(C-169) Revisiting Representativeness

Background: Representativeness has long been an important topic in epidemiologic and pharmacoepidemiologic research. Yet, the definition of representativeness remains somewhat elusive and ambiguous.

Objectives: To provide a clearer definition of representativeness.

Methods: We revisit representativeness within epidemiologic research by differentiating two key concepts: sample representativeness and estimate representativeness. Drawing on recent advances in understanding selection bias, we explore the relationship between sample representativeness, estimate representativeness, and two types of selection bias.

Results: Sample representativeness is defined as the condition when the study sample is a random sample of (or equivalent to) the target population, and no or only random attrition occurs, and hence the distribution of all participant characteristics in the analytic sample will mirror that in the target population in expectation. Estimate representativeness is defined as the scenario in which the effect estimate from the analytic sample is asymptotically unbiased for the true causal effect in the target population. Sample non-representativeness under non-random sampling may lead to two types of selection bias—selection bias related to effect measure modifiers and selection bias related to colliders—with either yielding a biased effect estimate that leads to estimate non-representativeness.

Conclusions: A deeper understanding of “representativeness” is essential in pharmacoepidemiologic research. Although real-world data are often presumed to offer stronger representativeness than randomized trials, non-random sample selection can invalidate this assumption, potentially leading to non-representative estimates.