Developing a prognostic model in castration-resistant prostate cancer patients in the presence of missing data
Abstract
Purpose: To assess the use of tumor serum markers, such as circulating tumor cells (CTC), as independent predictors of overall survival in men with castration-resistant prostate cancer, and to find the best combination of such markers to use in predicting survival. To compare different methods of handling missing data - pairwise deletion, mean substitution, and multiple imputation (using either a Markov Chain Monte Carlo (MCMC) approach or Multiple Imputation by Chained Equations (MICE)) - in multivariable prognostic models. Methods: This retrospective study included 99 men with castration-resistant prostate cancer. The Cox proportional hazards model was applied to both measured and imputed data to test associations of survival with baseline tumor serum markers and clinical or pathological factors. The effectiveness of the four different strategies of handling missing data were compared, using regression coefficients and model performance measures. Results: In univariate analyses, high lactate dehydrogenase (LDH) levels, CTC counts, carcinoembryonic antigen (CEA) levels, and erythrocyte sedimentation rates (ESR), and low hemoglobin levels were associated with increased risk of death. In multivariable models, LDH, CTC counts, and CEA remained predictive of overall survival. In comparisons of methods for handling missing data, pairwise deletion (equivalent to complete case analysis in univariate analyses) produced inflated regression coefficient estimates for variables with 10% or more missingness. Mean substitution underestimated standard errors and p-values. With multiple imputation, models based on the MCMC approach did not differ much from the models derived from the MICE approach. Conclusion: Our results should be validated in a larger sample, but if verified, the multivariable model should aid in understanding and predicting survival time for patients with castration-resistant prostate cancer. For handling missing data, multiple imputation is superior to complete case analysis and mean substitution. Future simulation studies may lead to a better assessment of the relative merits of the MCMC and MICE methods for multiple imputation.
Subject Area
Biostatistics|Oncology
Recommended Citation
Ouyang, Fangqian, "Developing a prognostic model in castration-resistant prostate cancer patients in the presence of missing data" (2013). Texas Medical Center Dissertations (via ProQuest). AAI1549927.
https://digitalcommons.library.tmc.edu/dissertations/AAI1549927