An approach to addressing selection bias in survival analysis

Area of Science:

• Biostatistics
• Survival Analysis
• Health Economics

Background:

• Non-random treatment assignment is a common challenge in survival analysis, potentially biasing results.
• Existing methods like propensity scores or uncorrected models can yield biased estimates, even with large sample sizes.
• The two-stage residual inclusion (2SRI) method addresses endogeneity but has computational and interpretive limitations.

Purpose of the Study:

• To propose and validate a novel endogenous selection survival model (esSurv) that accounts for non-random treatment assignment.
• To demonstrate the computational and interpretive advantages of esSurv compared to 2SRI and other methods.
• To assess the impact of endogeneity on survival outcomes using real-world dialysis patient data.

Main Methods:

• Development of the endogenous selection survival model (esSurv) incorporating frailty.
• Monte Carlo simulations to compare esSurv performance against 2SRI, propensity score methods, and uncorrected models.
• Application of esSurv to dialysis patient data comparing mortality risk between mature arteriovenous grafts and those not yet ready for use.

Main Results:

• esSurv parameter estimates closely align with 2SRI results, indicating consistency.
• esSurv demonstrates significantly faster computation and generally smaller standard errors than 2SRI.
• esSurv explicitly estimates the correlation of unobservable factors (ρ=0.55), providing superior interpretability over 2SRI's residual parameter.
• Uncorrected models showed substantial bias (hazard ratio 0.630) compared to esSurv (0.197) and 2SRI (0.173).

Conclusions:

• The proposed esSurv model effectively handles non-random treatment assignment in survival analysis.
• esSurv offers significant computational and interpretive benefits over existing methods, including 2SRI.
• Accounting for endogeneity is critical for accurate survival outcome estimation, as demonstrated in the dialysis patient example.