Professor Department of Public Health, Shizuoka Graduate University of Public Health, Japan
Background: Target trial emulation (TTE) is a causal inference framework in which inverse probability weighting (IPW) is commonly used to control confounding. In 2024, overlap weighting (OW) can be reportedly as an alternative method to enhance confounding control. This proposal was based on a single COVID-19 vaccine effectiveness study, and knowledge about the utility of OW in TTE beyond this work remains limited.
Objectives: To test the reproducibility of past findings on OW, through re-analysis of our previously published study on TTE.
Methods: We previously investigated and reported the effectiveness of an annual health checkup program in Japan on the incidence of diabetes and hypertension within a TTE framework; stabilized IPW (sIPW) was used for confounder adjustment. The study period was from April 2008 to March 2020. Depression was pre-specified as benchmarking a negative control outcome (NCO). Unexpectedly, the hazard ratio (HR) for the NCO was shown to be non-null. In the present study, we repeated the analysis, replacing the sIPW method with overlap weighting (OW), under the assumption that OW would reduce residual bias in the NCO analysis and bring the HR closer to the expected null value. We considered depression—our benchmarking NCO—as meeting the criteria for an ideal NCO, including U-comparability.
Results: The primary and NCO analysis respectively included 293,174 and 284,444 individuals. The HR for the composite of incident diabetes and hypertension was 0.902 (95% confidence interval [CI], 0.888 to 0.916) in the original analysis and 0.874 (95% CI, 0.862 to 0.886) in the revised analysis with OW. In the NCO investigation, univariate analysis showed that health checkups were associated with a higher risk of depression (HR, 1.108: 95% CI, 1.058 to 1.161). In the adjusted analyses, the HR in the original analysis was 1.047 (95% CI, 1.022 to 1.073) and 1.057 (95% CI, 1.033 to 1.079) in the analysis using OW. This result deviated slightly from the null relative to the estimates from sIPW analysis.
Conclusions: In our case example, both sIPW and OW were useful for controlling confounding. However, compared with sIPW, OW may have limited utility in the context of reducing residual bias, according to our NCO analysis. Our findings suggest that no single weighting method fits all TTE studies.
