Department of Biostatistics Seminar/Workshop Series

Semiparametric Latent Variable Transformation Models for Multiple Outcomes of Mixed Types

Anna Snavely

PhD Candidate in the Department of Biostatistics at Harvard School of Public Health

Wednesday, March 21, 1:30-2:30pm, MRBIII Room 1220

In many research settings, the goal is to assess the association between treatment (or some other covariate) and an outcome that cannot be measured directly. In order to perform such an analysis, we must rely on several measurable quantities that, when taken together, provide information about this unobservable or latent outcome of interest. In the biomedical setting, the measurable quantities (measurable outcomes) are often of varying types, including failure times. To handle this situation, we propose a semiparametric latent variable normal transformation model. In this model, the measurable outcomes are assumed to be governed by an unobserved (latent) variable, which in turn may depend on covariates such as treatment. Through this structure we are able to study the effect of covariates on the latent variable, which is of primary interest. As an extension of traditional latent variable approaches, our method allows the relationship between the measurable outcomes and latent variable to be unspecified and allows for measurable outcomes of mixed types which includes accounting for potentially censored outcomes. The method is applied to a study of head and neck cancer patients from Dana-Farber Cancer Institute in which the unobservable outcome of interest is dysphagia.