key: cord-0695354-myp2bu4f
authors: Cadarette, Suzanne M.; Maclure, Malcolm; Delaney, J. A. Chris; Whitaker, Heather J.; Hayes, Kaleen N.; Wang, Shirley V.; Tadrous, Mina; Gagne, Joshua J.; Consiglio, Giulia P.; Hallas, Jesper
title: Control yourself: ISPE‐endorsed guidance in the application of self‐controlled study designs in pharmacoepidemiology
date: 2021-04-01
journal: Pharmacoepidemiol Drug Saf
DOI: 10.1002/pds.5227
sha: 22e792fec0328b7038be32ee8eecb8a92fae62fc
doc_id: 695354
cord_uid: myp2bu4f

PURPOSE: Consensus is needed on conceptual foundations, terminology and relationships among the various self‐controlled “trigger” study designs that control for time‐invariant confounding factors and target the association between transient exposures (potential triggers) and abrupt outcomes. The International Society for Pharmacoepidemiology (ISPE) funded a working group of ISPE members to develop guidance material for the application and reporting of self‐controlled study designs, similar to Standards of Reporting Observational Epidemiology (STROBE). This first paper focuses on navigation between the types of self‐controlled designs to permit a foundational understanding with guiding principles. METHODS: We leveraged a systematic review of applications of these designs, that we term Self‐controlled Crossover Observational PharmacoEpidemiologic (SCOPE) studies. Starting from first principles and using case examples, we reviewed outcome‐anchored (case‐crossover [CCO], case‐time control [CTC], case‐case‐time control [CCTC]) and exposure‐anchored (self‐controlled case‐series [SCCS]) study designs. RESULTS: Key methodological features related to exposure, outcome and time‐related concerns were clarified, and a common language and worksheet to facilitate the design of SCOPE studies is introduced. CONCLUSIONS: Consensus on conceptual foundations, terminology and relationships among SCOPE designs will facilitate understanding and critical appraisal of published studies, as well as help in the design, analysis and review of new SCOPE studies. This manuscript is endorsed by ISPE.

Pharmacoepidemiology bridges the fields of clinical pharmacology and epidemiology by targeting the effects of therapeutic drugs in humans. 1 Large healthcare databases that include drugs dispensed or prescribed, as well as medical claims (eg, diagnoses, procedures) are often utilized to study drug safety and effectiveness in the "real world." 2 Cohort and case-control studies are well established traditional epidemiologic designs used to estimate the effects of drug exposure on disease (outcome) incidence, by comparing different groups of patients. Cohort studies compare outcome measurements between patients exposed to a drug to patients unexposed, or patients exposed to a different drug or drugs; and case-control studies compare exposure histories between patients that experience the outcome to patients without the outcome.

However, missing clinical detail and lifestyle factor information often limit the ability to adjust for confounding factors that vary between groups, and is a commonly cited limitation of traditional epidemiologic designs in pharmacoepidemiology. [2] [3] [4] Unlike the cohort and case-control designs that compare different groups of patients, Self-controlled Crossover Observational PharmacoEpidemiologic (SCOPE, Box 1) 5 The application of SCOPE studies is increasing, yet inconsistent and ambiguous language has been used to describe methodological features that may hamper the reader's ability to understand, critique, or replicate. 10 We received funding from the International Society for Pharmacoepidemiology (ISPE) to develop guidance documents for the application and reporting of SCOPE designs. In this first paper, we start from first principles, briefly reviewing foundational concepts in causal inference, pharmacology and epidemiology that inform the design of SCOPE studies. We then introduce a common language, and SCOPE designs using published examples. Key study design features are summarized to help the reader remain mindful of potential exposure-, outcome-, and time-related issues that need to be considered in the design of a SCOPE study. This document aims to provide a solid foundation and introduction for those new to SCOPE designs as well as clarify concepts and encourage a common language for experienced methodologists. This manuscript is endorsed by ISPE.

This section briefly introduces foundational principles in causation, pharmacology and epidemiology that inform the design of SCOPE studies.

• Despite differences in terminology, Self-controlled Crossover Observational PharmacoEpidemiology (SCOPE) study designs share conceptual foundations and a common strategy for controlling time-invariant confounding.

• SCOPE designs are best suited to studying transient exposures in relation to abrupt outcomes, and are broadly split into: outcome-anchored (case-crossover, case-timecontrol and case-case-time control), and exposureanchored (self-controlled case series) that are suitable for slightly different research questions.

• A proposed common terminology and worksheet facilitate critical thinking in the design, analysis and review of SCOPE studies.

• The strength of SCOPE designs is influenced by exposure transiency, outcome abruptness, rapidity of the exposure-outcome association and degree of potential time-related issues.

BOX 1 What's in a Name?

We encourage use of Self-controlled Crossover Observational PharmacoEpidemiology (SCOPE) to describe all observational pharmacoepidemiologic applications of selfcontrolled study designs. SCOPE is a comprehensive label that clearly identifies the nature of the group of study designs, with a simple acronym to facilitate discussion. All designs use the patient as their own control (self-controlled), with a crossover analysis. Although it can be argued that "self-controlled" and "crossover" are redundant, including both speaks more broadly to other domains, including experimental research. The established benefits of crossover trials (eg, control for time-invariant within-person confounding) can thus be readily translated to the observational setting. The word "observational" also clarifies that the design is not experimental. Finally, the addition of "pharmacoepidemiology" helps to clarify the unique features that we review and are relevant when studying drugs that may not be as easily translated to non-drug exposures in the broader field of epidemiology. 11 Our intention is to create a common language to help clarify the distinction between biological (pharmacological) truths, and the phenomenon we wish to measure using imperfect measurement in the real-world. For example, we encourage induction period be used exclusively based on what is known based on pharmacology, and induction window as the investigator's window of observation that is earmarked for induction in the study. and pharmacodynamics (what the drug does to the body). Drug administration, absorption, distribution, metabolism and elimination influence the speed, intensity and duration of drug action. Therefore, pharmacologic reasoning about plausible temporal relations between causes and effects, and hypothesized durations of induction periods and carry-over effects, influence the choice of windows of observation. The clinical crossover trial design is often used to estimate these parameters. To exemplify pharmacologic reasoning, we use the case of fluoxetine, a selective serotonin reuptake inhibitor indicated for the treatment of depression. Below, we define induction, effect, and carry-over periods based on biological truths.

In an individual patient, the induction period is the length of time required for a drug to yield a detectable causal effect (either clinical benefit or adverse outcome). This often varies from one patient to the next and can make the population impact of a drug difficult to estimate. The induction period in a population is defined as the shortest individual induction period, labeled the minimum induction period.

According to clinical trial evidence, the minimum induction period for fluoxetine's clinical benefit in major depressive disorder (ie, improvement in depressive symptoms like suicide ideation or lack of appetite) is 2-4 weeks. 12 However, some patients who eventually report clinical benefit from fluoxetine do not achieve a full effect until as many as 8 weeks of continuous therapy. 13 In contrast, the minimum induction period from first intake of fluoxetine to other common adverse outcomes, such as gastrointestinal bleeding, insomnia, nervousness or sexual dysfunction, typically occur within 2 weeks of treatment F I G U R E 1 Simple representation of Self-controlled Crossover Observational PharmacoEpidemiologic (SCOPE) study design figures using the recommended language. For simplicity, we depict a point exposure that is administered as a single dose, such as an annual influenza vaccination. Similarly, only the main observation windows of interest (focal and referent) are depicted, yet transition (induction, lag, and washout) windows and other boundaries (eg, age groups) often need to be considered. (A) Outcome-anchored designs are typically uni-directional in pharmacoepidemiology (as depicted), with referent window defined only prior to the outcome-anchor. (B) Exposure-anchored designs are typically bi-direction meaning that referent windows before and after the exposure-anchored focal window are considered [Colour figure can be viewed at wileyonlinelibrary.com] initiation. 12 Understanding the minimum and maximum induction periods for the specific drug exposure-outcome relationship under investigation is critical to the design of a SCOPE study.

The effect period is the period of time within which it is biologically plausible that the drug causes the outcome. The effect period immediately follows the induction period and extends from the first to the last outcome attributable to the exposure. The effect period is only meaningful in the context of a specific exposure and outcome that the "effect" relates to. The carry-over period is the length of time when residual pharmacodynamic effects of the exposure occur after the effect period of primary interest. In a clinical crossover trial, the carry-over effects are defined as the residual effects from exposure in the first treatment phase that are carried over to, and impact, the causal effect estimates following treatment initiation in the second treatment phase. 14 Carry-over effects are mitigated by selecting a sufficiently long washout w. For example, fluoxetine has an elimination half-life of 4-6 days, and some of its active metabolites even longer.

Pharmacologic methods such as dose-response and dose-titration studies can be used to determine the washout window that is most appropriate for an experimental crossover clinical trial of the exposure of interest. 14 For drugs that follow a first-order pharmacokinetic model, a minimum washout window of five times the elimination halflife of drug from plasma levels for a 97% elimination from the body is recommended. 15 However, other factors such as tissue binding and other physiologic carryover effects may last longer and be considered when selecting the washout window. 16 The next section walks through the epidemic curve and symmetry analysis to help guide the selection of observation windows in a SCOPE study.

Epidemiology is the study of the occurrence, distribution and determinants of human health-related outcomes in specified populations. 17 The epidemic curve, and related "induction curve" at the population level, can be particularly informative in defining induction, effect and washout windows in the design of a new SCOPE study. An epidemic curve provides a graphical display of the number of incident cases in an outbreak of illness plotted over time. The shape of the curve helps to inform hypotheses about the nature of the disease. The epidemic curve is most commonly used to study the course of infectious diseases, yet can inform SCOPE studies by plotting the number of incident cases relative to the start of drug exposure. For example, the self-controlled case series (SCCS) design was motivated by an observation that risk for acute meningitis following Measles Mumps Rubella (MMR) vaccine exposure was elevated 15 to 35 days after immunization, Figure 2 . 18 In addition to epidemic curves, sequence symmetry analysis (SSA, Box 3), may be strategically employed when little is known about the true phases of the phenomenon under study. SSA can also screen for multiple associations simultaneously. In an extreme recent analysis, more than 200 billion sequences and 3 million hypotheses were screened. 19 What SSA lacks is a formal framework for handling timedependent confounders or flexible handling of exposure time. Therefore, SSA may be considered a great choice for signal detection and hypothesis-screening that can be further tested using a SCOPE design.

All SCOPE designs are case-only designs at their core. Extensions may be considered hybrids as they include non-case comparisons. Each SCOPE design uses different strategies for sampling comparison time.

At the heart of each approach is the comparison of an observed frequency during focal window(s) of observation to an observed frequency during referent window(s) of observation. The observed frequency (outcome incidence or odds of exposure) during a focal (hypothesized exposure-outcome risk) window is compared with the

All studies start with a single point in time from which all design features relate. This point in time is commonly referred to the "index date" in cohort studies, and "outcome date" in case-control studies. In an effort to define a common language across SCOPE studies, we use the term "anchor" to differentiate between SCOPE studies that define the primary windows of interest based on the outcome (outcome-anchored) or exposure (exposure-anchored).

We thus encourage "outcome-anchored" (case-crossover We also discourage the use of "prospective" and "retrospective;" "cohort" and "case-control;" or "forward looking"

and "backward looking." Our motivation relates to the tendency to consider anything "retrospective" as inferior to "prospective;" and thus also the case-control design as inferior to the cohort study design. 37 This black and white mentality hampers the value of a well-designed observational study. Although we and many others have previously categorized SCOPE designs as "prospective" and "retrospective," or "cohort" and "case-control;" or even "forward looking" and "backward looking;" we now encourage the adoption of "outcome-anchored" and "exposure-anchored" to differentiate between the main two groupings of SCOPE designs.

Our recommendation is also consistent with the STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) guidance document that refrains from using the terms "prospective" and "retrospective." 38 "usual" outcome incidence, or the "usual" odds of exposure, in referent windows that are outside the focal and transition (induction, lag and washout) windows. Depending on the specific design, experience (person-time) from non-cases can contribute to effect estimates to combat time-trends or to adjust for time varying confounding. Nonetheless, at their core, all SCOPE designs compare observed frequencies (exposure or outcome) within predefined focal and referent windows, conditioned on the individual patient.

In the following sections, we adopt the proposed language in Table 1 to describe and compare SCOPE designs, Table 2 . Readers are encouraged to consult a recent comparative summary for considerations of the strengths and limitations of SCOPE designs in the broader field of epidemiology. 20, 21 We appreciate that variation in terminology will persist in practice, yet believe that the common terminology proposed here may help in better understanding the similarities and subtle differences among alternative SCOPE approaches.

T A B L E 1 Distinction between causal and observational worlds of knowledge and proposed common language for time-related phases "Truth" " Study" or Estimation

The phenomenon being studied is… True relation (causation) Estimated association (observation)

…which is produced by… Biology (nature) Measurement (investigator's design)

…which is limited or modified by… Modifiers in human populations (pharmacology)

What is measurable and measured (data available) within the constraints of a healthcare (structural) system …which form the bases for defining… Hypothesized phases of the cause-effect process

Time-related phases of the phenomenon

Effect period (period of time when it is biologically plausible that exposure causes outcome, that is, exposureoutcome effect period)

Focal window (window of interest when it is hypothesized to be biologically plausible that exposure causes outcome; temporally linked to a study design anchor) a,b

Baseline Baseline period (period where outcome risk is determined by factors other than exposure)

Referent window (window outside the focal and transition windows chosen to estimate baseline risk when people are at biological risk for the outcome due to factors other than the exposure of interest) a

Transition time:

Induction period (period of time after a person is exposed to the drug and before the outcome is biologically possible)

Induction window (window of observation hypothesized to capture the induction period; can be modeled as a separate referent window or excluded)

Not applicable (healthcare system or data issues and thus only contextual for a SCOPE study) c

Lag window (healthcare system issue that increases or decreases exposure or outcome; can be adjusted for or excluded) c Carryover Carry-over effect period (period of time after stopping the drug until drug effects on outcome risk are gone) a Washout window (time window during which individual variation in carry-over effect is thought to be complete; can be modeled as a separate referent window or excluded) a Ranges between individuals, can be estimated based on population-based incidence curves; based on current knowledge of the effect period. b Focal window is proposed in SCOPE studies as the "suspected" window of interest. Although exposure-outcome risk window is more explicit, it may be technically more challenging when there is no biology to support an association, yet a safety signal is under investigation. "Hypothesized exposureoutcome risk window" could be used, yet lengthy. "Focal" does not give value to effect, is short, and thus also strategic for inclusion in study figures and tables.

c Healthcare system issues are not biological, rather structural issues that impact drug exposure (eg, healthy vaccine effect). 

Outcome-anchored designs are best thought of as "trigger" designs since they are suited to the study of transient exposures and abrupt onset outcomes, Figure 3 . In pharmacoepidemiology, outcomeanchored designs typically only consider referent windows prior to the outcome-focal window. This is an important consideration because the probability of drug exposure often changes following an outcome. However, some outcomes, like an unknown adverse effect, would not change exposure probability and thus a bidirectional design with reference windows pre-and post-outcome can be considered. Understanding the study context and local healthcare system and norms of data under study are critical in design of any SCOPE study. In other fields, such as environmental epidemiology that consider ambient environmental exposures, referent windows before and after the outcome-focal window are common. 22 

The CCO design was originally developed through an interview study investigating the relationship between myocardial infarction (MI) and various acute exposures. 5 The concept that control-person selection bias could be avoided if cases served as their own controls motivated the gradual development of the design. In one of the first CCO studies, participant interviews considered medications, illicit drugs, alcohol, coffee, smoking, extreme exertion, sexual activity, anger and bereavement in the hours, days and weeks before the onset of MI. 23 Questions were structured in several ways to explore different durations of effect periods, ranging from minutes to days between potential triggers (causal exposures) and MI onset.

The basic research question was: "Did anything unusual happen just before?" For example, among 3882 MI patients interviewed, 9 had been exposed to marijuana (now commonly referred to as cannabis) within the hour before MI symptoms, and 3 in the preceding hour. 23 Assuming that cannabis' effect on MI risk dissipated within 1 hour, the 60 minutes before MI onset was chosen as the focal window and the preceding hour (60 to 120 minutes before MI), the referent window. This yielded a baseline (observation during the referent window) of 3 exposed patients, and thus produced a relative risk estimate of 3 (9 patients exposed in the focal window divided by 3 patients exposed in the referent window).

Since the maximum induction period was unknown, it was possible that 1 or 2 of the 3 MIs within the 60-to-120 minutes prior to MI were triggered by cannabis. Therefore, use of referent windows further away from the outcome were considered. In total, 25 patients reported cannabis use within the 2 to 24 hours before MI. The expected number of exposed in the 2 to 24 hours before MI observation window was thus closer to 1 per hour (25/22) than the estimate of 3 per hour from the 60-to-120-minute window prior to MI.

A much larger sample of referent windows was obtained by asking patients about their usual frequency of cannabis in preceding days, weeks and months. 23 In this analysis, each patient's data were treated as if they were an n-of-1 study and a stratified analysis was performed with one patient per stratum; that is, the analysis was conditioned on the individual patient. The Mantel-Haenszel estimates of the cannabis exposure odds ratios (OR) were 4.8 (95% confidence interval [CI]: 2.9, 9.5) and 1.7 (95% CI: 0.6, 5.1) for the first and second hour before MI onset compared to other reported person-time. The paper concluded that the risk of MI increased within an hour of cannabis (marijuana) exposure, and then dissipated. 23 A key distinguishing feature of this broader analysis is that it is conditioned on the individual patient. CCO analytical options include Mantel-Haenszel estimates, conditional logistic and Poisson regression. 21 

The prescription sequence symmetry analysis (SSA) was introduced in 1996 as a screening tool for unknown, unsuspected associations in large datasets. 39 The first paper examined whether cardiovascular medication trig- 

The CTC design is an extension of the CCO that includes matched risk-set sampled controls (ie, non-case at the time of sampling) as well as cases in the analysis. The CTC was proposed to adjust for population level exposure time-trends in the CCO design. 6 24 We focus on exposure time trends here, yet diagnostic trends in defining the outcome can also be adjusted using CTC.

Population level exposure time trends are especially pronounced in new-to-market medical products. 25 In a CCO evaluating the relationship between aripiprazole and MI, the analysis was repeated every quarter for 4 years, starting 6 months after aripiprazole entered the market. 25 In the first observation window following market entry, the estimated OR CCO(cases) was 2.7. Over time, the estimated OR declined, with OR CCO(cases) = 1.4 by the 15th quarterly assessment. The pattern of estimated ORs in matched (age, sex, calendar time) non-cases ran parallel to the estimates in cases; OR CCO(non-cases) = 3.1 in the first referent window and declined to 1.5 by the 15th quarterly assessment.

After adjusting the CCO estimates in each observation window with the estimated exposure time-trend in matched non-cases, the OR CTC were consistently null. Until the incidence of exposure reaches a steady state in the population, CCO estimates may be biased by population trends in exposure probability, and thus a CTC design is more appropriate.

The CCTC design is implemented in the same way as the CTC design;

however, non-cases are sampled exclusively from future cases. 7, 26 Because future cases are at-risk for the event during historical person-time, they are eligible to be sampled as non-cases in the CCTC design. This is true in the CTC, with the distinction here that the CCTC only considers future cases. CCTC may be the only option if

The self-controlled risk interval (SCRI) and exposure crossover designs are other types of exposure-anchored designs.

In particular, the SCRI is a variant of the self-controlled case series (SCCS) that is more restrictive about observation time. To our knowledge, the name was first used by Lee et al (2011) in an application of vaccine safety. 43 SCRI is indexed on exposure, with focal and referent windows defined in relation to exposure. The analysis estimates the relative incidence during the focal (exposure) window to the referent (unexposed) window(s) using only cases identified in either window. 18 The design may be "bi-directional" with two referent windows, one before and one after the focal Figure 3 , maximizing the potential for cross-over time between windows of exposure and non-exposure. Only comparisons that are discordant (either exposed in the focal or referent window, yet not both) contribute to the analysis. Chronic (persistent) drug use leads to a considerable amount of concordant exposure, that is, subjects who are exposed during the focal window and all referent windows, and thus do not contribute to the overall effect estimate. Chronic (persistent) use can also induce hypersensitivity towards exposure misclassification. 35 Care must be taken to remove the observation window immediately prior to vaccination (lag window) otherwise estimates of relative incidence will be inflated. We use the term period to refer to time-related biological phases of the true cause-effect process (summarized under pharmacology), and the term window to refer to investigators' choices of time-related intervals in which to observe phenomena.

Pharmacology may be thought of studying drug effects in the causal world (truth), whereas epidemiology may be considered studying drug effects in the observational world (study). Reducing our fallibility is the immediate purpose of this document. The long-term goal is to reduce imperfections in the evidence.

Mina Tadrous

Jesper Hallas

Guidelines for good pharmacoepidemiology practice (GPP): guidelines for good pharmacoepidemiology practice

A review of uses of health care utilization databases for epidemiologic research on therapeutics

Reliable assessment of the effects of treatment on mortality and major morbidity, II: observational studies

Sensitivity analysis and external adjustment for unmeasured confounders in epidemiologic database studies of therapeutics

The case-crossover design: a method for studying transient effects on the risk of acute events

The case-time-control design

Future cases as present controls to adjust for exposure trend bias in case-only studies

Relative incidence estimation from case series for vaccine safety evaluation

Signal identification and evaluation for risk of febrile seizures in children following trivalent inactivated influenza vaccine in the Vaccine Safety Datalink Project

Diffusion of methodological innovation in pharmacoepidemiology: self-controlled study designs

Three worlds

Canadian Network for Mood and Anxiety Treatments (CANMAT) 2016 clinical guidelines for the management of adults with major depressive disorder: section 3. Pharmacological treatments

What constitutes an adequate antidepressant trial of fluoxetine?

Carryover bias in clinical investigations

Pharmacokinetics 101. Paediatr Child Health

Handbook of Bioequivalence Testing

Dictionary of Epidemiologyj

Tutorial in biostatistics: the self-controlled case series method

Hypothesis-free screening of large administrative databases for unsuspected drug-outcome associations

A vaccine study design selection framework for the postlicensure rapid immunization safety monitoring program

Analysis of observational selfmatched data to examine acute triggers of outcome events with abrupt onset

Case-crossover analysis of air pollution health effects: a systematic review of methodology and application

Triggering myocardial infarction by marijuana

Handbook of Statistical Methods for Case-Control Studies

First-wave" bias when conducting active safety monitoring of newly marketed medications with outcome-indexed self-controlled designs

Case-crossover studies of therapeutics: design approaches to addressing time-varying prognosis in elderly populations

Age, antipsychotics, and the risk of ischemic stroke in the veterans health administration

Risk of aseptic meningitis after measles, mumps, and rubella vaccine in UK children

Self controlled case series methods: an alternative to standard epidemiological study designs

Self-controlled case series methodology

Case series analysis for censored, perturbed, or curtailed post-event exposures

Self-controlled case series analysis with eventdependent observation periods

Sample sizes for selfcontrolled case series studies

Evaluating efficiency and statistical power of self-controlled case series and self-controlled risk interval designs in vaccine safety

Use of self-controlled designs in pharmacoepidemiology

ROBINS-I: a tool for assessing risk of bias in non-randomised studies of interventions

Prospective versus retrospective: what's in a name?

Strengthening the Reporting of Observational Studies in Epidemiology (STROBE): Explanation and Elaboration

Evidence of depression provoked by cardiovascular medication: a prescription sequence symmetry analysis

The validity of sequence symmetry analysis (SSA) for adverse drug reaction signal detection

Adverse events with use of antiepileptic drugs: a prescription and event symmetry analysis

Sequence symmetry analysis in pharmacovigilance and pharmacoepidemiologic studies

H1N1 and seasonal influenza vaccine safety in the vaccine safety datalink project

Quantifying the impact of time-varying baseline risk adjustment in the self-controlled risk interval design: SCRI design with time-varying baseline risks

The exposure-crossover design is a new method for studying sustained changes in recurrent events

Physician warnings in psychiatry and the risk of road trauma: an exposure crossover study

Statistical analysis of correlated data using generalized estimating equations: an orientation

Pregnancy and the risk of a traffic crash

Causation of bias: the episcope

How to cite this article: Cadarette SM, Maclure M

Control yourself: ISPE-endorsed guidance in the application of self-controlled study designs in pharmacoepidemiology