Background: Multi-site studies, including distributed networks, are becoming a mainstay of pharmacoepidemiologic research. While standard meta-analytic methods combine less precise site-specific estimates to generate a network-wide effect estimate, other target populations and subpopulations may also be clinically relevant. If there are effect measure modifiers (EMMs), network-wide estimates may not reflect effects in all populations of interest, including smaller participating sites.
Objectives: To demonstrate how standardization using inverse odds weights (IOW) can increase the precision of estimates within a specific site of a multi-site study, even in the presence of EMMs, using a real-world case study.
Methods: We created an artificial multi-site study in data from the United Kingdom’s Clinical Practice Research Datalink (CPRD) Aurum by dividing it into 11 regional “data partners” and assumed these partners could not share data on treatment, outcomes, or the timing of diagnoses or office visits. Our primary target population was the smallest region (#11, Northern Ireland). Within each region, we mimicked a previously-conducted CPRD study comparing metformin vs sulfonylurea initiation as first-line treatment for type 2 diabetes. Exposure was defined using an intention-to-treat approach and the outcome was all-cause mortality. We accounted for confounders with inverse probability of treatment weights (IPTW) and standardized each site to resemble region 11 with respect to EMMs by combining IPTW with IOW. Finally, we pooled standardized estimates of the 1-year risk difference (RD) with inverse variance weights and used the bootstrap to estimate standard errors and 95% confidence intervals (CIs).
Results: We identified 813,156 metformin initiators and 193,978 sulfonylurea initiators, with 18-21% initiating sulfonylureas across all regions. When analyzed as one data set, the IPTW 1-year RD for all-cause mortality for sulfonylureas vs metformin was 2.7% (95% CI 2.5, 2.8). Region 11 consisted of only 3,146 initiators, and its site-specific RD was 5.8% (95% CI 1.0%, 9.8%). This estimate was the farthest from the null; RDs in other regions ranged from 1.9% to 3.5%. After standardization to resemble region 11, RDs in other regions increased to range from 2.7% to 3.9%. Inverse variance weighting these standardized RD estimates resulted in a final estimate of 3.2% (95% CI 3.0%, 3.3%) in region 11.
Conclusions: Standardization can increase the precision of treatment effect estimates in specific sites of a multi-site study or distributed network without assuming homogeneous treatment effects. This approach increases the ability of knowledge users from smaller sites to draw inferences about safety and effectiveness within their site.
.jpg "Michael Andrew Webster-Clark, PharmD, PhD photo")
