The Deterrent Effect of Capital Punishment: An Analysis of Daily Homicide Counts
Grogger, Jeffrey, Journal of the American Statistical Association
Social scientists and legal scholars have been troubled for decades by whether capital punishment deters murder. Although researchers from several disciplines have used widely varying methodologies to test different aspects of the deterrence issue, a clear consensus on whether and how capital punishment deters has yet to emerge. To a great extent, research on this important social-policy issue has suffered from a lack of high-quality data, and to a somewhat lesser extent from the use of weak or inappropriate statistical techniques to analyze what data do exist.
In this article, Poisson and negative binomial regression models are used to analyze daily homicide data from California from the period 1960-1963 (the last period during which frequent executions were carried out there). Such models have been frequently used in the natural sciences (Frome, Kuttner, and Beauchamp 1973; Gart 1964), and have recently found application in the social sciences as well (Cameron and Trivedi 1986; Hausman, Hall, and Griliches 1984; Mullahy 1986; Portney and Mullahy 1986). Based as they are on discrete probability distributions, such models are the logical choice in analyzing daily homicide counts. The use of these techniques allows me to use a large set of high-quality data to address a question posed previously by several authors: Does the occurence of executions exert a short-term deterrent effect on homicides (Graves 1967; Phillips 1980; Phillips and Hensley 1984; Savitz 1958)? In basing my research on a long series of data from one legal jurisdiction and on statistical models consistent with the stochastic nature of the data, I hope to eliminate many of the ambiguities of those previous studies.
The rest of the article is organized as follows. In Section 2, a brief background on previous research into the deterrent effect of capital punishment is given. The methodology is then compared with previous studies, with particular attention paid to what I believe to be this article's improvements in both data and technique. The data are then described. After a brief discussion of Poisson and negative binomial regression models, I discuss the actual estimations and present test results. Conclusions are drawn in the final section.
2. PREVIOUS RESEARCH
The efficacy of capital punishment as a deterrent to homicide has been debated for at least a century [see Stephen (1864) and Beccaria (1764) for contrasting views]. In the last decade, a considerable body of research on the deterrence question has emerged in the economics literature (see Ehrlich 1975, 1977; McManus 1985; Passell 1975). Such studies have generally been highly aggregative in nature, focusing on the relationship between homicide rates and rates of arrest, conviction, and execution in different jurisdictions or over time. This approach has yielded ambiguous results. Findings of a purported deterrent effect in U.S. time series data (Ehrlich 1975) were harshly criticized (Baldus and Cole 1975; Bowers and Pierce 1975; Passell and Taylor 1976; Zeisel 1976), and aspersions were cast on Ehrlich's (1977) and Passell's conflicting cross-sectional results in analyses by Brier and Fienberg (1980) and McManus (1985).
Recently, sociologists Phillips and Hensley (1984) suggested that the failure to find consistent deterrent effects in such studies may stem from the analysis of highly aggregated data. They posited that such effects are likely to be small, and persist for a relatively short period following a capital punishment. They proposed a return to the earlier tradition of analyzing daily data to detect short-term decreases in homicides following the occurrence of such publicized punishments as death sentences and executions. They employed linear regression models, however, to provide greater precision than those earlier studies (Graves 1967; Savitz 1958).
Their findings of a short-term deterrent effect have been questioned because of the sensitivity of their results to the statistical specification employed (Grogger 1987). …