Future-case control crossover analysis for adjusting bias in case crossover studies

Control crossover analysis with “future case” patients

Case-time-control analysis and case-case-time-control analysis were developed to adjust for exposure time trends in case crossover studies.45 These methods include a control group for a control crossover analysis conducted analogously to the case crossover analysis, for the purpose of assessing the exposure time trends. The main difference between the two control crossover designs is the source of the controls; the case-time-control design includes non-case external controls, whereas the case-case-time-control design uses “future cases.” Specifically, the future cases are selected from the original cases (or “current cases”) using risk set sampling, such as matching by age, sex, or other characteristics, to make the risk for events during the person time of the two groups similar. In this design, the control crossover analysis basically shifts both the hazard and the reference windows backwards by a specified time period from the original event date. Because the index date transitions from the original event date to an earlier one, it appears as if the patients are to experience the events subsequently, in the near future, and they are therefore referred to as “future cases.” Because both the current cases and the future cases experience the outcome of interest, the exposure time trends at individual level are assumed to be similar. Population level time trends can also be minimised, as we usually consider matching the calendar time when sampling control populations of future cases. Figure 1 schematically presents case crossover analysis, different types of exposure time trend bias, and how control crossover analyses minimise exposure time trend effects.

Fig 1

Schematic presentation of different case crossover designs and the exposure time trend bias

• Download figure
• Open in new tab
• Download powerpoint

Specifically, the method assumes that the odds ratio observed from the case crossover analysis is influenced by both the causal effect and the time trend effect (ie, OR_case=OR_causal×OR_{time trend}), and that the odds ratio observed from the future-case control crossover analysis estimates the time trend effect (OR_control=OR_{time trend}). The causal effect odds ratio can then be derived by dividing the odds ratios from the case crossover analysis and the control crossover analysis (ie, OR_causal=OR_case÷OR_{time trend}=OR_case÷OR_control). Given that both current and future cases experience the events, we can assume that they share a comparable individual level time trend of exposure. Moreover, this approach can also mitigate population level trend bias, given that the index date for a future case is the same as the date of the matched current case, sharing identical calendar time spans.

Control crossover analysis with future case controls could be an effective approach to resolve protopathic bias, which derives from exposure time trends at the individual level. For example, Huang and colleagues assessed the association between recently raised anticholinergic burden and cardiovascular events.6 Because a clinician might prescribe drugs (such as anti-vertigo drugs for dizziness) for the symptoms of cardiovascular events before a confirmed diagnosis, the use of the drugs during hazard windows might appear elevated, compared with the reference windows. This individual level time trend could generate reverse causality between drug use and outcomes (ie, protopathic bias). Control crossover analysis of future cases enables us to estimate the time trend effect and subsequently derive adjusted odds ratios that take the time trend effect into account.

Unlike case-time-control analysis using non-case external controls, an advantage of using future cases for control crossover analysis is that the controls are sampled from the original cases. This eliminates concerns about availability of data for the controls, or about avoiding selection bias when controls do not accurately reflect the target population. The major limitation of this approach is that there are no standard criteria for an appropriate time gap between the event dates of current and future cases. Sensitivity analyses with various time intervals may be considered to examine the assumptions.