Statistical Analysis in Longitudinal Mental Health Study

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): it is often difficult or impossible, to "randomize" treatment (e.g. to assign a representative payee or not) or the risk factors of interest (e.g., socioeconomic status) that may be associated with longitudinal mental health outcome. Observational studies thus play an important role in many crucial areas of mental health (services) research. However, in the presence of concurrent time-dependent or longitudinal confounding, statistical analysis with standard approaches (e.g., using multiple regression to adjust for baseline or time-dependent differences) is likely to be biased. In this proposal, we develop new analytic strategies for the analysis of longitudinal mental health data, extending the marginal structural framework of Robins et al, to address situations in which baseline, and most importantly, longitudinal confounding covariates may bias the observed relationships between independent variables of interest and outcomes. The proposed methods are primarily developed for use in observational outcome studies, but they also have applications in studies that use experimental designs as for example, when the amount of missing data varies by treatment group, or when treatment noncompliance occurs. In addition, we propose an extension of the latent class approach to allow description of interacting patterns among multiple longitudinal variables so as to improve our understanding of the dynamic relationship among such variables in mental health research. We want to emphasize two important conceptualizations in this proposal. First, we make the distinction between longitudinal variables and concurrent time-dependent variables. The former may only be available intermittently and may be measured with error. The value of the latter is assumed to be known whenever variables of primary interest or outcomes measured at multiple time points are available. Second, we recognize that assertions of causality in observational study must rely on the untestable assumptions of "no unmeasured confounding". As a result, adjustments for confounding can only address factors that have been measured. Therefore, we state that outcomes are "attributable" to the defined risk factors or treatments. We assert only that our findings are consistent with causal hypotheses, and not that they demonstrate causality itself.