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

https://doi.org/10.1016/j.jclinepi.2013.05.003Get rights and content

Abstract

Objectives

To introduce a new design that explores how an acute exposure might lead to a sustained change in the risk of a recurrent outcome.

Study Design and Setting

The exposure-crossover design uses self-matching to control within-person confounding due to genetics, personality, and all other stable patient characteristics. The design is demonstrated using population-based individual-level health data from Ontario, Canada, for three separate medical conditions (n > 100,000 for each) related to the risk of a motor vehicle crash (total outcomes, >2,000 for each).

Results

The exposure-crossover design yields numerical risk estimates during the baseline interval before an intervention, the induction interval immediately ahead of the intervention, and the subsequent interval after the intervention. Accompanying graphs summarize results, provide an intuitive display to readers, and show risk comparisons (absolute and relative). Self-matching increases statistical efficiency, reduces selection bias, and yields quantitative analyses. The design has potential limitations related to confounding, artifacts, pragmatics, survivor bias, statistical models, potential misunderstandings, and serendipity.

Conclusion

The exposure-crossover design may help in exploring selected questions in epidemiology science.

Introduction

Clinical epidemiology is sometimes chastised as the science of unfair comparisons [1], [2], [3]. As a consequence, observational studies usually endeavor to identify cases and controls that are reasonably similar so that inferences are not unduly slanted by hidden confounding [4], [5], [6]. One method for ensuring equivalence is randomization, although doing so typically requires substantial sample size and individual cooperation [7]. Alternative methods include regression modeling, propensity score stratification, individual pair matching, subgroup stratification, or other analytical methods for making separate patient groups appear similar [8], [9], [10]. None of these methods is ideal, and methodological work developing new designs remains a priority for future progress.

One major advance in clinical epidemiology was the development of the case-crossover design in 1991 [11]. The main strength of this repeated-exposure approach is to define each patient as their own control and explore the transient effects of a brief exposure on the onset of an acute outcome [12]. An early contribution from this design examined heart attack patients (n = 1,228) and identified that 54 patients had exercised in the hour before the onset, whereas only nine patients had exercised in the same hour 1 day before the onset, equivalent to a sixfold temporary increase in heart attack risk associated with exercise [13]. Scientists in subsequent years expanded on the case-crossover design with further theory, modeling, and practical applications [14], [15], [16], [17].

The purpose of this article is to introduce a new approach called the “exposure-crossover” design. Similar to other epidemiology methods, the intent of the design is to test for a potential link between exposure and outcome [18], [19]. The name is selected to convey a notion that each patient serves as their own control and undergoes observation during a time with an exposure and a time without an exposure. The name is also intended to be reminiscent of the case-crossover design, the epidemiologic approach that the exposure-crossover design most closely resembles. Some additional names that were considered but rejected include the exposed crossover design, sustained impact design, interventional analysis design, and individual-level interrupted time-series design.

Section snippets

Background

The exposure-crossover design was first inspired by the methods for examining large economic changes [20], [21]. For example, the 23% decrease on Monday October 19, 1987, in the US stock market is apparent when evaluated as a time-series graph (Fig. 1). Such time-series analytical studies are not limited to financial markets and have extended to several other nonmedical fields, for example, a study on the effect of contaminated milk leading to a sustained decrease in dairy product consumption

Case study 1

Consider the following pilot data based on a validation study. The purpose was to test the hypothesis that formal warnings about sexually transmitted infections might alter the subsequent risk of a motor vehicle crash of a patient. This simple hypothesis is implausible and serves to check for the lack of an association when no association would be anticipated. That is, patients who develop sexually transmitted infections might have additional risky behaviors related to lifestyle [41], [42].

Confounding

The inescapable limitation of the exposure-crossover design is that the intervention is not randomly assigned. Hence, a transient factor might cause the patient to both receive treatment and subsequently change. The intervention, therefore, might be a marker but not a mechanism for subsequent differences. Such within-person confounding can be controlled if sufficient data are available on relevant factors before and after the intervention, yet such details are rarely complete. Tracer analyses

Summary

The exposure-crossover design is essentially a before-and-after comparison. Just as an electron microscope shares similarities with an optical microscope, however, the new epidemiology design has distinct features (Table 1). The structure of the exposure-crossover design provides an opportunity for separating the observations before intervention into discrete baseline and induction intervals (each with their own insights). The statistical power further segments each interval into finer

Acknowledgments

The views expressed are those of the author and do not necessarily reflect the Ontario Ministry of Health or the Institute for Clinical Evaluative Sciences. The following individuals merit thanks for providing helpful comments: Peter Austin, Eric Ho, Andrew Lustig, John Staples, Therese Stukel, Deva Thiruchelvam, Robert Tibshirani, Jason Woodfine, Wu, and Christopher Yarnell.

References (64)

  • G.A. Satten et al.

    Sample size determination for pair-matched case-control studies where the goal is interval estimation of the odds ratio

    J Clin Epidemiol

    (1990)
  • M.H. Merson et al.

    The history and challenge of HIV prevention

    Lancet

    (2008)
  • P. Martínez-Camblor et al.

    Repeated measures analysis for functional data

    Comput Stat Data Anal

    (2011)
  • D.A. Redelmeier et al.

    Interpretation and bias in case-crossover studies

    J Clin Epidemiol

    (1997)
  • C. Thus et al.

    Does alcohol protect against the formation of gallstones? A demonstration of protopathic bias

    J Clin Epidemiol

    (1991)
  • E. Loder et al.

    Registration of observational studies

    BMJ

    (2010)
  • M.B. Bracken

    Preregistration of epidemiology protocols

    Epidemiology

    (2011)
  • A. Dechartres et al.

    Single-center trials show larger treatment effects than multicenter trials: evidence from a meta-epidemiologic study

    Ann Intern Med

    (2011)
  • F. Dal-Pizzol et al.

    Functional disability 5 years after ARDS

    N Engl J Med

    (2011)
  • L. Gordis

    Epidemiology

    (2009)
  • P.C. Austin

    A critical appraisal of propensity-score matching in the medical literature between 1996 and 2003

    Stat med

    (2008)
  • G. Cortese et al.

    Flexible survival regression modelling

    Stat Methods Med Res

    (2010)
  • M. Maclure

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

    Am J Epidemiol

    (1991)
  • M. Maclure et al.

    Should we use a case-crossover design?

    Annu Rev Public Health

    (2000)
  • M.A. Mittleman et al.

    Triggering of acute myocardial infarction by heavy physical exertion—protection against triggering by regular exertion

    N Engl J Med

    (1993)
  • A. Dewanji et al.

    Choice of stratification in Poisson process analysis of recurrent event data with environmental covariates

    Stat Med

    (2002)
  • W. Lee et al.

    An easy-to-implement approach for analyzing case-control and case-only studies assuming gene-environment independence and Hardy-Weinberg equilibrium

    Stat Med

    (2010)
  • K.W. Mahaffey et al.

    The impact of postrandomization crossover of therapy in acute coronary syndromes

    Circ Cardiovasc Qual Outcomes

    (2011)
  • K.J. Rothman

    Epidemiology: an introduction

    (2002)
  • S. Bahng et al.

    Interactions of stock markets within the greater China economic bloc

    Glob Econ Revs

    (2004)
  • R. Masih et al.

    Common stochastic trends and the dynamic linkages driving European stock markets: evidence from pre-and post-October 1987 crash eras

    Eur J Finance

    (2004)
  • M.E. Smith et al.

    Sales loss determination in food contamination incidents: an application to milk bans in Hawaii

    Am J Agric Econ

    (1988)
  • Cited by (28)

    • Risks of Serious Injury with Testosterone Treatment

      2021, American Journal of Medicine
      Citation Excerpt :

      Statistical testing used longitudinal generalized estimating equations (GEE) to yield estimates of relative risk (similar to McNemar test adapted for the unequal durations of the baseline and subsequent intervals).67,68 Longitudinal GEE with a Poisson link function was also used to correct for possible temporal trends because all patients age forward in time.69,70 Findings were considered positive if significant results were obtained both before and after adjustment for time effects.

    • Association Between Burn Injury and Mental Illness among Burn Survivors: A Population-Based, Self-Matched, Longitudinal Cohort Study

      2017, Journal of the American College of Surgeons
      Citation Excerpt :

      The key strength of our study is its design; the exposure-crossover method facilitates comparisons of pre- and post-burn mental health rates within individuals. This approach minimizes confounding due to personality, genetics, socioeconomic status, and other stable characteristics that can be difficult to measure.8 In addition, use of administrative data facilitated capture of a large cohort of burn-injured individuals with complete longitudinal follow-up over time.

    • Weight loss surgery and subsequent emergency care use: A population-based cohort study

      2016, American Journal of Emergency Medicine
      Citation Excerpt :

      The analysis excluded emergency visits in the immediate perioperative interval defined as the 3 months before and 3 months after surgery to avoid confounding by surgery preparation and complications. This analytical approach, called the exposure crossover design, has been previously used to assess the effects of distinct medical interventions [22]. Self-matching is usually robust in reducing bias related to stable individual factors that might influence emergency care such as genetics, education, and personality [22,23].

    • Cross-linked survey analysis is an approach for separating cause and effect in survey research

      2015, Journal of Clinical Epidemiology
      Citation Excerpt :

      Cross-linked survey analysis investigates data under the assumption that the available survey responses are valid and important. Cross-linked survey analysis was first inspired by the exposure-crossover design, a method for examining sustained changes in the risk of recurrent events [3]. The main commonality of the two approaches is in linking large databases through individual identifiers to provide a before-and-after comparison after a defined exposure.

    View all citing articles on Scopus

    Conflicts of interest: The funding organizations had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the article. The author has no financial or personal relationships or affiliations that could influence the decisions and work on this article. The author (D.A.R.) had full access to all the data in the study, takes responsibility for the integrity of the data, and is accountable for the accuracy of the analysis.

    Funding: This project was supported by the Canada Research Chair in Medical Decision Sciences and the Canadian Institutes of Health Research.

    View full text