-up (Wave 2) Characteristics Baseline (Wave 1) Follow-up (Wave 2) (n=2,051) (n=1,753) Age (years, mean (SD)) 74.8 (6.17) 76.5 (6.04) Age group (years, n (%)) (53.0) 754 (43.0) (47.0) 999 (57.0) Sex (Female, n (%)) 1107 (54.0) 953 (54.4) Number of repeat drug classes (median (IQR)) 5 (3-8) 6 (3-9) Number of reported conditions (n (%)) (10.4) 88 (5.0) (20.6) 268 (15.3) (24.3) 370 (21.1) 3 or more 916 (44.7) 1027 (58.6) Level of education attainment (n (%)) None/primary 1056 (51.5) 879 (50.2) Secondary 642 (31.3) 565 (32.3) Third/higher 351 (17.1) 308 (17.6) Living arrangements (n (%)) Living alone 718 (35.0) 626 (30.5) Living with spouse 965 (47.1) 793 (38.7) Living with others 368 (17.9) 632 (30.8) Private health insurance (n (%)) 891 (43.4) 760 (43.4) Diagnosed mental health condition (n (%)) 129 (6.3) 157 (9.0) Any hospital admission (n (%)) 354 (17.3) 366 (20.9) Moderate activity (n (%)) 799 (39.0) 751 (42.8) Depressive symptoms (n (%)) a None 1172 (58.2) 1248 (74.9) Sub-clinical 613 (30.4) 277 (16.6) Clinical 230 (11.4) 141 (8.5) Social participation (n (%)) 943 (46.0) 826 (47.1) a Depressive symptoms measured by Centre for Epidemiological Studies Depression scale missing for 36 participants at baseline and 87 participants at follow-up. The most common (prevalence >2%) individual PIM criteria and PPO criteria are reported in Table 4-4 and Table 4-5 respectively. Results for all applied criteria are presented in Appendix III. The most prevalent baseline PIM criteria were aspirin with no history of coronary, cerebral or peripheral arterial symptoms or occlusive arterial event (STOPP, 19.6%), proton pump inhibitor (PPI) at full therapeutic dosage for >8 weeks (STOPP, 17.2%) and prescribing a medication with strong anticholinergic effects (ACOVE indicators, 11.9%). The most prevalent baseline PPO criteria were calcium and vitamin D supplement omission in patients with self-reported osteoporosis (ACOVE indicators and START, 14.7%), omission of a laxative in an older person with persistent pain treated with 100

103 opioids (ACOVE indicators, 11.0%) and omission of a gastro-protective agent with an NSAID in participants with a risk factor for bleeding (ACOVE indicators, 10.1%). Figure 4-1 Flow diagram of study participants from TILDA cohort aged 65 years Change in prevalence over time The prevalence of PIMs and PPOs increased significantly (p < 0.05) between the baseline and follow-up study periods for each screening tool (also shown in Table 4-3). PIMs determined using STOPP were most prevalent (56.1%), and PPO prevalence was highest when determined by ACOVE indicators (49.3%). At follow-up, 64.8% of participants received a PIM and 56.6% received a PPO defined by any of the screening tools, while the proportion of the sample with both a PIM and PPO increased to 41.1% (843 participants). 101

104 Table 4-3 Number and percentage of participants with PIMs and PPOs at baseline and two year follow-up Baseline Follow-up Screening tool n % (95% CI) n % (95% CI) STOPP 1, (50.5, 54.8) 1, (54.0, 58.3) (27.8, 31.8) 13.2 (11.7, 14.6) 9.8 (8.5, 11.0) (27.4, 31.3) 15.0 (13.4, 16.5) 11.8 (10.4, 13.2) Beers criteria (28.5, 32.5) (31.0, 35.1) (14.3, 17.4) 9.4 (8.1, 10.6) 5.3 (4.3, 6.2) (15.4, 18.6) 9.9 (8.6, 11.2) 6.1 (5.1, 7.2) ACOVE indicators (18.1, 21.6) (20.2, 23.8) (14.8, 18.0) 3.0 (2.3, 3.8) 0.4 (0.2, 0.7) (16.5, 19.8) 3.1 (2.3, 3.8) 0.8 (0.4, 1.2) Any above PIM a 1, (59.3, 63.5) 1, (62.8, 66.9) START (36.1, 40.3) (38.4, 42.6) (24.9, 28.8) 8.3 (7.1, 9.5) 3.0 (2.2, 3.7) (26.6, 30.5) 8.3 (7.1, 9.5) 3.6 (2.8, 4.4) ACOVE indicators (42.6, 46.9) 1, (47.1, 51.5) (20.9, 24.5) 13.6 (12.1, 15.1) 8.5 (7.3, 9.7) (22.2, 25.9) 16.1 (14.5, 17.7) 9.1 (7.8, 10.3) Any above PPO a 1, (51.2, 55.5) 1, (54.5, 58.8) a PIM and PPO screening tools are not mutually exclusive, overall prevalence of PIMs and PPOs accounts for any overlap. 102

105 103 Table 4-4 Prevalence of individual PIM criteria (prevalence 2%) at baseline and two year follow-up Baseline Criteria description n % of sample STOPP Cardiovascular system n Follow-up % of sample Change in prevalence (95% CI) Loop diuretic for dependent ankle oedema only (-0.2, 1.9) Aspirin with history of PUD without H2 receptor antagonist or PPI (-0.7, 0.5) Aspirin with no history of coronary, cerebral or peripheral arterial symptoms or occlusive arterial event Central nervous system (-2.6, 0.2) TCAs with an opiate or calcium channel blocker (0.4, 1.9)** b Long-term (>1 month), long-acting benzodiazepines (-1.4, -0.1)* b Gastrointestinal System PPI at full therapeutic dosage for >8 weeks (3.2, 6.3)*** b Musculoskeletal system NSAID with history of PUD, unless with concurrent H2 receptor antagonist, PPI or misoprostol (-0.5, 0.9) NSAID with moderate-severe hypertension >160/100 mmhg a (-0.4, 2.3) Long-term use of NSAID (>3 months) (-0.6, 1.4) Long-term corticosteroids (>3 months) as monotherapy for rheumatoid arthrtitis/osteorarthritis Drugs that adversely affect fallers (-0.1, 1.1) Benzodiazepines in those prone to falls (-1.3, 0.5) Neuroleptic drugs in those prone to falls (-0.1, 1.5) Analgesic drugs

106 104 Baseline Criteria description n % of sample n Follow-up % of sample Change in prevalence (95% CI) Regular opiates for >2 weeks without concurrent use of laxatives (-1.4, 0.8) Duplicate drug classes Any regular duplicate drug class prescription (-0.1, 1.6) Beers criteria (2012) Anticholinergics Antispasmodics (-0.4, 1.4) Central nervous system Tertiary TCAs (0.2, 2.0)* b Benzodiazepines, short, intermediate and long acting (-2.3, 0.2) Non-benzodiazepine (Z drug) hypnotics, avoid chronic use >90 days (-0.3, 1.1) Gastrointestinal Metoclopramide (0.2, 1.8)** b Pain Non-COX-selective NSAIDs, avoid chronic use (-1.6, 0.4) Drug-Disease interactions Avoid with history of falls/fractures (fracture and fall or >1 fall or >1 fracture) (total) (0.8, 3.5)** b Anticonvulsants (0.5, 2.0)*** b Benzodiazepines (-0.4, 1.6) Z drugs (0.0, 1.7)* b SSRIs (-0.5, 1.0) ACOVE indicators Falls and Mobility Problems If 2 falls (or 1 fall with injury) in previous year discontinue benzodiazepine (-0.8, 0.9)

107 105 Baseline Criteria description n % of sample Hypertension n Follow-up % of sample Change in prevalence (95% CI) If a vulnerable elder (VE) has HTN discontinue NSAID or COX-2 inhibitor a (-0.9, 1.0) Medication Use Discontinue benzodiazepine if taking for >1 month (-1.4, -0.1)* b Avoid medication with strong anticholinergic effects (0.6, 3.6)** b In iron-deficiency anaemia, prescribe no more than one low-dose oral iron tablet daily (0.0, 1.6)* b Abbreviations: ACE, angiotensin converting enzyme; COPD, chronic obstructive pulmonary disease; COX, cyclo-oxygenase; GI, gastrointestinal; HTN, hypertension; MI, myocardial infarction; NSAID, non -steroidal anti-inflammatory drug; PPI, proton pump inhibitor; PUD, peptic ulcer disease; SSRI, selective serotonin reuptake inhibitor; TCA, tricyclic antidepressant; TIA, transient ischaemic attack; VE, vulnerable elder. * McNemar s test p < 0.05, ** McNemar s test p < 0.01, *** McNemar s test p < a Hypertension defined using objectively measured blood pressure or self-reported hypertension diagnosis with antihypertensive medication. b McNemar s test calculated using non-exact p value if >20 individuals changed exposure status between time periods (i.e. had criterion during one time period only).

108 106 Table 4-5 Prevalence of individual PPO criteria (prevalence 2%) at baseline and two year follow-up Baseline Criteria description n % of sample START Cardiovascular System n Follow-up % of sample Change in prevalence (95% CI) Warfarin (or another oral anticoagulant) in the presence of chronic atrial fibrillation (0.9, 2.6)*** c Aspirin/clopidogrel with a history of atherosclerotic coronary, cerebral or peripheral vascular disease (-0.5, 0.8) Antihypertensive therapy where systolic blood pressure >160 mmhg a (-3.0, -1.1)*** c Statin therapy with a history of coronary, cerebral or peripheral vascular disease (0.0, 1.4)* c ACE inhibitor following acute MI (-0.6, 0.3) β blocker with chronic stable angina (-0.7, 0.4) Respiratory System Regular inhaled β2 agonist or anticholinergic agent for mild to moderate asthma or COPD (-0.1, 1.5) Musculoskeletal System Bisphosphonates if taking oral corticosteroids for >3 months (-0.5, 1.4) Calcium and vitamin D supplement with osteoporosis (0.1, 2.6)* c Endocrine System Antiplatelet therapy in diabetes mellitus if 1 major CV risk factor (hypertension, hypercholesterolaemia, smoking history) (-0.2, 1.0) Statin therapy in diabetes mellitus if 1 major CV risk factor (-0.6, 0.6) ACOVE indicators COPD If a VE has COPD, prescribe a rapid-acting bronchodilator (0.5, 1.8)*** c

109 107 Baseline Criteria description n % of sample If a VE with COPD has 2+ exacerbations requiring antibiotics/oral corticosteroids in the previous year, then (in addition to a long-acting bronchodilator) prescribe inhaled steroids (if not taking oral steroids) Diabetes If a VE with diabetes mellitus not on anticoagulant or antiplatelet, then daily aspirin should be prescribed Hypertension n Follow-up % of sample Change in prevalence (95% CI) (0.4, 1.7)*** c (-0.4, 0.7) If a VE with HTN has IHD, prescribe a β blocker b (-0.5, 0.6) If a VE with HTN has a history of HF, IHD, chronic kidney disease, or CV accident, prescribe (-0.5, 0.7) an ACE inhibitor/arb b Ischaemic heart disease If a VE has had an MI, prescribe a β blocker (-0.7, 0.5) If a VE has IHD, prescribe an ACE inhibitor/arb (-0.6, 0.7) Medication use If a VE with a risk factor for GI bleeding (aged 75, PUD, warfarin use, chronic glucocorticoid use) is prescribed a non-selective NSAID, treat concomitantly with misoprostol/a PPI Osteoporosis If a VE without osteoporosis is taking 7.5 mg/d of prednisone (or equivalent) for 1 month, prescribe calcium and vitamin D (-2.8, 0.4) (-0.9, 0.9) If a VE has osteoporosis, prescribe calcium and vitamin D supplements (0.1, 2.6)* c If a female VE has osteoporosis, treat with bisphosphonate, raloxifene, calcitonin, HRT, or teriparatide If a male VE has osteoporosis, treat with bisphosphonate, calcitonin, parathyroid hormone, or testosterone (2.0, 4.2)*** c (0.4, 1.6)*** c

110 108 Baseline Follow-up Criteria description n % of sample Pain n % of sample Change in prevalence (95% CI) If a VE with persistent pain is treated with opioids, prescribe a stool softener/laxative (0.3, 3.6)* c Stroke If a VE has had a TIA or stroke, prescribe antiplatelet/anticoagulant therapy (0.2, 1.4)** c Abbreviations: ACE, angiotensin converting enzyme; ARB, angiotensin II receptor blocker; COPD, chronic obstructive pulmonary disease; COX, cyclo-oxygenase; CV, cardiovascular; GI, gastrointestinal; HF, heart failure; HRT, hormone replacement therapy; HTN, hypertension; IHD, ischaemic heart disease; MI, myocardial infarction; NSAID, non-steroidal anti-inflammatory drug; PPI, proton pump inhibitor; PUD, peptic ulcer disease; TIA, transient ischaemic attack; VE, vulnerable elder. * McNemar s test p < 0.05, ** McNemar s test p < 0.01, *** McNemar s test p < a 661 participants (32.2%) had missing data for measured blood pressure. b Hypertension defined using objectively measured blood pressure or self-reported hypertension diagnosis with antihypertensive medication. c McNemar s test calculated using non-exact p value if >20 individuals changed exposure status between time periods (i.e. had criterion during one time period only).

111 The prevalence of PIMs increased between waves and the unadjusted odds ratio (OR) for the presence of any PIM, comparing follow-up to baseline, was 1.08 (95% CI 1.03, 1.13), using unadjusted GEE analysis. A multivariate GEE model (Table 4-6) showed that female sex, age and higher number of medicines were significantly associated with change in PIM prevalence and the change in prevalence at follow-up compared to baseline was not significant after adjusting for these covariates. Similarly for PPO prevalence, the association for follow-up compared to baseline in the unadjusted analysis (OR 1.07, 95% CI 1.02, 1.11) was no longer significant in the multivariable model (Table 4-6), where age and higher numbers of medicines and chronic conditions were found to be significantly associated with change in PPO prevalence. When included, level of education was not significant in the models and adjusting for it did not alter any of the other parameter estimates. Table 4-6 Population-averaged GEE models for change in sample prevalence of PIMs and PPOs Any PIM Adjusted odds ratio (95% CI) (n=2046 a ) Any PPO Follow-up (vs baseline) 1.00 (0.95, 1.06) 0.97 (0.92, 1.02) Age (years) 1.03 (1.02, 1.04)* 1.03 (1.02, 1.04)* Female (vs male) 1.27 (1.07, 1.5)* 0.86 (0.72, 1.01) Number of medicines b 1.20 (1.17, 1.24)* 1.04 (1.01, 1.07)* Number of chronic conditions b 1.05 (0.99, 1.11) 1.47 (1.39, 1.56)* * z score p < 0.05 a Self-reported number of medicines was missing at both time points for 5 (0.2%) participants. b Continuous variables with odds ratios for each one unit increase in the number of medicines/chronic conditions. The most common PIM and PPO criteria (prevalence >2%) during the follow-up period are also presented in Table 4-4 and Table 4-5 respectively. A number of individual criteria showed highly significant (p < ) increases in prevalence between baseline and follow-up, including prescription of PPIs at full therapeutic dosage for >8 weeks (STOPP, 17.2% to 21.9%), use of contraindicated medicines in dementia (Beers, 0.3% to 1.3%), omission of warfarin in atrial fibrillation (START, 7.5% to 9.3%) and omission of treatment for osteoporosis in females (ACOVE, 9.1% to 12.1%). The only criteria that significantly decreased in prevalence were prescription of long-term (>1 month) long-acting benzodiazepines (STOPP/ACOVE, 3.9% to 3.1%) and omission of antihypertensives in participants with elevated blood pressure (START, 5.5% to 3.5%). 109

112 4.3.3 Association of STOPP and START with patient outcomes Primary outcome - Healthcare utilisation In the 12 months preceding Wave 2 interview, 16.1% of participants reported one ED visit, 3.8% reported two visits and 1.8% reported three or more (see Figure 4-2). Figure 4-3 shows that during the same time frame, 96.1% of participants reported visiting a GP (median 4 visits, IQR 2.5-6) Number of ED visits in 12 months before Wave 2 Figure 4-2 Distribution of number of ED visits in previous 12 months reported by participants at follow-up interview Results of the healthcare utilisation analysis are presented in Table 4-7. In the multivariate model for ED visits including exposure to both STOPP and START adjusted for covariates, presence of any STOPP PIM was significantly associated with higher rates of visits while the presence of a START PPO was not significantly associated. When number of criteria was considered, there was a statistically significant increase in the rate of ED visits for those with two or more STOPP criteria (adjusted IRR 1.42, 95% CI 1.06, 1.91) as well as for multiple START criteria (adjusted IRR 1.45, 95% CI 1.03, 2.04) relative to those with no criteria. For GP visits (Table 4-7), having any STOPP PIM was associated with an increased rate of visits and having any PPO determined by START was not associated with a significant increase. In the model including number of criteria, the relationship of STOPP 110

113 persisted for having one STOPP PIM (IRR 1.14, 95% CI 1.05, 1.25) and two or more STOPP PIMs (IRR 1.16, 95% CI 1.06, 1.28) while two or more START PPOs were also significantly associated with increased GP visits (adjusted IRR 1.13, 95% CI 1.01, 1.27). Table 4-7 Number (percentage) with an ED visit and median (IQR) GP visits the 12 months preceding follow-up by subgroup and adjusted incident rate ratios (95% CI) for ED visits (n=1,748) and GP visits (n=1,741) n (%) Emergency department visits Adjusted IRR (95% CI) a PIP exposure Model 1 b Model 2 c Any STOPP PIM (vs none) 246 (25.6) 1.30 (1.02, 1.66)* - Number of STOPP PIMs 0 (reference) 134 (17.0) (22.4) (0.94, 1.62) (29.6) (1.06, 1.91)* Any START PPO (vs none) 174 (26.1) 1.23 (0.98, 1.53) - Number of START PPOs 0 (reference) 206 (19.0) (24.3) (0.90, 1.46) 2 56 (30.9) (1.03, 2.04)* GP visits Median (IQR) Adjusted IRR (95% CI) a PIP exposure Model 1 b Model 2 c Any STOPP PIM (vs none) 4 (2-5) 1.15 (1.06, 1.24)* - Number of STOPP PIMs 0 (reference) 4 (2-5) 1 4 (3-6) (1.05, 1.25)* 2 5 (3-8) (1.06, 1.28)* Any START PPO (vs none) 4 (3-6) 1.04 (0.97, 1.12) - Number of START PPOs 0 (reference) 4 (2-6) 1 4 (2-6) (0.93, 1.09) 2 5 (3-10) (1.01, 1.27)* * p < 0.05 a Adjusted for age group, sex, number of repeat drug classes, number of reported conditions, level of educational attainment, living arrangements, private health insurance status and number of ED/GP visits reported at baseline. b PIP exposure assessed using binary variables for presence or absence of STOPP and START. c PIP exposure assessed using categorical variables for presence of 0, 1 and 2 STOPP and START criteria. 111

114 Number of GP visits in 12 months before Wave 2 Figure 4-3 Distribution of number of GP visits in previous 12 months reported by participants at follow-up interview Secondary outcome - Functional decline Difficulties with ADLs were reported by 7.7% of participants at baseline and 8.3% of participants reported an increase in ADLs which caused difficulty at follow-up. In the multivariate logistic regression analysis having any START PPO was significantly associated with functional decline, with a larger effect in the dose-response model for those with multiple criteria (adjusted OR 2.06, 95% CI 1.25, 3.39), however no evidence of an effect due to STOPP was found (Table 4-8) Secondary outcome - Quality of life CASP-R12 scores at follow up ranged from 5 to 36 and the mean score was 26.2 (SD 5.2) and the distribution of scores is shown in Figure 4-4. Although there was a degree of right-hand skewness, the distribution was approximately normal (with skewness and excess kurtosis of and 0.24 respectively, both within the conventionally used limits for normality)[227] and data did not violate any assumption of linear regression as confirmed during model checking. Multivariate linear regression found that neither presence of any STOPP nor any START criteria was significantly associated with CASP-R12 score (Table 4-9). In Model 2, although there was no evidence of a relationship with 112

115 STOPP PIMs, exposure to two or more START PPOs was associated with a small but statistically significant reduction in QoL (adjusted β coefficient -1.05, 95% CI -1.83, -0.26). Table 4-8 Number (percentage) with an increase in ADL difficulties (functional decline) between baseline and follow-up by subgroup and adjusted odds ratios (95% CI) for functional decline (n=1,753) n (%) Functional decline Adjusted OR (95% CI) a PIP exposure Model 1 b Model 2 c Any STOPP PIM (vs none) 110 (11.0) 1.23 (0.78, 1.92) - Number of STOPP PIMs 0 (reference) 35 (4.6) 1 45 (8.5) (0.75, 2.02) 2 65 (13.8) (0.75, 2.06) Any START PPO (vs none) 84 (11.5) 1.55 (1.07, 2.25)* - Number of START PPOs 0 (reference) 61 (6.0) 1 50 (9.8) (0.89, 2.04) 2 34 (15.4) (1.25, 3.39)* * p < 0.05 a Adjusted for age group, sex, number of repeat drug classes, number of reported conditions, level of educational attainment, living arrangements, reporting diagnosis of a mental health conditions, reporting a hospital admission in the 12 months preceding follow-up and reporting moderate activity at baseline b PIP exposure assessed using binary variables for presence or absence of STOPP and START c PIP exposure assessed using categorical variables for presence of 0, 1 and 2 STOPP and START criteria Table 4-9 Mean (SD) of CASP-R12 quality of life score at follow-up by subgroup and adjusted β coefficient (95% CI) for CASP-R12 score (n=986) Mean (SD) CASP-R12 score Adjusted β coefficient (95% CI) a PIP exposure Model 1 b Model 2 c Any STOPP PIM (vs none) 25.5 (5.4) (-0.81, 0.29) - Number of STOPP PIMs 0 (reference) 27.0 (5.0) (5.1) (-0.81, 0.39) (5.5) (-1.16, 0.27) Any START PPO (vs none) 25.5 (5.4) (-0.75, 0.26) - Number of START PPOs 0 (reference) 26.7 (5.1) (5.2) (-0.48, 0.64) (5.7) (-1.84, -0.27)* * p <

116 a Adjusted for age group, sex, number of repeat drug classes, number of reported conditions, level of educational attainment, living arrangements, level of depressive symptoms, reporting social participation and CASP-R12 score at baseline b PIP exposure assessed using binary variables for presence or absence of STOPP and START c PIP exposure assessed using categorical variables for presence of 0, 1 and 2 STOPP and START criteria CASP-R12 score at Wave 2 Figure 4-4 Distribution of participant CASP-R12 scores at follow-up Sensitivity analysis and marginal structural models Variables for the interaction between STOPP and START showed no statistically significant association (p > 0.05) with any of the outcomes and LR tests provided no evidence of improved model fit and therefore, interactions terms were not included. Capping the number of GP visits at follow up at a maximum value of 52 or using CASP-19 as alternative outcomes in the analysis of healthcare utilisation and QoL respectively had no significant impact on the results. In the marginal structural models for STOPP, analyses were weighted by the inverse probability of exposure to a STOPP PIM, conditional on baseline and time-varying covariates (including exposure to STOPP before Wave 1), and the probability of censoring to account for loss to follow-up at Wave 2. In this analysis for ED visits, the IRR decreased in magnitude and became marginally non-significant (IRR 1.27, 95% CI 0.99, 1.64) while 114

117 for GP visits the estimate also decreased slightly, but remained significant (IRR 1.15, 95% CI 1.06, 1.25) (see Table 4-10). For the outcomes of functional decline and quality of life, adjusting for time-varying confounding and censoring also had little effect on the parameter estimates. Table 4-10 Sensitivity analysis comparing parameter estimates with 95% CIs by outcome for multivariate regression and marginal structural models (MSMs) Outcome (parameter) Unweighted analysis p Weighted analysis (MSMs) a STOPP ED visits (IRR (95% CI)) 1.31 (1.02, 1.67) (0.99, 1.64) GP visits (IRR (95% CI)) 1.15 (1.06, 1.24) (1.06, 1.25) Functional decline (OR (95% CI)) 1.21 (0.77, 1.89) (0.71, 2.01) CASP-R12 score (β coeff (95% CI)) (-0.81, 0.29) (-0.92, 0.30) START ED visits (IRR (95% CI)) 1.24 (0.99, 1.54) (0.99, 1.60) GP visits (IRR (95% CI)) 1.04 (0.97, 1.12) (0.95, 1.14) Functional decline (OR (95% CI)) 1.54 (1.06, 2.24) (1.1, 2.34) CASP-R12 score (β coeff (95% CI)) (-0.76, 0.26) (-0.58, 0.50) a Weighted by product of stabilised inverse probability of exposure and probability of remaining uncensored at follow-up p Accounting for potential time-varying confounding of the relationship between exposure to START and healthcare utilisation showed little impact in both the ED and GP visit models. The IRR for ED visits with any STOPP criterion did decrease and became marginally non-significant in the MSM, which could have been due to a lack of effect or reduced power to detect an effect in this model. With regard to functional decline, the adjusted odds ratio increased slightly from 1.54 (95% CI 1.06, 2.24) in the standard multivariate regression analysis to 1.61 (95% CI 1.10, 2.34) following adjustment for timevarying confounding. This increase may suggest a degree of confounding by indication by a time-dependent covariate. 4.4 Discussion Principal findings This study showed in a cohort of 2,051 community-dwelling people aged 65 and over, more than 61% received a PIM in a one year period defined by a subset of STOPP criteria, Beers criteria and ACOVE indicators, while 53% had a PPO defined by a subset of START criteria and ACOVE indicators. Aspirin with no record of previous coronary, cerebral or 115

118 peripheral arterial symptoms or occlusive arterial event (i.e. apparent use for primary prevention), prolonged use of PPIs at full therapeutic dosage and strong anticholinergic drugs were the most common PIMs while common omissions were calcium and vitamin D in osteoporosis, laxatives for patients on opioids and gastro-protection with an NSAID. The increase in PIM and PPO prevalence between baseline and follow-up was associated with patient characteristics (age, female sex, numbers of prescribed medicines and chronic conditions) rather than being a function of time. The most applicable screening tool for PIMs was STOPP and START for PPOs given these had the fewest criteria with a prevalence of zero. Older people in this study who were prescribed a STOPP PIM visited the ED and their GP more often (for those with two or more PIMs, 42% and 16% increases in rate respectively), however no evidence of a relationship with functional decline and QoL was found. Participants with multiple PPOs had higher rates of healthcare utilisation (45% higher rate of ED visits and 13% more GP visits) and a small reduction in QoL. Having a START PPO was also associated with higher odds of functional decline over a two year period. Time-varying confounding did not appear to play a role in these associations Findings in the context of previous research Prevalence of PIMs/PPOs and change over time It is not surprising that the prevalence varies depending on the screening tool used as each tool includes different types of prescribing in what is classified as potentially inappropriate. Also the Beers criteria and ACOVE indicators were first developed for use in the United States whereas STOPP and START are more widely applicable, as was found here. A number of studies have estimated the prevalence of PIP using multiple screening tools. Prevalence estimates have been highly variable as research has been carried out across settings (hospitals, residential care, community), in a number of countries and using data ranging from full clinical records to administrative data meaning only a subset of PIP criteria have been applied.[202] A systematic review of studies applying the STOPP and/or START criteria found prevalence ranging from 21% to 79% for PIMs and from 22% to 74% for PPOs.[115] PIM prevalence according to the Beers criteria varied from 3-40% in studies included in a review of the 1991 Beers criteria, and up to 53.4% using more recent iterations of Beers criteria.[228,229] The ACOVE indicators have not been applied extensively, with two 116

119 studies reporting the prevalence of ACOVE PPOs giving estimates of 58% and 58.5%.[80,230] A previous study of TILDA participants using STOPP/START reported lower PIM and PPO prevalence at baseline interview than the current study (14.6% and 30%),[204] however fewer criteria were applied and prevalence was measured at one time point (the date of TILDA interview) rather than over a period of 12 months which may explain the higher prevalence in the present analysis. A number of studies have assessed the prevalence of both PIMs and PPOs, however the present study appears to be the first to report on the proportion of study participants with concurrent PIMs and PPOs.[125,179, ] An association has been demonstrated between polypharmacy (using 5 medications concomitantly) and underprescribing,[157] so the high rate of concurrent PIMs and PPOs is not unexpected. Few epidemiological studies have reported the longitudinal prevalence of PIP within a cohort and findings have shown a trend of PIP decreasing over time.[100,231] A cohort study which, like the present study, controlled for numbers of prescribed medicines and co-morbidities found that sub-optimal prescribing remained unchanged or decreased over a 4 year follow-up period after adjusting for these factors.[232] Association of STOPP/START with outcomes Studies on the impact of PIP on patient outcomes have predominantly used crosssectional or retrospective cohort designs so this study is one of few to examine the prospective relationship between STOPP and START and patient outcomes.[115] One prospective study of older hospitalised patients found a significant association between STOPP and avoidable ADEs,[102] and this is supported by other work on ADEs.[ ] For the outcomes examined in the present study, the findings appear to be consistent with previous research, in that the weight of evidence supports an association of STOPP with hospital visits,[106,107] while fewer studies have shown an effect of STOPP on health-related QoL,[106] vulnerability,[107] and functional decline during hospital stay,[104] and START on non-cardiovascular mortality.[108] All studies that have applied STOPP and START together in the same study have been hospital based with limited research on older populations in the primary care setting. Secondary analysis of data from a trial of a hospital pharmacist intervention found the only significant association was between number of STOPP criteria and number of medication-related hospital readmissions,[110] and that both STOPP and START had poor 117

120 discriminative ability to identify older patients at risk of unplanned rehospitalisation or death.[111] A study of patients following hip fracture showed higher all-cause mortality among patients with a greater combined number of STOPP and START criteria.[112] Independent of confounders, each additional criterion was associated with a 28% increase in the odds of death (adjusted OR 1.28, 95% 1.07, 1.52). A case-control study of medication-related hospital admissions found an association with STOPP criteria and a composite of STOPP and START.[113] Studies that have used different measures of inappropriate prescribing have also found an association with adverse outcomes in community-dwelling older people, such as the Medication Appropriateness Index and high-risk prescribing classified using the Drug Burden Index.[58,233] A recent trial in general practice targeting high-risk use of NSAIDs and antiplatelet drugs significantly reduced not only the targeted prescribing but also the rate of hospitalisations for related adverse events.[83] Practice and policy implications Prevalence of PIMs/PPOs and change over time Potentially inappropriate prescribing in older people is a common issue and warrants attention to improve the quality of care provided to this age group. However, this complex problem may not be fully captured using administrative data.[59,234] It is possible that patients did not respond or had contraindications to the first-line treatment resulting in a PIM being prescribed or an indicated medicine being omitted. Prescribers may have to weigh up the incremental benefit of one additional indicated medicine against increasing the treatment burden in older patients already taking multiple medications. The strength of evidence of inappropriateness varies across criteria and the risk-benefit ratio may have changed since PIP screening tools were developed. For example, aspirin with no previous coronary, cerebral or peripheral arterial symptoms or occlusive arterial event is included in STOPP but evidence is conflicting on the net benefit of aspirin for primary prevention in people with cardiovascular risk factors but without previous cardiovascular events/symptoms.[235,236] The structure of the health system in Ireland, in particular the lack of implementation of a co-ordinated chronic disease management policy across primary and secondary care, may also contribute to the rate of PIP, and future implementation of this policy could have a positive impact on prescribing.[136] 118

121 A high proportion of study participants had both prescribing issues relating to acts of commission (PIMs) and omission (PPOs). This suggests that reviewing both suitability of current medicines and assessing the need for additional indicated therapies is necessary to optimise prescribing for older people. The long-term prescription of full therapeutic dosage PPIs has been identified previously as a particularly common issue in older people and represents a significant cost burden [160]. Though the cost-effective use of PPIs in Ireland has been promoted through policies such as reference pricing and the HSE Medicines Management Programme s preferred drug iniative, a focus on prescribing appropriate dosages and durations may provide clinical benefits to patients as well as cost-savings [139,180]. Although long-term benzodiazepine use declined in this study, this may be explained by substitution with Z-drug hypnotics which showed a non-significant increase in prevalence. Omission of antihypertensive therapy also declined at follow up, which may have been due to participants with high blood pressure at baseline interview being advised to discuss this with their doctor Association of STOPP/START with outcomes Patients with either STOPP PIMs or START PPOs appear to have poorer outcomes, so incorporating review of these criteria into the care of older people and acting to rectify situations defined as inappropriate may benefit patients. When screening tools such as STOPP and START were developed, criteria were included if deemed by expert consensus to be potentially inappropriate with a marginally unfavourable risk-benefit ratio. This study provides evidence to support that this is the case and that there is an association between such prescribing and harm for patients. This is independent of the effect of number of medications, lending credence to the view that polypharmacy itself is not necessarily detrimental, but can be if it includes inappropriate prescribing.[32] However given the limited time available to healthcare professionals to review and optimise treatment, the modest size of the effect of PIP should be considered when prioritising issues to spend time on with patients. If reviewing PIP can be incorporated easily into routine clinical practice, for example through clinical decision support systems or by streamlining explicit measures to focus on fewer high-risk criteria, using these screening tools may be an efficient way to avoid extra healthcare utilisation, functional decline or reduced QoL. Further research should consider the cost-effectiveness of such approaches and large-scale prospective cohort studies or economic modelling would provide evidence to identify the most clinically significant prescribing issues to focus on in practice. 119

122 For patients who are identified as having PIP, discussing advantages and disadvantages of any medication change with the patient themselves is important. A recent trial in general practice to reduce PIP found that more changes were made when patients were present for medication review.[193] Any discussion should put particular emphasis on the patient s own priorities as they may place different weights on various benefits and risks. This is especially important when considering starting a new medication to address a PPO, as it may be preferable to both prescriber and patient to not start a preventive treatment in advancing older age despite it being indicated.[237] Rigidly applying treatment guidelines can be ineffective as they do not take account of neither comorbidities nor patients preferences and evidence often comes from trials which did not include older patients.[20,238] This is in contrast to the process of addressing PIMs which may require consideration of stopping a medicine.[239] Both types of PIP present distinct challenges and different approaches may be needed to address potential errors of commission and omission.[172] Adjustment for time-dependent confounding using MSMs did not alter the results here, possibly because factors such as number of medicines were relatively time-stable over the study period. This may relate to therapeutic inertia, failure to start new drugs,[240] and conversely due to prescriber and patient reluctance to deprescribe treatments for fear of negative consequences.[182] Although MSMs may provide better evidence for causal relationships than conventional regression analyses, associations from longitudinal studies should also be interpreted in the context of other criteria for causation such as the Bradford Hill criteria.[101] Strengths and limitations This study s participants were community-dwelling older people from a nationallyrepresentative study on ageing, which improves the generalisability of these findings. Although only participants with eligibility to the means-tested GMS scheme were included, a high proportion (73.5%) of TILDA participants aged over 65 years reported GMS eligibility (see Figure 4-1). This is one of the first longitudinal studies in the community setting to determine the prospective relationship between PIP and adverse outcomes. This robust design allowed for baseline differences to be accounted for and also addressed a number of criteria for inference of causality in epidemiological studies, including temporality and biological gradient.[101] 120

123 The use of administrative pharmacy claims data in this study may provide more accurate information on medication exposure than self-reported medication use, although good agreement has been found between such sources.[210] It also allows medication exposure to be determined over a 12 month time period to provide a more accurate assessment of PIM and PPO prevalence, as opposed to using medication data from one point in time which could underestimate PIMs and overestimate PPOs. A limitation of pharmacy claims data is patients may not have actually consumed medications dispensed (i.e. if patients are non-adherent) and a lack of information on medicines purchased over-the-counter may lead to an overestimation of some prescribing omissions (e.g. calcium and vitamin D, laxatives). Additionally there may have been a clinically justified reason why some participants had a PIM/PPO, however as no clinical notes were available it is not possible to determine clinicians rationale for such prescribing decisions. The combination of data from a pharmacy claims database and a longitudinal study on ageing to provide medications and clinical information allowed for a greater number of criteria from each screening tool to be applied than with either source alone.[204] However a number of criteria from each of the PIM and PPO screening tools could not be included in this analysis due to the required information not being available. For those that were applied, this study did not have sufficient power to determine if any individual criteria were more strongly associated with adverse patient outcomes. Some information on diagnoses was based on participants self-report and so may not accurately reflect the presence/absence of the conditions of interest. The outcomes of healthcare utilisation and functional decline were patient reported rather than objective measures which may have affected accuracy, however the range of covariates adjusted for should have addressed any systematic reporting bias amongst participant subgroups.[217,219] The CASP-R12 is a measure of quality of life rather than healthrelated quality of life specifically and so may not have been sensitive to changes in the health status of participants. Although there is no minimally important or clinically significant difference defined for the CASP-R12, previous analysis of the CASP-19 measure found differences of between two and eight units associated with important life circumstances such as living alone or having difficulty walking a quarter of a mile.[241] This suggests the one unit reduction in QoL associated with having two or more START 121

124 criteria is relatively modest. While this cohort was well characterised, there is still potential for the presence of unmeasured or unknown confounders and hence relationships observed could be subject to residual confounding. Although reverse causality could explain the relationships of START with adverse outcomes, i.e. preventive treatments being omitted in frailer patients with limited life expectancy who then experience functional decline and reduced QoL, controlling for baseline outcome differences and time-varying covariates should reduce this possibility.[208] Conclusions Although prevalence of PIMs and PPOs can vary depending on the screening tool used, such prescribing issues are common and become more prevalent in patients with more medicines or chronic illnesses. This underlines the importance of ongoing prescribing review for older patients, both to assess the appropriateness of current drug therapy as well as to evaluate the need for additional clinically-indicated treatments. PIP determined by STOPP and START was associated with adverse outcomes in this prospective older community-dwelling cohort. If application of these criteria can be integrated into routine medication review, they may help to support prescribers in optimising treatment and improve patient outcomes. Although such prescribing is only potentially inappropriate, the independent effects identified add weight to the suggestion that PIP is a marker of healthcare quality and patient safety and should be minimised if possible. As well as the impact of PIP on patient outcomes, the effects on the wider health systems and healthcare costs should also be considered. 122

125 Chapter 5 Evaluating the economic impact of PIP and related adverse events in older people in Ireland using a Markov modelling approach 123

126 5.1 Introduction As has been discussed and illustrated previously in Chapter 2 and Chapter 4, PIP determined by explicit criteria such as STOPP is prevalent among older people. STOPP contains 65 criteria which were included based on existing evidence and expert consensus. These comprise a heterogeneous mix of indicators, some of which have a high propensity to cause harm and some less so, with a diverse range of potential adverse effects.[42] There is also evidence that being exposed to these PIP criteria is associated with adverse outcomes in older people, increasing the risk of adverse drug events and rate of healthcare use,[106,242] the latter of which was illustrated in Chapter 4. While the effect of PIP defined by the STOPP criteria on patients has been evaluated, less is known about the economic impact of such prescribing and its implications for the health service.[115] In two systematic reviews, one of studies assessing the effect of STOPP and another on the economic impact of inappropriate drug prescribing more generally, only direct medication costs of PIP drugs were assessed.[115,118] The annual expenditure on PIP defined by STOPP amongst those over 70 years of age in Ireland was estimated at 45 million in 2007 and 6 million in Northern Ireland, 9% and 5.4% of all drug costs in this age group respectively.[92,116] However, switching patients from a PIP medication to a more appropriate alternative may actually result in increased drug costs.[118,122] For instance, the costs of inappropriate medications in those aged 65 years to the German health system in 2009 was million, but substituting these inappropriate medications with a recommended drug would incur additional costs of 20 million using the least expensive alternative or up to 504 million if the most expensive options were substituted.[122] Pharmaceutical expenditure is only one of the economic consequences of PIP, and the costs of managing any resulting adverse events have yet to be quantified. Iatrogenic costs, that is those relating to harm caused by medical treatment, have only been assessed for individual medication classes to date, such as benzodiazepines and NSAIDs.[126,128] Furthermore, the methods used to conduct these analyses have included regression to determine extra attributable costs associated with use of these drugs in observational studies, possibly subject to confounding by indication.[46] Costs of other medication-related problems such as medication errors have been assessed by categorising according to the error site (prescribing, dispensing, administration) and type (wrong drug, dose, route, or frequency).[130] While this does 124

127 provide an estimate for the case of non-specific medication errors, using such an approach for distinct drug classes involved in PIP may be an oversimplification reducing face validity. Benzodiazepines and NSAIDs are two of the more frequently identified drug classes in STOPP and there is clear evidence of potential harm.[243,244] For others such as PPIs, although a number of possible risks have been reported, evidence may only support an effect on fractures and Clostridium difficile infection.[147] Moreover, STOPP only defines long-term use of a PPI at maximal dose rather than maintenance dose as potentially inappropriate, and a dose-response effect on these events has not been clearly demonstrated.[245,246] An empirical assessment of which criteria are most significant in terms of patient safety or economic impact has not been conducted, and would be challenging to perform as a large-scale observational study would be required in order to detect differences in patient outcomes between a large number of criteria. Also, the range of potential adverse effects associated with the different criteria, with variable levels of severity, make direct comparisons difficult. Information on the differential impact of types of PIP would be important not only at the health system level to address the most important criteria with prescriber-level interventions but also at the individual level if a doctor is considering which issues to prioritise in a patient with several PIPs. Although there appears to be no primary studies that have addressed these questions, an economic modelling framework may be used to synthesise existing evidence and information sources to examine these issues. Economic modelling can be used to extrapolate beyond the results of a primary study to address a research question by combining the best available sources of evidence.[247] A number of methods can be used, for example decision tree models, state-transition (or Markov) models and discrete event simulation.[146] Quantifying the overall cost and quality of life impact of PIP criteria does not necessarily justify the introduction of interventions to address these issues. As was found in Chapter 4, individuals with criteria from STOPP and START had higher healthcare use but this increase was modest. Therefore, the question arises as to whether targeting these medication problems is an efficient use of health professionals limited time and healthcare resources in general. Even effective interventions cannot be introduced at any cost as there are finite resources available to fund healthcare.[248] There is an 125

128 opportunity cost to any resources devoted to medicines optimisation interventions as they could be used to fund other health services where potentially a greater benefit could be produced. For this reason, new interventions that are developed must usually demonstrate cost-effectiveness at a threshold deemed acceptable by decision makers to ensure health gains are achieved at an appropriate cost. It is therefore important to determine whether interventions to reduce PIP are a cost-effective use of finite healthcare resources. The aim of this study was to estimate and compare the economic impact of three of the most common forms of potentially inappropriate prescribing identified by STOPP in older people in Ireland. This aim was achieved through a number of objectives: to develop and populate with data from the literature a Markov model for each of the three most prevalent PIP criteria which includes the principle adverse consequences of interest, to use these models to compare each PIP to a more appropriate non-pip scenario in terms of the effects on patient outcomes, namely quality and quantity of life, and costs to the health service of the primary adverse events, and to apply these models to evaluate the cost-effectiveness of hypothetical interventions to reduce PIP. 5.2 Methods Markov models A Markov modelling approach was used, which involves a number of steps: Development of a model which maps out health states which a patient may enter as a result of the disease, intervention or medication of interest. Assignment of transition probabilities, the likelihood of moving from one health state to another, and the effect of the intervention of interest on these probabilities. Assigning cost and effect payoffs for each state. The model is then run over a certain timeframe (or horizon) of discrete time periods or cycles where a cohort of hypothetical patients can transition between states once per cycle according to the probabilities specified and accrue costs and effects based on how many cycles they spend in each state. An example of a typical Markov model is shown in 126

129 Figure 5-1 where patients can move from being well to progressive disease where quality of life may be poorer and healthcare costs higher and there is a probability of dying in both of these states. Models are run for each strategy under consideration to estimate the expected cost and effects and to calculate the difference in these outcomes between alternative strategies. For example, a new drug which reduces the probability of developing progressive disease could be evaluated in the model below compared to standard care. Markov models have most often been employed in the economic evaluation of interventions, however, for the purposes of this study the models being developed aim to capture the adverse consequences of a patient having a PIP compared to an alternative appropriate prescription. Figure 5-1 Illustrative example of a Markov model structure The three PIP criteria considered here were the most prevalent issues identified in Chapter 2 among those aged 65 years and over in the Eastern Health Board region of Ireland by a subset of STOPP criteria applied to HSE-PCRS administrative pharmacy claims data from 2012 (see Table 5-1).[176] A model was developed for each of these to assess the cost and effect differences between each PIP and a non-pip alternative. An annual cycle length was used for the models and, as states are required to be mutually exclusive (an individual can only be in one state per cycle) and collectively exhaustive (all individuals in the cohort must fall into one of the states), only one event could occur per cycle.[248] Table 5-1 Description of included PIP criteria from STOPP PIP Non-PIP alternative Prevalence Adverse events represented NSAID >3 months Paracetamol 4.1% Dyspepsia Gastrointestinal bleed Myocardial infarction Benzodiazepine >4 No medication 4.25% Hip fracture weeks PPI maximal dose >8 weeks Maintenance dose PPI Other fall injuries 23.57% Hip fracture Clostridium difficile infection 127

130 The overall model structure for each PIP consisted of a decision tree between the two alternative scenarios. A Markov model attached to these branches represents the adverse events that patients may have as a result of each instance of PIP, to capture the medication costs and costs and effects of the consequences of inappropriate compared to appropriate prescribing. Health states were included to reflect the most likely adverse effects of each PIP and were only included if there would be a difference in the transition probabilities, cost, or effect between the PIP and non-pip scenarios. The models were populated with parameter estimates (see Table 5-2) derived from published sources which are described in Appendix IV. A broadly inclusive search of the literature was conducted to identify studies which included individuals 65 years and over and the PIP exposures and adverse events of interest. Studies which were used to provide parameter inputs for previous Markov models in these therapeutic areas were also assessed for their suitability. In cases were multiple potential sources of parameter inputs were available, these were appraised for study quality to select the most robust estimate. Model structures were assessed for face validity by the research team (who have expertise in pharmacy, general practice, clinical pharmacology, pharmacoepidemiology, and statistics) and cross validity was assessed by comparison to other published models concerning these therapeutic areas.[249] The research team also assessed the face validity of the parameter inputs and model outcomes used. A description of the model structure for each PIP in turn is presented in Section Following this, the general approaches used for deriving model inputs for transition probabilities, costs, and utilities are outlined with further detail provided in Appendix IV. Point estimates were extracted from the published literature and distributions for each parameter were derived from data available from these sources. Details of how these were derived and distributional assumptions are provided in Section and Appendix V. The analysis that was undertaken is then described in Section Description of model structures The states included in each model capture the possible consequences for a patient with a PIP and the typical resource use and increased risks following an event are described. The same model structures were used for both the PIP and non-pip scenarios with the only differences being transition probabilities and cost of the PIP or non-pip treatment. 128

131 NSAID model All patients start in the Well (no previous event) state and remain here until they have a GI event (dyspepsia or GI bleed), an MI, or die (see Figure 5-2). Patients are on diclofenac 75mg twice daily in the PIP arm or paracetamol 1,000mg four times daily in the non-pip arm. In the non-pip arm, the transition probabilities reflect the rates of the adverse events in the general NSAID non-user population, and in the PIP arm, the relative risk in NSAID users was applied to these probabilities. Figure 5-2 State transitions of NSAID Markov model Patients can transition to the Dyspepsia state where individuals have persistent dyspepsia causing GI discomfort requiring consultation with a doctor and so they attend their GP for an extra visit, are switched from diclofenac to paracetamol and receive a prescription for a proton pump inhibitor (lansoprazole 15mg once daily for four weeks). They return to the baseline (non-pip) risk of further dyspepsia and if no further event occurs in the following cycle, they transition to the Well, GI event history state. Patients who transition to the GI bleed state attend the emergency department (ED), are admitted to hospital for investigation and management of upper GI bleeding, are switched from diclofenac to paracetamol and receive a prescription for lansoprazole 15mg once daily for four weeks. After discharge, they are expected to have additional healthcare use as a result of their GI bleed, namely two GP visits and two outpatient department (OPD) visits.[250,251] As with dyspepsia, they return to baseline risk of a further GI bleed and transition to the Well, GI event history state if they have no further 129

132 event in the following cycle. In the Well, GI event history state, patients therapy has been switched from diclofenac to paracetamol, so the cost of medication (paracetamol) and transition probabilities for further GI events or an MI from this state is equal in both the PIP and non-pip arms. Patients transition to the MI state following an MI and remain here for one cycle unless they have a further MI in the following cycle. Patients who have an MI incur inpatient treatment costs, are switched from diclofenac to paracetamol and commence medications for secondary cardiovascular prevention. They also have an additional 11 OPD visits and attend their GP an extra 8 times in the year of an MI.[252] During this year patients are also at increased risk of a further MI.[253] If no event occurs in the subsequent cycle then patients transition to the Well, previous MI state, where the probability of a subsequent MI falls, although it remains higher than in patients with no previous MI.[253] Patients in any previous MI state incur the costs of attending two extra OPD appointments and two GP appointments per year,[252] as well as the cost of secondary preventive medicines and paracetamol Benzodiazepine model All patients start in the Well, no fall injury, community state as the cohort is communitydwelling and are assumed to have had no fall injury in the previous 12 months (see Figure 5-3). The only cost incurred by patients in this state is the cost of the PIP medication, diazepam 5mg twice daily in the PIP arm, whereas no pharmacotherapy is prescribed in the non-pip arm. Patients in the PIP arm remain on this medication with its associated cost and increased adverse events risk throughout the model i.e. no therapy switch occurs after an adverse event. From this state, a transition can occur following a hip fracture or some other fall injury that a patient seeks healthcare for. Hip fractures were divided into (i) those where the patient returns home and (ii) those which result in the patient being permanently admitted to a nursing home setting. Other events that can occur independently of falls are death and admission to a nursing home. On having a hip fracture, patients transition to one of the two hip fracture states, depending on where they are discharged to following this event and remain here for one cycle, unless they suffer a further hip fracture. All hip fracture patients present at an ED, are admitted as inpatients and are discharged either back to the community or to a residential care setting. After discharge, hip fracture patients attend an average of 9 130

133 additional OPD appointments and have an excess of 10 visits to their GP.[254] For those discharged to the residential setting, there is the additional cost of nursing home residence. For 12 months following a hip fracture patients are at an increased risk of a further fall due to their recent injurious fall.[255] If they have no hip fracture or other fall injury in the following cycle, they transition back to the Well, no fall injury state (either community or residential) and return to baseline fall risk. Figure 5-3 State transitions of benzodiazepine Markov model All patients with a fall injury requiring healthcare that is not a hip fracture (such as bruising, soft tissue injuries or other types of fractures) transition to the Other fall injury state. The costs incurred in this state are based on a weighted average of the prevalence of different injury types and typical healthcare use taken from an Irish costing study.[256] Half of patients with other falls injuries have one additional visit to their GP, 22% attend an ED, are not admitted and are referred to their GP for a follow-up visit. Twenty percent attend ED with a non-hip fracture, are admitted as inpatients, and are discharged to community where they have 9 additional OPD visits and 6 extra GP visits.[254] The remaining 8% attend ED with other fall injuries, are admitted as inpatients and following discharge, are referred for one OPD visit and one GP visit for follow-up.[257] The only difference between community and nursing home setting is the additional cost of nursing home residence. As with the hip fracture states, patients remain in this state for one cycle unless they suffer another fall injury and are at an increased risk of a further fall while in this state. 131

134 Patients from all of the community-based states transition to the Well, no fall injury, residential state based on the annual probability of being admitted to a nursing home. This background probability of nursing home admission is included as otherwise the number of admissions attributed to hip fracture in benzodiazepine users would be overestimated. Patients also transition to this state in the cycle following a hip fracture which results in permanent nursing home admission, or if they are nursing home residents who suffer a hip fracture or other fall injury. As only permanent admissions are represented in this model, no transitions occur from residential states back to community states PPI model The model structure (Figure 5-4) is similar to the benzodiazepine model. All individuals start in the Well, no event, community where the only resource use is cost of the PIP or non-pip medication (i.e. maximal dose PPI or maintenance dose PPI). Patients in each arm remain on these medications, with their associated costs and increased adverse events risk, throughout the model i.e. no therapy switch occurs after an adverse event. A number of events can then occur, those that are affected by PIP exposure (Clostridium difficile infection and hip fracture) and those that are unaffected (death and admission to a nursing home). Similarly, following a transition to a residential state, patients remain there and no transition back to community can occur. Figure 5-4 State transitions of PPI Markov model 132

135 Following a hip fracture, patients transition to one of the Hip fracture states (again depending on the setting they are discharged to) and remain in this event state for one cycle, unless they suffer a further hip fracture. Regarding healthcare utilisation, the same pattern that applied to this state in the benzodiazepine model was used here, including the additional cost of nursing home care for residential states. Patients who develop C. difficile infection transition to the C difficile infection state for one cycle where the healthcare resource use is the cost of inpatient management attributable to the infection, as community-dwelling patients aged 65 years or over are likely to be admitted as a result of an infection.[258] No further healthcare costs are incurred, and there is no increased risk of recurrence following a case (as recurrent cases were included in the baseline probability used) or being in a residential setting Transition probabilities Probabilities of transitions between states for the three models were taken from published literature sources which reported rates or probabilities of the adverse events of interest. Further details of these are provided in Appendix IV. Population-based epidemiological studies with study samples representative of older community-dwelling adults were used, whenever possible. This reflected the baseline rate of adverse events for individuals in the non-pip models. In the PIP models, a relative risk was applied to the baseline probability for the effect of the PIP medication on each of the adverse events. These were taken from experimental studies of NSAIDs, benzodiazepines and PPIs, ideally meta-analyses of randomised controlled trials, or in the absence of interventional studies, meta-analyses of observational studies or single observational studies were used as the source of estimates. Annual probability of death from all causes was based on agespecific population rates for 2014 from the Central Statistics Office (CSO) and this was included in all models. Excess mortality following adverse events was taken from observational studies and was independent of PIP exposure (i.e. the same post-event mortality was applied in both PIP and non-pip scenarios) Costing Each state was assigned a cost reflecting the annual costs to the health system per patient in that health state. Costs from 2014 were used, and where not available historical costs were inflated using the applicable Consumer Price Index (CPI) Health sub index from the CSO. Where data on costs in Ireland were unavailable, international 133

136 estimates were used (see Section ). In line with the Irish Health Information and Quality Authority (HIQA) recommendations for health technology assessment,[259] historical costs from other countries were inflated to 2014 costs using the CPI from the origin country, and were then converted to Irish costs using the Purchasing Power Parity index (PPPI). Discounting was used to devalue all costs and effects occurring after the first year to represent time preferences.[248] The rate recommended by HIQA is 5% with a range of 0% to 6% proposed for sensitivity analysis and these recommendations were implemented in this analysis.[259] There are two predominant approaches for costing in economic evaluations. These are micro-costing, which uses detailed information on treatment inputs for each patient type and a unit cost per input is used to determine an overall cost, and macro-costing, which uses aggregated costs for example associated with diagnosis-related groups (DRGs).[248] A micro-costing approach was used for each state, while some hospital inpatient cost inputs were based on macro-costing using DRGs. Further details are provided in Appendix IV Utility values The primary outcome of effect was quality-adjusted life years (QALYs), which is a composite measure of both quality and quantity of life.[248] Each period of time, for example each year, is assigned a weight relating to the health-related quality of life of that health state, typically on a scale of 1 being full health to 0 being equivalent to death. These weights are often referred to as utilities or utility values and are derived by obtaining preferences for the health state relative to the best and worst possible health states from members of the general population. These quality-weighted time periods are summed to yield QALYs. The same value was applied to the Well or no event health states in all three models considered. In studies of individual disease areas, different baseline utilities can be used to reflect the quality of life of patients with the pathology of interest. As the medications in these models are not indicated for one discrete disease and to increase comparability between the models, the same baseline utility was used. Age-specific utility values were UK population norms for the EQ-5D for people aged and 75 years and over which were derived using a visual analogue scale.[260] Utility decrements or disutilities, the reduction in utility in the year following an adverse event, were subtracted from the Well utility for older people to give the utility value for each state. The secondary outcome of effect was life years (LYs) which is the quantity of life with no weighting for quality. 134

137 5.2.6 Analysis Economic impact of PIP relative to non-pip The first stage of analysis was to estimate the cost and effect differences between the PIP and non-pip scenarios. Analyses were conducted separately for each PIP in TreeAge Pro 2015 (TreeAge Software, Inc., Williamstown, MA). This analysis was of a closed cohort of patients aged 65 years using an annual cycle length over a horizon of 35 cycles (i.e. up to age 100) and parameter inputs for each model are shown in Table 5-2. An example of the model structure in TreeAge Pro for the NSAID PIP is shown in Figure 5-5, and for the other PIP models in Appendix VI. A half-cycle correction was included in the first and last cycles to prevent overestimation of survival as the models were not run over a lifetime horizon.[261] Models were validated by double-programming (i.e. replicating each model using an alternate software program) in Microsoft Excel 2010 (Microsoft Corp., Redmond, WA) to ensure no structural or coding errors were included, and extreme value testing and comparison of cohort traces between TreeAge Pro and Excel were also conducted.[249] Results are presented as point estimates of mean differences in cost and in effect (both QALYs and LYs) between the PIP and non-pip scenarios. An incremental cost-effectiveness ratio (ICER) was also calculated for each PIP by dividing the mean difference in costs by the mean difference in QALYs. The ICERs indicate the expected additional cost per additional QALY in the PIP scenario relative to the non-pip scenario. These can be plotted on the cost-effectiveness (CE) plane as shown in Figure 5-6, and the quadrant which an ICER falls into is informative as to which strategy is preferred. In the southeast quadrant, the strategy of interest dominates the comparator as it is less costly and provides more benefit, while conversely in the northwest quadrant the strategy of interest is dominated. For the other two quadrants, the preferred strategy depends on the numerical value of the ICER and the maximum acceptable ICER or willingness to pay (WTP) for additional effect. In the northeast quadrant, the strategy of interest is preferred if the ICER is below the WTP threshold, i.e. the cost per additional QALY gained is less than the maximum cost decision makers are willing to pay. In the southwest quadrant, the strategy of interest is preferred if the ICER is above the WTP threshold, i.e. the cost-saving per additional QALY foregone is greater than the acceptable threshold. 135

138 Figure 5-5 Decision tree structure for NSAID Markov model in TreeAge Pro 136

139 Figure 5-6 Cost-effectiveness plane As only adverse effects were assessed in these models, it was expected that the PIP scenario would be dominated in each case by the non-pip scenario (i.e. be more costly and generate fewer QALYs). Differences in the total number of adverse events for the PIP scenario compared to the non-pip scenario were also determined. These analyses were conducted using the point estimate for each parameter input One-way sensitivity analysis To assess the impact of varying specific parameter inputs, one-way deterministic sensitivity analyses were conducted. The inputs investigated in this sensitivity analysis were the inpatient cost of C. difficile (as unlike other costs, an estimate of this was not available from an Irish source), the cost of PIP and non-pip medications, and discount rates (which were varied between 0% and 6% for costs and outcomes as recommended by the HIQA guidelines).[259] Probabilistic sensitivity analysis Using point estimates for parameter inputs does not provide information on the precision of or confidence in the results. Uncertainty associated with imprecision of the parameter inputs (parameter uncertainty) was incorporated into the model using probabilistic sensitivity analysis (PSA) to allow confidence intervals (CIs) to be fitted for the incremental cost and effect estimates. 137

140 This involved specifying a distribution of possible values for each parameter to be varied. The distributions were fitted under the assumption of a homogenous sample of patients informing parameter estimates (i.e. heterogeneity between patient sub-groups was not investigated). The distribution type used for each parameter reflected the form of data the parameter takes and reflected the standard distributional assumptions used when estimating CIs i.e. beta distributions for probabilities, log-normal distributions for relative risks, gamma distributions for costs, beta distributions for utility values and gamma distributions for disutilities.[262] The distributions fitted for each parameter were calculated from data available in published sources and these are reported in Table 5-2. General methods of how these distributions were specified from data presented in the literature are detailed in Appendix V. Then, 10,000 iterations of each model were run and a random value for each parameter input was sampled from the specified distribution for each run. The outputs of each iteration were recorded to provide a distribution of cost and effect differences and the 2.5 th and 97.5 th percentiles for these differences were used to estimate 95% CIs. Statistical significance was assumed if the 95% CI for the incremental costs and effects did not include zero. The outputs for each iteration were also plotted on a CE plane to compare the distribution of ICER estimates for each PIP Structural sensitivity analysis Sensitivity analysis was also conducted on an assumption of the NSAID model. This model differed from the others in that it was assumed that in the PIP scenario, following an initial case of dyspepsia, a GI bleed, or an MI, patients would be switched from an NSAID to paracetamol and would then be subject to non-pip probabilities of adverse events for the remainder of the time horizon. Structural uncertainty was assessed by applying the relative risks of adverse events in NSAID users to transition probabilities through all states in the PIP arm rather than applying non-pip probabilities after an adverse event Cost-effectiveness of hypothetical interventions Having estimated the incremental costs and effects of each PIP, in the second stage of the analysis each model was used to evaluate the cost-effectiveness of a hypothetical intervention to reduce the prevalence of that PIP criterion. A new decision was framed between having an intervention to reduce PIP or no intervention (i.e. usual care), as shown for NSAIDs in Figure 5-7. The intervention was delivered once at the beginning of 138

141 the model (with a one-off cost) to all individuals on a long-term NSAID and resulted in a proportion of these people being switched to paracetamol for the duration of the model time horizon. The prevalence of long-term NSAID use determined in the 2012 EHB GMS population in Chapter 2 was used for the proportion starting in this arm at the beginning of the model run.[176] The cost of this intervention per person and the effectiveness (i.e. the relative reduction in the proportion on a long-term NSAID) were varied over a range to determine circumstances in which the intervention would be preferred to no intervention (i.e. usual care) at a WTP threshold of 45,000 per QALY, which is the conventionally used threshold in Ireland.[259] These results were plotted and this was then repeated for the other PIPs relating to benzodiazepines and PPIs. For a fixed intervention cost of 500 per person (selected as an arbitrary figure for illustrative purposes), the minimal effectiveness at which the intervention would be cost effective was determined for each PIP. Next, further threshold analysis was conducted using published estimates of the effectiveness of trials targeting these PIPs to determine maximal costs at which such medicines optimisation interventions would be cost effective. In the OPTI-SCRIPT trial of a complex intervention in general practice, the relative risk of being on a long-term maximal dose PPI post-intervention was 0.45 (i.e. a 55% reduction) compared to usual care.[193] For NSAIDs, a recent trial of education, informatics and incentives in general practice demonstrated a significant reduction of 49.8% in high-risk prescribing relating to NSAIDs and gastroprotection (i.e. a risk reduction of 0.498).[83] A trial to reduce inappropriate prescribing of benzodiazepines using direct patient education demonstrated an additional 23% of those in the intervention group had discontinued benzodiazepines compared to control (i.e. a risk reduction of 0.23).[263] This was evaluated at WTP thresholds of 45,000, 20,000, and 0 per QALY. Figure 5-7 Decision tree structure of hypothetical intervention analysis 139

142 Table 5-2 Point estimates for each parameter input and distributions used in probabilistic sensitivity analysis Parameter description Value Distribution Source Transition probabilities NSAID model Probability of dyspepsia in non-nsaid users Beta (4,058, 75,513) [264,265] Probability of GI bleed in non-nsaid users Beta (99.71, 76,601.91) [266,267] Probability of death following GI bleed by age group Beta (156, 1,265) (174, 698) [268] Probability of an MI in non-nsaid users Beta (419, 50775) [269] Probability of an MI in the 12 months following an MI Beta ( , ) [253] Probability of an MI in subsequent years after an MI Beta ( , ) [253] Probability of death following an MI Beta (209, 1942) [270] Probability of death by age group Effect Relative risk of dyspepsia in long-term NSAID users 1.4 Log-normal (0.336, 0.126) [264,265] Relative risk of GI bleed in long-term NSAID users 3.07 Log-normal (1.122, 0.114) [267] Relative risk of MI in long-term NSAID users 1.53 Log-normal (0.425, 0.174) [267] Relative risk of death in people >1 year post-mi 2 Log-normal (0.693, 0.088) [253] Utility Utility of being in well state Beta (129.13, 38.57) (108.51, 38.13) Annual disutility of dyspepsia state Gamma (129.13, 38.57) [133,271, 272] Annual disutility of GI bleed state Gamma (108.51, 38.13) [133,271, 272] Annual disutility of MI state Gamma (74.37, ) [273,274] Annual disutility of a previous MI Gamma (4, ) [ ] Costs Cost of NSAID treatment Gamma (100, 0.668) [277] Cost of paracetamol treatment Gamma (100, 1.024) [277] Cost of managing dyspepsia Gamma (100, 0.655) [277] Cost of managing a GI bleed 4, Gamma (44.44, 0.009) [ ] Cost of managing an MI 9, Gamma (100, 0.010) [252,278, 280] Cost of a previous MI Gamma (100, 0.122) [252,277] Transition probabilities Probability of an injurious fall requiring healthcare utilisation Benzodiazepine model Probability of a hip fracture Beta (95, 1,905) (200, 1,800) Beta (197, 140,517) [156] [260] [ ] [266,267]

143 Parameter description Value Distribution Source Probability of being in nursing home at 12m following a hip fracture Probability of being admitted to nursing home in general population Effect Relative risk of an injurious fall in long-term benzodiazepine users Relative risk of injurious fall in 12 months post-fall injury Relative hazard of death in 12 months following a hip fracture relative to people without fracture Utility (357, 114,804) (597, 89,858) (961, 62,263) (1,071, 42,289) 0.11 Beta (224, 1,810) [268] Beta (301, 143,095) (393, 118,759) (601, 91,865) (980, 63,904) (1,093, 44,254) [285] Log-normal (0.440, 0.043) [286] 2.0 Log-normal (0.693, 0.039) [255] 3.26 Log-normal (1.182, 0.062) [287] Utility decrement in 12m following a hip fracture Gamma (209.33, 1,031.2) [288,289] Utility decrement in 12m following other fall injury 0.06 Gamma (22.13, ) [288,289] Utility decrement of being resident in nursing home 0.06 Gamma (0.58, 9.72) [ ] Costs Cost of benzodiazepine treatment Gamma (100, 1.283) [277] Cost of hip fracture 17, Gamma (385.34, 0.022) [254,277, 293] Cost of other fall injury 2, Gamma (25, 0.009) [254,256, 257,277] Cost of residence in nursing home 42, Gamma (9,407.98, 0.220) [294] Transition probabilities PPI model Probability of having C. difficile infection Beta (1839, 511,848) [258] Effect Relative risk of hip fracture in maximal dose PPI users relative to non-users and maintenance dose PPI users relative to non-users Relative risk of C. difficile infection in maximal dose PPI users relative to non-users and in maintenance dose PPI users relative to nonusers Log-normal (0.380, 0.097) Log-normal (0.221, 0.050) Log-normal (0.854, 0.140) Log-normal (0.551, 0.114) [245] [246] Relative hazard for death in 12m post C. difficile 1.23 Log-normal (0.207, 0.089) [295] Utility Utility decrement in 12m post C. difficile Gamma (0.530, 20.38) [ ] Costs Cost of max dose PPI treatment Gamma (25, 0.155) [277] Cost of maintenance dose PPI Gamma (25, 0.213) [277] Cost of C. difficile 5, Gamma (19.3, 0.003) [258,299, 300] 141

144 5.3 Results Economic impact of PIP relative to non-pip For all three models, the PIP scenarios were dominated by the non-pip scenario (i.e. they generated higher costs and lower QALYs) over a time horizon up to the age of 100 years. The largest difference in costs and QALYs was in the benzodiazepine model, where being on a PIP generated an average of 3,470 more in costs and 0.07 fewer QALYs per patient compared to the non-pip scenario (see Table 5-3). For costs, this was followed by the PPI model where patients on a long-term maximal dose accrued additional costs of 989 and had a reduction of 0.01 QALYs relative to those on a maintenance dose. In the case of NSAIDs, the PIP model resulted in 806 more in costs and 0.07 fewer QALYs per patient compared to the non-pip model. The ICER for each model was negative and all fell within the northeast quadrant of the cost-effectiveness plane, which indicates the PIP scenario was more costly and resulted in a worse outcome than the appropriate alternative. The excess adverse events occurring in each PIP model relative to the non-pip model are reported in Appendix VII. Table 5-3 Cost, effect, and ICER outputs for PIP compared to non-pip scenarios for each model Model Incremental Cost ( ) Incremental QALYs ICER ( per QALY) Incremental LYs NSAID model , Benzodiazepine model 3, , PPI model , Abbreviations: ICER, incremental cost-effectiveness ratio; LYs, life years; NSAID, non-steroidal anti-inflammatory drug; PPI, proton pump inhibitor; QALYs, quality-adjusted life years One-way sensitivity analysis In the one-way deterministic sensitivity analysis, the PIP arm of each model was still dominated by the non-pip across the range of values for all variables investigated and the rankings of the different models in terms of incremental costs and incremental effects did not change. Full details of this are presented in Table A-2 in Appendix VII. With respect to the outcomes discount rate, at 6% (the high point of the range used), the incremental effect in the PPI model was closest to zero and for costs, at a discount rate of 6% the lowest incremental cost was in the NSAID model. Over the plausible range for cost of managing C. difficile infection there was little effect on incremental cost in the PPI model. 142

145 Probabilistic sensitivity analysis Introducing parameter uncertainty into the models in the probabilistic sensitivity analysis, the incremental costs and QALY loss for the NSAID model were statistically significant (95% CI 415, 1,346 costs; , QALYs) and this was also the case for the benzodiazepine model (95% CI 2,434, 5,001 costs; , QALYs). In the probabilistic analysis of the PPI model, there was no evidence of a significant difference between the PIP and non-pip scenarios in terms of incremental costs (95% CI - 69, 2,127) or QALYs (95% CI , 0.003). The relative magnitude and distribution of cost and QALY differences for each model is plotted in Figure 5-8. All ICERs for the benzodiazepine model were dominated and fell within the northwest quadrant, i.e. the PIP scenario was more costly and generated fewer QALYs than the non-pip scenario. For the NSAID model, 99.96% of estimates were also in this quadrant, while 0.04% were in the southwest quadrant (less costly and fewer QALYs). Lastly ICERs for the PPI model fell into all quadrants, 91.9% were dominated (northwest), 2.38% were in the southwest, 1.01% were dominant and fell within the southeast (i.e. the PIP was less costly and yielded more QALYs than the non-pip scenario) and 4.71% were in the northeast (more costly and more QALYs). Cost difference between PIP and non-pip ( ) 8000 PPI NSAID BDZ Utility difference between PIP and non-pip (QALYs) 0 Figure 5-8 Incremental costs and utilities for PIP compared to non-pip from probabilistic sensitivity analysis for each model (northwest quadrant)

146 Structural sensitivity analysis Relaxing the assumption of a switch from the PIP to non-pip medication in the NSAID model after an adverse event (i.e. patients remain on the PIP of a long-term NSAID with its increased risks regardless of adverse events suffered) resulted in a larger cost difference ( 1,494, 95% CI 756, 2,493) and effect difference (-0.11 QALYs, 95% CI , ) between PIP and non-pip strategies. The incremental cost was still lower than that for the benzodiazepine model; however, it was greater than that for the PPI model. The relative spread of probabilistic estimates from this model relative to the other models is shown in Appendix VIII (99.97% in the northwest quadrant, 0.03% in the southwest quadrant) Cost-effectiveness of hypothetical interventions These models were applied to hypothetical interventions to reduce the prevalence of each of the PIPs to investigate their cost-effectiveness at a WTP threshold of 45,000 per QALY. The relationship between once-off intervention cost, effectiveness and preferred option (intervention or no intervention) is represented graphically for each model in Figure 5-9. An intervention costing 500 per patient would be the preferred option in the case of NSAIDs if the relative reduction in PIP was at least 0.126, whereas for benzodiazepines and PPIs an intervention would be preferred to no intervention at an effectiveness above and respectively. These can be obtained from the plots in Figure 5-9 by finding the point on the effectiveness axis at which the line of no difference between intervention and no intervention line at a cost of 500, as illustrated in Figure 5-10 for the NSAID model. Considering published effectiveness estimates, an intervention which reduced potentially inappropriate NSAID use by 49.8% would be cost effective up to a cost of 1,970 per person at a WTP of 45,000. An intervention that resulted in 23% discontinuation in benzodiazepine users would be cost effective up to costs per person of 1,480 and for PPIs, in the case of a 55% reduction in PIP the corresponding threshold cost would be 831 per person at a WTP of 45,000. Threshold effectiveness values at a cost of 500 and threshold costs at published levels of effectiveness using different WTP values are presented in Table 5-4. Taking the extreme case of a WTP threshold of 0 per QALY (i.e. willing to pay nothing additional for any QALY gain), cost-effectiveness would be achieved 144

147 for interventions targeting NSAIDs, benzodiazepines, and PPIs up to once-off intervention costs per patient of 401, 798, and 544 respectively. Figure 5-9 Two-way sensitivity analyses of intervention cost and effectiveness at willingness-topay threshold of 45,000 per QALY for a) benzodiazepines, b) PPIs, and c) NSAID models Figure 5-10 Threshold effectiveness value for NSAID intervention at intervention cost of 500 and willingness-to-pay threshold of 45,000 per QALY 145