MISSING DATA METHODS FOR SUBSTANCE USE SURVEYS

点击翻译按钮获取中文摘要

Missing data pose special problems in surveys of substance use. High rates of missingness on key surgery variables make it necessary to develop principled, statistically sound approaches for data analysis. New, highly effective missing-data techniques have appeared in the statistical literature, including multiple imputation and algorithms for iterative simulation but these have not yet found their way into mainstream prevention research. The first goal of this proposed work is to introduce state-of-the-art technology for missing data into athe field of prevention by creating user-friendly software implementations of Schafer's algorithms, and to demonstrate their use in substantive analyses of data from the Adolescent Alcohol Prevention Trial (AAPT) A second goal is to extend the available missing-data technology to address unique feature of survey data that frequently arise in prevention research: (a) semicontinuous variables, which have a proportion of responses equal to zero and a continuous distribution among the nonhero values; (b) interactions, in which the relationship between predictor and response variables varies among subgroups; (c) nonignorable missingness mechanisms, in which the probability that variables are missing depends on the missing variables, or on other unrecorded variables related to them; (d) multilevel structure, in which individuals are tested within larger units such as classrooms or schools; and (e) longitudinal structure, in which data are collected for a group of individuals on multiple occasions. Substantive analyses of data from the AAPT will be conducted using these new methods, to confirm the validity of previous statistical conclusions and extend scientific knowledge relating to substance-use onset and prevention.