Detransformation Bias in Nonlinear Trip Generation Models

Abstract

In recent years, there have been substantial efforts from researchers and practitioners to improve site-level trip generation estimation methods to address some of the pitfalls of conventional approaches for applications such as traffic impact analyses. These new trip generation models often adopt sophisticated nonlinear model forms to utilize new information and incorporate new factors influencing trip generation. However, if sufficient caution is not taken in their application, these new predictive models may introduce severe bias. This paper focuses on a typical source of biases in the applications of such models arising from detransformation of predictions from models with a nonlinearly transformed dependent variable in the prediction process (for example, predicting from a semilog model). While such biases are well known and corrections have been proposed in other disciplines, they have not been adopted in site-level trip generation models to the authors’ knowledge. The detransformation bias is described and demonstrated—focusing on log-transformed models—with numeric simulations and empirical studies of trip generation models, before discussing their implications for trip generation applications and research.