Under the current project Missing-Data Methods for Substance-Use Surveys, we have created state-of-the-art methods for missing data in prevention studies. Our major accomplishments include: injecting the new technology of multiple imputation into prevention research through software products, publications, and workshops nationally and internationally; extending Latent Transition Analysis or LTA to allow missing values and to calculate standard errors; and developing a new class of growth models for semicontinuous substance-use variables which are a mixture of zeros and positive values. In this continuation, we propose to develop statistical tools for prevention research in five important areas: (1) Multistage multiple imputation. This sequential imputation technique will allow us to isolate the effects of missing information arising from multiple sources, such as nonresponse and measurement error. (2) Latent Transition Analysis with concomitant predictors. This extension of LTA will support richer models of stage-sequential development. (3) New imputation methods for multivariate multilevel data. For the first time, researchers will be able to properly impute missing covariates in multi-site longitudinal studies-for example, where repeated measurements are taken on students nested within schools. (4) Pattern-mixture models for nonignorable nonresponse. These new models will help us to check the sensitivity of study results to departures from the assumptions usually made about missing values, assumptions which in practice are often unverifiable. (5) Software development and technology dissemination. Our new methods will be validated through simulation and applied to data drawn from major prevention studies. Working together with the staff of the Software Development and Computer Support Core, the new methods will be implemented in user-friendly and well-documented software and distributed free of charge to the prevention research community.
Showing the most recent 10 out of 443 publications
