Walking is a form of active transportation with numerous benefits, including better health outcomes, lower environmental impacts and stronger communities. Understanding built environmental associations with walking behavior is a key step towards identifying design features that support walking. Human mobility data available through GPS receivers and cell phones, combined with high resolution walkability data, provide a rich source of georeferenced data for analyzing environmental associations with walking behavior. However, traditional techniques such as route choice models have difficulty with highly dimensioned data. This paper develops a novel combination of a data-driven technique with route choice modeling for leveraging walkability audits. Using data from a study in Salt Lake City, Utah, USA, we apply the data-driven technique of random forests to select variables for use in walking route choice models. We estimate data-driven route choice models and theory-driven models based on predefined walkability dimensions. Results indicate that the random forest technique selects variables that dramatically improve goodness of fit of walking route choice models relative to models based on predefined walkability dimensions. We compare the theory-driven and data-driven walking route choice models based on interpretability and policy relevance.