Statistical Methodology for Correlated Data in Health Sciences

点击翻译按钮获取中文摘要

The goal of the proposed research is to develop novel statistical methodology that is directly applicable to clustered data arising from clinical and health services research. Specific questions addressed are: 1) estimation of risk ratios in cluster randomization trials involving a small number of large clusters; 2) inference for the concordance probability of survival models with clustered data; 3) design and analysis of 2-stage preference randomization trials with binary outcomes; and 4) inference procedures for correlations and their differences with clustered data. New methods are needed for estimating risk ratios in cluster randomization trials with a small number of large clusters, because existing methods focus on odds ratio which are uneasy to interpret for non-statisticians. Procedure for the estimation of risk ratios will be developed to address the issue of small number of large clusters. Associated methods for sample size estimation and model selection will also be developed. Prediction of outcomes that occur over time is important in population health and health services research as well as clinical practice. There exists a variety of techniques to quantify the discrimination ability of prediction models for survival data, but few have been developed for clustered survival data. New methods will be developed for such cases by extending the concept of concordance probability that is widely applied in practice. Confidence interval procedures for concordance probabilities and their differences will be developed and evaluated. Treatment preferences reflect individuals' choice of treatment and influence their adherence to treatment and outcomes. Multiple effects including patient preference and self-selection can be estimated using a 2-stage randomized design that allows a randomly selected subset of participants to choose their own treatment, with the remainder randomized to treatment groups in the usual way. Statistical methods are now well-developed for continuous outcomes with less attention paid to binary outcomes in this design. In the proposed research, novel procedures for confidence interval construction and sample size estimation will be developed and evaluated for the 2-stage performance randomized trials with binary outcome. Correlation coefficients are widely used to measure the association between two variables. Oftentimes one needs to assess correlations in sittings where multiple measurements are available on each of the variables within a cluster, whereas virtually all existing methods assume independence of data for pairs of variables. New confidence interval procedures for common correlations such as Pearson's and Spearman's in clustered data settings will be developed with a focus on improving the performance in the case of small number of clusters.