The dangers of natural gas distribution and aging infrastructure are well known, but determining the risk for incidents related to natural gas is a difficult statistical, engineering, and civic challenge. Furthermore, investigating and repairing infrastructure can be prohibitively expensive. These authors use publicly-available data and compare several statistical methods to identify an optimal predictive model. Given a broad approach to feature selection, a naive prediction based only on historical leaks performed better than more complex models, such as regularized linear regression, random forests, and multi-layer perceptron neural networks; however, these more complex models did demonstrate the ability to detect complex structures in the data, improve feature selection and possibly address more precisely defined objectives, such as prioritizing inspections over select geographic areas.