41
Durbin–Watson statistic is close to 2.
The key outcome variables in our assessment of the STD intervention were the monthly
number of confirmed shots fired and robberies. Since the underlying data were counts, a Poisson
regression in a log-linear model was selected to analyse the time series data. Poisson regression
applies where the dependent variable is a count (e.g. crime incidents, cases of a disease) rather
than a continuous variable. It assumes the response variable has a Poisson distribution whose
expected value (mean) is dependent on one or more predictor variables. Typically the log of the
expected value is assumed to have a linear relationship with the predictor variables. As Crawley
(2007, p. 527) notes, linear regression is not appropriate for such data since (1) the linear model
might lead to the prediction of negative counts; (2) the variance of the response variable is likely
to increase with the mean; (3) the errors will not be Normally distributed, and (4) zeros are
difficult to handle in transformations.
The aim of the time series approach was to isolate and evaluate the direct impact of the
implementation of the STD only and STD plus interventions on reported offenses in Joliet. To
accomplish this
three dummy variables were created to represent the three distinct program
periods (i.e., pre-STD, STD only and STD + probation/parole) in order to estimate the effects of
the intervention on the monthly counts of shots fired and robberies.
It also is important to
consider that potential reductions in shots fired and robbery associated with the STD intervention
could be influenced by other factors. Therefore
we
also included covariates
to control for any
changes in the monthly counts of shots fired and robberies that could be associated with other
factors such as changes in Joliet’s percent of minority population, the unemployment rate and the
number of police officers. Furthermore, the models control for the number of drug arrests across
the pre and post intervention periods. Over the 7.75 year period there were 5,619 drug arrests
42
(1,567 pre-intervention and 4,052 intervention drug arrests). Controlling for drug arrests may be
of consequence in light of research on the impact of zero tolerance policing strategies on crime
reduction.
Studies have shown that aggressive enforcement activities by uniformed patrol
officers targeting illicit drug sale locations and illegal drug activity can produce short-term
reductions in street level drug dealing as well as reductions in many other types of criminal and
disorderly behavior (Braga et al., 1999; Sherman and Rogan, 1995; Sherman and Weisburd,
1995; Weisburd and Green, 1995). For example, aggressive patrolling of suspected drug
locations in Jersey City involving stops and searches by patrol officers resulted in reductions in
crime not only at the drug locations, but for several surrounding blocks as well (Braga et al.,
1999; Weisburd and Green, 1995).
Results
T
able 5 presents the results of the Poisson regression models for shots fired and robbery.
Poisson regression models have the defining characteristic that the conditional mean of the
outcome is equal to the conditional variance.
Goodness of fit tests based on the Deviance and
Pearson residuals were used to assess whether the model assumptions have been violated. Each
should approximately equal its degrees of freedom and so Value/df (value divided by degrees of
freedom) should be close to one. The data show that the Poisson regression model is a good fit.
Next, the Omnibus Test, a test to determine if all of the estimated coefficients are equal to zero
(a test of the model as a whole), was statistically significant.
The parameters for the independent variables were expressed as incidence rate ratios (i.e.,
exponentiated coefficients). Incidence rate ratios are interpreted as the rate at which things occur;
for example, an incidence rate ratio of 0.65 would suggest that, controlling for other independent
variables, the selected independent variable was associated with a 35% decrease in the rate at
43
which the dependent variable occurs. To ensure that the coefficient variances were robust to
violations of the homoskedastic errors assumption of linear regression models,
Huber/White/sandwich robust variance estimators were used.
Table 5
Poisson Regression Results for Shots Fired and Robbery
Shots
Fired
Robbery
Variable
B (SE)
IRR
p-value
B (SE)
IRR
p-value
Trend
.006
(.006)
1.01 .170 -.003
(.007)
.997 .672
unemployment
.000
(.016)
1.00 .975 -.025
(.021)
.976 .241
% minority pop.
.294 (1.07)
1.34
.783
1.63 (1.40)
5.12
.245
# police officers
.001 (.002)
1.00
.477
.000 (.002)
1.00
.920
Monthly # drug
arrests
.001 (.002)
1.00
.595
.004 (.003)
1.00
.136
STD only
-.123 (.126)
.885
.329
.453 (.191)
1.57
.018
STD +
probation/parole
-.368
(.235)
.692 .118 .436
(.355)
1.55 .220
intercept
2.89 (.133)
---
.000
1.61 (.236)
---
.000
Shots fired: Deviance = 205.08, Pearson Chi-square = 202.45 Value/df = 2.41/2.38
Likelihood ratio Chi-Square = 22.75, df = 7, p = .002
Robbery:
Deviance = 166.09, Pearson Chi-square = 154.18 Value/df = 1.95/1.81
Likelihood ratio Chi-Square = 42.40, df = 7, p = .000
Controlling for the covariates, the STD and STD+ program interventions were associated
with a statistically significant decrease in the monthly number of shots fired. According to the
incidence rate ratios, the STD only and STD + probation and parole components were associated
with a 11% and 31% decrease in the monthly number of shots fired events, respectively.
However, these decreases did not differ significantly from the pre-intervention period (p = .329