Tiago Villanueva Editor BMJ. Dear Dr Villanueva,

Transcription

1 Tiago Villanueva Editor BMJ Dear Dr Villanueva, RE: Ms BMJ : Recent trends and future projections for dementia in England and Wales to 2040: Estimates from the IMPACT-Better Ageing Model First revision Revised Title: Temporal trend in dementia incidence since 2002 and projections for prevalence in England and Wales to 2040: a modelling study Many thanks for sending us the report from your discussions and the referees assessments of our manuscript. We are very pleased to have the opportunity to address the expert feedback and to be able to resubmit our manuscript to the BMJ. The process has been positive and enormously helpful in refining the substance and presentation of the manuscript. We look forward to hearing from you when you and colleagues have had the opportunity to review the revised version of the paper. Yours sincerely Sara Ahmadi-Abhari, on behalf of co-authors 1

2 **Report from The BMJ s manuscript committee meeting** These comments are an attempt to summarise the discussions at the manuscript meeting. They are not an exact transcript. Members of the committee were: John Fletcher (chair), Gary Collins (statistician), Daoxin Yin, Georg Roggla, Rubin Minhas, Jose Merino, Amy Price, Tiago Villanueva, Elizabeth Loder Detailed comments from the meeting: POINT 1: Our statistician made the following comments: For a modelling study this seems reasonably well done and reported the Methods in the paper are a lay summary of what they did, with a good, more detailed description of the Methods in the supplementary material. However, these studies largely depend on their assumptions and whilst I think they have this information in there (and in the supplementary material), it could be made much clearer. A clear summary box would be helpful. OUR RESPONSE: Thank you for your suggestion. A summary table of the model assumptions and evidence supporting the assumptions has been added to the manuscript (Table 1) and presented at the end of comments from the Editors committee on Page 26 of this document. We hope this Table helps clarify the strength of the evidence and assumptions underlying our predictions. We realize the Table may be large and would be happy to move it to the Supplementary material. POINT 2: Observed dementia cases of by wave and sex-specific age groups are quite small (Supplementary figure 1, Table 2); I wonder whether some of this should be in the main paper, as this underpins the study. OUR RESPONSE: Thank you. We agree with your point and have added information about numbers of dementia cases to the results section, Page 15: Between 2002 and 2013 dementia was ascertained for a total of 1,448 ELSA participants. 634 participants had dementia at time of recruitment and 814 incident cases were identified over the course of the follow up. 16.5% of cases occurred in participants below age 65. Of the 1,448 cases of dementia, 466 received a doctor diagnosis of dementia, 245 were identified via IQCODE 2

3 questionnaire plus ADL functional impairment, and 1,078 were identified by impairment in cognitive tests and impairment in one or more ADLs. Of the 245 cases based on IQCODE and functional impairment, 171 (70%) reported a doctor diagnosis of dementia. Among those with impairment in cognitive tests and ADL function, 263 (25%) reported a doctor diagnosis of dementia at later stages or were positive on the IQCODE questionnaire. The number of incident dementia cases ascertained at each data collection wave is shown in Supplement-Table 3. Age- and sex-specific prevalence of dementia at each data collection wave of ELSA provides the starting values for the IMPACT-BAM model (Supplement Table 5). To increase power in calculating age- and sex-specific dementia prevalence and incidence rates, the 2- year epochs between successive waves in ELSA were pooled together so that each individual contributed as many observations as corresponded to the number of 2-year epochs in which they participated in the study until being censored. Hence, although age-specific numbers at each wave may be small, analyses to obtain age, sex, and calendar year specific dementia incidence were based on all 814 incident cases, rather than individual age, sex and wave specific numbers. Similarly, analyses involving dementia prevalence include all 1,448 cases. Although the number of cases by age and sex is limited, previous studies of the epidemiology of dementia are affected by a similar limitation. None had more than 371 cases (References 1-3 below). The six waves of ELSA data used here provide for larger number of cases and greater statistical power. The Cognitive Function and Ageing Study (CFAS), which to date serves as the reference for dementia incidence and prevalence in the UK (1), is based on 367 prevalent and 86 incident cases of dementia at CFAS-I, and 275 prevalent and 173 incident cases of dementia at CFAS-II. The Framingham study (2), reporting on trends in dementia incidence over three decades, is based on a total of 371 incident cases of dementia across all waves of data collection. The Rotterdam study (3) reported trends in dementia incidence based on 286 dementia cases observed in the first and 49 cases observed in the second sub-cohort. References: 1. Matthews FE, Stephan BC, Robinson L, Jagger C, Barnes LE, Arthur A et al. A two decade dementia incidence comparison from the Cognitive Function and Ageing Studies I and II. Nat Commun 2016; 7: doi: /ncomms11398.: Satizabal CL, Beiser AS, Chouraki V, Chene G, Dufouil C, Seshadri S. Incidence of Dementia over Three Decades in the Framingham Heart Study. N Engl J Med 2016; 374(6):

4 3. Schrijvers EM, Verhaaren BF, Koudstaal PJ, Hofman A, Ikram MA, Breteler MM. Is dementia incidence declining?: Trends in dementia incidence since 1990 in the Rotterdam Study. Neurology 2012; 78(19): POINT 3: The validation seems to only include those aged , why were those <65 omitted? Too few events? Or no data in CFAS II? OUR RESPONSE: Thank you. Data for persons below age 65 were not available from CFAS-II. POINT 4: One of the reviewers has indicated that Wave 7 (2015 data) are available - and they raise a good point to see how were their model predicts with this new 2015 data. OUR RESPONSE: Thank you. We agree with the reviewer s suggestion that addition of ELSA wave 7 data, released in October 2016, may strengthen our analysis. The IMPACT-BAM model predictions compared to dementia prevalence at wave 7 ELSA are presented in Supplement Figure 11 shown below. Overall, the model predictions fit well with the dementia prevalence at wave 7 ELSA. In assessment of cognitive and functional impairment at wave 7, we were unable to exclude cases of transient cognitive or functional impairment due to absence of data at later (post wave 7) dates. Without detection of the transient cases, we would tend to overestimate prevalence of dementia at ELSA wave 7, especially in the older age-groups. We therefore calculated the numbers of persons who would have been marked as dementia cases based on transient cognitive and functional impairment at previous waves of ELSA and applied the age-specific proportion to the dementia prevalence at wave 7 to correct for the false positives. Wave 7 ELSA data are also incorporated in the revised manuscript to identify transient impairments in cognition and function among wave 6 participants in order to obtain an accurate prevalence of cognitive and functional impairment at wave 6. 4

5 Supplement Figure 11: IMPACT-BAM predicted prevalence of dementia in men and women compared to estimates from wave 7 of the English Longitudinal Study of Ageing ( ) POINT 5: One editor said that you are claiming to have a better prediction model than what has gone before. He was not sure you do. He made the following comments: 1. One reason you say yours is better is that they model competing risks of death from CVD. But they don't model deaths from cancer or COPD to name the two next biggest killers. OUR RESPONSE: Thank you. We apologize that these important points were not clearly explained in the manuscript. Competing risks due to death from non-cardiovascular causes, which includes cancer and COPD, have indeed been modelled as well as competing risks due to cardiovascular disease incidence and mortality. The assumptions related to incorporation of cardiovascular and noncardiovascular mortality rates in the model are presented in the table of assumptions (Table 1, Assumption 15, as well as assumptions 9, 10, 13, and 14) which is shown at the end of the response to Editors comments (Page 26 of this document). We have also elaborated on the competing risks with mortality in the methods and discussion as follows: Methods Page 12-13: 5

6 Probabilities of cardiovascular and non-cardiovascular mortality up to 2040 in 5-year age bands were calculated using the Bayesian Age Period Cohort (BAPC) model with ONS mortality and population estimates from for England and Wales as inputs (Supplement Figure 3). 33 Health state specific transition probabilities to cardiovascular and non-cardiovascular mortality were calculated by estimating the odds ratios of death from each health state using ELSA data (Supplement Table 1) and using the odds ratios to obtain the probability of cardiovascular and non-cardiovascular death from each health state compared to that in the general population. The probability ratios were applied to overall cardiovascular and non-cardiovascular mortality rates shown in Supplement Figure 3 to obtain state specific cardiovascular and non-cardiovascular mortality rates. Mortality rates from each health state were assumed to change in parallel with overall cardiovascular and noncardiovascular mortality rates. Discussion Page 18: Changes in mortality rates are important determinants of the numbers of people living with dementia as they govern both life-expectancy and the pool of individuals susceptible to dementia, as well as survival in those affected by the condition, 5 and thus, are meticulously incorporated in the model. Cardiovascular and non-cardiovascular mortality rates have shown steady and linear downward trends in the past decades and we assumed this trend is likely to continue. IMPACT-BAM shows that the decline in age-standardised dementia prevalence, corresponding to the decline in incidence, is outweighed by population ageing in the near future and numbers living with dementia are likely to increase rapidly between 2015 and In the following decade, however, the number of people with dementia will level out. In Figure 1 below illustrating the IMPACT-BAM model, the green arrows represent transitions to death from cardiovascular causes and the red arrows represent transitions to death from all noncardiovascular causes. The transition probabilities to non-cardiovascular death from each health state are in effect a weighted average of the probabilities of death from all non-cardiovascular causes, including cancer and COPD, for each age and sex stratum. Once a member of the population dies of either cardiovascular or non-cardiovascular causes they are no longer at risk of developing dementia thus, competing risks due to mortality from any cause (including cancer and COPD) are accounted for in our model. 6

7 Figure 1: IMPACT-BAM model. Competing risks due to cardiovascular (Green arrows) and non-cardiovascular causes of death (Red arrows) have been accounted for. To demonstrate how the model accounts for the effect of competing risks due to cancer and COPD mortality numerically, we provide here a detailed explanation of how projections for mortality rates are derived and present results from a hypothetical scenario in which age and sex-specific mortality rates due to cancer and COPD increase over time and double by year We will show that this hypothetical increase in cancer and COPD mortality rates results in an increase in overall noncardiovascular disease mortality which, through the IMPACT-BAM model, affects the projections of future dementia cases. Calendar trends and changes in mortality rates over time are accounted for in the IMPACT-BAM model by projecting observed trends over the past two decades to the future. The UK Office for National Statistics (ONS) registers all deaths and provides mortality rates by age, sex, and cause of death. These data show that age-standardised mortality rates from cardiovascular disease, all causes combined, and from cancer and COPD have shown steady, linear trends over the past two decades 7

8 (Figure 2 below). Since trends over the past two decades are steady and linear, we hypothesized the most likely scenario would be that these trends will continue in the future. To project age-and sex specific rates of death to the future we applied the Bayesian Age Period Cohort (BAPC) model explained in detail previously (Guzman-Castillo M, et al. Future declines of coronary heart disease mortality in England and Wales could counter the burden of population ageing. PLoS One 2014; 9(6):e99482). The observed and projected rates of mortality from cardiovascular and noncardiovascular causes by sex and age-group are shown in Figure 3 below (added as supplement Figure 3 to the manuscript). Figure 2: Age-standardised mortality rates by cause, Data from UK Office for National Statistics. 8

9 Figure 3: Observed ( ) and projected ( ) cardiovascular and non-cardiovascular mortality rates by age and sex. Observed rates are data from the UK office for National Statistics (ONS). Projected rates are obtained from application of the Bayesian Age-Period-Cohort method. The mortality rates are displayed on a logarithmic scale. 9

10 To describe more specifically, Figure 3 above shows overall cardiovascular and non-cardiovascular mortality rates. In the actual IMPACT-BAM model, the rates of mortality entered as input data are health-state specific. For example, death rates for the group of people with cardiovascular disease and dementia (state 6 in Figure 1) are higher than the death rates in people with the same age and sex but free of cardiovascular disease, and cognitive or functional impairment (state 1 in Figure 1). The health-state specific rates of death were obtained by applying the odds ratio of death for each health state, obtained from ELSA (shown in Supplement Table 1), to the projected rates of death as explained in detail in the methods section Page 13. Age-specific rates of death due to cancer and COPD have not changed considerably in recent years whilst deaths from other causes have declined (Figure 2 above). As a result, the proportion of deaths attributable to cancer and COPD has risen over the past two decades. Our model assumes this trend will continue in the future (Figure 4). Figure 4: Proportion of deaths due to cancer and COPD

11 In the table below we present two scenarios: Scenario 1 is based on projection of current trends in age, sex, and cause specific mortality rates to the future and forms the basis of the predictions presented in our paper as it appears to be the most likely scenario. This projection is then compared with Scenario 2 which is the hypothetical and unlikely scenario in which age and sex-specific mortality rates due to cancer and COPD increase over time and double by year 2040; the findings demonstrate the substantial impact of competing risks due to non-cardiovascular causes on modelled predictions (Figures below). Scenario 1: Mortality rates follow the current trends to the future (This is presented in the paper as the most likely scenario) Scenario 2: Mortality rates from cancer and COPD increase from 2014 such that the age-specific cancer and COPD mortality rates double by The observed and projected cancer and COPD mortality rates in each of the two scenarios are demonstrated below. Scenario 1 Scenario 2 11

12 Non-cardiovascular mortality rates are calculated as a weighted average of probabilities of deaths due to non-cardiovascular causes including deaths from cancer and COPD. For the two scenarios, we have computed the overall non-cardiovascular mortality rates and we show these rates by age-group below. For input to IMPACT-BAM these estimates are calculated separately for each health state in the model. 12

13 Non-cardiovascular mortality rates Scenario 1 Non-cardiovascular mortality rates Scenario 2 13

14 Estimates of numbers of cases of dementia based on each scenario: *Dotted lines represent 95% uncertainty intervals. IMPACT-BAM prediction Number of Dementia Cases Scenario 1 IMPACT-BAM prediction Number of Dementia Cases Scenario 2 14

15 In the hypothetical scenario 2 (assuming mortality rates due to cancer and COPD rise from 2014 and double by 2040) there would be considerably fewer people reaching old age to be susceptible to dementia, and the number living with dementia would decline. Since the non-cardiovascular mortality rates incorporated in the model are a weighted average of all non-cardiovascular causes of death, modelling cancer and COPD deaths as separate death-states would not alter the predictions of dementia prevalence. The former and latter approaches are mathematically equivalent. The effect of competing risks of cardiovascular disease and mortality on dementia prevalence would become especially prominent in the next stage of the IMPACT-BAM project where we model the effect of vascular risk reduction on future projections for numbers of people living with dementia. Vascular risk reduction would have a considerable effect on cardiovascular disease, and thus life-expectancy. Therefore, cardiovascular disease should be explicitly modelled alongside dementia to obtain accurate estimates for the effects of vascular risk reduction scenarios. 15

16 The modelling strategy in the present study follows prevailing research recommendations in the UK. The UK Government has now endorsed the Blackfriars Consensus statement by leading UK health charities, Public Health England and NICE, plus WHO. They all advocate prevention strategies which address the shared determinants of non-communicable diseases. Thus by reducing CVD burden, the dementia burden might also be reduced (UK ministers sign Blackfriars Consensus Statement on promoting brain health and reducing risks for dementia in the population as part of Dementia Awareness Week. POINT 5.2: Dementia prevalence and crude incidence are largely dependent on survival to old age so any changes in mortality from other causes will have a knock on effect on dementia frequency and they don't appear to have modelled this. OUR RESPONSE: Thank you. In fact, we have modelled this issue. Survival to old age in IMPACT- BAM is modelled through application of mortality rates. The prevalence of each health state in the model (e.g., dementia) is determined by numbers in that state from the previous year (i.e., previous iteration of the model), plus the number added due to incident cases, minus the number who die from cardiovascular or non-cardiovascular causes (Table 1, assumption 5). Please see our response to POINT 5.1 above where we describe the methods we used to account for changes in mortality rates and survival to old age. We have also accounted for changes in survival with dementia or cardiovascular disease in the model. Survival in a particular health state equals the number of years the individual remains in that health state. Trends in incidence and mortality rates are incorporated in the model, therefore, changes in survival over time are accounted for. This assumption is explained in the table of assumptions (Table 1, assumption 13), presented at the end of the response to Editors comments. We have clarified this by adding the text to the discussion Page 20: Survival with cardiovascular disease and dementia has improved over time Changes in lifeexpectancy and survival with cardiovascular disease, functional impairment, or dementia are accounted for in the model for by applying calendar-time specific mortality rates. This method is based on the assumption that survival with health conditions change in proportion to changes in overall life-expectancy. This assumption, although reasonable and commonly applied in modelling studies, lacks verification, as evidence on the exact age and sex specific survival with dementia is rare, 5 and hard to obtain given uncertainty over date of onset. 16

17 POINT 6: They are predicting a long way out and with the large rise in frequency of diabetes it is by no means certain that CVD will decline further and may even increase again. OUR RESPONSE: Thank you for raising this important point. The net effect of changes in cardiovascular risk factors in the past two decades, including changes in prevalence of obesity/overweight, type 2 diabetes, smoking, and other risk factors, has been a steady decline in incidence of cardiovascular disease. For example, in the Whitehall II cohort between 1985 and end of 2004, the incidence of major coronary events declined by some 74% (6.5% p.a.) and rising BMI was estimated to have reduced the scale of the decline by about 11% (Hardoon SL, et al. Rising adiposity curbing decline in the incidence of myocardial infarction: 20-year follow-up of British men and women in the Whitehall II cohort. Eur Heart J 2012; 33(4):478-85). It is thus reasonable to assume continuation of the trend in the future is the most likely, although not a certain, scenario. To address the uncertainty raised by the reviewer, we conducted a sensitivity analysis assuming age-specific incidence of cardiovascular disease does not decline any further and will remain constant from (results are shown in Supplement Figure 16 and presented below). 17

18 . Scenario 1: Forecast for number of cases of dementia assuming CVD incidence rates follow downward trends observed in the past decades and continue to decline to 2040 (This is presented in the paper as the most likely scenario) Scenario 3: Forecast for number of cases of dementia assuming CVD incidence rates do not decline further and remain constant from (Scenario based on the assumption that rising diabetes prevalence halts future decline in CVD incidence) 18

19 If CVD incidence rates cease to decline and remain constant until 2040, whilst mortality rates continue to decline, there will be 1.1 million (95% uncertainty interval 1.0, 1.2) people with dementia in 2040, 8% fewer cases compared to the scenario where cardiovascular incidence declines over time. 19

20 POINT 7: They should present a table of assumptions and starting values that drive their model. Without this we can't judge if the assumptions are reasonable. The methods are not sufficiently detailed to allow someone to repeat their study. OUR RESPONSE: Thank you for this helpful suggestion. We have now added a Table of assumptions to the manuscript (Table 1), shown at the end of the response to Editors comments on Page 26 of this document. In the Table we have added details of our methods as suggested by the reviewers. We have also added Tables and Figures to present starting values for the model. Briefly, the IMPACT-BAM model consists of 8 states that include living persons, 2 terminal (absorbing) states, and 39 transition probabilities. Since all input values to the model are by single year of age and sex, the starting values of model inputs consist of 920 prevalence estimates and 4140 transition probabilities. Due to the number of individual starting values, we added Tables and Figures showing starting values in summary format as below: Starting values for prevalence of dementia were presented in Supplement Table 5 and Supplement Figure 9. Dementia incidence has changed over time, and dementia incidence rates by age and sex for years 2005, 2010, and 2015 are shown in Supplement Figure 6. These values are derived from pooling 6 waves of ELSA data, corrected for the effect of selective dropout of study participants and used as input to the IMPACT-BAM model. Crude incidence of dementia by age and sex at each wave of ELSA is shown in Supplement Table 3, and the effect of correction for selective dropout shown in Supplement Figures 7 and 8. Starting values for prevalence of cardiovascular disease are shown in Supplement Figure 2. Starting values for overall cardiovascular and non-cardiovascular mortality rates are shown in supplement Figure 3. Odds ratio of deaths from each health state, which is used to derive health state specific rates of death, are presented in Supplement Table 1. Presenting mortality rates from each of the eight individual health states would require 32 Figures and for brevity we have not shown them in the manuscript. We can present these data if required. The main text points to all these data. POINT 8: In their reporting they should stick to presenting the model and the results of the model. For example, the "conclusion" that incidence of dementia is falling is not based on the model but is input data from another study that drives the model. 20

21 OUR RESPONSE: Thank you. We apologise that the aims of the study have not been presented clearly. We have re-written the last paragraph of the introduction (Pages 5-6) to make the aims of the study clear as below: In the present study, we developed a novel probabilistic Markov method (labelled IMPACT-Better Ageing Model) to simultaneously model the transitions of the population through states of health, cardiovascular disease, cognitive and functional impairment, dementia, to death, to obtain projections for the prevalence of dementia in England and Wales up to To account for the effect of selective dropout of study participants and to obtain unbiased estimates for dementia incidence over the past 15 years, with which to inform the model, we applied a robust joint modelling technique 26 to data from the English Longitudinal Study of Ageing. Although previous studies have shown a decline in dementia incidence, these studies had shortcomings in estimating the magnitude of the decline in dementia incidence. The Framingham heart study had not accounted for the bias in estimating calendar trends due to selective loss to follow up of study participants with dementia (reference 1 below). The CFAS-I and II studies (2) used multiple imputation methods which are based on a missing-at-random assumption. Participants with dementia are more likely to drop out from the study (thus are missing-not-at-random) limiting the applicability of multiple imputation methods. In the present study, we have addressed this limitation by applying a statistical joint modelling technique to obtain probability of developing dementia for those lost to follow up based on trajectories of cognitive function. We found that the decline in dementia incidence is steeper than that shown in previous studies and significant for both men and women. This is an important finding in its own right, as well as serving a vital input to the IMPACT- BAM model. Please also see our response to POINT 9 below. References: 1. Satizabal CL, Beiser AS, Chouraki V, et al. Incidence of Dementia over Three Decades in the Framingham Heart Study. N Engl J Med 2016; 374(6): Matthews FE, Stephan BC, Robinson L, et al. A two decade dementia incidence comparison from the Cognitive Function and Ageing Studies I and II. Nat Commun 2016; 7: doi: /ncomms11398.: POINT 9: Another editor said that the finding that incidence is declining is not novel and has been shown in many studies from different countries. The fact that they estimate that prevalence will increase depends on the inputs to the model but he was not convinced that the assumptions hold. OUR RESPONSE: Thank you for your comment. We recognise that predictions derived from our model are sensitive both to input data and the assumptions we make. As mentioned in response to the 21

22 previous comment (Point 8), although decline in dementia incidence has been demonstrated in previous studies, we have further reduced bias in the estimate of the magnitude of the decline in incidence by accounting for the effect of non-random loss to follow up of study participants, including those with dementia. This correction, based on the method developed by Dr Dimitris Rizopoulos (Rizopoulos D. Joint models for longitudinal and time-to-event data, with applications in R. Chapman and Hall / CRC press; 2012), is novel and has not been used in previous studies estimating the calendar trend in dementia incidence. The importance of deriving an unbiased calendar trend for dementia incidence is demonstrated in Figure 6 (below) in the manuscript by showing how the magnitude of the calendar trend in dementia incidence affects predictions for dementia prevalence. Assumptions underlying the IMPACT-BAM model and evidence supporting those assumptions are presented in Table 1 shown at the end of the response to the Editors comments on Page 26 of this document. We hope this Table helps clarify the strength of the evidence and assumptions underlying our predictions. Figure 6: Sensitivity analysis for number of cases of dementia under alternative assumptions for calendar trend in incidence of dementia Dashed lines represent 95% uncertainty intervals. 22

23 POINT 10: Another editor agreed that if you haven't modelled COPD and cancer that could be a problem. OUR RESPONSE: Thank you. Deaths from cancer, COPD, and other causes have been accounted for in the model, based on appropriate assumptions. A detailed explanation is provided in response to POINT 5 above. POINT 11: Another editor had two concerns: you should discuss the international context. Moreover, he wondered why changing lifestyle patterns including diet and physical activity don t play a role in this paper. OUR RESPONSE: Thank you. We agree that the paper lacked an international perspective and we have now added a paragraph on policy to the Introduction (Page 5) and Discussion on Page 21. By 2050, well in excess of 100 million people worldwide are estimated to be affected by dementia. 1 Current costs of dementia to the UK economy are estimated at 23 billion/year. 2 Burden of disability and years of life lost due to dementia in the UK have increased by 76% between 1990 and Accurate projections for burden of dementia is a key step for planning to meet future needs. Basic research efforts to understand the causes of dementia have increased markedly in recent years, but to date drug trials have failed to show modification of disease processes. 54 The WHO and other expert bodies have identified prevention, identification and reduction of risk as a top research priority, in response to the lack of therapy Debate continues about the relative importance of vascular and neurodegenerative causes. In the context of this uncertainty the IMPACT-BAM model is a means to understand how the dementia burden will evolve, and provides a platform to measure how the burden might be reduced through various policy interventions Our analyses of six waves of ELSA data ( ) confirm previous studies in the UK and elsewhere that age-specific dementia incidence has declined along with the decline in CVD and vascular risks, excluding obesity and diabetes (reference 1, 2 below). Incidence trend studies and other lines of evidence link mid-life health behaviours and vascular risk factors, particularly physical activity and hypertension, with cognitive decline and dementia risk (3, 4). Among risk factors affecting incidence of cardiovascular disease or dementia such as diabetes, smoking, diet, and physical activity, the population levels of some have improved, others deteriorated over time. The sum of the effect of changes in risk factors on changes in mortality rates and incidence of cardiovascular disease and dementia have presented as steady and linear calendar trends (Supplement 23

24 Figure 3, shown on page 9 of this document). In the present paper we present results of a scenario in which past trends will continue to the future. We assumed continuation of such trends would be the most likely scenario. The observed decline in incidence of dementia, cardiovascular disease, and mortality rates, may partly be due to improvements in diet and physical activity. Assuming the steady downward trends in incidence and mortality observed over the past decade will continue to the future, we are indirectly accounting for improvements in diet, physical activity and other risk factors (Please see assumption 14, Table 1 presented below). The impact of changes in diet and physical activity, as well as changes in other vascular risk factors, and the effect of public health policies targeting changes in the risk factors at population level, will be explicitly and carefully modelled in the next phase of the IMPACT-BAM project and are the subject of future papers. These are detailed investigations and reporting the results beyond the scope of this paper. The contribution of educational attainment (3) as well as vascular risk factors to the preventable burden of dementia, underlines the potentially pivotal role of wider determinants as well as interventions based in health care for dementia prevention. There is a need to compare future policy options for prevention, whether based on fiscal measures to change behaviour or improved midlife blood pressure control in primary care (5, 6). The IMPACT-BAM model is a means to measure the impact of such interventions and policy scenarios. References: 1. Matthews FE, Stephan BC, Robinson L, et al. A two decade dementia incidence comparison from the Cognitive Function and Ageing Studies I and II. Nat Commun 2016; 7: doi: /ncomms11398.: Satizabal CL, Beiser AS, Chouraki V, et al. Incidence of Dementia over Three Decades in the Framingham Heart Study. N Engl J Med 2016; 374(6): Norton S, Matthews FE, Barnes DE, et al. Potential for primary prevention of Alzheimer's disease: an analysis of population-based data. Lancet Neurol 2014; 13(8): Ngandu T, Lehtisalo J, Solomon A, et al. A 2 year multidomain intervention of diet, exercise, cognitive training, and vascular risk monitoring versus control to prevent cognitive decline in at-risk elderly people (FINGER): a randomised controlled trial. Lancet 2015; 385(9984): Lincoln P, Fenton K, Alessi C, et al. The Blackfriars Consensus on brain health and dementia. Lancet 2014; 383(9931): NICE guideline. Dementia, disability and frailty in later life: mid-life approaches to delay or prevent onset

25 - Several other editors were supportive. In your response please provide, point by point, your replies to the comments made by the reviewers and the editors, explaining how you have dealt with them in the paper. OUR RESPONSE: Thank you. We have provided point by point responses to each of the reviewers comments, below. NOTE: recent modification to the IMPACT-BAM model While this paper was under review we modified the structure of the IMPACT-BAM model. In order to use IMPACT-BAM to also predict numbers of people living with disability we added a tenth health state to the model to represent persons with disability without cardiovascular disease and dementia. Addition of this health state to the model had a minor effect on projections for dementia prevalence. Results in the revised paper reflect this modification. Parts of the methods section were described in the supplementary material which we have moved to the main manuscript at the request of the reviewers. We are happy to move sections to the supplement material if a shorter manuscript is favourable. 25

26 Table 1: Summary of assumptions underlying the IMPACT-BAM model. Assumption Justification IMPACT-BAM models health transitions of the population of England and Wales aged 35 and over through to death. The input data for the probabilistic Markov model are the population size in each age and sex stratum, initial health state prevalence values, and transition probabilities by age, sex, and calendar year. Population numbers by age and sex Estimates for population numbers by sex and 5- year age-groups at model baseline were obtained from the UK Office for National Statistics (ONS). At each 1-calendar-year iteration of the model, men and women reaching age 35 were entered. The predictions for number of people aged 35 by year were obtained from the ONS. The UK Office for National Statistics provides official estimates for population demographics. Assumption 1: ONS predictions are realistic; Assumption 2: migration is not a major source of bias. Starting prevalence values Initial prevalence of health states in the model by age and sex were obtained from the English Longitudinal Study of Ageing (ELSA). Assumption 3: ELSA is a representative sample of the population of England and Wales. To improve statistical power, six waves of ELSA data were pooled. Prevalence estimates of CVD, cognitive, and functional impairment that define the health states were obtained from pooled data and attributed to year 2006 which Accuracy of prevalence values depends on how well ELSA represents the population of England and Wales. ELSA study participants aged 50 and over were selected at random. The core participant s cohabiting partners, including adults aged below 50, were also enrolled in the study. The overall response rate was 67%. To ensure study participants form a representative sample, survey weights are applied. To maintain representativeness at every phase of data collection, refreshment samples are recruited to the study periodically. Comparisons of the socio-demographic characteristics of participants against results from the national census indicated that the ELSA sample was broadly representative of the English population The prevalence values obtained from the pooled ELSA data matched the prevalence values obtained at the mid-point (2006, wave three). Estimates for prevalence of cardiovascular disease are displayed as an example in Supplement Figure 2. 26

27 the mid-point of the ELSA data collection timeframe and the baseline of the model. Assumption 4: Prevalence estimates from 6 pooled waves of data provide a precise and accurate estimate of prevalence at mid-point of the data collection time-frame. Assumption 5: The prevalence of each health state at each calendar year from the starting point (2006) onwards, equals number of persons who were in that health state in the previous year, plus new incident cases, minus those who made the transition to another health state or died from Epidemiologic concept applied to Markov models. 27

28 any cause. Number of new incident cases and numbers of death were determined by transition probabilities to and from that condition. Transition probabilities Transition probabilities were obtained as a function of age and sex from incident cases between wave n and n+1 in ELSA. As with estimates of prevalence values, the transition probabilities obtained from pooling ELSA epochs were attributed to the mid-point of the data collection period. Incidence of cardiovascular disease and dementia by age and sex were consistent with age, and sex specific incidence values obtained from independent external sources for the corresponding calendar time. Estimates of dementia incidence were available from the Cognitive Function and Ageing Study II (CFAS II, ). Incidence of dementia in the corresponding time-frame is compatible with CFAS-II estimates (Supplement Figure 5). Assumption 6: Transition probabilities, (equivalent to incidence by age/sex/calendar year) for cardiovascular disease, dementia, functional impairment, and mortality in ELSA are similar to those for England and Wales. 28

29 Deaths predicted by IMPACT-BAM matched with observed and predicted mortality rates from the ONS (Supplement Figure 13). Assumption 7: Transition probabilities are equivalent to a weighted average across the spectrum of the severity of each condition thus varying severities among people in each health state is accounted for. Similarly, survival of persons with each condition is assumed to be equivalent to the weighted average of survival of persons with different levels of severity. Assumption 8: The effects of comorbidity (such as diabetes) are accounted for in the model. Modelling is based on a single transition probability for each age/sex/calendar-year strata and health transition. The probability of death or development of functional impairment among those with cardiovascular disease or cognitive impairment is dependent on the severity of cardiovascular disease or cognitive impairment. Under the assumption that ELSA participants are a representative sample of the population (see above), the spectrum of the severity of conditions (e.g. cardiovascular disease, or cognitive impairment) observed in ELSA is proportionate to that at population level. As such, transition probabilities obtained from ELSA are a weighted average of the transition probabilities across the spectrum of the severity of the conditions. The weighted average transition probability multiplied by the total number of people in a health state is mathematically equivalent to the sum of the product of severity specific transition probabilities and severity specific numbers of people in that health state. Since ELSA participants are assumed to be a representative sample of the population of England after weighting (see above), estimates for risks of dementia, cardiovascular disease, functional impairment and death obtained from ELSA reasonably represent a weighted average of risk levels across the spectrum of the severity of these conditions and comorbidities. 29

30 Calendar trends: Transition probabilities (mortality rates and incidence of cardiovascular disease and dementia) change over time. Assumption 9: The observed downward calendar Data obtained from the UK Office for National Statistics show that cardiovascular and noncardiovascular mortality rates followed steady and linear downward trends over the past two decades. trend in mortality rates over the past two decades will continue to the future. We assumed the most likely scenario would be that these trends will continue (Supplement Figure 3). Assumption 10: Life-expectancy and maximum life-span are amenable to being increased Changes in life-expectancy are accounted for by application of mortality rates. As mortality rates continue to decline, life-expectancy will increase. 30

31 Assumption 11: Trends in incidence of cardiovascular disease over time are parallel to cardiovascular mortality. Age and sex standardized cardiovascular incidence and mortality rates declined in parallel in ELSA (Supplement Figure 4). To examine the uncertainty of this assumption, we conducted a sensitivity analysis assuming incidence of cardiovascular disease does not decline any further (results shown in Supplement Figure 16). Assumption 12: Dementia incidence declines over time. A decline in dementia incidence has been reported in studies in England, Holland and USA. The magnitude of the calendar trend in England and Wales is less certain. We determined the calendar trend corrected for deaths and loss to follow up of study participants utilizing a robust statistical technique to model ELSA data. We conducted sensitivity analyses with calendar trends estimated from other studies (including CFAS I and II, and the Framingham study), and an alternative scenario in which dementia incidence does not decline any further (Figure 6). 31

32 Assumption 13: Survival with cardiovascular disease, dementia, or functional impairment change over time in parallel to changes in overall life-expectancy Assumption 14: The most likely net effect of future changes in risk factors will be the continuation of calendar trends in mortality rates and incidence of dementia and cardiovascular disease observed over the past two decades. Competing risks Assumption 15: Deaths due to any cause (including cancer and COPD) and changes in cause-specific mortality rates act as competing risks to development of dementia. Survival in IMPACT-BAM is indirectly modelled as a function of changing mortality rates. It is assumed that the ratio of mortality rates for each health state in the model compared to the general population is similar to that observed in ELSA. Thus mortality and survival for each health state in the model changes in parallel to mortality and survival in the general population. Current evidence suggests survival with cardiovascular disease and dementia is improving over time. We did not find evidence suggesting this improvement to be over and beyond improvement in overall survival. Population levels of risk factors affecting incidence of cardiovascular disease and dementia such as diabetes, smoking, diet, and physical activity have changed over time. The net effect of recent changes in risk factors on changes in mortality rates and incidence of cardiovascular disease and dementia has been steady and linear declining calendar trends. The present analysis forms the baseline modelling scenario. IMPACT-BAM will be utilized in future to model the health impacts of changes in risk factors and public health interventions compared with the baseline scenario. Results of such analysis are extensive and beyond the scope of this paper. Cardiovascular and non-cardiovascular causes of death are the terminal health states in the model. Once a person dies from any cause they are no longer at risk of disease. Thus, competing risks due to both cardiovascular and non-cardiovascular causes are accounted for in the model. 32

33 Comments from Reviewers Reviewer: 1 Comments: TO THE EDITOR This is a very interesting study presenting data from the English Longitudinal Study of Aging (ELSA) on trends in dementia incidence and prevalence. Although previous studies already have shown that dementia incidence is declining, the present study used the Markov model IMPACT-BAM to predict future dementia prevalence taking into account the effects of competing risks (cardiovascular disease, mortality). Yet, the study has some minor issues that should be addressed or at least discussed more critically. I expect the authors to be capable of making such minor revisions. TO THE AUTHORS GENERAL COMMENTS In this manuscript, Ahmadi-Abhari and coworkers present the results of an estimation of dementia incidence and future dementia prevalence based on the ELSA dataset ( ). A probabilistic Markov model was used that takes into account the effects of cardiovascular disease. The manuscript is well written and the topic seems relevant to this journal. Major strengths include the large population-based sample, and the validation of results against independent reports. I hope that the following comments help strengthening the manuscript further. MINOR ESSENTIAL/DISCRETIONARY REVISIONS POINT 1: Abstract (sentence 44): Try to avoid the use of we. Perhaps change to: A recent decline in age-specific dementia incidence was confirmed. OUR RESPONSE: We thank the reviewer for his evaluation of our paper and thoughtful comments. In response to this point, we have edited the sentence in the abstract and avoided the nominative case throughout the manuscript. 33

34 POINT 2: Introduction: Information on the societal importance of this study is lacking in this section (e.g. policy implications, effects of health costs, healthcare burden, input for public health campaigns, etc.). OUR RESPONSE: We thank the reviewer for this valid suggestion. This information has been added to the Introduction, Page 5. By 2050, well in excess of 100 million people worldwide are estimated to be affected by dementia. 1 Current costs of dementia to the UK economy are estimated at 23 billion/year. 2 Burden of disability and years of life lost due to dementia in the UK have increased by 76% between 1990 and Accurate projections for burden of dementia is a key step for planning to meet future needs. Addressing shared determinants of disease follow recommendations from the National Institute for Health and Care Excellence (NICE) 10, Public Health England, and WHO in the Blackfriars consensus. 11 POINT 3: Introduction (sentence 14): check citation 12;13 should be without the ; (throughout whole manuscript) OUR RESPONSE: We thank the reviewer for pointing out the style required and we have now corrected all the references throughout the manuscript. POINT 4: Introduction (sentence 14) Thus, vascular risk reduction is likely to drive Did you see the results of this 6-year multi-domain vascular care intervention to prevent dementia? ( Results did not lead to reduced incidence of all-cause dementia. Perhaps rephrase sentence. OUR RESPONSE: Thank you. We appreciate your cautious perspective on dementia prevention. There are several lines of evidence on this unresolved issue (please see our response to the BMJ committee, point 11 on page 23 of this document). The prediva trial which is referred to, produced an important null result, with design limitations which are instructive. prediva recruited participants in their eighth decade, follow up was of limited duration, and the setting involved high levels of usual care for vascular risk factors. The intervention did not lead to a difference in levels of risk factors between the intervention and control group. As expected this did not lead to a difference in incidence 34

35 of cardiovascular disease, although the effect of the vascular risk factors targeted by the intervention and risk of cardiovascular disease is well established and has been shown in several trials previously. Therefore, we think that the prediva trial results do not preclude a beneficial role for midlife vascular risk factor control. Whilst several observational longitudinal studies have shown an association between vascular risk factors and incident dementia (references 1-3 below), uncertainties on the value of vascular risk reduction for dementia prevention remain. We have now summarised this point in the discussion Page 21: Basic research efforts to understand the causes of dementia have increased markedly in recent years, but to date drug trials have failed to show modification of disease processes. 54 The WHO and other expert bodies have identified prevention, identification and reduction of risk as a top research priority, in response to the lack of therapy Debate continues about the relative importance of vascular and neurodegenerative causes. In the context of this uncertainty the IMPACT-BAM model is a means to understand how the dementia burden will evolve, and provides a platform to measure how the burden might be reduced through various policy interventions References: 1: Norton S, Matthews FE, Barnes DE, Yaffe K, Brayne C. Potential for primary prevention of Alzheimer's disease: an analysis of population-based data. Lancet Neurol Aug;13(8): : Barnes DE, Yaffe K. The projected effect of risk factor reduction on Alzheimer's disease prevalence. Lancet Neurol Sep;10(9): Deckers K, van Boxtel MP, Schiepers OJ, et al (2015) Target risk factors for dementia prevention: a systematic review and Delphi consensus study on the evidence from observational studies. Int J Geriatr Psychiatry. 30: POINT 5: Methods: ELSA Wave 7 data (up to 2015) is available ( Perhaps this can be used as a confirmation of your estimation of dementia incidence in 2015 (if available)? OUR RESPONSE: We thank you for bringing this to our attention. We have responded to this point in detail in our response to POINT 4 of the additional comments by the Editors committee, please see our response on pages 4 and 5 of this document. Briefly, we have followed your advice and 35

36 provided analysis comparing model predictions with dementia prevalence at wave 7 ELSA (results shown in Supplement Figure 11) POINT 6: Methods: participants were recruited around age 50-55: how was information collected regarding early life events/health outcomes (retrospectively?)? Is the ELSA population a true representative of the general population in England and Wales (e.g. wealth, minority/race, and educational level)? OUR RESPONSE: Thank you for pointing out the lack of detail in our methods. We have now added text to the Methods (Page 7) as well as a flow-diagram to better describe the sampling method for ELSA (Supplement Figure 1): The English Longitudinal Study of Ageing (ELSA) sample was recruited in from participants of the Health Surveys for England (HSE) It was drawn by postcode sector, stratified by health authority and proportion of households in non-manual socioeconomic groups. A total of 12,099 men and women participated (response rate 67%) including 11,392 persons aged 50+ selected through the random sampling and 707 cohabiting partners. To ensure study participants form a representative sample, survey weights were applied. Comparison of sociodemographic characteristics against national census indicated the ELSA sample was broadly representative of the population of England. 25 To maintain representativeness of the study sample, refreshment participants were recruited to the study at wave 3 ( ; ages 50-55), wave 4 ( ; ages 50-74), and wave 6 ( ; ages 50-55) all drawn from the HSE of the preceding years. A total of 17,906 participants were recruited to the study between waves 1 to 6 (Supplement-Figure 1). As described above the ELSA sample includes participants aged 18+. Data from these participants are used to derive transition probabilities to cardiovascular disease and functional impairment below age 50. POINT 7: Methods: the cognitive tests administered in ELSA are limited (as you already mention in the Discussion). In section 1.2 of the Supplement the cognitive domains are described. Orientation in time, day, month, year was used to assess concentration and the numeracy test as measure of executive function. This can be discussed/questioned of course. Are the studies that used the same procedure for establishing cognitive domains? Please cite these. 36

37 OUR RESPONSE: Thank you. We agree that the set of cognitive function tests administered in ELSA is not comprehensive, however, they have been validated and are among standard tests for assessment of impairment in cognitive domains and are used in other studies including the Health and Retirement Study (HRS) in USA (below references). To avoid enlarging the paper excessively, we added references to previous studies assessing the utility and validity of the individual cognitive function tests rather than a detailed discussion on each individual test (Discussion Page 19): Cognitive assessment in ELSA is based on a set of standard and validated cognitive function tests; nonetheless the list is not comprehensive. Cognitive impairment in domains other than those tested may have been missed, leading to underestimation of dementia cases. Similarity of age-specific estimates of dementia incidence and prevalence with those of CFAS-II suggests this source of bias is small. References for validation of cognitive assessment employed in the present study: 1. Chertkow H, Nasreddine Z, Joanette Y, Drolet V, Kirk J, Massoud F, Belleville S, Bergman H. Mild cognitive impairment and cognitive impairment, no dementia: Part A, concept and diagnosis. Alzheimers Dement Oct;3(4): Stephan BC, Kurth T, Matthews FE, Brayne C, Dufouil C. Dementia risk prediction in the population: are screening models accurate? Nat Rev Neurol Jun;6(6): Lonie JA, Tierney KM, Ebmeier KP. Screening for mild cognitive impairment: a systematic review. Int J Geriatr Psychiatry Sep;24(9): Langa KM, Llewellyn DJ, Lang IA, Weir DR, Wallace RB, Kabeto MU, Huppert FA. Cognitive health among older adults in the United States and in England. BMC Geriatr Jun 25;9: Howieson DB, Dame A, Camicioli R, Sexton G, Payami H, Kaye JA. Cognitive markers preceding Alzheimer's dementia in the healthy oldest old. J Am Geriatr Soc May;45(5): Masur DM, Sliwinski M, Lipton RB, Blau AD, Crystal HA. Neuropsychological 37

38 prediction of dementia and the absence of dementia in healthy elderly persons. Neurology Aug;44(8): Welsh KA, Butters N, Hughes JP, Mohs RC, Heyman A. Detection and staging of dementia in Alzheimer's disease. Use of the neuropsychological measures developed for the Consortium to Establish a Registry for Alzheimer's Disease. Arch Neurol May;49(5): Knopman DS, Ryberg S. A verbal memory test with high predictive accuracy for dementia of the Alzheimer type. Arch Neurol Feb;46(2): POINT 8: Methods: participants were compared with a reference population aged with the same level of education regarding cognitive tests scores. Do you also have different reference groups for males and females? There are some differences regarding cognitive performance between men and women, especially in the older age groups. OUR RESPONSE: Thank you. We are grateful for this note about the lack of clarity in our Methods section (Page 9). We have added the following text: We did not find considerable differences between men and women in the distribution of cognitive function scores after adjustment for age and education. The sex specific cut-off values to determine impairment on each cognitive test, defined as 1.5 SD below the mean in the participants with the same level of education, were similar in men and women. POINT 9: Methods: the IQCODE was used to determine cognitive impairment in persons who were unable to take part in the study. A proxy informant is asked about the participant s memory abilities, daily situations, etc. What was the relation between the participant and the proxy informants asked to complete this questionnaire? Partners, children, neighbors, nurses? I think this will make a difference. Additionally, are there any references that validate using this instrument for this purpose? OUR RESPONSE: Thank you for raising this point. The proxy informant responding to the IQCODE questionnaire was an immediate family member (partner/children) for the majority of participants. The IQCODE questionnaire for identifying cases of dementia has been validated in 38

39 several studies and a systematic review has summarized the findings (referenced in the paper). We have added both points and relevant references to the methods section (Page 8). For participants unable to take part in the study who provided consent in advance or through a consultee, the short version of the informant questionnaire for cognitive decline (IQCODE) was administered. 29 The IQCODE comprises 16 questions asking a proxy informant (usually an immediate family member) how the participant s state of memory, ability to learn new tasks, judgement, and handling of key everyday situations (e.g., making decisions on every day matters, or handling money for shopping) are compared to two years ago. The answers are graded on 5-point scales from muchimproved to much-worse. Use of the IQCODE questionnaire to identify cognitive impairment has been previously validated. 29 The participant or proxy informants were asked about any doctor diagnosis of dementia. Reference 29: Harrison JK, Fearon P, Noel-Storr AH, McShane R, Stott DJ, Quinn TJ. Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE) for the diagnosis of dementia within a secondary care setting. Cochrane Database Syst Rev 2015 March 10;3:CD POINT 10: Methods: dementia cases were defined on either cognitive + functional impairment or self-reported doctor diagnosis of dementia. What was the overlap between both? Why not also using a combination (very sure cases of dementia) in a sensitivity analysis? Which % of dementia cases is based on self-report of a doctor s diagnosis and which % is based on objective measurements of cognitive and functional abilities? OUR RESPONSE: Thank you. We appreciate that this key information should have been included in the paper. It has been added to the first paragraph of the revised results section (Page 15) as follows: Between 2002 and 2013 dementia was ascertained for a total of 1,448 ELSA participants. 634 participants had dementia at time of recruitment and 814 incident cases were identified over the course of the follow up. 16.5% of cases occurred in participants below age 65. Of the 1,448 cases of dementia, 466 received a doctor diagnosis of dementia, 245 were identified via IQCODE questionnaire plus ADL functional impairment, and 1,078 were identified by impairment in cognitive tests and impairment in one or more ADLs. Of the 245 cases based on IQCODE and functional impairment, 171 (70%) reported a doctor diagnosis of dementia. Among those with impairment in cognitive tests and ADL function, 263 (25%) reported a doctor diagnosis of dementia at later stages or were positive on the IQCODE questionnaire. 39

40 The modest (25%) overlap between cases of dementia based on impairments in cognitive tests plus ADLs with doctor diagnosis of dementia is not unexpected. A proportion of dementia cases remain clinically undiagnosed, especially at less severe stages, but may be captured by administration of cognitive function tests. It is also possible that a proportion of participants with cognitive and functional impairment will have been diagnosed with dementia by a doctor at a later stage, however, having withdrawn from the study a record of the doctor diagnosis may not have been established. On the other hand, in ELSA cognitive function tests were not administered for participants who had already been diagnosed with dementia. POINT 11: Methods: definition of cardiovascular disease? No information available on atrial fibrillation given it association with cognitive impairment ( It s debatable whether stroke should be included in a cardiovascular compound, perhaps better to exclude persons with stroke to have a more clearly compound of cardiovascular disease. What is you rationale behind this? How was death from cardiovascular disease established? Is it based on death certificate data or from the ONS data, and what is the accuracy? OUR RESPONSE: Thank you for raising this point about atrial fibrillation and stroke. We have included stroke, but not atrial fibrillation, in the definition for cardiovascular disease for two reasons. Firstly, because of the overlap in the pathophysiology and shared risk factors leading to stroke and other cardiovascular disease presentations; and secondly, to allow comparison of our findings with other studies on prevalence of cardiovascular disease, including the Health Survey for England. Cause of death was ascertained based on the ICD-9 and ICD-10 codes on death certificates. That is the most accurate data that is available for cause of death data in the UK. Atrial fibrillation and stroke are indeed risk factors for dementia. It is not feasible to explicitly model all co-morbidities in the model. However, in IMPACT-BAM, initial prevalence estimates and transition probabilities to and from each health state in the model are, in effect, a weighted average of the conditions included in that health state. If stroke is excluded from the cardiovascular states and included in a separate state or combined with state 1 (disease free state), both the numbers of people with stroke and the weight stroke has on transition probabilities to dementia would be transferred from the cardiovascular disease state to the new state. Although this would affect the projections for numbers of people with cardiovascular disease, it would not greatly affect projections for numbers of people with dementia. 40

41 Since calculations of prevalence and transition probabilities are based on weighted averages in a population representative sample (ELSA), explicitly modelling comorbidities will not considerably improve predictions. This in now addressed in the Table of assumptions in the paper (Table 1) that is also presented at the end of the response to comments from the Editors committee on Page 26 of this document To display mathematically, we call total number of people with CVD or stroke N Total, numbers of people with stroke N Stroke and numbers of people with CVD (excluding stroke) N CVD; transition probability to dementia for people with stroke P Stroke and transition probability to dementia from CVD P CVD. If we create separate health states for CVD and stroke, numbers of new cases of dementia from these two states would be: New dementia cases among those with CVD or stroke = (P Stroke x N Stroke ) + (P CVD x N CVD ) If we multiply and divide the above equation by total number of people with stroke and CVD we get: New dementia cases = N Total x [(P Stroke x N Stroke /N Total ) + (P CVD x N CVD /N Total )] [(P Stroke x N Stroke /N Total ) + (P CVD x N CVD /N Total )] is the weighted average transition probability to dementia from people with CVD or stroke. The equations show grouping health states together and applying weighted average transition probabilities is mathematically equivalent to modelling a comorbidity as a separate health state. POINT 12: Statistical methods (sentence 32): we assumed the decline ; perhaps not entirely clear for the reader which decline you are referring to. OUR RESPONSE: Thank you for raising this point. This section has been corrected to make it clear as below (Page 13): The decline in cardiovascular incidence paralleled the decline in cardiovascular mortality in ELSA (Supplement Figure 4). We hence used the annual percentage change in cardiovascular mortality to estimate temporal change in cardiovascular incidence. POINT 13: Results: decline in the future prevalence of dementia in women is reported and in the Discussion (page 9, sentence 39) you mention as the current narrowing of the life expectancy gap between the sexes continues. Can you hypothesize or explain why this is about to occur? 41

42 OUR RESPONSE: Thank you. We agree that this point requires some explanation. We edited the text in the discussion (Page 19) as below: the numbers of men and women living with dementia is set to converge within the next 15 years. This finding can partly be explained by a faster decline in mortality rates in men compared to women (shown in Supplement Figure 3) and narrowing of the life expectancy gap between the sexes. 9 According to data from the UK Office for National Statistics, the decline in mortality rates in men is faster than that in women and the gap is narrowing (Figures below). Bayesian forecasting has showed that between 1981 and 2012, life expectancy increased by 8 2 years for men and 6 0 years for women, closing the female-male gap from 6 0 to 3 8 years (Reference 9: Bennett JE et al. The future of life expectancy and life expectancy inequalities in England and Wales: Bayesian spatiotemporal forecasting. Lancet 2015; 386(9989): ). Detailed investigations into the reasons for the faster decline in mortality rates in men compared to women was beyond the scope of this paper. POINT 14: Results/Discussion: Changes over time in available risk factors accounted for about 22% of the calendar effect in dementia incidence (Supplement Table 3). Why did you choose for these risk factors? Why not also include depression? Depression is well-known risk factor of cognitive 42

43 impairment or dementia ( Other possible risk factors to include are healthy diet and socio-economic status/wealth (both available in ELSA based on the study website) OUR RESPONSE: Thank you for your suggestion. The risk factors chosen are known vascular risk factors that are also probable risk factors for dementia. Following your suggestion, we have now added socioeconomic status, net wealth and income, and depression as covariates in regression models with calendar trends in dementia incidence over time as the outcome. These results are now included as Supplement Table 4. Supplement Table 4: Relative annual change in incidence of dementia adjusted for change in level of risk factors. Annual Change in Incident Dementia ( ) Odds Ratio (95% CI) Relative Annual Change (%) (95% CI) Calendar Trend (per year) (0.971, 0.976) -2.7 (-2.9, -2.4) % Adjusted for Education (0.974, 0.979) -2.3 (-2.6, -2.1) % Physical Activity (0.980, 0.985) -1.7 (-2.0, -1.5) % Smoking (0.964, 0.970) -3.3 (-3.6, -3.0) % Alcohol Intake (0.973, 0.979) -2.4 (-2.7, -2.1) % Body Mass Index (0.975, 0.980) -2.2 (-2.5, -2.0) % Blood Pressure (0.975, 0.981) -2.2 (-2.5, -1.9) % LDL and HDL Cholesterol (0.977, 0.982) -2.1 (-2.3, -1.8) % Diabetes (0.970, 0.975) -2.8 (-3.0, -2.5) % Stroke (0.970, 0.976) -2.7 (-3.0, -2.4) % Depression (0.971, 0.977) -2.7 (-2.9, -2.3) % Social Class (0.968, 0.974) -3.0 (-3.2, -2.6) % Total Income and Net Wealth (0.974, 0.979) -2.4 (-2.6, -2.1) % Multivariable Adjusted (0.977, 0.982) -2.0 (-2.3, -1.7) % We have commented on these results in the manuscript as follows (Discussion Page 21): There are several plausible explanations supporting a decline in dementia incidence over time. Improvement in vascular risk factors, as well as in education levels, can partly account for the decline in incidence. In the present study, increased physical activity accounted for the largest proportion of the decline in dementia incidence between 2002 and Changes in prevalence of diabetes, smoking, and social class over time had negative confounding effects, such that the 43

44 downward incidence trend increased after controlling for respective changes. Adjustment for stroke and depression did not have a significant effect on the calendar trend in dementia POINT 15: Results/Discussion: check the correct use of abbreviations CFAS and ELSA throughout the manuscript (e.g. Discussion, page 10, sentence 9 OUR RESPONSE: Thank you for spotting the inconsistencies. All abbreviations have been checked and inconsistencies corrected. POINT 16: Discussion: Cognitive decline starts at a younger age than the 65+ or 70+ age-groups recruited in previous studies. I agree with this point, but some of these studies only look at dementia as an outcome and therefore focused on a group >60/65 due to limited number of people with dementia under age 60/65. This leads directly to the following question: How many people (%) in the ELSA cohort showed significant cognitive decline before the age of 60? OUR RESPONSE: Thank you for asking this important question. The numbers of cases of dementia in each age-group are presented in Supplement Tables 3 and 5. In summary, of the 1,448 cases of dementia observed in ELSA from waves 1 to 6, a total of 239 (16.5 %) developed dementia below age 65, and a total of 158 (10.9 %) developed dementia below age 60. This information has been added to the results (please see our response to POINT 10 above). POINT 17: Discussion, sentence 12; which CFAS study? 1, 2 or both? Provide reference as evidence OUR RESPONSE: Thank you for raising this point. We have clarified which CFAS study we refer to throughout the manuscript. For the purpose of validation of estimates for age-specific incidence and prevalence of dementia we compared our findings with those of CFAS-II study. The previously published calendar trend in dementia incidence was derived from CFAS I and II (Matthews FE et al. A two decade dementia incidence comparison from the Cognitive Function and Ageing Studies I and II. Nat Commun 2016; 7: doi: /ncomms11398.:11398). 44

45 POINT 18: Discussion: you mention declining cardiovascular disease incidence and Improvement in vascular risk factors. Please elaborate on this important matter (e.g. better cardiovascular disease management programs) OUR RESPONSE: Thank you for highlighting our rather cursory statement. We have further elaborated on this matter in the discussion Page 21: There are several plausible explanations supporting a decline in dementia incidence over time. Improvement in vascular risk factors, as well as in education levels, can partly account for the decline in incidence. In the present study, increased physical activity accounted for the largest proportion of the decline in dementia incidence between 2002 and Changes in prevalence of diabetes, smoking, and social class over time had negative confounding effects, such that the downward incidence trend increased after controlling for respective changes. Adjustment for stroke and depression did not have a significant effect on the calendar trend in dementia. Basic research efforts to understand the causes of dementia have increased markedly in recent years, but to date drug trials have failed to show modification of disease processes. 54 The WHO and other expert bodies have identified prevention, identification and reduction of risk as a top research priority, in response to the lack of therapy Debate continues about the relative importance of vascular and neurodegenerative causes. In the context of this uncertainty the IMPACT-BAM model is a means to understand how the dementia burden will evolve, and provides a platform to measure how the burden might be reduced through various policy interventions. In line with our results, previous studies, including the Framingham heart study (reference 1 below), showed that improvements in vascular risk factors partly accounted for the decline in dementia incidence and the study by Kontis V et al (reference 2) showed the effect of vascular risk reduction on declining mortality rates. The decline in cardiovascular incidence and improvements in vascular risk factors offer possible explanations for changes observed in the epidemiology of dementia. Examining the reasons for changes in vascular risk factors or factors driving the decline in cardiovascular disease, although of high importance, was beyond the scope of the present study. 1: Satizabal CL, Beiser AS, Chouraki V, Chene G, Dufouil C, Seshadri S. Incidence of Dementia over Three Decades in the Framingham Heart Study. N Engl J Med 2016; 374(6): : Kontis V, Mathers CD, Rehm J, Stevens GA, Shield KD, Bonita R, Riley LM, Poznyak V, Beaglehole R, Ezzati M. Contribution of six risk factors to achieving the non-communicable 45

46 disease mortality reduction target: a modelling study. Lancet Aug 2;384(9941): doi: /s (14) Point 19: Table 1(page 5): Do you have an explanation why the number of people from Wave 1 and Wave 4 with diabetes and cardiovascular disease increase with almost 100%? These are substantial changes. OUR RESPONSE: Thank you for identifying the error in the table. The correct figures are 7.2% and 9.9% for prevalence of diabetes at waves 1 and 4 respectively, and 14.9% and 14.0% for prevalence of CVD. The cause of the error has been identified and numbers have now been corrected in Supplement Table 2. The tabulation error was independent of the IMPACT-BAM input data. Additional Questions: Please enter your name: Kay Deckers 46

47 Reviewer: 2 Recommendation: Comments: The authors analyzed data from the ELSA to explore whether the incidence of dementia in the UK has changed in recent years, and to estimate how this would impact on the projections of dementia prevalence in the country in the next decades. The aging of societies is expected to lead to catastrophic increases in the number of people with dementia under the assumption that the age- and sex-specific prevalence of dementia will not vary over time. Recent epidemiological studies in Europe and the US suggested that this assumption may not hold, but results are not unequivocal. Despite this research is extremely important for policy makers and clinicians, it is not POEM. Changes over time, or trends, in prevalence and incidence of dementia have been reported in recent years in the MRC-CFAS study. The MRC-CFAS was purposely designed and powered to detect any such changes. Further evidence is warranted, but the present study is not original and has important limitations. Previous studies have been conducted in the UK on larger and better defined samples, using more advanced research protocols; in previous papers the potential issues related to differential attrition and the variability of response rates were also addressed (the latter was not explicitly addressed by the authors of the present study) [See Matthews F et al. The Lancet 2013]. There are several important limitations in the study design and in the presentation and interpretation of the results, the internal validity of the study itself may be not adequate. OUR RESPONSE: We thank the reviewer for this thoughtful assessment of our study. We offer a point by point response below: MAIN COMMENTS POINT 1: Constant methods and diagnostic criteria to ascertain dementia in community dwelling samples are required to explore changes over time in dementia incidence and prevalence. Although the criteria to establish dementia caseness were kept constant in the present study, they were not formally validated, and they were not equal for all participants. Cognitive assessment was limited to 47

48 questions on time orientation (but not space) and two cognitive tests, one of memory and one of verbal fluency. The informant questionnaire consisted in the short version of the IQCODE, which was used only in a subsample of those with no cognitive assessment. The outcome of the present study is probably best defined as cognitive impairment (objective or reported by a next of kin) rather than dementia, because cognitive assessment was too limited and evidence of significant decline from a previous level of performance [ ] (DSM) was not sought. Further, learning effects in cognitive tests between the 2-year intervals were likely. This could have introduced bias in outcome ascertainment through the study follow-ups, with repercussions on the validity of the comparisons of incident cognitive impairment over time. OUR RESPONSE: Thank you for your insightful comments on our case definition of dementia. Below we provide a detailed response: (i) Cognitive assessment and bias in outcome ascertainment Cognitive assessment in the present study included orientation to time, two tests for memory function, separate tests for verbal fluency, and numeracy function. We do acknowledge that the cognitive assessment based on cognitive function tests was not comprehensive and thus we may have missed cases with cognitive impairment in domains other than the ones being tested. However, assessment of cognitive impairment was complemented by the IQCODE questionnaire and doctor diagnosis of dementia. To account for cases that may not have been identified by any of the above methods or among participants who were lost to follow up, we applied a joint modelling of longitudinal and time to event method to predict the occurrence of dementia based on the trajectory of cognitive function. We compared our estimates for incidence and prevalence of dementia with those obtained from the CFAS-II study (Supplement Figure 5 (below)). The comparison suggests bias due to underascertainment of dementia cases is small. The cornerstone of clinical diagnostic criteria for dementia (including the DSM-IV) is impairment in two or more cognitive domains that result in considerable loss of function. We explain in Section 1 of the supplement material how the definition applied in this study resembles these criteria: Methods Page 9: Dementia caseness was defined either as a combination of cognitive impairment (according to the definitions provided on Methods Page 9) and functional impairment (difficulty in performing one or more activities of daily living), or self-reported doctor diagnosis of dementia. We adapted the dementia case definition to resemble DSM-IV and other criteria (such as NINDES- AIREN and NINCDS-ADRDA) for diagnosis of dementia. The cornerstone of clinical diagnostic criteria for dementia is impairments in two or more cognitive domains that result in considerable loss of function. Thus, we defined cognitive impairment as a score of equal to or lower than 1.5 standard deviations below mean, standardised to the population aged with the same level of education, 48

49 similar to criteria used for defining cognitive impairment no dementia (CIND). Loss of function was defined as impairments in conducting activities of daily living. We aimed for a set of criteria to encompass all types of dementia and not merely Alzheimer s disease. Although memory impairment is a key element in the diagnosis of Alzheimer s disease, memory is affected to varying degrees in vascular, fronto-temporal and Lewy body types of dementia. Thus, memory impairment was not included as a necessary criterion in defining cognitive impairment in this study. DSM-IV criteria specify that the disturbances do not occur exclusively during the course of delirium and are not better accounted for by another mental disorder. For the criteria to hold, and to increase specificity, transient impairments in cognitive function or conducting ADLs, were not classified as cognitive or functional impairment. Following your comments, we have now also added the following to the limitations of the study section in the discussion (Page 19) ELSA participants were not clinically screened for dementia, rather, an operational case definition based on standardised assessments of cognition and function was applied. The standardised assessment is comparable across time, and thus more informative of dementia trends than clinical assessments which are likely to be affected by changes in diagnostic criteria and clinical practice over time. 23 The case definition applied in this study, follows DSM-IV and other clinical criteria (NINDS-AIREN, NINCDS-ADRDA) in that it hinges on non-transient impairment in two or more cognitive domains resulting in functional impairment (Supplement section 1). Cognitive assessment in ELSA is based on a set of standard and validated cognitive function tests; nonetheless the list is not comprehensive. Cognitive impairment in domains other than those tested may have been missed, leading to underestimation of dementia cases. Similarity of age-specific estimates of dementia incidence and prevalence with those of CFAS-II suggests this source of bias is small. The case definition for dementia used in this study required moderate to severe impairments in cognition and function to minimise false-positives. This results in inclusion of moderate to severe, rather than mild, cases of dementia. The main aim of this exercise is to inform future societal and healthcare needs. To this end, the dementia case definition is relevant to health and social policy as it forecasts numbers of individuals who would require supportive care due to moderate or severe cognitive and functional impairment. Although individuals residing in care homes were not included in ELSA, we took account of this group using the statistical joint-modelling approach. Further, data from carers and self-reported doctor diagnosis of dementia identified cases among those unable to take part in the study. 49

50 (ii) Seeking evidence of significant decline from a previous level of performance This aspect of the DSM-IV criteria for diagnosis of dementia is integral to our definition. An individual will only count toward incidence of dementia if they have participated in the ELSA study, are free from dementia, which means they do not show either cognitive impairment on the cognitive function tests, the IQCODE questionnaire, or functional impairment on the activities of daily living scale, but over time develop low scores on the cognitive function tests and develop functional impairment on the ADL scale. With this definition, a decline in performance from a previous level of functioning is central to the definition. The IQCODE questionnaire clearly asks how the subject performs compared to a previous level of functioning and thus establishes a decline from previous level of performance. In addition, for the DSM-IV criteria to hold, and to increase specificity, transient impairments in cognitive function or conducting ADLs, were not classified as cognitive or functional impairment. (iii) Learning effect We agree that learning effects could bias the trend estimate, and paid close attention to this issue in our analysis. We did not find considerable learning effects in consecutive cognitive function assessments now mentioned in the Methods section on Page 9. We did not find evidence of large learning effects in the cognitive function tests. The annual ageand sex-specific decline observed among participants who conducted the tests 4 or more years apart was similar to participants who conducted tests two years apart. (iv) Validation To validate the case definition of dementia, we compared the age and sex specific dementia incidence obtained from ELSA with that from the independent Cognitive Function and Ageing study (CFAS-II). The incidence of dementia in ELSA based on the above definition was similar to the age-specific values observed in CFAS-II. We conducted the joint model of longitudinal and time to event data to correct for the effect of selective dropout of study participants, and as expected, estimates were higher after correction (Supplement Figure 5 (below)). 50

51 Supplement Figure 5: Comparison of incidence (per 1000 person years) of dementia in English Longitudinal Study of Ageing (waves 4 to 6 ( )) with Cognitive Function and Ageing Study II ( ) CFAS is a well conducted study and suitable for comparing our estimates, but as always not without limitations. Participants in CFAS that were included in the analysis of dementia incidence and prevalence were selected from three areas (Cambridgeshire, Newcastle, and Nottingham), whereas ELSA participants are a representative sample of the English population as a whole. CFAS results were based on 367 dementia cases at CFAS I (86 incident cases) and 275 cases at CFAS II (173 incident cases). ELSA has both a larger sample size and a larger number of identified cases of dementia (1,448 cases of dementia in total of which 814 were incident cases among a total of 17,906 participants), resulting in greater statistical power. Variability of response rates in ELSA has been addressed during the data collection phase of the study by recruiting refreshment samples so that 51

52 every wave of data collection constitutes a representative sample and also by application of survey weights. Multiple imputation was used to correct for loss to follow up of study participants in estimating the calendar trend in dementia incidence in CFAS-I and II studies (Matthews FE et al. A two decade dementia incidence comparison from the Cognitive Function and Ageing Studies I and II. Nat Commun 2016; 7:11398). Multiple imputation methods are based on a missing-at-random assumption. Participants with dementia are more likely to drop out from the study (thus are missing-not-at-random) limiting the applicability of multiple imputation methods. The joint modelling statistical technique and corrections for loss to follow up applied in the present study require more than two waves of data collection and hence were not possible to apply to the CFAS-I and II studies. As such, findings in the present study are novel in that aspect. Without applying corrections for loss to follow up, results obtained from ELSA for age-specific dementia incidence are similar to those of CFAS-II. As expected, after correcting for the bias due to selective dropouts, we observe higher values for age specific dementia incidence (see Supplement Figure 5 above) and a steeper decline in incidence rates over time. Having six waves of data in ELSA we were able to show decline in dementia incidence in both men and women whilst in CFAS-I and II studies looking at calendar trends in dementia incidence the decline was only significant for men. POINT 2: The paper does not comply with the standards of reporting recommended in the STROBE Statement. OUR RESPONSE: Thank you. The STROBE statement was submitted with the paper last year in respect of the ELSA data which supports the IMPACT-BAM model. We have rechecked it and can confirm that it does comply with the STROBE guidelines. Further, the paper cites the cohort profile published in: Ref 27: Steptoe A, Breeze E, Banks J, Nazroo J. Cohort profile: the English longitudinal study of ageing. Int J Epidemiol 2013; 42(6): Ref 28: Mindell J, Biddulph JP, Hirani V, Stamatakis E, Craig R, Nunn S et al. Cohort profile: the health survey for England. Int J Epidemiol 2012; 41(6):

53 POINT 3: Descriptive data of dementia prevalence are reported as supplementary materials only; and the statistical methods used to model the expected projections of dementia prevalence, which do not increase the validity of the diagnosis itself, seem unnecessarily complicated (probably more appropriate for a set of sensitivity analyses). These techniques apparently relied on a number of assumptions that may not hold, including the posited 2.7% annual reduction in dementia incidence, which is somewhat unsupported (below). In addition, the internal validation of this statistical method is likely prone to circularity, and it is confusing (below). OUR RESPONSE: We thank the Reviewer for highlighting several uncertainties about our modelling method. We provide a point by point response below: (i) Descriptive data of dementia prevalence We have added the following information on dementia prevalence to the Results section of the paper (page 15) as below: Between 2002 and 2013 dementia was ascertained for a total of 1,448 ELSA participants. 634 participants had dementia at time of recruitment and 814 incident cases were identified over the course of the follow up. 16.5% of cases occurred in participants below age 65. Of the 1,448 cases of dementia, 466 received a doctor diagnosis of dementia, 245 were identified via IQCODE questionnaire plus ADL functional impairment, and 1,078 were identified by impairment in cognitive tests and impairment in one or more ADLs. Of the 245 cases based on IQCODE and functional impairment, 171 (70%) reported a doctor diagnosis of dementia. Among those with impairment in cognitive tests and ADL function, 263 (25%) reported a doctor diagnosis of dementia at later stages or were positive on the IQCODE questionnaire. The number of incident dementia cases ascertained at each data collection wave is shown in Supplement-Table 3. Age- and sex-specific prevalence of dementia at each data collection wave of ELSA provides the starting values for the IMPACT-BAM model (Supplement Table 5) (ii) Complexity of methods for projection of dementia prevalence We agree the model may appear complicated, however, the method is based on simple epidemiological concepts governing the relationships between incidence, prevalence, and survival. Dementia prevalence is a function of dementia incidence and survival. Modelling trends in dementia incidence and mortality rates, whilst taking appropriate account of competing risks of cardiovascular disease and deaths, required probabilistic Markov modelling with the appropriate set of health states. This method provides an improvement to simple projections of dementia prevalence. Such projections only account for the effect of population ageing rather than for changes in dementia incidence that 53

54 also incorporate trends in cardiovascular disease and cardiovascular and non-cardiovascular mortality. We have elaborated on this point further in our response to POINTs 6 and 7 below. (iii) Model assumptions The model assumptions and evidence supporting these assumptions are collected together in Table 1 (shown at the end of the response to additional comments from Editorial committee on Page 26 of this document). We hope this Table helps clarify the strength of the evidence and assumptions underlying our predictions. With regards to the assumption related to the calendar decline in dementia incidence, the 2.7% annual decline in incidence was estimated from data collected over the 6 waves of ELSA using an elaborate model that takes into account losses to follow-up and mortality, both of which are dependent upon participants current cognitive function (please see our response to POINT 8 of the comments from the Editor s committee on Page 20 of this document). To address uncertainty in the 2.7% calendar effect, three alternative scenarios based on calendar effects observed in previous studies have been modelled in sensitivity analyses (Figure 6, see below). Two are based on the magnitude of decline in dementia incidence suggested by the CFAS and Framingham Heart study. A further scenario assumes that future dementia incidence remains constant. 54

55 Figure 6: Sensitivity analysis for number of cases of dementia under alternative assumptions for calendar trend in incidence of dementia Dashed lines represent 95% uncertainty intervals. (iv) Validation of the method To assess the validity of our method we compared model predictions to data from independent sources. We used ELSA data at year 2006 to inform the model and predict dementia prevalence in England and Wales in Predicted prevalence of dementia was compatible with dementia prevalence observed in CFAS-II in 2011 (Supplement Figure 10). Being able to accurately predict observed values from an independent study underpins the valid functionality of the IMPACT- BAM model. The model functionality was also tested by comparing prevalence of cardiovascular disease and mortality rates with those observed in independent studies (Supplement Figures 12 and 13). 55

56 POINT 4: The research question of the present study is not clearly stated. It is not clear whether the authors set out to explore whether potential differential attrition may explain the variability in dementia incidence over one decade in the ELSA, or to present descriptive data of a potential competing scenario, with the current one supported by Alzheimer s Disease International and the UK Alzheimer Society, regarding trends of dementia prevalence based on that evidence to inform policy makers and relevant stakeholders. OUR RESPONSE: We agree that the aims of the study could have been made clearer. The aim of the study is to accurately predict the number of people living with dementia in England and Wales up to Sufficient external evidence exists with regard to the calendar trends in cardiovascular incidence and mortality rates that were observed over the past two decades. Although a decline in dementia incidence over the past two decades has been shown in several studies, the magnitude of the calendar trend in incidence of dementia is uncertain and has not been estimated with the degree of accuracy required for the modelling strategy in this paper. The considerable effect of the calendar trend in dementia incidence on projections for dementia prevalence is shown in Figure 6 above. Therefore in this study we aimed to: 1. Obtain an accurate, unbiased estimate of the calendar trend in incidence of dementia using data from ELSA. This required models that control for the bias due to differential loss to follow-up in the study, as outlined in our response to POINT 3 above. 2. Use this corrected estimate of calendar trend in dementia incidence together with the recent dynamics of key health states including declines in cardiovascular disease, cardiovascular and noncardiovascular mortality to predict future prevalence of dementia in England and Wales to The end goal of this exercise is to inform policy: Accurate projections for burden of dementia is a key step for planning to meet future needs. (Introduction Page 5) We have re-written the last paragraph of the introduction (page 6) to present the aims of the study more explicitly: In the present study, we developed a novel probabilistic Markov method (labelled IMPACT-Better Ageing Model) to simultaneously model the transitions of the population through states of health, cardiovascular disease, cognitive and functional impairment, dementia, to death, to obtain projections for the prevalence of dementia in England and Wales up to To account for the effect of selective dropout of study participants and to obtain unbiased estimates for dementia incidence over the past 15 years, with which to inform the model, we applied a robust joint modelling technique 26 to data from the English Longitudinal Study of Ageing 56

57 POINT 5: Title The title is unfocused and somewhat misleading and does not reflect the content of the paper, rather it unnecessarily overemphasizes the statistical model used (presumably) to calculate dementia trends, with no specifications on whether these trends refer to incidence or prevalence or both. Several implicit claims in the title are inaccurate or confusing. What does recent mean in this context, recent birth cohorts or findings? Further, because this study cannot count on a validated dementia diagnosis, the authors should refrain from making such claim in the title (see comments above). Finally, the use of an acronym (i.e. IMPACT), and the fact that it refers to a statistical technique should be discouraged. OUR RESPONSE: Thank you for pointing out shortcomings in the Title of the paper. We have now revised the title as below: Temporal trend in dementia incidence since 2002 and projections for prevalence in England and Wales to 2040: a modelling study POINT 6: Introduction Prevalence depends on (and is actually proportional to) incidence and disease duration, and in the case of dementia the latter equals survival with the disease. This important consideration is absent in the introduction, is not considered in the methods, nor are its implications adequately considered in the Discussion and to interpret and contextualize the study s main findings, that is the future numbers of those affected by dementia. It is obvious that the numbers presented in the paper would be greatly affected by any proportional increment in survival in people with dementia compared to cognitively healthy counterparts in the UK population. This is key for policy makers, because improvements in care and more timely diagnosis of dementia are highly recommended [See the Lancet Neurology Commission: Defeating Alzheimer s Disease and Other Dementias, by Winblad et al.], which would presumably lead to significant longer survival in people with dementia. OUR RESPONSE: We thank the reviewer for these important comments and acknowledge that our description of the IMPACT-BAM model could be clearer. We have now added the assumptions related to survival in the Table of assumptions (Table 1, Assumption 13, also assumptions 5 and 7). Survival with dementia and changes in survival over time are a function of incidence and mortality thus have been carefully accounted for in the model by applying the appropriate incidence and mortality rates. Survival with dementia equals the number of years the individual remains in that 57

58 health state from entry to death. Since changes in incidence and mortality rates are incorporated in the model, changes in survival over time are also accounted for. Following your comment we have added the limitations of the method with regards to survival to the discussion (Page 20) as below: Survival with cardiovascular disease and dementia has improved over time Changes in lifeexpectancy and survival with cardiovascular disease, functional impairment, or dementia are accounted for in the model for by applying calendar-time specific mortality rates. This method is based on the assumption that survival with health conditions change in proportion to changes in overall life-expectancy. This assumption, although reasonable and commonly applied in modelling studies, lacks verification, as evidence on the exact age and sex specific survival with dementia is rare, 5 and hard to obtain given uncertainty over date of onset. The reviewer is correct in pointing out that the numbers of dementia cases could greatly increase if survival rates increased only in those with dementia. This beneficial change in survival among individuals with dementia would imply a reduction in the odds ratio for mortality of this group compared to other healthy parts of the UK population. As elaborated in the paper by Prince M et al. Alzheimer s Research & Therapy 2016, and mentioned above evidence on the exact age and sex specific survival with dementia is rare, and hard to obtain given uncertainty over date of onset. Therefore, our model assumes that the odds ratio for mortality comparing individuals with dementia compared to the general population remains fixed over the period to 2040 (pointed out in the Table of assumptions) which means survival with dementia improves in parallel to survival in the general population. We have now highlighted this point in the methods (Page 13): Health state specific transition probabilities to cardiovascular and non-cardiovascular mortality were calculated by estimating the odds ratios of death from each health state using ELSA data (Supplement Table 1) and using the odds ratios to obtain the probability of cardiovascular and noncardiovascular death from each health state compared to that in the general population. The probability ratios were applied to overall cardiovascular and non-cardiovascular mortality rates shown in Supplement Figure 3 to obtain state specific cardiovascular and non-cardiovascular mortality rates. Mortality rates from each health state were assumed to change in parallel with overall cardiovascular and non-cardiovascular mortality rates. 58

59 POINT 7: The criticisms made to the current, and widely accepted, projections of dementia prevalence for the next decades is flawed. These projections are not inaccurate, they are based on demographic projections of changes in age structures due to population aging in all world regions, and on the assumption that age-specific prevalence of dementia will not change. There is insufficient evidence to support the use of varying age-standardised risks over time to calculate the future trends of dementia prevalence in the population (see Prince M et al. Alzheimer s Research & Therapy 2016). The authors present an alternative scenario based on a 2.7% annual reduction in dementia incidence. As said, this should be presented as a possible scenario, because it is based on one possible model specification of the ELSA data only. For instance, in the MRC-CFAS study the incidence rates of dementia did not significantly change in men between CFAS I and II [Matthews F et al. 2016]. OUR RESPONSE: We thank the reviewer for raising this point. Evidence on trends in the agespecific prevalence of dementia as reported in the paper by Prince et al, were inconsistent across studies and world regions. In the present study we found that age-standardised prevalence of dementia is declining over time (Supplement Figure 9, shown on page 68 of this document). This finding is consistent with other reports in the UK, including the CFAS I and II studies which showed dementia prevalence declined over two decades (Matthews FE et al. A two-decade comparison of prevalence of dementia in individuals aged 65 years and older from three geographical areas of England: results of the Cognitive Function and Ageing Study I and II. Lancet 2013; 382(9902): ). Dementia prevalence is largely determined by incidence and survival, both of which are changing over time in the UK. Thus, a prediction model accounting for trends in incidence and mortality is likely to yield more accurate estimates of prevalence of dementia in the future compared to models that assume a constant prevalence. The non-linearity of our predictions of dementia prevalence underpins this point. As mentioned in our response to POINT 3 above, the 2.7% relative decline in incidence of dementia is presented as one scenario which the authors believe, based on the scientific evidence, is the most likely scenario. The results based on other scenarios suggested by the reviewer are also presented in the paper (Figure 6, shown above). POINT 8: Further, even assuming that dementia incidence is truly declining in Western countries, the reasons and explanations of these secular trends are not clear. Thus, the argument of the potential competing effect of cardiovascular risk is speculative, and unsubstantiated. There is no evidence that changes in vascular risk profiles explained the observed reductions in dementia incidence in the past 59

60 decades in both the Framingham and the MRC-CFAS studies (women only) (see Satizabal et al. NEJM 2016 and Matthews F et al. Nature Comm. 2016). OUR RESPONSE: Thank you. We agree that the reasons for a decline in dementia incidence are not clear and more research is needed. Cardiovascular mortality has been declining steadily over the past two decades (Supplement Figure 3, shown below). As cardiovascular mortality declines and lifeexpectancy increases, there will be more people at older ages at risk of dementia, naturally resulting in an increase in the overall prevalence of dementia in the population. On the other hand, dementia incidence is observed to be declining in several studies, which results in a decline in prevalence. To be able to accurately predict the prevalence and number of cases of dementia in the future these two opposing effects need to be taken into account. The evidence on both these competing effects is rather strong. The figures below show changes in age-specific cardiovascular mortality rates over the past two decades based on data from the UK Office for National Statistics. The effect of competing risks of cardiovascular disease on future projections for dementia prevalence will become especially prominent in the next stage of the IMPACT-BAM project where we model the effect of vascular risk reduction on future projections for numbers of people living with dementia. Vascular risk reduction will have a considerable effect on cardiovascular disease, and thus lifeexpectancy. Therefore, cardiovascular disease needs to be explicitly modelled alongside dementia to obtain accurate estimates for the effects of vascular risk reduction scenarios. The modelling strategy in the present study follows prevailing research recommendations in the UK. The UK Government has now endorsed the Blackfriars Consensus statement by leading UK health charities, Public Health England and NICE, plus WHO. They all advocate prevention strategies which address the shared determinants of non-communicable diseases. Thus by reducing CVD burden, the dementia burden might also be reduced (UK ministers sign Blackfriars Consensus Statement on promoting brain health and reducing risks for dementia in the population as part of Dementia Awareness Week. For evidence on declining dementia incidence over time we reference the below papers: 1. Wu YT, Fratiglioni L, Matthews FE, Lobo A, Breteler MM, Skoog I et al. Dementia in western Europe: epidemiological evidence and implications for policy making. Lancet Neurol 2015;(15):

61 2. Matthews FE, Stephan BC, Robinson L, Jagger C, Barnes LE, Arthur A et al. A two decade dementia incidence comparison from the Cognitive Function and Ageing Studies I and II. Nat Commun 2016; 7: doi: /ncomms11398.: Satizabal CL, Beiser AS, Chouraki V, Chene G, Dufouil C, Seshadri S. Incidence of Dementia over Three Decades in the Framingham Heart Study. N Engl J Med 2016; 374(6): Schrijvers EM, Verhaaren BF, Koudstaal PJ, Hofman A, Ikram MA, Breteler MM. Is dementia incidence declining?: Trends in dementia incidence since 1990 in the Rotterdam Study. Neurology 2012; 78(19):

62 Following your suggestion we have elaborated on the vascular hypothesis for dementia in the discussion (Page 21) as below: There are several plausible explanations supporting a decline in dementia incidence over time. Improvement in vascular risk factors, as well as in education levels, can partly account for the decline in incidence. In the present study, increased physical activity accounted for the largest proportion of the decline in dementia incidence between 2002 and Changes in prevalence of diabetes, smoking, and social class over time had negative confounding effects, such that the downward incidence trend increased after controlling for respective changes. Adjustment for stroke and depression did not have a significant effect on the calendar trend in dementia. Basic research efforts to understand the causes of dementia have increased markedly in recent years, but to date drug trials have failed to show modification of disease processes. 54 The WHO and other expert bodies have identified prevention, identification and reduction of risk as a top research priority, in response to the lack of therapy Debate continues about the relative importance of vascular and neurodegenerative causes. In the context of this uncertainty the IMPACT-BAM model is a means to understand how the dementia burden will evolve, and provides a platform to measure how the burden might be reduced through various policy interventions. Please also see our response to POINT 11 of the additional comments by the Editorial committee on the same topic. POINT 9: The final paragraph of the introduction is unfocused. Attrition is not defined, nor is it explained how it may impact trends in dementia incidence and/or prevalence. OUR RESPONSE: Thank you. We agree there were shortcomings in the introduction. We have now replaced the word attrition by dropout, and re-written the final paragraph of the introduction (Page 6) to address your valid comments, as below: In the present study, we developed a novel probabilistic Markov method (labelled IMPACT-Better Ageing Model) to simultaneously model the transitions of the population through states of health, cardiovascular disease, cognitive and functional impairment, dementia, to death, to obtain projections for the prevalence of dementia in England and Wales up to To account for the effect of selective dropout of study participants and to obtain unbiased estimates for dementia incidence over the past 15 years, with which to inform the model, we applied a robust joint modelling technique 26 to data from the English Longitudinal Study of Ageing. 62

63 The effect of dropout of study participants from the study on estimates for dementia incidence is shown in Supplement Figures 7 and 8. POINT 10: The relevant literature is not appropriately cited, there is a continuing confusion between studies (and evidence) of dementia prevalence and incidence trends. A review of European studies is cited along with the original papers, this is redundant. Also, more recent and comprehensive reviews are not cited (i.e. Prince M et al. 2016). References 1 to 6 are mixed, prevalence and incidence studies are erroneously combined. Reference number 6 is redundant. Reference number 17 is not pertinent. References number 12 is about modelling the attributable risks of dementia to a selection of risk factors, not pertinent. OUR RESPONSE: Thank you for pointing out these issues. The list of references used has now been updated and the referencing corrected throughout the manuscript. POINT 11: Methods The major concern is about the validity of the ascertainment of dementia diagnosis. As said the validity of the diagnostic procedure was not previously demonstrated, cognitive assessment is very limited and learning effects very likely. OUR RESPONSE: Thank you. We have addressed these concerns in our earlier response to POINT 1 above. POINT 12: The description of the IMPACT-BAM models should be presented in the statistical methods section. In addition, the validation of this model is problematic, as it assumes adequate comparability between the ELSA and the MRC-CFAS studies, which have remarkably different designs, including sampling and outcome ascertainment. OUR RESPONSE: Thank you for your suggestion. As suggested, the description of IMPACT-BAM model has now been moved to the statistical analysis section (Pages 11-14). We agree that ELSA and 63

64 MRC-CFAS employ very different study designs, but both are based in the older population, and both may reasonably be used to estimate prevalence of dementia in England and Wales. The independent MRC-CFAS is the best study available in the UK for comparison of dementia incidence and prevalence estimates. Similarity of age-specific incidence rates and prevalence of dementia obtained from the two studies despite differences in the sampling design and method of outcome ascertainment is reassuring and indicates that the methods used are indeed robust and valid. POINT 13: I am not a statistician, but the modelling of changes of dementia incidence across the waves of the study does not look adequate, and the results expressed in the form of HR are difficult to interpret. OUR RESPONSE: Thank you. We acknowledge that the modelling of changes in dementia incidence across the waves of ELSA is not straightforward. We have added detail in the statistical methods of the Methods section (Pages 10-11) and outlined the process here: To obtain unbiased estimates of the calendar trends in incidence of dementia in ELSA, it was necessary to account for the competing risks of mortality and bias due participants lost to follow up from the study. Change in dementia incidence across waves of ELSA was modelled in three stages as explained in the paper: 1. First, a simple Cox-proportional hazards model with incident dementia as the outcome and terms for age, age-squared, sex, and calendar time was fitted. The coefficient for calendar time obtained from this model is the age- and sex-adjusted hazard ratio of dementia in each year compared to the year before. The HR of calendar year from this method was which means (dementia risk in year n+1 / dementia risk in year n) = However, this model does not account for the competing effects of death or loss to follow up (both of which are also related to the presence of dementia) and the result obtained for the magnitude of the calendar effect in incidence of dementia is similar to that observed in previous studies (CFAS-I and II) which have applied a similar method to obtain calendar trends in dementia incidence. 2. In the second stage we fitted a competing risks model to account for the competing risk of mortality on the calendar trend in dementia incidence. The estimate for the calendar trend from this model was HR 0.973, and is interpreted as above. 3. In the third stage, to account for non-random loss to follow up as well as competing effect of mortality, we used a regression analysis with the log-odds of dementia incidence obtained from the 64

65 joint modelling of longitudinal and time to event data to correct for the bias due to loss to follow-up as the outcome, and terms for age, age-squared, sex, and calendar time. The coefficient for calendar time obtained from this method is corrected for the effect of mortality and loss to follow up. The coefficient obtained from this analysis, similar to above methods, represents the dementia risk in year n+1 / dementia risk in year n. The estimate of the calendar trend was similar to the competing risks method (OR: 0.973), but was statistically significant. Since incidence of dementia is low (<10%) both the OR and HR approximate the risk ratio of dementia in year n compared to year n-1 adjusted for age and sex (the magnitude of the calendar trend). We hope that the description of these methods added to the paper is now clear (Statistical methods Pages 10-11) POINT 14: Results Numbers and reasons for non-participation at each stage are not reported, neither is there a flowdiagram of the analytic sample derivation. OUR RESPONSE: Thank you for your suggestion. A flow diagram showing the numbers of participants, those who died or were lost to follow up is shown in Supplement Figure 1 (below): 65

66 Supplement Figure 1: Flow-diagram of participants recruited to the English Longitudinal Study of Ageing ( ). POINT 15: The characteristics of study participants should be presented in a comprehensive table and salient aspects reported in the text in the main paper not as supplementary materials. Differences in socio-demographic and health characteristics between wave 1 and 4 should be formally tested and reported using appropriate statistics. OUR RESPONSE: Thank you. Following the reviewer s suggestion we have provided the baseline characteristics of study participants at each wave of the ELSA study and reported the P-value for trends in the characteristics over time (Supplement Table 2, shown below). 66

67 Supplement Table 2: Baseline Characteristics of English Longitudinal Study of Ageing participants Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 Wave 6 P for temporal trend ( ) ( ) ( ) ( ) ( ) ( ) N=12,085 N=9,424 N=9,763 N=10,942 N=10,142 N=10,594 Age 64.2 (11.0) 65.8 (10.6) 64.5 (11.3) 65.2 (10.3) 66.8 (10.0) 66.6 (10.3) Male ( %) 5,332 (44 %) 4,124 (44 %) 4,293 (44 %) 4,862 (44 %) 4,490 (44 %) 4,742 (45 %) Education <0.001 No qualification 6,034 (50.0 %) 4,414 (46.9 %) 3,630 (37.3 %) 4,154 (38.1 %) 3,918 (38.8 %) 3,910 (38.1 %) A level / O level / equivalent 3,315 (27.5 %) 2,706 (28.8 %) 3,017 (31.0 %) 3,379 (31.0 %) 3,096 (30.7 %) 3,224 (31.4 %) University / Higher 2,716 (22.5 %) 2,290 (24.3 %) 3,075 (31.6 %) 3,366 (30.1 %) 3,078 (30.5 %) 3,126 (30.5 %) Social Class <0.001 I/II 2,719 (23.6 %) 1,957 (21.7 %) 1,940 (20.8 %) 2,110 (20.2 %) 1,883 (19.5 %) 1,727 (19.4 %) III-M/III-NM 5,109 (44.3 %) 3,971 (44.1 %) 4,017 (43.1 %) 4,348 (41.7 %) 4,017 (41.7 %) 3,669 (41.3 %) IV/V 3,719 (32.2 %) 3,082 (34.2 %) 3,356 (36.0 %) 3,967 (38.1 %) 3,736 (38.8 %) 3,496 (39.3 %) Current Smoker (%) 2,159 (18.2 %) 1,472 (15.6 %) 1,507 (15.4 %) 1,548 (14.2 %) 1,315 (13.0 %) 1,199 (11.3 %) <0.001 Daily alcohol intake (%) 3,318 (27.9 %) 3,185 (33.8 %) 3,204 (32.8 %) 3,726 (35.9 %) 3,466 (34.2 %) 3,516 (33.2 %) <0.001 Sedentary or low Physical activity (%) 2,789 (33.2 %) 2,573 (27.3 %) 2,947 (30.2 %) 3,374 (30.9 %) 3,219 (31.7 %) 2,756 (26.0 %) <0.001 Body Mass Index (Kg/m2) (4.9) (5.3) (5.1) <0.001 Systolic Blood Pressure (mmhg) (18.9) (17.8) (17.5) <0.001 Diastolic Blood Pressure (mmhg) (11.2) (11.0) (10.8) <0.001 LDL Cholesterol (mmol/l) (1.00) (1.03) (1.04) <0.001 HDL Cholesterol (mmol/l) (0.39) (0.4) (0.48) <0.001 Diabetes (%) 866 (7.2 %) 809 (8.6 %) 917 (9.4 %) 1,086 (9.9 %) 1,186 (11.7 %) 1,285 (12.1 %) <0.001 Cardiovascular Disease (%) 1,804 (14.9 %) 1,565 (16.6 %) 1,489 (15.3 %) 1,526 (14.0 %) 1,579 (15.6 %) 1,556 (14.7 %) <0.001 Cerebrovascular Disease/Stroke (%) 516 (4.3 %) 475 (5.0 %) 474 (4.9 %) 506 (4.6 %) 517 (5.1 %) 537 (5.1 %) 0.34 Values are mean (standard deviation) or number (%) 67

68 POINT 16: Dementia prevalence (not rates) should be reported providing the 95% CI intervals by gender and across the age groups; crude and age-, sex- and education-standardised prevalences could also be presented as the total for each wave of the study, to allow comparisons. OUR RESPONSE: Thank you for your advice. The estimates for dementia prevalence are educationstandardised because the cut-off points to determine cognitive impairment for each of the cognitive function tests were specific for level of education. Prevalence of dementia by age and sex at each wave of ELSA and the 95% confidence intervals are now presented in Supplement Table 5. Age, sex standardised values are presented in Supplement Figure 9, both are displayed below. Supplement Figure 9: Age- and sex-standardised prevalence of dementia at each data collection wave of the English Longitudinal Study of Ageing (ELSA)