CONTROLLING UNMEASURED CONFOUNDING AND BIAS IN EVALUATION OF TREATMENT EFFECTIVENESS USING REAL-WORLD DATA, WITH APPLICATION TO PALLIATIVE CARE
Author(s)
XIANGMEI MA, PhD1, Mihir Gandhi, MSc, PhD1, Grace Yang, MBBS, MPH, PhD2, Yin Bun Cheung, PhD1.
1Duke-NUS Medical School, Singapore, Singapore, 2National Cancer Center Singapore, Singapore, Singapore.
1Duke-NUS Medical School, Singapore, Singapore, 2National Cancer Center Singapore, Singapore, Singapore.
OBJECTIVES: We propose and evaluate a statistical model for recurrent time-to-event analysis to control for unmeasured confounding and bias arising from unobserved heterogeneity - also known as omitted covariate bias - in the use of real-world data to evaluate treatment effects on event rate outcomes, with focus on the context of palliative care and acute healthcare utilization.
METHODS: We embed the prior event rate ratio (PERR) method in the Andersen-Gill (AG) model for recurrent time-to-event analysis. The PERR method features control of unmeasured time-constant confounding whereas the AG model mitigates the bias arising from unobserved heterogeneity and allows adjustment for unobserved time-varying confounding that are correlated with time-to-death by using "reverse time-to-death" as the time-scale. We evaluate the method using simulation and illustrate it with real-world data of patients with advanced cancer.
RESULTS: Simulation shows unbiased estimation of treatment effect and accurate coverage probability of confidence interval in the presence of various sources of confounding and bias. Using typical method for recurrent event times analysis with covariate adjustment and palliative care as a time-varying exposure, palliative care was found associated with about 3-fold increase of emergency department visit rate; hazard ratio (95% CI) was 3.2 (2.8-3.6). Using our novel method, the hazard ratio (95% CI) was 0.51 (0.42-0.63).
CONCLUSIONS: Our proposed method of embedding PERR in the AG model for recurrent events analysis combines the strengths of the two methods in controlling unmeasured time-constant confounders and bias arising from unobserved heterogeneity. Furthermore, in the context of palliative care and advanced diseases, it allows the use of reverse time-to-death as the model's time-scale to control unmeasured time-varying confounders that are correlated with time-to-death. The sharp contrast of different hazard ratio estimates obtained from regression with covariates adjustment versus PERR embedded AG model indicates practical value of the proposed method. <!--EndFragment-->
METHODS: We embed the prior event rate ratio (PERR) method in the Andersen-Gill (AG) model for recurrent time-to-event analysis. The PERR method features control of unmeasured time-constant confounding whereas the AG model mitigates the bias arising from unobserved heterogeneity and allows adjustment for unobserved time-varying confounding that are correlated with time-to-death by using "reverse time-to-death" as the time-scale. We evaluate the method using simulation and illustrate it with real-world data of patients with advanced cancer.
RESULTS: Simulation shows unbiased estimation of treatment effect and accurate coverage probability of confidence interval in the presence of various sources of confounding and bias. Using typical method for recurrent event times analysis with covariate adjustment and palliative care as a time-varying exposure, palliative care was found associated with about 3-fold increase of emergency department visit rate; hazard ratio (95% CI) was 3.2 (2.8-3.6). Using our novel method, the hazard ratio (95% CI) was 0.51 (0.42-0.63).
CONCLUSIONS: Our proposed method of embedding PERR in the AG model for recurrent events analysis combines the strengths of the two methods in controlling unmeasured time-constant confounders and bias arising from unobserved heterogeneity. Furthermore, in the context of palliative care and advanced diseases, it allows the use of reverse time-to-death as the model's time-scale to control unmeasured time-varying confounders that are correlated with time-to-death. The sharp contrast of different hazard ratio estimates obtained from regression with covariates adjustment versus PERR embedded AG model indicates practical value of the proposed method. <!--EndFragment-->
Conference/Value in Health Info
2026-09, ISPOR Asia Pacific 2026, Bangkok, Thailand
Value in Health, Volume 55, Issue S1
Code
MSR5
Topic
Methodological & Statistical Research
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference
Disease
SDC: Oncology, STA: Multiple/Other Specialized Treatments