Inclement winter weather such as snow, sleet, and freezing rain significantly impacts roadway safety. To assess the safety implications of winter weather, maintenance operations, and traffic operations, various crash frequency models have been developed. In this study, several datasets, including for weather, snowplow operations, and traffic information, were combined to develop a robust crash frequency model for winter weather conditions. When developing statistical models using such large-scale multivariate datasets, one of the challenges is to determine which explanatory variables should be included in the model. This paper presents a feature selection framework using a machine-learning algorithm known as the Boruta algorithm and exhaustive search to select a list of variables to be included in a negative binomial crash frequency model. This paper’s proposed feature selection framework generates consistent and intuitive results because the feature selection process reduces the complexity of interactions among different variables in the dataset. This enables our crash frequency model to better help agencies identify effective ways to improve roadway safety via winter maintenance operations. For example, increased plowing operations before the start of storms are associated with a decrease in crash rates. Thus, pretreatment operations can play a significant role in mitigating the impact of winter storms.