Case-Time-Control Design
- Introduced by
- Samy Suissa; Sander Greenland
- Year
- 1995
- Type
- Self-controlled observational design with a time-trend control series
- Family
- Pharmacoepidemiology / case-only designs
What does it do?
The case-time-control design is a pharmacoepidemiologic study design that repairs a specific weakness of the case-crossover study: bias from a secular trend in exposure.
In a case-crossover analysis each case acts as their own control, comparing exposure in a short hazard window just before the event to exposure in earlier reference windows, which automatically removes all fixed, time-invariant confounders. But if the prevalence of exposure is rising or falling over calendar time for reasons unrelated to the outcome, this within-person comparison is biased. Samy Suissa's 1995 design adds a separate control series, analyzed the same way, to estimate that pure time trend; dividing the case-crossover odds ratio by the control odds ratio cancels the trend and leaves the exposure effect. Sander Greenland's 1996 analysis clarified the assumptions: the correction works only if the controls share the same exposure trend and there is no within-subject confounder, and it can introduce new bias if those conditions fail.
The intuition behind the method
Suppose you study whether starting a drug triggers a sudden adverse event. The case-crossover idea is elegant: for each patient who had the event, compare whether they were on the drug right before it versus during an earlier period. Because it is the same person, anything stable about them — genetics, chronic conditions, baseline habits — cancels out. The catch is calendar time. If the drug was simply being prescribed more and more over the years, patients are more likely to be exposed in the recent hazard window than in an older reference window for reasons that have nothing to do with the event, and the case-crossover wrongly reads that upward drift as risk. The fix is to look at people who did not have the event over the same two periods: any apparent within-person change in their exposure must be the pure time trend. You then subtract that trend (on the log-odds scale, by dividing the odds ratios) from the case estimate, leaving the part attributable to the event itself.
History
The case-crossover design, introduced by Malcolm Maclure in 1991, gave epidemiologists a self-controlled way to study transient triggers of acute events, but it remained vulnerable to time trends in exposure. Samy Suissa proposed the case-time-control design in his 1995 Epidemiology paper, adding a control series to estimate and remove that trend and illustrating it with a study of inhaled beta-agonists and fatal or near-fatal asthma. Sander Greenland's 1996 Epidemiology paper provided the critical theoretical scrutiny, showing precisely when the design corrects bias and when the time controls can introduce confounding, especially in the presence of within-subject confounders like disease severity. The exchange shaped subsequent self-controlled methodology, including refinements such as the case-case-time-control design and connections to the self-controlled case series.
When to use it
Use the case-time-control design for abrupt, transient effects of an intermittent exposure on acute events — the classic case-crossover setting — when you also suspect that exposure prevalence is changing over calendar time and would bias the within-person comparison. It is well suited to pharmacoepidemiology, for example studying whether starting or escalating a drug triggers an acute event when prescribing patterns are trending. It requires that exposure be measurable at two or more points in time within each subject and that a comparable control series sharing the exposure trend be available. Avoid it when the suspected confounding is time-varying and tied to disease severity (confounding by indication), because the time controls cannot remove that and may worsen bias; in such cases consider the self-controlled case series, active-comparator new-user cohorts, or negative-control designs instead.
Limitations and considerations
- Cannot remove within-subject (time-varying) confounders such as changing disease severity; confounding by indication remains a threat.
- As Greenland showed, using time controls can introduce new confounding if the controls do not share the cases' exposure trend or if a within-subject confounder is present.
- Depends on a valid, comparable control series and on correct specification of hazard and reference windows.
- Suited only to acute, transient effects of intermittent exposures; it is inappropriate for chronic exposures or outcomes with long induction periods.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Removes all fixed, time-invariant confounding by using each subject as their own control, inheriting the core strength of the case-crossover design.
- Additionally corrects for secular time trends in exposure, the main weakness of the plain case-crossover analysis.
- Requires no measurement of the trend-inducing factor itself, since the control series estimates the trend directly from observed exposure changes.
- Fits as a single conditional logistic regression with a case-by-exposure interaction, giving a familiar and tractable estimation route.
Applications
- Studying whether initiation or escalation of a medication triggers an acute adverse event when prescribing is trending over time.
- Re-analyzing case-crossover studies suspected of bias from secular changes in exposure prevalence.
- Investigating transient triggers of acute cardiovascular or respiratory events with intermittent exposures.
- Pharmacoepidemiologic safety signals where a randomized trial is infeasible and time-invariant confounding is a major concern.
- Methodological comparisons against the self-controlled case series and active-comparator cohort designs for trigger questions.
How it works
How it works
The pipeline begins with a standard case-crossover analysis among the cases. For each case, exposure during the hazard window immediately preceding the event is compared with exposure during one or more reference windows for the same person. Only discordant person-periods (exposed in one window, unexposed in the other) carry information, and the case-crossover odds ratio is the ratio of the two kinds of discordance. This estimate is automatically free of all time-invariant confounding but is contaminated by any systematic trend in exposure over time.
Formula
\widehat{OR}_{CC} = \frac{\sum_i \mathbb{1}[\text{exposed hazard},\,\text{unexposed ref}]}{\sum_i \mathbb{1}[\text{unexposed hazard},\,\text{exposed ref}]}The defining step is to assemble a control series — people who did not experience the event — and apply the identical within-subject window comparison to them. Because controls had no event, any discordance between their hazard-equivalent and reference windows reflects only the background change in exposure over calendar time, not outcome-related risk. The control odds ratio therefore estimates the secular exposure trend that biases the case-crossover analysis, using the same self-controlled logic so that it is measured on a comparable scale.
Formula
\widehat{OR}_{ctrl} = \frac{b_{10}}{b_{01}}\ \ \text{(discordant pairs among controls over the same windows)}The case-time-control estimate is then obtained by dividing the case-crossover odds ratio by the control odds ratio. On the log scale this subtracts the time-trend log-odds from the case log-odds, so the residual reflects the excess within-person change in exposure that is specific to the cases — the exposure effect net of the secular trend. Suissa showed this is exactly the interaction between case status and exposure in a conditional logistic model, which is why the design is fit as an ordinary case-control comparison of within-subject exposure changes.
Formula
\widehat{OR}_{CTC} = \frac{\widehat{OR}_{CC}}{\widehat{OR}_{ctrl}} = \exp(\beta_{\text{case}} - \beta_{\text{time}})In practice the whole design is estimated with conditional logistic regression stratified on the individual, with terms for exposure and for the interaction between exposure and case status. The main exposure coefficient captures the common time trend, and the case-by-exposure interaction coefficient is the trend-adjusted exposure effect of interest. This regression formulation accommodates multiple reference windows, matched controls, and additional time-varying covariates, and yields valid confidence intervals from the conditional likelihood.
Formula
\log\frac{\Pr(Y_{it}=1)}{1-\Pr(Y_{it}=1)} = \alpha_i + \beta\,X_{it} + \gamma\,(\text{case}_i \times X_{it})Finally, the analyst must check the assumptions that make the correction valid, following Greenland. The control series must experience the same exposure trend as the cases, and there must be no within-subject (time-varying) confounder such as worsening disease severity that changes exposure differently for cases. If a time-varying confounder like confounding by indication is present, the time controls do not remove it and can even introduce new bias, so the case-time-control estimate is trustworthy only when these conditions are argued to hold and ideally supported by sensitivity analysis.
Formula
\widehat{OR}_{CTC} \to OR_{\text{true}}\ \text{iff}\ \text{controls share the exposure trend and have no within-subject confounder}Common pitfalls
- Assuming the control series experiences the same exposure trend as cases without checking, which invalidates the trend correction.
- Applying the design when confounding by indication (a time-varying severity-driven confounder) is the real concern, where time controls fail and may add bias.
- Choosing hazard or reference windows that misalign with the exposure's induction and persistence, distorting the within-person contrasts.
- Interpreting the main exposure coefficient as the effect, when the trend-adjusted effect is the case-by-exposure interaction term.
Frequently asked questions
How does the case-time-control design differ from a plain case-crossover study?
A case-crossover study uses only cases, each serving as their own control across hazard and reference windows, which removes time-invariant confounding but is biased by any secular trend in exposure. The case-time-control design adds a separate control series analyzed the same way to estimate that trend, then divides the case-crossover odds ratio by the control odds ratio to remove it. Equivalently, it is the case-by-exposure interaction in a conditional logistic model. The result is a trend-adjusted effect estimate, valid under stronger assumptions than the case-crossover alone.
When can the design make things worse rather than better?
Greenland's 1996 paper showed that the time-control correction is valid only if the controls share the same exposure trend as the cases and there is no within-subject (time-varying) confounder. If a time-varying confounder such as worsening disease severity drives both the exposure change and the event — classic confounding by indication — the control series does not capture it, and adjusting by the controls can introduce new bias. So the design fixes secular trends but cannot fix severity-driven confounding, and using it blindly in that situation can degrade the estimate.
How is it related to the self-controlled case series?
Both are self-controlled designs that use within-person comparisons to eliminate fixed confounders, and both target transient effects of intermittent exposures. The case-time-control design extends the case-crossover by borrowing an external control series to net out exposure time trends, whereas the self-controlled case series models the within-person rate of events across exposed and unexposed time using only cases and an explicit time-trend term. When time trends are the concern, both can address them; the SCCS handles trends through its model specification, while the case-time-control handles them through the added control group.
Related methods
Sources
Suissa, S. (1995). The case-time-control design. Epidemiology, 6(3), 248-253.
Open DOI referenceCopied.Greenland, S. (1996). Confounding and exposure trends in case-crossover and case-time-control designs. Epidemiology, 7(3), 231-239.
Open DOI referenceCopied.