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

Area of Science:

• Epidemiology
• Biostatistics
• Health Services Research

Background:

• Health administrative databases are valuable but often lack crucial confounder data, leading to residual confounding.
• Residual confounding can bias estimates of causal effects, particularly in survival analyses.

Purpose of the Study:

• To evaluate the performance of high-dimensional disease risk score (hdDRS) methods in mitigating residual confounding bias.
• To assess causal effects on survival outcomes using health administrative data.

Main Methods:

• Utilized health administrative data from 49,197 individuals in British Columbia.
• Employed a plasmode simulation to compare eight hdDRS methods, high-dimensional propensity score (hdPS), and traditional regression adjustment.
• Investigated the relationship between tuberculosis infection and time-to-cardiovascular disease (CVD) development, targeting a log-hazard ratio of log(3).

Main Results:

• hdDRS methods demonstrated reduced bias compared to traditional methods, especially under strong unmeasured confounding.
• Bias ranged from -0.045 to -0.058 for hdDRS methods, while hdPS showed a bias of -0.047.
• Data analysis indicated an 8%-11% increased CVD risk associated with tuberculosis infection after addressing residual confounding.

Conclusions:

• Selected hdDRS methods are effective in addressing residual confounding bias for survival outcome estimation.
• The hdDRS method using rate-based risk score modeling on unexposed individuals showed minimal bias.
• The hdPS method performed comparably, and reproducible R code is provided for broader adoption.