Statistical methods for knowledge discovery in adverse drug reaction surveillance

Size: px
Start display at page:

Download "Statistical methods for knowledge discovery in adverse drug reaction surveillance"

Transcription

1 Statistical methods for knowledge discovery in adverse drug reaction surveillance

2

3 Statistical methods for knowledge discovery in adverse drug reaction surveillance G. Niklas Norén Stockholm University

4 c G. Niklas Norén, Stockholm 2007 Cover photography by G. Niklas Norén ISBN pp Typeset by L A T E X Printed in Sweden by Universitetsservice AB, Stockholm 2007 Distributor: Department of Mathematics, Stockholm University

5 Abstract Collections of individual case safety reports are the main resource for early discovery of unknown adverse reactions to drugs once they have been introduced to the general public. The data sets involved are complex and based on voluntary submission of reports, but contain pieces of very important information. The aim of this thesis is to propose computationally feasible statistical methods for large-scale knowledge discovery in these data sets. The main contributions are a duplicate detection method that can reliably identify pairs of unexpectedly similar reports and a new measure for highlighting suspected drug drug interaction. Specifically, we extend the hit-miss model for database record matching with a hit-miss mixture model for scoring numerical record fields and a new method to compensate for strong record field correlations. The extended hit-miss model is implemented for the WHO database and demonstrated to be useful in real world duplicate detection, despite the noisy and incomplete information on individual case safety reports. The Information Component measure of disproportionality has been in routine use since 1998 to screen the WHO database for excessive adverse drug reaction reporting rates. Here, it is further refined. We introduce improved credibility intervals for rare events, post-stratification adjustment for suspected confounders and an extension to higher order associations that allows for simple but robust screening for potential risk factors. A new approach to identifying reporting patterns indicative of drug drug interaction is also proposed. Finally, we describe how imprecision estimates specific to each prediction of a Bayes classifier may be obtained with the Bayesian bootstrap. Such case-based imprecision estimates allow for better prediction when different types of errors have different associated loss, with a possible application in combining quantitative and clinical filters to highlight drug ADR pairs for clinical review.

6

7 List of Papers This thesis is based on the following original publications, which are referred to in the text by their Roman numerals. I II Norén, G. N., Orre, R., Bate, A., Edwards, I. R. (2007). Duplicate detection in adverse drug reaction surveillance. Data Mining and Knowledge Discovery. Published on-line. Norén, G. N., Bate, A., Orre, R., Edwards, I. R. (2006). Extending the methods used to screen the WHO drug safety database towards analysis of complex associations and improved accuracy for rare events. Statistics in Medicine, 25(21): III Hopstadius, J, Norén, G. N., Bate, A., Edwards, I. R. (2007). Adjustment for potential confounders in adverse drug reaction surveillance. Submitted for publication. IV V Norén, G. N., Sundberg, R., Bate, A., Edwards, I. R. (2007). A statistical methodology for drug drug interaction surveillance. Submitted for publication. Norén, G. N., Orre, R. (2005). Case based imprecision estimates for Bayes classifiers with the Bayesian bootstrap. Machine Learning, 58(1): Reprints of I, II and V were made with kind permission from the publishers.

8

9 Contents Part I: Thesis summary 1 Introduction Aim Adverse drug reaction surveillance Individual case safety reports The WHO database Adverse drug reaction signal detection Knowledge discovery in adverse drug reaction surveillance Context Process Disproportionality Shrinkage Pattern discovery and detection Facilitating interpretation Future directions Overview of the papers Paper I Paper II Paper III Paper IV Paper V Acknowledgements Bibliography

10

11 Part I: Thesis summary

12

13 1. Introduction It is in the nature of pharmaceutical development that the full safety profile of a new medicinal product will not be known at the time it is introduced to the general public. Because randomised clinical trials are limited in both the types and numbers of patients exposed, continued safety monitoring of drugs is in the interest of patients, regulatory authorities and pharmaceutical companies (Finney 1966, Evans 2000). Individual case safety reports are submitted by health professionals based on suspected adverse drug reaction (ADR) incidents (Edwards and Aronson 2000) observed in real world clinical practice. They remain one of the best resources for early post-marketing discovery of potential public health or patient safety issues. They are rich sources of information, but anecdotal in nature. The reliance on voluntary submission, the variation in quality of information and the large number of new reports submitted to national and international organisations every year provide a range of interesting statistical challenges. 1.1 Aim The overall aim of this thesis is to propose improved statistical methods for knowledge discovery in collections of individual case safety reports. I proposes a new method for automated duplicate detection based on the hit-miss model introduced for statistical record linkage (matching records across data sets) by Copas and Hilton (1990). An extended hit-miss model that handles numerical record fields and compensates for correlations between record fields is implemented for the WHO database and demonstrated to be useful in real world duplicate detection. II proposes improved credibility intervals, a poststratification approach to adjustment for confounding variables and an extension to higher order associations for the Information Component (IC) measure of disproportionality used to screen the WHO database for excessive ADR relative reporting rates. III demonstrates that the post-stratification adjustment of the observed-to-expected ratio for suspected confounders adopted for the IC in II may lead to spurious underestimation in the presence of any very small strata in a stratified data set. A comparison to a literature reference indicates that while routine adjustment for some potential confounders in first pass screening of collections of individual case safety reports does improve 1

14 performance, the magnitude of this improvement is modest compared to the improvement from a triage (prioritisation) criterion requiring reports from at least two countries before a drug ADR pair is highlighted for clinical review. This suggests that confounding may have less impact on the analysis of individual case safety reports than previously believed. IV introduces a new measure of drug drug interaction for collections of individual case safety reports. Unlike methods proposed previously for this purpose, it defines interaction as departure from a baseline model with independent attributable risk. V introduces a Bayesian bootstrap method for estimating the uncertainty in Bayes classification associated with each individual prediction. We demonstrate how this information can be used to improve performance, when different types of errors have different associated loss, with a possible application in selecting drug ADR pairs for detailed clinical review. 2

15 2. Adverse drug reaction surveillance The analysis of individual case safety reports is the cornerstone of early postmarketing ADR detection (Rawlins 1988). Whereas large, formal drug safety studies are useful to test specific hypotheses related to drug safety, they are not suitable for continuous monitoring with the aim of detecting previously unsuspected ADRs, as early as possible. In the context of this PhD thesis, ADR surveillance refers exclusively to drug safety monitoring based on individual case safety reports. It thus excludes other post-marketing efforts such as the intensive monitoring programs of New Zealand and the United Kingdom, as well as safety monitoring based on health registries and hospital-based safety monitoring. For comprehensive overviews of post-marketing ADR surveillance, see Lindquist (2003) and Bate (2003). 2.1 Individual case safety reports Individual case safety reports communicate genuine clinical concerns from observant health professionals (Edwards 1999). As they are based on actual patients in real world clinical practice, their collection and analysis increase the chance to discover ADRs that are due to drug drug interaction, affect patients with certain medical predispositions or that belong to patient subgroups that tend to be excluded from pre-marketing clinical trials, such as children or pregnant women. In addition, the large numbers of patients exposed and the unlimited follow-up time available considerably increase the chance to detect ADRs that are rare or that occur only after extended periods of use. An example of an authentic individual case safety report is provided in Figure 2.1. Much of the information on these reports can be originally provided as free text, some of which is later encoded as structured information upon database entry. This is usually done by trained personnel at pharmaceutical companies or at national authorities. The encoding of observed ADR incidents in terms of standardised terminology is a critical part of the preprocessing. One potential pitfall is the risk of misinterpretation when the ADR encoding is performed by someone who has never actually met the patient. Variation in coding across regions and time periods may lead to systematic differences that can affect subsequent data analysis. A general problem is that several ADR terms are often applicable to a given incident. Thus, exploratory analysis 3

16 Figure 2.1: Sample individual case safety report. Reprinted with kind permission of the Adverse Drug Reactions Unit at the Therapeutic Goods Administration of Australia 4

17 focusing on single ADR terms may fail to include all relevant reports a phenomenon which has been referred to as signal fragmentation (Purcell 2003). While in the follow-up of specific issues, this can be remedied by specifying groups of relevant ADR terms for the issue of interest, it is not obvious how such strategies can be easily automated for routine exploratory analysis. Individual case safety reports refer to suspected ADR incidents and some adverse events observed in association with drug prescription will in reality be coincidental, due to concomitant medication or natural progression of the underlying disease. At the same time, not all ADR incidents that actually occur are identified as such and eventually reported to the national drug safety centres. The degree of under-reporting is unknown but can be expected to vary with the severity of the suspected ADR, across geographical regions and time periods. There may also be variation in the propensity to report suspected ADRs during the life-span of a drug and in response to any attention to suspected drug safety issues in the public or scientific media. The categories of health professionals who are allowed to submit reports also differ over time and between regions. Some countries allow only medical doctors to submit reports, whereas others accept reports from medical nurses and pharmacists as well. In addition, some countries encourage direct consumer reporting. Unsurprisingly, the propensity to report suspected ADRs of different types varies considerably between different categories of reporters (Savage 1985). An important characteristic of individual case safety report submission is that separate reports sometimes have a common origin and therefore cannot be considered as independent pieces of information (Finney 1973). This may distort automated knowledge discovery and mislead clinical review. The most obvious cause of non-independent reports is report duplication, where a single suspected ADR incident results in several reports. This phenomenon is discussed at some length in I. More subtle examples include groups of reports provided by the same health professional, such as those from the Norwegian dentist discovered in I, reports from the same clinical study (sometimes mislabelled as spontaneous reports) or separate reports for the same patient at different points in time. If single individuals are responsible for encoding large numbers of reports, this may also induce superficial similarity between reports. A potential example of this is the group of over 600 very similar reports originally collected by a single law firm, discovered in IV. Violated independence assumptions differ from other data quality issues in that they do not relate to the quality of single reports, but to the quality of collections of reports. Even upon the confirmation that a pair of reports are indeed duplicates it is not obvious how to proceed: should the suspected duplicates be flagged or should one of them perhaps be removed from the data set (if so, which one)? 5

18 Drugs Anatomical Class Therapeutic Selective serotonin reuptake inhibitors (SSRI) Antiinfl. prep. nonsteroids for topical use (NSAID) Platelet aggregation inhibitors excl. heparin ADRs Reports System Organ Class Reports 193,939 Body as a whole - general disorders 180,770 Skin and appendages disorders 179,226 Gastro-intestinal system disorders ACE inhibitors, plain 171,706 Central & peripheral nervous system disorders deriva- Benzodiazepine tives 1,218,425 1,070, , , ,898 Psychiatric disorders 677,227 Table 2.1: The most commonly reported groups of drugs and ADRs in the WHO database (note that each report may list more than one drug and more than one ADR). 2.2 The WHO database The Uppsala Monitoring Centre maintains and analyses the world s largest collection of individual case safety reports. As of December 2006, the WHO database contained over 3.8 million reports, with a current yearly growth of over 200,000 reports (see Figure 2.2). The database is held on behalf of the countries participating in the WHO Programme for International Drug Monitoring, whose number has continued to grow from the founding 10 countries in 1968 to over 80 member countries at the end of The international coverage allows rare but important public health or patient safety issues to be detected earlier after drug launch than if based on isolated analysis of national data sets (Olsson 1998). Variation between countries in the range of available drug substances, populations at risk, reporting culture and regulation may influence relative reporting rates and make knowledge discovery more complicated in international data sets. At the same time, this diversity is an invaluable asset in detecting public health or patient safety issues related to for example ethnic or dietary ADR risk factors. Thus, even though most reports in the WHO database come from the USA and other industrialised nations, the worldwide coverage of the WHO programme is perhaps its greatest strength. As is clear from Figure 2.2, the vast majority of reports in the WHO database are so-called spontaneous reports that refer to observations in regular clinical practice. However, a small minority are from intensive monitoring programs or clinical studies. Such atypical reports should in principle be labelled as 6

19 Number of reports 4,000,000 3,000,000 2,000,000 1,000, Year a. Database growth United States United Kingdom Germany Canada France Australia Other countries 20% 5% 5% Spontaneous reports Other 5% 47% 12% 6% 5% 95% b. Biggest contributors c. Types of reports 4,000,000 2,000,000 Number of reports 3,000,000 2,000,000 1,000,000 Number of reports 1,500,000 1,000, , Number of drugs per report Number of ADRs per report d. Number of drugs per report e. Number of ADRs per report Number of reports 50,000 40,000 30,000 20,000 10, Patient age (years) f. Patient age distribution Figure 2.2: Characteristics of the WHO database 7

20 such, but occasional mislabellings do occur. Thus, they cannot reliably be excluded from the analysis. Table 2.1 indicates what groups of drugs and ADRs have been reported most often during the entire life span of the WHO database. From Figure 2.2, it is clear that most reports list only one suspected drug and between one and four ADRs, but there are reports that deviate from this general pattern, and list very large numbers of drugs and ADRs. The most striking aspect of the empirical age distribution in Figure 2.2 is perhaps the large number of reports for children less than two years of age. A large proportion of these relate to suspected adverse reactions to vaccines. Another interesting phenomenon is the digit preference on 0 and 5 for encoding patient age. 2.3 Adverse drug reaction signal detection The detection of early warnings related to potential public health or patient safety issues is the main aim of collecting and analysing individual case safety reports. In the context of ADR surveillance, the WHO defines a signal as: "Reported information on a possible causal relationship between an adverse event and a drug, the relationship being unknown or incompletely documented previously. Usually more than a single report is required to generate a signal, depending upon the seriousness of the event and quality of the information." (Edwards and Biriell 1994) As is clear from the definition, single reports in isolation rarely motivate the communication of an early warning of a potential ADR, but there are exceptional examples where single reports of very high quality do (Meyboom et al. 1997). Particularly valuable pieces of information in this respect are those that indicate the effect on the ADR of withdrawing the suspected medication (socalled dechallenge intervention), and the effect of re-exposing the patient to the suspected treatment, after a successful dechallenge (so-called rechallenge intervention) (Edwards et al. 1990). Moreover, Aronson and Hauben (2006) argue that there are certain types of ADRs for which single, well documented incidents may motivate early warning, much in the spirit of the triage algorithms proposed by Ståhl et al. (2004). Early warning of a potential ADR is possible even in the absence of any individually very strong reports, if there is a large enough number of reports on the drug ADR pair of interest (Edwards et al. 1990). This is true in particular when alternative systematic explanations to excessive reporting rates, such as reporting biases or strong confounding, can be dismissed and the relative 8

21 Figure 2.3: Signal detection process reporting rate remains excessive even after suspected duplicates have been removed. The aim of ADR signal detection is to generate, strengthen and refine hypotheses related to suspected drug toxicity. Hypothesis testing is not possible on account of the inherently non-systematic nature of data collection and the lack of proper comparison groups. In-depth clinical evaluation and scrutiny of reports remain at the core of the ADR signal detection process. However, the WHO database receives tens of thousands of reports every month and this massive inflow of reports require efficient computational methods to help clinical experts focus on the groups of reports most likely to represent important public health or patient safety issues (Meyboom et al. 2002). As indicated in Figure 2.3, the signal detection process in routine use on the WHO database consists of a combination of automated knowledge discovery methods (Bate et al. 1998), triage (prioritisation) algorithms and clinical review (Ståhl et al. 2004). The knowledge discovery methods highlight drug ADR pairs with unexpectedly large numbers of reports relative to the average reporting rates in the database. Triage algorithms use a combination of quantitative and qualitative information to focus attention on the most urgent issues for follow-up (Ståhl et al. 2004). Reports related to drug ADR pairs singled out by the triage algorithm are forwarded to a panel of international experts for clinical review. In the context of the clinical review, pattern discovery methods may often be useful to profile larger groups of reports and suggest alternative explanations to observed excessive reporting rates. Hypotheses of suspected ADRs first highlighted in automated knowledge discovery that remain after clinical review are routinely communicated to the drug safety community, and some have been published in the mainstream medical literature (Coulter 9

22 et al. 2001, Sanz et al. 2005). However, the risk of distortion from undiscovered data quality problems and the difficulty of obtaining complete, detailed information on reported ADR incidents mean that signals of suspected ADRs often remain tentative, even after clinical review. 10

23 3. Knowledge discovery in adverse drug reaction surveillance Vast improvements in data storage capacity over the last decades have spurred ever increasing ambitions to analyse large, complex data sets not originally collected for the purpose of statistical analysis. Such investigations require data analysis methodology that scales well with increasing amounts of data and that focuses on discovery and exploration rather than on inference. This area of research and application, on the border between mathematical statistics and computer science, is referred to as knowledge discovery or data mining. Fayyad et al. (1996) describe data mining as one step in a more general knowledge discovery process. Mannila (1996) and Hand (1998) emphasise the similarity between data mining and exploratory statistical analysis, the latter characterising the difference as one primarily related to data set size and properties: in data mining, contamination, nonstationarity and biases are standard. On account of the complex data sets involved, interpretability is often a main consideration, which may favour simplicity at the expense of prediction accuracy (Glymour et al. 1997). An important dividing line is the choice between model based inference and algorithmic approaches (Breiman 2001). Whereas much of the research on knowledge discovery has been driven by computer science, key contributions from the statistical community include the clarification of inferential processes underlying algorithmic methods, insight into the bias variance trade-off in determining model complexity, methods for quantifying uncertainty and placing emphasis of the impact on interpretation of potential distortions such as confounding or selection biases (Elder and Pregibon 1996, Glymour et al. 1997, Efron 2001). In contrast with the more rigid framework for hypothesis testing, knowledge discovery is usually an interactive and iterative process of increasingly refined hypothesis generation. In my view, it should combine an unintimidated attitude towards the analysis of problematic and complex data sets with a proper understanding and clear statement of the limitations in nature and strength of the conclusions that can be drawn. 11

24 3.1 Context Collections of individual case safety reports clearly contain important pieces of rich and very useful information (Finney 1973, Edwards 1997), but they constitute an inherently non-random sample. The presence of reporting biases and violated independence assumptions discussed in Section 2 render summary statistics potentially deceptive. In particular, the presence of nonindependent reports can lead to optimistic precision estimates and invalidate standard tests for association (Finney 1971). As a consequence, the place for statistical methodology in the analysis of collections of individual case safety reports is somewhat out of the ordinary. Its main focus is on providing a framework for effective hypothesis generation and refinement, rather than on hypothesis testing (Bate 2003). Methods for reliably identifying elevated ADR reporting rates in collections of individual case safety reports are already part of routine drug safety signal detection (Ståhl et al. 2004). In the future, methods for highlighting suspected drug drug interaction, groups of nonindependent reports or reporting patterns involving larger sets of drugs and ADRs should allow for even more sophisticated use of this valuable source of information. The emphasis on hypothesis generation and refinement applies throughout this thesis: the aim of the record matching algorithm in I is to highlight likely duplicates for manual review and the aim of II, III and IV is to determine the most effective approach to highlighting apparently excessive ADR reporting rates for further follow-up. The purpose of implementing the methods in V for prioritisation of drug ADR pairs for further follow-up as discussed in Section 4.5 would also be effective hypothesis generation. In knowledge discovery, large numbers of possible associations and patterns are considered simultaneously. Familywise error rates that reflect the probability that any highlighted association corresponds to a false positive are usually less relevant in this context, because all open-ended investigations are bound to produce some false positives. Performance is better evaluated in terms of measures that indicate the proportion of false positives that can be anticipated in a specific study, such as false discovery rates. In our work, we have used two related measures of performance from the literature on Information Retrieval: precision (the number of true positives over the sum of true and false positives) and recall (the number of true positives over the sum of true positives and false negatives). Precision recall graphs that indicate how the precision and recall vary by the threshold for clinical review are used in both I and III. They provide an informative overview of performance, independent of the selected threshold. 12

25 Figure 3.1: Exploratory analysis of collections of individual case safety reports 3.2 Process Fayyad et al. (1996) define knowledge discovery as: "The nontrivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data." The knowledge discovery process is not limited to actual data analysis but includes: data collection, cleaning and preparation; reduction and projection; data analysis and interpretation, and finally dissemination, incorporation into existing structures and action based on discovered knowledge. It thus entails the entire ADR signal detection process outlined in Section 2.3, from the collection of reports and their pooling in an international database, through data preparation and transformation including conversion from free text to structured information, data cleaning and duplicate detection, via disproportionality analysis and triage algorithms to clinical review, and finally communication to national centres, pharmaceutical companies and the general public. The statistical methodology developed in the context of this thesis is applied at two different stages of the knowledge discovery process for ADR surveillance, as indicated in Figure 2.3. On one hand, disproportionality analysis is a core component in screening for excessive ADR reporting rates in first pass analysis of the database. On the other hand, pattern discovery methods are 13

26 useful in assisting clinical review and highlighting interesting aspects of specific groups of reports in more detailed investigations. Figure 3.1 proposes a general framework for such exploratory analysis. For the purpose of illustration, assume that the data subset of interest consists of all reports involving a particular drug D. At the outset of the exploratory analysis, simple descriptive information such as the total number of reports listing D and from what countries and during what time periods they have been submitted, may be very useful. Together with lists of the most commonly co-reported drugs and ADRs, as well as empirical distributions for patient age and gender, this provides a descriptive overview of the reporting of D which can serve as a useful reference for subsequent discoveries. Experienced data analysts may react directly to descriptive information that contradicts their subject matter knowledge. For example, a domain expert familiar with the WHO database may react to the observation that a suspiciously large proportion of the reports in a subgroup of interest have been submitted from a country with a low overall reporting rate. The middle box in Figure 3.1 is an attempt to formalise such comparative data analysis. Contrasts between the group of reports of interest and a comparison group (e.g. the database as a whole or all reports involving a drug in the same class of drugs as D) provide insight into what properties of the data subset differentiate it from the comparison group. For example, it may turn out that the relative reporting rate of a rare ADR for D by far exceeds that in the database as a whole. Such discrepancies may well be more enlightening than information on what the most commonly reported ADR is in absolute terms. The discussion of such disproportionality analysis is further extended in Section 3.3. Both descriptive and comparative studies may be misleading when the group of interest contains distinct subgroups. For example, if D is prescribed on one hand to young males and on the other hand to elderly females, the summary information that the average patient age on reports listing D is 43 years and that the overall proportion of females is 52%, conveys a very insufficient overview. Clustering algorithms allow for automated partitioning of data, with the aim of detecting latent structure, and may allow for much more relevant subsequent descriptive or comparative data analysis, as indicated by Figure 3.1. In addition to the iteration of automated partitioning, description and comparison described above, there are other methods for pattern discovery in collections of individual case safety reports. Record matching methods such as that adapted for duplicate detection in I can be used to detect groups of unexpectedly similar reports. Modified Hopfield networks and clustering algorithms such as those evaluated in Orre et al. (2005) may allow groups of often recurring ADRs (syndromes) to be identified. Similarly, interaction detection methods such as those in II and IV can be used to highlight suspected ADR risk factors. 14

27 It is rarely possible to specify at the outset of a large exploratory study, a fully automated, all-purpose approach to exploratory data analysis appropriate for all possible questions and patterns of potential interest. In addition, knowledge discovery often produces results that relate not to the primary study objective, but to fundamental properties of the data or of the data collection process. Thus, data cleaning and analysis are in practice intertwined, so that the correction of a data quality problem highlighted in initial data analysis allows for more refined subsequent data analysis. For example, in screening the WHO database for reporting patterns indicative of suspected drug interaction in IV, some larger groups of non-independent reports were highlighted. Their removal may allow for more accurate subsequent studies of drug drug interaction in the WHO database. 3.3 Disproportionality The frequency or relative frequency of a certain event (or set of events) in a database is sometimes of direct interest. However, in many knowledge discovery applications the discrepancy between the observed (relative) frequency and its expected value under some baseline model is of greater interest. An example from the analysis of purchasing patterns in supermarket sales data is that even if milk is the product most commonly purchased together with the product of interest, because this is true of most products, it may be more enlightening to point out that, for instance, grapefruit juice is purchased four times as often together with the product of interest as overall in the database. Such contrasts provide the basis of disproportionality analysis, which focuses on identifying events whose relative frequency in a given subgroup deviates substantially from the relative frequency of the same event in a given comparison group. Most modern methods for screening collections of individual case safety reports for excessive ADR reporting rates are based on disproportionality relative to the rest of the database. This is true of the Information Component (IC) (Bate et al. 1998), the Empirical Bayes Geometric Mean (EBGM) (DuMouchel 1999), the Proportional Reporting Ratio (PRR) (Evans et al. 2001) and the Reporting Odds Ratio (ROR) (Egberts et al. 2002). All these measures compare the number of reports on a certain drug ADR pair to an expected number of reports conditional on the overall reporting rates for the drug and the ADR in the database. The original idea of making comparisons with the database itself as reference goes back to the early days of ADR surveillance (Patwary 1969, Finney 1974). In addition to the lack of reliable external estimates for the international usage of different drugs, an advantage of disproportionality analysis is that marginal reporting biases that affect only the drug or only the ADR, cancel out (at least approximately) in 15

28 a measure of disproportionality. Thus, even though the reporting rates are likely to be higher for serious than for harmless ADRs, this does not have a considerable impact on the measures of disproportionality, as long as the reporting bias affects all drugs to an equal extent. The main drawback of disproportionality measures is that they rely on comparison to the reporting of other drug ADR pairs. Thus, if a particular drug ADR pair is massively reported, it will inflate the overall reporting rates for both the drug and the ADR, sometimes to the extent that excessive reporting rates for the same drug with another ADR or for another drug with the same ADR are masked (Evans 2004, Hauben et al. 2005). Assume the following contingency table based on the cross-classification of reports according to whether they involve a drug x and an ADR y: y not y x a b not x c d The basis for pairwise disproportionality analysis in the WHO database is an observed-to-expected ratio OE contrasting the relative reporting rate of y given x to the overall relative reporting rate of y in the database. With the annotation used in the above contingency table, the observed number of reports on y given x is a, and the expected number of reports conditional on the table marginals is the product of the marginal relative reporting rate of y and the total number of reports on x: is: a+c a+b+c+d (a + b). The observed-to-expected ratio OE = a/(a + b) (a + c)/(a + b + c + d) (3.1) The same measure of disproportionality has been used also in the context of association rule analysis (Agrawal et al. 1996), where it is referred to as the lift or the interest of an association rule involving x and y (Silverstein et al. 1998, Hastie et al. 2001). The similarity between the observed-to-expected ratio and other measures of disproportionality proposed for the analysis of individual case safety reports is clear. The Proportional Reporting Ratio (PRR) based on the above contingency table is (Evans et al. 2001): PRR = a/(a + b) c/(c + d) (3.2) and the corresponding Reporting Odds Ratio (ROR) is (Egberts et al. 2002): ROR = a/b c/d (3.3) 16

29 The IC measure of disproportionality used in routine knowledge discovery for the WHO database is essentially a conservative version of log 2 OE, that tends to 0 for rare drug ADR pairs. The moderation in magnitude is referred to as shrinkage (for details see Section 3.4 below). The availability of thoroughly evaluated shrinkage measures is the main advantage of the OE ratio over the PRR and the ROR. Other strengths are the link to Bayes classifiers described in Norén (2005) and the somewhat better robustness to zero counts in the contingency table than for the PRR and ROR (van Puijenbroek et al. 2002). The main limitation is that the observed-to-expected ratio provides a less distinct contrast between the group of interest and the reference group by including the group of interest in the reference. Another limitation is that the observed-toexpected ratio for a given pair of events by definition cannot exceed the inverse of the marginal relative reporting rate for each event. For example, if one of the events has an overall relative reporting rate of 0.5, then observed-to-expected ratios involving this event can at most reach 2 (if the relative reporting rate of the first event conditional on the other event is 1.00). In practice this limits the usefulness of observed-to-expected ratio as a measure of disproportionality to events that are reasonably rare. While disproportionality analysis is usually carried out at an early stage of the exploratory analysis of collections of individual case safety reports, there are sometimes requests to compute a measure of disproportionality for a drug ADR pair highlighted for review based on clinical judgement or on account of one or a few very strong reports. If the drug ADR pair turns out to be disproportionally reported, this may indeed lend added support. However, observed disproportionality must always be interpreted with caution. The possibility of alternative explanations such as report duplication, violated independence assumptions, publication biases or confounding must always be analysed and clearly stated. 3.4 Shrinkage Shrinkage is an attempt to regularise and reduce the volatility of a measure or parameter estimate of interest, by trading an increase in bias for a decrease in variance. In large and sparse data sets such as national or international collections of individual case safety reports, raw measures of disproportionality tend to sometimes yield very large values based on extremely low numbers of reports, but disproportionality based on just 1 or 2 reports is rarely of practical interest. The problem is that for rare drugs and ADRs, the expected number of reports may be very close to 0, relative to which even a single observed report may constitute a substantial deviation. Very low expected numbers of reports occur in the analysis of collections of individual case safety reports because the 2 by 2 contingency table of Section 3.3 is usually very unbalanced. Even 17

30 Figure 3.2: The simplified IC shrinkage measure plotted against the standard IC shrinkage measure for 10,000 randomly selected drug ADR pairs in the WHO database for the most common drugs and ADRs in the WHO database, the number of reports that do not involve either the drug or the ADR, d, is around 3,000,000, whereas b and c are generally in the order of 100 or 1,000 and a is even smaller (a 10 for 80% of the drug ADR pairs in the database). In order to reduce the vulnerability to spurious associations, two shrinkage measures of disproportionality have been proposed for the analysis of collections of individual case safety reports: the IC (Bate et al. 1998) and the EBGM (DuMouchel 1999). These measures of disproportionality are versions of the (logarithm of the) observed-to-expected ratio in (3.1) moderated towards a baseline value in the absence of large amounts of data. For the IC, the baseline value is 0 which corresponds to an observed-to-expected ratio of 1. Such shrinkage provides a robust measure of disproportionality moderated towards less extreme values for rare drugs and ADRs. However, as data accumulates it tends to log 2 OE as desired. The IC shrinkage measure is defined in II as a Bayesian maximum à posteriori estimate of a parameter related to the logarithm of the observed-to-expected ratio in (3.1). It is well approximated by the following simplified shrinkage measure based on observed and expected counts O xy and E xy : IC log 2 O xy + 1/2 E xy + 1/2 (3.4) A comparison between the IC shrinkage measure in II and that in (3.4) for 10,000 randomly selected drug ADR pairs in the WHO database is presented in Figure 3.2. Clearly, the difference between the two shrinkage measures is negligible. The main advantages of the simplified IC shrinkage measure are 18

31 that it is easier to compute and that it provides a general recipe for shrinkage that can be applied to any measure expressed in terms of an observed-toexpected ratio, such as the Ω measure of drug drug interaction in IV. This shrinkage can also be implemented for the PRR and ROR, after re-expression in terms of observed-to-expected ratios with O xy = a and E xy = (a+b)c c+d for the PRR, and with O xy = a and E xy = bc d for the ROR. Empirical Bayes estimation provides an alternative framework for shrinkage, where the prior distribution for a group of parameters is estimated based on the empirical distribution of maximum likelihood estimates for the group. The main advantage of empirical Bayes estimators is that they borrow strength from similar observations to improve the overall accuracy. However, with respect to each parameter, its estimate will only improve under the assumption that it is indeed related to the other parameters. Unlike the IC prior distribution, an empirical Bayes prior for the observed-to-expected ratio will not necessarily be centred at 1, and thus may inflate individual disproportionality measures rather than shrink them towards less extreme values. A practical issue is that for drug ADR pairs that have never been co-reported, the maximum likelihood estimate of the observed-to-expected ratio is 0. In practice, these drug ADR pairs appear to be ignored in the estimation of the empirical prior distribution in DuMouchel (1999), and the potential bias due to this is unclear. Berry and Berry (2004) propose a hierarchical empirical Bayes estimator for the observed-to-expected measure of disproportionality, where each measure of disproportionality is shrunk towards the group mean for a smaller group of more closely related ADRs. This should allow for more sophisticated empirical Bayes shrinkage, but the identification of appropriate groups of related ADR terms remains a challenging research problem in its own right. 3.5 Pattern discovery and detection Pattern recognition is the attempt to partition a group of data points into classes, based on a given set of explanatory variables (Webb 2002). Distinction is made between supervised and unsupervised pattern recognition: in supervised pattern recognition (or discrimination) a classifier is constructed based on training data consisting of labelled data points with the aim of accurately categorising unseen data points; in unsupervised classification (or clustering), the aim is to identify a natural partitioning of the available data set, without labelled training data available, or even a specification of what the classes of interest may be. For our purposes, the distinction between patterns and models in the context of pattern discovery and detection is more relevant. Hand and Bolton (2004) characterise patterns as related to local features of a data set involving only 19

32 subsets of the data points and/or subsets of the variables. Whereas a global model provides a high level description of the most important general features of a data set, a pattern may highlight one or a few outlying observations or a strong correlation between two variables. Hand and Bolton (2004) propose the following general definition: "A pattern is a local structure that generates data with an anomalously high density compared with that expected under the (global) baseline model." The focus on deviation from a global baseline model applies broadly to the methods described in this PhD thesis. The very aim of disproportionality analysis, is to identify groups of events that are co-reported more often than would be expected, based on a baseline independence model. Similarly, in duplicate detection and other record matching applications, the aim is to identify pairs (or small subsets) of unexpectedly similar reports whose similarity deviates from a global baseline model assuming all reports have been submitted independently. With the exception of the work on Bayes classifiers in V (which relates primarily to supervised pattern recognition by the above definition), this thesis focuses on unsupervised pattern discovery. The aim is to discover structure in data, without strict à priori specification of what the structure of interest is. At the same time, completely open-ended hypothesis generation is not possible as the type of potential patterns is determined by the choice of pattern discovery method, as well as implicitly by a range of other choices such as the variables considered in a given study (Hand 1994, p 319). Thus, while disproportionality analysis may highlight a variety of patterns related to anything from a suspected drug ADR association to an elevated reporting rate of a certain drug in one particular country, the type of patterns in such studies is restricted to unexpectedly high (or low) relative reporting rates. Similarly, record matching may highlight a variety of non-independent reports, but all highlighted patterns will refer to unexpected report similarity. 3.6 Facilitating interpretation Interpretation is one of the final steps in the knowledge discovery process (Fayyad et al. 1996), and a key component of the ADR signal detection process. Transparency is of particular importance in the analysis of non-systematically collected data such as individual case safety reports, where the use of overly complex statistical methodology may give a false sense of security and distract domain experts from limitations with the data (Hauben et al. 2005). Breiman (1985) refers to the application of 20

33 advanced statistical methodology to hide inadequacies with the data as edifice building ; in the ADR signal detection process, the use of overly complex statistical methods may divert clinical experts from careful consideration of alternative explanations to apparently excessive ADR relative reporting rates. Since the primary aim of applying knowledge discovery methods to collections of individual case safety reports is to guide and support domain experts in their manual review, better transparency is a strong argument in favour of choosing a simple method over a more complicated one. Indeed, better transparency is perhaps the strongest argument for choosing the simple IC shrinkage measure over the more complicated one as discussed in Section 3.4. Statistical sophistication does not necessarily rule out transparency, however. The hit-miss model record matching algorithm in I is based on a rather intricate probabilistic model, but its basis for highlighting a given record pair as suspected duplicates is immediately clear from an overview such as that presented in Figure 4.2 of Section 4. While sophisticated statistical methods are sometimes required to make the most of the available data, knowledge discovery results should always be presented as transparently as possible. For example, while shrinkage measures of disproportionality have proved a very powerful basis for filtering individual case safety reports for interesting reporting patterns, they may confuse domain experts, with little interest in the statistical methodology. Moreover, it is difficult to evaluate the impact of data quality issues such as suspected duplication or reporting biases on shrinkage measures of disproportionality. Observed and expected counts provide a more transparent explanation for why certain drug ADR pairs have been highlighted for manual review. In the presence of suspected data quality issues, simple arithmetic will indicate to what extent an excessive reporting rate may be due to a group of suspected duplicates, for instance. At the same time, domain experts often do want a sense of whether an observed disproportionality is likely to be due to chance or not. Credibility intervals around measures of disproportionality give some such indication, although the potential for violated independence assumptions means that precision can be overestimated. Adjustment for potential confounders may complicate interpretation of shrinkage measures of disproportionality. However, as commented on in III, adjusted observed-to-expected ratios sometimes correspond closely to stratum specific ones, and translating adjusted observed-to-expected ratios to stratum specific ones may simplify interpretation. For example, in the example on hypertension and zimeldine in Appendix E.4 of Hopstadius (2006), the IC increases from to when adjusted for time of reporting and country of origin. A closer investigation of the detailed data available in Appendix F.2 of the same thesis indicates that the discrepancy 21

34 is due to hypertension being more than twice as common on US reports (1.7%) as on reports from other countries (0.7%), whereas zimeldine was never used in the USA. Additionally the overall relative reporting rate of hypertension has increased in recent years, whereas zimeldine was primarily used in the early 1980 s. Thus, the crude IC which contrasts the observed relative reporting rate of hypertension given zimeldine to the overall relative reporting rate of hypertension in the entire database underestimates the disproportionality. Arguably, the best information to present to clinical experts in this case would be the observed number of reports on hypertension for zimeldine, and the expected number of such reports based on the relative reporting rate of hypertension in the countries and time period in which it was available. To guide domain experts to appropriate interpretation is clearly as important a challenge as method development in knowledge discovery research. 3.7 Future directions The new methodology proposed in this thesis provides a strong basis for future improvement and further research on knowledge discovery methods for collections of individual case safety reports. The method for drug drug interaction detection goes beyond simple drug ADR disproportional reporting rates, and could potentially be used also to screen for other types of ADR risk factors, such as related to patient gender or age. In general, we must aim to make better use of the rich information available on individual case safety reports. Virtually all knowledge discovery methods of today (including those described in this thesis) are based on raw numbers of reports (Hauben et al. 2005). They do not account for the amount or quality of information on each report nor for suspected duplication. This is in stark contrast with clinical review, in which both the quality of single reports and the quality of sets of reports as a group is carefully scrutinised (Meyboom et al. 1997). Indeed, given that the overall aim of applying knowledge discovery methods to collections of individual case safety reports is to assist and direct clinical review, an important challenge for the future is to achieve better alignment between automated knowledge discovery and clinical review. A first step may be to develop new and improved quality criteria for individual case safety reports similar to those discussed by Edwards et al. (1990). Based on such quality criteria, the number of high quality, distinct reports referring to a particular drug ADR pair can be identified and potentially provide a useful triage criterion. The possibility to highlight single high quality reports is interesting in its own right. The extended hit-miss model record matching algorithm has proved very useful for duplicate detection in the WHO database. Its importance is likely to 22

A Data Mining Approach for Signal Detection and Analysis

A Data Mining Approach for Signal Detection and Analysis SHORT COMMUNICATION Drug Safety 2002; 25 (6): 393-397 0114-5916/02/0006-0393/$25.00/0 Adis International Limited. All rights reserved. A Data Mining Approach for Signal Detection and Analysis Andrew Bate,

More information

Chapter 1. Introduction

Chapter 1. Introduction Chapter 1 Introduction 1.1 Motivation and Goals The increasing availability and decreasing cost of high-throughput (HT) technologies coupled with the availability of computational tools and data form a

More information

Principles of signal detection. Marie Lindquist

Principles of signal detection. Marie Lindquist Principles of signal detection Marie Lindquist Pharmacovigilance endpoints What we do should support good decision making in relation to medication therapy to maximise treatment effectiveness to reduce

More information

How Does Analysis of Competing Hypotheses (ACH) Improve Intelligence Analysis?

How Does Analysis of Competing Hypotheses (ACH) Improve Intelligence Analysis? How Does Analysis of Competing Hypotheses (ACH) Improve Intelligence Analysis? Richards J. Heuer, Jr. Version 1.2, October 16, 2005 This document is from a collection of works by Richards J. Heuer, Jr.

More information

Practitioner s Guide To Stratified Random Sampling: Part 1

Practitioner s Guide To Stratified Random Sampling: Part 1 Practitioner s Guide To Stratified Random Sampling: Part 1 By Brian Kriegler November 30, 2018, 3:53 PM EST This is the first of two articles on stratified random sampling. In the first article, I discuss

More information

A method for data driven exploration to pinpoint key features in medical data and facilitate expert review

A method for data driven exploration to pinpoint key features in medical data and facilitate expert review Received: 16 December 2016 Revised: 16 May 2017 Accepted: 14 July 2017 DOI: 10.1002/pds.4285 ORIGINAL REPORT A method for data driven exploration to pinpoint key features in medical data and facilitate

More information

COMMITTEE FOR PROPRIETARY MEDICINAL PRODUCTS (CPMP) POINTS TO CONSIDER ON MISSING DATA

COMMITTEE FOR PROPRIETARY MEDICINAL PRODUCTS (CPMP) POINTS TO CONSIDER ON MISSING DATA The European Agency for the Evaluation of Medicinal Products Evaluation of Medicines for Human Use London, 15 November 2001 CPMP/EWP/1776/99 COMMITTEE FOR PROPRIETARY MEDICINAL PRODUCTS (CPMP) POINTS TO

More information

International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use

International Conference on Harmonisation of Technical Requirements for Registration of Pharmaceuticals for Human Use Final Concept Paper E9(R1): Addendum to Statistical Principles for Clinical Trials on Choosing Appropriate Estimands and Defining Sensitivity Analyses in Clinical Trials dated 22 October 2014 Endorsed

More information

BACKGROUND + GENERAL COMMENTS

BACKGROUND + GENERAL COMMENTS Response on behalf of Sobi (Swedish Orphan Biovitrum AB) to the European Commission s Public Consultation on a Commission Notice on the Application of Articles 3, 5 and 7 of Regulation (EC) No. 141/2000

More information

GUIDELINE COMPARATORS & COMPARISONS:

GUIDELINE COMPARATORS & COMPARISONS: GUIDELINE COMPARATORS & COMPARISONS: Direct and indirect comparisons Adapted version (2015) based on COMPARATORS & COMPARISONS: Direct and indirect comparisons - February 2013 The primary objective of

More information

Time series disturbance detection for hypothesis-free signal detection in longitudinal observational databases

Time series disturbance detection for hypothesis-free signal detection in longitudinal observational databases Time series disturbance detection for hypothesis-free signal detection in longitudinal observational databases Andrew Bate Senior Director, Epidemiology Group Lead, Analytics On behalf of Manfred Hauben,

More information

Funnelling Used to describe a process of narrowing down of focus within a literature review. So, the writer begins with a broad discussion providing b

Funnelling Used to describe a process of narrowing down of focus within a literature review. So, the writer begins with a broad discussion providing b Accidental sampling A lesser-used term for convenience sampling. Action research An approach that challenges the traditional conception of the researcher as separate from the real world. It is associated

More information

Recent developments for combining evidence within evidence streams: bias-adjusted meta-analysis

Recent developments for combining evidence within evidence streams: bias-adjusted meta-analysis EFSA/EBTC Colloquium, 25 October 2017 Recent developments for combining evidence within evidence streams: bias-adjusted meta-analysis Julian Higgins University of Bristol 1 Introduction to concepts Standard

More information

ICH E9(R1) Technical Document. Estimands and Sensitivity Analysis in Clinical Trials STEP I TECHNICAL DOCUMENT TABLE OF CONTENTS

ICH E9(R1) Technical Document. Estimands and Sensitivity Analysis in Clinical Trials STEP I TECHNICAL DOCUMENT TABLE OF CONTENTS ICH E9(R1) Technical Document Estimands and Sensitivity Analysis in Clinical Trials STEP I TECHNICAL DOCUMENT TABLE OF CONTENTS A.1. Purpose and Scope A.2. A Framework to Align Planning, Design, Conduct,

More information

Active TB drug-safety management & monitoring

Active TB drug-safety management & monitoring Active TB drug-safety management & monitoring Global TB Programme, WHO, Switzerland 6 August 2015 Pharmacovigilance: definition of science and activities relating to the detection, assessment, understanding

More information

Risk Monitor How-to guide

Risk Monitor How-to guide Risk Monitor How-to guide Contents INTRODUCTION... 3 RESEARCH METHODS... 4 THE RISK MONITOR... 5 STEP-BY-STEP GUIDE... 6 Definitions... 7 PHASE ONE STUDY PLANNING... 8 Define the target population... 8

More information

Title: A robustness study of parametric and non-parametric tests in Model-Based Multifactor Dimensionality Reduction for epistasis detection

Title: A robustness study of parametric and non-parametric tests in Model-Based Multifactor Dimensionality Reduction for epistasis detection Author's response to reviews Title: A robustness study of parametric and non-parametric tests in Model-Based Multifactor Dimensionality Reduction for epistasis detection Authors: Jestinah M Mahachie John

More information

Chapter 02. Basic Research Methodology

Chapter 02. Basic Research Methodology Chapter 02 Basic Research Methodology Definition RESEARCH Research is a quest for knowledge through diligent search or investigation or experimentation aimed at the discovery and interpretation of new

More information

EPI 200C Final, June 4 th, 2009 This exam includes 24 questions.

EPI 200C Final, June 4 th, 2009 This exam includes 24 questions. Greenland/Arah, Epi 200C Sp 2000 1 of 6 EPI 200C Final, June 4 th, 2009 This exam includes 24 questions. INSTRUCTIONS: Write all answers on the answer sheets supplied; PRINT YOUR NAME and STUDENT ID NUMBER

More information

Evaluation Models STUDIES OF DIAGNOSTIC EFFICIENCY

Evaluation Models STUDIES OF DIAGNOSTIC EFFICIENCY 2. Evaluation Model 2 Evaluation Models To understand the strengths and weaknesses of evaluation, one must keep in mind its fundamental purpose: to inform those who make decisions. The inferences drawn

More information

extraction can take place. Another problem is that the treatment for chronic diseases is sequential based upon the progression of the disease.

extraction can take place. Another problem is that the treatment for chronic diseases is sequential based upon the progression of the disease. ix Preface The purpose of this text is to show how the investigation of healthcare databases can be used to examine physician decisions to develop evidence-based treatment guidelines that optimize patient

More information

A Comparison of Collaborative Filtering Methods for Medication Reconciliation

A Comparison of Collaborative Filtering Methods for Medication Reconciliation A Comparison of Collaborative Filtering Methods for Medication Reconciliation Huanian Zheng, Rema Padman, Daniel B. Neill The H. John Heinz III College, Carnegie Mellon University, Pittsburgh, PA, 15213,

More information

Technical Specifications

Technical Specifications Technical Specifications In order to provide summary information across a set of exercises, all tests must employ some form of scoring models. The most familiar of these scoring models is the one typically

More information

Reporting Patterns Indicative of Adverse Drug Interactions

Reporting Patterns Indicative of Adverse Drug Interactions ORIGINAL RESEARCH ARTICLE Drug Saf 2011; 34 (3): 253-266 0114-5916/11/0003-0253/$49.95/0 ª 2011 Adis Data Information BV. All rights reserved. Reporting Patterns Indicative of Adverse Drug Interactions

More information

Data and Statistics 101: Key Concepts in the Collection, Analysis, and Application of Child Welfare Data

Data and Statistics 101: Key Concepts in the Collection, Analysis, and Application of Child Welfare Data TECHNICAL REPORT Data and Statistics 101: Key Concepts in the Collection, Analysis, and Application of Child Welfare Data CONTENTS Executive Summary...1 Introduction...2 Overview of Data Analysis Concepts...2

More information

Statistical Considerations for Post Market Surveillance

Statistical Considerations for Post Market Surveillance Statistical Considerations for Post Market Surveillance 30 th International Conference on Phamacoepidemiology and Therapeutic Risk Management Ted Lystig, Ph.D. MEDTRONIC INC. Distinguished Statistician

More information

Estimands and Sensitivity Analysis in Clinical Trials E9(R1)

Estimands and Sensitivity Analysis in Clinical Trials E9(R1) INTERNATIONAL CONCIL FOR HARMONISATION OF TECHNICAL REQUIREMENTS FOR PHARMACEUTICALS FOR HUMAN USE ICH HARMONISED GUIDELINE Estimands and Sensitivity Analysis in Clinical Trials E9(R1) Current Step 2 version

More information

The Regression-Discontinuity Design

The Regression-Discontinuity Design Page 1 of 10 Home» Design» Quasi-Experimental Design» The Regression-Discontinuity Design The regression-discontinuity design. What a terrible name! In everyday language both parts of the term have connotations

More information

Response to the ASA s statement on p-values: context, process, and purpose

Response to the ASA s statement on p-values: context, process, and purpose Response to the ASA s statement on p-values: context, process, purpose Edward L. Ionides Alexer Giessing Yaacov Ritov Scott E. Page Departments of Complex Systems, Political Science Economics, University

More information

Hypothesis-Driven Research

Hypothesis-Driven Research Hypothesis-Driven Research Research types Descriptive science: observe, describe and categorize the facts Discovery science: measure variables to decide general patterns based on inductive reasoning Hypothesis-driven

More information

ISPOR Task Force Report: ITC & NMA Study Questionnaire

ISPOR Task Force Report: ITC & NMA Study Questionnaire INDIRECT TREATMENT COMPARISON / NETWORK META-ANALYSIS STUDY QUESTIONNAIRE TO ASSESS RELEVANCE AND CREDIBILITY TO INFORM HEALTHCARE DECISION-MAKING: AN ISPOR-AMCP-NPC GOOD PRACTICE TASK FORCE REPORT DRAFT

More information

THIS CHAPTER COVERS: The importance of sampling. Populations, sampling frames, and samples. Qualities of a good sample.

THIS CHAPTER COVERS: The importance of sampling. Populations, sampling frames, and samples. Qualities of a good sample. VII. WHY SAMPLE? THIS CHAPTER COVERS: The importance of sampling Populations, sampling frames, and samples Qualities of a good sample Sampling size Ways to obtain a representative sample based on probability

More information

Guidance Document for Claims Based on Non-Inferiority Trials

Guidance Document for Claims Based on Non-Inferiority Trials Guidance Document for Claims Based on Non-Inferiority Trials February 2013 1 Non-Inferiority Trials Checklist Item No Checklist Item (clients can use this tool to help make decisions regarding use of non-inferiority

More information

Citation for published version (APA): Ebbes, P. (2004). Latent instrumental variables: a new approach to solve for endogeneity s.n.

Citation for published version (APA): Ebbes, P. (2004). Latent instrumental variables: a new approach to solve for endogeneity s.n. University of Groningen Latent instrumental variables Ebbes, P. IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from it. Please check the document

More information

Controlled Trials. Spyros Kitsiou, PhD

Controlled Trials. Spyros Kitsiou, PhD Assessing Risk of Bias in Randomized Controlled Trials Spyros Kitsiou, PhD Assistant Professor Department of Biomedical and Health Information Sciences College of Applied Health Sciences University of

More information

INTERNATIONAL STANDARD ON AUDITING (NEW ZEALAND) 530

INTERNATIONAL STANDARD ON AUDITING (NEW ZEALAND) 530 Issued 07/11 INTERNATIONAL STANDARD ON AUDITING (NEW ZEALAND) 530 Audit Sampling (ISA (NZ) 530) Issued July 2011 Effective for audits of historical financial statements for periods beginning on or after

More information

A Brief Introduction to Bayesian Statistics

A Brief Introduction to Bayesian Statistics A Brief Introduction to Statistics David Kaplan Department of Educational Psychology Methods for Social Policy Research and, Washington, DC 2017 1 / 37 The Reverend Thomas Bayes, 1701 1761 2 / 37 Pierre-Simon

More information

Applying Machine Learning Methods in Medical Research Studies

Applying Machine Learning Methods in Medical Research Studies Applying Machine Learning Methods in Medical Research Studies Daniel Stahl Department of Biostatistics and Health Informatics Psychiatry, Psychology & Neuroscience (IoPPN), King s College London daniel.r.stahl@kcl.ac.uk

More information

Principles of Signal Detection. Pia Caduff-Janosa MD Annual Course May 2013

Principles of Signal Detection. Pia Caduff-Janosa MD Annual Course May 2013 Principles of Signal Detection Pia Caduff-Janosa MD Annual Course May 2013 Outline What is a signal? Signal sources Triggers Assessment of potential risk Signal evaluation Outcome 2 Pia Caduff-Janosa,

More information

Design and Analysis Plan Quantitative Synthesis of Federally-Funded Teen Pregnancy Prevention Programs HHS Contract #HHSP I 5/2/2016

Design and Analysis Plan Quantitative Synthesis of Federally-Funded Teen Pregnancy Prevention Programs HHS Contract #HHSP I 5/2/2016 Design and Analysis Plan Quantitative Synthesis of Federally-Funded Teen Pregnancy Prevention Programs HHS Contract #HHSP233201500069I 5/2/2016 Overview The goal of the meta-analysis is to assess the effects

More information

Surveillance strategies for Foot and Mouth Disease to prove absence from disease and absence of viral circulation

Surveillance strategies for Foot and Mouth Disease to prove absence from disease and absence of viral circulation Surveillance strategies for Foot and Mouth Disease to prove absence from disease and absence of viral circulation Vincenzo Capora Asuncion, 25 June 20 Introduction In FMD control the need to prove absence

More information

MODEL SELECTION STRATEGIES. Tony Panzarella

MODEL SELECTION STRATEGIES. Tony Panzarella MODEL SELECTION STRATEGIES Tony Panzarella Lab Course March 20, 2014 2 Preamble Although focus will be on time-to-event data the same principles apply to other outcome data Lab Course March 20, 2014 3

More information

Sawtooth Software. The Number of Levels Effect in Conjoint: Where Does It Come From and Can It Be Eliminated? RESEARCH PAPER SERIES

Sawtooth Software. The Number of Levels Effect in Conjoint: Where Does It Come From and Can It Be Eliminated? RESEARCH PAPER SERIES Sawtooth Software RESEARCH PAPER SERIES The Number of Levels Effect in Conjoint: Where Does It Come From and Can It Be Eliminated? Dick Wittink, Yale University Joel Huber, Duke University Peter Zandan,

More information

Understanding Uncertainty in School League Tables*

Understanding Uncertainty in School League Tables* FISCAL STUDIES, vol. 32, no. 2, pp. 207 224 (2011) 0143-5671 Understanding Uncertainty in School League Tables* GEORGE LECKIE and HARVEY GOLDSTEIN Centre for Multilevel Modelling, University of Bristol

More information

11/18/2013. Correlational Research. Correlational Designs. Why Use a Correlational Design? CORRELATIONAL RESEARCH STUDIES

11/18/2013. Correlational Research. Correlational Designs. Why Use a Correlational Design? CORRELATIONAL RESEARCH STUDIES Correlational Research Correlational Designs Correlational research is used to describe the relationship between two or more naturally occurring variables. Is age related to political conservativism? Are

More information

Adverse Events Monitoring (aka Pharmacovigilance)

Adverse Events Monitoring (aka Pharmacovigilance) Adverse Events Monitoring (aka Pharmacovigilance) Presentation to American Conference Institute s FDA Boot Camp William W. Vodra May 16, 2007 Agenda What is pharmacovigilance (PV)? How does PV use adverse

More information

Further data analysis topics

Further data analysis topics Further data analysis topics Jonathan Cook Centre for Statistics in Medicine, NDORMS, University of Oxford EQUATOR OUCAGS training course 24th October 2015 Outline Ideal study Further topics Multiplicity

More information

Chapter 11: Advanced Remedial Measures. Weighted Least Squares (WLS)

Chapter 11: Advanced Remedial Measures. Weighted Least Squares (WLS) Chapter : Advanced Remedial Measures Weighted Least Squares (WLS) When the error variance appears nonconstant, a transformation (of Y and/or X) is a quick remedy. But it may not solve the problem, or it

More information

Cochrane Pregnancy and Childbirth Group Methodological Guidelines

Cochrane Pregnancy and Childbirth Group Methodological Guidelines Cochrane Pregnancy and Childbirth Group Methodological Guidelines [Prepared by Simon Gates: July 2009, updated July 2012] These guidelines are intended to aid quality and consistency across the reviews

More information

Glossary From Running Randomized Evaluations: A Practical Guide, by Rachel Glennerster and Kudzai Takavarasha

Glossary From Running Randomized Evaluations: A Practical Guide, by Rachel Glennerster and Kudzai Takavarasha Glossary From Running Randomized Evaluations: A Practical Guide, by Rachel Glennerster and Kudzai Takavarasha attrition: When data are missing because we are unable to measure the outcomes of some of the

More information

Regression Discontinuity Analysis

Regression Discontinuity Analysis Regression Discontinuity Analysis A researcher wants to determine whether tutoring underachieving middle school students improves their math grades. Another wonders whether providing financial aid to low-income

More information

UNIT 5 - Association Causation, Effect Modification and Validity

UNIT 5 - Association Causation, Effect Modification and Validity 5 UNIT 5 - Association Causation, Effect Modification and Validity Introduction In Unit 1 we introduced the concept of causality in epidemiology and presented different ways in which causes can be understood

More information

Confounding and Bias

Confounding and Bias 28 th International Conference on Pharmacoepidemiology and Therapeutic Risk Management Barcelona, Spain August 22, 2012 Confounding and Bias Tobias Gerhard, PhD Assistant Professor, Ernest Mario School

More information

Problem solving therapy

Problem solving therapy Introduction People with severe mental illnesses such as schizophrenia may show impairments in problem-solving ability. Remediation interventions such as problem solving skills training can help people

More information

MS&E 226: Small Data

MS&E 226: Small Data MS&E 226: Small Data Lecture 10: Introduction to inference (v2) Ramesh Johari ramesh.johari@stanford.edu 1 / 17 What is inference? 2 / 17 Where did our data come from? Recall our sample is: Y, the vector

More information

Study 2a: A level biology, psychology and sociology

Study 2a: A level biology, psychology and sociology Inter-subject comparability studies Study 2a: A level biology, psychology and sociology May 2008 QCA/08/3653 Contents 1 Personnel... 3 2 Materials... 4 3 Methodology... 5 3.1 Form A... 5 3.2 CRAS analysis...

More information

You can t fix by analysis what you bungled by design. Fancy analysis can t fix a poorly designed study.

You can t fix by analysis what you bungled by design. Fancy analysis can t fix a poorly designed study. You can t fix by analysis what you bungled by design. Light, Singer and Willett Or, not as catchy but perhaps more accurate: Fancy analysis can t fix a poorly designed study. Producing Data The Role of

More information

Psychology 2019 v1.3. IA2 high-level annotated sample response. Student experiment (20%) August Assessment objectives

Psychology 2019 v1.3. IA2 high-level annotated sample response. Student experiment (20%) August Assessment objectives Student experiment (20%) This sample has been compiled by the QCAA to assist and support teachers to match evidence in student responses to the characteristics described in the instrument-specific marking

More information

A Spreadsheet for Deriving a Confidence Interval, Mechanistic Inference and Clinical Inference from a P Value

A Spreadsheet for Deriving a Confidence Interval, Mechanistic Inference and Clinical Inference from a P Value SPORTSCIENCE Perspectives / Research Resources A Spreadsheet for Deriving a Confidence Interval, Mechanistic Inference and Clinical Inference from a P Value Will G Hopkins sportsci.org Sportscience 11,

More information

Fundamental Clinical Trial Design

Fundamental Clinical Trial Design Design, Monitoring, and Analysis of Clinical Trials Session 1 Overview and Introduction Overview Scott S. Emerson, M.D., Ph.D. Professor of Biostatistics, University of Washington February 17-19, 2003

More information

How to interpret results of metaanalysis

How to interpret results of metaanalysis How to interpret results of metaanalysis Tony Hak, Henk van Rhee, & Robert Suurmond Version 1.0, March 2016 Version 1.3, Updated June 2018 Meta-analysis is a systematic method for synthesizing quantitative

More information

IAASB Main Agenda (February 2007) Page Agenda Item PROPOSED INTERNATIONAL STANDARD ON AUDITING 530 (REDRAFTED)

IAASB Main Agenda (February 2007) Page Agenda Item PROPOSED INTERNATIONAL STANDARD ON AUDITING 530 (REDRAFTED) IAASB Main Agenda (February 2007) Page 2007 423 Agenda Item 6-A PROPOSED INTERNATIONAL STANDARD ON AUDITING 530 (REDRAFTED) AUDIT SAMPLING AND OTHER MEANS OF TESTING CONTENTS Paragraph Introduction Scope

More information

Introduction to diagnostic accuracy meta-analysis. Yemisi Takwoingi October 2015

Introduction to diagnostic accuracy meta-analysis. Yemisi Takwoingi October 2015 Introduction to diagnostic accuracy meta-analysis Yemisi Takwoingi October 2015 Learning objectives To appreciate the concept underlying DTA meta-analytic approaches To know the Moses-Littenberg SROC method

More information

Unit 1 Exploring and Understanding Data

Unit 1 Exploring and Understanding Data Unit 1 Exploring and Understanding Data Area Principle Bar Chart Boxplot Conditional Distribution Dotplot Empirical Rule Five Number Summary Frequency Distribution Frequency Polygon Histogram Interquartile

More information

White Paper Estimating Complex Phenotype Prevalence Using Predictive Models

White Paper Estimating Complex Phenotype Prevalence Using Predictive Models White Paper 23-12 Estimating Complex Phenotype Prevalence Using Predictive Models Authors: Nicholas A. Furlotte Aaron Kleinman Robin Smith David Hinds Created: September 25 th, 2015 September 25th, 2015

More information

Representing Association Classification Rules Mined from Health Data

Representing Association Classification Rules Mined from Health Data Representing Association Classification Rules Mined from Health Data Jie Chen 1, Hongxing He 1,JiuyongLi 4, Huidong Jin 1, Damien McAullay 1, Graham Williams 1,2, Ross Sparks 1,andChrisKelman 3 1 CSIRO

More information

Detecting Suspect Examinees: An Application of Differential Person Functioning Analysis. Russell W. Smith Susan L. Davis-Becker

Detecting Suspect Examinees: An Application of Differential Person Functioning Analysis. Russell W. Smith Susan L. Davis-Becker Detecting Suspect Examinees: An Application of Differential Person Functioning Analysis Russell W. Smith Susan L. Davis-Becker Alpine Testing Solutions Paper presented at the annual conference of the National

More information

3 CONCEPTUAL FOUNDATIONS OF STATISTICS

3 CONCEPTUAL FOUNDATIONS OF STATISTICS 3 CONCEPTUAL FOUNDATIONS OF STATISTICS In this chapter, we examine the conceptual foundations of statistics. The goal is to give you an appreciation and conceptual understanding of some basic statistical

More information

Lecture Outline. Biost 590: Statistical Consulting. Stages of Scientific Studies. Scientific Method

Lecture Outline. Biost 590: Statistical Consulting. Stages of Scientific Studies. Scientific Method Biost 590: Statistical Consulting Statistical Classification of Scientific Studies; Approach to Consulting Lecture Outline Statistical Classification of Scientific Studies Statistical Tasks Approach to

More information

Psych 1Chapter 2 Overview

Psych 1Chapter 2 Overview Psych 1Chapter 2 Overview After studying this chapter, you should be able to answer the following questions: 1) What are five characteristics of an ideal scientist? 2) What are the defining elements of

More information

AUG Expert responses to issues raised by British Gas during the query period for the first draft 2017/18 AUG Statement, 14 March 2017.

AUG Expert responses to issues raised by British Gas during the query period for the first draft 2017/18 AUG Statement, 14 March 2017. AUG Expert responses to issues raised by British Gas during the query period for the first draft 2017/18 AUG Statement, 14 March 2017. Below are the issues raised by British Gas. Following each issue is

More information

Smith, J., & Noble, H. (2014). Bias in research. Evidence-Based Nursing, 17(4), DOI: /eb

Smith, J., & Noble, H. (2014). Bias in research. Evidence-Based Nursing, 17(4), DOI: /eb Bias in research Smith, J., & Noble, H. (2014). Bias in research. Evidence-Based Nursing, 17(4), 100-101. DOI: 10.1136/eb- 2014-101946 Published in: Evidence-Based Nursing Document Version: Peer reviewed

More information

observational studies Descriptive studies

observational studies Descriptive studies form one stage within this broader sequence, which begins with laboratory studies using animal models, thence to human testing: Phase I: The new drug or treatment is tested in a small group of people for

More information

Animal-assisted therapy

Animal-assisted therapy Introduction Animal-assisted interventions use trained animals to help improve physical, mental and social functions in people with schizophrenia. It is a goal-directed intervention in which an animal

More information

PLS 506 Mark T. Imperial, Ph.D. Lecture Notes: Reliability & Validity

PLS 506 Mark T. Imperial, Ph.D. Lecture Notes: Reliability & Validity PLS 506 Mark T. Imperial, Ph.D. Lecture Notes: Reliability & Validity Measurement & Variables - Initial step is to conceptualize and clarify the concepts embedded in a hypothesis or research question with

More information

Transmission to CHMP July Adoption by CHMP for release for consultation 20 July Start of consultation 31 August 2017

Transmission to CHMP July Adoption by CHMP for release for consultation 20 July Start of consultation 31 August 2017 1 2 3 30 August 2017 EMA/CHMP/ICH/436221/2017 Committee for Human Medicinal Products 4 ICH E9 (R1) addendum on estimands and sensitivity 5 analysis in clinical trials to the guideline on statistical 6

More information

Understandable Statistics

Understandable Statistics Understandable Statistics correlated to the Advanced Placement Program Course Description for Statistics Prepared for Alabama CC2 6/2003 2003 Understandable Statistics 2003 correlated to the Advanced Placement

More information

2 Critical thinking guidelines

2 Critical thinking guidelines What makes psychological research scientific? Precision How psychologists do research? Skepticism Reliance on empirical evidence Willingness to make risky predictions Openness Precision Begin with a Theory

More information

CARE Cross-project Collectives Analysis: Technical Appendix

CARE Cross-project Collectives Analysis: Technical Appendix CARE Cross-project Collectives Analysis: Technical Appendix The CARE approach to development support and women s economic empowerment is often based on the building blocks of collectives. Some of these

More information

Incorporating Clinical Information into the Label

Incorporating Clinical Information into the Label ECC Population Health Group LLC expertise-driven consulting in global maternal-child health & pharmacoepidemiology Incorporating Clinical Information into the Label Labels without Categories: A Workshop

More information

WELCOME! Lecture 11 Thommy Perlinger

WELCOME! Lecture 11 Thommy Perlinger Quantitative Methods II WELCOME! Lecture 11 Thommy Perlinger Regression based on violated assumptions If any of the assumptions are violated, potential inaccuracies may be present in the estimated regression

More information

Measuring impact. William Parienté UC Louvain J PAL Europe. povertyactionlab.org

Measuring impact. William Parienté UC Louvain J PAL Europe. povertyactionlab.org Measuring impact William Parienté UC Louvain J PAL Europe povertyactionlab.org Course overview 1. What is evaluation? 2. Measuring impact 3. Why randomize? 4. How to randomize 5. Sampling and Sample Size

More information

Polypharmacy and Deprescribing. A special report on views from the PrescQIPP landscape review

Polypharmacy and Deprescribing. A special report on views from the PrescQIPP landscape review Polypharmacy and Deprescribing A special report on views from the PrescQIPP landscape review Introduction and background In recent years, we have seen the emergence of an international discussion around

More information

ISA 540, Auditing Accounting Estimates, Including Fair Value Accounting Estimates, and Related Disclosures Issues and Task Force Recommendations

ISA 540, Auditing Accounting Estimates, Including Fair Value Accounting Estimates, and Related Disclosures Issues and Task Force Recommendations Agenda Item 1-A ISA 540, Auditing Accounting Estimates, Including Fair Value Accounting Estimates, and Related Disclosures Issues and Task Force Recommendations Introduction 1. Since the September 2016

More information

Results. NeuRA Hypnosis June 2016

Results. NeuRA Hypnosis June 2016 Introduction may be experienced as an altered state of consciousness or as a state of relaxation. There is no agreed framework for administering hypnosis, but the procedure often involves induction (such

More information

NeuRA Sleep disturbance April 2016

NeuRA Sleep disturbance April 2016 Introduction People with schizophrenia may show disturbances in the amount, or the quality of sleep they generally receive. Typically sleep follows a characteristic pattern of four stages, where stage

More information

Dose and Response for Chemicals

Dose and Response for Chemicals Dose and Response for Chemicals 5 5 DOSE AND RESPONSE FOR CHEMICALS All substances are poisons; there is none which is not a poison. The right dose differentiates a poison and a remedy. Paracelsus, 16th

More information

ISC- GRADE XI HUMANITIES ( ) PSYCHOLOGY. Chapter 2- Methods of Psychology

ISC- GRADE XI HUMANITIES ( ) PSYCHOLOGY. Chapter 2- Methods of Psychology ISC- GRADE XI HUMANITIES (2018-19) PSYCHOLOGY Chapter 2- Methods of Psychology OUTLINE OF THE CHAPTER (i) Scientific Methods in Psychology -observation, case study, surveys, psychological tests, experimentation

More information

Chapter-2 RESEARCH DESIGN

Chapter-2 RESEARCH DESIGN Chapter-2 RESEARCH DESIGN 33 2.1 Introduction to Research Methodology: The general meaning of research is the search for knowledge. Research is also defined as a careful investigation or inquiry, especially

More information

The Impact of Relative Standards on the Propensity to Disclose. Alessandro Acquisti, Leslie K. John, George Loewenstein WEB APPENDIX

The Impact of Relative Standards on the Propensity to Disclose. Alessandro Acquisti, Leslie K. John, George Loewenstein WEB APPENDIX The Impact of Relative Standards on the Propensity to Disclose Alessandro Acquisti, Leslie K. John, George Loewenstein WEB APPENDIX 2 Web Appendix A: Panel data estimation approach As noted in the main

More information

Statistical analysis DIANA SAPLACAN 2017 * SLIDES ADAPTED BASED ON LECTURE NOTES BY ALMA LEORA CULEN

Statistical analysis DIANA SAPLACAN 2017 * SLIDES ADAPTED BASED ON LECTURE NOTES BY ALMA LEORA CULEN Statistical analysis DIANA SAPLACAN 2017 * SLIDES ADAPTED BASED ON LECTURE NOTES BY ALMA LEORA CULEN Vs. 2 Background 3 There are different types of research methods to study behaviour: Descriptive: observations,

More information

CHECKLIST FOR EVALUATING A RESEARCH REPORT Provided by Dr. Blevins

CHECKLIST FOR EVALUATING A RESEARCH REPORT Provided by Dr. Blevins CHECKLIST FOR EVALUATING A RESEARCH REPORT Provided by Dr. Blevins 1. The Title a. Is it clear and concise? b. Does it promise no more than the study can provide? INTRODUCTION 2. The Problem a. It is clearly

More information

This document was created on 07 January 2005 and has been printed from

This document was created on 07 January 2005 and has been printed from 1 Spontaneous Adverse Event Reporting Information Introduction This document provides information regarding spontaneous adverse event reports received concerning patients taking rosuvastatin, which relates

More information

Psychology, 2010, 1: doi: /psych Published Online August 2010 (

Psychology, 2010, 1: doi: /psych Published Online August 2010 ( Psychology, 2010, 1: 194-198 doi:10.4236/psych.2010.13026 Published Online August 2010 (http://www.scirp.org/journal/psych) Using Generalizability Theory to Evaluate the Applicability of a Serial Bayes

More information

Results. NeuRA Treatments for internalised stigma December 2017

Results. NeuRA Treatments for internalised stigma December 2017 Introduction Internalised stigma occurs within an individual, such that a person s attitude may reinforce a negative self-perception of mental disorders, resulting in reduced sense of selfworth, anticipation

More information

Psychology Research Excellence Framework 2014 Impact Analysis

Psychology Research Excellence Framework 2014 Impact Analysis Psychology Research Excellence Framework 2014 Impact Analysis Thom Baguley Acknowledgement The Research Board sincerely thanks Professor Thomas Baguley for preparing this report on its behalf. Thom Baguley

More information

DRAFT (Final) Concept Paper On choosing appropriate estimands and defining sensitivity analyses in confirmatory clinical trials

DRAFT (Final) Concept Paper On choosing appropriate estimands and defining sensitivity analyses in confirmatory clinical trials DRAFT (Final) Concept Paper On choosing appropriate estimands and defining sensitivity analyses in confirmatory clinical trials EFSPI Comments Page General Priority (H/M/L) Comment The concept to develop

More information

Safeguarding public health Subgroup Analyses: Important, Infuriating and Intractable

Safeguarding public health Subgroup Analyses: Important, Infuriating and Intractable Safeguarding public health Subgroup Analyses: Important, Infuriating and Intractable The views expressed herein are not necessarily those of MHRA, EMA, EMA committees or their working parties. Rob Hemmings

More information

Department of International Health

Department of International Health JOHNS HOPKINS U N I V E R S I T Y Center for Clinical Trials Department of Biostatistics Department of Epidemiology Department of International Health Memorandum Department of Medicine Department of Ophthalmology

More information