The source: “The Uses and Abuses of Empirical Evidence in the Death Penalty Debate” by John J. Donohue and Justin Wolfers, in The Stanford Law Review, 58:3.

At the heart of the debate about whether the United States should retain capital punishment is the question of whether it deters murder. Some argue that executing murderers may actually cause more murders by desensitizing society at large to killing. But over the years, several studies have shown that killing convicted murderers does deter future murders. After reanalyzing the data used in the most prominent of these studies, how­ever, Yale law professor John J. Donohue and Wharton business professor Justin Wolfers conclude that none of them demonstrates a clear deterrent effect.

Donohue and Wolfers tested the findings of original studies by covering a different time period, intro­ducing comparison groups, changing the variables, and using other alternative analytical techniques. The fundamental difficulty with all these studies is that executions occur so rarely in the United States, they write. Thus, the number of homicides the death penalty can plausibly have caused or deterred cannot be reliably disentangled from the large year-to-year changes in the homicide rate caused by other factors.

One of the most often cited capital punishment studies is by economist Isaac Ehrlich, who changed the American debate with a 1975 analysis of national time-series data that led him to claim that each execution saved eight lives. The Supreme Court had ruled three years earlier that existing death penalty statutes were unconstitutional, but a year after Ehrlich released his study, the Court ended the death penalty moratorium in Gregg v. Georgia.

Ehrlich’s results have been questioned over the years. Though his study covered the years 1935 to 1969, his conclusion that the death penalty is a deterrent relied heavily on an upsurge in the homicide rate after 1962, combined with a fall in the execution rate during the same period. A 1978 National Academy of Sciences report pointed out that this “simple pairing” of more murders and fewer executions between 1963 and 1969 explained his results. For all of his sophisticated econometric analyses, Ehrlich did not fully take into account other influences on the homicide rate.

In a 2004 study, Hashem Dezhbakhsh and Joanna M. Shepherd analyzed the same kind of data Ehrlich considered for the period 1960 to 2000 and suggested that around 150 fewer homicides occur per execution. But this study included the same distorting mid-1960s period. And Dezhbakhsh and Shepherd’s case was also helped by the fact that homicide rates were higher during the death penalty moratorium in the mid-1970s than during the early or late years of the decade. The obvious implication that lifting the death penalty explains the difference, however, is contradicted by the fact that there was also an upsurge in murders in states where the death penalty laws did not change.

Another problem with studies such as these two is that their conclusions don’t hold up when examined against comparison cases, say Donohue and Wolfers. Canada hasn’t executed anyone since 1962, though narrow death penalty statutes remained on the books until 1998. Yet Canada’s homicide rate has moved in virtual lockstep with that of the United States. And within the United States, homicide rates in the six states that had no death penalty between 1960 and 2000 moved in close concert with those of states that did have death penalty statutes in effect during at least some portion of that period.

Despite efforts to control for a range of social and economic trends, say Donohue and Wolfers, the studies failed to capture some of the factors that influence homicide rates. Of the half-dozen or so studies that Donohue and Wolfers scrutinized, none produced statistically significant evidence of deterrence upon re-examination.

Noting the impact of such studies on public policy, the authors caution against rushing to change the law based on any study that hasn’t stood the test of time and rigorous scientific validation.