lighting crime: an empirical investigation of the impact of ... city’s 30,333 cobra head...
TRANSCRIPT
July August September. October November December
Lighting Crime: An empirical investigation of the impact of an LED streetlight conversion project on crime rates in Oakland, California
Monica Testa University of California, Berkeley | College of Environmental Design CP204c | Fundamentals of GIS in City Planning | Spring 2014 Prof. John Radke | GSIs: Lauren Heumann & David Von Stroh
Motivation: In June 2013, Oakland, California began the rollout of a citywide streetlight retrofit project intended to
enhance public safety by providing a better lighting quality. The city’s 30,333 cobra head streetlights will be
converted from High Pressure Sodium (HPS) fixtures to Light Emitting Diodes (LEDs). The new LED
fixtures provide a directional, broad spectrum white-light which renders color more accurately than the
yellow spectrum light emitted by the HPS lamps they are replacing. The improved visibility associated with
the LED technology is anticipated to deter crime by increasing the likelihood that the perpetrator may be
seen and subsequently caught. By comparing Oakland’s LED rollout to changes in crime rates, this study
tests the hypothesis that LEDs reduce crime.
Modeling Techniques: Processing, modeling and integrating raw data: geo-referencing and spatial joins
• Geo-reference LED installations and reported crime for 2012 and 2013
• Spatial join: LED installation points by month and reported crime by month to census blocks
Vector-based Overlay and Neighborhood Modeling of Discrete Objects: erase, spatial joins, & unions
• Erase water polygons from census block shapefile, join monthly LED installation point data to census
blocks, creating a count of LEDs on each block each month, union monthly LED installation polygons,
generate map of LED rollout over time. Repeat steps using Crime data. Union LED & Crime datasets.
Quantifying Relationships & Analyzing Patterns: spatial statistics
• Ordinary Least Square (OLS) Regression: Is there a relationship between the change in opportunistic
crimes and LED installations areas?
• Spatial Autocorrelation (Global Moran’s I): Are the under- and over- predictions randomly distributed?
• Hot Spot Analysis (Getis-Ord Gi*): Is there a spatial pattern associated with the under- and over-
predictions?
Modeling the 3rd Dimension
• Use ArcGlobe to represent both spatial and temporal dimensions of crimes.
Implications and Conclusions: This investigation found no correlation between the installation of LED Fixtures and variation in crime.
This is surprising considering that the City of Oakland, California explicitly stated the “primary goal [of the
LED Streetlight Conversion Project] is to support public safety with brighter and better lighting.”
Furthermore, there was very little change in monthly opportunistic crimes in between 2012 and 2013 and
no change whatsoever when normalized by population. This is perhaps even more surprising in the context
that Oakland had one of the highest crime rates in the country in 2012 and was on track to top the charts
again in 2013, until the Oakland police department re-launched Operation Ceasefire to target violent crime
East of High Street, hired more officers, and began making raids as of mid-2013. Although Ceasefire
focused on violent crime, the increase in police activity appears not to have had any impact or spillover
effect on opportunistic nighttime crime, despite being concentrated in the same general area as the LED
installations.
Limitations: There are a number of limitations to this study which warrant further investigation. These include the
limited timeframe of the project and the limited number of crime records which were successfully geo-
coded. Including a greater number of records reflecting a longer period after the LED Street Conversion
Project has completed would increase the reliability of the results.
LED Fixture HPS Fixture
Directional white light Dispersed yellow light
LED Installations Completed in 2013 Assumptions, Data Selection, and Modeling Decisions: To model the installations over time and
create LED treatment areas, each lamp was first geo-referenced based on recorded street address.
28,207 of the original 30,333 records were successfully geo-referenced. The LED point shapefile was
then joined to a block shapefile, generating a count of the total lampposts on each block. Next, using
the select by attributes feature, LED points were selected by installation date and new shapefiles were
created from each selection representing LED installations completed by the end of each month. I
then added a field to the attributes table and calculated the percentage of converted streetlights by
dividing the number of LEDs installed by the total number of lamp posts successfully geo-referenced.
Using the total number of LED installations as a proxy for total streetlights was a reasonable
assumption, as all cobrahead streetlights in Oakland will ultimately be converted from HPS to LED
as part of the conversion project. I created a binary variable indicating LED treatment areas if blocks
had ≥ 75% LED installations and repeated the process for each month. This was based on the
assumption that the LED streetlights would only produce an effect on crime once each block was
close to completely lit, whereas partial conversion or intermittent lighting hot spots may produce a
different result. Finally, the 6 months of LED installation were unioned and LED indicator fields
summed to produce a map illustrating the rollout of LEDs overtime.
June July August September October November
Ordinary Least Square (OLS) Regression & Spatial Autocorrelation (Global Moran’s I) of Regression Residuals. This brings us to the moment of truth: is there a relationship between the change in opportunistic crime between 2012 and 2013 and the LED installations? Are the over- and under-predictions of changes in crime randomly distributed? The following maps model the relationship between the year-over-year difference in
crime counts. The percent of LEDs installed by the end of the previous month is treated as the independent variable and the opportunistic crime count of each block is treated as the dependent variable. The red and blue areas represent under- and over- predictions of the model respectively. Stated another way, red
represents blocks where more crimes occurred than the OLS regression model predicted and blue represents blocks where less crime occurred than predicted.
Visualizing Spatial & Temporal Dimensions of Crime Using 3D Analysis
Results & Findings LED Rollout
One of the key takeaway from the map of the LED rollout is
that the fixtures appear to have been rollout out somewhat
randomly. The spatial pattern may have been influenced by the
circuitry or another technological consideration, however, the
variation Although a general South to North trend emerges, the
time at which each block reaches 75% or greater LED
installations complete varies widely. This is important to
consider when using the installation of LEDs as an experimental
treatment.
Monthly Variation in Crime
Average monthly variation in nighttime opportunistic crimes
ranged from a decrease of 2 crimes per block per month up to
an increase of 2 crimes per block per month. Three outlier
observations (values of 4, 4, and -6) were cleaned from the
dataset. Normalizing by population removed much of the
noise from the total count and revealed that there is virtually no
change in average monthly crimes per block per person, with
only a few exceptions.
OLS Regression and Spatial Autocorrelation
For all months, regression residuals are clustered. Given the Z-
scores far above 2.58 and p-values of essentially 0 for each
month, there is a less than 1% likelihood that spatial
distributions of over- and under- predictions of change in crime
are due to random chance. Furthermore, the adjusted R-
squared values from the OLS regressions fell well below 1% for
all months, indicating that the percent of LEDs installed on
each block explain less than 1% of the variation in crime for
that block. These results indicate that LEDs cannot be expected
to predict to influence crime in any measurable way. Instead,
the highly clustered pattern of the over- and under- predictions
of crime suggest that there are spatial influences other than
lighting responsible for driving the changing pattern of
opportunistic crimes committed.
Hot Spot Analysis (Getis-Ord Gi*)
The hot spot analysis revealed interesting results regarding the
spatial pattern of the under- and over- predictions in crime. For
every month, areas were identified where high and low crime
values cluster spatially. Particularly striking is that the over-
predicted areas where crime was significantly lower than
expected (represented in blue) stay relatively constant over time
and are consistently concentrated to the North-West of Lake
Merritt. This is in contrast to the under-predicted areas where
crime was higher than expected (represented in red) which vary
significantly from month to month and display drastically
different patterns. These patterns open many more
opportunities for analysis. One potential hypothesis for the
different patterns of high and low crime rates might be that the
areas with lower than anticipated crime may be most influenced
by changing demographic patterns which tend to evolve slowly
whereas the high crime areas may be driven primarily by
changing police beats which adapt quickly.
3D Visualization of both Space and Time
The intention of this graphic was to demonstrate that crimes
occur at different points in space, but also at different points in
time. Unfortunately, with so many points, this map may have
just added more confusion than clarity. One insight that can be
gleaned, however, is that a disproportionate number of auto-
burglaries appear to dominate the very end of the study period
in December, which likely reflects a problem with the data.
July August September. October November December
Quantifying Relationships & Analyzing Spatial Patterns
Monthly Crime Variation by Block (2012 – 2013)
Data Sources Crime Data: Oakland Police Department ftp:///crimewatchdata.oaklandnet.com; LED Data: Oakland City Administration Public Records Request records.oaklandnet.com; Census Data: Professor John Radke.
Assumptions, Data Selection, and Modeling Decisions: The main goal for this part of the model
was to create a visual illustration of the change in crime rates from year to year by month and
block. In this case, LEDs are assumed only to effect nighttime crimes when the lights are on
and only opportunistic crimes or crimes otherwise likely to be effected by enhanced visibility on
the street and sidewalks. Daily sunrise and sunset times were appended to the crime report
(using STATA), from which a night variable was created indicating whether each particular
crime occurred at night. Opportunistic crimes were selected from the total crime report dataset
including robbery, auto burglary, grand theft auto, petty theft, and vandalism.
117, 647 out of 252,367 crimes were successfully geo-referenced for 2012 and 2013, of which
8023 were opportunistic crimes which occurred at night. Individual point shapefiles of
nighttime opportunistic crime were created for each month of the study (July through
December) and joined to the census block shapefile with the LED installations by month,
generating a count of crimes on each block every month. To calculate the change in crime
counts, 2012 monthly crime by block was subtracted from 2013 monthly crime counts by block,
resulting in a year-over-year difference in total crimes by month. Each month was modeled
separately to account for correlations with weather and temperature.
For every block, the average monthly year-over-year change in crime following 75% or greater
LED installations for that particular block was calculated and then normalized by population.
Assumptions, Data Selection, and Modeling Decisions: The main goal for the 3D Analysis of crime was to illustrate both spatial and temporal dimensions of crime. This map was generated using 3D analyst, where a time lapse variable was
created for each crime point and treated as the z-coordinate in the 3D analysis. In other words, the height of each crime reflects the time passed since the beginning of the study period on July 1st, 2013.
Hot Spot Analysis (Getis-Ord G*): Spatial patterns of over- and under-predictions of crime. Hot Spot Analysis maps were developed from the regression residuals (the under- and over- predictions of crime) identified in the OLS regression. Hot spots represent statistically significant spatial clusters of high and low values, where higher than expected crime rates are red and lower than expected crime rates are blue.