(296) Bias Amplification of Unobserved Confounding in Pharmacoepidemiological Studies: There Is No Free Lunch in Restricting the Sample to Those with a Particular Drug Indication

Publication Number: 970

Background:
Pharmacoepidemiological studies are often biased by confounding by indication when utilizing real-world data. To circumvent such bias, researchers often choose to restrict their study sample to individuals with a particular drug indication (i.e., indication-based sampling). This strategy may, however, be problematic.

Objectives:
We here showcase how indication-based sampling may exaggerate, rather than alleviate, bias in studies examining drug effects.

Methods:
We describe bias amplification under indication-based sampling, which we highlight using simulations of varying levels of confounding and applied examples, overall and in relation to effect-heterogeneity.

Results:
If we allow the indication to vary freely, the indication will explain some of the differences in drug use. However, if we constrain the indication, a larger share of the variation in drug use must be due to some other factor (e.g., an unmeasured confounder). In simulations, we find that indication-based sampling is more biased than alternative approaches (e.g., regression adjustment) to control for confounding by indication. This suggests that indication-based sampling is often not warranted, especially since the approach also depletes the sample size. Stratum-specific effects are, however, important to consider when effect-heterogeneity can be assumed or the research question under study concerns the drug effect in a specific population (e.g., it is relevant to stratify by/restrict to individuals with depression in studies of the effect of antidepressants on suicide). Nonetheless, it can be detrimental to perform indication-based sampling when there is no reason to expect effect-heterogeneity (e.g., it is not warranted to stratify by/restrict to individuals with epilepsy in studies of antiseizure medication teratogenicity).

Conclusions:
Restricting a study sample to a particular drug indication should not be considered unbiased or an optimal strategy to address confounding by indication. In fact, indication-based sampling is generally more biased than alternative modes of analysis. We suggest that future studies using real-world data stay wary of bias amplification when considering drug indication, in spite of recent enthusiasm for active-comparators and target trial emulation.