Baltimore, Maryland has long been plagued by high rates of homicides and non-fatal penetrating firearm injuries (non-fatal shootings). The homicide rate in Baltimore, Maryland, has been one of the highest among the US cities with over 500 000 people, with the vast majority of these killings involving firearms.¹ From 2003 to 2017, nearly 4000 Baltimore residents lost their lives to homicide and over 8000 were victims of non-fatal shootings.² City leaders in Baltimore, as in many urban locales, have sought to implement community programmes to prevent violence among individuals at high risk for firearm violence.

Cure Violence is an intervention that employs community members to interrupt violence through conflict mediation, model and promote non-violence, and link individuals at high risk of violence to needed services.^3–5 The Cure Violence model has been implemented and evaluated in multiple cities around the US.^5–8 However, studies examining the programme’s impact on homicides and non-fatal shootings, while generally positive, have produced varied programme effect estimates across sites and cities, with some studies reporting no beneficial effects.^4–10

The Cure Violence model was first implemented in Baltimore, Maryland, in 2007. Named ‘Safe Streets’, the programme was started in police posts whose total homicides and non-fatal shootings were in the top 25th percentile (see online supplemental figure 1 for a map with locations and operational periods for all programme sites). The impact of Safe Streets was first assessed in an evaluation of the original four programme sites in Baltimore using 2007–2010 operational data,¹¹ and then again in 2018 using data on the then-seven Safe Streets sites operating between 2007 and 2017.¹² Difference-in-differences regression strategies in both evaluations yielded wide-ranging effect sizes across sites in both homicides and non-fatal shootings. The 2018 evaluation found no aggregate protective effects of Safe Streets on homicides or non-fatal shootings. Models that estimated site-specific effects on homicides found that only one site (Cherry Hill) experienced a sizeable and statistically significant reduction, while the direction and effect sizes of homicide estimates for the other six Safe Streets sites ranged from null to a statistically significant increase in homicides in one site. Site-specific effects on non-fatal shootings showed that four of the seven sites had reductions in non-fatal shootings, while three experienced increases; none of the estimates were statistically significant.

The previous Safe Streets evaluations, like other Cure Violence evaluations, were limited in their ability to identify suitable comparison units for estimating programme effects because they used either traditional regression analyses or interrupted time-series approaches that do not use statistical methods to determine the most appropriate non-intervention controls for estimating the counterfactual for intervention areas. The objectives of the current study were to improve on the 2012 and 2018 evaluations of Safe Streets and extend the literature on the effectiveness of Cure Violence replications by using the synthetic control method (SCM), which has noted advantages over regression approaches for estimating intervention effects of non-randomised interventions.¹³ We hypothesised that by using the SCM method, we would generate more accurate estimates of intervention effects, allowing us to better assess the effectiveness of Safe Streets.

Table 1 compares the preintervention average monthly values of outcome variables and predictors in the treated police post, its synthetic control, and the unweighted pool of comparison units which were applied in the previous regression analyses. As illustrated by the larger variance between the unweighted comparison units and the treated sites in relation to the synthetic controls, the weighted approach used to generate the SCM provides closer estimates of the predictors than do the pooled unweighted comparison units. Furthermore, the prediction errors in the SCM, shown in table 2 as the RMSPE, were substantially smaller than the corresponding prediction errors from the regression models (RMSE), demonstrating that the SCM produces a better fit for estimating the counterfactuals and programme effects for each intervention site. This improved model fit can also be visually noted in graphs comparing the SCM preintervention predicted versus observed values for homicides (online supplemental figure 2A) and non-fatal shootings (online supplemental figure 2B) with the regression model preintervention predicted versus observed values for homicides (online supplemental figure 2C) and non-fatal shootings (online supplemental figure 2D). The graphs for the placebo tests showing the observed minus predicted values are shown in online supplemental figure 3A for homicides and online supplemental figure 3B for non-fatal shootings.

The SCM estimates of programme effects are shown in table 3 as per cent change increases or decreases in homicides and non-fatal shootings post programme implementation. Estimated effects varied widely by both site and outcome. For example, the Cherry Hill and Sandtown-Winchester sites saw reductions in homicides (−21% and −9%, respectively) but not non-fatal shootings (+11% and +15%, respectively). In contrast, the Madison-Eastend and Mondawmin sites experienced increases in homicides (+69% and +76%, respectively) and non-fatal shootings (+153% and +27%, respectively) during the study period. Lower Park Heights experienced non-significant decreases in both homicides and non-fatal shootings, while the homicide increases in Mondawmin and Elwood Park and the non-fatal shootings in Madison-Eastend were significantly different from the placebo test results.

We found inconsistent programme effect estimates over time for the two outcomes in McElderry Park and Cherry Hill (table 4). In McElderry Park, the programme was associated with homicide reductions for the 3-year (−62%), 5-year (−48%) and 7-year (−24%) estimates, indicating attenuated protective effects on homicide over time, while non-fatal shootings increased relative to the counterfactual over the same time periods. In contrast, Cherry Hill saw relatively stable programme-related decreases in homicides and greater reductions in non-fatal shootings at each duration tested.

The results of the preunrest and postunrest analyses suggest that the civil unrest had varied impacts in the Safe Streets locations (table 5). For example, compared with its preunrest estimate for non-fatal shootings, Lower Park Heights appears to have had a postunrest reduction in non-fatal shootings relative to the estimated counterfactual, while Mondawmin, which was at the centre of the civil unrest, saw dramatic increases in both homicides and non-fatal shootings following the unrest relative to its synthetic control.

This study estimated the effects of Safe Streets, a public health programme based on the Cure Violence model, in seven intervention sites in Baltimore, Maryland, using SCM. The findings indicate that Safe Streets has had disparate impacts on violence across implementation sites with more evidence of harm than benefit. Three sites—Madison-Eastend, Elwood Park and Mondawmin—experienced substantial increases in both fatal and non-fatal violence during the study period. These sites lost their funding after experiencing poor outcomes. While other sites did see programme-related reductions in at least one violence outcome, none of the seven Safe Streets sites experienced violence reductions outside the norm when compared with placebo tests.

The varied effect estimates provide additional context and nuance to the existing literature on Cure Violence interventions. The findings that Safe Streets in Baltimore was associated with significant increases in violence in three locations raise great concern, and when considered along with the null or negative findings in a number of past Cure Violence evaluation studies, the results of this research suggest that closer examination of implementation and contextual factors that influence programme effects is warranted. These findings may also suggest that programme enhancements or modifications are needed to achieve significant reductions in firearm violence in the neighbourhoods with exceptionally high rates of violence.

This study offered interesting comparisons with the 2018 Safe Streets evaluation. Both analyses showed that the effects of the programme varied widely by location. However, the regression analyses estimated that Cherry Hill experienced a significant reduction in homicide, while the SCM analysis in the current study found that, although Cherry Hill did experience homicide reductions, the effects were not significant. The two analytic approaches also yielded some differences in the magnitude of effect estimates. Given the model fit comparisons highlighted in this study, the findings from the SCM analyses are likely better estimates of Safe Streets’ site-specific effects than were the previous regression analyses.

Taken in aggregate, the evaluations of Safe Streets suggest that the initial potentially positive effects of the programme attenuated over time. Prior studies found that Safe Streets increased youth’s preferences for non-violent conflict resolution³¹ and was associated with increases in preferences for non-violent responses to interpersonal conflicts.³² Additionally, the Cure Violence model has been associated with reductions in youth’s willingness to use violence to settle conflict,³³ improved confidence in police³⁴ and increased confidence in a community’s ability to reduce firearm violence.⁹ However, it is important to note that community outreach and violence interruption programmes, regardless of branding, vary from site to site and city to city in their implementation, investment and management strategy. There is limited information on implementation differences across Safe Streets sites; each received the same training, city contracts and technical assistance.¹¹ The differential programme effects found in this and other studies and the attenuation of early protective effects underscore the importance of strong implementation, appropriate oversight and meaningful support of community-level violence interventions like Safe Streets.^{2 3} Geographical and contextual differences across sites may also influence the model’s probability for success. For example, the Cherry Hill area is geographically isolated, which may offer unique protection that other sites do not, and a cross-sectional study of Baltimore found that Safe Streets-associated reductions in homicides were associated with a higher proportion of gang-related conflict mediations, while neighbourhoods without similar programme-associated homicide reductions saw more mediations for weapons and retaliatory conflicts.³⁵ The underlying drivers of potentially violent interpersonal disputes vary by community, and they must be understood and appropriately matched with interventions that can adequately address those drivers. Future research should also examine how the theoretical concepts of the Cure Violence model apply to various communities in Baltimore and how the strategy might be adapted or enhanced to better achieve the desired outcomes.

Furthermore, promising findings from studies of Cure Violence programmes in New York suggest that programme success may be associated with increased financial, mental health and training resources for staff and programme participants, as well as the availability of wraparound services for the individuals engaged by outreach workers—things that have been lacking for Safe Streets Baltimore.^{2 9 33 34} Future research should examine distinct variations in implementation across sites and focus on how discrepancies in programme impact may be explained by factors such as programme oversight, outreach worker training and support, collaboration with community-based organisations, and the type, number, quality and utilisation of support and social services available to programme participants.

This study is the first known research to assess how civil unrest and increases in community violence, particularly following incidents of police violence, may influence the outcomes of violence interventions. While deeper investigation of unrest-related effects was beyond the scope of this study, the increases in violence in three of the four sites in operation when the unrest occurred suggest that Safe Streets may have been ill-equipped to address the flared tensions and increased violence that frequently occur in communities experiencing unrest following incidents of police violence.^{36 37} This finding importantly sheds light on similar challenges experienced during the summer of 2020, when dozens of cities across the US saw both civil unrest and coincident increases in violence, even in places where community-based violence prevention strategies existed but were hamstrung by the COVID-19 pandemic.

There are several study limitations to note. First, due to large differences in operational periods across the programme sites, the amount of data available to estimate preintervention trends or postintervention effects differed greatly and affected model fit. For instance, according to their respective RMSPEs, the models for the Safe Streets sites in Mondawmin and Sandtown-Winchester had better fit than did those for the longest-standing sites in McElderry Park and Cherry Hill. Conversely, the earlier sites benefited from much more postintervention data than did the sites that opened more recently. There was no appealing strategy for addressing this limitation. However, the RMSPEs were small for all sites, and the examination of programme effects over different time periods provided insight into variance across sites over time, irrespective of programme length.

An additional limitation is that the SCM is unable to account for breaks in the intervention. The intervention’s effects following the April 2015 civil unrest could not be isolated, although we attempted to account for the effects by estimating Safe Streets’ impact on violence preunrest and then for the entire duration of the study period. Similarly, temporary breaks in the programmes in McElderry Park and Mondawmin as a result of site reviews following programme worker arrests could not be modelled with the SCM. However, those programme interruptions were brief and thus unlikely to have had a major impact on the intervention’s effects in those areas.

Despite these limitations, the current study makes important contributions to the body of literature on community-level violence interventions. To our knowledge, this is the first study to use SCM to examine site-specific and time-varying effects of a Cure Violence programme model on violence outcomes. It is also the longest examination of a Cure Violence programme’s operation in multiple cities within one city, as well as the first to consider differential programme impacts following civil unrest spurred by police violence.

The current study illuminates challenges associated with achieving and sustaining reductions in violence using a programme like Safe Streets. It also elevates the urgent need to provide greater support for the workers who are tasked with outreach and violence interruption and risk their lives to reduce violence in communities. The model has demonstrated some success in other prior quantitative and qualitative evaluations, but variation in effectiveness highlights the importance of closer examination of various aspects of the programme. The results from this study emphasise careful implementation and continuous evaluation of violence prevention efforts. This research also points to a greater need to more deliberately identify and replicate critical components of effective community outreach and violence interruption work, and to more directly address systemic drivers of community violence, in order to reduce violence in communities nationwide.