Semiparametric Gaussian Copula Regression modeling for Mixed Data Types (SGCRM)
share
Many clinical and epidemiological studies encode collected participant-level information via a collection of continuous, truncated, ordinal, and binary variables. To gain novel insights in understanding complex interactions between collected variables, there is a critical need for the development of flexible frameworks for joint modeling of mixed data types variables. We propose Semiparametric Gaussian Copula Regression modeling (SGCRM) that allows to model a joint dependence structure between observed continuous, truncated, ordinal, and binary variables and to construct conditional models with these four data types as outcomes with a guarantee that derived conditional models are mutually consistent. Semiparametric Gaussian Copula (SGC) mechanism assumes that observed SGC variables are generated by - i) monotonically transforming marginals of latent multivariate normal random variable and ii) dichotimizing/truncating these transformed marginals. SGCRM estimates the correlation matrix of the latent normal variables through an inversion of "bridges" between Kendall's Tau rank correlations of observed mixed data type variables and latent Gaussian correlations. We derive a novel bridging result to deal with a general ordinal variable. In addition to the previously established asymptotic consistency, we establish asymptotic normality of the latent correlation estimators. We also establish the asymptotic normality of SGCRM regression estimators and provide a computationally efficient way to calculate asymptotic covariances. We propose computationally efficient methods to predict SGC latent variables and to do imputation of missing data. Using National Health and Nutrition Examination Survey (NHANES), we illustrate SGCRM and compare it with the traditional conditional regression models including truncated Gaussian regression, ordinal probit, and probit models.
READ FULL TEXT