Large-Sample Logistic Regression with Latent Covariates in a Bayesian Networking Context

We considered the problem of predicting student retention using logistic regression when the most important covariates such as the college variables are latent, but the network structure is known. This network structure specifies the relationship between pre-college to college variables and then from college to student retention variables. Based on this structure, we developed three estimators, examined their large-sample properties, and evaluated their empirical efficiencies using WKU student retention database. Results show that while the hat estimator is expected to be most efficient, the tilde estimator was shown to be more efficient than the check method. This increased efficiency suggests that utilizing the network information can improve our predictions.