STATISTICAL METHODOLOGY FOR MENTAL HEALTH RESEARCH

点击翻译按钮获取中文摘要

Modern mental health studies often result in large complicated data sets. In particular, longitudinal studies that follow individuals over time measure a variety of variables at each time point, yielding data that are often complicated by the presence of missing values or attrition from the sample. Survey data to detect the incidence and prevalence of depression involve lengthy questionnaires to measure mental disorders, and often are analyzed by methods that ignore complexities in the sample design. The effort and expense required to assemble these data sets is considerable, so it seems prudent to analyze them using the best methods available. However, often the analysis is confined to relatively basic statistical methods. Our objectives are to develop statistical methods for efficient and appropriate analysis of mental health data, and to make these methods accessible to other researchers. One of the most common problems facing NIMH researchers is the analysis of unbalanced repeated measures data. Data with this structure are often treated by inefficient and inappropriate methods, and state of the art methods, although an improvement, are usually based on assumptions that may not be appropriate for mental health outcomes. Statistical tests are largely based on large-sample theory, which is inappropriate for the small data sets that are often collected in NIMH studies. We propose to tackle these two problems, and to develop further the class of useful repeated measures models for NIMH data. A second focus of our research concerns the analysis of mental health surveys, with emphasis on nonresponse adjustments. Problems of outliers and missing data are often a significant reason why survey data are left incompletely analyzed. We propose to develop missing-data adjustments that improve considerably on naive methods of imputation, or methods that simply discard the incomplete cases. Our methods will build on the recently developed methodology of multiple imputation, and on the common technique of weighting adjustment for unit nonresponse.