Longitudinal studies occupy an important role in modern psychiatric research; however, the statistical methods for analyzing these data are rarely commensurate with the level of effort and expense required to obtain such data. Often, a simple """"""""endpoint"""""""" analysis is performed or records with missing data are discarded. The results from either compromise can be extremely misleading. The purpose of this proposal is to develop and make available, a general statistical model for the analysis of within-subject psychiatric data. Based on the concepts of """"""""random regression"""""""" we will develop a general model for continuous, discrete and ordinal data that permit missing data, irregularly spaced observations, correlated errors, time varying covariates (e.g., plasma level) and time invariant covariates (e.g., treatment group). Unlike traditional methods based on average change, the random regression approach can also provide estimates of the rate of change for each individual subject. This particularly useful in the psychiatric setting where a proportion of subjects may respond to therapy in quite different ways from the average response. In addition to the statistical derivation, the general model will be tested using simulated data and illustrated using 5 existing psychiatric data sets. We will develop user-friendly public-domain software running on a micro-computer and a primer designed to illustrate the use of the model on the five data sets. Finally, the statistical properties of the method will be examined in terms of power, robustness, model fit and model selection.
