fewer completed robberies and less injury. Two forms of self-defense, namely using force without a weapon and trying to get help or attract attention, are associated with higher injury rates than taking no self-protective action.

The results suggest interesting associations: victims who use guns defensively are less likely to be harmed than those using other forms of self-protection. Whether these findings reflect underlying causal relationships or spurious correlations remains uncertain. Much of the existing evidence reports simple bivariate correlations, without controlling for any confounding factors. Kleck and DeLone (1993) rely on multivariate linear regression methods that implicitly assume that firearms use, conditional on observed factors, is statistically independent of the unobserved factors influencing the outcomes, as would be the case in a classical randomized experiment.¹² Is this exogenous selection assumption reasonable? Arguably, the decisions to own, carry, and use a firearm for self-defense are very complex, involving both individual and environmental factors that are related to whether a crime is attempted, as well as the outcomes of interest.¹³ The ability of a person to defend himself or herself, attitudes toward violence and crime, emotional well-being, and neighborhood characteristics may all influence whether a person uses a firearm and the resulting injury and crime. Thus, in general, it is difficult to be confident that the control variables account for the numerous confounding factors that may result in spurious correlations. Furthermore, the committee is not aware of any research that considers whether the finding is robust to a variety of methodological adjustments. Without an established body of research assessing whether the findings are robust to the choice of covariates, functional form, and other modeling assumptions, it is difficult to assess the credibility of the research to date.

The most obvious and fundamental limitation, however, is that the data on defensive gun uses are, as described above, potentially error ridden. Without reliable information on the prevalence of defensive gun use, researchers are forced to make implausible and unsubstantiated assumptions about the accuracy of self-reported measures of resistance. For example, Kleck, one of the most vocal critics of DGU estimates derived from the NCVS, assumes these data are fully accurate when measuring the efficacy of resistance (Kleck, 2001b; Kleck and DeLone, 1993).