Methodology: Comparisons within Outage Affected Areas - dssg/streetlights-crime GitHub Wiki

The aim of this study is to determine if reducing street light outage durations would lead to reductions in crime.

This is a difficult question to answer! There may be areas of the city that are both high crime and susceptible to outages. We can take raw correlations between outages and crime across the city, but if there are other factors that simultaneously cause outages and crime, then the correlations do not help in understanding how reducing street light outage durations affects crimes. Further, anticipating what factors affect outages and crime rates and controlling for these factors can be difficult.

We get around this problem by comparing the crime rate in each outage-affected block during the outage to the crime rate in the same block for periods immediately before and after each outage. This approach has the advantage that each outage-affected area is its own control, and we do not need to control for factors that affect crime rates across different areas of the city. Using comparisons from similar time periods also helps as crime rates can vary over time, particularly as the weather changes.

We compare the crime rate during each outage to the crime rate over a 30-day period before the outage and over a 30-day period after the outage.

Specifically, we use Poisson Generalized Linear Models (GLMs). Learn more about these models here. Poisson GLMs have good properties for studying count data, such as the number of crimes. Outage duration is used as an offset in order to model crime rates.

Each observation in our regression is a time period (either before, during, and after) / outage combination. Regressors in the model include an outage indicator variable and fixed effects for each outage-affected area to estimate the average difference within each area. We also examine a model that includes monthly indicator variables to control for the crime time trend. Estimates of the average percentage difference between the crime rate during outages and crime rate in the before and after periods can be determined using a transformation of the regression coefficient for the outage indicator.

Finally, we estimate Poisson GLMs for each community area of Chicago to determine for which community areas is the relationship between outages and crime particularly strong. Community areas are big enough geographic areas to be able to detect statistically significant differences in crime rates.