High-Dimensional Disease Risk Score for Dealing With Residual Confounding Bias in Estimating Treatment Effects With a Survival Outcome.

Purpose: Health administrative databases often contain no information on some important confounders, leading to residual confounding in the effect estimate. We aimed to explore the performance of high-dimensional disease risk score (hdDRS) to deal with residual confounding bias for estimating causal effects with survival outcomes.

Methods: We used health administrative data of 49 197 individuals in British Columbia to examine the relationship between tuberculosis infection and time-to-development of cardiovascular disease (CVD). We designed a plasmode simulation exploring the performance of eight hdDRS methods that varied by different approaches to fit the risk score model and also examined results from high-dimensional propensity score (hdPS) and traditional regression adjustment. The log-hazard ratio (log-HR) was the target parameter with a true value of log(3).

Results: In the presence of strong unmeasured confounding, the bias observed was -0.11 for the traditional method and -0.047 for the hdPS method. The bias ranged from -0.051 to -0.058 for hdDRS methods when risk score models were fitted to the full cohort and -0.045 to -0.049 when risk score models were fitted only to unexposed individuals. All methods showed comparable standard errors and nominal bias-eliminated coverage probabilities. With weak unmeasured confounding, hdDRS and hdPS produced approximately unbiased estimates. Our data analysis, after addressing residual confounding, revealed an 8%-11% higher CVD risk associated with tuberculosis infection.

Conclusions: Our findings support the use of selected hdDRS methods to address residual confounding bias when estimating treatment effects with survival outcomes. In particular, the hdDRS method using rate-based risk score modeling on unexposed individuals consistently exhibited the least bias. However, the hdPS method showed comparable performance across most evaluated scenarios. We share reproducible R codes to facilitate researchers' adoption and further evaluation of these methods.