Segmentation modeling: Applications of Finite Mixture Regression Models in University Fundraising and Management of Transportation Infrastructure

One critical issue that arises in the analysis of panel data (or cross-sectional data) is the need to account for the unobserved heterogeneity. When not rigorously accounted for in statistical models, unobserved heterogeneity can cause large portions of unexplained variation, or biased parameter estimates because the associated variation is incorrectly attributed to the explanatory variables. In this dissertation, three Finite Mixture Regression models are proposed to account for the unobserved heterogeneity on a group-level and to provide a post hoc segmentation approach that is based on objective criteria. The proposed models allow for classification of individuals into segments, and the simultaneous estimation of a linear regression model within each segment. The models are illustrated in two applications, one using the alumni contribution data collected by a PhD-granting university in the Midwestern United States and the other using the pavement deterioration data from AASHO Road Test. The estimation results suggest radically different coefficients describing the effect of explanatory variables for alumni/pavements in different segments, which reinforce the ability of the proposed models to explain the unobserved heterogeneity. Such differences can be exploited by decision makers to make effective fundraising or Maintenance and Rehabilitation (M&R) strategies, e.g., different school effects suggest university fundraisers may do well to send solicitation materials with tailored school information; different traffic loading effects can motivate highway administrations to impose customized traffic/load restrictions. More implications of the results are discussed in the context of the two applications. Limitations and future research directions are provided in the end.