Burnin’ Down the House: The 2007 Recession and the Effect on Arson

ABSTRACT

In the economics and housing literature, the narrative of the “arsonist for profit” is popularized, where a would-be offender destroys personal property for financial gain. More of than not, arson is expected to increase during periods of economic decline. Over a 12-year period, data on this crime and other demographic variables were collected for 264 counties to study the impact of the 2007 recession on this firesetting. The analyses herein find no support for the profiting arsonist. Implications for future research in this area are discussed.

Acknowledgments

I would like to thank the Criminology faculty at the University of Texas at Dallas for their useful feedback at the Spring 2016 ACJS colloquium. I would also like to extend my gratitude to Alex Piquero for his helpful comments on a previous version of this article. This manuscript has improved dramatically due to his efforts. All errors, of course, are mine alone.

Notes

¹ One could argue that arson for profit is a potentially selfish crime; burning personal property could affect the lives of family members and employees. However, Friedman (1995) notes making a rational choice to commit crime is perhaps more in one’s self-interest, rather than a selfish decision.

² Prior studies have used the number of dollars paid out for house fire insurance in a given year (e.g., Hershbarger and Miller 1978). This measure was not used in this study because (1) the data are not readily available over this course of time, (2) they are not available at the county level, (3) insurance payouts are likely to vary depending on the type of fire and the extent of the damage that occurs, and (4) insurance payouts also change from state-to-state depending on various laws. In short, using insurance payouts due to arson is subject to a large amount of bias. Using the UCR definition of arson helps ensure continuity and criminal intent across each county and state.

³ Note that for all variables gathered from the U.S. Census, linear interpolation was used to fill in missing values during intercensal years. Additionally, values from 2010 were imputed for 2011 as there is no 2020 data to set up a linear interpolation trend. Rather than dropping the year 2011 from analysis, it was retained to assess if a trend existed four years past the intervention year (2007).

⁴ Some readers may wonder why a measure of neighborhood continuity was not included, especially with its importance long established in the neighborhoods and crime literature. In preliminary analyses, a variable capturing this measure was included. However, multicollinearity diagnostics indicated that this variable was extremely harmful to the overall integrity of the model. The effect seemed to exist even after several different actions were taken to correct the problem (e.g., centering the variable). Despite these efforts, data manipulation was not successful in correcting this problem, leading to its exclusion from the model.

⁵ Initially, cluster-robust standard errors were estimated at the county level. In a separate analysis, each county’s state was also used as the cluster. No substantive differences were observed in the standard errors between the county and state clusters, although the reported standard errors were slightly larger than clustering at the state level. To lower the risk of committing a Type I error, the analyses presented in this article used cluster-robust standard errors at the county level as they are more stringent than those used when setting the cluster at the state level.

⁶ Using this type of standard error does pose a small problem. Given the clustering effect, and the large number of dummy variables used with the FENB, there are not enough degrees of freedom to perform a Wald test to determine whether the model is significantly different from zero. One possible solution is to estimate the model without the cluster-robust standard errors, take note of the Wald statistic, and report the results. Another possible solution is to use a fixed-effect Poisson regression with bootstrapped errors; results from this type of analysis have been shown to be similar to those from FENB (Allison 2009). In this article, as a check against the reported results, a Poisson regression was used with bootstrapped standard errors sampled 10,000 times (Mooney and Duval 1993). The results are substantively similar to those presented in this article.

⁷ Raftery (1995) provides helpful guidelines for assessing model fit based on AIC and BIC.

⁸ The abrupt permanent variable proved to be highly collinear with the year and id dummies and produced highly questionable results after forcing model convergence. As such, these results are suppressed in order to avoid a spurious conclusion.

Additional information

Notes on contributors

Zachary A. Powell

ZACHARY A. POWELL is a doctoral student in the Criminology program at the University of Texas at Dallas. Recently, his work has appeared in Aggression and Violent Behavior and Policing: An International Journal of Police Strategies and Management. His research interests include offender decision making, economic crime, and public policy.