On the consequences of ignoring unobserved heterogeneity when estimating macro-level models of crime

Abstract

The majority of criminologists have ignored unobserved heterogeneity in macro-level models of crime. We used time-series–cross-section (TSCS) data from all 58 counties in California between the years of 1990 and 1998 to show that heterogeneity should not be ignored. The paper begins with a discussion of some of the problems inherent in TSCS data—contemporaneous correlation, serial autocorrelation, spatial autocorrelation, panel heteroskedasticity, and nonstationarity—and the techniques for detecting them. To correct these problems we estimated dynamic TSCS models (with panel corrected standard errors) for property and violent crime. The resulting parameter estimates served as a base of comparison for the additional models. Next, we controlled for heterogeneity by adding fixed effects for county and year to the dynamic TSCS models. Finally, we used equality-of-coefficients tests to demonstrate the consequences of ignoring unobserved heterogeneity in macro-level models of crime.