Measuring the Effectiveness of Private Security Patrols in Lower Rockridge
Paul Liu
First Version: October 22, 2013
This Version: February 1, 2014
The use of private security patrols is increasing in areas where traditional law enforcement has been insufficient to ensure public safety. In this report I analyze the effectiveness of private security patrols in the Oakland neighborhood of Lower Rockridge. The data show that robberies and burglaries in Lower Rockridge doubled from July to September 2013, against a generally flat to declining trend seen in other similar neighborhoods. But trends in October were similar, and the results in the three months since the patrols started in November suggest that crime in Lower Rockridge is down by a statistically significant 46% relative to what it would have been absent the patrols, 45% after accounting for possible displacement.
Background
On September 23, 2013, with the support of the neighbors on my block and the adjoining block I launched a crowd-funding campaign for a private security patrol in Lower Rockridge, a neighborhood in the city of Oakland, CA. The specific area I proposed for the patrol was the area north of Highway 24, west of College Ave, east of Telegraph Ave., and south of Woolsey St (“Lower Rockridge North/West” in Figure 1 below). Subsequent to the announcement of my campaign, two other campaigns were formed in immediately adjacent neighborhoods (“Lower Rockridge South” and “Lower Rockridge North/East”).
Figure 1: Lower Rockridge Patrol Areas
Each of the CrowdTilt campaigns successfully funded. Subsequently, the patrol for Lower Rockridge North/East was combined with North/West (forming Lower Rockridge North) and the boundary for Lower Rockridge South was slightly extended. Patrol services began on November 4, 2013 and will run through at least February 2014. Many in the community have asked whether these private security patrols will be effective; this report aims to answer that question by analyzing the available crime data.
Data
Oakland Police Department provides crime data to the public through the website Crimemapping.com. Crimemapping.com also contains crime data for Berkeley, CA. Besides providing maps and summary charts, it can list individual crime data for the past six months for either the area displayed on the map or for a radius around a specified address. The data fields are as follows:
Crime type/Crime description (e.g., “robbery - firearm”, “burglary - forcible entry”)
Case number
Location
Agency
Date and time
Note that the number of affected persons is unavailable, so these statistics will understate the impact on the community of recent crimes such as the casual carpool robbery involving up to 20 persons. In addition, while downloaded data for Berkeley contain block address for location, downloaded data for Oakland do not: this is per the Oakland Police Department’s configuration. Fortunately, it is sufficient to place crimes within circles of a given radius. The Lower Rockridge patrols can reasonably be captured through the union of four circles (three of radii 0.5 miles and one of radius 0.2 miles). By defining these circles, we can capture essentially all crimes on the Crimemapping.com website within the Lower Rockridge patrol areas in the past six months. Then one can compare crime in the patrol areas with a set of “control” areas. Control areas should typically be areas which are otherwise comparable to the areas of interest but do not have the proposed treatment. I include in my set of control areas the following neighborhoods, which have a roughly similar mix of residential and commercial use:
Elmwood (Berkeley)
Temescal (Oakland)
Piedmont Ave. (Oakland)
Grand Ave./Lakeshore Ave. (Oakland)
Each of these neighborhoods was defined using a 0.5 mile radius circle entirely outside of the patrol area. To the extent that these circles overlapped with the circles of the Lower Rockridge patrol areas, any crimes in the overlapping area were excluded from the count for the given Lower Rockridge patrol area. The map of these areas and their relation to the Lower Rockridge patrol areas is shown in Figure 2 below.
Figure 2
I focus the analysis on robberies and burglaries, as these are the crimes that are of greater concern to residents in the Lower Rockridge area. More serious crimes such as homicides are not generally an issue in the Lower Rockridge area.
Results
Weekly results for each neighborhood are shown in Figure 3. Robberies and burglaries in Rockridge, both North and South, roughly doubled from July to September 2013. The vertical red line indicates the start date for the Rockridge patrols.
Figure 3
If we aggregate the other areas to compare against Lower Rockridge as a whole, then we obtain the results in Figure 4. Because the Lower Rockridge area is defined by approximately 2.8 circles of 0.5 mile radii and the other control areas in the aggregate are defined by approximately 3.8 circles of 0.5 mile radii, I normalize the counts, after de-duplication by case number and date, by the area to get comparable figures per square mile. One can again see the doubling of robberies and burglaries in Lower Rockridge from July to September 2013, while the control consisting of the other areas in the aggregate declined by about half in the same period.
Figure 4
In the period from the last part of September through the end of October, the rolling four-week count of robberies and burglaries was quite similar, enough that I think it reasonable to use the last data point (for the four weeks from October 7 to November 3) as a “before” baseline from which to compare subsequent changes in crime in the treatment and control areas. Note that it does not make sense to use the entire 5.5 month period preceding the patrols as a baseline. Whether the increase in crime over the summer was the result of statistical noise can be debated but the fact is that it was real and palpable to those living in the community. As such, it makes no sense to compare crime levels after the patrols started with crime levels before or as crime was increasing, which is what would be the case if the entire 5.5 month period were used.
Comparing the three months since the patrols started to the four weeks prior for a difference-in-difference calculation gives the following calculation:
Table 1: Impact Calculation, Overall
Weekly Average Burglaries/Robberies,
Per Square Mile
|
Oakland/ Berkeley Control
|
Lower
Rockridge
|
Before: Oct. 7-Nov. 3
|
3.10
|
3.06
|
After: Nov. 4-Jan. 26
|
3.85
|
2.04
|
Before/After Change
|
+24.3
|
-33.3%
|
Lower Rockridge Relative to Control
|
-46.4%
(=0.667/1.243-1)
|
From these data, one can see that Lower Rockridge has seen a 33.3% decrease in burglaries/robberies in the three months since the patrol started compared to the four weeks prior, while the control group has seen an increase of 24.3%. Comparing Lower Rockridge relative to the control group suggests that burglaries/robberies are down 46.4% from what they would have been absent the patrols.
A similar result is found when comparing the most recent four weeks to the four weeks prior to the start of the patrol. Lower Rockridge burglaries/robberies were down 60.0% while crime in the control group was down by 10.8%. This suggests that burglaries/robberies in the most recent 4 week period were down 55.2% from what they would have been absent the patrols.
Table 2: Impact Calculation, Most Recent 4 Weeks
Weekly Average Burglaries/Robberies,
Per Square Mile
|
Oakland/ Berkeley Control
|
Lower
Rockridge
|
Before: Oct. 7-Nov. 3
|
3.10
|
3.06
|
After: Dec. 30-Jan. 26
|
2.76
|
1.22
|
Before/After Change
|
-10.8%
|
-60.0
|
Lower Rockridge Relative to Control
|
-55.2%
(=0.40/0.892-1)
|
Some have raised the prospect that crime may have been displaced to other areas. To estimate the effect of displacement on the difference-in-difference calculation, recall that to the extent that criminals moved to other areas, the area easily accessible by car in this region is quite broad. An area ranging from University Avenue in Berkeley south to Fruitvale Avenue in Oakland is all within a ten minute drive of Lower Rockridge. In the four-week baseline period of October 7 to November 3, CrimeMapping.com shows a total of 424 burglaries and robberies in this broader area. See Figure 5 below.
Figure 5
Note that CrimeMapping.com does not include data from the city of Piedmont, so the actual total would be slightly higher. This area includes the 25 burglaries and robberies observed in Lower Rockridge. With a 46.4% reduction in crime, I estimate there would have been 13.4 burglaries and robberies had the patrols been in place during the baseline period. The difference is 11.6 burglaries and robberies. If these burglaries and robberies were all displaced instead of deterred and spread evenly across the broader area, crime in the broader area less Lower Rockridge would have increased by less than 3%. Adjusting the control area for this maximum displacement gives similar results as before, with an overall three month reduction of 44.8% and a most recent four week reduction of 53.8%.
Thus far I have made no mention of statistical significance. This is in part intentional -- in this case we are working with the population data, not a sample as is typically the case, so whatever one might say about the randomness of crime, the fact is that the number of burglaries and robberies declined in Lower Rockridge relative to the control areas by a substantial amount. But was this decline just noise? This is where statistical modeling can step in to assist. Crime data are by their nature relatively low frequency count data, so the natural modeling to use is to fit a Poisson regression model to the data. Using the weekly counts of burglaries and robberies for the treatment and control areas, from four weeks prior to the start of the patrols through the last full week in January, I estimate the following model:
burglaries+robberiestreat,aft=Poisson(+treat+aft+treat*aft)
where treat = 1 for an observation in the treatment area, 0 otherwise, aft = 1 for an observation after the start of the patrols, 0 otherwise, and the coefficient for the treatment * aft variable represents the “difference-in-difference” estimate of the impact. The results are in Table 3 below:
Table 3: Poisson Regression Results
Poisson regression Number of obs = 32
LR chi2(3) = 45.19
Prob > chi2 = 0.0000
Log likelihood = -89.350566 Pseudo R2 = 0.2018
------------------------------------------------------------------------------
burgrob | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
treat | -.3920421 .2588958 -1.51 0.130 -.8994685 .1153843
aft | .2177235 .1851308 1.18 0.240 -.1451262 .5805731
treataft | -.6231886 .3070397 -2.03 0.042 -1.224975 -.0214018
_cons | 2.224624 .164399 13.53 0.000 1.902407 2.54684
------------------------------------------------------------------------------
Note that the “Pseudo R2” is not particularly of interest in this context; the relevant variable is the z-stat, which is -2.03 for the impact variable “treataft”, indicating statistical significance at the usual 95% confidence level (this is the case for z-stats greater than 1.96 in magnitude). This can be confirmed as well by observing that the 95% confidence interval in Table 3 does not contain zero for the coefficient to the “treataft” variable.
Translating Poisson regression results is most easily done by calculating the difference in prediction results between treataft=0 (“notreat”) and treataft=1 (“withtreat”). The results are in Table 4 below:
Table 4
Variable | Obs Mean Std. Dev. Min Max
-------------+--------------------------------------------------------
notreat | 32 9.163851 1.990199 6.25 11.5
withtreat | 32 4.913949 1.067208 3.351449 6.166667
Then the impact can be calculated as withtreat/notreat -1, or 4.91/9.16-1 = -46.4%, consistent with the point estimates found earlier in Table 1.
Appendix: Data Details
I used the following criteria for pulling the data from Crimemapping.com. The specific addresses are not significant in any way; they simply formed the most convenient center for the circle desired. The data were initially pulled on October 13, 2013 and pulled again on November 12, November 19, November 27, December 16, January 2, and January 28.
Treatment areas:
Lower Rockridge North/West:
- 0.5 mile radius circle around 6075 Colby, excluding anything in controls
Lower Rockridge North/East:
- 0.5 mile radius circle around 6400 Harwood, excluding anything in controls
Lower Rockridge South:
- 0.5 mile radius circle around 400 Clifton, excluding anything in controls
- 0.2 mile radius circle around 500 Taft, excluding anything in controls
Control areas:
Elmwood:
- 0.5 mile radius circle around 2750 Stuart
Temescal:
- 0.5 mile radius circle around 4200 Webster
Piedmont Ave.:
- 0.5 mile radius circle around 4100 Piedmont
Grand/Lakeshore:
- 0.5 mile radius circle around 3600 Grand
No comments:
Post a Comment
Be nice!