Gun Ownership as a Risk Factor for Homicide in the Home

Kellermann, Arthur, et al., "Gun Ownership as a Risk Factor for Homicide in the Home," New England Journal of Medicine, October 7, 1993, Vol. 329 No. 15, pp. 1084-91.



Kellermann and colleagues conducted a case-control study (see also this brief tutorial) claiming a gun in the home makes it 2.7 times more likely that a family member will become a homicide victim in the home.

The study examined 25 potential risk factors and six others related to gun ownership. Only six factors retained a signficant association in Kellermann's final model and they are shown below with their odds ratios:

Variable                     Adjusted Odds Ratio
Illicit drug use                5.7
Home rented                     4.4 
Any household member hit or     4.4     
  hurt in a fight in the home    
Case subject or control         3.7
  lived alone
Gun or guns kept in the home    2.7
Any household member arrested   2.5

Absolute risk

Before examining the weaknesses of Kellermann's study, for argument's sake, let's assume the 2.7 odds ratio is a reasonable estimate of the risk associated with a gun and homicide in the home. But, what is the absolute risk of this association? (For a basic primer on absolute and relative risk, and why critical readers should be alert to the distinction, see

Even if (and that's a big if), Kellermann's estimate is in the ballpark, a very conservative estimate of the actual homicide risk to each household member being killed per year, where no family member has a criminal record, is in the range of three-eighths of one-thousandth of 1 percent to three-quarters of one-thousandth of 1 percent (.000375 - .00075 percent). Over a forty year period that risk translates to between one-and-one-half hundredths to three one-hundredths of 1 percent of homicide risk for each family member (.015 - .03 percent). (See Calculation Derivations below.)

These absolute risk estimates can be reduced further by two more factors. Kellermann found gun homicide risk is 4.8 times greater with a gun kept in the home (p. 1089). "Homicide by other means was not significantly linked to the presence or absence of a gun in the home" (p. 1087). So, although we've already taken into account homicides where there was a gun in the home, the estimates have not been reduced by those who were killed by guns only. That lowers the absolute risk by 30 percent. (Roughly 70 percent of homicides involve a firearm according to FBI Uniform Crime Reports. GunCite's analysis of Kellermann's data found only 61 percent of these homicides involved a firearm after considering the factors already taken into account.) The other factor is – "Gun ownership was most strongly associated with homicide at the hands of a family member or intimate acquaintance (addjusted odds ratio, 7.8 ...). Guns were not significantly linked to an increased risk of homicide by acquaintances, unidentified intruders or strangers" (p.1087). 48% of the matched cases were murdered by a family member or intimate (again after considering the factors already taken into account). (50% is used in the calculation that follows.)

Applying the above two factors lowers the annual risk range to .000131 - .000262 percent and the 40 year range to .00525 - .0105 percent.

These estimates could be reduced yet again if we factor in previous violence and illicit drug use since Kellermann's dataset contains that information, but hopefully the reader already gets the point, and these unaccounted factors will be discussed below in other contexts. Regardless, we can see the risk of a gun in the home being used to kill a resident in an arrest-free home is quite small.

Whose gun?

In a letter to the editor in the New England Journal of Medicine, "The students of Dr. Mark Ferris's Mathematical Statistics 460" class ask, "In how many of the homicides was the victim killed with a gun that was kept in the house rather than a gun that was brought to the house by the perpetrator?" The question is a relevant one since, as the letter also notes, the study's authors had stated in part based on their findings that "people should be strongly discouraged from keeping guns in their homes [p. 1090]." In other words, advising people against keeping a gun in the home doesn't make sense unless it causes an increase in homicide risk.

Kellermann's first response to the students was incorrect: "Ninety-three percent of the homicides involving firearms occurred in homes where a gun was kept, according to the proxy respondents." In a follow-up letter (four years later) Kellermann acknowledges his error, but still fails to directly answer the question.

Kellermann's own data suggests that for all gun homicides of matched cases no more than 34% were murdered by a gun from the victim's home. (GunCite's analysis of Kellermann's data.) (The data, such as it is, is available at 34% is probably on the charitable side since it assumes all family member or intimate homicides were commited by offenders living with the victim which is highly unlikely given that not all intimates (as defined in the Kellermann dataset: spouse, parents, in-laws, siblings, other relatives, and lovers) were likely to have lived with an adult victim.

A subsequent study, again by Kellermann, of fatal and non-fatal gunshot woundings, showed that only 14.2% of the shootings involving a gun whose origins were known, involved a gun kept in the home where the shooting occurred. (Kellermann, et. al. 1998. "Injuries and deaths due to firearms in the home." Journal of Trauma 45:263-267) ("The authors reported that among those 438 assaultive gunshot woundings, 49 involved a gun 'kept in the home where the shooting occurred,' 295 involved a gun brought to the scene from elsewhere, and another 94 involved a gun whose origins were not noted by the police [p. 252].") (Kleck, Gary. "Can Owning a Gun Really Triple the Owner's Chances of Being Murdered?" Homicide Studies 5 [2001].)

Additional analysis of Kellermann's ICPSR dataset shows that just over 4½ percent of all homicides, in the three counties Kellermann chose to study, involved victims being killed with a gun kept in their own home (see derivation). This supports the conclusion that people murdered with a gun kept in their own home are a small minority of all homicides, precisely the opposite of what an uncritical reader of Kellermann's study would likely conclude. The mis-citations of Kellermann's study serve as examples: "In homes with guns, a member of the household is almost three times as likely to be the victim of a homicide compared to gun-free homes (source)." Or this page, which attempts a vigorous defense of Kellermann's study, claims, "A gun in the home make [sic] homicide 2.7 times more likely," and "the risk of getting killed was 2.7 times greater in homes with a gun than without them." Perhaps these mis-citations are inadvertent, but Kellermann attempted to identify and measure the risk factors for being murdered in the victim's home as opposed to an overall risk of gun owners or their families being murdered. The risks are different. Stated another way, murders in the home of victim residences are a subset of all murders. Kellermann's study claims a murder is roughly 3 times more likely to occur in this subset (the victim's home) to gunowners rather than non-gunowners. That is quite different from claiming a gun in the home triples one's chances of becoming a homicide victim.

Applying the above 4½ percent figure to national gun homicide totals puts the number murdered in this fashion in the realm of fatal gun accidents (the link supplies perspective).

Who's at higher risk for homicide?

(The percentages in this paragraph are based on an examination of Kellermann's ICPSR dataset.)
As mentioned, a reasonable estimate of gun victims killed by a gun from the victim's home is 34%. However, this number drops to 12.6% when households having a prior arrestee are excluded, and drops further to 7% when households with prior arrests, illicit drug use, or a history of violence are excluded. (That's 3.5% of all matched cases. Likewise, the previously mentioned 4½ percent figure of all homicides involving a victim killed by a gun in the home falls to 2.1%.)

These percentages indicate Kellermann's study essentially shows that households with guns in the hands of residents having criminal records, illicit drug use, or prior histories of violence, are at a higher risk of experiencing domestic homicides.

As a Dr. Pat Baranello writes in a letter to the editor in the New England Journal of Medicine, "What the article failed to address is that gun ownership by responsible people is not a risk factor (source)." Kellermann's response (contained in the same source) although a true statement, sidesteps the letter writer's point. Kellerman's response was, "Although we noted a degree of association among several behavioral risk factors, each contributed independently to the risk of homicide."

Households with persons having a criminal history or violence prone personality are at an increased risk for homicide, and a gun in the hands of these kinds of persons also most likely independently increases homicide risk more so than it does for law-abiding gun owning households.

Mathematically speaking, logistic regression calculates only one co-efficient per risk factor (which can be converted into an odds-ratio). If a gun in the hands of persons with criminal records or a history of violence are much more prone to commit homicide than unarmed persons without those risk factors, and the large majority of cases in a regression model had a history of violence and arrests, the odds-ratio is going to reflect the increased risk of a gun in the hands of a volatile group, rather than representing a risk factor for the general population. It's also possible that the risk of homicide by law-abiding persons could be extremely small, yet those same people with guns have a much higher risk of homicide, resulting in an odds ratio higher than what Kellermann's final model showed. Kellermann's study simply can't tell us which is the case (or neither).

Kellermann's defenders may try to claim that a link was found between guns and homicide for all 14 subgroups he studied (p. 1089), however each one of those subgroups still contained a majority of high-risk cases. (For an example to the contrary, even though living alone was found to be riskier than owning a gun, examining the ICPSR dataset shows there were 46 matched-pair cases who lived alone and had no history of arrest or violent activity. 15 cases were gun owning households versus 19 of the controls, giving a crude odds-ratio of 0.688. In this group, gunowners had a 31.2% lower risk of being murdered. But these numbers aren't conclusive of gun ownership being protective due to the lack of controls for any other factors that influence homicide victimization. It's simply an example of what might be a low-risk subgroup. Further study would be necessary.)

Counting guns

According to Gary Kleck, "The observed gun-homicide association is so weak that it could easily be due entirely to a higher rate of concealing gun ownership among controls than among cases." It would take just 2.7% of cases and 13.1% of controls (7.7% combined) falsely denying gun ownership to render the association nonsignificant. (Kleck, Gary. Targeting Guns. Hawthorne, NY: Aldine de Gruyter; 1997. P. 245.)

Why should there be a higher rate of false denial among the controls? As this law review article states:

While the problem of unwillingness to admit gun ownership is not entirely absent as to the homicide case households, it is much less acute. NEJM-1993's authors had police reports as to these households. Id. at 1084. In cases where the murder weapon was left near the body the police report would presumably so indicate. In cases where it was not, the report would presumably indicate whether the home was searched for guns, whether other occupants, if any, were asked about gun ownership, and whether registration records were consulted to see if a gun was registered to a person living in the household. The family of the deceased in the case-subject home also had time between the homicide and the interview to go through the effects of the deceased and to discover a gun, if one was owned. None of this, however, eliminates the possibility that a gun was kept in the homicide household. That possibility is far better minimized as to the homicide case households than as to the control households. There, the accuracy of NEJM-1993's gun ownership finding is entirely dependent on the truthfulness of the interviewees. (Source)

Kellermann attempts to address this potential problem, of controls underreporting gun ownership, by citing a study conducted by him and others. Kellermann writes, "A pilot study of homes listed as the addresses of owners of registered handguns confirmed that respondents' answers to questions about gun ownership were generally valid" (p. 1089).

However, another study, conducted in part due to "the small sample used in the Kellermann study," found that 10.3 percent of hunting license holders and 12.7 percent of handgun registrants denied household gun ownership in interviews. (Rafferty, Ann P. et. al. "Validity of a household gun question in a telephone survey." Public Health Reports. May-June 1995 v110 n3 p282(7).) "Thus, 11 percent would seem to be a conservative estimate of the level of false denial of gun ownership to be found among gun owners. Among the residents of the high-crime areas from which the case-control samples were drawn, the denial level would almost certainly be higher." (Gary Kleck, Michael Hogan. "National case-control study of homicide offending and gun ownership," Social Problems, May, 1999, Vol. 46 Iss. 2, pp. 275-93.)

Unaccounted confounders

Kleck believes the 2.7 association is at least partly attributable to confounding factors known to be strongly associated with both gun ownership and homicide victimization. "For example, Kellermann et al. failed to control for whether subjects were drug dealers or members of street gangs, persons who are both much more likely to own guns and far more likely to become victims of homicide." Kleck cites research claiming, "street gang members were 8.8 times more likely to own handguns than other youths, and that those who sold illicit drugs were 3.7 times more likely to own a handgun. In turn, gang members are 19 times more likely, and drug dealers at least six times more likely to be homicide victims. These risk factors would easily be large enough to create a spurious odds ratio..." (Targeting Guns, pp. 244-45.) Another source cites a paper finding "gang members are 60 times more likely than members of the general population to die through homicide." (Hutson, H. Range, et al. Feb. 3, 1994. "Adolescents and children injured or killed in drive-by shootings in Los Angeles." The New England Journal of Medicine. Vol 330, no. 5. P. 326.)


For roughly four years Kellermann refused to honor requests from legitimate scholars to examine his data, prompting law professor Daniel Polsby to comment that it was seriously debatable whether "the Kellermann results should be credited at all, because the data on which their work rests was neither deposited with the New England Journal nor otherwise made available to independent researchers" (Firearms Costs, Firearms Benefits and the Limits of Knowledge, p. 210).

Kellermann finally, purportedly, released some of his data in the form of the aforementioned ICPSR dataset as Kleck comments (personal correspondence with GunCite, Jan. 1999):

"Kellermann did finally release his dataset, or at least some version of it, submitting it to the Interuniversity Consortium for Political and Social Research (ICPSR), a data archive used by researchers whose universities belong to ICPSR. Two things are notable about the data. First, there is no way to tell if all of Kellermann's data are in this archived version of the dataset. Most conspicuously, there is not a scrap of evidence in this dataset indicating whether the guns used to kill homicide victims were guns kept in the victim's home, even though K's whole point was that keeping a gun in one's home raises one's risk of becoming a homicide victim. Second, I would have been able to determine whether all of the pieces of information gathered by K. and his team were included in the ICPSR version of the dataset if I could have examined the questionaires used to interview victims and matched controls, and the coding forms used to record information from official files. When I requested, in writing, that K. send me these materials, he did not reply. Speculation: K. did in fact have his staff code information as to whether the murder weapon was kept in the victim's home, and found virtually no evidence of homicides involving such guns. I can think of no legitimate reason why K. would not provide his interview and coding forms, and so suspect that his archived dataset does not completely reflect all of the information he gathered or tried to gather."

As for not being able to identify the source of the murder weapon, another Kellermann response was, "Few of the police reports we reviewed documented whether the weapon used in the homicide was kept in the home or brought to the scene. This was true whether the victim was shot, stabbed, or beaten to death. In most cases, the officers who completed the initial offense report were more concerned with documenting the circumstance of the homicide and the individuals involved than the origin of the weapon" (Kellermann, Arthur. "Response to Kleck." Homicide Studies 8 [2001]. Pp. 276-77). The "initial offense report" may not have been primarily concerned with locating the murder weapon, but Kellermann's original article states police detectives conducted investigations at the homicide scene (p.1085). It's hard to believe these detectives would not have made an effort to identify and resolve possesion of the murder weapon. Regardless, police reports should have documented whether the offender was living with the victim, which would also indicate the likelyhood of whether a gun in the home was employed as a murder weapon. And finally, Kellermann seemed to have less trouble documenting whether a gun was involved in a homicide in the home in his Journal of Trauma study, mentioned above.

Whether or not Kleck's speculation is correct, Kellermann has not made his original source data available for inspection and verification. This alone raises genuine credibility problems, especially in light of studies that initially received enthusiastic approval from peers, but upon closer scrutiny either serious citational or coding errors were discovered. (See GunCite's page on Michael Bellesiles for examples of the former, and this link to papers, some of which document coding errors in John Lott's concealed carry research.) (Incidentally, Kellermann was a Bellesiles defender [source].)


Kellermann writes in this previously quoted letter in the New England Journal of Medicine that he had expected his research to be corroborated by subsequent case-control studies and cites a study by Cummings and colleagues ("The association between the purchase of a handgun and homicide or suicide." 1997. American Journal of Public Health, Vol 87, Issue 6 974-978. [Abstract]) claiming that it was "generally consistent" with his results. However, Kleck claims the Cummings study "has serious methodological problems, including even fewer controls for confounding factors [and] fewer homicide victims... The authors found that handgun purchases were as strongly associated with nongun homicide as with gun homicide, suggesting that these associations were attributable to uncontrolled confounding risk factors associated with handgun purchase, rather than a causal effect of handguns on homicide victimization" (source). (Also, "this association may be, in part, attributable to confounding factors because handgun owners were also at increased risk of non-gun homicides" [Howard, Kim Ammann et al. 1999. "Beliefs about the risks of guns in the home: analysis of a national survey." Injury Prevention (full text)].)

A more recent case-control study also found an association between higher homicide rates and a gun in the home, however it too has very serious shortcomings. That study will be critiqued and added to GunCite's Gun Control Research page, but in the meantime this article provides criticism of the study.


Calculation Derivations

Yearly individual homicide rate:
    6/100,000 homicides per year = .00006

Yearly homicide rate in the home:
    .00006 × .25 homicides in the home = .000015

Yearly homicide rate in a criminal record free home:
    .000015 × .5 = .0000075

Yearly homicide homicide rate where a gun is kept in the home:
    .0000075 × .5 = .00000375 (Kellermann [p. 1088] reports 45.4 percent of cases had a gun in the home. 50 percent was used.)

Lower bound of increased risk over forty year period (upper bound = .00021 [see homicides in the home]):
    1 - (1.00000375)40 = .00015

Rates, sources, and rationales:

Homicide rate: 6 homicides per 100,000 per year (rounded up from 5.6/100,000) (2001 FBI Uniform Crime Report, p. 19)

Homicide occurring in victim's home: 25-50%
Kellermann (p. 1085) reported 23.9% of homicides occurred in the home of the victim in his three-county study. The state of California, for 1992-2001, has reported between 26 and 34 percent of victims were killed in their residence. (Only one year was at 34%, the rest were below 30%.) 19.3% for New York City in 1990 and 1991 (Tardiff, Kenneth et al., Jan. 1995, "A profile of homicides on the streets and in the homes of New York City." Public Health Reports Vol. 110 Iss. 1, pp. 13-18). Wisconsin reported 19.6% for years 1994-96. Oakland, California found 19% occurred in victim's residences in 1997. Massachusetts reported 43.7% of homicides occurring at a residence, in 1995, but provided no breakdown of victim/offender/other status. (A 50% figure is used in the final calculation to give the upper-bound of the risk estimate.)

Allowance for previously existing criminal record in the home: 50
Kellermann (Table 3, p. 1088) reports that almost 53 percent of case households had an adult member arrested, 36 percent of the cases themselves had been arrested, and almost 32 percent of case households had a household member hit or hurt in a fight in the home. (52% of the matched case households had someone arrested [GunCite's analysis of Kellermann's data].) Don Kates claims, "Looking only to prior crime records, roughly 90 percent of adult murderers had adult records, with an average adult criminal career of six or more years, including four major adult felony arrests." (Armed: New Perspsectives on Gun Control, Gary Kleck, and Don B. Kates, Prometheus Books, Amherst, 2001, p. 20.) For more statistics see GunCite's, "Gun homicides."

Derivation of percent of homicides from Kellermann's study involving a gun kept in the home of the victim:
Total homicides: 1860 (p. 1085)

Homicides in the home: 444 (p. 1085)

Homicides that Kellermann reported in his ICPSR dataset: 388

Gun-death homicides of intimates/family members where a gun was kept in the home among the 388 cases in the ICPSR dataset: 75

Gun-death homicides of intimates/family members where a gun was kept in the home among the other 56 cases not in Kellermann's ICPSR dataset:
444 - 388 = 56 cases
75 / 388 = 19.3 percent

Extrapolating .193 × 56 cases = 11 additional cases

(75 + 11) / 1860 × 100 = 4.6 percent