We examined BID effects by modelling the pre–post changes in the incidence or rate of violent crimes from 1994 to 2005 for all neighbourhood areas exposed to BID. Between 1996 and 2003, a total of 30 separate BID were implemented in Los Angeles. The unit of analysis is any of the 30 neighbourhood areas that eventually adopted a BID in Los Angeles during the period of observation. Table 1 reports the number of BID areas that became fully operational at any given year. We consider a BID fully operational if its implementation occurred for the entire calendar year. The formulation of BID in Los Angeles requires a formalised and uniform planning and adoption stage that is structured by law. Details are provided elsewhere.12 For all the areas that eventually adopted a BID in Los Angeles, there are at least 2 years' worth of data during which no BID (pre) was operational and, similarly, at least 2 years of data during which all the BID were fully operational (post). We make use of this type of interrupted time series to estimate the average BID effect on the rate of violent crimes.

The data consist of the yearly counts of officially recorded violent felony crimes (homicide, rape, robbery and aggravated assault) by the Los Angeles Police Department that were aggregated from police reporting districts to corresponding BID areas. If a reporting district is present in a BID this district is counted as receiving the BID intervention. A total of 179 reporting districts are present in the 30 BID neighbourhood areas. We focus primarily on the counts of robbery, because this crime is less susceptible to underreporting by the police, is more likely than other crimes to occur in public settings between strangers and just over 30% of people who are robbed experience an injury.13–15 Nearly one out of five robberies nationally result in a serious injury that requires medical treatment, such as a gunshot wound, knife/stab/slash wound, broken bones or teeth, internal injury and/or loss of consciousness.16 17 The estimated social costs of an average robbery are high, with an average cost of US$39 287 (in 2005 dollars) if one includes monetary costs associated with medical and emergency services, lost productivity, mental health and general quality of life.18 By focussing our analysis primarily on violent crimes, and in particular robbery, we are offering a closer look at the effective role that BID expend on crime prevention efforts that attempt to improve the social control of public space areas through environmental design modifications and spending on private security. We do not examine property crime offences because they are less likely to be reported to the police, many occur in private settings (eg, larceny/theft) and are not ‘street crimes’ directly subject to BID efforts to enhance social control in public spaces.

Table 2 presents a summary of the average frequency of robbery and all violent crimes for areas exposed to BID compared with non-BID areas that incorporate the rest of Los Angeles. The simple linear trend of these data indicates that BID areas experienced greater, on average, yearly reductions in the incidence of robbery and violence than non-BID areas. For example, the average yearly reduction in robbery was 1.9 in BID areas, compared with 1.2 in non-BID areas. The log of the average robbery counts indicates a 7% reduction in BID areas compared with a 5.7% reduction in non-BID areas. During the 12-year period, however, the average yearly count of reported violent offences dropped by 58% for Los Angeles as a whole, suggesting that it is important to take into account this secular trend in assessing the effects of BID on violent crimes.

We used a Bayesian hierarchical model to assess the pre–post effects of BID adoption in areas that were exposed to BID. We estimate the BID effect in each individual area and the average BID effect across all areas. We model the number of reported robberies and violent crimes with a Poisson distribution with mean (λ_it) for each of the 30 BID areas (i) over a 12-year (t) time period. We include a random effect parameter for areas to scale the time trend for the volume of crimes in each BID area, and to account for time invariant differences across the 30 BID areas. We account for the overall 12-year crime time trend in Los Angeles with a natural cubic spline.19 As the yearly incidence of violent crimes for Los Angeles as a whole was declining over the 12-year study period, it is important to account for the secular trend in the model so as not to overestimate the BID effect. The population at risk of violent crimes in each BID area is unknown, representing a mix of residents and non-residents (eg, shoppers). We bypass the problem of having to estimate the population at risk of violent crimes by comparing every BID area with itself. Under the assumption that the population at risk does not change over time, we propose a model that has as the main parameter of interest for area i the ratio (K_i) of the crime rate if a BID was adopted divided by the crime rate that the trends predict would have happened if the BID had not been adopted. If K_i is less than 1, this indicates that the presence of a BID in area (i) is associated with a reduction in the incidence of violent crimes for that area. We use a Bayesian hierarchical model to assess individual effects across each BID area (i) and to approximate the average overall BID effect (μ_K) across all 30 areas.20 Full technical details on the model and specification of priors are available elsewhere.12

The yearly violent crime trends for a sample of 12 of the 30 BID areas are displayed in figure 1 and show the timing of each of the BID interventions (denoted by a vertical line). Figure 1 provides a visual sense for the fact that the Bayesian hierarchical model is estimating an interrupted time series for each unique BID area and then pooling the effect across all areas. The Poisson model implies the counterfactual that the rate of violent crimes in an area after the BID becomes fully operational is proportional to what the rate would have been in that area had the BID not been implemented.

We also conducted several additional tests on our measures of violent crime and the methods for approximating BID effects. Homicides, for example, are the most accurately reported violent crimes, but we did not discuss the yearly trends in homicide because the counts are so low. The average number of homicides per year in neighbourhoods associated with BID is less than one, and the median (50th percentile) is zero. The point estimates from replicating our model for homicide counts varies widely, and the probability of detecting a BID effect is low (P(μ_K<1)=0.43)), suggesting that BID have no appreciable effect on homicide. In particular, the homicide model indicates a 5% increase associated with BID adoption but with a large CI (95% posterior probability interval −50 to 29), resulting from low counts and imprecision in our estimate. A combined estimate of the count of robbery and homicide together was statistically identical to that of robbery alone, suggesting that the BID adoption effect observed is driven by the rate of robberies.

We also constructed alternative model specifications in line with previous work that Brooks5 used on an analysis of the effects of BID on reported serious crimes in Los Angeles during earlier years. We compared the estimated effect of BID implementation on robberies and all reported violent crimes using all police reporting districts in Los Angeles as the unit of analysis, including those that do not intersect BID areas. We then included a dummy variable denoting the timing of BID, adjacent neighbourhoods to BID as control variables, and fixed-effect terms (dummy variables) for each individual reporting district, year, and their interactions (reporting district*year). Our results from these specifications were sensitive to the parameterisation of the outcomes. If we relied on ordinary least squares regression we found a statistically significant BID effect in reducing the mean incidence of robberies (b=−4.63; p<0.001) and all violent crimes (b=−7.33; p<0.001) by approximately 16% and 11%, respectively. However, if we relied on a Poisson regression model the results were substantially lower and were only marginally significant for robberies (b=−0.02; p=0.07) and total violence (b=−0.01; p=0.09), reducing the respective incidence by 3% and 2%. We think this sensitivity test provides further justification for our use of a simpler model that estimates only the BID effects for those areas that eventually adopted BID, rather than assigning BID effects to the entire city of Los Angeles.

The Bayesian hierarchical model provides an estimate of the effect of BID on the incidence of robbery and violent crimes in areas that were exposed to BID. Like all models this approach has several limitations. First, the model assumes that the population at risk of violence does not change once a BID starts. It is, however, possible that the establishment of a BID could change the population at risk of violent crimes in a number of ways. If, for example, there is a substantial increase in the number of shoppers or new residents because a BID was implemented then even a substantial decline in the incidence of violence would be offset by an increase in the denominator for the population at risk. Such an increase in the population at risk of violence would lead one to conclude that the adoption of the BID did not have an effect in reducing violence. Assuming that the population at risk coincides with the residential population in an area would be incorrect, as it is almost certain that successfully implemented BID attract a larger number of people for commerce to areas.

Second, the model assumes that the level to which violent crimes are reported to the police does not change systematically with the adoption of a BID. If, however, the adoption of a BID implies an increase in local merchants' and residents' willingness to report crimes to the police and an increased response from the police to combat crime, then the violent crime reports may actually increase as a function of the implementation of a BID. If this were the case, the adopted model would suggest that adopting a BID has the effect of increasing the incidence of violent crimes. Given that the findings suggest an overall effect of BID on reducing the robbery rate and marginal effects for all violent crimes, we have some confidence in these results.

In addition, if one assumes that BID areas have unique features in terms of the businesses that operate and the communities that encourage their establishment, constructing a group of comparison areas that are matched to the BID areas with respect to certain demographic features of area residents would represent a less conservative test of the effects of BID. We think that the areas that will eventually adopt a BID are the best comparison group for those areas that have already adopted a BID, as there are clearly features of BID areas that are unique in their ability to get a majority of landowners and merchants interested in their adoption. At the same time, our analysis offers no prescription for the various mechanisms by which BID impact robbery rates. BID adopt a variety of tactics, such as mobilising the police, hiring private security officers, street cleaning and environmental redesign to increase a sense of cleanliness and safety of BID areas. Unfortunately, the tactics adopted by each BID area are complex and not easy to approximate in a statistical model.