Considerations for outcome‐dependent biased sampling in health databases

Is There a Case for Case-Control Studies in the Exploration of Retrospective Data Sets?

“…In some situations, matching may actually introduce bias. 4,5 However, bias resulting from matching can be attenuated through adjustment at the analysis stage, provided that matching was based on patient characteristics at the time of (or prior to) exposure. 5 To make full use of the rich information that observational studies can provide, biostatisticians have recently started recommending that analyses be approached with a hypothetical target trial in mind.…”

“…6 Case-control studies involving data from administrative claims, registries, and electronic health records have been shown to overestimate the magnitude of association between a medication and the future risk of a condition (unassociated with the medication's initial indication) when the inclusion criteria and analysis methods deviate from those expected in a target clinical trial. 3,4,6,7 In 1 example, 6 failure to adjust for loss to follow-up, adjustment for variables measured at the index date (ie, the date of assessment of disease status, rather than at the time of exposure), and the inclusion of individuals prevalently using the medication at baseline were cited as reasons for the large discrepancy between estimates. These potential sources of bias are also present in the metformin-AMD study 1 published in this issue of JAMA Ophthalmology.…”

“…Rather, they can stem from the quality of the data, the choice of inclusion criteria, and paradoxically, methods intended to reduce confounding bias. In some situations, matching may actually introduce bias . However, bias resulting from matching can be attenuated through adjustment at the analysis stage, provided that matching was based on patient characteristics at the time of (or prior to) exposure …”