Bayesian Case-Crossover Design — Self-Matched Epidemiological Study with Bayesian Inference
Bayesian Case-Crossover Study Design · Also known as: Bayesian case-crossover, BCCO, Bayesian self-matched design, Bayesian within-person crossover
The Bayesian case-crossover design is a self-matched epidemiological method that estimates the transient effect of a time-varying exposure on the risk of an acute event. Each case serves as their own control, eliminating confounding by time-stable individual characteristics. Bayesian inference replaces or supplements the classical conditional logistic regression, enabling the incorporation of prior knowledge, more stable estimation in sparse data, and full uncertainty quantification via posterior distributions.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use the Bayesian case-crossover design when studying the short-term effect of a transient, time-varying exposure (air pollution, temperature, medication use, physical exertion) on an acute, well-defined event (myocardial infarction, asthma attack, injury). The self-matched structure is ideal when time-stable confounders (genetics, socioeconomic status, chronic disease burden) are a serious concern and individual-level longitudinal records are available. The Bayesian extension is particularly beneficial when events are rare and data are sparse, when informative priors from prior studies can stabilise estimates, or when full uncertainty quantification and probabilistic interpretation of results are required. Do not use it when the exposure itself is time-stable (e.g., a genetic variant or lifetime smoking status), when the event is chronic rather than acute, or when individual-level temporal data are unavailable and only aggregate ecological data exist.
Strengths & limitations
- Eliminates confounding by all time-stable individual characteristics without needing to measure them — a major advantage over conventional case-control designs.
- Bayesian inference provides full posterior uncertainty quantification and allows direct probability statements (e.g., Pr(OR > 1 | data)).
- Informative priors from previous studies can stabilise estimates when events are rare or samples are small.
- Hierarchical Bayesian extensions readily accommodate heterogeneity across sites, cities, or subgroups.
- Results are naturally communicated on the posterior distribution scale, which is more informative than a point estimate with confidence interval.
- Requires individual-level time-stamped exposure and event data — aggregate or cross-sectional data cannot support the design.
- The choice of control window strategy (time-stratified, bidirectional, etc.) can introduce bias if chosen inappropriately; no universally correct strategy exists.
- Prior specification requires justification; a poorly chosen informative prior can bias posteriors, and sensitivity analyses add analytic burden.
- MCMC computation can be slow for large datasets or complex hierarchical models; convergence must be verified.
- Cannot account for time-varying individual-level confounders that change within the study period (e.g., medication changes, disease progression).
Frequently asked
How is the Bayesian case-crossover different from the classical (frequentist) case-crossover?
The classical case-crossover uses conditional logistic regression and reports a point estimate with a frequentist confidence interval. The Bayesian version places a prior on the exposure effect and computes a full posterior distribution. Key practical differences: the Bayesian approach allows informative priors from prior studies, provides direct probability statements about the effect size, and handles sparse-data settings more gracefully through regularisation. Both approaches share the same self-matched data structure and control window logic.
Which referent window strategy should I choose?
Time-stratified referent selection — using fixed days of the same week within the same month and year as control periods — is generally recommended because it eliminates seasonal and long-term trends as sources of bias. Bidirectional and symmetric designs can introduce bias when exposure levels trend over time. Always justify your referent strategy in the methods section and consider a sensitivity analysis under an alternative strategy.
Can I use INLA instead of MCMC for Bayesian estimation?
Yes. Integrated nested Laplace approximation (INLA), implemented in the R-INLA package, provides fast approximate Bayesian inference and is well suited to the conditional logistic likelihood used in case-crossover models. For large multi-city datasets or complex hierarchical structures, INLA is often preferred over MCMC for its computational efficiency, though it involves additional approximation error that should be acknowledged.
What software is available for Bayesian case-crossover analysis?
Stan (via RStan or CmdStanR) and JAGS allow fully flexible MCMC-based Bayesian case-crossover models. R-INLA supports fast approximate Bayesian inference. The R packages 'survival' and 'gnm' can fit the classical conditional logistic likelihood, which can then be wrapped in a Bayesian framework. For environmental epidemiology specifically, the 'dlnm' package (distributed lag non-linear models) is often used alongside Bayesian pooling routines.
Does the self-matched design control for all confounding?
It controls for all time-stable (fixed) confounders — age, sex, genetic factors, chronic conditions, and socioeconomic status — because these do not vary between the hazard and control windows for the same individual. However, time-varying confounders that change within the study period (e.g., a new medication, a concurrent acute illness) are not controlled and must be handled by design restrictions or sensitivity analyses.
Sources
- Maclure, M. (1991). The case-crossover design: a method for studying transient effects on the risk of acute events. American Journal of Epidemiology, 133(2), 144–153. DOI: 10.1093/oxfordjournals.aje.a115853 ↗
- Janes, H., Sheppard, L., & Lumley, T. (2005). Case-crossover analyses of air pollution exposure data: referent selection strategies and their implications for bias. Epidemiology, 16(6), 717–726. DOI: 10.1097/01.ede.0000181315.18836.9d ↗
How to cite this page
ScholarGate. (2026, June 3). Bayesian Case-Crossover Study Design. ScholarGate. https://scholargate.app/en/epidemiology/bayesian-case-crossover-design
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Bayesian Hierarchical ModelBayesian↔ compare
- Case-crossover designEpidemiology↔ compare
- Self-Controlled Case SeriesSocial Epidemiology↔ compare