Determining the Number of Market Segments Using an Experimental Design

The aim of this work is to determine how well criteria designed to help the selection of the adequate number of mixture components perform in mixture regressions of normal data. We address this research question based on results of an extensive experimental design. The simulation experiment compares several criteria (26), including information criteria and classification-based criteria. In this full factorial design we manipulate 9 factors and 22 levels, namely: true number of segments (2 or 3), mean separation between segments (low, medium or high), number of consumers (100 or 300), number of observations per consumer (5 or 10), number of predictors (2, 6 or 10), measurement level of predictors (binary, metric or mixed), error variance (20% or 60%), minimum segment size (5-10%, 10-20% or 20-30%) and error distribution (normal versus uniform). The performance of the segment retention criteria is evaluated by their success rates; we also investigate the influence of experimental factors and their levels on success rates. The best results were obtained for the criteria AIC3, AIC4, HQ, ICLBIC and ICOMPLBIC. BIC and CAIC also perform well with large samples and a large number of market segments.