Sanford School of Public Policy

Does Gun Prevalence Affect
Teen Gun Carrying After
Philip J. Cook
Jens Ludwig
Working Papers Series
August 15, 2003
Does Gun Prevalence Affect Teen Gun Carrying After All?
Philip J. Cook
Duke University
Jens Ludwig
Georgetown University
Philip J. Cook
Sanford Institute of Public Policy
Box 90245
Duke University
Durham, NC 27708
(919) 613-7360
fax (919) 681-8288
[email protected]
Thanks to Bob Malme, Jason Ost and Stacey Phillips for excellent research assistance, to Freya
Sonnenstein for facilitating our access to the NSAM data, to the faculty and affiliated experts of
the Johns Hopkins Center for Gun Policy and Research for helpful comments, and to the Joyce
Foundation for financial support. All opinions and any errors are our own.
Author biographies
Philip J. Cook is the ITT / Terry Sanford Professor of Public Policy at Duke University, a Fellow
of the American Society of Criminology, a Member of the National Academy of Sciences
Institute of Medicine, and a research associate with the National Bureau of Economic Research.
His recent books include Gun Violence: The Real Costs (with Jens Ludwig) and The WinnerTake-All Society (with Robert H. Frank).
Jens Ludwig is Associate Professor of Public Policy at Georgetown University, Affiliated Expert
of the Johns Hopkins Center for Gun Policy and Research, and a member of the National
Consortium on Violence Research. His research focuses on gun violence, as well as peer and
neighborhood effects. He is co-editor (with Cook) of Evaluating Gun Policy.
Does Gun Prevalence Affect Teen Gun Carrying After All?
Previous research suggests that American adolescents tend to have ready access to guns, and that
the extent of adolescent misuse of guns is not much affected by local gun prevalence or
regulation. This “futility” claim is based on one interpretation of survey data from several cities,
but has not been tested directly. Here we test it using micro-data from a nationally representative
survey, the 1995 National Survey of Adolescent Males. Using the restricted geo-coded version of
these data, and conditioning on an extensive set of covariates, we find (among other results) that
the likelihood of gun carrying increases markedly with the prevalence of gun ownership in the
youth’s community. The most plausible mechanism behind this pattern is that for some
adolescents the decision to carry a gun is affected by the ease of obtaining one, which in turn is
related to their general prevalence in the community. We also analyze the propensity to carry
other types of weapons, finding that it is unrelated to the local prevalence of gun ownership. The
prevalence of both gun and other-weapon carrying is positively related to the local rate of youth
violence (as measured by the robbery rate), confirmatory evidence that weapons carrying by
youths is motivated in part by self-protection.
When it comes to gun policy, one of the few uncontroversial assertions is that unsupervised
adolescents should not carry them in public. This consensus has been codified in both national
and state laws. The Gun Control Act of 1968 prohibits licensed dealers from selling to minors
under 18 (for rifles and shotguns) or under 21 (for handguns). A subsequent amendment in 1994
established 18 as the national minimum age for purchase or possession of handguns or handgun
ammunition, with limited exceptions. Most states have also imposed minimum age requirements
for handgun purchase, and state laws governing the issuance of concealed-carry permits always
specify a minimum age (Vernick and Hepburn, 2003). These laws highlight the problem, but,
needless to say, have not solved it.
The “problem,” in essence, stems from the fact that adolescents tend to be irresponsible
and violence-prone. When their conflicts are settled with guns rather than less lethal weapons,
death is the more likely result (Cook, 1991; Felson and Messner, 1996; Kleck and McElrath,
1991; Wells and Horney, 2002; Zimring, 1968). By most accounts the national epidemic of
youth violence, which began in the mid 1980s and peaked in 1993, was so deadly because of
greatly increased gun carrying and use; during that period the homicide victimization rates for
males age 13-17 tripled, and all of the increase was due to gun assaults (Blumstein, 1995). The
percentage of youth homicides involving guns, which climbed rapidly during the epidemic, has
remained high in the years following the 1993 peak (Cook and Laub, 2002).
Efforts to reduce gun involvement by violence-prone adolescents could in principle focus
on either “demand” -- the motivation and legal consequences for carrying a gun – or “supply” -the availability of guns. The former approach was adopted in Boston’s well-known and
apparently successful Operation Ceasefire (Braga et al., 2001).1 The supply-side approach, on
the other hand, has not been well tested and remains controversial. A baseline issue is whether
the prevalence of guns in a community has any effect on teen involvement. While it may seem
obvious to some observers that guns will be more readily available to teens in communities
where guns are more common, prominent scholars in this area have argued that those teens who
are at risk for criminal gun use are in fact unaffected by local gun prevalence, since (1) people,
including teens, who need a gun for self-protection or other high-priority uses are insensitive to
the time, money or risk associated with acquiring a gun, and (2) in any case guns are sufficiently
common, and underground markets are sufficiently efficient, that such people do not have much
trouble acquiring guns (Kates and Polsby, 2000; Polsby, 1994; Wright and Rossi, 1994). We
dub this perspective the “futility hypothesis,” since it supports a view that restrictions on gun
markets that are intended to reduce misuse are futile (Cook and Leitzel, 1996).
To date the empirical evidence on whether prevalence matters is surprisingly limited. In
their pioneering study Sheley and Wright (1995) conclude that American teens generally have
ready access to guns and that feasible regulations on gun ownership or transactions would have
little effect on their involvement with guns. But that conclusion is speculative, and not based on
a direct test such as a comparison of gun-involvement rates across jurisdictions with different
levels of gun ownership.
In this paper we utilize the 1995 National Survey of Adolescent Males (NSAM), which
included items on weapons carrying and use, to conduct such a test. These data indicate that gun
carrying is remarkably prevalent among 15-17 year old American males -- ten percent reported
carrying at least once in a month -- and positively associated with the individual’s drug use,
violence involvement, and criminal activity. Our key finding is that gun carrying by this group
increases with the prevalence of gun ownership in the community. That result obtains after
controlling for individual and county-level characteristics. A plausible interpretation for why
gun carrying increases with the prevalence of gun ownership is that youthful involvement with
guns is directly influenced by gun availability.
Of course there is a possibility of reverse causation, which would be true if teen gun
carrying affected the household demand for gun ownership. To rule out the possibility of
reverse causation we estimated our gun-carrying regressions using two-stage least squares, where
the second-stage estimate is based on the variation in county-level gun prevalence in the 1990s
that is explained by variation in the proportion of the state population that lived in rural areas in
1950. It turns out that “percent rural in 1950” is strongly predictive of county-level gun
prevalence in 1995. Second-stage estimates are qualitatively similar to our original estimates,
and in particular suggest a positive effect of gun prevalence on adolescent gun carrying.2
We also analyze the propensity by adolescent males to carry other types of weapons, or
any type of weapon (including a gun). It turns out that the likelihood of carrying a weapon is not
related to community gun prevalence; in other words, gun prevalence affects the decision of
what type of weapon to carry, but not whether to carry one. The propensity to carry a weapon is
associated with indicators of the threat facing youthful males, increasing with the local robbery
rate and with residence in public housing.
We begin with a review of previous findings and interpretations, and then proceed to
introduce NSAM, reporting national patterns for gun and other weapon carrying estimated from
the NSAM data. We then develop and implement a multivariate analysis to explore the effects
of the community context on the likelihood of gun carrying while controlling for an array of
individual, household, and county-level characteristics. The concluding section relates the key
results to the policy context.
Much of what is known about adolescent gun involvement comes from the 1991 Youth Gun
Survey (Sheley and Wright, 1995). This survey interviewed 758 male students in ten inner city
high schools located in five cities across four states. It was a sample of convenience, not
intended to be representative of any well-defined population. Items on gun possession, carrying,
and access were included. Overall, 22% of respondents said they owned a gun, while 35%
carried a gun at least on occasion – 4% carried a gun all the time, 8% “most of the time,” and
23% “now and then” (p. 43). Regular carrying was typically associated with an expressed need
for self-protection (p. 70). The higher prevalence for carrying than ownership reflected the fact
that some respondents had access to guns owned by their gang, or could borrow a gun from other
sources (p. 99). When asked “How would you go about getting a gun if you decided you wanted
one?,” a majority said that they would borrow or buy from family or friends (p. 47). But Sheley
and Wright believe that theft is also an important source of guns to this group (p. 151).
Sheley and Wright were impressed by the apparent ease with which their respondents
could obtain a gun if they needed one, and concluded that gun control, as a strategy for reducing
adolescent involvement, is futile. For example, they speculate that if the general prevalence of
gun ownership were somehow reduced, and hence there were “fewer guns to steal from the gunowning public (p. 151),” then other informal and illicit sources would provide ready substitutes.
Wright has developed this “futility” argument in other writings as well. For example, he argues
that “the 200 million guns now in circulation would be sufficient to sustain roughly another
century of gun violence at the current rates…Because of the large number of guns already in
circulation, the violence-reductive effects of even fairly Draconian gun-control measures might
not be felt for decades (Wright, 1995: 64).” Wright goes on to say that “The survival motive
among the bad guys means exactly that the ‘wrong kinds of people’ will be carrying guns pretty
much all the time (p. 66).”
What might be called the “futility hypothesis,” then, is that changes in the general
prevalence of guns will have no effect on gun possession or carrying by adolescents (and, Wright
would add, dangerous adults). In economic terms, either their demand for guns is inelastic (so
the quantity demanded is unaffected by changes in the cost of obtaining one), or the supply to
this group is unaffected by gun prevalence and regulation -- or both. Arguments along these
lines are frequently offered in the national debate over gun control (Cook and Leitzel, 1996;
Jacobs and Potter, 1995), although they are rarely subjected to direct test.
There is more evidence available on how youths respond to availability, price, and
consequences in other areas of problematic behavior. Smoking and drinking by teens are both
highly sensitive to prices (Gruber and Zinman, 2001; Cook and Moore, 2001). Minimum-age
laws have been somewhat effective in reducing highway fatality rates among underage drivers
(Dee and Evans 2001). The decision by 18-19-year-old low-income women to obtain an
abortion is heavily influenced by the out-of-pocket cost (Cook et al., 1999). Levitt (1998) has
demonstrated that youths reduce criminal activity in the year that they make the transition from
juvenile to adult status. Such findings are suggestive, since they indicate that teenagers as a
group are responsive to incentives when making decisions concerning risky or criminal behavior,
but not directly on the mark.
What little is known about the elasticity of demand for guns by teens and dangerous
adults comes from surveys of convenience samples. In addition to the Sheley-Wright survey of
youths, Wright and Rossi’s (1994) survey of incarcerated adults provides evidence that at least
some high-risk adults are sensitive to the money, time or risk associated with acquiring a gun. 3
Of those prisoners who committed their crimes armed with something other than a gun, 45
percent report “against the law for me to own a gun” as at least a little important in their decision
to not use a firearm, while 28 percent report this for “too much trouble to get one” and 28 percent
report this for “costs too much” (p. 128-129). Similarly, 27 percent of those who were unarmed
during their most recent crime say that the money cost was at least a little important in this
decision, and 27 report that their ignorance about how to acquire a gun was relevant.
To test the “futility hypothesis” more directly requires observation on adolescent gun
involvement under different conditions of gun availability. The only previous direct test of this
hypothesis comes from Wintemute’s (2003) analysis of data from the 1999 Youth Risk Behavior
Surveillance survey, which provides state-level aggregate data for a total of 30 states. High
school students enrolled in grades 9 through 12 were asked whether they had carried a gun on
one or more of the past 30 days prior to the survey. Wintemute estimates a bivariate regression
and finds that each 1% increase in the prevalence of private gun ownership within a state is
associated with a 0.18% increase in teen gun carrying. However as Wintemute notes, causality
cannot be reliably inferred from these findings because the analysis does not control for any
other across-state differences. A test with greater power to rule out rival explanations for an
association between gun carrying and gun prevalence would begin with micro- rather than
aggregate-level data and control for observable heterogeneity among teens. The NSAM data
provide the basis for just such a test.
Our empirical analysis utilizes data from the National Survey of Adolescent Males (NSAM), and
in particular the cohort of participants interviewed in 1995. A sample of eligible participants –
defined as all adolescent males aged 15 to 19 living in households in the United States between
February and November 1995 – was drawn through a multi-stage sampling procedure that
assigned a higher probability of participation to Hispanic and African-American males.
Sampling weights are provided to adjust for this over-sampling process, and are used in all of the
analyses reported here. A total of 1,729 respondents completed the general NSAM
questionnaire for a response rate of 75%. Most of our empirical analysis focuses on the sample
of respondents who are under age 18 at the time of the survey, in part to focus on the population
for whom unsupervised handgun possession is unambiguously illegal. This age restriction
reduces the available sample to 1,151.
After completing a 68-minute interview-administered questionnaire about peers, social
background, and sexual attitudes and activity, 98% of respondents agreed to complete a 99-item
self-administered questionnaire regarding unusually sensitive risky behaviors.4 Of primary
interest for our analysis are two questions on the self-administered questionnaire concerning
weapons: “During the past 30 days, on how many days did you carry a gun?” (emphasis in the
original), and “During the past 30 days, on how many days did you carry some other type of
weapon such as a knife, razor or club?” We also examine items that capture the frequency with
which the respondent has been in physical fights, has been threatened with a weapon, or has
threatened others with a weapon during the 12 months prior to the interview.5
The NSAM is designed to produce unbiased estimates for the relevant segment of the
American public. Table 1 reports such estimates for males age 15-19, and for the subset of
males age 15-17.
/ Table 1 about here /
Focusing on the latter group, we see that a remarkably high fraction, 10%, carried a gun at least
once a month in 1995, and that those who carried at least once carried an average of 6 times
(0.63/0.104).6 A still greater fraction, 27%, of adolescent males age 15-17 carried a knife, razor,
or other weapon in the previous 30 days, and those who carried at least once averaged nearly 12
times. These weapons are often put to use, as suggested by the statistics on threats during the
previous 12 months: 20% had been threatened by a gun or knife, and nearly 25% had threatened
someone else.7
/ Table 2 about here /
Table 2 provides breakdowns for weapons carrying and threats along various dimensions.
Gun carrying is more likely among blacks than non-Hispanic whites, and relatively likely among
those involved in drugs, violence, and the criminal-justice system. The same patterns are evident
for carrying other weapons, with some notable exceptions; there is little difference in prevalence
between blacks and whites, and there are only small differences in other-weapon carrying
between those who are in trouble with the law and those who are not. In any event, it is
noteworthy that carrying a gun is not necessarily an alternative to carrying another weapon –
74% of those who carried a gun at least once also carried another weapon at least once.
The item on threats that is tabulated here reads “In the last 12 months, how many times
have you threatened to beat up someone or to hurt them with a weapon, such as a gun or knife,
not including any members of your own family?” Again the prevalence rates tend to be much
higher for those involved in violence, or who are in trouble with the law. The differences across
racial groups are small. Those who carried a gun (in the previous 30 days) were nearly five
times as likely to have threatened someone in the preceding year as those who did not.
It is important to note that the NSAM item on the frequency of gun carrying does not ask
the respondent’s motivation, and it is possible that some who responded in the affirmative were
engaged in lawful hunting or target shooting with responsible adults. But the prevalence patterns
tabulated here suggest that such innocuous circumstances are the exception, dominated by
carrying for self-protection or criminal purposes by delinquent youths. That conclusion is further
reinforced by the seasonal pattern of carrying. While most hunting seasons occur in the fall,8 fall
is the season with the lowest prevalence of carrying in the NSAM data.
According to the futility hypothesis, local gun prevalence and availability have no effect on
adolescent involvement with guns. The decision to possess or carry a gun will be driven by other
considerations, especially self-protection. To test this hypothesis we utilized a geo-coded
version of the NSAM data,9 which includes an indicator of the county of residence for each
respondent. The prevalence of gun ownership at the county level cannot be measured directly
from available data, but an excellent proxy is available for most large counties – the percentage
of suicides committed with guns (FS/S).
Recent research demonstrates that among the readily computed proxies that have been
used for this purpose, FS/S has the highest correlation with survey-based estimates of gun
prevalence. This proxy “outperforms” such measures as the percentage of homicides committed
with a gun, the subscription rate to gun-oriented magazines, or the prevalence of NRA
membership (Azrael, Cook and Miller, 2004). As an example, the cross-section correlation
between this proxy and survey-based estimates available for 21 states (from the Behavioral Risk
Factor Surveillance System) is .90; the corresponding correlation for the subscription rate to
Guns & Ammo is .67, and to the NRA membership prevalence is .55, and to the percent homicide
with guns, just .19.
The limitation of FS/S as a county-level proxy stems from the fact that suicide is a rare
event. Only larger counties have enough suicides to produce a reliable result. To ensure
reliability we combine suicide data for the period 1987 to 1996, and only use the measure for
counties that have at least 50 suicides during that period and at least 100,000 residents. (The
latter limitation is dictated by the fact that the National Center for Health Statistics suppresses
the county of residence in the Vital Statistics mortality data if the county has fewer than 100,000
residents.) Combining a number of years of data in this fashion is a reasonable procedure given
that the cross-section structure of gun prevalence in the United States has been highly stable over
time (Azrael, Cook, and Miller, 2004). For example, the correlation across sounties between
FS/S computed for our 10-year period, with FS/S computed on just a 5-year period (1992-6) is
As it turns out, the prevalence of gun ownership differs widely across jurisdictions in the
United States. At the state level, estimated prevalence ranges from 12 or 13% in Hawaii and
Massachusetts to over 55% in Louisiana, Alabama, and Mississippi (see Table A1, derived from
Azrael, Cook and Miller, 2004). While much of this cross-sectional variation is regional in
nature, there is also considerable variation across states within the same region, and variation
among counties within the same state.
In what follows, the effects of county gun prevalence (FS/S) on the individual NSAM
respondent’s likelihood of carrying a gun and other outcome variables are analyzed using
multivariate logistic regression that includes a long list of individual characteristics. These
individual characteristics are intended to control for his demographic and socioeconomic
circumstances. Included are the respondent’s school status, age, race, and region. Household
income is not included as such due to data problems,10 but a number of other measures of family
socioeconomic status are included: mother’s education, whether the respondent’s first language
as a child was English, whether there was an adult male living in the household when the
respondent was 14, and whether the respondent’s mother had her first child as a teenager. We
also control for the respondent’s frequency of church attendance at age 14.
Indicators of the respondent’s current behavior, as opposed to his ascribed circumstances,
are omitted from our preferred specification in order to preserve a clear causal ordering. The
exception to this rule is our measure of the respondent’s school enrollment status and grade of
enrollment. Given the strong correlation between family socio-economic status and educational
attainment (Mayer, 1997), conditioning on schooling helps limit the scope for bias from
unmeasured SES. School status is also likely to have important implications for each
respondent’s “routine activities,” which may be relevant for his opportunities and motivation for
carrying a firearm. In any case, as a check on the robustness of our findings we also present the
results of more parsimonious specifications that exclude school status.
The key findings concern the prevalence of gun ownership in the community. Also of
interest is how community context may influence the motivation for carrying a gun or other
weapon. According to Sheley and Wright (1995) the predominant motivation for gun carrying
expressed by their sample of youthful males was self-protection against violence. While this sort
of “self-help” may not be effective in practice (Wilcox, 2002), it is certainly plausible.
Presumably the felt need for self-protection differs depending on the threat of violence, which,
other things (such as the violence-proneness of the individual, or tendency to associate with bad
company) equal will be objectively greater in some communities than others. Since the robbery
rate tracks serious youth violence well (Blumstein, 2001), all regressions include the countylevel robbery rate.
We also examine the sensitivity of our results to more comprehensive specifications that
include other behavioral choices, and in particular those related to the NSAM respondent’s need
for self-protection. The challenge for more formally testing the importance of external threats as
a motivator for firearm use is that gun carrying and exposure to dangerous people or situations
are both likely to stem from the same underlying individual choices and characteristics. Below
we conduct a crude test of the “self-protection” hypothesis by expanding our baseline regression
specification to include explanatory variables measuring whether the respondent has ever been
threatened, and has ever had a weapon pulled on him.
The regression model also includes two measures of neighborhood disadvantage, which
recent research suggests may be an important predictor of gun involvement in crime (Baumer et
al., 2003). The neighborhood proxies available with the NSAM include the respondent’s
housing type (free-standing private home, row house, private apartment, public housing, or
trailer), and the condition of buildings in the respondent’s neighborhood (very rundown,
rundown, a little rundown, well kept). Both of these variables are coded by the NSAM
interviewer based on her own observations. Census-tract characteristics are unfortunately not
available even with the geo-coded version of the NSAM.
All regression results reported in this section are estimates of the parameters of equations
of the following form, calculated using the standard logit maximum-likelihood procedure:
(1) ln (pics/ 1- pics ) = b0 + b1 FS/Scs + b2 Robcs + b3 Xics + eics
where p is the (unobserved) probability of individual i, living in county c in state s, exhibiting the
behavior in question. FS/S is the proxy for the prevalence of gun ownership, namely the fraction
of suicides committed with a gun. (We also include an indicator for missing values on FS/S,
which occur if the respondent lives in a county with small population.) “Rob” is the 1995 county
robbery rate from the Uniform Crime Reports. X is a vector of measured individual
characteristics, while e is an “error” term that accounts for residual variation. If the parameter
estimates are to be consistent, this error term must be uncorrelated with the independent
variables. After reporting the results of the baseline specification, we go on to estimate
alternative specifications in order to explore the possibility of unmeasured but systematic
variation at the individual, county, or state level.
/ Table 3 about here /
The baseline results are reported in Table 3. The most notable result is that controlling
for individual characteristics, the likelihood of gun carrying is strongly positively related to gun
prevalence in the county. On the other hand, the likelihood of carrying another type of weapon,
or any type of weapon (including a gun), is unrelated to the prevalence of gun ownership -- the
coefficients on FS/S are just a fraction of the standard errors.11 The county robbery rate is
positively associated with both gun carrying and other-weapons carrying, especially the former.
Is the estimated magnitude of the effect of gun prevalence on carrying large enough to be
of interest? Using the standard formula to derive the marginal effect (dp/dx) implied by a logit
coefficient evaluated at the sample means (Greene, 1993: 639), the coefficient of 4.903 implies a
marginal effect of +.46.12 The mean values for the gun-carrying and gun-prevalence variables
are .104 and .580 respectively, so the logit coefficient implies that a 50 percent increase in
county gun ownership is associated with approximately a doubling of the prevalence of teen gun
carrying (.29 × .46 = .13). Put differently, the estimated elasticity of teen gun carrying with
respect to county gun prevalence equals +2.5.13
While we are unable to detect an effect of gun prevalence on the frequency with which
gun-carriers carry guns, our lack of supporting evidence does not rule out the possibility of such
a relationship. When the analytic sample is restricted to NSAM respondents under 18 who report
having carried a gun during the past 30 days, only 131 respondents remain. Using this sample,
the natural log of the number of times a gun was carried was regressed against FS/S and
covariates shown in Table 3. The point estimate was equal to -.3 but with a large standard error,
so that the t-statistic is only 0.2. In what follows we focus our analyses on the dichotomous
indicator for whether the teen has carried at all.
The second column of Table 3 shows that gun prevalence has little systematic
relationship with the likelihood that the teen has carried a knife or other weapon during the past
30 days. Gun prevalence also has little effect on the likelihood that the teen carries any type of
weapon, either a gun, knife, or something else (Table 3, column 3). While FS/S does not affect
the likelihood that a teen carries a weapon, the availability of guns clearly increases the
likelihood that those teens who do carry weapons choose guns (Table 3, column 4).
Table 3 also reports results on individual characteristics. Surprisingly, there is little
differentiation along these dimensions. Gun carrying is not strongly associated with age, grade,
or household socioeconomic status. Blacks and Hispanics are more likely to carry a gun than
others, although in the case of Hispanics, the effect appears to be limited to those who grew up in
English speaking homes.14 School-enrollment status has little effect on gun carrying overall,
which is the result of conflicting tendencies; students enrolled in school are less likely to carry a
weapon, but more likely to carry a gun if they do.
The results in Table 3 also show that our measures of neighborhood disadvantage are not
systematically related to gun carrying, which contrasts with the strong results reported by
Baumer et al. (2003) based on the geocoded National Crime Victimization Survey. It should be
noted that their analysis is different in several ways that may account for the apparent differences
in results: they analyze the likelihood of gun use in crimes of personal violence, rather than gun
carrying; they include adult perpetrators in their study; and their spatial analysis is based on the
location of the crime rather than the location of the perpetrator’s residence. But the difference in
results may also stem from differences in just what is being controlled for in the regression
analysis. Where we control for a large set of background characteristics, the only offender
characteristics controlled for by Baumer and his colleagues are age, race and gender (as reported
by the victim). When we estimate a parsimonious model that controls only for age, race, gender,
gun prevalence and neighborhood disadvantage, the effect on gun carrying of one of our two
neighborhood proxies -- type of the respondent’s housing (public housing, trailer, etc.) – emerges
as statistically significant. The findings of important “neighborhood effects” on gun carrying
reported by Baumer et al. could reflect to some degree the confounding of neighborhood with
offender characteristics.15
/ Table 4 about here /
Table 4 reports the coefficient estimates for FS/S from a number of different regression
specifications on weapons carrying, as well as for other dependent variables related to the
respondent’s involvement in violence. Each cell of this table includes the estimated coefficient
and standard error for FS/S; each is taken from a different logistic regression. All specifications
include the full set of individual characteristics and the county robbery rate, as with the baseline
specification in Table 3. The specifications here differ with respect to which additional control
variables are included, or which sample is used.
The first pattern to note is that our baseline results on gun carrying are remarkably robust.
Every coefficient estimate for gun carrying is highly significant, with little variation in absolute
A line-by-line synopsis of the variants on the baseline specification, reported in Table 4,
1. Repeats the result for the baseline model from Table 3
2. Replaces FS/S with the natural logarithm of FS/S.
3. Adds two county-level variables to the baseline specification -- an indicator for whether
the county of residence is included in a metropolitan area, and the population density in
1995. Both of these variables serve as further controls for urbanicity.
4. Adds to the baseline specification several additional county-level UCR crime and arrest
rates to further account for the amount of crime and disorder that might be motivating
gun carrying.
5. Adds to the baseline specification indicators for the month of the interview to control for
a possible chance confounding between the location and season of interview.
6. Adds to the baseline specification three indicators of the respondent’s behaviors that are
not directly related to weapons and violence, but which may reflect something about his
character: ever had sex with a female, ever made a woman pregnant, had five or more
drinks on at least one occasion in the previous 30 days.
7. Adds to the baseline specification three other indicators of the respondent’s behavior:
suspended from school in the last 12 months, ever arrested, ever jailed.
8. Estimates the baseline specification using a sample of just those residents living in urban
areas, where sporting uses of guns are least likely to be a factor.
9. Estimates the baseline specification using the full sample of NSAM respondents,
including those age 18 and 19.
10. Expands the baseline specification to include a full set of state fixed effects in place of
the regional indicators, in order to account for the possible importance of state legislation,
state-level differences in criminal-justice practice, or culture. With state fixed effects
included in the regression specification, the estimated effect of county gun prevalence on
teen gun carrying is identified from within-state across-county variation in gun
11. Excludes measures of the respondent’s educational attainment and school-enrollment
12. Estimates the baseline specification using only those NSAM respondents who have either
been stopped by the police or suspended from school. The coefficient on FS/S for this
“high risk” sub-sample is somewhat larger than the full-sample result, but the difference
in estimates is not statistically significant.
13. Includes “ever been threatened” as an additional explanatory variable. The estimated
effect of FS/S on gun carrying is almost identical to what is reported in our baseline
model. The “threatened” variable itself has a positive and statistically significant
relationship with gun carrying.
14. Includes both “ever threatened” and “ever had gun or knife pulled on them” as
explanatory variables in the model.
As a final specification check, we re-estimated the base-line specification against five other
types of risky behavior that have no clear causal relationship to gun prevalence, including
measures of sexual activity, alcohol bingeing, marijuana use, and cocaine use. A finding that
FS/S is statistically related to one or more of these outcomes would suggest the influence of an
unobserved variable that is correlated with FS/S, calling into question whether it is gun
prevalence per se that is influencing gun carrying. As it turns out, none of the coefficient
estimates are statistically different than zero.16
Returning to Table 4, the third column reveals that the baseline result on the likelihood of
carrying other weapons – namely, that the prevalence of gun ownership in the county has little or
no effect -- is also robust to the various alternative specifications.
Table 4 also reports the results of regressions in which the dependent variables are various
indicators of the respondent’s violence-involvement in the preceding year: got in a fight,
threatened with harm, had a gun or knife pulled on him, and threatened someone else with a gun
or knife. For the last three of these there is some indication that gun prevalence is statistically
relevant, although the results are sensitive to the specification and sample, and only occasionally
significantly different from zero at even the 10% level. In any event, it is noteworthy that all
coefficient estimates are negative. These results provide some evidence, then, that when guns
are prevalent, youths are more restrained about issuing threats. But this restraint does not
necessarily engender greater safety; the NSAM does not provide any items from which we can
adjudge the incidence of violent injury or death.
It is at least logically possible that the positive association between the prevalence of gun
ownership and the likelihood of gun carrying by adolescents is the result of reverse causation,
whereby the demand by householders for guns is influenced by gun carrying by local
adolescents. We explore this possibility by use of the standard two-stage least-squares
estimation procedure. This procedure requires an “instrument” for gun prevalence that is not
plausibly correlated with the error term in the gun-carrying regression. The ideal instrument
must pass three tests: highly correlated with gun prevalence, not affected by the current rates of
adolescent gun carrying, and uncorrelated with any omitted variables that might affect gun
carrying. The instrument that we use here exploits the fact that the cross-section structure of gun
ownership rates has been highly stable over time, and is driven in large part by each area’s local
rural tradition (Azrael, Cook and Miller, 2004). The instrument is the fraction of a state’s
population that lived in a rural area in 1950. It passes the first two tests: It is highly predictive of
each state’s gun ownership rate in 1995, and is presumably not influenced by adolescent guncarrying in 1995. We have less confidence in how it does by the third test; “rural tradition” in a
state may be correlated with other factors that influence gun-carrying rates, not all of which are
necessarily captured by the covariates in our specification. (Our specification checks described
in the previous section help allay this concern.) In any event, the second-stage estimates
utilizing the “percent rural in 1950” provide support for a conclusion that the prevalence of gun
ownership affects adolescent gun carrying.
The “instrumental variables” estimates come from estimating equations (2) and (3) using
two-stage least squares.17 In the equations Gi represents the gun ownership rate for the period
1987-1996 in respondent i’s state of residence, Ci represents whether or not individual i reports
carrying a gun in the preceding 30 days, Xi represents the vector of covariates from the baseline
specification (Table 3), and Ri represents the fraction of the state population that lived in rural
areas in 1950 for individual i’s state of residence. The first-stage equation (2) yields a predicted
value for the gun ownership rate, which is then substituted for FS/S in the second-stage equation
Gi = ∀0 + ∀1 Ri + ∀2 Xi + vi
Ci = 20 + 21 ^Gi + 22 Xi + ei
/ Insert Table 5 about here /
Table 5 shows that the instrument R has a very strong relationship with cross-sectional
variation in gun ownership rates. The F-statistic for the significance of the instrument in the firststage equation is equal to 12.70 (p<.01), while the partial R-squared is equal to .0495. The
second stage estimate (last column of Table 5) of the effect of gun prevalence is positive and
significant at the 6% level. We conclude that the key results regarding gun carrying and
prevalence are not the result of reverse causation.18
Our key results can be briefly summarized. In 1995, one in ten adolescent males nationwide
carried a gun at least once a month. The likelihood of carrying by this group differed widely
across counties according to the rate of robbery and the general prevalence of gun ownership,
even after controlling for individual and household characteristics. On the other hand, the county
prevalence of gun ownership had essentially no effect on the likelihood of an adolescent male
carrying a knife, razorblade, or some other sort of weapon -- gun prevalence affected only the
choice of weapon type. These findings are robust to a variety of statistical challenges that
attempt to distinguish the effects of gun prevalence from those of other confounding factors.
And based on the two-stage least-squares estimates, it is possible to rule out an explanation in
terms of reverse causation. Thus our findings provide suggestive evidence of a direct causal
effect of community gun ownership on adolescent involvement with guns.19
The nature of that causal influence is not identified by the statistical results, but it seems
plausible that the mechanism is gun availability. Where guns are prevalent, adolescents will find
it easier to borrow or steal or buy them from family members or other people.20 An alternative
interpretation is that in counties where guns are more common, teens tend to be more
experienced, knowledgeable, or comfortable with guns. Both these explanations grant a direct
causal role to gun prevalence, whether it operates through availability (as in the first explanation)
or learning (the second). In either case, adolescent behavior is closely linked to the context of
gun prevalence created by adults, and would be modified in response to a change in that context.
Strictly speaking, we cannot rule out the possibility that both the prevalence of gun
ownership and the propensity of youths to carry guns are influenced by some latent cultural
factor that prizes gun ownership and gun use. Note that our ancillary results rule out a number of
related mechanisms by demonstrating that gun prevalence is not related to weapons carrying,
violence involvement or various sorts of delinquency, and hence is not a proxy for a “culture of
violence” or a “scofflaw culture” that might encourage greater gun involvement by teens. Nor
(as we have shown) is the effect of prevalence on gun carrying mediated by whether the youth is
living in an urban or rural area. But there remains a logical possibility that it is a cultural factor
that we might call “gun appreciation” rather than the actual prevalence of guns that is driving our
results. That distinction is relevant in predicting the effect of regulations designed to change the
prevalence of gun ownership.
A more subtle “demand side” explanation is that teens are more likely to carry guns when
other potential predators or victims have guns. This “teen arms race” mechanism may amplify
the magnitude of the effect of gun prevalence, but would not logically cause it in the first place.
Our central finding that county gun prevalence is closely associated with teen gun
carrying appears to be at odds with the interpretation that Joseph Sheley and James Wright made
of their survey results concerning gun involvement by inner-city high-school males. But their
conclusion -- that the community context of gun ownership and availability was irrelevant to
youthful involvement --was speculative, and not based on direct evidence from comparing
different communities. Our analysis provides the first rigorous and nationally representative
evidence on this issue. Adolescents’ propensity to get involved with guns appears to be strongly
influenced by the prevalence of guns in their community, a result that holds even for the highestrisk sub-set of our NSAM sample. If this relationship is driven by the link between gun
prevalence and gun availability, which we believe to be the most plausible interpretation of our
results, then supply-side interventions are not inherently “futile” and should not be ruled out a
priori by policy makers.
Azrael, Deborah, Philip J. Cook, and Matthew Miller
State and Local Prevalence of Firearms Ownership: Measurement, Structure, and
Trends. Journal of Quantitative Criminology (forthcoming).
Baumer, Eric, Julie Horney, Richard Felson, and Janet L. Lauritsen
Neighborhood Disadvantage and the Nature of Violence. Criminology 41(1): 39-72.
Blumstein, Alfred
Youth violence, guns, and the illicit-drug industry. Journal of Criminal Law and
Criminology 86(1): 10-36.
Blumstein, Alfred
Disaggregating the violence trends. In Alfred Blumstein and Joel Wallman (eds.), The
Crime Drop. New York: Cambridge University Press: 13-44.
Braga, Anthony A., David M. Kennedy, Ilin J. Waring, and Anne Morison Piehl
Problem-oriented policing, deterrence, and youth violence: An evaluation of Boston’s
Project Ceasefire. Journal of Research in Crime and Delinquency 38(3): 195-225.
Cook, Philip J.
The technology of personal violence. In Michael Tonry (ed.), Crime and Justice: A
Review of Research (Vol. 14). Chicago: University of Chicago Press. 1-71.
Cook, Philip J. and John H. Laub
After the epidemic: Recent trends in youth violence in the United States. In Michael
Tonry, (ed.), Crime and Justice: A Review of Research, Vol. 29. Chicago: University
of Chicago Press: 1-37.
Cook, Philip J. and Jim Leitzel
‘Perversity, futility, jeopardy’: An economic analysis of the attack on gun control. Law
and Contemporary Problems 59(4): 91-118.
Cook, Philip J. and Jens Ludwig
Guns in America: Results of a Comprehensive Survey of Gun Ownership and Use.
Washington, DC: Police Foundation.
Cook, Philip J. and Jens Ludwig
“Guns and Burglary” in Jens Ludwig and Philip J. Cook, eds., Evaluating Gun Policy:
Effects on Crime and Violence. Washington, DC, Brookings Institution Press: 74-107.
Cook, Philip J., Stephanie Molliconi and Thomas B. Cole
Regulating Gun Markets. Journal of Criminal Law and Criminology. 86(1): 59-92.
Cook, Philip J. and Michael J. Moore
“Environment and Persistence in Youthful Drinking Patterns” in Jonathan Gruber, ed.,
Risky Behavior Among Youths: An Economic Analysis. Chicago: University of Chicago
Press, 375-437.
Cook, Philip J., Allan Parnell, Michael J. Moore, Deanna Pagnini.
The Effects of Short-Term Variation in Abortion Funding on Pregnancy Outcomes.
Journal of Health Economics 18(2), 241-258.
Dee, Thomas and William N. Evans
2001 Teens and Traffic Safety. in Jonathan Gruber, ed. Risky Behavior Among Youths: An
Economic Analysis. Chicago: University of Chicago Press, 121-165.
Felson, Richard and Steven F. Messner
To kill or not to kill? Lethal outcomes in injurious attacks. Criminology 34: 519-545.
Greene, William H.
Econometric Analysis (2nd ed.). New York: Macmillan.
Gruber, Jonathan and Jonathan Zinman
“Youth Smoking in the United States: Evidence and Implications. in Jonathan Gruber,
ed., Risky Behavior Among Youths: An Economic Analysis. Chicago: University of
Chicago Press, 69-120.
Huber, P.J.
The behavior of maximum likelihood estimates under non-standard conditions.
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and
Probability. Berkeley, CA: University of California Press, Volume 1: 221-233.
Jacobs, James B. and Kimberly A. Potter
Keeping guns out of the ‘wrong’ hands: The Brady Law and the limits of regulation.”
Journal of Criminal Law and Criminology 86(1): 93-120.
Kates, Don B. and Daniel D. Polsby
The Myth of the “Virgin Killer”: Law-Abiding Persons Who Kill in a Fit of Rage.
Paper delivered at the American Society of Criminology annual meeting, San
Kleck, Gary and Karen McElrath
1991 The impact of weaponry on human violence. Social Forces 69: 669-692.
Levitt, Steven
Juvenile Crime and Punishment. Journal of Political Economy 106 (December) 1156-85.
Mayer, Susan E.
What Money Can’t Buy: Family Income and Children’s Life Chances. Cambridge, MA:
Harvard University Press.
Polsby, Daniel P.
The false promise of gun control. Atlantic Monthly. March.
Sheley, Joseph F. and James D. Wright
Acquisition and Possession in Selected Juvenile Samples. NCJ 145326.
Washington, DC: National Institute of Justice, Research in Brief.
Sonenstein, Freya L., L. Ku, L.D. Lindberg, C.F. Turner, and J.H. Pleck
Changes in sexual behavior and condom use among teenaged males: 1988 to 1995.
American Journal of Public Health 88(6): 956-959.
Vernick, Jon S. and Lisa Hepburn
State and federal gun laws: Trends for 1970-99 In Jens Ludwig and Philip J. Cook
(eds.), Evaluating Gun Policy: Effects on Crime and Violence. Washington, DC:
Brookings Institution Press: 345-411.
Wells, William and Julie Horney
Weapon effects and individual intent to do harm: Influences on the escalation of
violence. Criminology 40(2): 265-296.
White, Halbert
A heteroskedasticity-consistent covariance matrix estimator and a direct test for
heteroskedasticity. Econometrica 48: 817-830.
White, Halbert
Maximum likelihood estimation of misspecified models. Econometrica 50: 1-25.
Wilcox, Pamela
2002 Self-Help? Examining the Anti-Crime Effectiveness of Citizen Weapon Possession.
Sociological Focus 35(2): 145-167.
Wintemute, Garen
Gun carrying among male adolescents as a function of gun ownership in the general
population. Annals of Emergency Medicine 41(3): 428-429.
Wright, James D.
Ten essential observations on guns in America. Society March/April: 63-68.
Wright, James D. and Peter Rossi
Armed and Considered Dangerous (2nd Ed.). Hawthorne, NY: Aldine de Gruyter.
Zimring, Franklin E.
Is gun control likely to reduce violent killings? University of Chicago Law Review 35:
Table 1. National Estimates of Risky Behaviors by Adolescent Males, 1995
Full NSAM sample
NSAM sample <18
% carried gun last 30 days
11.1 (0.8)
10.4 (0.9)
# days carried
0.80 (0.10)
0.63 (0.10)
% carried other weapon, last
30 days
# days carried
24.9 (1.3)
27.3 (1.8)
3.27 (0.26)
3.22 (0.32)
% in physical fight last 12
# fights
42.5 (1.6)
47.5 (1.7)
% been threatened with
beating or weapon, last 12
32.2 (1.4)
34.1 (2.1)
% had gun or knife pulled on
them, last 12 months
20.1 (1.2)
20.2 (1.7)
% threatened someone else
with beating or weapon, last
12 months
24.3 (1.2)
24.6 (1.5)
Weapons Carrying
Violence involvement
# times
2.33 (0.34)
2.77 (0.41)
3.32 (1.05)
3.40 (0.88)
Sex and Drugs
% ever had sex with female
55.3 (1.9)
42.6 (2.0)
% ever made girl pregnant
19.3 (1.3)
12.8 (1.3)
% ever smoked marijuana
43.1 (2.3)
39.7 (2.2)
5.5 (0.7)
5.1 (0.7)
% used cocaine, last 12
NOTES: Statistics calculated using NSAM sampling weights; parentheses present standard
errors that are adjusted for the cluster-sampled design of the NSAM.
Table 2. National Estimates for Prevalence of Weapon Carrying and Use Subgroups of
Males Aged 15-17
(Estimated prevalence for designated subgroups, and SE)
Carried gun last Carried other
30 days
weapon last 30
days (%)
someone else
with beating or
weapon, last 12
months (%)
8.8 (1.3)
15.2 (2.4)
12.4 (1.9)
14.0 (6.4)
27.3 (2.5)
24.5 (2.5)
28.2 (3.6)
32.0 (7.1)
22.6 (2.1)
29.7 (3.0)
24.6 (3.4)
35.8 (9.9)
9.4 (1.5)
14.6 (1.4)
8.4 (1.6)
29.2 (3.4)
30.0 (3.3)
23.6 (2.8)
26.6 (2.4)
23.2 (2.0)
23.7 (2.7)
Carried gun?
73.6 (5.2)
21.9 (1.9)
53.9 (5.9)
21.2 (1.4)
Fight in last
12 months?
16.7 (1.9)
4.7 (1.0)
44.3 (2.6)
11.8 (1.8)
40.6 (2.5)
10.1 (1.6)
17.2 (1.9)
6.9 (1.1)
46.8 (3.3)
43.5 (2.9)
17.1 (2.1)
14.7 (1.9)
26.8 (3.6)
6.2 (0.9)
57.6 (4.3)
51.7 (3.5)
19.6 (1.8)
17.7 (1.7)
Ever use
15.4 (2.1)
7.3 (0.9)
41.2 (2.6)
18.1 (2.0)
36.4 (2.5)
17.0 (1.8)
Size of
pulled on
Ever use
27.2 (8.1)
9.5 (0.9)
57.1 (10.6)
25.6 (1.8)
47.4 (10.3)
23.3 (1.6)
Suspended in
last 12
18.2 (2.6)
8.4 (1.0)
42.1 (3.4)
23.1 (2.0)
36.2 (3.9)
21.7 (1.7)
Ever stopped
by police?
15.6 (2.0)
7.8 (1.1)
38.1 (3.2)
21.8 (2.0)
39.1 (3.3)
17.2 (2.0)
21.8 (3.7)
7.9 (0.9)
35.3 (4.0)
25.6 (2.0)
45.0 (3.7)
20.1 (1.5)
Ever in jail?
29.9 (5.7)
8.7 (0.9)
28.7 (5.7)
27.2 (1.9)
50.0 (5.6)
22.4 (1.6)
6.1 (2.9)
12.1 (3.2)
10.5 (1.3)
10.1 (2.1)
14.3 (7.4)
37.4 (4.9)
24.6 (2.6)
31.5 (3.8)
16.7 (7.4)
35.0 (6.2)
25.4 (2.0)
19.8 (2.6)
Season of
NOTES: Point estimates are calculated using NSAM sampling weights. Parentheses contain standard errors, which
are adjusted to account for the cluster sampling design of the NSAM.
Table 3. Determinants of Gun and Other Weapon Carrying (Carried at least once in
previous 30 days), NSAM Respondents, ages 15-17, Logit Regression Results
(Estimated coefficients and standard errors)
Carried gun
Carried other
Carried either
gun or other
Carried gun
(for those
who carried a
FS/S (% Suicides with gun
in county 1987-1996)
4.903 (1.305)**
-.266 (1.156)
.079 (1.110)
7.250 (1.981)**
Indicator for suicide data
3.717 (0.847)**
-.030 (.768)
.142 (.739)
5.294 (1.248)**
County robbery rate per
1,000 residents (from UCR)
6.010 (1.877)**
2.921 (1.612)*
2.771 (1.450)*
6.310 (2.468)**
Age (15 omitted)
-.197 (.351)
.074 (.466)
-.175 (.267)
-.178 (.329)
-.127 (.264)
-.156 (.301)
-.398 (.505)
. 012 (.728)
Respondent ed attainment
(8th grade omitted)
9th grade
10th grade
11th grade
12th grade
> High school
.382 (.375)
.392 (.299)
.181 (.496)
.381 (.923)
-2.397 (1.409)*
-.081 (.271)
-.439 (.337)
.108 (.481)
-.124 (.914)
-4.085 (1.184)**
.083 (.273)
-.360 (.315)
.287 (.430)
-.072 (.897)
-4.011 (1.187)**
.581 (.527)
.978 (.581)*
-.081 (.785)
.896 (1.213)
School status (out of school
Full time
Part time
.250 (.462)
.267 (.991)
-1.083 (.305)**
-1.647 (.697)**
-1.052 (.299)**
-1.259 (.586)**
1.109 (.542)**
1.258 (1.149)
Race (White or other
.565 (.287)*
.923 (.325)**
-.453 (.310)
.105 (.315)
-.295 (.289)
.236 (.297)
1.274 (.482)**
.907 (.454)**
Mother ed attainment
(Less than HS omitted)
High school or GED
Some college
College degree (BA)
More than college degree
Mother ed missing
-.159 (.453)
-.074 (.473)
-.739 (.627)
-.072 (.569)
-.105 (.509)
-.132 (.290)
.047 (.316)
-.291 (.381)
-.010 (.418)
-.199 (.486)
-.027 (.298)
.100 (.285)
-.292 (.367)
-.015 (.401)
-.052 (.478)
-.123 (.755)
-.218 (.734)
-.699 (.919)
.314 (.946)
-.444 (.700)
Father / other adult male
lived with R at age 14
-.010 (.362)
-.183 (.235)
-.199 (.245)
-.159 (.439)
R’s mother had first child
before age 20
.293 (.282)
.114 (.173)
.095 (.161)
.317 (.390)
Urbanicity (urban omitted)
.617 (.315)*
.462 (.322)
-.065 (.241)
.009 (.315)
-.088 (.229)
.215 (.275)
1.126 (.432)**
.660 (.527)
Region (Northeast omitted)
-.106 (.445)
.012 (.437)
.409 (.454)
.319 (.364)
.446 (.400)
.690 (.331)**
.368 (.320)
.524 (.361)
.753 (.295)**
-.754 (.566)
-.590 (.642)
-.477 (.614)
R’s housing type (singlefamily home omitted)
Private apartment
Public housing
-.124 (.362)
-.722 (.508)
.761 (.731)
-.168 (.509)
-.117 (.348)
-.263 (.468)
.964 (.649)
.276 (.360)
-.163 (.335)
-.308 (.463)
1.579 (.666)**
.277 (.375)
.164 (.702)
-.322 (.719)
-.718 (.987)
-.325 (.588)
Condition of buildings in R’s
neighborhood (very rundown
A little rundown
Well kept
1.125 (.957)
.724 (1.013)
.691 (.873)
.968 (.612)
.755 (.556)
.570 (.586)
.950 (.656)
.679 (.597)
.551 (.600)
.317 (.946)
-.269 (.946)
-.032 (.833)
.907 (.370)**
.373 (.425)
.412 (.421)
.150 (.499)
-.130 (.546)
-.353 (.402)
-.055 (.360)
.189 (.292)
.101 (.281)
-.157 (.233)
.226 (.275)
-.007 (.288)
-.130 (.229)
-.494 (.739)
-.578 (.444)
.042 (.443)
-8.332 (1.586)**
-.905 (1.090)
-1.286 (1.078)
-7.167 (1.583)**
-2 × Log likelihood
R’s first language as a child
was English
Church-going at age 14
(never attended church
Less than 1 / month
1-3 times / month
Once / week
NOTES: Table presents logit coefficients from a maximum-likelihood model that uses the indicator for gun carrying
(column 1) or other-weapon carrying (column 2) as the dependent variable. Figures in parentheses are Huber-White
standard errors, adjusted for cluster sampling design of the NSAM survey (see text). All estimates are calculated
using NSAM sampling weights. R= “respondent.” ** = Statistically significant at 5%. * = Statistically significant
at 10%.
Table 4. Effects of Gun Prevalence on Weapon Carrying and Violence Involvement, Logit
Regression Results, NSAM Respondents age 15-17a
(Note: Each entry is from a different regression. Entries include the estimated coefficient and SE on “gun
prevalence” for given dependent variable and model specification)
In fight
last 12
with harm,
last 12 mo
Had gun or
knife pulled
on them last
12 mo
someone with
weapon last
12 mo
1. Base model
(as in Table 3)
2. Use log of
3. Control for
MSA residence
& county pop.
4. Control for
other UCR crime (1.512)
5. Control for
month of
6. Control for
sex, pregnant,
7. Control for
arrested, jailed
8. Sample:
respondents only
9. Sample: all
ages (15-19)
10. Include state
fixed effects
12. Sample:
Stopped by
police or
13. Control for
14. Control for
“threatened” &
“had weapon
pulled on him”
11. Exclude
for schoolingc
NOTES: a. Except in the row designated “Sample: all ages.”
b. In addition to controlling for county’s UCR crime rate for robbery as in Table 3, model includes controls for rates
per 100,000 of murder, rape, burglary, motor vehicle theft, aggravated assault, simple assault, larceny / theft,
weapons offenses, and drug offenses. Table presents logit coefficients from a maximum-likelihood model, using a
model specification similar to that reported in Table 3 unless otherwise noted. Figures in parentheses are HuberWhite standard errors, adjusted for cluster sampling design of the NSAM survey (see text). All estimates are
calculated using NSAM sampling weights. c. Excludes controls for respondent’s educational attainment and school
enrollment status.
* = Statistically significant at the 10 percent level. ** = Statistically significant at the 5 percent level.
Table 5. Determinants of Gun Carrying, NSAM Respondents, ages 15-17, Two-Stage Least
Squares Estimates (Estimated coefficients and standard errors)
First stage:
Gun Prevalence
Second stage:
% Rural, 1950
IV Gun Prevalence, 1987-1996
.00278 (.00078)**
-.778 (.397)*
County robbery rate per 1,000 residents
(from UCR)
-.143 (.137)
.545 (.268)*
Age (15 omitted)
.018 (.009)*
.026 (.012)**
-.031 (.025)
.002 (.048)
Respondent educational attainment
(8th grade omitted)
9th grade
10th grade
11th grade
12th grade
> High school
-.010 (.008)
-.027 (.010)**
-.026 (.019)
-.075 (.023)**
.050 (.030)
.057 (.024)**
.029 (.029)
.040 (.049)
.027 (.071)
.043 (.084)
School status (out of school omitted)
Full time
Part time
-.020 (.012)
-.031 (.020)
.037 (.027)
.063 (.073)
Race (White or other omitted)
-.004 (.010)
.003 (.012)
.090 (.027)**
.077 (.020)**
Mother educational attainment
(Less than HS omitted)
High school or GED
Some college
College degree (BA)
More than college degree
Mother education missing
.003 (.009)
.017 (.015)
.011 (.020)
-.018 (.015)
.001 (.014)
.036 (.034)
.019 (.053)
-.014 (.032)
.061 (.043)
.040 (.036)
Father / other adult male lived with R at
age 14
-.003 (.009)
-.014 (.031)
R’s mother had first child before age 20
.010 (.006)
.029 (.047)
Urbanicity (urban omitted)
.008 (.010)
.120 (.056)**
.035 (.027)
-.114 (.107)
Region (Northeast omitted)
.141 (.042)**
.062 (.051)
.181 (.052)**
-.138 (.067)**
-.104 (.057)*
-.105 (.096)
Carried Gun
R’s housing type (single-family home
Private apartment
Public housing
-.010 (.022)
-.004 (.010)
-.012 (.016)
.003 (.012)
-.017 (.058)
-.032 (.045)
.050 (.120)
-.034 (.064)
Condition of buildings in R’s
neighborhood (very rundown omitted)
A little rundown
Well kept
.020 (.020)
.026 (.025)
.016 (.021)
-.011 (.101)
-.024 (.077)
-.029 (.081)
R’s first language as a child was English
.015 (.010)
.016 (.020)
(never attended church omitted)
Less than 1 / month
1-3 times / month
Once / week
.004 (.010)
.004 (.010)
.013 (.009)
-.021 (.048)
-.012 (.033)
.024 (.033)
.345 (.045)**
-.409 (.183)**
Church-going at age 14
NOTES: The table presents coefficients from a two-stage least squares model, where the first stage involves
calculating the predicted value of % suicides w/ guns 1987-1996 in respondent's county as a function of % state
rural in 1950 together with the other control variables shown here. The regression coefficients can thus be
interpreted as with a standard linear probability model. Figures in parentheses are Huber-White standard errors,
adjusted for cluster sampling design of the NSAM survey (see text). All estimates are calculated using NSAM
sampling weights. R= “respondent.” ** = Statistically significant at 5%. * = Statistically significant at 10%.
Table A1. Household Prevalence of Firearms by State
From FS/S
New Jersey
New York
Rhode Island
New Hampshire
South Dakota
New Mexico
North Dakota
North Carolina
South Carolina
West Virginia
Source: Azrael, Cook and Miller (2004).
In response, policymakers sought methods for separating youths and guns. The seminal
example is Operation Ceasefire in Boston, in which a consortium of law-enforcement
agencies banded together beginning in 1996 to deliver a credible message to youth gangs
that any gun use would lead to serious legal consequences (Braga et al., 2001). Other
jurisdictions have adopted variants on this strategy. Despite these efforts, the percentage
of youth homicides involving guns, which climbed rapidly during the epidemic, has
remained high in the years following the 1993 peak (Cook and Laub, 2002).
As discussed below, this estimation approach does not logically rule out the possibility
of omitted-variable bias.
Cook, Molliconi and Cole (1995) interviewed a small group of incarcerated teens in
North Carolina, several of whom reported frequent firearms transactions influenced in
part by their financial circumstances.
Sonenstein et al. (1998), provide additional details about the survey.
These questions read: “During the last 12 months, how many times were you in a
physical fight?” “In the last 12 months, how many times has someone, not including a
member of your own family, threatened to beat you up or hurt you with a weapon, such
as a gun or knife?” “In the last 12 months, how many times has someone pulled a gun,
knife or razor on you?” “In the last 12 months, how many times have you threatened to
beat up someone or to hurt them with a weapon, such as a gun or knife, not including any
members of your own family?” “In the last 12 months, how many times have you pulled
a gun, knife or razor on someone else?”
These results cannot be compared directly with those from the Sheley-Wright survey
due to differences in item wording. As noted above, 12% of their inner-city sample
carried “all” or “most” of the time.
The fact that more of them report threatening than being threatened has its parallel in
the homicide statistics: Adolescent males kill more often than they are killed (Cook and
Laub, 2002).
For example, in the state of Wisconsin the seasonal hunting dates are as follows: white
tailed deer (gun), 11/23-12/1; black bear, 9-4/10/8 (dates vary depending on whether dog
is used); ring-necked pheasant, 10/19-12/31; ruffed grouse, 9/14-1/31; sharp-tailed
grouse, 10/19-11/10; bobwhite quail, 10/19-12/11; gray partridge, 10/19-12/31; jack
rabbit, 10/19-11/15; cottontail rabbit, 9/14-2/28; gray fox squirrels, 9/14-1/31; raccoon,
10/19-1/31; fox (all species), 10/19-2/15; bobcat, 10/19-12/31; ducks and other migratory
game birds, 9/28-12/8; American woodcock, 9/21-11/4; and early Canada goose, 9/39/15. The hunting seasons for wild turkeys and crows include periods in both the fall and
spring, while coyote, opposum, skunk, weasel, and snowshoe hare may be hunted yearround in the state. (Wisconsin Department of Natural Resources,
Because the geo-coded NSAM data are restricted-use, we sub-contracted with a
research associate on staff at the Urban Institute to conduct the analyses that we specified
on site. Our thanks to Freya Sonnenstein for her assistance in developing this
The NSAM’s measure of family income is problematic because it is reported by teens
rather than parents, and therefore perhaps not reliable, and because of potential coding
problems in the creation of the variable within the NSAM itself. (Personal
communication, Jason Ost with Freya Sonnenstein, September 10, 2002.)
The fact that the coefficient on the missing-data indicator is positive and significant
suggests that the prevalence of gun ownership is above average in counties for which the
gun-prevalence proxy cannot be computed. That makes sense, since these counties tend
to be more rural than average – lower population density, and outside of a metropolitan
area. Nationwide, gun ownership tends to higher in rural areas (Cook and Ludwig,
In comparison, the marginal effect implied by a linear probability model is around
three-fourths the size of the marginal effect implied by the logit maximum-likelihood
This elasticity is defined as the percentage change in teen gun carrying associated with
a one percent increase in gun prevalence.
Note that the coefficient on Hispanic in the second column is almost canceled by the
effect of growing up in a home where English is not the first language.
Since Baumer et al. rely on data from the National Crime Victimization Survey they are
able to control for a rich set of characteristics of the crime victim, which may be
correlated – but only imperfectly so -- with the (unobserved) socio-demographic
characteristics of the offender. In principle an alternative explanation for the discrepancy
in results between studies is that our measures of the local housing stock are imperfect
proxies for neighborhood disadvantage. But it is not clear how or even whether our
neighborhood measures are inferior to those employed by Baumer et al. Their census
tract variables presumably provide a more detailed portrait of the socio-economic
characteristics of the tract’s residents. On the other hand, our measures capture what the
NSAM coder takes to be the respondent’s “neighborhood,” which may be more closely
tied to the geographic area that comprises the respondent’s true neighborhood than does a
census tract.
For the five risky behaviors, the FS/S coefficient estimates and standard errors are as
Ever had sex with a female: -1.17 (.77)
Ever made someone pregnant: 1.14 (1.11)
Had 5 or more drinks at least once in the previous 30 days: -.52 (.82)
Every smoked marijuana: -.67 (1.46)
Used cocaine in the previous year: 3.27 (2.13)
It should be noted that we were forced to omit observations from counties of less than
100,000 population, since FS/S cannot be estimated for those counties. Hence these
results are based on a smaller sample than those from Table 3.
The instrumental variables results for our other outcome measures are qualitatively
similar to those shown in Table 4: The predicted value of FS/S only has a statistically
significant relationship with gun carrying.
Of course we do not claim that we have proven such an effect, since it is not possible to
entirely rule out the possibility that there are one or more unmeasured variables that have
an effect on the likelihood of gun carrying and happen to be correlated with gun
Transactions that do not involve licensed dealers make up the “secondary” market;
supply to this market is closely linked to the prevalence of gun ownership (Cook,
Molliconi, and Cole 1995). Direct evidence that the theft rate in residential burglary
increases with the prevalence of gun ownership is provided in Cook and Ludwig (2003).