Project Safe Neighborhoods and Violent Crime Trends in US Cities: Assessing Violent Crime Impact

Abstract

Since the mid-1990s, a number of initiatives intended to address gang, gun and drug-related violence have arisen and demonstrated promise in reducing levels of violent crime. These initiatives have employed some combination of focused deterrence and problem-solving processes. These strategies formed the basis for Project Safe Neighborhoods (PSN), a national program implemented by the Department of Justice and coordinated by US Attorneys’ Offices. This paper is an initial attempt to assess the potential impact of the nationally implemented PSN initiative through an analysis of violent crime trends in all US cities with a population of 100,000 or above. While a number of site specific studies exist examining the potential impact of locally implemented PSN programs, to date no national-level study has examined whether PSN may have had an impact on violent crime trends. Cities included in the current study are distinguished on the basis of whether they were considered a treatment city by the PSN task force and by the level of implementation dosage of the PSN program. This allowed a comparison of 82 treatment cities and 170 non-treatment cities as well as a variable of dosage level. Hierarchical Generalized Linear Models (HGLM) were developed that controlled for other factors that may have affected the level of violent crime across the sample of cities. The results suggested that PSN treatment cities in higher dosage contexts experienced statistically significant, though modest, declines in violent crime whereas non-target cities and low dosage contexts experienced no significant changes in violent crime during the same period. The limitations of this initial analysis are noted but the evidence seems to suggest that the multi-agency, focused deterrence, problem solving approach holds promise for reducing levels of violent crime. At a minimum, these findings call for continued programmatic experimentation with data-driven, highly focused, deterrence-based violence reduction strategies.

Appendices

Appendix A

Table 4 Detailed description of PSN dosage components

Appendix B

The HGLM model presented in Table 2 can be written as follows (Note: All level 1 measures were group-centered, all level 2 measures were grand-mean centered, and no random variance components were estimated for the annual dummy variables because all cities had the same fixed value for this measure):

$$ \begin{aligned} \eta_{ti} = & \pi_{0i} + \pi_{{ 1i}} \left( {\text{PSN Dosage}} \right) + \pi_{{ 2i}} \left( {\text{Prison Rate}} \right) + \pi_{{ 3i}} \left( {\text{Police Density}} \right) + \pi_{{ 4i}} \left( { 200 1} \right) \\ & + \pi_{{ 5i}} \left( { 200 2} \right) + \pi_{{ 6i}} \left( { 200 3} \right) + \pi_{ 7i} \left( { 200 4} \right) + \pi_{{ 8i}} \left( { 200 5} \right) + \pi_{{ 9i}} \left( { 200 6} \right) + e_{ti} \\ \end{aligned} $$

where η _ti  = the expected violent crime rate, π_0i  = β₀₀ + β₀₁(Disadvantage) + β₀₂(PopDensity) + r _oi , π_1i  = β₁₀ + r _1i , π_2i  = β₂₀ + r _2i , π_3i  = β₃₀ + r _3i , e π_4i  + ··· + π_9i  = β₄₀+ ··· + β₉₀. Thus, the reduced two-level equation can be written as:

$$ \eta_{ti} = \left[ {\pi_{0i} + \pi_{{ 1i \, }} + \pi_{{ 2i}} + \pi_{{ 3i}} + \pi_{ 4i \, } + \ldots + \pi_{{ 9i}} } \right] + \left[ {r_{oi} + r_{ 1i} + r_{ 2i} + r_{ 3i} + e_{ti} } \right] $$

The outcome in the HGLM is the count of violent crimes at level 1 and includes time-varying covariates within the cities (i.e., level 1 measures) and time invariant measures at level 2. As seen in the above equations, PSN dosage is included as a time-varying measure at level 1. Here, the HGLM model uses an overdispersed Poisson sampling model at level 1 and a log link function to equate the transformed count into a linear structural model. The log link function in the HGLM is used to equate the transformed count into a linear structural model, consistent with the regression-based analytic approach.

In the analysis, the HGLM Poisson model assumes an expected violent crime count

$$ E\left( {Y_{it} |\lambda } \right) = m_{it} \lambda_{it} , $$

where λ _it is expressed as the violent crime rate of a city i at time t and m _it is the exposure measure, which is expressed as the city population in 100,000 s. The expected violent crime rate for a city is transformed through a natural logarithmic function, where η _ti  = ln(λ _it ). The logged event rate, η _ti , becomes the dependent variable in the level 1 model.

Appendix C

The HGLM model presented in Table 3, which was the alternative growth curve model where the PSN treatment city designation is a static measure at level 2, is written as follows (Note: all level 1 measures were group-mean centered, and level 2 measures were grand-mean centered):

$$ \eta_{ti} = \pi_{0i} + \pi_{{ 1i}} \left( {\text{Prison Rate}} \right) + \pi_{{ 2i}} \left( {\text{Police Density}} \right) + \pi_{{ 3i}} \left( {\text{Time}} \right) + e_{ti} $$

where π_0i  = β₀₀ + β₀₁(PSN Treatment City) + β₀₂(Disadvantage) + β₀₃(PopDensity) + r _oi , π_1i  = β₁₀ + r _1i , π_2i  = β₂₀ + r _2i , π_3i  = β₃₀ + β₃₁(PSN Treatment City).

Thus, the reduced two-level equation can be written as:

$$ \eta_{ti} = \left[ {\pi_{0i} + \pi_{{ 1i \, }} + \pi_{{ 2i}} + \pi_{{ 3i}} } \right] + \left[ {r_{oi} + r_{ 1i} + r_{ 2i} + e_{ti} } \right] $$

The outcome in the HGLM is the count of violent crimes at level 1 and includes time-varying covariates within the cities (i.e., level 1 measures) and time invariant measures at level 2. As seen in the above equations, PSN treatment is a level 2 static measure. Here, the HGLM model uses an overdispersed Poisson sampling model at level 1 and a log link function to equate the transformed count into a linear structural model. The log link function in the HGLM is used to equate the transformed count into a linear structural model, consistent with the regression-based analytic approach.

In the analysis, the HGLM Poisson model assumes an expected violent crime count

$$ E\left( {Y_{it} |\lambda } \right) = m_{it} \lambda_{it} , $$

where λ _it is expressed as the violent crime rate of a city i at time t and m _it is the exposure measure, which is expressed as the city population in 100,000 s. The expected violent crime rate for a city is transformed through a natural logarithmic function, where η _ti  = ln(λ _it ). The logged event rate, η _ti , becomes the dependent variable in the level 1 model.