This is a preprint of an article published in Addiction © 2000 Carfax Publishing.
Fergusson DM, Horwood LJ. Alcohol abuse and crime: A fixed effects regression analysis.
Addiction, 2000; 95(10): 1525-1536.
Alcohol Abuse and Crime: A Fixed Effects Regression Analysis
DAVID M FERGUSSON & L JOHN HORWOOD
Christchurch Health and Development Study,
Christchurch School of Medicine,
Christchurch, New Zealand
Short title: Alcohol Abuse and Crime
Correspondence to:Professor David Fergusson
Christchurch Health and Development Study
Christchurch School of Medicine
P O Box 4345
Christchurch, New Zealand
Alcohol Abuse and Crime: A Fixed Effects Regression Analysis
D M Fergusson & L J Horwood
Aims: To examine linkages between patterns of alcohol abuse and crime in a New Zealand birth
cohort studied to the age of 21 taking into account confounding factors through the use of fixed
effects regression methods.
Measurements: Over the period from age 15 to 21 years assessments were made of: a)
involvement in violent and property crime; and b) extent (if any) of alcohol abuse/dependence
symptoms. In addition, information was gathered on a number of social and contextual factors that
were likely to be related to alcohol abuse or crime.
Findings: Increasing alcohol abuse was associated with clear and significant (p<.0001) increases
in rates of both violent and property crime. Control for observed and non observed confounding
through the use of fixed effects regression models indicated that much of this association was
attributable to the effects of confounding factors that were associated with both alcohol abuse and
crime. Nonetheless, even after such control alcohol abuse remained significantly related to both
violent and property offending.
Conclusions: The findings suggest the presence of a possible causal association between alcohol
abuse and juvenile offending that is evident after control for both observed and non observed
sources of confounding.
It is widely believed that the use and abuse of alcohol by young people is a factor that
contributes to criminal behaviours (Mott, 1990; Greenfield & Weisner, 1995). In particular, it may
be proposed that the disinhibiting effects of heavy alcohol consumption may increase tendencies to
risk taking, antisocial and violent behaviour, placing young people who engage in heavy alcohol
consumption at increased risk of criminal behaviours (Collins, 1981). These beliefs have generally
been supported by research studies that have variously shown that: (a) amongst those who offend,
rates of alcohol abuse are elevated; or (b) amongst those who abuse alcohol, rates of criminal
offending are elevated (Swanson et al., 1990; Pernanen 1991; Milgram, 1993; Fergusson Lynskey
& Horwood, 1994; Shepherd, 1994; Greenfield & Weisner 1995; Fergusson, Lynskey & Horwood,
1996; Kerner et al., 1997). Findings from correlational studies have also been supported by
laboratory based research which has shown in controlled settings that antisocial behaviours, and
particularly aggression, tend to increase with increasing alcohol consumption (Bushman & Cooper,
1990; Taylor & Chermack, 1993). Although there is little doubt that consistent linkages between
alcohol consumption patterns and antisocial behaviours can be demonstrated, the extent to which
these linkages reflect causal processes in which the heavy consumption of alcohol leads to an
increased risk of offending remains controversial (Fergusson, et al., 1996; Wagner, 1996).
A major threat to the validity of this conclusion comes from the possibility that linkages
between alcohol consumption patterns and crime may reflect the presence of third or confounding
factors that are associated with both criminality and increased risks of the heavy use of alcohol.
There is good prior evidence to support this view to the extent that the literature on crime and
alcohol use has identified a similar set of risk factors that are related to both outcomes. In general,
rates of alcohol use and/or crime have been found to be increased amongst young people exposed to
social disadvantage, family dysfunction, parental deviance and impaired childhood environments
and amongst young people characterized by the early onset of behavioural problems (Farrington et
al, 1990; Hawkins, Catalano & Miller, 1992; Farrington and Loeber, 1999). Given this evidence, it
may be argued plausibly that the linkages between crime and alcohol use arise because the risk
factors and life processes that encourage the heavy use of alcohol also independently increase risks
of involvement in crime. The extent to which linkages between alcohol use and crime can be
explained by common confounding factors has been explored in a number of studies in which these
associations have been adjusted for the effects of such factors (Swanson et al., 1990; Ensor &
Godfrey, 1993; White, Brick & Hansell, 1993; Otero-Lopez et al., 1994; Greenfield & Weisner,
1995; Fergusson, et al., 1996). In general, the results of these analyses have suggested that
although control for confounding reduces linkages between crime and alcohol use, even after such
control associations between alcohol use and crime persist. For example, in an analysis of a birth
cohort of New Zealand subjects studied to the age of 16, Fergusson, et al. (1996) found that young
people classified as exhibiting abusive or hazardous use of alcohol had odds of property and violent
crimes that were in the region of 5-6 times higher than those not meeting these diagnostic criteria.
However, much of this association was explained by common confounding factors and after
adjustment for these factors, the odds ratios between alcohol abuse and crime ranged from 1.4 to
3.2. The authors concluded that a large amount of the association between alcohol abuse and crime
arose because of the common effects of shared risk factors but that there was nonetheless evidence
of a specific linkage between alcohol abuse and crime which was particularly evident for crimes of
However, all of these studies face a common difficulty in that associations between alcohol
use and crime have been adjusted for observed confounding factors and the possibility remains that
these associations may be accounted for by non observed sources of confounding that have not been
adequately represented in the analysis. For example, one potential source of uncontrolled
confounding arises from common genetic factors that may be associated with increased risks of
both alcohol abuse and crime. For these reasons, it is likely that studies that estimate the
association between alcohol use and crime by adjusting for observed covariate factors may lead to
an overestimation of the direct causal contribution of alcohol use to crime as a result of the under
control of confounding influences arising from omitted covariate factors (Greenfield & Weisner,
In this paper, we develop and apply a method of analysis that permits the estimation of the
effects of alcohol use on crime, taking into account both observed and non observed sources of
confounding, through the use of the fixed effects regression model. This method is applied to data
gathered on alcohol abuse and crime in a birth cohort of over 1,000 New Zealand young people
studied from age 15 to 21. The background to the analysis approach is developed below.
The Application of the Fixed Effects Regression Model
Consider a situation in which a cohort of N subjects is observed on T occasions on measures
of crime and alcohol use. Let Yit denote the measure of crime for the ith subject observed at time t
and Xit denote the corresponding measure of alcohol use. Assume also that the cohort has been
assessed on, or before, the first point of observation on a series of covariate factors Z1....ZJ and let
Zij denote the response of the ith subject on the jth covariate. Given this specification, we seek to
model the relationship between the response variable Yit and the exposure variable Xit taking into
account the observed covariates Zij and other non observed sources of confounding that could
influence the relationship between Yit and Xit.
One model that could be used to represent this state of affairs is:
Yit = B0 + B1Xit + ΣBj Zij + ui + eit EQ 1
where the coefficients B1, Bj are the regression coefficients linking the observed variables Xit, Zij to
the response variable Yit. The disturbance of this equation involves the sum of two components ui +
eit. The first term ui, represents the individual specific component of Yit that is not explained by the
model. This individual specific term represents unobserved sources of variation that influence Yit
and are specific to individual i and are assumed to remain constant over time. These sources of
variation will include the effects of any non observed covariate factors that influence the response
of individual i. The second part of the disturbance is the random error term eit. It is clear that the
estimation of the above equation is straightforward if it is assumed that the disturbance terms ui, eit
are uncorrelated with the variables Xit and Zij. In these circumstances, B1 provides an estimate of
the effect of Xit on the outcome Yit. However, in the situation in which ui is correlated with Xit the
estimate of B1 will be biased by a failure to take account of this correlation (Judge et al., 1980;
Greene, 1993). Such bias could, for example, occur if variations in ui largely represented the
consequences of uncontrolled covariate factors that were correlated with Xit. The analytic
challenge is, thus, to estimate the parameters of the model in a way that takes account of the
possibility of a correlation between the individual specific term ui and the observed explanatory
variables Xit, Zij. (By assumption, the random error term eit is uncorrelated with Xit; Zij). This may
be achieved by the application of a fixed effects regression model. The logic of the estimation of
this model is outlined below. A more detailed account of the derivation and application of fixed
effects regression models may be found in the econometric literature (for example Judge et al.,
1980; Chapter 8).
Consider a situation in which all terms in equation 1 are summed over all (T) points of
observation for each subject i and the mean response for each subject is computed. Applying this
approach to equation 1 gives:
yi = B0 + B1 xi + ΣBj Zij +ui + ei EQ 2
where yi is the mean response value for the ith subject on Y; xi is the mean response for the ith
subject on X and the covariates Zij remain unchanged. As noted above, ui is the individual specific
response and ei is the mean of the random error term eit over the (T) time periods.
Subtracting EQ 2 from EQ 1 gives:
(Yit - yi) = B1 (Xit - xi) + (eit - ei) EQ 3
Inspection of equation 3 shows that it has two important properties. First, it leads to an
estimation of the regression parameter of interest (B1). Second, the process of differencing has
eliminated from the equation all observed (Zij) and non observed (ui) sources of confounding that
could influence the estimation of B1. This latter property implies that the estimate of B1 obtained
from solving equation 3 yields an estimator of the effects of Xit on Yit that is corrected for both the
observed sources of confounding Zij and other sources of confounding that were contained within
the individual specific term ui.
Although the above describes the basic form of the fixed effects regression model, to provide
the necessary background for later analysis, it becomes useful to consider some extensions of this
model. As can be seen from the above, the model assumes that the confounding influences reflect
fixed variables that were assessed at or before the first point of observation. However, the use of
fixed variables may not represent all sources of confounding since it is possible that time dynamic
factors that are concurrent with the assessment of Yit and Xit may confound the relationship
between these variables. In the present context one such time dynamic confounding factor is peer
affiliations. It could be argued that linkages between crime and alcohol use arise because of the
common effects of peer group affiliation on both outcomes. Under these circumstances, failure to
take account of peer affiliations over time could lead to the association between alcohol use and
crime being mis-estimated. To address such issues the fixed effects model can be extended to
include observed time dynamic covariates. In this form, the model controls for the effects of all
(observed and non observed) fixed covariates and for the effects of time dynamic covariate factors.
A complication that arises with fitting the fixed effects model to data on crime rates is that the
distribution of offences in the population is markedly non normal raising issues about the suitability
of such data for analysis using ordinary least squares regression methods. These issues have been
resolved by extension of the fixed effects model to fit this model to logistic, Poisson and negative
binomial distributions (Hausman, Hall & Grilliches, 1984; Hamerle and Ronning, 1995). The
negative binomial form of the model is likely to be particularly applicable to offending data as it
allows offending rates to be modelled as an over dispersed Poisson distribution (Hausman et al.,
1984). As we will show later, this model also makes it possible to interpret fitted model parameters
as rate ratios.
Against this general background the present paper reports on a longitudinal study of the
relationships between symptoms of alcohol abuse and rates of crime in a birth cohort of over 1,000
New Zealand young people studied over the period from age 15 to 21 years. The aims of this
1. To document the association between symptoms of alcohol abuse and rates of property
and violent crime in the cohort.
2. To estimate the associations between symptoms of alcohol abuse and rates of crime taking
into account confounding factors using both a fixed effects regression approach and a conventional
regression adjustment for observed covariates.
In general, the aims of this analysis are to describe the linkages between alcohol abuse and
crime and to explore the extent to which these linkages may arise from both observed and non
observed sources of confounding.
The data described in this report were gathered during the course of the Christchurch Health
and Development Study (CHDS). The CHDS is a longitudinal study of an unselected birth cohort of
1265 children who were born in the Christchurch (New Zealand) urban region in mid 1977. This
cohort has been studied at birth, 4 months, 1 year, annual intervals to age 16, at 18 and 21 years. An
overview of the study design and methodology has been given in previous papers (Fergusson et al.,
1989; Fergusson et al., 1996; Fergusson and Lynskey, 1998). The present analysis is based on 1063
subjects for whom information on alcohol use and offending was available on at least one
assessment between age 15 and 21. This sample represented 84% of the original cohort of 1265
subjects. However, since not all subjects were observed at all occasions, the numbers with non-
missing data in a given year range from 953 to 1025. These variations in sample size are shown in
Table 1 of the Results. The following measures were used in the analysis.
At ages 15, 16, 18 and 21 years, subjects were questioned concerning their use of alcohol in
the preceding 12 months and their experience of problems related to alcohol use (Fergusson et al.,
1996). At ages 15 and 16 alcohol related problems were assessed using the Rutgers Alcohol
Problem Index (RAPI, White & Labouvie, 1989), whereas at ages 18 and 21 questioning was based
on the Composite International Diagnostic Interview (CIDI, World Health Organization, 1993)
items relating to alcohol abuse/dependence. On the basis of this information, subjects were
assessed on the extent to which they met standardized diagnostic criteria for alcohol abuse or
dependence using DSM-III-R (American Psychiatric Association, 1987) criteria at ages 15 and 16
years and DSM-IV (American Psychiatric Association, 1994) criteria at ages 18 and 21. At each
point of observation a scale score was constructed based on the number of symptom criteria for
alcohol abuse or alcohol dependence that the subject met during the preceding 12 month period,
with this score ranging from 0 for those meeting no criteria to a maximum of 11 for those meeting
all criteria. In the interests of brevity, the measure of alcohol abuse/dependence symptoms is
described throughout the text as alcohol abuse since all those meeting criteria for alcohol
dependence criteria also met criteria for alcohol abuse.
At ages 15 and 16 years, self report measures of the frequency of criminal offending over the
preceding 12 months were obtained using the Self Report Early Delinquency Scale (SRED, Moffitt
& Silva, 1988). At 18 and 21 years similar information was obtained using the Self Report
Delinquency Inventory (SRDI, Elliott & Huizinga, 1989). At each age two offence scores were
constructed representing the total number of offences reported by the young person in each of the
following areas: (a) violent offences including assault, fighting, using a weapon, physical coercion
and cruelty to animals; (b) property offences including vandalism, firesetting, shoplifting, burglary
and other theft. Both measures of offending were highly skewed as a result of the fact that the great
majority (between 80% to 95%) of the cohort did not report violent or property offences at each
age. To address the skewed nature of the distribution of the offending measures, these scores were
transformed to natural logarithms for all analyses. However, experimentation with alternative
scalings produced very similar conclusions (see Results).
Prior to age 14 cohort members were assessed on a wide range of social, family and
individual factors that were likely to have bearing on their future risks of alcohol abuse and crime.
These factors were assumed to be fixed at the age of 14 and included: (a) measures of family
socioeconomic background (maternal age, maternal education, family socioeconomic status, family
living standards); (b) measures of family functioning (parental change, parental conflict, frequency
of adverse family life events, exposure to childhood physical or sexual abuse); (c) measures of
parental deviance (parental history of criminality, parental history of alcoholism/problems with
alcohol, parental illicit drug use); (d) individual factors (gender, early onset conduct problems, the
frequency of truancy, the frequency of alcohol use and offending diversity prior to age 14). Despite
the large number of potential fixed covariates that were known to be correlated with both alcohol
abuse and juvenile offending, only five of these measures (gender, maternal age at the subject’s
birth, family living standards, early onset conduct problems and offending diversity prior to age 14)
were shown to make statistically significant contributions in regression analyses involving fixed
covariates. The measures of family living standards, early conduct problems and offending diversity
are described in greater detail below.
Family living standards (0-10 years). At each assessment from age 1 to 10 years, interviewer
ratings of the quality of family living standards were obtained using a five point scale ranging from
1 = ‘very good’ to 5 = ‘very poor’. In the present analysis, these ratings were averaged over the ten
year period to provide an overall measure of the quality of family living standards during this
Early onset conduct problems (8 years). When subjects were aged 8 years, parallel parental
and teacher reports of the child’s tendencies to disruptive or conduct disordered behaviours were
obtained using an instrument that combined items from the Rutter (Rutter, Tizard & Whitmore,
1970) and Conners (1969, 1970) parent and teacher child behaviour questionnaires. Parental and
teacher reports were combined to produce a unidimensional scale representing the extent to which
the child displayed disruptive behaviours (Fergusson et al., 1991; Fergusson & Lynskey, 1998). The
reliability of this scale, assessed using coefficient alpha, was .93.
Offending diversity (12-14 years). At ages 13 and 14 parental and self reports of involvement
in violent or property offending over the past year were obtained using the SRED (Moffitt & Silva,
1988). At these ages offending reports assessed only whether the young person had committed
different types of offence, not the frequency of offending. To provide an overall measure of the
young person’s history of offending prior to age 14, the parental and self report measures were
combined to obtain an offending diversity score representing the total number of offence types
reported by either the subject or his/her parents over the period from 12-14 years.
Time Dynamic Covariates
In addition to the fixed covariate factors described above, time dependent observations were
made of a series of aspects of adolescent behaviours and life style. These time dynamic covariates
Illicit drug abuse. Parallel to the assessment of alcohol abuse, at each age subjects were
assessed on standardized symptom criteria for illicit drug abuse/dependence. For each subject at
each age, a measure of the extent of illicit drug abuse/dependence was derived based on the number
of symptom criteria the subject met for substance abuse or substance dependence over the
preceding year. At ages 15 and 16 years assessment was based on DSM-III-R criteria, and at ages
18 and 21 years on DSM-IV criteria. The symptom scores ranged from 0 for subjects meeting no
symptom criteria to a maximum of 11.
Deviant peer affiliations. At each assessment from age 15 to 21 years, subjects were
questioned on the extent to which their friends used tobacco, alcohol, cannabis or other illicit drugs,
were suspended or truant from school, or were involved in property, violent or other offending.
These items were combined to produce scale score measures representing the extent of the young
person’s reported affiliations with delinquent or substance using peers at each age (Fergusson and
Horwood, 1996, 1999).
Adverse life events. At each assessment from age 15 to 21 years, subjects were questioned
concerning their exposure to adverse or stressful life events over the preceding year using an
instrument based on the Feeling Bad Scale (Lewis, Seigel & Lewis, 1984). The life event items
spanned such areas as relationship problems and difficulties; serious illness, accident or death in the
family; educational or employment difficulties; victimization; and financial difficulties. At each age
a total life events score was obtained by summing the number of life events reported by the subject
in the previous year.
Conditional fixed effects regression models for negative binomial data were fitted using
methods of maximum likelihood estimation and the analysis program STATA (StataCorp, 1997).
The results of fixed effects regression models were contrasted with negative binomial regression
models for observed data, estimated within a generalised estimating equation (GEE) framework
(Liang & Zeger, 1986; Zeger & Liang, 1986). These models were fitted assuming an unstructured
correlation matrix of model disturbances over time. Again all models were fitted using STATA.
Association Between Alcohol Abuse and Crime Rates
Table 1 shows the CHDS cohort classified into 4 groups reflecting the number of symptoms
of alcohol abuse reported in each year of observation from age 14-15 years to age 20-21
years. These groups range from those who reported no symptoms of alcohol abuse to those who
reported five or more symptoms. The Table shows, for each age, the mean rate of violent and
property crimes for each alcohol abuse symptom group. Inspection of the Table shows the presence
of clear trends for the increasing abuse of alcohol to be associated with increasing rates of violent
and property crimes. To test the significance of the association between alcohol abuse symptoms
and rates of crime in each year, rates of crime were modelled as log linear function of the number of
symptoms of alcohol abuse under the assumption that rates of crime had a negative binomial
distribution. The table reports on tests of significance for log linear trend for each row of Table
1. This analysis shows that in all cases there was evidence of statistically significant increases
(p<.0001) in rates of violent and property crimes with increases in symptoms of alcohol abuse.
INSERT TABLE 1 HERE
Initial Fixed Effects Regression Model
To examine the extent to which the associations between alcohol abuse and crime shown in
Table 1 could be explained by non observed fixed confounding factors, the data were analysed
using a fixed effects regression model. This model assumed that the rate of crime was a log linear
function of the individual’s level of alcohol abuse. In addition, it was assumed that rates of crime
had a negative binomial distribution. In these and subsequent analyses, the measure of alcohol
abuse was scored as a continuous number of symptoms rather than in the class intervals depicted in
Table 1. In addition, to take account of the fact that rates of crime varied with age, increasing to age
18 and declining thereafter, the fitted model also included age and age2 terms. Table 2a reports the
parameters of the fitted models for violent and property offending. This Table shows that even after
adjustment for the effects of non observed fixed covariate factors and age variations in crime rates,
there were clear and statistically significant (p<.0001) associations between the extent of alcohol
abuse and rates of both violent crime and property crime. This result implies that the associations
between crime and alcohol abuse depicted in Table 1 cannot be explained by the effects of non
observed fixed confounding factors that were related to both alcohol abuse and crime.
For comparison purposes the Table also reports the unadjusted associations between alcohol
abuse and crime rates. The parameters for these models were obtained by fitting negative binomial
regression models to the data depicted in Table 1 using a generalised estimating equation (GEE)
approach. Comparison of the unadjusted associations with the adjusted associations shows that
correction for non observed fixed covariates reduced the associations between alcohol abuse and
crime quite substantiality. However, such adjustment failed to fully explain the associations.
One useful way of presenting the results of the regression models in Table 2a is through the
use of incidence rate ratios. In particular, it can be shown that the fitted model parameters
describing the effects of alcohol abuse on crime, when exponentiated, give an estimate of the
incidence rate ratio: this rate ratio has the interpretation of the proportionate increase in rates of
crime for a one symptom increase in the level of alcohol abuse. The incidence rate ratios and 95%
confidence intervals are shown in Table 2b. The results in Table 2b may be interpreted as follows:
1. Prior to adjustment for confounding factors the fitted model suggested that a one
symptom increase in alcohol abuse symptoms increased rates of violent crime by a factor of 1.6;
after adjustment for fixed confounding factors a one symptom increase was associated with a 1.23
2. Similarly, for property crimes a one symptom increase was associated with a 1.37 times
greater rate of crime whereas after adjustment for fixed confounding factors this reduced to a 1.21
times increase in the rate of crime .
Both sets of conclusions imply that whilst some component of the association between
alcohol abuse and crime was spurious and reflected the presence of non observed fixed confounding
factors, even after such adjustment symptoms of alcohol abuse made an appreciable contribution to
rates of offending.
INSERT TABLE 2 HERE
Extension to Include Time Dependent Covariate Factors
A limitation of the analysis in Table 2 is that it fails to take into account the confounding
effects of time dynamic factors that may be correlated with patterns of alcohol abuse. This in turn
could lead to the fitted model “under controlling” the association between alcohol abuse and crime.
To address this issue the fixed effects regression models fitted in Table 2 were extended to include
measures of observed time dynamic factors including: illicit drug abuse; deviant peer affiliations
and adverse life events. Table 3a reports on the fixed effects models including time dynamic
covariate factors. Table 3b reports on the incidence rate ratios implied by the fitted models.
The Tables show that control for both fixed non observed confounding factors and observed
time dynamic factors further reduced the associations between alcohol abuse and crime.
Nonetheless, even after such control there were still significant (p<.001) associations between
alcohol abuse and crime. The incidence rate ratios show that a one symptom increase in alcohol
abuse was associated with a 1.15 times increase in the rate of violent crime and a 1.10 times
increase in the rate of property offending. Both sets of results imply that even when due allowance
was made for non observed fixed covariate factors and observed time dynamic factors, increasing
alcohol abuse was associated with increasing rates of both violent and property crime.
INSERT TABLE 3 HERE
Comparison of Fixed Effects Regression With Conventional Covariate Adjustment
Although the fixed effects regression models in Tables 2 and 3 make it possible to adjust the
associations between alcohol abuse and crime for non observed fixed confounding factors, a
potential limitation of this method is that it gives little indication of the underlying confounding
processes that are being taken into account. This issue is addressed in Table 4 which reports on the
results of fitting negative binomial regression models in which the frequency of violent and
property offending was regressed on observed fixed and time dynamic covariate factors. As would
be expected, these models contained a similar array of significant time dynamic covariates to those
found for the fixed effects models. In addition, these models showed that rates of violent or
property offending were influenced by a number of fixed factors including: gender (p<.0001);
maternal age (p<.001); family living standards (p<.0001); early onset conduct problems (p<.001)
and offending diversity prior to age 14 (p<.0001).
Comparison of the results in Table 4 with the results in Table 3 shows that the fixed effects
regression models and the models adjusted for observed covariates lead to similar conclusions
about the associations between alcohol abuse and crime. Both sets of results show that: a) a
substantial amount of these associations reflects the presence of fixed and time dynamic
confounding factors; and b) nonetheless, following control for these confounding factors there were
still significant (p<.0001) tendencies for the abuse of alcohol to be related to increasing rates of
crime. As would be expected the fixed effects regression models produce estimates of the effects of
alcohol abuse on crime that are slightly weaker than the adjustments based on observed factors. The
reason for this is that the fixed effects regression model has taken into account all sources of fixed
confounding whereas the covariate adjustment approach takes into account only the observed fixed
sources of confounding.
INSERT TABLE 4 HERE
In this study we have used fixed effects regression models to estimate the association
between the frequency of alcohol abuse symptoms and rates of crime taking into account both
observed and non observed fixed covariates. This analysis was extended to consider: (a) time
dependent covariate factors and (b) a comparison of the fixed effects model with results of
conventional regression adjustment for observed confounders. The major findings and their
implications are reviewed below.
In confirmation of previous studies (Swanson et al., 1990; Pernanen, 1991; Fergusson et al.,
1994; Shepherd, 1994; Greenfield & Weisner, 1995; Fergusson et al., 1996; Kerner et al., 1997),
there was evidence of clear linkages between the extent of alcohol abuse by young people and rates
of offending: the incidence rate ratios for violent and property offending were estimated to increase
by a factor of 1.6 and 1.37 times respectively for every one symptom increase in the level of alcohol
abuse. Both sets of results suggest the presence of quite substantial relationships between alcohol
abuse and criminal behaviour. One interpretation of these results is clearly that involvement in
alcohol abuse increases the likelihood that individuals will engage in criminal behaviours.
However, as we have pointed out earlier, a major threat to the validity of this interpretation comes
from the possibility that the associations between alcohol abuse and crime arise from third or
confounding factors that lead to associations between alcohol abuse and criminal behaviours. In this
paper we have examined the role of potential confounding factors using a number of related
analysis approaches. These approaches include:
1. Fixed effects regression. In the initial analysis reported in Table 2 we examined the
linkages between alcohol abuse and crime using a fixed effects regression modelling approach. This
approach has the advantage of controlling for the effects of non observed fixed covariate factors
and thus at least partially addresses concerns that have been raised about “omitted covariates” in
previous analyses of the linkages between alcohol abuse and crime (Greenfield & Weisner, 1995).
The results of the fixed effects model suggested that although a substantial component of the
association between alcohol abuse and crime was non causal, even after control for non observed
fixed sources of confounding there were still substantial relationships between alcohol abuse and
rates of crime. Rate ratio estimates suggested that a one symptom increase in alcohol abuse
increased rates of crime by a factor of approximately 1.2 times for both violent and property
2. Fixed effects regression with observed time dynamic covariates. A limitation of the fixed
effects model is that it controls for only fixed covariate factors and fails to take into account time
dynamic covariate factors. To address this issue the initial fixed effects model was extended to
include further time dynamic covariate factors including illicit substance abuse, deviant peer
affiliations and adverse life events. The results of these adjustments further reduced the associations
between alcohol abuse and crime but, even after such adjustment, a substantial relationship
remained: the incidence rate ratio estimates derived from the fitted model suggested that every
symptom increase in alcohol abuse increased rates of crime by between 1.10 to 1.15 times, even
when non observed fixed covariates and observed time dynamic covariates were taken into account.
3. Conventional covariate adjustment. A potential limitation of the fixed effects model is
that it produces an estimate of the association between alcohol abuse and crime adjusted for non
observed covariate factors. This feature of the model means that the model estimates give no
indication of the nature of the confounding processes and this in turn may lead to questions about
model realism and validity. A useful way of checking model validity is to compare the results of the
fixed effects regression model with the results of a conventional regression adjustment using
observed covariate factors. Finding that both methods produce similar conclusions provides clear
evidence of model validity. In this instance, the results of the fixed effects regression analysis and
the conventional regression approach produced conclusions that were generally in agreement to the
extent that both approaches suggested that: (a) a substantial amount of the association between rates
of crime and alcohol abuse reflected confounding factors; and (b) even after adjustment for
confounding factors, a substantial relationship between alcohol abuse and rates of crime persisted.
This convergence of conclusions from two regression methods having different strengths and
weaknesses provides generally quite compelling evidence for the view that it is unlikely that the
relationship between alcohol abuse and crime is solely due to the effects of confounding factors.
The results support the view that increasing alcohol abuse is associated with increasing rates of
crime with every symptom of alcohol abuse being associated with in the region of a 1.1 to 1.2 fold
increase in rates of crime.
The conclusion that alcohol abuse is associated with increases in rates of crime appears to be
generally consistent with the results of laboratory based research which has suggested that
increasing consumption of alcohol is associated with corresponding increases in antisocial
behaviour and particularly aggression (Bushman & Cooper, 1990; Taylor & Chermack, 1993).
These parallels between the findings of longitudinal research and laboratory research clearly
reinforce the view that statistical linkages between alcohol abuse and crime, in part at least, reflect a
cause and effect association in which the heavy consumption of alcohol increases risks of criminal
Inevitably, there are a number of caveats that should be imposed on these conclusions.
Perhaps most importantly, although the fixed effects regression model provides a powerful
approach for controlling unobserved sources of confounding, the estimation process relies critically
on a number of key assumptions that are implicit in the fixed effects specification. These
assumptions are: (a) that the relationships between alcohol abuse and crime are such that rates of
crime can be adequately described by a log linear function of alcohol abuse and that there are not
complex, and particularly reciprocal, relationships between these measures over time; and (b) that
the effects of non observed confounding factors on the relationships between alcohol abuse and
crime do not vary with time. Whilst these are clearly strong assumptions, the appeal of the model is
that it provides a means of assessing the effects of observed and non observed covariates on the
association and is, accordingly, likely to give a better estimate of the causal contribution of alcohol
abuse to crime than that obtained in analyses which take into account only those covariate factors
that have been observed. To examine this issue we have contrasted the estimates from the fixed
effects model with those derived from a conventional regression model that included both fixed and
time dynamic covariates. The conventional model assumes that all sources of confounding are
adequately described by observed covariate factors. In this instance there was good agreement
between the findings of the fixed and conventional models, suggesting that the conventional
analysis contained a sufficient number of covariates to adequately control confounding. However,
there was a small tendency for the conventional model to produce slightly larger effect size
estimates than the fixed effects model. It is likely that these slightly larger effect sizes reflect
uncontrolled sources of confounding that were not taken into account in the conventional
model. Nonetheless, it is re-assuring to find that these alternative methods of covariate adjustment
lead to generally similar conclusions about the effects of alcohol abuse on crime.
This research was funded by grants from the Health Research Council of New Zealand, the
National Child Health Research Foundation, the Canterbury Medical Research Foundation and the
New Zealand Lottery Grants Board.
AMERICAN PSYCHIATRIC ASSOCIATION. (1987). Diagnostic and Statistical Manual of
Mental Disorders (3rd ed. - rev). Washington, DC: American Psychiatric Association.
AMERICAN PSYCHIATRIC ASSOCIATION. (1994). Diagnostic and Statistical Manual of
Mental Disorders (4th ed.). Washington, DC: American Psychiatric Association.
BUSHMAN, B.J., & COOPER, H.M. (1990) Effects of alcohol on human aggression: An
integrative research review. Psychological Bulletin, 107, 341-354.
COLLINS, J.J. (1981). Drinking and crime: Perspectives on the Relationship Between Alcohol
Consumption and Criminal Behavior. (New York, Guilford Press).
CONNERS, C. K. (1969). A teacher rating scale for use in drug studies with children. American
Journal of Psychiatry, 126, 884-888.
CONNERS, C. K. (1970). Symptom patterns in hyperkinetic, neurotic and normal children. Child
Development, 41, 667-682.
ELLIOTT, D.S., & HUIZINGA, D. (1989). Improving self-reported measures of delinquency, in:
M.W. Klein (Ed.), Cross-National Research in Self-Reported Crime and Delinquency, pp.
155-186 (Dordrecht, Kluwer Academic Publisher).
ENSOR, T., & GODFREY, C. (1993). Modelling the interactions between alcohol, crime and the
criminal justice system. Addiction, 88, 477-487.
FARRINGTON, D.P., & LOEBER, R. (1999). Transatlantic replicability of risk factors in the
development of delinquency, in: P. Cohen, C. Slomkowski & L.N. Robins (Eds.), Historical
and Geographical Influences on Psychopathology, pp. 299-329 (Mahwah, New Jersey,
Lawrence Erlbaum Associates).
FARRINGTON, D.P., LOEBER, R., ELLIOTT, D.S., HAWKINS, J.D., KANDEL, D.B., KLEIN,
M.W., MCCORD, J., ROWE, D.C., & TREMBLAY, R.E. (1990). Advancing knowledge
about the onset of delinquency and crime, in: B.B. Lahey & A.E. Kazdin (Eds.), Advances in
Clinical Child Psychology, Vol. 13, pp. 383-442 (New York, Plenum Press).
FERGUSSON, D.M., & HORWOOD, L.J. (1996). The role of adolescent peer affiliations in the
continuity between childhood behavioral adjustment and juvenile offending. Journal of
Abnormal Child Psychology, 24, 205-221.
FERGUSSON, D.M., & HORWOOD, L.J. (1999). Prospective childhood predictors of deviant
peer affiliations in adolescence. Journal of Child Psychology and Psychiatry, 40, 581-592.
FERGUSSON, D. M., HORWOOD, L. J., & LLOYD, M. (1991). Confirmatory factor models of
attention deficit and conduct disorder. Journal of Child Psychology and Psychiatry, 32, 257-
FERGUSSON, D.M., HORWOOD, L.J., SHANNON, F.T., & LAWTON, J.M. (1989). The
Christchurch Health and Development Study: A review of epidemiological findings.
Paediatric and Perinatal Epidemiology, 3, 278-301.
FERGUSSON, D.M., & LYNSKEY, M.T. (1998). Conduct problems in childhood and
psychosocial outcomes in young adulthood: A prospective study. Journal of Emotional and
Behavioral Disorders, 6, 2-18.
FERGUSSON, D.M., LYNSKEY, M.T., & HORWOOD, L.J. (1994). Alcohol consumption and
associated problems in a birth cohort of 15 year olds. New Zealand Medical Journal, 107,
FERGUSSON, D.M., LYNSKEY, M.T., & HORWOOD, L.J. (1996). Alcohol misuse and juvenile
offending in adolescence. Addiction, 91, 483-494.
GREENE, W.H. (1993). Econometric Analysis (New York, Macmillan Publishing Company).
GREENFIELD, T.K., & WEISNER, C. (1995). Drinking problems and self-reported criminal
behavior, arrests and convictions: 1990 US alcohol and 1989 county surveys. Addiction, 90,
HAMERLE, A., & RONNING, G. (1995). Panel analysis for qualitative variables, in: G. Arminger,
C. Clogg & M. E. Sobel (Eds.), Handbook of Statistical Modeling for the Social and
Behavioral Sciences, pp. 401-451 (New York, Plenum Press).
HAUSMAN, J., HALL, B. H., & GRILICHES, Z. (1984). Econometric models for count data with
an application to the patents-R & D relationship. Econometrica, 52, 909-938.
HAWKINS, J. D, CATALANO, R.F., & MILLER, J.Y. (1992). Risk and protective factors for
alcohol and other drug problems in adolescence and early adulthood: Implications for
substance abuse prevention. Psychological Bulletin, 112, 64-105.
JUDGE, G.E., GRIFFITHS, W.E., HILL, R.C., & LEE, T. (1980). The Theory and Practice of
Econometrics (New York, John Wiley & Sons).
KERNER, H., WEITEKAMP, E.G.M., STELLY, W., & THOMAS, J. (1997). Patterns of
criminality and alcohol abuse: Results of the Tuebingen criminal behaviour development
study. Criminal Behaviour and Mental Health, 7, 401-420.
LEWIS, C.E., SEIGEL, J.M., & LEWIS, M.A. (1984). Feeling bad: Exploring sources of distress
among pre-adolescent children. American Journal of Public Health, 74, 117-122.
LIANG, K.-Y., & ZEGER, S. L. (1986). Longitudinal data analysis using generalized linear
models. Biometrika, 73, 13-22.
MILGRAM, G.G. (1993). Adolescents, alcohol and aggression. Journal of Studies on Alcohol
Suppl, 11, 53-61.
MOFFITT, T.E., & SILVA, P.A. (1988). Self-reported delinquency: Results from an instrument for
New Zealand. Australian and New Zealand Journal of Criminology, 21, 227-240.
MOTT, J. (1990). Young people, alcohol and crime. Home Office Research & Statistics
Department, Research Bulletin No. 28: 24-28.
OTERO-LOPEZ, J.M., LUENGO-MARTIN, A., MIRON-REDONDO, L., CARRILLO-DE-LA-
PENA, M.T., & ROMERO-TRINANES, E. (1994). An empirical study of the relations
between drug abuse and delinquency among adolescents. British Journal of Criminology, 34,
PERNANEN, K. (1991). Alcohol in Human Violence (New York, Guilford Press).
RUTTER, M., TIZARD, J., & WHITMORE, K. (1970). Education, Health and Behaviour
SHEPHERD, J. (1994). Violent crime: The role of alcohol and new approaches to the prevention
of injury. Alcohol and Alcoholism, 29, 5-10.
STATACORP. (1997). Stata Statistical Software: Release 5.0 (College Station, TX, Stata
SWANSON, J.W., HOLZER III, C.E., GANJU, V.K., & JONO, R.T. (1990). Violence and
psychiatric disorder in the community: Evidence from the epidemiologic catchment area
surveys. Hospital and Community Psychiatry, 41, 761-770.
TAYLOR, S.P., & CHERMACK, S.T. (1993). Alcohol, drugs and human physical aggression.
Journal of Studies on Alcohol Suppl, 11, 78-88.
WAGNER, E.F. (1996). Substance use and violent behavior in adolescence. Aggression and
Violent Behavior, 1, 375-387.
WHITE, H.R., & LABOUVIE, E.W. (1989). Towards the assessment of adolescent problem
drinking. Journal of Studies on Alcohol, 50, 30-37.
WHITE, H.R., BRICK, J., & HANSELL, S. (1993). A longitudinal investigation of alcohol use
and aggression in adolescence. Journal of Studies on Alcohol Suppl, 11, 62-77.
WORLD HEALTH ORGANIZATION. (1993). Composite International Diagnostic Interview
(CIDI) (Geneva, World Health Organization).
ZEGER, S. L., & LIANG, K.-Y. (1986). Longitudinal data analysis for discrete and continuous
outcomes. Biometrics, 42, 121-130.
Table 1. Mean Number of Violent and Property Offences by Age and Concurrently Measured
Symptoms of Alcohol Abuse
Number of Symptoms of Alcohol Abuse
AgeOffence Type0 1-23-4 5+p*
14-15 Years (N)(864) (67) (22)(12)
Violent0.61 0.736.861.92 <.01
Property0.61 3.6111.827.67 <.001
15-16 Years(N)(828) (73)(31)(21)
Property 0.984.326.65 14.43 <.001
17-18 Years (N)(782)(138)(55)(50)
Property0.64 4.314.6226.08 <.0001
20-21 Years(N)(749)(161) (63)(38)
* P-value derived from negative binomial regression model of the association between alcohol
abuse and each offence type in each year.
Table 2. Comparison of Unadjusted and Fixed Effects Adjusted Regression Models for Violent and
a) Model Parameters
Violent Offending *
Property Offending *
a) Estimated Offending Incidence Rate Ratio
Violent Offending Property Offending
Unadjusted Fixed Effects Unadjusted Fixed Effects
Rate ratio (95% CI) for
a one symptom
increase in alcohol
* All coefficients statistically significant at p<.0001 level
Table 3. Fixed Effects Regression Models Including Time Dynamic Covariates
a) Model Parameters
Violent Offending Property Offending
Measure B (se)P B (se)P
Alcohol abuse .139 (.031)<.0001 .092 (.027) <.001
Age .947 (.100)<.0001.211 (.077)<.01
Age squared-.142 (.014)<.0001-.054 (.012) <.0001
Time Dynamic Covariates
Illicit drug abuse-* -.101 (.028)<.001
Deviant peer affiliations .039 (.007)<.0001 .047 (.006) <.0001
Adverse life events.089 (.026)<.001.090 (.022) <.001
b) Estimated Offending Incidence Rate Ratio
Violent OffendingProperty Offending
Rate ratio (95% CI) for a one
symptom increase in alcohol
* Parameter not statistically significant. Effect excluded from final fitted model.
29 Download full-text
Table 4. Fitted Negative Binomial Regression Models for Violent and Property Offending
Controlling for Observed Confounders
a) Model Parameters
Violent Offending Property Offending
Measure B (se)PB (se)P
Alcohol abuse .200 (.017)<.0001.163 (.018)<.0001
Age .928 (.053)<.0001.526 (.052) <.0001
Age squared-.142 (.008) <.0001-.126 (.008) <.0001
Time Dynamic Covariates
Illicit drug abuse -* -.220 (.019) <.0001
Deviant peer affiliations .062 (.003)<.0001 .089 (.003) <.0001
Adverse life events.258 (.013)<.0001 .208 (.013) <.0001
Gender -1.244 (.066)<.0001-1.275 (.072) <.0001
Maternal age-.029 (.007)<.001 -* -
Family standard of living
.108 (.008) <.0001-* -
Conduct problems (8 years).025 (.003) <.001 -* -
Offending diversity score
.162 (.012) <.0001.228 (.010)<.0001
b) Estimated Offending Incidence Rate Ratio
Violent OffendingProperty Offending
Rate ratio (95% CI) for a one
symptom increase in alcohol
* Parameter not statistically significant. Effect excluded from final fitted model.