Mixed and missing data: a unified treatment with latent graphical models

TL;DR

This paper introduces a unified latent Gaussian graphical model for handling mixed and missing data, enabling improved data analysis, imputation, and prediction across various applications.

Contribution

It develops a novel latent Gaussian model with a sparse inverse covariance estimation for mixed and missing data, outperforming existing methods in prediction and imputation.

Findings

Outperforms state-of-the-art methods on medical datasets

Better than random forest in prediction error when model is correct

More effective than hot deck imputation even if model is misspecified

Abstract

We propose to learn latent graphical models when data have mixed variables and missing values. This model could be used for further data analysis, including regression, classification, ranking etc. It also could be used for imputing missing values. We specify a latent Gaussian model for the data, where the categorical variables are generated by discretizing an unobserved variable and the latent variables are multivariate Gaussian. The observed data consists of two parts: observed Gaussian variables and observed categorical variables, where the latter part is considered as partially missing Gaussian variables. We use the Expectation-Maximization algorithm to fit the model. To prevent overfitting we use sparse inverse covariance estimation to obtain sparse estimate of the latent covariance matrix, equivalently, the graphical model. The fitted model then could be used for problems including re- gression, classification and ranking. Such an approach is applied to a medical data set where our method outperforms the state-of-the-art methods. Simulation studies and real data results suggest that our proposed model performs better than random forest in terms of prediction error when the model is correctly specified, and is a better imputation method than hot deck imputation even if the model is not correctly specified.