Leveraging Geographic Information for Causal Inference in Pharmacoepidemiology

Abstract

Randomized trials (RCT) with major adverse cardiovascular event (MACE) outcomes found no effect of dipeptidyl-peptidase-4 inhibitors (DPP-4i) medications compared to placebo or second-generation sulfonylureas (SU) while non-randomized database studies suggested a benefit of DPP-4i versus SU. Residual confounding by socioeconomic factors were thought to be a main reason. Aim 1 characterized the geospatial distribution of the adoption of DPP-4i antidiabetics versus SU. Aim 2 evaluated whether incorporation of small-area-level socioeconomic measures can reduce confounding in non-randomized comparisons of DPP-4i to SU as second-line antidiabetic therapies in the prevention of MACE. Aim 3 compared area-level prescribing density (APD) and physician prescribing preference (PPP) as instrumental variables (IVs) with the objective of finding a strategy for control of previously unmeasured confounders. Using Medicare claims data from 2012 to 2017, two cohorts were built emulating RCTs of sitagliptin or saxagliptin starters each compared to SU starters. For each ZCTA, the proportion DPP-4i prescribing in relation to total ZCTA cohort members was calculated and used for a local indicator of spatial association cluster analysis. Multilevel logistic models were used to quantify the variation in medication use at the individual, ZCTA, state, and region levels. DPP-4i utilization proportion was low (sitagliptin median = 0.22; interquartile range 0.15 to 0.33; saxagliptin median = 0.025; 0.00 to 0.069). High amounts of clustering were observed for sitagliptin proportion (Moran’s I = 0.32) and saxagliptin proportion (Moran’s I = 0.20). Sitagliptin utilization was high in the New York metro area and urban southern California. Saxagliptin had similar patterns with additional clusters in the upper Midwest. Regions, states, and ZCTAs accounted for 8.1% of variation in sitagliptin prescribing and 13.3% of saxagliptin prescribing. Variation across ZCTAs suggests neighborhood factors may have been important determinants of prescribing. Removing high co-payments may have improved equity in access to safer antidiabetics. Area-level covariates were obtained for ZCTAs from the American Community Survey. ZCTA-level socioeconomic covariates were incorporated into propensity scores for Cox proportional hazards models. Cox models stratified by ZCTAs were also fit. Adding area covariates improved propensity score model fit and treatment discrimination. Unadjusted associations for receiving sitagliptin or saxagliptin compared to SU showed a decreased hazard of MACE occurrence (sitagliptin hazard ratio (HR) = 0.86; 95% confidence interval 0.83 to 0.88; saxagliptin HR = 0.68; 0.64 to 0.73). Adjusting for individual-level covariates moved estimates towards the null (sitagliptin HR = 0.89; 0.86 to 0.92; saxagliptin HR = 0.78; 0.73 to 0.83). Adding area covariates moved estimates minimally closer to the null. Adjusted stratified Cox models produced similar results (sitagliptin HR = 0.90; 0.87 to 0.93; saxagliptin HR = 0.76; 0.71 to 0.81). Incorporation of area-level covariates in survival analyses did not meaningfully reduce confounding beyond individual-level covariates. The proportion of DPP-4i prescribing in relation to all cohort members in a zip code tabulation area defined the APD IV at various cutoffs (0% vs. 100% to % vs. ≥50%). The same proportions were calculated for each patient’s physician prescribing history as the PPP IV. An instantaneous physician preference (iPPP) IV used a physician’s most recent prescription. Two-stage IV regression models were adjusted for propensity score quintiles. All IVs were strong and reduced covariate imbalance. APD IV analyses found no meaningful difference for sitagliptin (0% vs. 100% HR = 1.11; 0.79 to 1.57). PPP IV analyses showed reduced risk for sitagliptin (% vs. ≥50% HR = 0.69; 0.48 to 0.98). iPPP analyses showed little to no difference for sitagliptin (HR = 0.86; 0.60 to 1.10) and saxagliptin (HR = 0.98; 0.56 to 1.72). Instruments focusing on short-term prescribing preference like the iPPP IV hold promise over area-based instruments to improve confounding control in comparative effectiveness analyses.