(273) Comparing the performance of nested case-control and case-crossover study designs for assessing adverse outcomes of concomitant drug use: a simulation study

Publication Number: 606

Background: Nested case-control studies are a pragmatic approach to assess the risk of adverse outcomes associated with concomitant drug use in a well-defined cohort. Conversely, case-crossover studies offer inherent adjustment for time-invariant confounding, though they carry strong assumptions. Assessing the impact of violating assumptions of both designs may inform appropriate study selection.

Objectives: To compare the robustness of nested case-control and case-crossover study designs to violation of their assumptions.

Methods: We generated 1,000 cohorts of 50,000 patients exposed to an object drug (oral anticoagulants) for a mean duration of 1 year and a precipitant drug (selective serotonin reuptake inhibitors) for a fixed length of 30 days, with follow-up limited to one year. Daily probability of the outcome, major bleeding, was generated as a function of the two exposures. The estimand, the OR of the outcome associated with concomitant use of both drugs, compared with object drug use only, had a theoretical value of 1.35. We conducted nested case-control and case-crossover analyses in all cohorts. In the base case simulation, no assumptions were violated. In alternative scenarios, we introduced the following: non-transient precipitant exposure (3 scenarios with mean durations of 60, 90, and 120 days), time trend in precipitant use (2 scenarios where 50% of exposure begins in the first and last 30% of follow-up, respectively), time-varying confounding (3 scenarios with a moderately positive, strongly positive, and moderately negative confounder, respectively), and residual time-invariant confounding (3 scenarios with a moderately positive, strongly positive, and moderately negative confounder, respectively).

Results: In the base case scenario, estimates from both analyses had minimal bias (OR 1.33 for both). In scenarios with non-transient exposure, there was some bias in the case-crossover analyses only (OR 1.48, 1.16, and 1.97). Estimates from the case-crossover analysis were also biased in scenarios with time-varying confounding (OR 1.54, 1.63, and 1.24) and time-trend in exposure (OR 0.92 and 1.91), while remaining unbiased in nested case-control analyses. However, nested case-control analyses were somewhat biased in scenarios with residual time-invariant confounding (OR 1.38, 1.42, and 1.30), while case-crossover analyses were unbiased. Finally, case-crossover analyses had higher empirical standard errors in all scenarios.

Conclusions: Nested case-control analyses appear more robust than case-crossover in most scenarios, though strong residual confounding leads to some bias. Nonetheless, greater bias may result from time-varying confounding in case-crossover analyses.