Genetic algorithm search for large logistic regression models with significant variables

A genetic algorithm (GA) is described which searches the space of all possible subsets of predictor variables for the best logistic regression model containing only significant variables. The method has been shown to be effective on a data set with eighteen variables and on a larger data set of two hundred variables. For the smaller data set an exhaustive search revealed only seven valid models, of which the GA found five. The method has been applied to linear regression with equal success. As GAs never guarantee to find an optimal solution the method is best described as an exploratory tool.