Studying rapidly changing phenomena like substance use, its determinants, and its correlates requires a fine-grained approach involving measurement at frequent intervals over time. Formerly, designs involving such intensive measurement were impractical. Today new computer and communication technologies are providing exciting alternatives to traditional data collection procedures and allowing more frequent, more flexible, and less intrusive measurement. One example is Ecological Momentary Assessment, where subjects in a study of smoking cessation are given a hand-held Palm Pilot computer which beeps them at some combination of random and predetermined times and presents them with questions pertaining to mood and urge to smoke. In this type of data collection protocol, it is not unusual for each subject to provide more than 12,500 responses over the course of several weeks' participation. We refer to such data as """"""""massively multivariate."""""""" In theory, these data can provide answers to important questions such as: What is the relationship between mood and substance use? What environmental cues trigger smoking? How does the sensation of withdrawal vary over a day or a week, and how does it vary according to environmental cues? In practice, however, it is not immediately clear what statistical procedures should be used to address these questions with massively multivariate data. Methods currently used by prevention researchers for multivariate, longitudinal, and time series data simply were not designed for irregular serial and clustered data, or for assessing time-varying effects. We will draw on new methods from the statistical literature, in particular from spatial statistics, to develop new methods of data analysis tailored to massively multivariate prevention data. We will develop a factor analysis approach for irregular serial and clustered data, and we will develop nonparametric methods for time-varying mixed effects, correlations, and variance. In collaboration with prevention researchers, we will apply the new methods to massively multivariate prevention data. We will also produce software to enable prevention scientists to apply the new procedures.
Showing the most recent 10 out of 443 publications
