Content uploaded by Mark E. Schaffer

Author content

All content in this area was uploaded by Mark E. Schaffer

Content may be subject to copyright.

DISCUSSION PAPER SERIES

ABCD

www.cepr.org

Available online at: w

ww.cepr.org/pubs/dps/DP5357.asp and www.ssrn.com/abstract=878132

www.ssrn.com/xxx/xxx/xxx

No. 5357

GUN PREVALENCE, HOMICIDE

RATES AND CAUSALITY:

A GMM APPROACH TO

ENDOGENEITY BIAS

Tomislav Kovandzic,

Mark E Schaffer and Gary Kleck

PUBLIC POLICY

ISSN 0265-8003

GUN PREVALENCE, HOMICIDE

RATES AND CAUSALITY:

A GMM APPROACH TO

ENDOGENEITY BIAS

Tomislav Kovandzic, University of Alabama at Birmingham

Mark E Schaffer, Heriot-Watt University and CEPR

Gary Kleck, Florida State University

Discussion Paper No. 5357

November 2005

Centre for Economic Policy Research

90–98 Goswell Rd, London EC1V 7RR, UK

Tel: (44 20) 7878 2900, Fax: (44 20) 7878 2999

Email: cepr@cepr.org, Website: www.cepr.org

This Discussion Paper is issued under the auspices of the Centre’s research

programme in PUBLIC POLICY. Any opinions expressed here are those of

the author(s) and not those of the Centre for Economic Policy Research.

Research disseminated by CEPR may include views on policy, but the

Centre itself takes no institutional policy positions.

The Centre for Economic Policy Research was established in 1983 as a

private educational charity, to promote independent analysis and public

discussion of open economies and the relations among them. It is pluralist

and non-partisan, bringing economic research to bear on the analysis of

medium- and long-run policy questions. Institutional (core) finance for the

Centre has been provided through major grants from the Economic and

Social Research Council, under which an ESRC Resource Centre operates

within CEPR; the Esmée Fairbairn Charitable Trust; and the Bank of

England. These organizations do not give prior review to the Centre’s

publications, nor do they necessarily endorse the views expressed therein.

These Discussion Papers often represent preliminary or incomplete work,

circulated to encourage discussion and comment. Citation and use of such a

paper should take account of its provisional character.

Copyright: Tomislav Kovandzic, Mark E Schaffer and Gary Kleck

CEPR Discussion Paper No. 5357

November 2005

ABSTRACT

Gun Prevalence, Homicide Rates and Causality: A GMM Approach

to Endogeneity Bias*

The positive correlation between gun prevalence and homicide rates has been

widely documented. But does this correlation reflect a causal relationship?

This study seeks to answer the question of whether more guns cause more

homicide, and unlike nearly all previous such studies, we properly account for

the endogeneity of gun ownership levels. We discuss the three main sources

of endogeneity bias – reverse causality (higher crime rates lead people to

acquire guns for self-protection), mis-measurement of gun levels, and

omitted/confounding variables – and show how the Generalized Method of

Moments (GMM) can provide an empirical researcher with both a clear

modeling framework and a set of estimation and specification testing

procedures that can address these problems. A county level cross-sectional

analysis was performed using data on every US county with a population of at

least 25,000 in 1990; the sample covers over 90% of the US population in that

year. Gun ownership levels were measured using the percent of suicides

committed with guns, which recent research indicates is the best measure of

gun levels for cross-sectional research. We apply our procedures to these

data, and find strong evidence of the existence of endogeneity problems.

When the problem is ignored, gun levels are associated with higher rates of

gun homicide; when the problem is addressed, this association disappears or

reverses. Our results indicate that gun prevalence has no significant net

positive effect on homicide rates: ceteris paribus, more guns do not mean

more homicide.

JEL Classification: C51, C52 and K42

Keywords: counties, crime, endogeneity, GMM, gun levels and homicide

Tomislav V. Kovandzic

Department of Justice Sciences

University of Alabama at Birmingham

UBOB 210

1530 3rd Avenue South

Birmingham, Alabama 35294-4562

USA

Email: tkovan@uab.edu

For further Discussion Papers by this author see:

www.cepr.org/pubs/new-dps/dplist.asp?authorid=163631

Mark E Schaffer

Centre for Economic Reform

and Transformation, Dept of

Economics

Heriot-Watt University

Riccarton

EDINBURGH

EH14 4AS

Tel: (44 131) 451 3494

Fax: (44 131) 451 3294

Email: m.e.schaffer@hw.ac.uk

For further Discussion Papers by this author see:

www.cepr.org/pubs/new-dps/dplist.asp?authorid=110682

Gary Kleck

College of Criminology& Criminal

Justice

Florida State University

Tallahassee

Florida 32306-1127

USA

Email: gkleck@mailer.fsu.edu

For further Discussion Papers by this author see:

www.cepr.org/pubs/new-dps/dplist.asp?authorid=163630

*We are grateful to seminar audiences in Edinburgh, Moscow and Aberdeen

for helpful comments and suggestions. The usual caveat applies.

Submitted 02 November 2005

1. Introduction

As is well known, guns are heavily involved in violence in America, especially homicide. In

2002, 63.4 percent of homicides were committed by criminals armed with guns (U.S. Federal Bureau of

Investigation 2002, p. 23). Probably an additional 100,000 to 150,000 individuals were medically treated

for nonfatal gunshot wounds (Kleck, 1997, p. 5; Annest et al., 1995). Further, relative to other

industrialized nations, the United States has higher rates of violent crime, both fatal and nonfatal, a larger

private civilian gun stock (about 90 guns of all types for every 100 Americans), and a higher fraction of

its violent acts committed with guns (Killias, 1993; Kleck, 1997, p. 64). These simple facts have led many

to the logical conclusion that America's high rate of gun ownership must be at least partially responsible

for the nation's high rates of violence, or at least its high homicide rate (e.g., Sloan et al., 1990; Killias,

1993; Zimring and Hawkins, 1999).

1

This belief in a causal effect of gun levels on violence rates, and not

merely on criminals' choice of weaponry, has likewise inclined some to conclude that limiting the

availability of guns would substantially reduce violent crime, especially the homicide rate (e.g., Clarke

and Mayhew, 1988, p. 106).

While there is a considerable body of individual-level research relevant to these questions

(summarized in Kleck, 1997), macro-level research on the possible links between gun availability and

homicide is also essential to assessing these assumptions. This is true partly because it is obviously useful

to have multiple approaches to testing a given hypothesis. Perhaps more importantly, macro-level analysis

enables estimation of the net effects of community gun availability on homicide rates. While gun

possession among aggressors in violent incidents may serve to increase the probability of a victim’s

death, gun possession among victims may reduce their chances of injury or death. Individual-level

research (e.g. Kleck and McElrath, 1990; Kleck and DeLone, 1993; Tark and Kleck 2004) can assess such

effects of gun use in crime incidents, but it is less useful for detecting deterrent effects of gun ownership

among prospective victims. Because criminals usually cannot visually distinguish people carrying

1

Detailed studies using cross-national data are, however, generally unsupportive of this conclusion, and suggest

instead that there is no significant association between national gun ownership rates and rates of homicide, suicide,

robbery, or assault (Kleck, 1997, p. 254; Killias, van Kesteren, and Rindlisbacher, 2001, pp. 436, 440).

1

concealed weapons from other people, or residences with gun-owning occupants from other residences,

deterrent effects would not be limited to gun owners, and might not even differ between owners and

nonowners (Kleck, 1988; Kleck and Kates, 2001, pp. 153-154; Lott, 2000). Because the protective effects

of gun ownership may spill over to nongun owners, the aggregate net impact of homicide-increasing and

homicide-decreasing effects of gun availability can be quantified only through macro-level research.

Such macro-level studies must, however, take account of a number of potential pitfalls. The most

important of these are reverse causality in the guns-crime relationship, errors in and validation of

measures of gun prevalence, and omitted and confounding variables.

First, gun levels may affect crime rates, but higher crime rates may also increase gun levels, by

stimulating people to acquire guns, especially handguns, for self-protection. At least ten macro-level

studies have found effects of crime rates on gun levels (Kleck, 1979; Bordua and Lizotte, 1979;

Clotfelter, 1981; McDowall and Loftin, 1983; Kleck, 1984; Magaddino and Medoff, 1984; Kleck and

Patterson, 1993; Southwick, 1997; Duggan, 2001; Rice and Hemley, 2002), and individual-level survey

evidence (not afflicted by simultaneity problems) directly indicates that people buy guns in response to

higher crime rates (summarized in Kleck, 1997, pp. 74-79). Alternatively, higher violent crime rates,

especially gun crime rates, could discourage some people from owning guns, by reminding them of the

dangers of guns.

Thus, causality in the guns-crime relationship may run in either or both directions. If such a

simultaneous relationship exists, but analysts fail to take account of it using appropriate methods, their

results will be almost meaningless. What is asserted to be the impact of gun levels on crime rates will in

fact also include the impact of crime rates on gun levels. Indeed, in their estimations the crime=>guns

relationship could quantitatively dominate the guns=>crime relationship, in which case the analysts will

misinterpret an effect of crime on gun levels as an effect of gun levels on crime.

Second, direct measurement of gun levels is subject to well-documented problems (Kleck, 2004),

and many researchers have responded by using a diverse set of proxy measures. At the most basic level,

researchers must either validate their chosen proxy against other measures – i.e., establish that it is

correlated with other measures of gun levels – or rely on the validation investigations of others. Even a

2

valid proxy will, however, still suffer from measurement error. Measurement error can lead to biased

estimates of the impact of gun levels on crime. Again, unless appropriate variables and methods are used,

the analyst may commit either Type I or Type II errors when testing whether gun levels have an impact

on crime rates.

Third, analysts must be aware of and take measures to accommodate possible confounding

variables. Omitted variable bias is a particular problem for macro-level studies of the guns-crime

relationship. Omission of confounding variables that are known to be correlated with both gun prevalence

and crime rates (e.g., poverty and unemployment) will contaminate any estimate of the impact of gun

levels on crime. Investigators need at least to include an appropriate range of controls, and ideally should

adopt estimation methods that address the problem of confounding variables that are unobservable (e.g.,

“social capital”).

This study seeks to answer the question of whether there is a causal effect of gun levels on

violence rates. Unlike nearly all previous such studies, we properly account for the endogeneity of gun

ownership levels. We set out a formal analysis of the main sources of endogeneity bias, and discuss how

an empirical researcher can address these problems using estimation and specification testing procedures

in a Generalized Method of Moments (GMM) framework for a linear model. A county-level cross-

sectional analysis was performed using data on every U.S. county with a population of at least 25,000 in

1990. Gun ownership levels were measured using the percent of suicides committed with guns, which

recent research indicates is the best measure of gun levels for cross-sectional research. Our estimation

techniques allow us to address the problems of reverse causality, measurement error and unobservable

confounding variables, and we include a wide range of controls in our estimations.

The paper is organized as follows. Section 2 formalizes the three sets of problems discussed

above – reverse causality, mismeasurement of gun ownership levels, and omitted/confounding variables –

using a simple modeling framework, and discusses how to address these problems. Section 3 critically

reviews the macro-level research on the gun-homicide relationship in light of these three sets of problems.

Section 4 sets out a GMM-based estimation strategy and describes the tests to be used, and Section 5

3

discusses the data and the specific estimation strategy used in this study. The results are presented in

Section 6, and Section 7 concludes.

2. Reverse Causality, Mismeasurement, and Confounding Variables

Consider a researcher who wants to estimate the impact of gun availability on the homicide rate.

The researcher has available cross-sectional data on localities (we consider the potential alternative of

longitudinal data in the next section). The researcher estimates the following simple linear model using

ordinary least squares,

hom

i

= β

0

guns

i

+ β

1

control

i

+ u

i

(1)

where, in self-evident notation, hom

i

is the homicide rate in locality i, guns

i

is the level of gun ownership,

control

i

is a variable that controls for some characteristic of locality i, u

i

is the error term for the homicide

equation, and for expositional convenience the constant term is suppressed. The key parameter of interest

to the researcher in equation (1) is β

0

, the impact of gun levels on the homicide rate. What are the

potential pitfalls facing this estimation strategy?

Reverse causality

Say, for the moment, that equation (1) is well-specified, but also that it captures only part of the

picture – causality in the guns-crime relationship also runs from crime to guns. In equation (2),

guns

i

= γ

0

hom

i

+ γ

3

X

i

+ u

gi

(2)

gun levels are influenced both by homicide rates – we expect γ

0

>0, i.e., people buy guns in response to

higher crime rates – and by some other covariate X.

If the researcher estimates (1) by OLS, but the true set of relationships is captured by (1) and (2)

together, then the estimated

will not be “consistent” – it will suffer from “endogeneity” or

“simultaneity” bias.

0

β

ˆ

2

The reason is that the regressor guns is itself endogenous in a system of

2

The term “bias” is used here as a shorthand for “asymptotic bias”, i.e., the difference between the probability limit

(as the sample size goes to infinity) of an estimator and the true value of the parameter. An estimator is “consistent”

if its asymptotic bias is zero. IV/GMM estimators are in general unbiased only asymptotically. See e.g., Hayashi

(2000), pp. 94-5 and chapter 3.

4

simultaneous equations, making it correlated with the error term u

i

in (1). In this case, will be biased

upwards by the positive γ

0

β

ˆ

0

. Indeed, if the reverse causality is strong enough, i.e., γ

0

is large relative to β

0

,

the researcher could find that

>0 and conclude that more guns means more crime even if the true

impact of guns on crime is negligible or negative.

0

β

ˆ

The standard answer to this problem is to estimate equation (1) using the method of instrumental

variables (IV) or the more modern framework of GMM. This requires the researcher to have a variable

that is correlated with guns (“instrument relevance”) and that is also uncorrelated with the error term u

i

in

the homicide equation (“instrument validity”). The covariate X is potentially such a variable because it

appears in equation (2) as a determinant of gun levels, but is excluded from equation (1) (hence the term

“exclusion restriction”). Note that even if equation (2) is misspecified and itself suffers from endogeneity

or other problems, the researcher can still obtain consistent estimates of the parameters of equation (1) so

long as X satisfies these requirements.

Measurement and mismeasurement of gun levels

Let us maintain the assumption that equation (1) is a well-specified description of the impact of

guns on homicide rates. Say also that there is no reverse causality issue – homicide levels have no impact

on the propensity of people in the locality to acquire guns, i.e., γ

0

=0. However, the level of gun ownership

cannot be measured exactly, and the researcher must make use of a proxy. Instead of estimating

equation (1), the researcher estimates equation (1a),

hom

i

= b

0

prox

i

+ b

1

control

i

+ u

i

(1a)

again by OLS. The (unobservable) relationship between guns and prox is given by equation (3):

prox

i

= δ

0

guns

i

+ u

pi

(3)

where the term u

pi

is the measurement error that degrades the proxy.

The consequences of measurement error for the OLS estimate

depend on the nature of u

0

b

ˆ

pi

. If it

is a textbook case of purely random measurement error, then the estimated relationship between the gun

proxy and homicide will be biased towards zero. It is also possible, however, that the proxy for guns is

5

itself endogenous, via its direct dependence on an endogenous guns or some other route that generates a

correlation between prox and u

i

.

3

If so, the researcher faces an endogeneity bias problem as well, and the

net bias on the OLS estimate

can be positive or negative.

0

b

ˆ

The researcher must in any case bear in mind that the coefficient b

0

is not the quantitative impact

of gun levels on crime rates; this is given by b

0

δ

0

, and of course δ

0

is typically not observed directly. A

test of the estimated

may enable the researcher to say if there is a statistically significant non-zero

impact of guns on crime; but without an estimate of δ

0

b

ˆ

0

, the researcher will be unable to say anything

about the practical significance of the impact.

The usual approach when employing a proxy subject to measurement error is threefold. First, the

researcher needs to validate the proxy against other, albeit imperfect, measures of gun availability.

Second, estimating equation (1a) using IV/GMM techniques will generate a consistent estimate

,

assuming that instruments for guns are available that satisfy the conditions of relevance and validity.

Third, if the estimated coefficient on the guns proxy

is significantly different from zero, the researcher

will have a qualitative estimate of the impact of guns on crime; but s/he must also have some idea of the

magnitude of δ

0

b

ˆ

0

b

ˆ

0

(e.g., from a validation exercise) in order to obtain a quantitative estimate of the impact.

Omitted/confounding variables

Now let us abandon the assumption that equation (1) is a well-specified description of the impact

of guns on homicide rates. The true relationship is one in which there is an additional characteristic of

localities that determines of homicide rates, confound, as shown in equation (1b):

hom

i

= β

0

guns

i

+ β

1

control

i

+ β

2

confound

i

+ u

i

(1b)

Say again that there is no reverse causality issue. Gun levels are, however, also influenced by the variable

confound, as in equation (2a):

guns

i

=

γ

2

confound

i

+ γ

3

X

i

+ u

gi

(2b)

3

For example, omitted variable bias as discussed below could also operate via a confounder that is omitted from the

homicide and proxy equations.

6

If the researcher estimates the original equation (1) using OLS, the estimated will again be

biased. This time, the endogeneity bias is an omitted variable bias. The absence of confound from the

estimated homicide equation (1) and its role as a determinant of guns in equation (2b) means that guns

will be correlated with the error term in (1). In other words, omitting confound from the homicide

equation makes guns an endogenous regressor.

0

β

ˆ

How to address this problem depends on whether the omitted variable is observable. If confound

is a variable that is available to the researcher, then equation (1b) may be estimated using OLS; the

researcher simply includes it alongside control as a second control variable. Often, however, confounding

variables are unobservable (e.g., pro-violence subcultural norms or social capital) and hence cannot be

included as explicit regressors. In this case, the standard approach would be the same as that for the

reverse causality problem: estimate (1) using IV/GMM, with the covariate X as the excluded instrument.

Specification testing

In all three sets of problems discussed above – reverse causality, measurement error, and

omitted/confounding variables – IV/GMM methods can, in principle, enable the researcher to avoid the

biases that would contaminate the OLS estimate of the impact of guns levels on homicide rates. Given the

mass of empirical evidence and theoretical considerations cited above, the natural starting point of the

researcher should be that these biases are likely to be present and that IV or GMM estimation is the

appropriate technique.

For the results to have any credibility, the IV specification should be both plausible and subject to

rigorous testing. The ex ante exclusion restrictions that identify the model must be consistent with prior

theory and evidence – here, the exclusion of the covariate X from the homicide equation and its presumed

correlation with guns or prox. For example, in a model of burglary rates, the percent of an area’s

population that resides in rural areas should not be omitted if prior research and theory indicates that this

variable influences burglary rates (e.g., Cook and Ludwig 2003, pp. 106-107). The ex post specification

7

testing should include tests of both instrument relevance and, if the model is overidentified,

4

instrument

validity; the latter test is sometimes called a test of overidentifying restrictions.

Although the presumption of the researcher should be that OLS estimates are likely to be biased,

the possibility that these biases are small or negligible cannot be ruled out. If this is indeed the case, then

guns (or prox) can be treated as an exogenous regressor, and estimation by OLS would be preferred to IV

because it is the more efficient (lower variance) estimator. The standard approach to this question is to

conduct a test of the endogeneity of guns. Such a test relies implicitly on a comparison of an estimation in

which guns is treated as exogenous and one in which it is treated as endogenous. For the test to have any

meaning, it is therefore essential that the OLS estimation be contrasted with a well-specified IV

estimation.

Summary

The discussion above indicates that macro-level investigations of the guns-homicide relationship

need to take account of a range of potential pitfalls. Both the empirical evidence and standard practice

suggest that the starting point should be the presumption of possible reverse causality, measurement error,

and omitted variables. We can summarize the basic requirements and procedures for addressing these

pitfalls as follows:

1. The researcher requires a proxy for gun levels that has been properly validated against other

available measures.

2. One or more instruments for guns is required. Any such instrument needs to satisfy the two

requirements of instrument validity and instrument relevance.

3. Instrument validity means an instrument should be exogenous in the econometric sense of the

term, i.e., uncorrelated with the error term in the homicide equation. Prior reasoning should also

suggest that the variable has no direct impact on homicide rates, i.e., that the variable is properly

excluded from the homicide equation. If the researcher has more than one instrument available

and the model is overidentified, then the validity of the instruments can and should be tested and

the results reported.

4. Instrument relevance means an instrument should be plausibly correlated with gun levels. The

relevance should be tested and results reported.

4

I.e., the number of excluded instruments exceeds the number of endogenous regressors.

8

5. If gun levels are to be considered exogenous, such a specification should be supported by a

properly conducted endogeneity test that uses a comparison with a well-specified IV/GMM

estimation, i.e., one that uses instruments for guns that are both relevant and valid. Testing for the

endogeneity of guns by comparing OLS to a misspecified IV estimation cannot provide evidence

that OLS is acceptable.

6. The homicide equation should include a reasonably full set of control variables, so as to reduce

the problem of omitted variables and confounding factors.

In the next section, we use these criteria to review the macro-level studies that investigate the

guns-homicide relationship.

3. Prior Research

Table 1 summarizes macro-level studies of the impact of gun ownership levels on crime rates.

Many of these studies found a significant positive association between crime or violence rates and some

measure of gun ownership, but all of these studies share at least one, and usually several, of the

methodological problems we have outlined (see Kleck, 1997, Chapter 7 for a more extensive review of

the pre-1997 research). The studies are also commonly characterized by small samples and other

problems, but it is the flaws in measurement and modeling that are most clearly consequential.

Invalid measurement of levels and changes in gun ownership

Table 1 (see note b) shows that past macro-level guns-violence studies have used a large and

diverse set of proxies for gun levels. With few exceptions (e.g., Cook, 1979; Kleck and Patterson, 1993),

researchers using these measures failed to validate them using any criterion, such as establishing that they

correlate well with more direct survey measures of gun prevalence. The validity of over two dozen of

these measures has recently been systematically assessed by measuring correlations between the proxies

and direct survey measures (Kleck, 2004). The results indicate that most of the measures used in prior

cross-sectional research and all of those used in time-series or pooled cross-section studies have poor

validity. The best indicator of gun levels was the percent of suicides committed with guns (PSG), which

correlated strongly (with correlations ranging between 0.85 to 0.95) with direct survey measures across

cities, states, and nations.

9

Despite its excellence as an indicator of cross-sectional variation in gun levels, PSG has no

validity whatsoever as a cross-temporal indicator. Kleck (2004, p. 24) demonstrated that not only do

changes in PSG fail to strongly correlate with direct survey measures of changes in gun prevalence over

time; PSG actually generally has weak negative correlations with these criterion measures. Further, Kleck

found that none of the measures used in past research or any of the rest of the two dozen indicators he

assessed were valid measures of cross-temporal variation in gun levels (pp. 20-21). The implication for

the potential alternative approach of longitudinal analysis is clear: unless some new, heretofore unknown

proxies are developed that are valid indicators of gun trends, meaningful longitudinal analysis of the

impact of gun levels on violence rates is impossible.

The recent study by Moody and Marvell (2005) illustrates this. The authors generated a state-

level panel dataset covering 1977-98 by combining direct survey measures of gun levels taken from the

General Social Surveys (GSS) with imputed values based on PSG. Their analysis used both IV methods

and Granger causality tests to detect whether changes in gun levels led to changes in violence rates or

visa-versa. This use of direct survey measures of gun levels fails, however, because (a) some states

contribute few or no respondents to the GSS sample in a given year, (b) the GSS asked gun ownership

questions in only 17 of the 22 years analyzed, and (c) since only about 1,400 people are asked the gun

questions in a typical year (not 3,000, as the authors report), GSS samples for any one state in any one

year average only about 29. As a result, random sampling and response error could easily account for

virtually all observed cross-temporal variation in the state gun prevalence figures; indeed, it is unlikely

that any of the year-to-year changes in state gun prevalence are statistically meaningful.

This becomes apparent if we examine and extend the validation exercise the authors use to justify

imputing missing state-level GSS data on percent of households with a gun (HHG) or handgun (HHGG)

for a given year using state-level data on PSG;

5

the exercise is crucial to their study since imputation of

missing HHG and HHGG observations effectively doubles their estimation sample. The authors report

that in their state-level panel, PSG is highly correlated with both HHG and HHGG, and much more

5

The data were downloaded from http://cemood.people.wm.edu/research.html.

10

correlated than alternative proxies such as gun magazine subscriptions. What the authors fail to note is

that this correlation is driven entirely by the cross-sectional correlation between PSG and the GSS gun

measures; the cross-temporal correlations are tiny. This can be easily seen by examining separately the

cross-sectional and cross-temporal correlations of HHG and PSG using their data. The left-hand panel of

Figure 1 shows that state-average PSG and state-average HHG are highly correlated – the correlation

coefficient is 0.83 and statistically highly significant. The right-hand panel, on the other hand, shows that

the correlation of annual changes (first-differences) in state PSG and state HHG is essentially nil – the

correlation coefficient is 0.06 and statistically insignificant. Note also that the very small GSS state

sample sizes mean that most of the annual changes in GSS state-level gun prevalence are implausibly

large (the standard deviation is 18). Since the authors’ analysis is based entirely on cross-temporal

variation in HHG/HHGG

6

with missing values imputed from the cross-temporal variation in PSG, their

null findings on the guns/crime link are not surprising. The authors were essentially modeling noise in the

gun data, and their analysis says little about the merits of the guns-cause-crime hypothesis.

7

Figure 1: PSG and General Social Survey HHG, State-level Data 1977-98

r = 0.8311

30

40

50

60

70

80

PSG

Average Level 1977-98

0

20

40

60

80

GSS HHG

Average Level 1977-98

r = 0.0589

-10

0

10

20

PSG

Annual Changes 1977-98

-100

-5 0

0

50

GSS HHG

Annual Changes 1977-98

6

The estimations are either in first-differences or in levels with fixed effects; in the latter case, the fixed effects

absorb all the between-state variation and leave only the within-state (cross-temporal) variation in crime levels to be

explained by the corresponding variation in guns.

7

The authors try in the paper to address the issue of attenuation bias with a calibration exercise, but they

underestimate the scale of the bias by calibrating to the wrong gun measure (gun levels, instead of changes in gun

levels as used in their regressions.)

11

In sum, with the exception of the few studies that used PSG, indexes including PSG, or direct

survey measures in cross-sectional research (e.g., Cook, 1979; Kleck and Patterson, 1993), the supposed

gun-crime associations estimated in nearly all past research are uninterpretable on the simple grounds that

gun levels were not actually measured.

Many measures have flaws that go beyond merely being imperfect indicators of gun levels.

Measures such as the percent of homicides (or robberies, or aggravated assaults) committed with guns are

vulnerable to the possibility of artifactual associations with crime rates. For example, Hemenway and

Miller (2000) used Cook’s (1979) “gun density” index, which is the average of (1) the percent of

homicides committed with guns, and (2) PSG. The reason for the significant associations found between

the Cook measure and homicide rates across 26 nations (and the absence of such when just PSG was

used), is that both national homicide rates ([gun homicides + nongun homicides]/population) and the

percent of homicides committed with guns ([gun homicides/total homicides] x 100%) contain a common

component in their numerators: the number of gun homicides. In other words, the “gun density” index is

endogenous: when used as a proxy for gun levels, u

p

, the error term that degrades the proxy in

equation (3), is correlated by construction with the homicide rate hom, and hence with u, the error term in

the homicide equation (1). Had the authors employed instrumental variables or a related technique, they

might, in principle, have been able to obtain unbiased estimates. But of course the better approach,

regardless of estimation technique employed, is to use a proxy that is not biased by construction.

The “percent of crimes with a gun” proxy has another flaw. This indicator reflects not only the

availability of guns but also the preference of the criminal population for using guns (Brill, 1977, pp. 19-

20). While availability certainly affects how often criminals use guns, the “lethality” of offenders, i.e.,

their willingness to inflict potentially lethal injury on others, affects weapon choice as well – criminals

willing to use lethal weapons are also more willing to inflict lethal injury (Cook, 1982, p. 248; Wright and

Rossi, 1983, pp. 189-211). Consequently, these macro-level indicators can confound gun availability with

the average lethality of the criminal population, producing guns/homicide associations that are virtual

tautologies, reflecting nothing more than the truism that populations that are more lethal are more likely

12

to commit lethal acts. Here we have an example in which the same problem can be described as either

measurement error or omitted variable bias. Average lethality is an omitted and unobservable variable

(confound in equation (1b)), and is also a component of the measurement error that degrades the proxy (u

p

in equation (3)). Again, this problem can be countered using instrumental variable estimation techniques,

but again, it is preferable to use a better proxy whatever the estimation technique employed.

Finally, virtually all studies using a proxy for gun levels have failed to calibrate the proxy used to

a survey-based measure of gun prevalence. That is, they failed to adjust for the fact that there is not

necessarily a one-to-one correspondence between the gun proxy and actual gun levels. Moody and

Marvell (2003) showed that the failure of two such studies (Duggan, 2000; Cook and Ludwig, 2003) to

calibrate their proxy led the authors to make claims of policy relevance that were simply unfounded.

Endogeneity and reverse causality

Most guns-violence studies do nothing whatsoever to deal with this problem (e.g., Hemenway

and Miller, 2000; Killias, 1993; Miller, Azrael, and Hemenway, 2002). Table 1 shows that of 31 total

studies, twenty had no research design features that would help the analyst distinguish the possible

positive effects of gun levels on crime rates from the possible positive effects of crime rates on gun levels.

Three studies (Southwick, 1999; Duggan, 2003; Moody and Marvell, 2005) used the weaker notion of

Granger causality, in which longitudinal data are used to establish whether past gun levels help predict

current crime rates, but their findings are uninterpretable for the reasons cited above: there are no useable

longitudinal data on gun levels that have been shown to be reliable enough for statistical analysis.

Of the 10 studies that attempted to address the causality problem using a structural approach,

eight clearly failed because they used inappropriate methods, e.g., estimating unidentified models. This

means that if crime rates do influence decisions to acquire guns, the findings of all but a handful of prior

studies (e.g., Kleck and Patterson, 1993) are uninterpretable on the grounds that the statistical models on

which they were implicitly based were unidentified. Even the studies that did estimate identified models

using appropriate techniques failed to report tests of instrument relevance and instrument validity, a

problem that we might label “underreporting” of structural causality methods.

13

Hoskin (2001), for example, estimated a simultaneous equations model of homicide rates and gun

availability across 36 nations, but the model was almost certainly underidentified. The exclusion

restrictions used to identify the model were arbitrary and implausible (indeed, he never made them

explicit), and directly contradicted the author’s own theoretical assertions.

8

Nor did Hoskin report any

tests of instrument validity that would indicate their adequacy, or provide a discussion of the requirement

of instrument relevance.

The Stolzenberg and D’Alessio (2000) study is another example of the underreporting problem.

They used an appropriate test for endogeneity, the Hausman test, to support a specification of a crime

equation in which gun levels is treated as exogenous, but the meaning of their test is doubtful (p. 1475).

The test’s utility depends crucially on the specification of the IV estimation in which gun levels are

instrumented, and these authors did not report what instrumental variables they used, let alone whether

they passed tests of validity and relevance. As a result, it is impossible to place any confidence in their

Hausman test results, and therefore in the conclusions they draw from their estimated crime equation.

Omitted/confounding variables

Most studies also use minimal or no controls for possible confounding variables. As a point of

reference, Kleck and Patterson (1993, pp. 259-260) controlled for as many as 36 potential confounders,

beyond their gun level measure and 19 dummy variables representing gun laws. In contrast, Cook and

Ludwig’s (2003) county-level instrumental variable analysis included just three control variables (beyond

county and year dummies), Hoskin (2001, p. 586) included just three controls in his homicide equation,

and Hemenway and Miller (2000) did not control for a single confounding variable in their homicide

models.

Due to uncertainty about exactly which macro-level attributes of places affect crime rates, it is

impossible to know exactly which variables in a guns-violence study might be confounding variables, i.e.,

8

To achieve identification in his homicide equation, Hoskin excluded (1) population density, (2) the percent of the

population that was male and aged 15 to 34 (p. 584), and (3) an East Asia dummy. Yet just a few pages earlier he

had asserted, quite plausibly, that the first two of these variables should affect homicide rates, and his discussion of

the third is limited to the remark that both homicide and firearms ownership rates are low in East Asia (pp. 580-1).

14

factors that affect crime rates but are also correlated with gun levels or gun control laws. That is, it is

uncertain which variables might either generate spurious associations among these variables or suppress

or distort genuine causal effects. The consequences of failing to control confounders can be quite serious

– biased parameter estimates – while the consequences of wrongly including irrelevant variables are more

mild – somewhat inflated standard errors of coefficients. Thus, the most sensible procedure is to control

for as many likely and observable crime determinants as is reasonable, within the limits imposed by

sample size and assuming that multicollinearity does not preclude doing so, and to employ IV/GMM

techniques in order to address the problem of unobservable and hence omitted confounders.

4. IV/GMM, Instrument Validity, and Instrument Relevance

In this section we outline an estimation strategy using modern econometric techniques and GMM

methods in particular. GMM can be applied to nonlinear as well as linear problems; we use a linear model

for simplicity of exposition as well as ease of implementation. We begin by illustrating the two

requirements of instrument validity and instrument relevance in the simplest version of model (1). There

is a single explanatory variable, guns,

hom

i

= β

0

guns

i

+ u

i

(1c)

and a single excluded instrument, X. The standard IV estimator of β

0

is =

0

ˆ

β

∑

∑

iii

gunsXhomX

i

.

The proof that

is a consistent estimator of β

0

ˆ

β

0

is straightforward:

0

ˆ

plimβ

)plim(

ii

∑∑

= gunsXhomX

ii

))(plim(

0

∑

∑

+β=

iiiii

gunsXugunsX

)plim(

i0

∑∑

+β=

iii

gunsXuX

)plim()plim(

N

1

i

N

1

0

∑

∑

+β=

iii

gunsXuX

. This will be equal

to β

0

if the second term is zero, which will be the case if X satisfies two conditions. First, if E(X

i

u

i

)=0 (X

is uncorrelated with the error term; validity), then

0plim

N

1

=

∑

ii

uX

. Second, if E(X

i

guns

i

)≠0 (X is

correlated with guns; relevance), then

0plim

N

1

≠

∑

ii

gunsX

. Both of these are “moment conditions”;

they relate to statistical moments of guns, X and u.

15

A modern and increasingly popular approach to the problem of estimation with endogenous

regressors is the Generalized Method of Moments or GMM (Hansen, 1982).

9

GMM provides a unified

framework for estimation and testing that is naturally suited to empirical problems where endogeneity and

instrument validity are central. The issue of instrument relevance is one that has attracted a great deal of

econometric research in recent years and new findings are appearing regularly. Nevertheless, there are

enough established results to provide empirical researchers with some practical guidelines for how to

detect and address problems of instrument relevance.

Estimation, testing and instrument validity in the Generalized Method of Moments

As its name implies, GMM is a generalization of the method of moments (MM), a much older

technique introduced by Karl Pearson in 1894. The essence of MM is straightforward. The researcher has

theoretical priors which imply theoretical or population moments, i.e., characteristics of the population

that are implied by the researcher’s model. The researcher also has data available, and can calculate

sample moments using these data. The sample moments depend not only on the data, but also on the

unknown parameters of the model that the researcher wants to estimate. The researcher’s estimates of the

parameters are the values that make the sample moments match the assumed population moments.

We use model (1c) to illustrate MM estimation. There is a single explanatory variable, guns. The

researcher also has a single theoretical moment condition, namely that the variable X is exogenous: it is

orthogonal to, i.e., uncorrelated with, the error term u, E(X

i

u

i

)=0. This is referred to as an “orthogonality

condition”. Note that the researcher has not imposed any priors about the exogeneity of guns; in

particular, by choosing E(X

i

u

i

)=0 as the single theoretical orthogonality condition, the researcher is

allowing for the possibility that that guns may be endogenous, i.e., E(guns

i

u

i

)≠0. The MM estimate of the

parameter β

0

, is the value that makes the sample moment corresponding to E(X

0

ˆ

β

i

u

i

)=0 also equal to

zero. The sample counterpart to the error term u

i

is the residual, defined in the usual way as

9

The literature on GMM is now vast and many good expositions are available. Wooldridge (2001) provides an

easily accessible introduction to GMM and its applications. Hayashi (2000) is an advanced text that sets out many of

the tests and results used and cited in this paper. Baum et al. (2003) set out the basics of IV and GMM estimation

and specification testing, and describe the set of extended Stata estimation and testing routines used here.

16

i0ii

ˆ

ˆ

gunshomu β−≡

. The sample moment condition is therefore that the sample mean of ( ) is

zero, i.e.,

ii

uX

ˆ

0

ˆ

N

1

=

∑

ii

uX

. This is just one equation in one unknown, i.e., the model is exactly identified.

To solve the equation for

we substitute for the residual to obtain

0

ˆ

β

0)

ˆ

(

0i

N

1

=β−

∑

gunshomX

i

, and

after simplifying and rearranging, we have the MM estimate of the impact of guns on homicide:

∑∑

=β

iii

gunsXhomX

i0

ˆ

. The MM estimator in this single-regressor model allowing for possible

endogeneity of guns is, in fact, the standard IV estimator.

The above illustrates two features of MM which carry over to GMM and which make the method

attractive to empirical researchers and to users of their work. First, many commonly used estimators are

special cases of GMM estimators, and the apparatus of GMM can be used with these estimators. Both IV

and OLS, for example, are GMM estimators. Second, the MM/GMM approach makes very clear what the

assumptions of the researcher are. Thus estimating model (1) above with two regressors using OLS is

equivalent to using MM with two theoretical orthogonality conditions, namely E(guns

i

u

i

)=0 and

E(control

i

u

i

)=0. These conditions must hold – both guns and the control variable must be exogenous – if

consistent estimates of the parameters β

0

and β

1

are to be obtained in this way. If either theoretical

orthogonality condition does not hold, then the estimates of both parameters will be inconsistent.

A natural question is, what if there are more orthogonality conditions than there are parameters to

be estimated? Say that the researcher wants to estimate (1), believes that control (but not guns) is

exogenous, and has two more variables, X1 and X2, that s/he believes are also exogenous. (In the

terminology of IV estimation, control is an “included instrument” and X1 and X2 are “excluded

instruments”.) This gives three orthogonality conditions, E(control

i

u

i

)=0, E(X1

i

u

i

)=0 and E(X2

i

u

i

)=0,

but still only two parameters to estimate, β

0

and β

1

. Because the model is overidentified – there are three

sample moment equations but only two unknowns – in general it will be impossible to find estimates of β

0

and β

1

that set the sample moment conditions exactly to zero.

In GMM, estimates of β

0

and β

1

are chosen such that the three sample moment conditions are as

“close” to zero as possible. More precisely, GMM proceeds by defining an objective function J(.) that is a

function of the data, the parameters and a set of weights in a weighting matrix W. The GMM objective

function is a quadratic form in the sample moment conditions, i.e., the sample moment conditions are

17

weighted using the weights in W and summed to produce a scalar that is minimized. J(.) can be thought of

as the “GMM distance” – the distance from zero, which is the value the objective function would take if

all the sample moment conditions were satisfied – and the definitions of the GMM estimators of β

0

and β

1

are those values that minimize J given W and the data.

There are as many different GMM estimators as there are different possible Ws to use in J, and in

fact any GMM estimator using a non-trivial W will generate consistent estimates of the parameters.

Where GMM comes into its own is when an optimal weighting matrix is chosen. The optimal GMM

weighting matrix is the inverse of the covariance matrix of moment conditions S; in our example, the

variances and covariances of (control

i

u

i

), (X1

i

u

i

) and (X2

i

u

i

). GMM estimation with an optimal

weighting matrix has the following features. First, when S

-1

is used as the optimal weighting matrix, the

GMM estimator is efficient, i.e., it has the smallest asymptotic variance. Although the true covariance

matrix S is unknown, it can be easily estimated in a prior step,

10

and the two-step GMM estimator that

uses the estimated

as the weighting matrix is also efficient. Second, more orthogonality conditions

mean more efficient estimation, as long as the additional orthogonality conditions are satisfied. Third,

efficient GMM readily accommodates forms of errors that are often encountered by empirical researchers

in social science: heteroskedasticity, autocorrelation, and clustering (spatial correlation). Thus if the

estimated

is robust to arbitrary heteroskedasticity, i.e., heteroskedasticity of unknown form, the

efficient GMM estimator that uses

as the weighting matrix will be efficient in the presence of

arbitrary heteroskedasticity, and the GMM standard errors of the parameters will be robust to arbitrary

heteroskedasticity. The same applies to autocorrelation and clustering. This robustness to arbitrary

violations of homoskedasticity and independence is appealing to empirical researchers, not least because

it means obtaining valid estimation results does not require a researcher to model these violations

explicitly and correctly.

1

S

ˆ

−

S

ˆ

1

S

ˆ

−

10

The reason that it is easy is that an estimate of S can be obtained from any consistent estimator of the equation.

The usual GMM estimator is therefore “two-step feasible efficient GMM”: (1) in the first step, some consistent but

possibly inefficient estimator of the parameters (e.g., IV) is used to obtain

S ; in (2) in the second step, is used

to minimize J and obtain the efficient GMM estimates of the parameters.

ˆ

1

S

ˆ

−

18

Fourth, and most importantly for this paper, efficient GMM provides a straightforward

framework for testing the validity of orthogonality conditions when the equation is overidentified, i.e.,

when there are more orthogonality conditions than there are parameters to be estimated. Under the null

hypothesis that the all the orthogonality conditions are valid – all the variables that were assumed to be

exogenous are indeed exogenous – the minimized value of J is distributed as χ

2

with degrees of freedom

equal to the degree of overidentification.

11

This test of overidentifying restrictions is known in the

literature as the Hansen or Sargan-Hansen J statistic and is, conveniently, an automatic by-product of

GMM estimation.

Note that the orthogonality condition corresponding to an excluded instrument may fail because

the exclusion restriction itself is invalid. Say the model of equation (1) is misspecified in the sense that X1

should actually be an included exogenous regressor because it has a direct impact on homicide rates. If

the researcher estimates equation (1) and uses X1 and X2 as excluded instruments, X1 will be correlated

with the error term of (1), and the J statistic will be tend to be large, indicating a failure of orthogonality

conditions. In other words, the incorrect exclusion of X1 from equation (1) makes X1 endogenous.

The same framework can be used to test the validity of a subset of orthogonality conditions, i.e.,

to test whether or not selected instruments are exogenous. Consider the J statistic resulting from two

different efficient GMM estimations: J1 is the J statistic from an efficient GMM estimation that uses all

the moment conditions, and J2 is the J statistic from an efficient GMM estimation that does not use the

moment conditions corresponding to the suspect instruments (if the suspect variables are included

instruments, i.e., regressors, in the J2 estimation they are treated as endogenous; if they are excluded

instruments, in the J2 estimation they are not used at all). Under the null hypothesis that the suspect

variables are valid instruments, the quantity J1-J2 is distributed as χ

2

with degrees of freedom equal to the

number of instruments being tested. This test is known variously in the literature as a C test, a distance

GMM test, or a difference-in-Sargan test. It is important to note that for the C test to be valid, the

11

In the exactly-identified case, the minimized value of J is zero, and the orthogonality conditions cannot be tested.

19

orthogonality conditions corresponding to the instruments not being tested must also be valid, i.e., the J2

statistic (which is a test of the validity of these orthogonality conditions) should be small.

The GMM framework usefully encompasses and extends a number of older and well-known

procedures. For example, if the error term u is homoskedastic and independent:

12

(a) OLS is the efficient

GMM estimator when regressors are exogenous; (b) IV is the efficient GMM estimator when some

regressors are endogenous; (c) if regressors are exogenous, OLS is more efficient (has a lower variance)

than IV because it makes use of more orthogonality conditions;

13

(d) the J statistic for the IV estimator is

numerically identical to the Sargan’s (1958) NR

2

overidentification statistic;

14

(e) the Hausman-Wu test

for the endogeneity of regressors is numerically identical to the C test that uses the Sargan-Hansen J

statistics from IV and OLS estimations of a model. GMM provides a straightforward method to

generalizing the above estimators and tests to situations where homoskedasticity and/or independence do

not hold.

Useful as it is, GMM is not a panacea, and several caveats to its use should be mentioned. First,

the use of large numbers of orthogonality conditions in the form of excluded instruments can generate

finite sample bias problems, and the general advice here is to be parsimonious with excluded instruments.

Second, there is some evidence that the standard errors for some efficient GMM estimators may be biased

downwards in finite samples, i.e., testing coefficients may be prone to Type I errors. A conservative

estimation strategy adopted by some researchers is consequently to use inefficient GMM estimators (e.g.,

OLS or IV in the presence of heteroskedasticity), rely on robust standard errors for inference, and use the

efficient GMM J statistic for specification testing.

15

Third, Sargan-Hansen tests can have limited power,

i.e., may be prone to Type II errors, and a J or C test that fails to reject the null should be treated with

caution.

12

See Hayashi (2000) for details.

13

More precisely, when errors are homoskedastic and independent, the additional orthogonality conditions

corresponding to the endogenous regressors in IV improve the efficiency of OLS, and the conditions corresponding

to the excluded instruments in IV become redundant for OLS.

14

Basmann’s (1960) overidentification statistic is a close relative of, and asymptotically equivalent to, Sargan’s

statistic, and is also invalid when errors are not homoskedastic.

15

Thus Wooldridge’s (1995) robust overidentification statistic for IV estimation is numerically equal to the J

statistic for two-step efficient GMM (Baum et al. 2003).

20

The GMM procedure that our hypothetical researcher would employ for testing the exogeneity of

variables, and that we use below, is as follows:

1. Estimate equation (1) using the full set of four orthogonality conditions: E(guns

i

u

i

)=0,

E(control

i

u

i

)=0, E(X1

i

u

i

)=0 and E(X2

i

u

i

)=0, and obtain the J statistic for the efficient GMM

estimator. Allow for possible heteroskedasticity or spatial correlation if suggested by priors or

evidence from the data. (The corresponding efficient GMM estimator is due to Cragg (1983) and

is also known as HOLS, “heteroskedastic OLS”.) If the J statistic is large, take this as evidence

that one or more moment conditions is invalid, i.e., one or more of the four variables is

endogenous, and proceed to Step 2. If the J statistic is small, take this as evidence that the

orthgonality conditions are jointly valid – all the variables are exogenous. However, prior

research suggests that the assumption that guns is exogenous is questionable, so proceed to Steps

2 and 3.

2. Estimate equation (1) without the assumption that guns is exogenous, i.e., using the three

orthogonality conditions E(control

i

u

i

)=0, E(X1

i

u

i

)=0 and E(X2

i

u

i

)=0, and obtain the J statistic

for the efficient GMM estimator. Allow for possible heteroskedasticity or spatial correlation. If

the J statistic is small, take this as evidence that the moment conditions are valid – control, X1,

and X2 are all exogenous – and proceed to Step 3. If the J statistic is large, take this as evidence

that one or more of the remaining orthogonality conditions is invalid, i.e., either the control

variable or one of the excluded instruments is invalid, and stop – consistent estimation is not

possible.

3. Test whether guns is endogenous using a C test using J1-J2, where J1 is the J statistic using the

full set of orthogonality conditions (Step 1) and J2 is the J statistic that does not assume

exogeneity for guns. If the C statistic is small, take this as evidence that guns may be exogenous

along with the other variables, and proceed to Step 4. If the C statistic is large, take this as

evidence that guns is endogenous and go to Step 5.

4. (guns is exogenous) Consider estimating the equation by efficient GMM (i.e., HOLS) or OLS.

The former is consistent and efficient; the latter is consistent, but inefficient if errors are not

homoskedastic and independent. Alternatively, because of the danger of a Type II error and

because prior evidence and research suggests that gun levels may be subject to endogeneity bias

from various sources, treat guns as endogenous and go to Step 5.

5. (guns is endogenous) Estimate the equation by efficient GMM or by IV. The former is consistent

and efficient, but may be more prone to Type I errors; the latter is consistent, but inefficient if

errors are not homoskedastic and independent.

Instrument relevance and the weak instruments problem

As the simple IV example above showed, instrument relevance is another type of moment

condition required for an IV-type estimator to be consistent. The minimal relevance requirement is just

the rank condition for identification of the model; if the rank condition is not satisfied, the model is

unidentified. The intuition behind instrument relevance is simply that the excluded instruments must be

21

correlated with any endogenous regressors. If there are multiple endogenous regressors, measuring

instrument relevance is not straightforward because it requires estimation of the rank of the covariance

matrix of regressors and instruments.

16

In the case of a single endogenous regressor, however, the

recommendation in the literature is straightforward and easy to implement: the statistic for instrument

relevance is the F test of the excluded instruments in the “first-stage” regression, i.e.,

guns

i

=

θ

1

control

i

+ θ

3

X1

i

+ θ

4

X2

i

+ v

i

(4)

and the researcher examines θ

3

and θ

4

for their consistency with theory and prior evidence, and for their

joint significance using a standard F statistic; a large value indicates that the model is identified. If

heteroskedasticity or clustering is suspected, a heteroskedastic- or cluster-robust test statistic can be used.

The researcher should also examine the individual significance of both θ

3

and θ

4

in the estimation of the

first-stage equation (4). Adding instruments does not come without a cost; in particular, the finite sample

bias of the IV/GMM estimator is increasing in the number of instruments. If the first-stage regression

suggests that one instrument is strongly relevant and the other is irrelevant, the researcher should consider

dropping the irrelevant instrument.

17

An important practical problem arises if the excluded instrumental variables are correlated with

the endogenous regressor but only weakly. Recent research (e.g., Bound et al., 1995; Staiger and Stock,

1997; Stock et al. 2002 provide a survey) has shown that when instruments are weak, IV/GMM estimates

of parameters will be badly biased (in the same direction as OLS), estimated standard errors will be

unreliable, and therefore so will the Wald-based hypothesis tests and confidence intervals that use these

estimates; in particular, standard errors will be too small and the null will be rejected too often. This is an

area of ongoing research in econometrics, and guidelines for how to detect a “weak instrument” problem

are not yet well established. There is, however, a consensus that it is not enough for the first-stage F

16

Two statistics that have been suggested for this purpose are the Cragg-Donald statistic and Anderson’s canonical

correlations statistic. Both statistics require homoskedastic and independent errors to be valid. See Hall et al. (1996)

and Stock and Yogo (2002) for further discussion.

17

The formal approach would be to test the instrument for “redundancy”. An excluded instrument is redundant if the

efficiency of the estimation is not improved by its use. See Hall and Peixe (2000) for further discussion.

22

statistic to be significant at conventional levels of 5% or 1%; higher values are required. Staiger and Stock

(1997) recommend an F statistic of at least 10 as a rule of thumb for the standard IV estimator

.

18

If weak instruments are detected or suspected, recent research suggests two possible approaches.

The investigator may use an estimator that is relatively robust to weak instruments. These estimators are

often non-standard and not widely available in commercial software packages, they may not be robust to

heteroskedasticity or clustering, their performance in the presence of weak instruments is still being

explored, and so we do not go down this route here. The second approach is to employ a procedure for

construction of a confidence interval that is robust to weak instruments. Various such methods have been

proposed; see, e.g., Dufour (2003) and Andrews and Stock (2005). One method that is easy to implement

for the single-endogenous-regressor case using standard regression software packages is based on the

Anderson-Rubin (1949) test.

19

The method is as follows. Consider a specific hypothesized value

0

~

β

for the coefficient on guns

in model (1). Subtract the quantity

i0

~

gunsβ

from both sides of (1) and then substitute using (4) to obtain

])

~

[(

)

~

()

~

(])

~

[()

~

(

00

40030011000

ii

iiiii

uv

X2X1controlgunshom

+β−β+

θβ−β+θβ−β+β+θβ−β=β−

(5)

If the null hypothesis H

0

:

00

~

β=β

is correct, both X1 and X2 will drop out, and equation (5) simplifies to

iiii

ucontrolgunshom +β=β−

10

)( (6)

The AR test of the null hypothesis is therefore to estimate

iiiiii

ηX2X1controlgunshom +π+π+π=β−

4310

)

~

(

(7)

and employ a standard F test of whether the coefficients on the excluded instruments π

3

and π

4

are indeed

both equal to zero. The AR test is suitable when a model is just-identified or the degree of

18

Stock and Yogo (2002) provide more detailed advice based on Monte Carlo studies, and show inter alia that the

Staiger-Stock rule of thumb is a reasonable guideline to follow for the single-endogenous-regressor case when the

number of excluded instruments is small.

19

Not to be confused with the Anderson-Rubin overidentification test, which is the test analogous to the Sargan-

Hansen J test for maximum-likelihood estimation.

23

overidentification is low, and it can easily be made heteroskedastic- or cluster-robust by using the

appropriate covariance estimator for (7).

20

Note that the AR test assumes that the orthogonality conditions

E(X1

i

u

i

)=0 and E(X2

i

u

i

)=0 are satisfied; if these did not hold, the AR test statistic would be significant,

not because

00

~

β≠β

, but because the excluded instruments were correlated with the composite error term

η

i

. If either X1 or X2 is endogenous, the AR test is invalid.

An AR confidence interval is simply the region where the AR test fails to reject the null. For

example, an AR 95% confidence interval for the coefficient on guns β

0

would be the range of specific

values for

0

~

β

such that the AR test statistic is below the 5% critical value for the F distribution. An AR

confidence interval is fully robust to weak instruments, and the intuition for this is clear from inspection

of equations (5) and (7). Note that π

3

=

300

)

~

( θβ−β

and similarly for π

4

. If X1 and X2 are only weakly

correlated with guns, then θ

3

and θ

4

will be small, so will π

3

and π

4

, and a test of H

0

: π

3

=π

4

=0 can fail to

reject the null even if the hypothesized

0

~

β

is far from the true β

0

. Weak instruments will therefore tend to

widen the AR confidence interval for the impact of guns β

0

, i.e., low instrument relevance will be

properly reflected in the imprecision of the estimated impact of gun levels on homicide.

21

5. DATA, MODEL AND ESTIMATION STRATEGY

Data

To estimate the impact of gun availability on homicide rates, we use cross-sectional data for all

U.S. counties which had a population of 25,000 or greater in 1990, and for which relevant data were

available (N=1,462). These counties account for about half of all U.S. counties but over 90% of the U.S.

population in that year. The use of 1990 data is dictated by two factors. First, most of the control variables

included in the homicide equations to mitigate omitted variable bias are available at the county-level only

during census years. A second reason for choosing 1990 is the fact that the firearm crime rate (homicide,

robbery, and assault) had reached its highest level in nearly 30 years by 1990. It is reasonably argued that

20

The AR test loses power as the number of excluded instruments goes up. More powerful tests have recently been

proposed; see Andrews and Stock (2005) for a survey.

21

We calculate our AR confidence intervals with a simple grid search over possible values β

0

. Note that some care

in calculating AR confidence intervals may be needed in practice because the AR test can sometimes generate

confidence intervals that are either empty or disjoint. See Zivot et al. (1998).

24

if gun availability is responsible for higher homicide rates, the high levels of firearm crime in 1990 should

provide one of the best opportunities to date for testing the gun availability-homicide relationship.

County-level data were chosen for several reasons. First, the use of counties provides for a

diverse sample of ecological units, including urban, suburban, and rural areas. Second, counties are more

internally homogenous than nations, states, or metropolitan areas, thereby reducing potential aggregation

bias. Third, counties exhibit great between-unit variability in both gun availability and homicide rates,

which is precisely what gun availability and homicide research is trying to explain. Fourth, county data

provide a much larger sample than previous gun level studies, which have focused mainly on nations,

states, or large urban cities (almost all with 50 or fewer cases; see Table 1). The large sample size

provides us with greater statistical power to detect more modest effects of gun availability on homicide

rates, while still permitting us to enter numerous control variables in the homicide equations to minimize

omitted variable bias. A large sample size is particularly important when IV/GMM methods are used,

because these estimators are only consistent, i.e., bias approaches zero only as the sample size increases.

Model

We follow the convention for crime policy studies and use a linear model in which most variables

are specified in logs. The dependent variables in our model are the gun, nongun, and total homicide rates

per 100,000 county population. Homicide data for each county were obtained using special Mortality

Detail File computer tapes (not the public use tapes) made available by the National Center for Health

Statistics (U.S. NCHS 1997). The data include all intentional homicides in the county with the exception

of those due to legal intervention (e.g., police shootings and executions). Homicide rates are averages for

the seven years 1987 to 1993, thus bracketing the census year of 1990 for which data on many of the

control variables were available. Seven years were covered to reduce the influence of random year-to-

year aberrations, e.g., misclassification of homicides as other kinds of deaths such as suicides or

unintentional deaths, and to allow the use of rates in a linear model as an approximation for count data

(see below).

25

It is desirable to separately assess rates of homicide with and without guns, to provide sharper

tests of the hypothesis that gun levels affect homicide rates. The decision criteria upon which we rely for

determining whether gun levels causes a net increase in homicide rates are as follows. If the gun level has

a net positive impact on homicide rates, it should have (1) a significant positive association with the gun

homicide rate, (2) a significant positive association with the total homicide rate, and (3) be less strongly

positively correlated with the nongun homicide rate than with the gun homicide rate. On the other hand, if

gun levels are as strongly positively associated with nongun homicides rates as with gun homicide rates

(or more so), this suggests that the gun level is merely a correlate of some omitted variable that affects

homicide in general, but the gun level has no effect of its own, since there is no strong reason why gun

levels should increase the rate of homicides committed without guns. If (1) is true, but (2) is not, it would

generally indicate that gun availability merely shifts criminals from nongun weapons to guns, but has no

net effect on the number of people murdered. If (2) is true, but (1) is not, it suggests that gun levels are

merely associated with some omitted variables that have an effect on total homicide rates but that gun

levels themselves have no effect, since they should have their effects by, at minimum, increasing

homicides committed with guns.

In our main regressions, the crime rate variables are specified in logs. This poses some minor

problems, because even though we are using 7-year averages and excluding the smallest counties, a small

number of counties have zero murders: of the 1,462 counties in the sample, 20 had no gun murders (about

1% of the sample), 52 had no nongun murders (about 4%), and 3 had no murders at all. Our approach is to

report in detail the results using the logged crime rates and dropping the observations for which the

dependent variable is undefined. Histograms of the logged crime rates after dropping zero-crime-rate

counties are shown in Figure 2; simple visual inspection suggests no particular skewness or outlier

problems after log transformation.

26

Figure 2: Distributions of Log Crime Rates

0

1

2

3

4

5

6

Percent

-3

-2

-1

0

1

2

3

4

5

Log gun homicide rate

0

1

2

3

4

5

6

Percent

-3

-2

-1

0

1

2

3

4

5

Log non-gun homicide rate

0

1

2

3

4

5

6

Percent

-3

-2

-1

0

1

2

3

4

5

Log total homicide rate

To check the robustness of the results, we also estimated using a variety of transformations of the

dependent variable that do not require us to drop observations: (1) logged “add 1” crime rates, where the

number of murders for the 7-year period has 1 added to it; (2) “one-sided winsorized”

22

log crime rates,

where prior to taking logs the zero crime rate counties are assigned a crime rate equal to the lowest non-

zero-crime-rate county; (3) “two-sided winsorized” log crime rates, where the treatment of the zero crime

rate counties is as in (2), and the same number of the highest crime counties are, prior to logging, assigned

a crime rate equal to that of the next highest (i.e., the highest non-transformed) county; (4) raw crime

rates (murders per 1,000 population).

While it might be desirable to measure gun ownership levels directly using survey-based

estimates, surveys asking people directly whether they own guns are usually limited to a single large area,

such as a nation or state. Instead, we use the best indirect measure of gun availability for cross-sectional

research, the percent of suicides committed with guns (PSG). Some of the smaller counties typically have

few or no suicides in a given year, and misclassification of a few suicides as homicides or accidents in

small counties could produce substantial measurement error in a single year’s PSG. Therefore, as was

22

See the help file for the Stata command “winsor” by Nick Cox (2003) for a discussion of “winsorizing” and for

references to the relevant statistical literature.

27

done with homicide rates, PSG was computed for the 7-year period 1987 to 1993, bracketing the

decennial census year of 1990. Similar to homicide, data for the percent of suicides committed with guns

were obtained using special Part III Mortality Detail File computer tapes made available by the National

Center for Health Statistics. Unlike widely available public use versions, the tapes permit the aggregation

of death counts for even the smallest counties (U.S. NCHS 1997).

Figure 3 shows histograms of our gun proxy PSG in raw percentage form and after logging.

There are no problems of zeros here; rather, the issue is whether the skewness to the left seen in the

distribution of log PSG is of any concern. Our approach is to report detailed results using log PSG, and to

confirm robustness of the findings with regressions using PSG in raw percentage form.

Figure 3: Distributions of PSG and Log PSG

0

1

2

3

4

5

6

7

Percent

0

25

50

75

100

PSG

0

1

2

3

4

5

6

7

Percent

3

4

5

Log PSG

We also need to calibrate our proxy to available survey-based measures of gun levels. The most

convenient calibration is to the mean percentage of households with guns (HHG) according to the General

Social Surveys (GSS). National gun survey prevalence figures have been available since 1959, though not

for every year. The mean HHG for 1959-2003 is 44.2% while the mean PSG for 1959-2002 (the latest

year available) is 54.9%. These figures imply a value of 54.9/44.2=1.24 for the calibration factor δ

0

by

which we should inflate or deflate the estimated coefficient on PSG b

0

so as to obtain an estimate of β

0

,

the impact of gun levels on homicide rates (see equation 3). Neither PSG nor HHG varied greatly during

this period, and the use of a different reference period would matter little.

Using state-level measures from surveys conducted by the Centers for Disease Control (CDC) in

2002 (Okoro et al., 2005) and PSG data for 1995-2002 taken from CDC’s WONDER service, simple OLS

28

regressions (based on 50 states) of PSG and log PSG and the corresponding HHG and log HHG measures

are as follows (standard errors are in parentheses):

Log PSG = 2.31 + 0.481 Log HHG + e

(0.12) (0.035)

PSG = 30.11 + 0.706 HHG + e

(2.85) (0.069)

The coefficients on log HHG and HHG can be interpreted as estimates of the log-log and level-level

calibration δ

0

, respectively. The figures suggest that in our main log-log estimations, the coefficient on

log PSG should be approximately halved (δ

0

=0.481) in order to be interpreted as the HHG-homicide

elasticity. In a levels-levels estimation, the calibration should be in the neighborhood of 0.706 (based on

the 50 states regression) to 1.24 (based on national means). These are, however, only approximations

based on limited data and simple linear calibrations. A more cautious conclusion would be that PSG is,

within an order of magnitude, already calibrated to HHG, and that interpretations of the estimated

coefficient on gun levels will be most robust in the neighborhood of mean PSG.

In addition to the gun prevalence measure, we included numerous county-level control variables,

paying particular attention to those that prior theory and research suggest are important determinants of

both gun ownership levels and homicide rates. Failing to control for confounders that affect both gun

availability and homicide rates would generate an endogeneity bias in the coefficient on PSG, as

discussed earlier. Decisions as to which control variables to include in the homicide equations were based

on a review of previous macro-level studies linking homicide rates to structural characteristics of

ecological units (see Kleck, 1997, Chapter 3; Kovandzic et al., 1998; Land et al., 1990; Sampson, 1986;

Vieraitis, 2000 and the studies reviewed therein).

We were particularly concerned to control for variables that had opposite-sign associations with

gun levels and homicide rates because such variables could suppress evidence of any positive effect of

gun levels on homicide rates. Thus, we controlled for the percent of the population that is rural because

rural people are more likely to own guns, but less likely to commit homicide. Likewise, we controlled for

the poverty rate, the share of the population in the high-homicide ages of 18 to 24 and 25 to 34, and the

29

African-American share of the population because people in these groups are less likely to own guns, but

more likely to commit homicide, than other people (Kleck, 1997; Cook and Ludwig, 1997; U.S. FBI,

2000). The other controls used were percent Hispanic, population density, average education level,

unemployment rate, transient population (born out-of-state), vacant housing units, female-headed

households with children, median household income, households earning less than $15,000, and

inequality (ratio of households earning more than $75,000 to households earning less than $15,000).

The sets of controls for rurality and age structure are, exceptionally, used in percentage rather

than log form. Because the raw percentages sum to 100, including all categories would generate a perfect

collinearity problem, and so one category must be omitted. Using raw percentages therefore has the

appealing feature that the results are invariant to whichever percentage is the omitted category. We omit

the percentage rural and the percentage aged 65+.

The use of county-level data has an additional advantage: by including state fixed effects, i.e.,

state dummy variables, we are able to control for any unobserved or unmeasured county characteristics

that vary at the state level and that could be expected to influence both gun levels and homicide rates.

Examples of such confounders would be state laws and judicial practice relating directly or indirectly to

homicide and gun ownership, state-level resources devoted to law enforcement, and incarceration rates in

state prisons. Testing of the fixed effects in the estimations below strongly supported their inclusion as

controls.

23

The disadvantage of this approach is that only variables available at the county level can be

used in the estimations, because state-level measures would be perfectly collinear with the fixed effects.

24

We do not include explicit state dummies and instead use an estimation routine that obtains numerically

equivalent results by applying the “within” transformation to all variables, i.e., expressing them as

deviations from state means. The R

2

reported with the regressions is the “within R

2

”, and measures how

much of the within-state variation in homicide rates our models explain.

The key challenge in using IV methods is finding a source of identifying variation: here, variables

that are correlated with gun levels, but that are exogenous with respect to homicide and that a priori

23

F tests of the significance of the state fixed effects vs. a specification using a set of 9 regional dummies.

24

It also means that the single observation for Washington, D.C. drops out of the fixed-effects regressions.

30

reasoning and evidence suggest should be excluded from the homicide equation. The excluded

instrumental variables used in this paper are (1) an index (RGUNMAG) comprised of subscriptions to

each of the three most popular outdoor/sport magazines (Field and Stream, Outdoor Life, and Sports

Afield) in 1993, per 100,000 county population (Audit Bureau of Circulations, 1993), and (2) the percent

of the county population voting for the Democratic candidate in the 1988 Presidential election

(PCTDEM88). The gun magazine index was created using principal components analysis; the analysis

yielded a factor that explained 84 percent of the cumulative variance in this latent construct.

25

Both excluded instruments are theoretically important correlates of gun ownership that are

plausibly otherwise unrelated to homicide. RGUNMAG serves as a measure of interest in outdoor sports

such as hunting and fishing, or perhaps as a measure of a firearms-related “sporting/outdoor culture”

(Bordua and Lizotte, 1979). PCTDEM88 serves as a measure of political liberalism and hence should be

negatively correlated with gun ownership. The 1988 election results were chosen in preference to the

1992 results because the date precedes the census year from which most our data are taken (and hence is

more plausibly exogenous), and because the choice between the two main candidates in 1988 maps more

closely to attitudes towards gun ownership.

26

Prior research suggests that both variables are important

predictors of gun ownership (Kleck, 1997, pp. 70-72; Cook and Ludwig, 1997, p. 35).

Table 2 lists and provides a brief description of each variable used along with their means and

standard deviations. Data for the control variables were obtained from the U.S. Bureau of the Census,

County and City Data Book, 1994, except for PCTDEM88, which is from ICPSR (1995), and rurality,

which is from U.S. Census Bureau (2000).

Estimation strategy

The estimation and testing procedure we follow is the GMM approach outlined above in

Section 4. It is natural to expect the presence of heteroskedasticity in a cross-sectional dataset, and indeed

25

The factor had an eigenvalue of 2.52, well above the conventional threshold of 1.00. The loading scores for each

magazine were as follows: SPAFIELD (0.824), LIFE (0.968), and STREAM (0.948).

26

In the 1992 election, unlike the 1988 election, the politically less conservative candidate (negatively correlated

with gun ownership) was also a southerner (positively correlated with gun ownership). The 1992 results are also less

easily interpreted because of the significant share of the vote that went to the third-party candidate, Ross Perot.

31

application of the Pagan-Hall (1983) test for heteroskedasticity in IV estimation to our data suggests it is

present here. We therefore use GMM estimation that is efficient in the presence of arbitrary

heteroskedasticity.

An oft-neglected issue with cross-sectional data on locations is the potential problem of spatial

correlation. It is reasonable to suspect that observations in two physically adjacent counties are more

likely to have correlated disturbance terms than two counties at opposite ends of the state or country. If

spatial correlation is present, standard errors that assume independence are likely to be underestimated

and test statistics will be invalid. A straightforward method of addressing this in the GMM context is the

“cluster-robust” approach, where clusters are defined as groups of counties – here, states.

27

The

corresponding GMM estimator is efficient in the presence of both arbitrary heteroskedasticity and

arbitrary within-state correlation of the disturbance term, and requires only the assumption of

independence across states, which is a reasonable assumption given our fixed-effects specification. The

main drawback to this approach is that we have only 50 clusters (states), and hence relatively few degrees

of freedom. Most of our regressions have 18 exogenous regressors and 2 excluded instruments, leaving us

with effectively only 50-18-2=30 observations on clusters for calculating cluster-robust standard errors.

We therefore use the fixed-effects GMM estimator that is efficient in the presence heteroskedasticity for

our benchmark results, and report the results using the cluster-robust fixed-effects GMM estimator as a

check on the sensitivity of the findings to the presence of spatial correlation.

Our initial specification treats PSG as exogenous; it is the two-step efficient GMM estimator with

exogenous regressors that allows for arbitrary heteroskedasticity (Cragg’s HOLS estimator). The

corresponding J statistic is a test of the full set of orthogonality conditions, i.e., the exogeneity of PSG,

RGUNMAG and PCTDEM88 (plus the other covariates). Our second specification is the two-step

efficient GMM estimator that treats PSG as endogenous, and the J statistic is a test of the reduced set of

orthogonality conditions, i.e., the exogeneity of RGUNMAG and PCTDEM88 (again plus the other

covariates). The C statistic reported with the initial specification is a test of the endogeneity of PSG and is

27

See Wooldridge (2003) for an overview and discussion of the case where the number of groups is small.

32

based on the difference of the two J statistics.

28

We test for instrument relevance using a heteroskedastic

or cluster-robust F test of the joint significance of the excluded instruments RGUNMAG and PCTDEM88

in an OLS estimation of the first-stage equation of the gun proxy (PSG or log PSG), and we also examine

the significance of RGUNMAG and PCTDEM separately using conventional t-statistics. Because these

instruments are, in some specifications, bordering on weak, we also report an Anderson-Rubin confidence

interval for the coefficient on PSG, and, for comparison, the usual Wald-based confidence interval

employing the GMM-estimated standard error. The versions of these tests that we use are all

heteroskedastic- or cluster-robust; non-robust test statistics would tend to be biased upwards and lead to

Type I errors. Lastly, as a check on the sensitivity of the results, we also estimate the main equations

using the corresponding inefficient GMM estimators, OLS and IV. With the exception of the F test-based

AR confidence intervals, we report large-sample standard errors and significance tests; i.e., we do not

make a finite-sample adjustment for the number of explicit regressors, and we use the normal or χ

2

distributions for significance tests, p-values and confidence intervals.

29

The statistical package Stata was used for all estimations. The main IV/GMM estimation

programs, ivreg2 and xtivreg2, were co-authored by one of us (Schaffer), and can be freely downloaded

via the software database of RePEc.

30

For further discussion of how the estimators and tests are

implemented, see Baum, Schaffer and Stillman (2003) and (2005), and the references therein.

6. RESULTS

Estimation results for the benchmark regressions using the logged gun homicide rate as the

dependent variable are in Table 3. Columns 1 and 2 report the results of 2-step GMM estimations that are

efficient in the presence of arbitrary heteroskedasticity and that treat PSG as exogenous and endogenous,

respectively, along with heteroskedastic-robust standard errors and test statistics. Columns 3 and 4 report

28

The difference in reported J statistics is not exactly equal to the reported C statistic, because we use a version of

the latter that is guaranteed to be non-negative in finite samples. See Hayashi (2000) and Baum et al. (2003).

29

The heteroskedastic-robust variances and statistics have a degrees-of-freedom adjustment for the number of fixed

effects (50); the adjustment is not required for the cluster-robust variances. See Wooldridge (2002), pp. 271-75.

30

http://ideas.repec.org/SoftwareSeries.html. ivreg2 is a general-purpose IV/GMM estimation routine for linear

models; xtivreg2 supports fixed-effects panel data models.

33

the corresponding estimates that are efficient in the presence of heteroskedasticity and within-state

clustering, plus heteroskedastic- and cluster-robust standard errors and statistics.

We consider first the heteroskedatic-robust results in columns 1 and 2. Most of the parameter

estimates for the 18 control variables are significant, and the significant coefficients have the expected

sign in both specifications. High gun murder rates are associated with high population density, lower

education levels, and the various poverty, low-income and inequality measures. The percentage of the

population that is black is associated with high gun crime rates, as is the percentage Hispanic, but in the

latter case only in the first specification, when PSG is treated as exogenous. The only modest surprise is

provided by the age structure controls; contrary to expectations, the 18-24 age group was associated with

relatively low gun homicide rates,

31

though this is a common finding (Marvell and Moody 1991). The

overall fit of the regressions is quite good, with the PSG-exogenous specification explaining 44% of the

within-state variation in county-level log gun homicide rates, and the PSG-endogenous specification

explaining 33%.

The key results concern the coefficient on and the exogeneity of PSG. Column 1 shows that in the

efficient GMM estimation where log PSG is treated as exogenous, the variable has a coefficient of 0.290

and is statistically significant at the 5% level. This confirms the oft-reported result that, when endogeneity

issues are ignored, gun levels are associated with higher gun crime rates. When log PSG is treated as

endogenous and instrumented with RGUNMAG and PCTDEM88, the picture changes dramatically.

Column 2 shows that log PSG has a negative coefficient of –1.50 that is statistically significant at the 1%

level.

We now apply the GMM-based procedure outlined in Section 4 for testing the exogeneity of

PSG. [Step 1] The J statistic in column 1 is 13.3. This is very large for a χ

2

(2) statistic; the p-value is only

0.001. We therefore reject the null hypothesis that the orthogonality conditions in the PSG-exogenous

estimation are satisfied, and take this as strong evidence that one or more variables – log PSG,

RGUNMAG, PCTDEM88, and/or the control variables – are endogenous. [Step 2] The J statistic for the

31

The effect is comparable to that of the omitted category of 65+ and lower than those for the remaining categories.

34

PSG-endogenous estimation in column 2 is 0.25, which small for a χ

2

(1) statistic – the corresponding p-

value is 0.61. We therefore cannot reject the null that RGUNMAG, PCTDEM88, and the control

variables are exogenous; in other words, the evidence suggests that our instruments are valid. [Step 3] We

now test explicitly whether log PSG is endogenous using a C test based on the J statistics for the PSG-

exogenous and PSG-endogenous estimations. The C statistic reported in column 1 is 13.0.

32

This is very

large for a χ

2

(1) statistic; the p-value is only 0.0003. We therefore have strong evidence that log PSG is

endogenous, and that some form of IV/GMM estimation is required. Before we turn to this last estimation

[Step 5], however, we must also consider whether our excluded instruments RGUNMAG and

PCTDEM88 are relevant as well as valid.

The first-stage heteroskedastic-robust F statistic reported in column 2 is 26.2, well above the

Staiger-Stock rule-of-thumb level of 10. Both the excluded instruments are correlated with log PSG in the

expected directions and at the 0.1% significance level: counties voting Democrat in 1988 tend to have

fewer guns as proxied by log PSG, and subscriptions to outdoor magazines are associated with higher gun

levels. We conclude that our instruments in this estimation are relevant and not weak. Nevertheless, as a

check we also report a heteroskedastic-robust AR 95% confidence interval for log PSG. The AR

confidence interval is [–3.05 , –0.34] only slightly wider than the conventional Wald-based confidence

interval of [–2.52 , –0.48].

Having shown that the GMM estimation in column 2 satisfies the requirements of both validity

and relevance, we turn to the issue of calibration. The calibration exercise above suggests we should

roughly halve coefficient estimate of –1.50 to obtain the elasticity of gun homicide with respect to HHG,

i.e., about -0.75. The estimate is therefore not only statistically significant, it is potentially also practically

significant. We hasten to add, however, that this is a very approximate estimate: after calibration to HHG

and allowing for uncertainty in the calibration itself, both the conventional Wald and weak-instrument-

robust AR confidence intervals would include low (practically insignificant albeit statistically significant)

HHG-gun homicide elasticities. In sum, we have strong evidence that gun levels are endogenous, and

32

The C statistic differs slightly from the difference between the relevant J statistics because we use a version of the

C test that guarantees a positive test statistic. See Hayashi (2000) or Baum et al. (2003) for details.

35

when this is accounted for in the estimation, the positive association of gun levels with gun homicide

completely disappears; if anything, gun levels are associated with lower, not higher, gun homicide rates.

These findings are essentially unchanged if we allow for intra-state spatial correlation as well as

heteroskedasticity. When we treat log PSG as endogenous (column 4), the two-step efficient GMM

estimate of the coefficient on log PSG is virtually the same as before: –1.53, again significant at the 1%

level. The large J statistic in the PSG-exogenous GMM estimation in column 3 again strongly suggests

that we should reject the hypothesis that PSG, RGUNMAG, PCTDEM88 and the controls are all

exogenous; the small J statistic in the PSG-endogenous GMM estimation in column 4 means we fail to

the reject the hypothesis that RGUNMAG, PCTDEM88 and the controls are all exogenous; and the large

C statistic in column 3 strongly implies we should reject the hypothesis that PSG in particular is

exogenous. Again, some form of IV/GMM estimation that treats PSG as endogenous is called for. With

respect to the requirement of instrument relevance, column 4 shows that both of the excluded instruments

are still significant in the first-stage regression. A potential cause for concern, however, is the first-stage F

statistic, which at 12.2 is somewhat low and indicative of a possible weak instruments problem. We

therefore refer to the Anderson-Rubin 95% confidence interval for log PSG. This is [–3.35 , –0.06], vs.

the Wald-based interval of [–2.52 , –0.48]. The reduced relevance of the excluded instruments when

allowing for spatial correlation noticeably widens the AR confidence interval, but we can still reject the

hypothesis that our estimated coefficient on log PSG is zero, albeit at a reduced level of statistical

significance. Our key finding is therefore clearly robust to within-state spatial correlation. We again

conclude that we must treat gun levels as endogenous, and when this is done we find that they are, if

anything, associated with lower, not higher, rates of gun homicide across counties.

Table 4 presents the corresponding results for the log nongun homicide rate. The patterns in the

coefficients on the covariates are similar to those for the gun homicide equations; the main differences are

that rurality is now significantly associated with lower nongun homicide rates, and the economic controls

are less significant. The main difference vis-à-vis the gun homicide coefficients is in the estimated impact

of gun levels: across all estimations for the nongun homicide rate, the coefficient on log PSG is

insignificantly different from zero. The tests of orthogonality conditions suggest, moreover, that guns

36

may be treated as exogenous in the nongun homicide equation: the J statistics are low for both the

specifications that treat guns as exogenous (columns 1 and 3) and the specifications that treat guns as

endogenous (columns 2 and 4), and the C statistics for the endogeneity of log PSG are also low. The

implication is that the estimates for nongun homicide that treat guns as exogenous are to be preferred on

efficiency grounds, and the estimates that treat guns as endogenous are preferable if we wish to avoid a

Type II error in concluding they are exogenous; but the choice matters little because both sets of

estimations suggest no impact of guns on nongun homicide rates. The tests of instrument relevance are

(naturally) very similar to those for the gun homicide equations. Again the low first-stage F statistic when

allowing for intra-state spatial correlation is a concern, and again the AR confidence intervals for log PSG

support the Wald test results: gun levels have no impact on nongun homicide rates.

If we take these estimates of gun effects seriously, they suggest that gun levels in the general

public may, on net, have a deterrent effect on gun homicide rates, but no such effect on nongun

homicides. Deterrent effects would be stronger for gun homicides if their perpetrators were more likely to

plan the killings (or crimes leading up to the attacks, such as robbery or a drug deal) than those who use

less lethal weapons. The fact that an aggressor chose a lethal weapon, better suited to lethal purposes,

rather than merely making use of whatever weapons happened to be available at the scene, may itself be

an indication of premeditation. Thus, people who kill with guns may be more easily deterred by the

prospect of confronting a gun-armed victim than those who kill with other weapons, because the former

are more likely to think about the potential costs of their actions.

The results in Table 5 are for the estimations using the log total homicide rate (gun+nongun) as

the dependent variable. The results are similar to, but slightly weaker than, those for the gun homicide

rate. In the specifications that treat log PSG as exogenous, its coefficient is positive and either

insignificant (column 1) or significant at the 5% level (column 3). When gun levels are treated as

endogenous, the coefficient on log PSG becomes negative (about –1.0) and significant at the 5% level,

and the AR and Wald confidence intervals are again similar. The J statistics for the PSG-exogenous

estimations are large and suggest that one or more orthogonality conditions is violated. When the

orthogonality condition for PSG is dropped the J statistics become very small, suggesting that the

37

remaining orthogonality conditions are valid, and the C test strongly rejects the hypothesis that gun levels

are exogenous. The first-stage statistics again indicate a possible weak instrument problem for the cluster-

robust specification, and in this case the weak-instrument-robust AR 95% confidence interval in

column 4, though mostly negative, includes zero, i.e., an AR test of the coefficient on log PSG fails to

reject a zero impact of gun levels on homicide at the 5% significance level.

33

Our key finding remains:

gun levels have at least no impact, and possibly a negative impact, on total homicide rates.

We tested the robustness of the coefficient estimates by using the corresponding inefficient but

consistent GMM estimators, OLS and IV, and heteroskedastic- and cluster-robust standard errors; the

results were essentially identical.

34

Finally, we checked robustness to the functional form by comparing

the results using the dependent variables in the four possible log measures (drop zero-crime-rate counties

as used in the results above, “add 1”, “one-sided winsorized”, and “two-sided winsorized”) plus a fifth,

unlogged crime rates, and by using both log PSG and unlogged PSG as our gun proxies. This gave us

5*2=10 possible combinations of specifications, and estimating with and without cluster-efficient GMM

gave us 10*2=20 estimations in total for each of the three categories of homicide rates. The qualitative

results were almost entirely unaffected by these variations in functional form with respect both to the

estimate of impact of gun levels and to the findings on endogeneity, instrument validity and instrument

relevance. When both crime rates and PSG were logged, the quantitative results were also essentially

unchanged, with estimated coefficients very similar to those reported above. The specifications that used

the unlogged nongun homicide rate (CRNGMUR) provided the sole exception to this pattern, and, if

anything, strengthen our general conclusions: both PSG and log PSG were found to be endogenous, and

when instrumented, the estimated coefficients were negative and statistically significant, suggesting that

higher gun levels are associated with lower nongun homicide rates.

With respect to calibration, in the level-level specifications using the heteroskedastic-robust

efficient GMM estimator, the coefficients on PSG in the gun, nongun, and total homicide equations were

33

The AR confidence interval is negative and excludes zero if the level of confidence is reduced from 95% to 89%.

34

The first-stage and J and C statistics were identical by construction, the latter two because efficient GMM was

needed to produce heteroskedastic- and cluster-robust J and C statistics.

38

–0.227, –0.066, and –0.294, all statistically significant with p-levels about 0.01 (the cluster-robust

estimates were very similar). Using the calibration to HHG suggested above (0.706<δ

0

<1.24) implies that

an increase in 10 percentage points in the number of households with guns in a county would reduce gun

homicides by about 2 persons per 100,000 population, nongun homicides by about 0.7 persons, and total

homicides by about 3 persons. The implied elasticities of homicide to gun levels are somewhat higher

those estimated directly with the log-log specifications in Tables 3-5.

35

Again, however, we hasten to add

that these are very approximate estimates, and confidence intervals would include low, i.e., practically

unimportant, impacts. Our main conclusions, after from all these robustness checks, remain that the

positive correlation between gun levels and homicide rates is driven by endogeneity bias, and when the

endogeneity of gun levels is properly addressed in the estimation, any positive correlation vanishes.

7. CONCLUSIONS

Most studies of the gun-crime relationship have ignored the endogeneity problem, and the few

that have tried to address it with IV methods have failed to perform the tests needed to tell whether their

estimation procedures were adequate. We have presented a formal analysis of the three main problems

facing researchers – reverse causality, mismeasurement of gun levels, and omitted/confounding variables

– and discussed how to address these problems using estimation and specification testing procedures in a

GMM framework. We applied these procedures to U.S. county level data, and found strong evidence of

the existence of endogeneity problems. When the problem is ignored, gun levels are associated with

higher rates of gun homicide; when the problem is addressed, this association disappears or reverses

(though the reversal, suggesting that higher gun levels lead to lower gun crime rates, should be treated

with caution).

Our findings provide no support for the more guns, more homicide thesis. The appearance of such

an effect in past research appears to be the product of methodological flaws, especially the failure to

35

E.g., mean county CRGMUR in the sample is 4.1 homicides per 100,000 (Table 2); average HHG in the U.S. is

44% (Section 5). With a 1:1 calibration of PSG to HHG, the implied elasticity is therefore (–0.227)*44/4.1 ≈ –2.4,

vs. about –1.5 in Table 3.

39

properly account for the possible effect of crime rates on gun ownership levels. Higher gun prevalence

may have a net violence-elevating effect, but one that is confined to criminals or perhaps other high-risk

subsets of the population. To be consistent with our generally null findings regarding the effects of gun

levels, however, if there were such a violence-increasing effect of guns among criminals, it would have to

be counterbalanced by violence-reducing effects among noncriminals. Guns among criminals may

increase homicide while guns among noncriminals decrease it, with the two opposite-sign effects

canceling each other out. The most straightforward policy implication of such a combination of effects

would be that gun control measures should focus on reducing gun prevalence among criminals while

avoiding reducing it among noncriminals.

It might be argued that we failed to find support for the more guns, more homicide thesis because

PSG serves primarily as an indicator of gun prevalence among noncriminals, especially in suburban/rural

counties. Even if this were true, however, one would still expect criminal gun prevalence to be positively

correlated with noncriminal gun prevalence, if for no other reason than that most criminals acquire guns

as a direct or indirect result of thefts from noncriminals (Wright and Rossi 1986, p. 196). Thus, PSG

would still measure criminal gun prevalence, but less strongly than more direct measures. Therefore, a

more precise variant of this speculation would be that PSG might be a weaker proxy for criminal gun

prevalence than it is for noncriminal gun prevalence, or that the excluded instruments that we use

(RGUNMAG and PCTDEM88) are more weakly correlated with criminal than with noncriminal gun

prevalence; either would lead to weaker associations between gun prevalence and violence rates than

would be obtained if we could more specifically measure criminal gun prevalence. It bears repeating,

however, that at this point this idea is nothing more than a plausible but empirically unsupported

speculation. Therefore, identifying proxies that can separately measure criminal and noncriminal gun

prevalence should be a top priority for future research.

40

Table 1. Macro-Level Studies of the Impact of Gun Levels on Crime Rates

a

Study Sample

Gun Measure

b

Crime Rates

c

Results

d

Simul?

e

Brearley (1932) 42 states PGH THR Yes No

Krug (1967) 50 states HLR ICR No No

Newton and Zimring (1969) 4 years, Detroit NPP THR,TRR,AAR,

GHR

Yes No

Seitz (1972) 50 states GHR,FGA,AAR THR Yes No

Murray (1975) 50 states SGR,SHR GHR,AAR,TRR No No

Fisher (1976) 9 years, Detroit NPP,GRR,PGH THR Yes (No)

Phillips et al. (1976) 18 years, U.S. PROD THR Yes No

Brill (1977) 11 cities PGC ICR

THR

TRR

No

Yes

No

No

Kleck (1979) 27 years, U.S. PROD THR Yes (No)

Cook (1979) 50 cities PGH,PSG TRR

RMR

No

Yes

No

Kleck (1984) 32 years, U.S. PROD THR

TRR

No

Yes

(No)

Maggadino and Medoff (1984) 31 years, U.S. PROD THR No (No)

Lester (1985) 37 cities PCS VCR No No

Bordua (1986) 102 counties

9 regions

GLR,SIR HAR,THR,GHR No No

McDowall (1986) 48 cities, 2 years PGH,PSG TRR No (No)

Lester (1988b) 9 regions SGR THR Yes No

McDowall (1991) 36 years, Detroit PSG,PGR THR Yes (No)

Killias (1993) 16 nations SGR THR,GHR Yes No

Kleck and Patterson (1993) 170 cities 5-item factor

incl. PSG

f

THR,GHR,TRR,

GRR,AAR,GAR

No Yes

Lester (1996) 12 nations PGH,PSG THR,GHR Yes No

Southwick (1997) 48 years, U.S. PROD THR,TPR,TRR,

AAR

No Yes

Southwick (1999) 34 years, U.S. HGS THR,TRR,AAR,

TPR,VCR,BUR

No No

Hemenway and Miller (2000) 26 nations PGH,PSG THR Yes No

Lott (2000) 15 states, 2 years

g

SGR THR,TPR,TRR,

AAR, 3 others

No No

Stolzenberg and D'Alessio

(2000)

4 years, 46

counties

CCW, GUNSTOL VCR Yes No

Duggan (2001) 19 years, 50 states GMR THR,TPR,TRR,

AAR

Yes No

Hoskin (2001) 36 nations PSG THR Yes (No)

Killias et al. (2001) 21 nations SGR THR,TRR,TAR,

GHR,GRR,GAR

No No

Sorenson and Berk (2001) 22 years, California HGS THR Yes (No)

Cook and Ludwig (2002) 22 years, 50 states PSG BUR Yes (No)

Miller et al. (2002) 10 years, 50 states

10 years, 9 regions

PSG,PHG

SGR

THR

THR

Yes

No

No

No

Ruddel and Mays (2005) 50 states 3-item factor incl.

PSG

h

THR Yes No

Moody and Marvell (2005) 50 states, 22 years PSG, SHR THR,TPR,TRR,

AAR,BUR

No Yes

41

Notes to Table 1:

a. Table covers only studies and findings where the dependent variable was a crime rate, as opposed to the

fraction of crimes committed with guns, and where gun ownership levels were actually measured, rather

than assumed. Studies that examined only gun violence rates (e.g., only gun homicides) are excluded.

b. Measures of Gun Level: CCW = concealed carry permits rate; FGA = Fatal gun accident rate; GLR =

Gun owners license rate; GMR = Gun magazine subscription rates; GRR = Gun registrations rate;

GUNSTOL = % of $ value of stolen property due to guns; HGS = handgun sales (retail); HLR = Hunting

license rate; NPP = Number of handgun purchase permits; PGA = % aggravated assaults committed with

guns; PGC = % homicides, aggravated assaults and robberies (combined together) committed with guns;

PCS = same as PGC, but with suicides lumped in as well; PGH = % homicides committed with guns;

PGR = % robberies committed with guns; PSG = % suicides committed with guns; PROD= Guns

produced minus exports plus imports, U.S.; SGR = Survey measure, % households with gun(s); SHR =

Survey measure, % households with handgun(s); SIR = Survey measure, % individuals with gun(s)

c. Crime Rates: AAR = Aggravated assault rate; BUR = burglary rate; GAR = Gun aggravated assault

rate; GHR = Gun homicide rate; GRR = Gun robbery rate; HAR = Homicide, assault and robbery index

(factor score); ICR = Index crime rate; RMR = Robbery murder rate; THR = Total homicide rate; TPR =

Total rape rate; TRR = Total robbery rate; VCR = Violent crime rate

d. Yes=Study found significant positive association between gun levels and violence; No=Study did not

find such a link.

e. Did research address possible simultaneous relationship between gun levels and crime rates with

properly identified model? (No) means researchers tried to address the issue, but model was still

underidentified.

f. Five-item factor composed of PSG, PGH, PGR, PGA, and the percent of dollar value of stolen property

due to stolen guns.

g. Panel design, two waves.

h. Three-item factor composed of PSG, the gun theft rate, and the fatal gun accident rate.

42

Table 2a: Descriptive Statistics Mean

Std.

Dev.

Min Max

Homicide variables, 1987-93 average

CRGMUR Gun homicides per 100,000 population 4.13 4.18 0.00 46.08

Log CRGMUR " logged 1.01 0.96 -2.02 3.83

CRNGMUR Nongun homicides per 100,000 pop. 2.17 1.73 0.00 13.39

Log CRNGMUR " logged 0.53 0.78 -1.74 2.59

CRMUR Total homicides per 100,000 pop. 6.30 5.62 0.00 57.67

Log CRMUR " logged 1.50 0.86 -1.01 4.05

Gun availability, 1987-93 average

PSG % suicides with guns 66.67 13.46 15.28 100.00

Log PSG " logged 4.18 0.23 2.73 4.61

Excluded instruments

PCTDEM88 % presidential vote Democrat, 1988 42.49 9.87 17.70 84.74

Log PCTDEM88 " logged 3.72 0.24 2.87 4.44

RGUNMAG

Principal components measure of 3 top

outdoor/sport magazine subscriptions

0.00 1.00 -4.79 2.39

Controls

DENSITY Persons per square mile 418 2075 2 53126

Log DENSITY " logged 4.78 1.27 0.47 10.88

PCTURBAN % urban (inside urbanized area) 28.32 36.95 0.00 100.00

PCTSUBURBAN % suburban (outside urbanized area) 25.52 22.22 0.00 100.00

PCTRURAL % rural (farm+nonfarm) 46.16 26.14 0.00 100.00

PCT0T17 % aged 17 and under 26.34 3.24 15.10 41.70

PCT18T24 % aged 18-24 10.70 3.72 5.10 37.10

PCT25T44 % aged 25-44 30.94 3.02 20.30 45.30

PCT45T64 % aged 45-64 18.95 2.13 8.40 27.10

PCT65PLUS % aged 65 and over 13.08 3.59 3.00 33.80

PCTBLK % African-American 9.24 12.67 0.01 72.13

Log PCTBLK " logged 1.05 1.82 -4.36 4.28

PCTHISP % Hispanic 4.43 10.25 0.14 97.22

Log PCTHISP " logged 0.41 1.30 -1.97 4.58

PCTFEM18 % female-headed HHs w/children < 18 58.54 7.11 33.40 84.10

Log PCTFEM18 " logged 4.06 0.12 3.51 4.43

PCTEDUC % aged 25+ with a BA degree or higher 16.06 7.36 4.60 52.30

Log PCTEDUC " logged 2.68 0.42 1.53 3.96

PCTTRANS % born out of state 31.22 15.85 5.09 86.54

Log PCTTRANS " logged 3.32 0.51 1.63 4.46

Log MEDHHINC Log median household income, 1989 10.18 0.24 9.23 10.99

PCTINCLT15K % households with income < $15,000 27.86 8.88 5.00 65.20

Log PCTINCLT15K " logged 3.27 0.36 1.61 4.18

INEQUALITY % HHs w/income <$15k / % income >$75k 0.32 0.50 0.02 6.74

Log INEQUALITY " logged -1.64 0.89 -4.08 1.91

PCTPOOR % persons below poverty line, 1989 14.26 6.89 2.20 60.00

Log PCTPOOR " logged 2.55 0.49 0.79 4.09

PCTUNEMP % persons unemployed 6.65 2.47 1.50 23.60

Log PCTUNEMP " logged 1.83 0.36 0.41 3.16

PCTVACANT % housing units vacant 11.03 7.65 2.70 66.20

Log PCTVACANT " logged 2.24 0.54 0.99 4.19

43

Table 2b: Alternative measures of logged homicide variables Obs Mean

Std.

Dev.

Min Max

Log CRGMUR

Logged, dropping zero-rate counties 1442 1.01 0.96 -2.02 3.83

"Add 1", logged 1462 1.11 0.88 -1.33 3.83

Winsorized lower tail, logged 1462 0.97 1.02 -2.02 3.83

Two-sided winsorized, logged 1462 0.96 1.01 -2.02 2.84

Log CRNGMUR

Logged, dropping zero-rate counties 1410 0.53 0.78 -1.74 2.59

"Add 1", logged 1462 0.65 0.70 -1.52 2.60

Winsorized lower tail, logged 1462 0.45 0.87 -1.74 2.59

Two-sided winsorized, logged 1462 0.44 0.86 -1.74 1.77

Log CRMUR

Logged, dropping zero-rate counties 1459 1.50 0.86 -1.01 4.05

"Add 1", logged 1462 1.58 0.79 -0.83 4.06

Winsorized lower tail, logged 1462 1.49 0.87 -1.01 4.05

Two-sided winsorized, logged 1462 1.49 0.87 -1.01 3.73

Sources for Tables 2a and 2b: U.S. Bureau of the Census, County and City Data Book (1994), except for

(a) homicide rates and PSG, from U.S. NCHS (1997); (b) PCTDEM88, from ICPSR (1995); (c) magazine

subscription rates used to construct RGUNMAG, from Audit Bureau of Circulations (1993); (d) rurality

measures, from U.S. Bureau of the Census (2000).

44

Table 3: Log gun homicide equation

Dependent variable: Log CRGMUR

2-step Efficient GMM

(heteroskedastic-efficient)

2-step Efficient GMM

(heteroskedastic- and cluster-efficient)

PSG-exogenous PSG-endogenous PSG-exogenous PSG-endogenous

Log PSG 0.290* -1.500** 0.389*** -1.530**

(0.115) (0.521) (0.112) (0.528)

Log DENSITY 0.241*** 0.145*** 0.226*** 0.143***

(0.027) (0.039) (0.030) (0.042)

PCTSUBURBAN -0.004* -0.005*** -0.003* -0.005***

(0.001) (0.002) (0.001) (0.001)

PCTURBAN -0.001 -0.002* -0.001 -0.002*

(0.001) (0.001) (0.001) (0.001)

PCT0T17 0.032** 0.043*** 0.030* 0.043**

(0.010) (0.012) (0.013) (0.015)

PCT18T24 -0.013 -0.003 -0.013 -0.003

(0.009) (0.010) (0.008) (0.009)

PCT25T44 0.026** 0.025** 0.024** 0.025**

(0.009) (0.010) (0.008) (0.008)

PCT45T64 0.053** 0.074*** 0.055*** 0.075***

(0.017) (0.020) (0.017) (0.016)

Log PCTBLK 0.149*** 0.161*** 0.150*** 0.161***

(0.014) (0.016) (0.016) (0.016)

Log PCTHISP 0.061* 0.016 0.041 0.015

(0.024) (0.030) (0.032) (0.033)

Log PCTFEM18 -0.118 0.393 -0.153 0.406

(0.192) (0.251) (0.262) (0.339)

Log PCTEDUC -0.375*** -0.468*** -0.333** -0.483***

(0.097) (0.108) (0.116) (0.124)

Log PCTTRANS -0.023 0.008 0.012 0.009

(0.055) (0.058) (0.065) (0.077)

Log MEDHHINC -0.294 -0.100 -0.297 -0.104

(0.310) (0.338) (0.411) (0.448)

Log PCTINCLT15K 0.508 0.900** 0.282 0.896*

(0.265) (0.302) (0.340) (0.402)

Log INEQUALITY 0.392*** 0.422*** 0.354** 0.433***

(0.093) (0.102) (0.112) (0.115)

Log PCTPOOR 0.726*** 0.519** 0.799*** 0.532**

(0.142) (0.162) (0.206) (0.206)

Log PCTUNEMP -0.020 0.097 0.060 0.090

(0.083) (0.096) (0.092) (0.092)

Log PCTVACANT 0.144*** 0.153*** 0.158*** 0.153***

(0.041) (0.045) (0.040) (0.045)

R

2

0.442 0.333 0.439 0.329

N 1441 1441 1441 1441

45

Table 3: Log gun homicide equation (continued)

Dependent variable: Log CRGMUR

2-step Efficient GMM

(heteroskedastic-efficient)

2-step Efficient GMM

(heteroskedastic- and cluster-efficient)

PSG-exogenous PSG-endogenous PSG-exogenous PSG-endogenous

J statistic χ

2

(2)=13.29** χ

2

(1)=0.254 χ

2

(2)=8.18* χ

2

(1)=0.162

p-value 0.0013 0.6145 0.0168 0.6874

C statistic χ

2

(1)=13.01*** χ

2

(1)=8.01**

p-value 0.0003 0.0047

95% confidence interval

for log PSG

Wald [-2.52, -0.48] [-2.56, -0.50]

Anderson-Rubin [-3.05 , -0.34] [-3.35 , -0.06]

First-stage regression:

F statistic 26.2 12.2

Log PCTDEM88 -0.075*** -0.075**

(0.022) (0.027)

RGUNMAG 0.046*** 0.046***

(0.008) (0.013)

Notes: * p<0.05, ** p<0.01, *** p<0.001.

Standard errors in parentheses.

Excluded instruments are log PCTDEM88 and RGUNMAG.

R

2

is the within-R

2

(see text).

All equations are estimated with 50 state fixed effects.

46

Table 4: Log nongun homicide equation

Dependent variable: Log CRNGMUR

2-step Efficient GMM

(heteroskedastic-efficient)

2-step Efficient GMM

(heteroskedastic- and cluster-efficient)

PSG-exogenous PSG-endogenous PSG-exogenous PSG-endogenous

Log PSG

-0.009 -0.490 0.013 -0.559

(0.107) (0.478) (0.111) (0.552)

Log DENSITY

0.135*** 0.110*** 0.134*** 0.107**

(0.022) (0.033) (0.021) (0.034)

PCTSUBURBAN

0.003* 0.002 0.003* 0.002

(0.001) (0.001) (0.001) (0.001)

PCTURBAN

0.003** 0.002* 0.003** 0.003**

(0.001) (0.001) (0.001) (0.001)

PCT0T17

0.017 0.020* 0.017 0.019*

(0.009) (0.010) (0.009) (0.009)

PCT18T24

-0.034*** -0.032*** -0.031*** -0.033***

(0.008) (0.009) (0.007) (0.008)

PCT25T44

0.004 0.003 0.007 0.005

(0.008) (0.008) (0.008) (0.009)

PCT45T64

-0.012 -0.007 -0.009 -0.010

(0.015) (0.016) (0.014) (0.014)

Log PCTBLK

0.152*** 0.154*** 0.150*** 0.154***

(0.013) (0.013) (0.010) (0.010)

Log PCTHISP

0.062** 0.046 0.061* 0.045

(0.023) (0.027) (0.030) (0.032)

Log PCTFEM18

0.073 0.229 0.042 0.235

(0.181) (0.238) (0.148) (0.256)

Log PCTEDUC

-0.196* -0.219* -0.232** -0.230**

(0.085) (0.089) (0.080) (0.084)

Log PCTTRANS

0.057 0.060 0.079 0.071

(0.048) (0.048) (0.050) (0.050)

Log MEDHHINC

0.324 0.395 0.334 0.377

(0.274) (0.283) (0.308) (0.303)

Log PCTINCLT15K

0.916*** 1.011*** 1.020*** 1.099***

(0.231) (0.251) (0.259) (0.270)

Log INEQUALITY

0.214* 0.220* 0.242* 0.228*

(0.092) (0.093) (0.101) (0.103)

Log PCTPOOR

0.355** 0.309* 0.320* 0.240

(0.124) (0.133) (0.138) (0.155)

Log PCTUNEMP

0.116 0.151 0.130 0.178

(0.083) (0.091) (0.095) (0.104)

Log PCTVACANT

0.045 0.044 0.047 0.039

(0.038) (0.038) (0.047) (0.046)

R

2

0.435 0.425 0.434 0.422

N

1409 1409 1409 1409

47

Table 4: Log nongun homicide equation (continued)

Dependent variable: Log CRNGMUR

2-step Efficient GMM

(heteroskedastic-efficient)

2-step Efficient GMM

(heteroskedastic- and cluster-efficient)

PSG-exogenous PSG-endogenous PSG-exogenous PSG-endogenous

J statistic

χ

2

(2)=1.99 χ

2

(1)=0.898 χ

2

(2)=2.43 χ

2

(1)=1.37

p-value

0.3703 0.3432 0.2964 0.2413

C statistic

χ

2

(1)=1.07 χ

2

(1)=0.977

p-value

0.3005 0.3230

95% confidence interval

for log PSG

Wald

[-1.43, 0.45] [-1.64, 0.52]

Anderson-Rubin

[-1.71, 0.62] [-2.10, 1.06]

First-stage regression:

F statistic

24.2 10.9

Log PCTDEM88

-0.071** -0.071*

(0.022) (0.027)

RGUNMAG

0.045*** 0.045**

(0.008) (0.013)

Notes: * p<0.05, ** p<0.01, *** p<0.001.

Standard errors in parentheses.

Excluded instruments are log PCTDEM88 and RGUNMAG.

R

2

is the within-R

2

(see text).

All equations are estimated with 50 state fixed effects.

48

Table 5: Log total homicide equation

Dependent variable: Log CRMUR

2-step Efficient GMM

(heteroskedastic-efficient)

2-step Efficient GMM

(heteroskedastic- and cluster-efficient)

PSG-exogenous PSG-endogenous PSG-exogenous PSG-endogenous

Log PSG

0.160 -1.023* 0.225* -1.023*

(0.095) (0.437) (0.094) (0.450)

Log DENSITY

0.212*** 0.147*** 0.200*** 0.147***

(0.021) (0.033) (0.023) (0.032)

PCTSUBURBAN

-0.001 -0.003* -0.001 -0.003*

(0.001) (0.001) (0.001) (0.001)

PCTURBAN

-0.000 -0.001 0.000 -0.001

(0.001) (0.001) (0.001) (0.001)

PCT0T17

0.024** 0.030** 0.022** 0.030**

(0.009) (0.010) (0.008) (0.010)

PCT18T24

-0.021** -0.015 -0.020** -0.015

(0.007) (0.008) (0.007) (0.008)

PCT25T44

0.020** 0.020* 0.023** 0.020*

(0.008) (0.008) (0.007) (0.008)

PCT45T64

0.034* 0.048** 0.037** 0.048***

(0.014) (0.016) (0.012) (0.013)

Log PCTBLK

0.154*** 0.161*** 0.153*** 0.161***

(0.012) (0.013) (0.013) (0.013)

Log PCTHISP

0.061** 0.027 0.041 0.027

(0.020) (0.025) (0.021) (0.024)

Log PCTFEM18

0.070 0.427 0.008 0.426

(0.166) (0.218) (0.185) (0.255)

Log PCTEDUC

-0.273*** -0.317*** -0.290** -0.317***

(0.081) (0.087) (0.088) (0.092)

Log PCTTRANS

-0.004 0.008 0.021 0.008

(0.047) (0.049) (0.058) (0.064)

Log MEDHHINC

-0.084 0.032 -0.126 0.032

(0.261) (0.276) (0.326) (0.343)

Log PCTINCLT15K

0.584** 0.848*** 0.502 0.848**

(0.223) (0.256) (0.279) (0.317)

Log INEQUALITY

0.309*** 0.325*** 0.340*** 0.325***

(0.080) (0.085) (0.088) (0.090)

Log PCTPOOR

0.599*** 0.441** 0.674*** 0.441*

(0.117) (0.137) (0.156) (0.175)

Log PCTUNEMP

0.115 0.207* 0.162* 0.208*

(0.073) (0.086) (0.081) (0.085)

Log PCTVACANT

0.080* 0.074 0.099** 0.074*

(0.036) (0.038) (0.031) (0.030)

R

2

0.527 0.463 0.525 0.463

N

1458 1458 1458 1458

49

Table 5: Log total homicide equation (continued)

Dependent variable: Log CRMUR

2-step Efficient GMM

(heteroskedastic-efficient)

2-step Efficient GMM

(heteroskedastic- and cluster-efficient)

PSG-exogenous PSG-endogenous PSG-exogenous PSG-endogenous

J statistic

χ

2

(2)=8.51 χ

2

(1)<0.001 χ

2

(2)=6.01 χ

2

(1)<0.001

p-value

0.0142 0.9907 0.0497 0.9922

C statistic

χ

2

(1)=8.51 χ

2

(1)=6.01

p-value

0.0035 0.0143

95% confidence interval

for log PSG

Wald

[-1.88, -0.17] [-1.90, -0.14]

Anderson-Rubin

[-2.35, -0.02] [-2.64, 0.21]

First-stage regression:

F statistic

24.8 11.3

Log PCTDEM88

-0.072*** -0.072*

(0.022) (0.027)

RGUNMAG

0.045*** 0.045***

(0.008) (0.013)

Notes: * p<0.05, ** p<0.01, *** p<0.001.

Standard errors in parentheses.

Excluded instruments are log PCTDEM88 and RGUNMAG.

R

2

is the within-R

2

(see text).

All equations are estimated with 50 state fixed effects.

50

REFERENCES

Anderson, Thomas W., and Herman Rubin. 1949. “Estimation of the parameters of a single equation in a

complete system of stochastic equations.” Annals of Mathematical Statistics 91: 46-63.

Andrews, Donald W.K., and James H. Stock. 2005. “Inference with weak instruments.” NBER Technical

Working Paper 313, August. http://www.nber.org/papers/T0313.

Audit Bureau of Circulations. 1993. Supplementary Data Report, covering county paid circulation for gun

and related sports magazines. Schaumburg, IL: Audit Bureau of Circulations.

Azrael, Deborah, Philip J. Cook and Matthew Miller. 2004. “State and local prevalence firearms

ownership: measurement, structure, and trends.” Journal of Quantitative Criminology (forthcoming).

Basmann, Robert L. 1960. “On finite sample distributions of generalized classical linear identifiability

test statistics.” Journal of the American Statistical Association 55:650-659.

Baum, Christopher F., Mark E. Schaffer, and Steven Stillman. 2003. “Instrumental variables and GMM:

Estimation and testing.” The Stata Journal 3:1-31.

Baum, Christopher, Mark E. Schaffer and Steven Stillman. 2005. “IVREG2: Stata module for extended

instrumental variables/2SLS and GMM estimation”. http://ideas.repec.org/c/boc/bocode/s425401.html.

Bordua, David J. 1986. “Firearms ownership and violent crime: a comparison of Illinois counties.” Pp.

156-88 in The Social Ecology of Crime, edited by James M. Byrne and Robert J. Sampson. N.Y.:

Springer-Verlag.

Bordua, David J., and Alan J. Lizotte. 1979. “Patterns of legal firearms ownership: a cultural and

situational analysis of Illinois counties.” Law and Policy Quarterly 1:147-75.

Bound, John, David A. Jaeger, and Regina Baker. 1995. “Problems with instrumental variables estimation

when the correlation between the instruments and the endogenous explanatory variable is weak.” Journal

of the American Statistical Association 90:443-450.

Brearley, H.C. 1932. Homicide in the United States. Chapel Hill: University of North Carolina Press.

Brill, Steven. 1977. Firearm Abuse: A Research and Policy Report. Washington, D.C.: Police

Foundation.

Clarke, Ronald V., and Pat Mayhew. 1988. “The British gas suicide story and its criminological

implications.” Pp 79-116 in Crime and Justice, Vol. 10, edited by Michael Tonry and Norval Morris.

Chicago: University of Chicago Press.

Clotfelter, Charles T. 1981. “Crime, disorders, and the demand for handguns.” Law & Policy Quarterly

3:425-446.

Cook, Philip J. 1976. “A strategic choice analysis of robbery.” Pp. 173-87 in Sample Surveys of the

Victims of Crime, edited by Wesley Skogan. Cambridge: Ballinger.

____. 1982. “The role of firearms in violent crime.” Pp.236-91 in Criminal Violence, edited by Marvin E.

Wolfgang and Neil Alan Weiner. Beverly Hills: Sage.

51

Cook, Philip J., and Jens Ludwig. 1997. Guns in America. Washington, D.C.: Police Foundation.

____. 2000. Gun Violence: The Real Costs. N.Y.: Oxford.

____. 2003. “Guns and burglary.” Pp. 74-118 in Evaluating Gun Policy, edited by Jens Ludwig and Philip

J. Cook. Washington, D.C.: Brookings Institution Press.

Cox, Nick. 2003. “WINSOR: Stata module to Winsorize a variable”.

http://ideas.repec.org/c/boc/bocode/s361402.html.

Cragg, J. 1983. “More efficient estimation in the presence of heteroskedasticity of unknown form.”

Econometrica 51:751-763.

Dufour, J.M. 2003. “Identification, weak instruments and statistical inference in econometrics.” CIRANO

Working Paper 2003s-49. http://www.cirano.qc.ca/pdf/publication/2003s-49.pdf.

Duggan, Mark. 2001. “More guns, more crime.” Journal of Political Economy 109:1086-1114.

Duwe, Grant. 2000. “Body-count journalism.” Homicide Studies 4:364-399.

Fisher, Joseph C. 1976. “Homicide in Detroit: The role of firearms.” Criminology 14:387-400.

Hahn, Jinyong and Jerry Hausman. 2003. “Weak instruments: Diagnoses and cures in empirical

econometrics.” American Economic Review 93(2):118-125.

Hall, A.R. and F.P.M Peixe. 2000. “A consistent method for the selection of relevant instruments.”

Econometric Society World Congress 2000 Contributed Papers.

http://econpapers.repec.org/paper/ecmwc2000/0790.htm

Hansen, Lars. 1982. “Large sample properties of generalized method of moments estimators.”

Econometrica 50: 1029-1054.

Hayashi, Fumio. 2000. Econometrics. Princeton: Princeton University Press.

Hemenway, David, and Matthew Miller. 2000. “Firearm availability and homicide rates across 26 high-

income countries.” Journal of Trauma 49:985-988.

Hoskin, Anthony W. 2001. “Armed Americans.” Justice Quarterly 18:569-592.

Inter-university Consortium for Political and Social Research (ICPSR). 1995. General Election Data for

the United States, 1950-1990. [Computer file]. ICPSR ed. Ann Arbor, MI: Inter-university Consortium

for Political and Social Research [producer and distributor], 1995.

____. 2000. Uniform Crime Reporting Program Data [United States]: Offenses known and Clearances by

Arrest, 1989 [1990, 1991] [Computer file]. Compiled by U.S Department of Justice, Federal Bureau of

Investigation. ICPSR ed. Ann Arbor, MI: Inter-university Consortium for Political and Social Research

[producer and distributor], 2000.

____. 2001. Uniform Crime Reporting Program Data [United States]: Supplementary Homicide Reports,

1976-1999 [Computer file]. Compiled by U.S Department of Justice, Federal Bureau of Investigation.

52

ICPSR ed. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [producer and

distributor], 2001.

Killias, Martin. 1993. “Gun ownership, suicide, and homicide: an international perspective.” Pp. 289-303

in Understanding Crime: Experiences of Crime and Crime Control, edited by Anna del Frate, Uglijesa

Zvekic, and Jan J. M. van Dijk. Rome: UNICRI.

Killias, Martin, John van Kesteren, and Martin Rindlisbacher. 2001. “Guns, violent crime, and suicide in

21 countries.” Canadian Journal of Criminology 43:429-448.

Kleck, Gary. 1979. “Capital punishment, gun ownership, and homicide.” American Journal of Sociology

84:882-910.

____. 1984. “The relationship between gun ownership levels and rates of violence in the United States.”

Pp. 99- 135 in Firearms and Violence: Issues of Public Policy, edited by Don B. Kates, Jr. Cambridge,

Mass.: Ballinger.

____. 1997. Targeting Guns: Firearms and their Control. N.Y.: Aldine.

____. 2001. “Modes of news media distortion of gun issues.” Pp. 173-212 in Armed: New Perspectives on

Gun Control, edited by Gary Kleck and Don B. Kates, Jr. Amherst, NY: Prometheus.

____. 2004. “Measures of gun ownership levels for macro-level crime and violence research.” Journal of

Research in Crime and Delinquency 41(1):3-36.

Kleck, Gary, Chester L. Britt, and David J. Bordua. 2000. “The emperor has no clothes: using interrupted

time series designs to evaluate social policy impact.” Journal on Firearms and Public Policy 12:197-247.

Kleck, Gary, and Miriam DeLone. 1993. “Victim resistance and offender weapon effects in robbery.”

Journal of Quantitative Criminology 9:55-82.

Kleck, Gary, and Don B. Kates. 2001. Armed: New Perspectives on Gun Control. Amherst, NY:

Prometheus.

Kleck, Gary, and Tomislav Kovandzic. 2001. “The impact of gun laws and gun levels on crime rates.”

Paper presented at the annual meetings of the American Society of Criminology, Atlanta, GA.

Kleck, Gary, and Karen McElrath. 1991. “The effects of weaponry on human violence.” Social Forces

69:669-92.

Kleck, Gary, and E. Britt Patterson. 1993. “The impact of gun control and gun ownership levels on

violence rates.” Journal of Quantitative Criminology 9:249-288.

Lester, David. 1985. “The use of firearms in violent crime.” Crime & Justice 8:115-20.

____. 1988b. “Firearm availability and the incidence of suicide and homicide.” Acta Psychiatrica

Belgium 88:387-393.

____. 1996. “Gun ownership and rates of homicide and suicide.” European Journal of Psychiatry 10:83-

85.

53

Lott, John R., Jr. 2000. More Guns, Less Crime. 2nd edition. Chicago: University of Chicago Press.

Marvell, Thomas B., and Carlisle E. Moody, Jr. 1991. “Age structure and crime rates: the conflicting

evidence.” Journal of Quantitative Criminology 7(3):237-273.

McDowall, David. 1986. “Gun availability and robbery rates: a panel study of large U.S. cities, 1974-

1978.” Law & Policy 8:135-48.

____. 1991. “Firearm availability and homicide rates in Detroit, 1951-1986.” Social Forces 69:1085-

1099.

McDowall, David, and Colin Loftin. 1983. “Collective security and the demand for handguns.” American

Journal of Sociology 88:1146-1161.

Miller, Matthew, Deborah Azrael, and David Hemenway. 2002. “Firearm availability and unintentional

firearm deaths, suicide, and homicide among 5-14 year olds.” Journal of Trauma Injury, Infection, and

Critical Care 52:267-275.

Moody, Carlisle E. 2001. “Testing for the effects of concealed weapons laws: Specification errors and

robustness.” Journal of Law and Economics 44:799-813.

Moody, Carlisle E., and Thomas B. Marvell. 2003. “Pitfalls of using proxy variables in studies of guns

and crime”. SSRN working paper. http://ssrn.com/abstract=473661.

Moody, Carlisle E., and Thomas B. Marvell. 2005. “Guns and crime.” Southern Economic Journal

71:720-736.

Newton, George D., and Franklin Zimring. 1969. Firearms and Violence in American Life. A Staff Report

to the National Commission on the Causes and Prevention of Violence. Washington, D.C.: U.S.

Government Printing Office.

Okoro, Catherine A., David E. Nelson, James A. Mercy, Lina S. Balluz, Alex E. Crosby, and Ali H.

Mokdad. 2005. “Prevalence of household firearms and firearm-storage practices in the 50 states and the

District of Columbia.” Pediatrics 116:e370-e376.

Pagan, A.R. and D. Hall. 1983. “Diagnostic tests as residual analysis.” Econometric Reviews 2(2):159-

218.

Phillips, Llad, Harold L. Votey, and John Howell. 1976. “Handguns and homicide.” Journal of Legal

Studies 5:463-78.

Rice, Douglas C., and David D. Hemley. 2002. “The market for new handguns.” Journal of Law and

Economics 45:251-265.

Ruddell, Rick, and G. Larry Mays. 2005. “State background checks and firearms homicides.” Journal of

Criminal Justice 33:127-136.

Sargan, D. 1958. “The estimation of econometric relationships using instrumental variables.”

Econometrica 26:393-415.

Sommers, Paul M. 1984. “Letter to the Editor.” New England Journal of Medicine 310:47-8.

54

Sorenson, Susan B, and Richard A. Berk. 2001. “Handgun sales, beer sales, and youth homicide,

California, 1972-1993.” Journal of Public Health Policy 22:183-197.

Southwick, Lawrence, Jr. 1997. “Do guns cause crime? Does crime cause guns?: a Granger test.” Atlantic

Economic Journal 25:256-273.

____. 1999. “Guns and justifiable homicide: deterrence and defense.” Saint Louis University Public Law

Review 18:217-246.

Staiger, Douglas, and James H. Stock. 1997. “Instrumental variables regression with weak instruments.”

Econometrica 65:557-586.

Stock, James H., Jonathan H. Wright and Motohiro Yogo, 2002. “A survey of weak instruments and weak

identification in generalized method of moments.” Journal of Business and Economic Statistics

20(4):518:529.

Stock, J.H., and M. Yogo. 2002. “Testing for weak instruments in linear IV regression.” NBER Technical

Working Paper 284. http://www.nber.org/papers/T0284.

Stolzenberg, Lisa, and Stewart J. D'Alessio. 2000. “Gun availability and violent crime.” Social Forces

78:1461-1482.

Tark, Jongyeon, and Gary Kleck. 2004. “Resisting crime: the effects of victim action on the outcomes of

crimes.” Criminology 42(4):861-909.

U.S. Bureau of the Census. 1994. County and City Data Book, 1994. Washington, D.C.: U.S.

Government Printing Office.

U.S. Bureau of the Census. 1990. “Census 1990 Summary File 3 (SF3) – Sample Data, Table P006 Urban

and Rural”. Retrieved 7 February 2005 from U.S. Census http://factfinder.census.gov.

U.S. Bureau of Justice Statistics. 2001. Criminal Victimization in the United States - Statistical Tables.

Tables available online at http://www.ojp.usdoj.gov/bjs/abstrat/cvusst.htm. Accessed 12-6-01.

U.S. Federal Bureau of Investigation (FBI). 1990-2000a. Crime in The United States 1989 [-1999]

Uniform Crime Reports. Washington, D.C.: U.S. Government Printing Office.

U.S. National Center for Health Statistics. 1997. Special versions of Mortality Detail Files, 1987-1993,

with location detail, supplied to third author.

Vieraitis, Lynne M. 2000. “Income inequality, poverty, and violent crime: A review of the empirical

evidence.” Social Pathology 6(1):24-45.

Wooldridge, Jeffrey M. 1995. “Score diagnostics for linear models estimated by two stage least squares.”

In Advances in Econometrics and Quantitative Economics: Essays in honor of Professor C. R. Rao, eds.

G. S. Maddala, P. C. B. Phillips, and T. N. Srinivasan. Pp. 66–87. Cambridge, MA: Blackwell Publishers.

____. 2002. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: MIT Press.

55

____. 2003. “Cluster-sample methods in applied econometrics.” American Economic Review 93(2):133-

138.

Wright, James D., and Peter H. Rossi. 1986. Armed and Considered Dangerous. New York: Aldine.

Wright, James D., Peter H. Rossi, and Kathleen Daly. 1983. Under the Gun: Weapons, Crime and

Violence. New York: Aldine.

Zimring, Franklin E., and Gordon Hawkins. 1997. Crime is Not the Problem. N.Y.: Oxford.

Zivot, Eric, Richard Startz, and Charles R. Nelson. 1998. “Valid confidence intervals and inference in the

presence of weak instruments”. International Economic Review 39(4):1119-1144.

56

BIOGRAPHICAL SKETCHES

Tomislav Kovandzic is Assistant Professor of Criminal Justice in Department of Justice Sciences at the

University of Alabama at Birmingham. His current research interests include criminal justice policy and

gun-related violence. His most recent articles have appeared in Criminology & Public Policy,

Criminology, and Homicide Studies. He received his Ph.D. in Criminology from Florida State University

in 1999.

Mark E. Schaffer is Professor of Economics at Heriot-Watt University, Edinburgh, U.K. His research

interests include economic reform in transition and developing economies, and the implementation of

econometric estimators. He is a member of the Executive Committee of the Association for Comparative

Economic Studies (ACES) and is an Associate Editor of the Stata Journal.

Gary Kleck is Professor of Criminology and Criminal Justice at Florida State University. His research

focuses on the links between guns and violence and the deterrent effects of punishment. He is the author

of four books, including Point Blank, which won the 1993 Hindelang Award, and, most recently, Armed

(Prometheus, 2001).

57