This is a FIRST proposal for five years and $340,000 direct expenditures. The primary focus of this proposal is to examine a number of issues related to incomplete data in mental health services. These incomplete data can occur either due to the nature of the research design, or due to non-response and attrition. The proposed study will develop more readily available algorithms for dealing with common patterns of missing data. There are five projects proposed over the period. The first study will develop a method for handling incomplete data in longitudinal studies where data may be missing on several (rather than just one) repeated measures, as well as baseline measures. The longitudinal outcomes will be modeled jointly, under the assumption that the nonresponse/missing data problems are ignorable. Marko-chain Monte Carlo techniques such as Gibbs sampling will be employed. The specific alternative approaches to be examined are described on page 27. Using the Medical Outcome Study (MOS) as a case example, the new approach will be compared with other ways of dealing with missing data. Dr. Belin has experience using this approach in a study by Liu, Taylor and Berlin (1995), applied to AIDS. An initial analysis using the MOS data is included in Preliminary Studies (pages 21-23). The second study will examine the use of follow-up data to deal with missing data issues, allowing for both ignorable and non-ignorable non-response. This approach depends on being able to partition initial non-responders into some who do respond to a second attempt at data collection and those who do not. Then an assumption is made about the relationship between the response for responders and non-responders, an assumption that can be tested using the follow-up responders. The application will be to a survey of patients in a study of team-based management in the public sector. Dr. Belin has employed a related model in a study (Belin et al., 193) related to undercounts in the Census, where there was some follow-up on some initial non-respondents. The third study will examine multiple sources of data in making quality of care determinations, where some of the data are only available for a subset of the cases. In many applications, it is not possible to classify a case based on a gold-standard method, but the study has done some validation work that provides a more accurate assessment using the higher quality data. Using data from multiple sources, log linear models can be estimated on the subset which has data from all sources. Based on these estimated relationships, on could then use multiple imputations or other approaches to deal with the missing data for the subset of individuals with fewer data sources. The application will be to a study of quality of care for schizophrenia in medication management. Dr. Belin has used the same approach in a cancer screening study. The fourth study will examine model selection based on incomplete categorical data in highly dimensional contingency tables, and in combining regression models across multiply-imputed data sets. The fifth study will deal with issues related to skewed outcomes and count data with incomplete data, as well as relying on normal theory models. The alternatives to be considered include Poisson and negative binomial models. A similar study will be conducted using the Box-Cox transformation for continuous measures. Attached are two papers that explore these approaches in the latter case. Each of these five products has a research plan, including the development of the estimator and some form of evaluation, either with a mental health data base or a simulation. Dr. Belin will also continue as an investigator in the Biostatistics Core of the UCLA Research Center on Managed Care for Psychiatric Disorders, (Dr. Kenneth Wells, PI).
| Yang, Xiaowei; Belin, Thomas R; Boscardin, W John (2005) Imputation and variable selection in linear regression models with missing covariates. Biometrics 61:498-506 |
| Song, Juwon; Belin, Thomas R (2004) Imputation for incomplete high-dimensional multivariate normal data using a common factor model. Stat Med 23:2827-43 |
| Rotheram-Borus, M J; Piacentini, J; Cantwell, C et al. (2000) The 18-month impact of an emergency room intervention for adolescent female suicide attempters. J Consult Clin Psychol 68:1081-93 |
| Liu, M; Taylor, J M; Belin, T R (2000) Multiple imputation and posterior simulation for multivariate missing data in longitudinal studies. Biometrics 56:1157-63 |
| Belin, T R; Hu, M Y; Young, A S et al. (1999) Performance of a general location model with an ignorable missing-data assumption in a multivariate mental health services study. Stat Med 18:3123-35 |
