Do execution moratoriums increase homicide? Re-examining evidence from Illinois

Abstract

This article revisits the event study by Cloninger and Marchesini (2006), who find that the declaration of the Illinois’ death penalty moratorium on 31 January 2000 had a homicide-promoting effect and resulted in 150 additional homicides over the period 2000–2003. We reassess the author’s identification strategy, which they refer to as ‘portfolio approach’ and which draws upon event studies in finance research. We argue that their methodology is not applicable in crime studies. Instead, we apply univariate time-series methods to test for a structural break at a known and unknown break date. We allow for unknown break points as the structural break might have occurred slightly earlier (criminals might have anticipated the moratorium) or later (due to persistence in criminal behaviour). In addition, we implement the synthetic control estimator which approximates the counterfactual homicide series by a weighted average of homicide outcomes in other US states. Based on various testing methods and two distinct data sets, we conclude that there is no empirical evidence to support the hypothesis that the Illinois’ execution moratorium significantly increased homicides.

Keywords:

• capital punishment
• event studies
• structural breaks
• homicides

JEL Classification:

• C22
• K42

Acknowledgement

The authors would like to thank Mark Schaffer, Justin Wolfers and two anonymous referees for helpful comments.

Notes

¹ We obtained these data from Justin Wolfer’s personal website. See http://users.nber.org/~jwolfers/deathpenalty.php (Accessed on 4 July 2013).

² All calculations and estimations, if not mentioned otherwise, were performed in Stata 12.0.

³ The data query tool is available at http://www.ncovr.heinz.cmu.edu/Docs/datacenter.htm (Accessed on 4 July 2013). The updated data set is available at the link provided in the supplementary data section of this paper.

⁴ To select $p$, that is, the number of lags of $\mathrm{\Delta }{y}_t$ on the right-hand side, we consider the BIC as well as the general-to-specific approach, where the lag length is successively reduced using t- and F-tests. Both BIC and the general-to-specific methodology indicate that $p=\hspace{0.17em}2$ provides best fit. Diagnostic checks suggest that the model is adequate in that the Box–Ljung statistics for residual serial correlation indicate no serial correlation and the White–Koenker test indicates homoscedasticity. The White–Koenker test was performed using ivhettest by Schaffer (2010).

⁵ These tests were carried out using the R packages forecast and urca (Hyndman and Khandakar, 2008; Pfaff, 2008; Hyndman, 2014).

⁶ Homicide counts for Kentucky and Wisconsin could not be included as controls due to missing data points during the time period studied.

⁷ More specifically, critical values depend on $\lambda ={\pi }_2\left(1-{\pi }_1\right)/\left({\pi }_1\left(1-{\pi }_2\right)\right)$ where ${\pi }_1=t_1/T$ and ${\pi }_2=t_2/T$. $T$ is the number of observations and $t_1$ and $t_2$ is the time index corresponding to the start and the end of the interval, respectively. In our case, $T=204$ and 1999:2 corresponds to $t_1=62$ and 2000:2 corresponds to $t_2=86$. Therefore, $\lambda =1.67$. Since we test ${\varphi }_i=0$ for $i=0,1,\dots ,3$, the number of restrictions is 4. The closest critical value at the 10% level tabulated in Andrews (2003) is for $\lambda =1.49$ and is equal to 10.41.

⁸ The CUSUM tests were performed using the strucchange package by Zeileis (2006) in R 2.15.1 and the cusum6 command by Baum (2000) in Stata 12.0.

⁹ We are grateful to an anonymous referee who suggested implementing the synthetic control method.

¹⁰ The donor pool includes Alabama, Arizona, Arkansas, California, Colorado, Connecticut, Delaware, Florida, Georgia, Idaho, Indiana, Kentucky, Louisiana, Mississippi, Missouri, Montana, Nebraska, Nevada, New Mexico, North Carolina, Ohio, Oklahoma, Oregon, Pennsylvania, South Carolina, South Dakota, Tennessee, Texas, Utah, Virginia, Washington and Wyoming (32 states).

¹¹ The pre-intervention mean squared error for state $i$ is defined as $1/T_0\sum _{t=1}^{{\mathrm{T}}_0}(y_{it}-\sum _{j\ne \mathrm{i}}{\hat{\omega }}_j{y}_{jt}{)}^2$ where ${\hat{\omega }}_j$ are the estimated synthetic weights and $T_0$ corresponds to the introduction of the moratorium in 2000.