Graduate Student ENAR PresentationsMarch 15, 2011 @ 3:30 pm - 4:30 pm
Location: 701 Blockley Hall
Title: A sensitivity analysis for the treatment effect on cost with unmeasured confounding
Abstract: Observational studies can be used to compare costs for treatment alternatives, but the estimated treatment effect is subject to bias from confounding. Even after adjustment for all known covariates, the results may still be subject to bias from unmeasured confounders. It is therefore advisable to assess the sensitivity of the treatment effect to various hypothesized unmeasured confounders. In some cases, closed-form relationships exist between the true and observed treatment effects. The investigator may use this relationship to adjust the estimated effect and corresponding confidence intervals for hypothesized distributions of the unknown confounder. We show how this adjustment can be used with cost models, and derive the adjustment for confounders that follow the Poisson and Gamma distributions. We assess the performance of the adjustment for cost data using simulation studies assuming a range of potential distributions of the confounder, and apply it to costs derived from SEER-Medicare for a stage II/III muscle-invasive bladder cancer cohort. We evaluate the costs for radical cystectomy versus combined radiation/chemotherapy, and find that the significance of the treatment effect is insensitive to unmeasured Bernoulli and Gamma confounders.
Poster Title: Parametric and Non-Parametric Methods for Estimating Conditional Survival
Abstract: There is an extensive body of literature in clinical oncology on conditional survival (CS) and on differential CS both over time and between groups. These papers focus on estimates of five-year CS, for example, for increasing patient survival time post-diagnosis. The statistical properties of estimators of CS, required for appropriate statistical inference, have not been studied. In this study, we compare the statistical properties of CS estimates using non-parametric and parametric methods. Non-parametric CS estimators are obtained from the survival distribution estimated using the Kaplan-Meier method and the parametric estimators are based on maximum likelihood theory assuming an underlying Weibull distribution. We developed estimators for the variances and covariances among the CS estimates required for multivariate analysis. Finally, we use our proposed methodology to evaluate changes in CS over time for patients with melanoma using survival data from the National Cancer Institute's Surveillance, Epidemiology and End Results (SEER) registry.
Poster Title: Comparison of Non-linear vs. linear models in detecting disease-modifying effects in Alzheimer's Trials
Abstract: Alzheimer's is a brain disease that causes problems with memory, thinking and behavior. Currently, drugs that available on the markets for people who suffer from Alzheimer's disease only approved explicitly to treat the symptoms but not the underlying disease progression. This presentation will compare non-linear vs. linear models in detecting the disease-modifying(DM) effect in Alzheimer's trials. A nonlinear model which is adopted from Ploeger B. and Holford N.(1) is used to generate the data. Non-linear models are from Ploeger B. and Holford N. (1) and Bhattaram V. et al (2). Linear, cLDA, model is from Liang and Zeger,(4). A delayed-start trial is used and three types of dropout rates are also implemented in combination with three different sample sizes. Simulation results show that non-linear models generally perform better in terms of power and bias in the estimate of the DM effect but may nevertheless have elevated type I error rates. They may also fail to converge, even when the assumed model is correct with starting parameters set to true values. The cLDA model, although typically biased, achieves near-nominal type I error rates, is simple to implement, and has no problems with convergence. Thus the cLDA model appears to be a good candidate for the detection of DM effects.
Poster Title: Adjustment for Measurement Error in Evaluating Diagnostic Markers
Abstract: Repeated measurements of a biomarker vary within a subject due to measurement error. Sources of measurement error include variability within individual over time and variability in the test. A naive approach ignores the error, biasing the sensitivity and specificity of the test and giving the erroneous impression that the biomarker is not effective. We propose bias-correction approaches for estimating sensitivity, specificity as well as positive and negative predictive values when the test is subject to measurement error. We derive their asymptotic properties. We then perform simulations to compare our approaches to naive approaches in estimates of sensitivity, specificity as well as positive and negative predictive values. The proposed methods have broad biomedical applications (e.g., renal disease, Alzheimer's disease) and are illustrated using a biomarker study in Alzheimer's disease.
Poster Title: Using SAS for Calculation of Prentice Constraints for GEE Analysis of Binary Data
Abstract: It is well known that in GEE analysis of binary data, that the correlations should satisfy additional constraints. We describe the constraints in general and present simplified versions for a logistic model. We then demonstrate our SAS macro that can be used to calculate the constraints. We recommend routine application of this macro after implementation of PROC GENMOD in GEE.