STATISTICAL MODELS FOR NESTED SERVICES UTILIZATION DATA

点击翻译按钮获取中文摘要

As noted in Program Announcement #94-060, {Research on Methods, Measurement, and Statistical Analysis and Mental Health Research}, "advances in mental health research are highly dependent on the quality of data analytic strategies available to investigators." With this in mind, our three-year project "Statistical Models for Nested Services Utilization Data" extended random-effects regression models (RRM) to allow for more general types of data collected in mental health services research. RRM are especially useful for analyzing data from designs that are longitudinal (observations nested within subjects) or clustered (subjects nested within clusters), both of which are quite common in mental health services research. Under the grant, we developed RRM for nominal outcomes and counts, and produced software and manuals (called MIXNO and MIXPREG) implementing these procedures and describing their use. This work build upon past work of this research team in which methods and programs for continuous, dichotomous, and ordinal outcome variables had been developed (programs MIXREG and MIXOR). Thus, methods and software are now available for a wide class of outcomes for designs that are either longitudinal or clustered. The focus of this competitive renewal is to further generalize RRM to handle data that are both clustered and longitudinal. For example, repeated observations (level-1) may be observed within subjects (level-2) who are nested within clusters (level-3, e.g., hospital, clinic, research unit). For such 3-level data, we proposed to generalize current statistical methodology of RRM, extent our freeware programs, enhance the user interface of these programs, and develop accompanying Primers. A second and more basic statistical research component of this proposal is to begin work on multivariate RRM. Specifically, we propose a general multivariate mixed-effects regression model that combines a random-effects variance component structure for cluster and/or person-specific time trends with a factor analytic model for association between multiple outcome variables (that might simultaneously measure multiple domains of the underlying response process). The model will also allow residual autocorrelation. This new area of statistical research will be explored in detail for the cases of continuous and binary outcome measures. Thus, the overall goal of this proposal is to further develop and generalize RRM to handle many of the challenges encountered in analyzing mental health services research data of various types (i.e., continuous, ordinal, nominal, counts), structures (i.e., univariate or multivariate) and from a variety of designs (i.e., 2-level or 3-level).