Evaluating two model reduction approaches for large scale hedonic models sensitive to omitted variables and multicollinearity

Hedonic models in environmental valuation studies have grown in terms of number of transactions and number of explanatory variables. We focus on the practical challenge of model reduction, when aiming for reliable parsimonious models, sensitive to omitted variable bias and multicollinearity. We evaluate two common model reduction approaches in an empirical case. The first relies on a principal component analysis (PCA) used to construct new orthogonal variables, which are applied in the hedonic model. The second relies on a stepwise model reduction based on the variance inflation index and Akaike’s information criteria. Our empirical application focuses on estimating the implicit price of forest proximity in a Danish case area, with a dataset containing 86 relevant variables. We demonstrate that the estimated implicit price for forest proximity, while positive in all models, is clearly sensitive to the choice of approach, as the PCA reduced model produces a parameter estimate double the size of the alternative models. While PCA is an attractive variable reduction approach, it may result in an important loss of information relative to the stepwise reduction information based approach. Copyright Springer-Verlag Berlin Heidelberg 2014