Risk-based crime statistics: A forecasting comparison for burglary and auto theft

A major criticism of official statistics on crime is that they use inappropriate bases for computing rates. Here we investigate whether computing crime rates that contain in their denominators the number of exposures to risk of a specific event (e.g., residential burglary and auto theft) provides more accurate forecasts than employing the traditional FBI denominators as a base (e.g., the number of auto thefts and burglaries per 100, 000 persons living in the United States). Single equation, macrodynamic structural models are fitted to both the “traditional” and “alternative” forms of computing auto theft and burglary rates over the twenty-seven-year period from 1947–1974, in order to determine how well they perform on statistical and substantive grounds over the estimation period. Ex-post forecasts of the 1975–1979 observed crime rates, used to gauge the accuracy of these models, reveal few differences between the two kinds of rates in terms of how well they forecast. Both types of rates forecast well with the exogenous variables employed here and lead to similar substantive conclusions. The forecasts of the “traditional” rates are consistently, but only slightly, more accurate than those of the “alternative” rates (in most cases the differences are less than 1 percent). It is argued that the criticism of official data may be overstated and that little benefit accrues from the modification of the rate base for some purposes.