STATISTICAL ANALYSIS OF DRUG ABUSE TREATMENT RESEARCH

点击翻译按钮获取中文摘要

The two-group, longitudinal comparison is currently one of the most common designs of clinical trials in drug abuse outcome research. To analyze data from such a study, there are several statistical procedures available whose theoretical properties are understood. What is not so well-known is which of these methods should be used when analyzing real data collected under the constraints often encountered in drug abuse studies. Such constraints include high rates of attrition, restricted sample sizes, and varying assessment-to-assessment correlations. This is a complex question, compounded by the fact that statistical theory and research frequently outpace the realities of data collection and analysis. The proposed research will address this by investigating five statistical methods using simulated data and then confirm those results using real data. Studies 1, 2, and 3 will compare the statistical properties of size and power for the analysis of a two-group longitudinal design, contrasting individual Least-Squares Estimates of Slope as a summary statistic to four alternative procedures: Repeated Measures Analysis of Variance, Random Regression Modeling, Pre-Post Change, and the Normalized Area Under the Curve. Study 1 will investigate the effects of four factors that control the effective sample size available for analysis: 1) initial sample size; 2) the minimum number of observations per subject required for analysis; 3) the rate of attrition and; 4) the shape of the attrition curve. In Study 2, we will examine four factors associated with the size and type of treatment effect as reflected by the distribution of group means at each assessment point. We will vary: 1) effect size; 2) presence or absence of an interaction; 3) the number of assessments and; 4) level of assessment to assessment intercorrelation. The third study will examine the extent to which the presence of nonrandom missing data biases the obtained probabilities. The validity of the simulation results will then be established in Study 4 by applying the five statistical methods to the analysis of data sets drawn from actual randomized studies of drug abuse. Conclusions from the simulations will be used to predict the behavior of each method as a function of the characteristics of the data set analyzed. Findings from these four studies will improve the analysis of data from longitudinal clinical trials in all fields, especially drug abuse treatment research.