on crime and injury found in other data. For example, Hemenway (1997a) points out that results from the NSDS imply that firearms are used defensively in every burglary committed in occupied households and in nearly 60 percent of rapes and sexual assaults committed against persons over 18 years of age; that defensive gun users thought they wounded or killed offenders in 207,000 incidents, yet only 100,000 people are treated in emergency rooms for nonfatal firearms injuries; and that hundreds of thousands of persons almost certainly would have been killed if they had not used a firearm defensively, implying that nearly all potentially fatal attacks are successfully defended against (Cook and Ludwig, 1998). Cook and Ludwig (1998), Hemenway (1997a), and others argue that these and other similar comparisons lead to “completely implausible conclusions” and go on to suggest that these inconsistencies “only buttress the presumption of massive overestimation” of defensive gun uses in the NSDS (Hemenway, 1997a:1444).
Although potentially troubling, the strong conclusion drawn about the reliability and accuracy of the DGU estimates seems premature. In some cases, it may be that the comparison statistic is subject to error. The reported prevalence of rape in the NCVS, for example, is believed to be biased substantially downward (National Research Council, 2003). More importantly, however, evidence on the apparent biases of the estimated incident rates, wounding rates, and counts of averted injuries does not directly pertain to the accuracy of the DGU estimates. Kleck and Gertz (1995), in fact, note that victimization estimates drawn using the NSDS, a survey designed to measure firearms use rather than victimization, are subject to potential reporting errors in unknown directions. Cook and Ludwig (1998) find evidence of reporting errors of crime in the firearms use surveys, with many respondents reporting that crime was involved on one hand, yet that no crime was involved on the other. Likewise, questions about whether a respondent thought he wounded or killed the offender and those eliciting subjective information on what would have happened had a gun not been used are also subject to substantial reporting biases. As noted by Kleck and Gertz (1998), respondents may be inclined to “remember with favor their marksmanship” and may tend to exaggerate the seriousness of the event.
In addition to invalid response errors, sampling variability may also play an important role in these conditional comparisons. Inferences drawn from the relatively small subsamples of persons who report using firearms defensively (N = 213 in the NSDS) are subject to high degrees of sampling error. Using data from the National Study of Private Firearms Ownership, a survey similar to the NSDS, Cook and Ludwig (1998), for example, estimate that firearms were used defensively in 322,000 rapes (rape, attempted rape, sexual assault) but report a 95 percent confidence interval of