Journal of Quantitative Criminology, Vol. 16, No.

1, 2000

Poisson-Based Regression Analysis of Aggregate

Crime Rates
D. Wayne Osgood 1

This article introduces the use of regression models based on the Poisson distri-
bution as a tool for resolving common problems in analyzing aggregate crime
rates. When the population size of an aggregate unit is small relative to the
offense rate, crime rates must be computed from a small number of offenses.
Such data are ill-suited to least-squares analysis. Poisson-based regression models
of counts of offenses are preferable because they are built on assumptions about
error distributions that are consistent with the nature of event counts. A simple
elaboration transforms the Poisson model of offense counts to a model of per
capita offense rates. To demonstrate the use and advantages of this method, this
article presents analyses of juvenile arrest rates for robbery in 264 nonmetropoli-
tan counties in four states. The negative binomial variant of Poisson regression
effectively resolved difficulties that arise in ordinary least-squares analyses.

KEY WORDS: Poisson; negative binomial; crime rates; aggregate analysis.

1. INTRODUCTION
The purpose of this paper is to introduce a statistical approach to ana-
lyzing aggregate crime rates that solves problems arising from small popu-
lations and low base-rates. In aggregate analyses, the units of the sample
are aggregations of individuals, such as neighborhoods, cities, and schools,
and the researcher is interested in explaining variation in crime rates across
those units. These crimes rates are defined as the number of crime events
(e.g., arrests, victimizations, crimes known to the police) divided by the
population size, often reported as crimes per 100,000.
The standard approach to analyzing per capita rates such as these is to
use the computed rates for each aggregate unit (or a transformed version of
them) as the dependent variable in an ordinary least-squares (OLS)
regression. For reasons discussed below, this least-squares approach is inap-
propriate when rates for many of the units must be computed from small
1
Crime, Law and Justice Program, Department of Sociology, 201 Oswald Tower, Pennsylvania
State University, University Park, Pennsylvania 16802. e-mail: [email protected].
21
0748-4518y00y0300-0021$18.00y0  2000 Plenum Publishing Corporation
22 Osgood

numbers of events. The present paper demonstrates how to resolve this

problem through Poisson-based regression models that are well suited to
such data. The statistical basis of this analytic approach is well established
(e.g., Cameron and Trevedi, 1998; Gardner et al., 1995; King, 1989; Liao,
1994; McCullagh and Nelder, 1989; Land et al., 1996). In criminology, Pois-
son-based regression models have become common in analyses of criminal
careers (Greenberg, 1991; Land et al., 1996; Nagin and Land, 1993; Rowe
et al., 1990), but they have rarely been applied to aggregate analysis of crime
or other social phenomena (for examples see Bailey et al., 1994; Sampson
et al., 1997). In the hope of making these techniques accessible to a broader
range of researchers, the present article is devoted to articulating the special
advantages of these models for solving problems in analyzing aggregate
data.

1.1. The Problem

Because arrests are discrete events, the possible crime rates for any
given population size are those corresponding to integer counts of crimes.
For instance, in a county of 200,000 individuals, every additional crime
increases the crime rate by half an arrest per 100,000, while in a neighbor-
hood of 5000 each crime corresponds to 20 crimes per 100,000. If the popu-
lation sizes of the aggregate units are large relative to the average arrest
rate, then the calculated rates will be sufficiently fine-grained that there is
no harm in treating them as though they were continuous and applying
least squares statistics. For almost any measure of offending, populations
of several hundred thousand should prove adequate in this regard. When
populations are small relative to offense rates, however, the discrete nature
of the crime counts cannot be ignored. Indeed, for a population of a few
thousand, even a single arrest for rape or homicide might correspond to a
high crime rate. Low counts of crimes are common for offense-specific
analyses, samples of small towns and rural areas, and analyses of subpopul-
ations, such as females versus males or specific age categories.
Crime rates based on small counts of crimes present two serious prob-
lems for least squares analysis. First, because the precision of the estimated
crime rate depends on population size, variation in population sizes across
the aggregate units will lead to violating the assumption of homogeneity of
error variance. We must expect larger errors of prediction for per capita
crime rates based on small populations than for rates based on large popu-
lations. Second, normal or even symmetrical error distributions of crime
rates cannot be assumed when crime counts are small. The lowest possible
crime count is zero, so the error distribution must become increasingly
skewed (as well as more decidedly discrete) as crime rates approach this
Poisson for Aggregate Data 23

lower bound. As populations decrease, an offense rate of zero will be

observed for a larger and larger proportion of cases. Thus, there is an effec-
tive censoring at zero that is dependent on sample size, which has consider-
able potential for biasing the resulting regression coefficients.
The standard solution to the problems of low offense counts has been
to increase the level of aggregation, such as analyzing only large cities or
combining specific offenses into broad indices. For instance, one rarely see
analyses of homicide for populations less than several hundred thousand.
Not only does this strategy preclude analyses about many interesting ques-
tions, it leads to coarser measurement of important explanatory variables,
such as being forced to assume that a single poverty rate applies equally
well to all neighborhoods in a city. Fortunately, there is an alternative data
analytic approach that resolves these problems.

2. THE POISSON REGRESSION MODEL

The Poisson distribution has been useful for many problems in crimi-
nology and criminal justice. Indeed, Poisson originally derived the distri-
bution for analyzing rates of conviction in France during the 1820s (see
Maltz, 1994). Maltz (1994) reviews many uses of the Poisson distribution
for modeling phenomena related to crime, such as assessing the potential
for selective incapacitation, projecting prison populations, and estimating
the size of the criminal population. The present paper focuses specifically
on Poisson-based regression models, which relate explanatory variables to
dependent variables that are counts of events. These models can solve the
problems described above because they allow us to recognize the depen-
dence of crime rates on counts of crimes. Several good, nontechnical descrip-
tions of Poisson regression are now available (e.g., Gardner et al., 1995;
Liao, 1994; Land et al., 1996), so my description of these models is brief,
emphasizing the features most relevant to aggregate analysis.
The Poisson distribution characterizes the probability of observing any
discrete number of events (i.e., 0, 1, 2, . . .), given an underlying mean count
or rate of events, assuming that the timing of the events is random and
independent. For instance, the Poisson distribution for a mean count of 4.5
would describe the proportion of times that we should expect to observe
any specific count of robberies (0, 1, 2, . . .) in a neighborhood, if the ‘‘true’’
(and unchanging) annual rate for neighborhood were 4.5, if the occurrence
of one robbery had no impact on the likelihood of the next, and if we
had an unlimited number of years to observe. Figure 1 shows the Poisson
distribution for four mean counts of arrests. When the mean arrest count is
low, as is likely for a small population, the Poisson distribution is skewed,
24 Osgood

with only a small range of counts having a meaningful probability of occur-

rence. As the mean count grows, the Poisson distribution increasingly
approximates the normal. The Poisson distribution has a variance equal to
the mean count. Therefore, as the mean count increases, the probability of
observing any specific number of events declines and a broader range of
values have meaningful probabilities of being observed.
Our interest is in per capita crime rates rather than in counts of offen-
ses, and Fig. 2 demonstrates the correspondence between rates and counts.
Figure 2 translates the Poisson distributions of crime counts in Fig. 1 to
distributions of crime rates. Given a constant underlying mean rate of 500
crimes per 100,000 population, population sizes of 200, 600, 2000, and
10,000 would produce the mean crime counts of 1, 3, 10, and 50 used in
Fig. 1. For the population of 200, only a very limited number of crime rates
are probable (i.e., increments of 500 per 100,000), but those probable rates
comprise an enormous range. As the population base increases, the range
of likely crime rates decreases, even though the range of likely crime counts
increases. The standard deviation around the mean rate shrinks from 500
crimes per 100,000 for a population of 200 to 71 crimes per 100,000 for a
population of 10,000. Thus, Fig. 2 illustrates the effect of population size
on the accuracy of estimated crime rates.
The basic Poisson regression model is
K
ln(λ i )G ∑ β k xik (1)
k G0

e −λ iλ iy i
P(Yi Gyi )G (2)
yi !

Equation (1) is a regression equation relating the natural logarithm of the

mean or expected number of events for case i, ln(λ i ), to the sum of the
products of each explanatory variable, xik , multiplied by a regression coef-
ficient, β k (where β 0 is a constant multiplied by 1 for each case). Equation
(2) indicates that the probability of yi , the observed outcome for this case,
follows the Poisson distribution (the right-hand side of the equation) for the
mean count from Eq. (1), λ i . Thus, the expected distribution of crime
counts, and corresponding distribution of regression residuals, depends on
the fitted mean count, λ i , as illustrated in Fig. 1. The role of the natural
logarithm in Eq. (1) is comparable to the logarithmic transformation of the
dependent variable that is common in analysis of aggregate crime rates. In
both cases, the regression coefficients reflect proportional differences in
rates. Liao (1994) provides a detailed discussion of the interpretation of
regression coefficients from Poisson-based models.
 Poisson for Aggregate Data

Fig. 1. Poisson distribution for four mean arrest counts.

25
 26

Fig. 2. Poisson distribution of arrest rates for four population sizes, given a mean rate of 500 arrests per 100,000 population.
Osgood
Poisson for Aggregate Data 27

Next we must alter the basic Poisson regression model so that it pro-
vides an analysis of per capita crime rates rather than counts of crimes. If
λ i is the expected number of crimes in a given aggregate unit, then λ i yni
would be the corresponding per capita crime rate, where ni is the population
size for that unit. With a bit of algebra, we can derive a variation of Eq. (1)
that is a model of per capita crime rates:

λi
1 2
k
ln G ∑ β k xik
ni k G1

K
ln(λ i )Gln(ni )C ∑ β k xik (3)
k G0

Thus, by adding the natural logarithm of the size of the population at risk
to the regression model of Eq. (1), and by giving that variable a fixed coef-
ficient of one, Poisson regression becomes an analysis of rates of events per
capita, rather than an analysis of counts of events. The same strategy can
be used to standardize event count models for other sources of variation
across cases, such as the length of the period of observation. Accordingly,
computer programs for Poisson regression routinely incorporate this
feature.
A Poisson-based regression model that is standardized for the size of
the population at risk acknowledges the greater precision of rates based on
larger populations, thus addressing the problem of heterogeneity of error
variance discussed above. This becomes apparent when we translate the
known variance of the Poisson distribution to a standard deviation for the
corresponding crime rates. Because the variance of the Poisson distribution
is the mean count, λ , its standard deviation will be SDλ G√λ . The mean
count of crimes, in turn, equals the underlying per capita crime rate, C,
times the size of the population: λ GCn. When a variable is divided by a
constant, its standard deviation is also divided by that constant. Therefore,
it follows that the standard deviation of a crime rate, computed from a
population of size n, will be

√λ √Cn √C √n √C
SDC G G G G
n n n √n

This equation shows that, in the expected distribution of observed crime

rates around the fitted mean crime rates produced by Eq. (3), the standard
deviation is inversely proportional to the square root of the population size.
Thus, Poisson regression analysis explicitly addresses the heterogeneous
residual variance that presented a problem for OLS regression analysis of
crime rates.
28 Osgood

2.1. Overdispersion and Variations on the Basic Poisson Regression Model

The basic Poisson regression model is appropriate only if the prob-
ability model of Eq. (2) matches the data. Equation (2) requires that the
residual variance be equal to the fitted values, λ i , which is plausible only if
the assumptions underlying the Poisson distribution are fully met by the
data. One assumption is that λ i is the true rate for each case, which implies
that the explanatory variables account for all of the meaningful variation
among the aggregate units. If not, the differences between the fitted and
true rates will inflate the variance of the residuals. It is very unlikely that
this assumption will be valid, for there is no more reason to expect that a
Poisson regression will explain all of the variation in the true crime rates
than to expect that an OLS regression would explain all variance other than
error of measurement.
Residual variance will also be greater than λ i if the assumption of inde-
pendence among individual crime events is inaccurate. Dependence will
arise if the occurrence of one offense generates a short-term increase in the
probability of another occurring. For aggregate crime data, there are many
potential sources of dependence, such as an individual offending at a high
rate over a brief period until being incarcerated, multiple offenders being
arrested for the same incident, and offenders being influenced by one
another’s behavior. These types of dependence would increase the year-to-
year variability in crime rates for a community beyond λ i , even if the under-
lying crime rate were constant.
For these two reasons, ‘‘overdispersion’’ in which residual variance
exceeds λ i is ubiquitous in analyses of crime data. Applying the basic Pois-
son regression model to such data can produce a substantial underestim-
ation of standard errors of the β ’s, which in turn leads to highly misleading
significance tests. There are several ways to allow for the possibility of over-
dispersion (Cameron and Trevedi, 1998; Land et al., 1996). Perhaps the
simplest is the quasi-likelihood approach (Gardner et al., 1995), which
retains coefficient estimates from the basic Poisson model but adjusts stan-
dard errors and significance tests based on the amount of overdispersion.
Other approaches explicitly incorporate a source of overdispersion in the
probability model, typically by adding a case-specific residual term to the
regression model [Eq. (1) or (3)], comparable to the error term in OLS
regression. These versions of Poisson regression are distinguished by the
specific assumptions made about the distribution of the residual variation
in underlying rates, which may be continuous or discrete (Cameron and
Trevedi, 1998).
We illustrate this approach with the negative binomial regression
model, which is the best known and most widely available Poisson-based
Poisson for Aggregate Data 29

Fig. 3. Negative binomial distributions with mean count of 3, for four levels of residual variance.

regression model that allows for overdispersion. Negative binomial

regression combines the Poisson distribution of event counts with a gamma
distribution of the unexplained variation in the underlying or true mean
event counts, λ i . This combination produces the negative binomial distri-
bution, which replaces the Poisson distribution of Eq. (2). The formula for
the negative binomial is

Γ( yiCφ ) φ φ λ iyi
P(Yi Gyi )G (4)
yi !Γ(φ ) (φ Cλ i )φ Ayi

where Γ is the gamma function (a continuous version of the factorial func-

tion), and φ is the reciprocal of the residual variance of underlying mean
counts, α (Gardner et al., 1995, p. 400).
Figure 3 demonstrates the impact of residual variance on the resulting
distribution for a mean count of three crimes. With α equal to zero, we
have the original Poisson distribution. For the Poisson, 5.0% of cases would
have zero crimes and 1.2% would have eight or more crimes. As α increases,
the distribution becomes more decidedly skewed as well as more broadly
dispersed. Even for a moderate α of 0.75, the change from the Poisson is
dramatic: 20.8% of cases would have zero crimes and 8.8% would have eight
or more crimes.
In negative binomial regression (as in almost all Poisson-based
regression models), the substantive portion of the regression model remains
Eq. (1) for crime counts or Eq. (3) for per capita crime rates. Thus, though
the response probabilities associated with the fitted values differ from the
basic Poisson regression model, the interpretation of the regression coef-
ficients does not.
30 Osgood

3. AN EXAMPLE: SOCIAL DISORGANIZATION AND RURAL

YOUTH VIOLENCE
I illustrate the use of Poisson-based regression to study aggregate crime
rates with an analysis of rates of juvenile violence in nonmetropolitan count-
ies of four states. This is an elaboration of part of the results presented
by (Osgood and Chambers, 2000), and the present article is intended as a
methodological companion to that article. Osgood and Chambers (2000)
provide a rationale for these analyses in terms of social disorganization
theory, offer a full description of the sample and measures, and present
analyses for a variety of specific offenses.
The sample consists of the 264 nonmetropolitan counties of Florida,
Georgia, South Carolina, and Nebraska, which have total populations rang-
ing from 560 to 98,000. The average population of these counties is roughly
10,000, which is comparable to average neighborhood populations in
research comparing neighborhoods within urban centers (Sampson et al.,
1997; Warner and Pierce, 1993).
The measure of offending for these illustrative analyses is the number
of juveniles (ages 11 through 17) arrested for robberies in each county,
pooled over the 5-year period of 1989 through 1993. The measures of the
explanatory variables are based primarily on 1990 census data (United
States Department of Commerce, 1992). They include (1) residential instabil-
ity, defined as the proportion of households occupied by persons who had
moved from another dwelling in the previous 5 years; (2) ethnic heterogen-
eity, computed as the index of diversity (Warner and Pierce, 1993), based
on the proportion of households occupied by white versus nonwhite per-
sons; (3) family disruption, indexed by female-headed households, expressed
as a proportion of all households with children; (4) poverty, defined as the
proportion of persons living below the poverty level; (5) the unemployment
rate (coded as proportion of the workforce); (6) proximity to metropolitan
counties, as indicated by a dummy variable with 1 being adjacent to a metro-
politan statistical area and 0 being nonadjacent.2 Also included in the analy-
sis was the number of youth 10 to 17 years of age, which is the population
at risk for juvenile arrests. Because states may differ in their statutes and in
the organization, funding, and policies of their justice systems, it was
important to eliminate from our analysis all variation between states and

2
To ensure that single cases did not have undue influence on our results, we recoded some
extreme values to values less deviant from the distribution as a whole. We set the maximum
for residential stability to 0.6 (formerly 0.76; three cases recoded), that for female-headed
households to 0.35 (formerly 0.42, four cases recoded), and that for unemployment to 0.12
(formerly 0.14; three cases recoded). This recoding had no substantive impact on the results,
and it increases our faith in their reliability.
Poisson for Aggregate Data 31

Table I. Descriptive Statistics

Mean SD
Robbery arrest rate per 100,000 per year 25.28 48.65
Population at risk in personyyears 11,346 10,776
Log population at risk 8.89 1.04
Residential instability 0.39 0.07
Ethnic heterogeneity 0.26 0.20
Female-headed households 0.18 0.08
Poverty rate 0.16 0.06
Unemployment 5.64 2.60
Adjacent to metropolitan area 0.45 0.50
N of counties 264

assess only within-state relationships pooled across the states. Therefore the
model includes dummy variables representing states (with Florida serving
as the omitted reference category).

3.1. The Distribution of Crime Rates

Table I presents descriptive statistics for all measures. During this
5-year period, there were 1212 arrests of juveniles for robbery in this sample
of counties, which corresponds to an annual arrest rate of 40.5 per 100,000,
or one arrest in 5 years for every 494 juveniles. The distribution of arrest
rates is highly skewed, with zero robbery arrests of juveniles recorded in
52% of the counties, while the highest annual arrest rates were 338 and 390
per 100,000. Counties with smaller populations tended to have lower arrest
rates, so the mean of robbery arrest rates across counties, 25.3, is lower
than the overall arrest rate. There were zero arrests in all but one of the 47
counties with the smallest populations (700 or less). The exception is a
county with two arrests in a population of 289 youths, which constitutes
the seventh highest annual arrest rate. With a population this small, even a
single arrest would place this county among the top 12% for arrest rate. It
is clear that the data provide very crude estimates of arrest rates for any
single county with a small juvenile population. Yet the lack of arrests across
many small counties is strong evidence that the per capita robbery rate is
lower in these counties than in counties with larger juvenile populations.

3.2. Ordinary Least-Squares Analysis

To demonstrate the purposes and use of Poisson-based regression
models, I compare five analyses of the same data. The results appear in
Table II. The first is an OLS regression analysis of the computed arrest rates
(per 100,000 per year) for each county. The full model explains 28.4% of
32 Osgood

Table II. Five Statistical Models of Juvenile Arrest Rates for Robbery in Nonmetropolitan
Counties of Four States
Statistical method
OLS, OLS, OLS, Basic Negative
Explanatory variable ratey100,000 log(rateC1) log(rateC0.2) Poisson binomial
Log population at risk
b 11.220 0.749 1.102 1.501a 1.718a
SE 3.838 0.128 0.177 0.061 0.188
t 2.923 5.852 6.226 8.213 3.819
P 0.004 0.000 0.000 0.000 0.000
Residential instability
b 35.573 3.017 4.366 1.567 0.162
SE 48.790 1.628 2.255 0.567 2.026
t 0.729 1.853 1.936 2.764 0.080
P 0.467 0.065 0.054 0.005 0.936
Ethnic heterogeneity
b 63.839 2.461 3.325 2.069 2.861
SE 32.711 1.091 1.512 0.419 1.156
t 1.952 2.256 2.199 4.938 2.475
P 0.052 0.025 0.029 0.000 0.013
Female-headed households
b 22.765 0.533 0.192 3.919 3.739
SE 71.679 2.391 3.313 1.030 2.937
t 0.318 0.223 0.058 3.805 1.273
P 0.751 0.824 0.954 0.000 0.203
Poverty rate
b 39.474 1.405 2.181 0.499 0.021
SE 81.162 2.708 3.752 1.009 3.381
t 0.486 0.519 0.581 0.495 0.006
P 0.627 0.604 0.561 0.621 0.995
Unemployment
b A42.658 5.246 8.137 A1.338 0.432
SE 203.957 6.804 9.428 1.810 6.568
t A0.209 0.771 0.863 A0.739 0.066
P 0.834 0.441 0.389 0.466 0.948
Adjacent to metropolitan area
b A3.944 A0.211 A0.267 A0.247 A0.458
SE 6.372 0.213 0.295 0.071 0.215
t A0.619 A0.991 A0.905 A3.479 A2.130
P 0.537 0.322 0.365 0.000 0.034
Constant
b A66.020 A6.645 A11.560 A13.750 A15.243
SE 43.732 1.459 2.022 0.630 1.722
t A1.510 A4.554 A5.717 A21.825 A8.852
P 0.132 0.000 0.000 0.000 0.000
Poisson for Aggregate Data 33

Table II. Continued.

Statistical method
OLS, OLS, OLS, Basic Negative
Explanatory variable ratey100,000 log(rateC1) log(rateC0.2) Poisson binomial

Model fit
Method specific criteria
Baseline model b
MSE c 1853.9 2.419 4.729 α e 1.263
R2 0.226 0.328 0.332 0.484 f 0.456 f
A2LLd 1584.5 950.8
Full model
MSE 1760.6 1.960 3.762 α 0.852
R2 0.284 0.471 0.483 0.585 0.548
A2LL 1420.9 901.5
Spearman r 0.653 0.708 0.710 0.671 0.687

Note: The models also included dummy variables representing differences between the four states.
a
t and P values computed for difference of b from 1 rather than difference of b from 0.
b
The baseline model controls for differences between states and, in the Poisson and negative
binomial models, includes log population at risk, with a fixed coefficient of 1.
c
Mean squared error for the OLS regression models.
d
A2 times the log likelihood for the Poisson and negative binomial models.
e
Reflects residual variance in true crime rates, which is overdispersion beyond that expected
from a simple Poisson process.
f
See footnote 5 for a description of the computation of R2 values for the Poisson and negative
binomial analyses.

the variance in these robbery rates, which is 5.9% more than a baseline
model that includes only differences between states [F(7,253)G2.97, PG
0.005].
There are several indications that this OLS model is very poorly suited
to the data. Though an arrest rate below zero would be impossible, this
model yielded negative fitted values for 42 cases, and these negative values
fall as much as 0.61 standard deviations below zero. Under this OLS model,
the two counties with the highest arrest rates constituted extreme outliers
with standardized residuals of 7.2 and 6.1, both far too large to be accept-
able at any sample size. These are strong indications that the fitted values
do not accurately track actual mean crime rates, so it is clear that a linear
model severely distorts the relationship between these explanatory variables
and county level arrest rates.
The critical assumption for the accuracy of standard errors and signifi-
cance tests in OLS analysis is that the residual variance does not vary
systematically across cases, and White’s test for heteroscedasticity
(McClendon, 1994, pp. 178–181) provides a simple and direct means of test-
ing this assumption. This test involves an OLS regression analysis in which
34 Osgood

the squared values of the residuals serve as the dependent variables, and the
fitted values of that regression will reflect mean levels of squared residuals.
The independent variables in this residual analysis can be any factors sus-
pected to be related to heterogeneity of the residuals. Because I expect that
residual variance will depend on population size, but not in a linear fashion,
I chose linear, squared, and cubed terms for population size as independent
variables. Using absolute values of residuals rather than squared residuals
as the dependent variable provided a better summary of the data. (When
squared, residuals of the outliers dominated the entire sample.) This analysis
indicated that the magnitude of residual variance varied widely by popu-
lation size: The squared values of the fitted absolute residuals ranged from
94 to 1162 [R 2 G0.050, F(3,260)G4.51, PG0.004].3

3.3. Ordinary Least-Squares Analysis of Logged Crime Rates

The most drastic shortcomings of the OLS model stem from the highly
skewed distribution of arrest rates. A common strategy for addressing this
problem is to transform the data so that they become less skewed. The
logarithmic transformation is a common choice for this purpose because it
reduces the skew and it also yields a straightforward conceptual interpret-
ation. Under a linear model of the untransformed data, the regression coef-
ficients indicate the difference in the mean of the dependent variable that
is associated with a unit difference on the explanatory variable. After the
logarithmic transformation, the regression coefficients reflect proportional
differences in the mean of the dependent variable, given a 1-unit difference
on the explanatory variable. For crime rates, proportional differences would
appear more plausible than constant differences. We do not expect a factor
that raises a crime rate from 40 per 100,000 to 60 per 100,000 to also raise
a crime of 1 per 100,000 to 21 per 100,000, as would be the case for a linear
model of the untransformed data. Under the proportional model produced
by the logarithmic transformation, the same percentage increase would hold
for both, such as 40 versus 60 and 1 versus 1.5.
The third column in Table II presents an OLS regression analysis with
the natural logarithm of the arrest rate as the dependent variable. One has
been added to the rates (per 100,000) before taking the logarithm because
the logarithm of zero is undefined (corresponding to minus infinity). The
OLS analysis of logged rates is a far better match to the data in several
respects. First, in this analysis the full model accounts for 47.1% of the
3
Strictly speaking, White’s test uses the squared residual as the dependent variable and uses a
χ2 significance test due to the likely heteroscedasticity of the residuals themselves. This test
would also be significant for this residual analysis [ χ2(3)G13.2, PG0.004] as well as for the
comparable analysis reported below [ χ2(3)G28.5, PG0.000].
Poisson for Aggregate Data 35

variance, which is a clear indication that the transformation puts the data
in a form that has a closer linear correspondence to these explanatory vari-
ables. Also, in this altered metric a larger share of the explained variance is
attributable to the explanatory variables rather than to differences between
states [increase in R 2 G0.142, F(7,253)G9.761, PF0.001]. Second, the
range of the fitted values is not problematic under this model because nega-
tive fitted values correspond to logarithms of crime rates between zero and
one. Third, the transformation also reduces problems of outliers, with the
most extreme standardized residual for the OLS analysis of the transformed
arrest rates now 3.2. The change in metric means that the coefficients of
these first two analyses are not comparable, but the benefit of the improved
correspondence between model and data is apparent in the higher t values
for the three variables most strongly related to crime rates (population size,
residential instability, and ethnic heterogeneity).
Though the logarithmic transformation renders the data more suitable
for OLS analysis, it also makes apparent the inherent problems that require
Poisson-based regression. First, rather than solving the problem of hetero-
scedasticity, the error variance has become even more strongly related to
population size. The cubic model of the absolute residuals now acounts
for 10.8% of their variance [F(3,260)G10.52, PF0.001], with fitted values
corresponding to squared residuals that range from 0.01 to 1.98. Further-
more, the specific cases that constitute outliers also have changed. In the
OLS analysis of untransformed rates, the two most extreme outliers were
the counties with the highest crime rates, both of which have larger than
median juvenile populations. The most extreme outlier in the OLS analysis
of the transformed rates is the smallest population with any recorded
arrests. The OLS assumption of homogeneity of residual variance implies
that the predictive accuracy of the model is independent of population size.
As we would expect from the inevitable unreliability of crime rate estimates
based on small population sizes, that assumption is clearly in error.
Second, the discrete and skewed nature of crime rates for small popu-
lations presents a special problem for analyzing log transformed crime rates.
Observed rates of zero will be common for small populations, in which
case the transformation can be computed only after adding a constant. The
common choice of adding one is highly arbitrary. The value could as easily
be 1 per 1000 or 1 per 1,000,000 as the 1 per 100,000 used in the analysis
just discussed, yet the choice of this constant may drastically effect the
results. To see this compare Columns 3 and 4 in Table II, which differ only
in that Column 3 reports an analysis resulting from adding a constant of 1
per 100,000 while the constant for Column 4 is 0.2 (corresponding to 1
arrest per 100,000 for the 5 years covered in the study, rather than for one
year). This arbitrary choice results in an increase of roughly 40% in most of
36 Osgood

the regression coefficients, which is a large difference in the implied effects

of these explanatory variables on mean crime rates. The reason for this
change is that decreasing the constant increases the variance of the depen-
dent variable by inflating the difference between the transformed value for
the rate of zero and the transformed value for the next higher observed rate.
Thus, the choice of this constant has great potential for biasing the coef-
ficient estimates. Interestingly, changing the additive constant had minimal
consequence for significance testing because standard errors grew pro-
portionately with the coefficients, with the result that t values were essen-
tially unchanged.

3.4. Poisson-Based Analyses

Poisson-based regression analyses successfully address the most serious
problems that arise in the OLS analyses. As discussed above, Poisson-based
models do not assume homogeneity of variance. Instead, residual variance
is expected to be a function of the predicted number of offenses, which is in
turn a function of population size. Furthermore, even though a logarithmic
transformation is inherent in Poisson-based regression, observed crime rates
of zero present no problem. Unlike the preceding OLS analyses of log crime
rates, Poisson-based regression analyses do not require taking the logarithm
of the dependent variable. Instead, estimation for these models involves
computing the probability of the observed count of offenses, based on the
fitted value for the mean count. As Figs. 1 and 2 demonstrate, observed
rates of zero become increasingly likely as the estimated mean rate
approaches zero.
The last two columns in Table II present results from a basic Poisson
regression and a negative binomial regression, both estimated with the
LIMDEP statistical package (Greene, 1995). Because these are maximum-
likelihood estimates, likelihood-ratio significance tests can be used to deter-
mine whether more complex models provide better fit to the data than do
simpler models. The test statistic is minus twice the difference between the
log likelihoods for the models, and the significance level is determined by
comparing this value to the χ2 distribution with degrees of freedom equal
to the number of additional parameters in the more complex model.

3.4.1. Poisson versus Negative Binomial

The negative binomial model differs from the basic Poisson by the
addition of the residual variance parameter, α . The likelihood-ratio test
value for the comparison between these models is 519.4 with 1 degree of
freedom, indicating that the data are far more consistent with the negative
binomial model than with the basic Poisson (PF0.001). This result means
Poisson for Aggregate Data 37

that, on average, differences between fitted and observed crime rates are
considerably larger than specified by the Poisson distribution. This overdis-
persion reflects some combination of unexplained variation in counties’ true
crime rates and positive dependence among crime events.
Comparing the fifth and sixth columns in Table II makes clear the
consequence of ignoring that overdispersion. The basic Poisson model gives
the impression of far greater precision in the estimated relationships than is
justified. The standard errors for the basic Poisson model average only
about one-third the size of the standard errors for the negative binomial, so
the basic Poisson model would produce highly erroneous significance tests
for the coefficients. In this example, the basic Poisson would lead us to
conclude that residential instability and female-headed households are sig-
nificantly related to rates of arrests of juveniles for robbery, while the nega-
tive binomial would not.

3.4.2. Fit of the Negative Binomial Model

I examined residuals from the negative binomial model to check the
match between the model and data in terms of potential outliers. Stan-
dardized residuals are less useful here because the residuals are not expected
to follow a normal distribution. Instead, an appropriate strategy is to use
the negative binomial distribution of Eq. (4) to compute the probability of
obtaining a value at least as extreme as the observed value, based on the
fitted values, λ i , and the estimate of residual variance in true crime rates,
α . Because this is a tedious calculation, even using a spreadsheet program,
I first computed standardized residuals4 to identify the cases most likely to
constitute outliers, and then computed probabilities for those cases. The
most extreme outlier was a county with three recorded arrests. Its fitted
value, λ , was 0.204, which yields a probability of 0.0043 of observing three
or more arrests. Only 1 in every 233 cases should have a probability this
small, but that is quite acceptable in our sample of 264. Probabilities for
four cases were less than 0.02, which is no more than would be expected by
chance in a sample of this size. It is especially encouraging that these four
counties varied widely in population size, from the eighth to the eightieth
percentile. Thus, there are good indications that the assumptions of the
negative binomial model are a good match to these data and that this model

4
Standardized residuals can be computed on the basis of the variance of the negative binomial,
which is λ iCαλ 2i (Gardner et al., 1995). Though the Poisson rapidly approaches the normal
distribution as λ increases, this is far less true of the negative binomial with the moderate
value of α obtained in this example. For instance, even with a value of λ as large as 32,
standardized residuals could differ from the normal deviates corresponding to negative
binomial probabilities by over 50%.
38 Osgood

successfully addresses the confounding of population size and accuracy of

crime rate estimates.
How well does the negative binomial model account for crime rates, in
comparison to the OLS analyses? The R 2 values in Table II are not very
helpful for this purpose because they reflect scaling differences in the depen-
dent variable as much as differences in the success of the models. Thus, the
higher R 2 values for the Poisson and negative binomial models are partly a
reflection that the outcome variable in these analyses is not the crime rate,
but rather the count of offenses. The Spearman rank order correlation,
which is unaffected by scaling, indicates that the five models have similar
success in ordering the counties by crime rate (values range from 0.66 to
0.71). When taking the metric of the outcome measure into account, we find
that the negative binomial model explains substantial portions of variance
in both untransformed crime rates (22.2% for negative binomial versus
28.4% for the OLS analysis of these rates) and log-transformed crime rates
(22.8% for negative binomial versus 47.1% for this OLS analysis).5 In sharp
contrast, each of the OLS analyses is surprisingly unsuccessful in accounting
for variance in the other metric: Fitted values from the OLS analysis of log-
transformed rates account for only 8.8% of the variance in untransformed
rates, while fitted values from OLS analysis of untransformed rates account
for only 5.2% of the variance in log-transformed rates. It is likely that these
mismatches reflect the shortcomings of both OLS approaches. A direct lin-
ear model is unsuitable for the untransformed rates, as evidenced in the
impossible negative fitted values that result. Yet the OLS analysis of log-
transformed rates yields fitted values that poorly match the original crime
rates, despite a strong rank order correlation between the two. It appears
that the arbitrary constant required for OLS analysis of log-transformed
rates results in a seriously distorted metric. Poisson-based models avoid
these problems and, as a result, generalize well to either response metric.

3.4.3. Results from the Negative Binomial Model

The likelihood-ratio test can be used to assess the overall contribution
of the explanatory variables, comparable to the test for increase in R 2 in
5
I computed these percentages of explained variance by transforming fitted values from each
model to the metric of the observed scores, summing the squared differences between these
and the observed scores, and calculating one minus the quotient of that sum divided by the
total sum of squares for the observed scores. This computation corresponds to the definition
of R 2 in OLS regression. Because R 2 is not part of the maximum-likelihood estimation of the
basic Poisson and negative binomial models, this procedure was also used to compute the R 2
values for those models. Negative fitted values from the OLS analysis of untransformed crime
rate were set to zero in order to compute the fit of that model to log-transformed crime rates.
The additive constant of 1 per 100,000 was used for all log transformations of crime rates.
Poisson for Aggregate Data 39

OLS models. For the Poisson-based analyses, the baseline model includes
not only dummy variables to control for differences between states, but also
the log of the population at risk with a fixed coefficient of one, as in Eq.
(3). This control for population size is necessary so that the regression will
be a model of per capita crime rates rather than a model of counts of crimes.
For the negative binomial, the full model yields a likelihood-ratio value of
49.4 in comparison to the baseline model, which is statistically significant
(dfG7, PF0.001). Thus, we can conclude that the explanatory variables
account for more variation in crime rates than would be expected by chance
alone.
By the conventional 0.05 standard of statistical significance, the nega-
tive binomial analysis indicates that higher juvenile arrest rates for robbery
are associated with larger populations at risk, greater ethnic heterogeneity,
and being adjacent to a metropolitan area.6 To interpret the regression coef-
ficients for these variables, we must take into account the logarithmic trans-
formation that intervenes between the linear model and fitted crime rates
[in Eqs. (1) and (3)]. Liao (1994) explains several useful strategies for inter-
preting these coefficients. One relatively straightforward approach to this
task follows the implication of Eqs. (1) and (3) that an increase of x in an
explanatory variable will multiply the fitted mean crime rate by the exp(bx).
Thus, given the coefficient of 2.861 for ethnic heterogeneity in the negative
binomial model, a 10% increase in ethnic heterogeneity would multiply the
rate by exp(0.286), which is 1.33. In plain English, a 10% increase in ethnic
heterogeneity is associated with a 33% increase in the juvenile arrest rate
for robbery. Because being adjacent to a metropolitan area is coded as a
dummy variable, an increase of one in this variable corresponds to the con-
trast between adjacent and nonadjacent counties. Thus, the statistically sig-
nificant coefficient of A0.458 indicates that counties adjacent to
metropolitan areas have a 37% lower rate of robbery than those that are
not because exp(−0.458 ∗ 1) equals 0.63. [This surprising result does not rep-
licate for analyses of other offenses reported by Osgood and Chambers
(2000).]
In interpreting the results for population size, we must take into
account the special role of this variable in Poisson-based analyses of aggre-
gate rates. When the coefficient for the log of the population at risk is fixed
at one [as in Eq. (3)], per capita crime rates are constant across counties
with different population sizes, controlling for the other explanatory vari-
ables. The analyses reported in Table II treat that coefficient as estimated

6
In the more extensive analyses reported by Osgood and Chambers (2000) population at risk,
ethnic heterogeneity, residential instability, and female-headed households proved to be
associated with most offenses, but adjacency to metropolitan areas did not.
40 Osgood

rather than fixed, which allows for the possibility that crime rates differ with
population size. In this case, however, it is necessary to subtract the value
of one from this coefficient in order to determine its implications for the
relationship of population size to per capita crime rates. Similarly, the stat-
istical significance of the relationship is gauged by comparing the estimate
to the value of one, rather than to zero as is the usual case.7 A coefficient
greater than one would indicate that counties with larger populations have
higher per capita crime rates, while a coefficient less than one would indicate
the opposite. Thus, the coefficient of 1.718 from the negative binomial
analysis agrees quite closely with the value of 0.749 from the OLS analysis
of the transformed crime rates. The first indicates that a doubling of the
population is associated with a 64% increase in per capita robbery rates
[exp(0.718 ∗ log(2))G1.645], while the second implies a 68% increase
[exp(0.749 ∗ log(2))G1.680].
I have argued that the coefficients and significance tests based on the
negative binomial (or another Poisson-based regression model that allows
for overdispersion) are preferable because the other models I have reviewed
rely on assumptions that are inconsistent with the data. Yet how much
difference does the choice of model make? We can get some idea by compar-
ing the coefficients and t values for the negative binomial analysis with those
for the other analyses in Table II. Other than the OLS analysis of untrans-
formed rates, all models specify a logarithmic relationship between fitted
values and mean crime rates, so coefficients have comparable meanings
across those models. In general, one would not expect an incorrect model
to introduce any systematic bias, so it is surprising that estimates for many
of the coefficients differ dramatically across the models. The absolute values
of coefficients for residential instability, poverty rate, and unemployment
are far larger in the OLS analyses than in the negative binomial analysis,
while the opposite is true for female-headed households. Differences of this
sort most likely are due to the role of population size in Poisson-based
analyses. OLS analyses place as much weight on small counties as on large
ones, but Poisson-based regression models expect error distributions in
small counties to have greater variance, with the consequence that results
are less influenced by small counties. This differential weighting has con-
siderable potential for changing results in a sample such as ours, where there
is a large range of population sizes.
The standard errors for the negative binomial model are most similar
to those of the OLS analysis of log transformed rates, using the additive
constant of one. Even here, however, standard errors for four of the seven

7
In other words, the test statistic to be compared to the normal distribution is not the usual
bySEb , but rather (bA1)ySEb .
Poisson for Aggregate Data 41

substantive variables are at least 20% larger in the negative binomial analy-
sis. There are far greater discrepancies in standard errors for the other
models, so it is clear that significance tests may be seriously affected by
applying an appropriate statistical model to aggregate data for small
populations.

4. CONCLUSIONS
Using Poisson-based regression models of offense counts to analyze per
capita offense rates is an important advance for research on aggregate crime
data. Standard analytical approaches require that data be highly aggregated
across either offense types or population units. Otherwise offense counts are
too small to generate per capita rates that have appropriate distributions
and sufficient accuracy to justify least-squares analysis. Poisson-based
regression models give researchers an appropriate means for more fine-
grained analysis. Poisson-based models are built on the assumption that the
underlying data take the form of nonnegative integer counts of events. This
is the case for crime rates, which are computed as offense counts divided by
population size. In our example analysis of juvenile arrest rates for robbery,
the Poisson-based negative binomial model provides a very good fit to the
data, while OLS analyses produce outliers and require arbitrary choices that
have a striking impact on results.
Poisson-based regression models free researchers to investigate a much
broader range of aggregate data because they are appropriate for smaller
population units and less common offenses. Yet these models are not magic.
The reason they are appropriate is that they recognize the limited amount
of information in small offense counts. The price one must pay in this trade-
off is that the smaller the offense counts, the larger the sample of aggregate
units needed to achieve adequate statistical power. For example, this sample
of 264 counties proved too small for a meaningful analysis of juvenile
homicide, the least common offense examined in this study (Osgood and
Chambers, 2000).
Though this article has concentrated on two of the most common
Poisson-based regression models, this approach to analyzing aggregate
crime rates can be implemented with virtually any of the Poisson-based
regression analyses. The numerous Poisson-based models reviewed by
Cameron and Trevedi (1998) offer many choices for finding a model with
assumptions that best match one’s data. Some models expand the range of
research questions that can be addressed, such as using finite-mixture mod-
els to identify homogeneous groups of counties. Other Poisson-based mod-
els have been developed for designs with repeated measures or nested data,
such as counties nested within states or multiple subpopulations nested
42 Osgood

within a sample of geographic areas. The semiparametric model of Nagin

and Land (1993), which has been so influential in research on criminal
careers, would be appropriate for such cases. Also, the recent version of
Bryk and co-workers’ (1996) HLM program implements a Poisson version
of their hierarchical linear modeling approach to analyzing nested data
(Bryk and Raudenbush, 1992). Thus, Poisson-based regression models
should have broad applicability for the study of crime at the aggregate level.

ACKNOWLEDGMENTS
This research was supported by Grant 94-JN-CX-0005 from the Office
of Juvenile Justice and Delinquency Prevention, Office of Justice Programs,
U.S. Department of Justice. The author thanks Jeff Chambers for his assist-
ance with this study, Chet Britt for comments on an early draft, and Gary
Melton and Susan Limber for their support of the entire project. Points of
view or opinions in this document are those of the authors and do not
necessarily represent the official position or policies of the U.S. Department
of Justice.

REFERENCES
Bailey, A. J., Sargent, J. D., Goodman, D. C., Freeman, J., and Brown, M. J. (1994). Poisoned
landscapes: The epidemiology of environmental lead exposure in Massachusetts. Soc. Sci.
Med. 39: 757–766.
Bryk, A. S., and Raudenbush, S. W. (1992). Hierarchical Linear Models: Applications and Data
Analysis Methods, Sage, Newbury Park, CA.
Bryk, A. S., Raudenbush, S. W., and Congdon, R. (1996). HLM: Hierarchical Linear and
Nonlinear Modeling with the HLM/2L and HLM/3L Programs, Scientific Software Inter-
national, Chicago.
Cameron, A. C., and Trivedi, P. K. (1998). Regression Analysis of Count Data, Cambridge
University Press, Cambridge.
Gardner, W., Mulvey, E. P., and Shaw, E. C. (1995). Regression analyses of counts and rates:
Poisson, overdispersed Poisson, and negative binomial. Psychol. Bull. 118: 392–405.
Greenberg, D. F. (1991). Modeling criminal careers. Criminology 29: 17–46.
Greene, W. H. (1995). LIMDEP: Version 7.0 Users Manual, Econometric Software, Plainview,
NY.
King, G. (1989). Unifying Political Methodology: The Likelihood Theory of Statistical Inference,
Cambridge University Press, Cambridge.
Liao, T. F. (1994). Interpreting Probability Models: Logit, Probit, and Other Generalized Linear
Models, Sage University Paper Series on Quantitative Applications in the Social Sciences,
07–101, Sage, Newbury Park, CA.
Maltz, M. D. (1994). Operations research in studying crime and justice: Its history and accom-
plishments. In Pollock, S. M., Rothkopf, M. H., and Barnett, A. (eds.), Operations
Research and the Public Sector, Volume 6 of Handbooks in Operations Research and
Management Science, North-Holland, Amsterdam, pp. 200–262.
Poisson for Aggregate Data 43

McClendon, McK. J. (1994). Multiple Regression and Causal Analysis, F. E. Peacock, Itasca,
IL.
McCullagh, P., and Nelder, J. A. (1989). Generalized Linear Models, 2nd ed., Chapman and
Hall, London.
Nagin, D. S., and Land, K. C. (1993). Age, criminal careers, and population heterogeneity:
Specification and estimation of a nonparametric, mixed Poisson model. Criminology 31:
327–362.
Osgood, D. W., and Chambers, J. M. (2000). Social disorganization outside the metropolis:
An analysis of rural youth violence. Criminology 38: 81–115.
Rowe, D. C., Osgood, D. W., and Nicewander, W. A. (1990). A latent trait approach to
unifying criminal careers. Criminology 28: 237–270.
Sampson, R. J., Raudenbush, S. W., and Earls, F. (1997). Neighborhoods and violent crime:
A multilevel study of collective efficacy. Science 177: 918–924.
United States Department of Commerce (1992). Summary Tape Files 1 and 3, 1990 Census.
Warner, B. D., and Pierce, G. L. (1993). Reexamining social disorganization theory using calls
to the police as a measure of crime. Criminology 31: 493–517.