Methods

Study design, setting and participants

We used a case–control study design; the study period was 1998–2003.

The California Department of Justice (CDOJ) provided records of handgun sales by licensed retailers and of proposed sales that CDOJ denied under laws prohibiting felons, violent misdemeanants and certain others from purchasing firearms. (In California, almost all transfers of firearms, including those between private parties, must be processed by a licensed retailer.) ATF provided records of all gun traces initiated during the study period, regardless of the location of the requesting law enforcement agency. (Supplementary material on these records is available online.)

We identified the 573 retailers who sold handguns for at least 1 year during 1998–2003 and averaged at least 50 handgun sales annually for the years they were in business during that time. For these retailers, we identified all handguns sold and then linked the datasets to identify all traces of these handguns by ATF. (Supplementary material on linkage procedures is available online.)

Although ATF traces ordinarily terminate with a gun’s first retail sale, at least 85% of recovered crime guns change hands again before their recovery by a police agency.10 Some of these transfers involve licensed retailers. We used an established procedure to link each traced handgun to the retailer whose transfer of that gun most closely preceded its recovery.1114 (Supplementary material on updating traces is available online.)

For all 573 retailers combined, we determined the rate at which handguns sold during 1998–2003 were later traced. This was expressed as traces per 1000 gun-years of exposure to the risk of being traced, analogous to person-years of exposure in a cohort study.11 (Supplementary material on the calculation of gun-years at risk is available online.) We then used a binomial distribution to identify retailers for whom the actual number of traced handguns exceeded the expected number (predicted by the aggregate tracing rate) by a margin that achieved statistical significance (p<0.05). These 60 retailers were defined as cases. We then chose a random sample of 240 controls, for a 4:1 control/case ratio.

The study was approved by the UC Davis Institutional Review Board.

Variables and data collection

Variables fell into four classes: relating to the retailer, the retailer’s handgun purchasers (including those who were denied), the local physical environment and the general socioeconomic environment. Appendix table 1 (online) gives the rationale for the inclusion of each and its hypothesised relationship to status as a case.

The records-based variables from our previous study11 are included along with variables based on data collected during site visits. Given the exploratory nature of the site visits, we were inclusive in identifying variables that might be related to status as a case. These variables were largely identified through a review of studies involving site visits to sellers of cigarettes,1516171819 alcohol202122232425 and guns2627 and preliminary visits to ∼30 retailers in Northern California. A few variables were included following the first ∼20 site visits conducted for the study. Where necessary, retailers were revisited to collect the additional data.

I conducted all site visits between August and December, 2004; I was unaware of retailers’ status as cases or controls. I assessed the local physical environment and the exterior of the premises first; this required ∼15 min. An interior inspection lasting 15–20 min was facilitated by a token purchase. Lastly, I made a subjective assessment of each retailer’s status as a case or control. I recorded my observations immediately after the visit on a pre-printed form. I did this before leaving the vicinity, to allow a return to add missing data.

Complete site visit data could not be collected when retailers had moved or gone out of business, were on restricted military installations, or were at residences or non-retail commercial premises. Exterior variables were therefore divided into two groups: “fixed”, not susceptible to modification by the occupant of the premises and likely to be valid at sites that were no longer occupied by the study subject (eg, proximity of nearest alcohol outlet); and “modifiable” (eg, security measures). Data for interior variables were collected only for retail premises that were still occupied by the study subject.

I prepared data for records-based variables for analysis. Site visit data were coded and entered by project staff, whom I trained. Dual data entry was used, with discrepancies resolved by me or a third staff member.

Statistical methods

As before,11 we replaced each retailer’s handgun sales volume with an estimate of gun-years of exposure to the risk of being traced during the study period for that retailer’s handguns. Records-based variables for subsets of handgun sales were expressed as percentages of total gun-years of exposure.

Some records-based and site visit variables correlated highly, such as sales of inexpensive handguns as determined from records and the prevalence of inexpensive handguns during the site visit. In such cases, site-visit variables were used given the study’s primary focus on the observational data. Many site visit variables were categorical. Given the relatively small number of study subjects, categories were combined in a manner that provided the best fit to the data.

Odds ratios (ORs) with 95% CIs were used to quantify relative risks. Regressions involving modifiable exterior variables were restricted to retailers who still occupied the premises; regressions involving interior variables were restricted to occupied premises for which an interior inspection could be performed. Forward stepwise regression was used to produce multivariate models, using entry and retention criteria of p⩽0.30 and p⩽0.10, respectively.

Terms for interactions between local policy on gun tracing and retailer licence type (dealer, pawnbroker) were added to the final multivariate models. A separate analysis was performed for retailers near a city with a policy of tracing recovered crime guns.

Logistic and linear regression were used to test whether denials met the criteria for status as a mediator for other variables that were associated with case status in bivariate regression.13 (Further information on mediation is in the supplementary material available online.)

The Hosmer–Lemeshow test was used to assess goodness of fit for multivariate models.

Results

Participants

The study population of 300 retailers (52.4% of 573 who met the eligibility criteria) included 60 cases and 240 controls (fig 1). We excluded 16 retailers at private residences, three at military installations, and one that could not be located. Residential retailers were less likely than others to be cases, but the difference was not statistically significant (OR 0.25, 95% CI 0.03 to 1.96). Additional restrictions applied to analyses including site visit data (fig 1).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Flow diagram showing numbers of study subjects with data available for analysis, and numbers of and reasons for exclusions.

Cases were linked to more traced handguns than controls were, whether traces were expressed as counts (cases, median 15, interquartile range (IQR) 7–32; controls, median 2, IQR 0–4) or per 1000 gun-years of exposure (cases, median 5.6, IQR 4.0–8.1; controls, median 1.1, IQR 0.0–2.0).

Main results

A wide array of variables was associated with case status (table 1). Among retailer variables, particularly large and significant associations were seen with status as a pawnbroker, median elapsed time from sale to recovery for traced guns, and the prevalence of several specific weapon types (handguns, inexpensive or used handguns, assault rifles and fantasy knives or swords), among others. Strong associations were also seen with the percentage of sales that were denied and environmental characteristics including the nature and maintenance of the business premises, the general nature of the location, and, particularly, proximity to a jurisdiction with a comprehensive gun tracing policy (OR 6.78, 95% CI 3.52 to 13.04).

Retailers who no longer sold handguns were not more likely than others to be cases. Proximity to a retail alcohol outlet was unrelated to case status, as were retailer hours of operation and most safety and security measures.

Table 2 gives multivariate regression results for records-based and exterior site visit variables. Sales of inexpensive handguns, urban location, the robbery rate and local gun tracing policy were associated with case status, whether just fixed exterior variables or all exterior variables were included. Case status was substantially more likely for every percentage point increase in denied sales.

When interior variables were included (table 3), handgun sales became a risk factor for case status; sales volume, sales of inexpensive handguns (now approximated by observed prevalence at the premises), denied sales, urban location, the robbery rate and local tracing policy persisted as risk factors.

Other analyses

Adding terms for interactions between local tracing policy or licence type did not affect the results.

Adding elapsed time from sale to recovery and restricting the analysis to retailers with traces yielded nearly identical results for the variables in tables 2 and 3 (data not shown). Elapsed time from sale to recovery was inversely associated with case status in the models that included all exterior variables (per year of increase, OR = 0.54, 95% CI 0.26 to 1.09) and both exterior and interior variables (per year of increase, OR = 0.21, 95% CI 0.05 to 0.88).

Given the importance of local tracing policy as a risk factor, a separate analysis was conducted for the 109 subjects (43 cases, 66 controls) located in or near a jurisdiction with a comprehensive tracing policy. Results were again similar (data not shown).

Table 4 summarises the mediation analysis. Of the variables that were associated with case status in a bivariate regression, seven met criteria for at least partial mediation.

Discussion

Key results

Several factors are strongly associated with handgun retailers’ risk of disproportionate sales of guns that are later used in crimes. Most are measured at the retailer level and describe the retailers themselves or their purchasers. Some, including status as a pawnshop, denied sales, sales of inexpensive handguns, multiple-gun sales, urban location and a short elapsed time from sale to recovery for traced guns, have been identified repeatedly.1112262829 Some of the risk factors identified in the observational data, including urban location and prevalence of inexpensive handguns, are available or have near analogues in the existing records.

The relationship to denied sales is particularly strong. The two models based on the most complete data also identify overall sales volume as a risk factor for disproportionate, not just frequent, sales of traced guns. As before,11 characteristics of the general socioeconomic environment seem to be less important, except for urban location (a large effect) and the robbery rate (a small effect). One further and important exception is local policy on gun tracing, which would disappear as a risk factor under the desired condition that all communities traced recovered crime guns.

Several risk factors that might represent a retailer’s attractiveness to high-risk purchasers also satisfy the computational criteria for mediation by the percentage of sales that are denied. These include sales of inexpensive handguns, licensure as a pawn broker, sales to women (who may act as straw purchasers8) and sales to law enforcement (intuitively a deterrent, and a protective factor in these data). Others, including multiple-gun sales, do not.

Such findings are consistent with the suggestion that the effects of some risk factors may be mediated by purchaser characteristics for which denied sales are a proxy measure. How might this occur? Retailers having certain characteristics may attract prospective purchasers who are at high risk of committing gun crimes themselves or who provide guns to others who have criminal intent. Such retailers might be described as “bad guy magnets.” To the extent that their prospective purchasers are prohibited from purchasing guns, their denied sales increase; denied sales become a measure of “bad guy magnetism.” Simultaneously, to the extent that the purchasers are not prohibited persons—with the results that their gun purchases are approved, with some of those guns being used in crime, and some of those crime guns being recovered and traced—these retailers risk becoming disproportionate sources of crime guns.

Factors we could not assess may also determine which retailers are disproportionate sources of crime guns. One may be word of mouth. Like other consumers, gun purchasers with criminal intent may rely on recommendations from peers (or superiors, in organised trafficking enterprises).30 Past performance is another; trafficking investigations often document long-term relationships between illegal buyers and sellers.83031

Limitations

Our findings are subject to several limitations. This is a single-state study, and generalisability may be limited. California has its own licensing and inspection programme; its handgun retailers may not be representative. Sales of handguns that fail specified safety tests, which tend to be inexpensive, were prohibited beginning in 2001. The sale of multiple handguns to an individual in any 30-day period became illegal in 2000. These changes may have affected our results for those variables. Only one set of site visit observations was collected, and by a single observer; the interobserver reliability and stability over time of the site visit data are unknown. For a variety of reasons, data were sometimes missing. We could not include long guns in the analysis, as CDOJ cannot retain sales records for them. (Such data might well show that another risk factor identified during site visits, handguns as a proportion of guns on display, could also be approximated from existing records.) The traced handguns used for our outcome measure are necessarily a subset of handguns sold by our subjects during 1998–2003 and used in crime. Most traced guns in California come from cities with comprehensive tracing policies, however, reducing the potential for selection bias.1432

Perhaps most important, we made no observational assessment of the sales practices of the retailers, and particularly of their willingness to engage in illegal sales such as straw purchases. Previous research suggests that such willingness is widespread.33 Other corrupt behaviours, such as willingness to sell guns without records, would also have been missed.3134

These limitations notwithstanding, our findings may assist efforts to prevent gun violence. The US Department of Justice stresses the need to identify retailers who are disproportionate sources of crime guns.35 Denied sales may be a useful initial screening tool, if paired with the number of background check queries to the National Instant Criminal Background Check System as a surrogate for gun sales.36 These data are available.

Screening and focused enforcement could help disrupt illegal gun commerce without unduly affecting the legitimate gun market. A focus on high-risk retailers can be effective,2829303738 although results may be only temporary (D Webster, personal communication, 2007). Given the importance of the United States as a source of crime guns elsewhere, such actions could have benefits in many countries.

Further research will be necessary to identify firmly risk factors for disproportionate sales of crime guns, particularly those involving retailer behaviour, and their mechanisms of action.