Selection of Minimal Variable Subsets with the help of Genetic Algorithms in Economic Analysis

In this paper we demonstrate the use of Genetic Algorithms in the selection of significant variables from a large set of available variables used to explain the observed behavior of subjects in an economic experiment. Standard regression analysis requires assumptions on a functional form and may thus prevent us from finding all relevant relationships. When using a more flexible functional form the number of coefficients (corresponding to different variables and their interactions) grows exponentially. Hence, our goal in this paper is to select the smallest set of variables with the largest “explanatory” value. We use a variation of a non-traditional type of Genetic Algorithm, CHC, to “evolve” this preferred minimal set of relevant variables. We compare this approach with estimation based on basic linear (least squares) regression models and regression models chosen by using the stepwise regression method. Additionally, we also evaluate the effectiveness of various fitness criteria in our genetic algorithm’s fitness function. We believe that an evolutionary computation approach provides a useful alternative and supplementary method to more traditional methods by offering potentially new useful subsets of significant variables warranting further exploration and a limit to the number of coefficients in outputs.