Browsing by Author "Fang, Xiao"
Now showing 1 - 2 of 2
- Results Per Page
- Sort Options
Item Comparing Different Methods of Applying Propensity Score in Controlling Selection Bias for Studies of Prostate CancerFang, Xiao; Kuo, Yong-Fang; Tan, Alai; Sheffield , KristinBackground: Prostate cancer is the most common cancer among men. There is growing interest in population-based comparative effectiveness research on different therapies for prostate cancer because of the limitations on generalizability of randomized controlled trials. The major challenge for population-based studies using observational data is how to control for selection bias. The objective of this study was to compare the performance of four common methods of applying propensity scores—covariate adjustment, stratification, matching, and inverse probability of treatment weighting—to adjust for confounding in analyses of treatment effects among prostate cancer patients. Methods: The linked Surveillance, Epidemiology, and End Results (SEER) Medicare database 1992-2007 was used for this research in which we studied two scenarios. In scenario 1, the overall mortality and cause-specific mortality for patients with local prostate cancer receiving active treatment (radical prostatectomy or radiation) vs. observation were compared. The known confounding factor, comorbidity, was removed in the analyses to evaluate the selection bias for general health. In scenario 2, we compared the overall mortality and cause-specific mortality between patients with and without primary androgen deprivation therapy. The known confounding factor, cancer grade, was removed in the analyses to evaluate the selection bias for cancer severity. Results: Among four propensity methods, matching (greedy matching and caliper matching) produced the estimates that were closest to the estimates when the removed confounder was included in both scenarios. Covariate adjustment and stratification yielded similar controlling effects in both scenarios. Inverse probability of treatment weighting showed a better performance in scenario 1 compared to scenario 2. Conclusions: Propensity score matching outperformed covariates adjustment, stratification, and inverse probability of treatment weighting. None of the propensity methods produced estimates that were identical to the estimates when the removed confounders, comorbidity or cancer grade, were included.Item Joint Models for Longitudinal Analysis and Competing Risks in Survival AnalysisFang, XiaoIn longitudinal studies involving assessing ordinal disease(s) outcome across multiple time points, ignoring the correlations between the development and transition of the disease status and any informative potential censoring event(s) may lead to bias in the estimation of covariate effects. To address the problems created by these dependencies, we propose and apply several models: 1. Linking the Cause-specific Hazards Model for competing risks analyses with a model for longitudinal ordinal measurements through a correlated random effects structure. To do this, we apply a 2009 approach by Li et al. (2009), and use this as the baseline model. 2. Developing a joint longitudinal competing risks survival model in which previous wave disease status history is an additional time-varying covariate in the survival component, through the same linkage mechanism as in the baseline model. 3. Applying the same linkage mechanism as in the baseline model, we model the longitudinal outcome using a continuous time Markov chain model. 4. Comparing the application results from applying the baseline approach to those obtained using available software. To assess the properties of these models and their effectiveness, we conduct simulation studies, in which we compare performance (likelihood and covariate parameter estimate accuracy) against the baseline model. The Aim 2 model achieves smaller bias in the estimation of the intercepts compared to separate model estimates, as well as estimates from the baseline approach when the speed of the change of the longitudinal ordinal outcome is appropriately estimable. Model 3 achieves more accurate estimates of the baseline transition probabilities and covariate effects, compared to the separate models, when the association between two sub-models is large. To assess the practical effects of these model modifications, we demonstrate the fit of these models for estimating measures of frailty (treated as a three-level factor) in the Hispanic Established Populations for the Epidemiologic Study of the Elderly data, which is a longitudinal survey study to estimate the prevalence of and risk factors for key health conditions in older Mexican Americans. Death and loss to follow-up are occurrences with potential to add bias to covariate effect estimates. We applied the baseline approach and available software to the working data, and covariate estimates and model fit were compared.