The EM algorithm for a linear regression model with application to a diabetes data
Xun Zhang, Jiale Deng, Rui Su · 2016
Linear modeling is the most common used statistical technique to discover hidden relationship between underlying random variables of interests because of its simplicity and interpretability. In this paper, we utilize linear models to study glycosylated hemoglobin, which a measure of the disease of diabetes. We want to find which other predictors or indicators have the most influential power on glycosolated hemoglobin. The dataset is collected from the American Diabetes Association. However, the dataset is incomplete due to missing data problem. We utilize EM algorithm to learn the linear model from the partial missing data. Convergence rate and robustness against initial values are examined. Moreover, we prove the convergence of EM algorithm in a more general setting. In addition, we evaluate the performance of EM at different missing rates and compare the results with two other methods that are typically used to deal with missing data. Experimental result shows that the EM algorithm have better performance than other methods in various missing rates in this application.