Suicide-related events in patients treated with antiepileptic drugs
Arana, A., Suissa, S., & Arellano, F. M. (2011). Suicide-related events in patients treated with antiepileptic drugs: Not an example of time-window bias. Epidemiology, 22(6), 876-7. DOI: 10.1097/EDE.0b013e31823198fc
Recently, Epidemiology published a brief report from Suissa and colleagues1 about time-window bias in case-control studies. This paper mentions that this bias is not uncommon, and refers to bias occurring in a recent case-control published by some of us on the risk of suicide associated with antiepileptic drugs.2 This was an incorrect attribution because the mentioned article used risk-set control sampling3 for the selection of the controls, and defined the end of the follow-up period for the controls (ie, the date when current use to antiepileptic drugs was assessed), as the date on which the case-event assigned to the control occurred2 precisely to ensure an equal time-window to measure exposure for cases and controls.1 In addition, and in contrast with the premises set in the definition of time-window bias, Arana et al defined current use of antiepileptic drugs as “the interval between the date of the prescription of an antiepileptic drug and 75 days after that date.” After that interval, as long as no other antiepileptic drug was prescribed, the patient was considered to be a previous user.
The confusion arose from the widely different findings when the data were analyzed as a cohort. In such analysis, the incidence rate of suicide-related events among patients with no epilepsy, depression, or bipolar disorder was 38 per hundred person-years (95% confidence interval [CI] = 26-54) for those who never used antiepileptic drugs and 48 (39-59) for those with any use of antiepileptic drugs. In contrast, the corresponding odds ratio by the case-control approach within this same cohort was 0.59 (95% CI = 0.35-0.98). Such discrepancies can be confusing and can mislead the reader in thinking that a certain bias, such as the described time-window bias, was present when in fact it was not. This was clarified in the discussion section where it was mentioned that relative risks were not calculated from the reported rates in the cohort analysis because they were subject to immortal time bias. As authors who represent both of these papers, we are glad to be able to provide this clarification.