We introduce a new test of preference stability across consumers, based on a discrete mixture model of feedforward neural networks, designed to detect clusters of consumers following distinct consumption functions. The procedure is tested on simulated data and applied to two sets of Canadian household consumption microdata (1969-1986 and 2004-2008). On both these datasets we find that two clusters best explain the data, thus rejecting preference stability. Finally we address elasticity estimation, and find on simulated data that for most practical uses simple demand systems suffice.
Classification JEL : C45, D12