Testing For Aggregation Bias in a Non-Linear Framework: Some Monte Carlo Results

Abstract

Researchers modeling the behavior of individual people or firms are often unable to utilize micro-level data because such data are unavailable or unreliable. Faced with this dilemma, researchers often resort to using aggregate-level data. When the individual-level variable of interest is dichotomous, however, the aggregate-level model is subject to a special form of aggregation bias. Kelejian [1994] provides a methodology for testing for the presence of this form of bias. This study uses Monte Carlo analysis to evaluate the usefulness of Kelejian’s test. Using 50 units, populated by, 100, 1000 and 10,000 individuals, we find that aggregation bias is almost universally present. Unfortunately, the Kelejian test identifies the bias in fewer than half of the cases.