Relationship between Police Killings, City Spending, and Violent Crime
An exploration on public investments and police brutality
In the United States, police officers are responsible for an average of 1,101 reported deaths per year since 2013. Protests against police brutality and excessive use of force were a landmark moment in 2020, spurred by the death of George Floyd. One of the key demands by protesters is for policy reform to defund the police; the idea is that reallocating funds from a stretched thin, overly violent police force to improving public and emergency services serving areas such as mental health, affordable housing, and addiction would simultaneously reduce crime and police killings.
A 2006 study by the Bureau of Justice Statistics found that police recruits spend more time training firearms skills and self-defense than any other skill, with a median training time of 60 hours and 51 hours respectively; the next highest median was for health and fitness training at 46 hours.¹ Police officers respond to a wide variety of emergencies which may involve domestic violence, drug use, homelessness, or mental illness. If their training is focused on use-of-force then they are not appropriately equipped to deal with such a wide range of situations.
Estimating the effect of police on crime has been a challenge for empirical studies in criminology, and the extensive literature has varying conclusions on the effect of police presence.² Many studies look at the number of officers or police patrol presence on the effect of crime,³ but instead of using police presence, I use police expenditures as an indicator of a city’s focus on bolstering the police force.
Critics against defunding the police claim that a divestment from the police would embolden would-be criminals and increase crime. The scope of this analysis is to examine expenditure factors and their relationships with police killings and violent crime through exploratory analysis and fixed-effects modeling with stepwise feature selection.
Data Sources
Lincoln Institute of Land Policy Fiscally Standardized Cities Database
Contains revenues and expenditures for government services including 150 of the largest cities in the US from 1977–2017.
FBI Uniform Crime Reporting
Instances and rates per 100,000 people for crime voluntarily reported by police departments, this work uses 1985–2014.
Mapping Police Violence
Includes over 8,000 police killings from 2013–2020; compiled from three crowdsourced databases.
Data Processing
For this work, crime rates reported by sheriff departments were removed because they represent counties instead of cities. The government expenditures include those from city, state, and federal government expressed in real per capita dollars (2017). Spend was combined into 5 categories: education, health, transport, policing, and environment/housing. There are 119 and 138 cities in the combinations of the UCR/FiSC and MPV/FiSC datasets respectively. Because the UCR data is voluntarily submitted by police agencies there could be some reporting bias.⁴
Exploratory Analysis
In this section, we’ll go through brief analyses of the individual datasets before delving into the merged data.
Police Killings between 2013 and 2020
From figure 1 we can see that police killings have been relatively constant since 2013. The months with the least and most killings were February, 2013 and May, 2020, coincidentally the same month as the death of George Floyd. On average, there are more killings in summer and spring months compared to fall and winter months.
The histogram in figure 2 tells us that young adults are killed by police at higher rates than other age groups. Victims between 21 and 35 make up 44% of all police killings. Though the majority are young adults, we can see that there doesn’t seem to be any age group that totally evades the brutality of our police force.
The bar chart in figure 3 details the 10 police agencies with the most reported killings. These 10 agencies account for 9.6% of all police killings in the time frame. Alarmingly, LAPD and LA County Sheriff’s Department are both in the top 3 agencies. This list isn’t too surprising, they’re mostly populous urban centers, but Phoenix with the second most killings makes me question their training techniques considering NYC has more than 5 times its population.
Government Expenditures between 1977 and 2014
Figure 4, on the left, displays average per capita spending as a percentage of total expenditures for different government services from 1977 to 2014. Education is firmly on top with the highest percentage of spend at an average of 34.56%. Police spending has steadily increased from an average of 6.74% in 1977 to 8.98% in 2014. On average, cities seem to be consistent in how they prioritize their allocation of funds.
Figure 5 visualizes the coefficient of variation (CV), which is used as a measure of dispersion to represent variability from the mean. Higher values of CV indicate more variation in a variable between cities in a particular year. For example, the CV for healthcare spending has increased over time which means that between cities there is a large disparity in how much they spend as a proportion of their total expenditures. Overall, cities are not varying much over time in how they are allocating funds; policing and education spending in particular have been highly consistent.
The below bar chart shows the ten cities which spend the highest average proportion of their funds on policing over the time frame. Florida has three cities in the top 10, with Ft. Lauderdale spending 0.8% more of their budget on police than the next highest city.
Crime Rates between 1985 and 2014
Violent crime rates have decreased 48.42% from the peak of 1,074 violent crimes per 100,000 people in 1993 to 520 violent crimes per 100,000 people in 2014. On the whole, the United States is safer than it ever has been in regards to crime.
The following scatterplot indicates a strong association between violent crime rates and population. It is interesting that crime rates are higher in densely populated cities, yet they don’t appear among those which spend a high proportion of their expenses on policing. Remember that an association does not mean that high populations cause violent crime or the other way around. (NYC and LA with average populations of 7.76 and 3.66 million and average crime rates of 1,262 and 1,410 respectively were removed from the plot for viewability, but they still provide weight to the trend line).
Crime Rates and Police Killings Merged with Expenditures
The below plots are a result of combining the datasets; they reveal associations between crime rates, police and education spending, and police killings. The merged data between crime rates and spending and killings and spending 1985–2014 and 2013–2017 respectively. To consolidate multiple time series, I use the average values for each city within the specified date ranges.
Figures 9 and 11, on the left, show that, on average, a city’s police spending, as a ratio of its total expenditures, has a positive relationship with both violent crime rates and police killings with correlation coefficients of 18.68% and 16.09%. Meanwhile, figures 10 and 12 indicate the reverse is true for education spending with correlations of -31.60% against crime rates and -24.39% against killings. This tells us that we can expect to find that cities that spend more of their budget on education have lower violent crime rates and instances of police killings.
There are a few key notes to take away from this exploratory analysis.
- Since 2013, police killings have been consistent every year
- Cities are relatively constant over time in their allocation of funds
- Police spending has a positive association with violent crime and police killings while the inverse is true for education spending
- Violent crime has been reduced substantially
Empirical Analysis
Between-fixed effects models were used to estimate the effect of government expenditure allocation on violent crime and police killings. The between fixed effects model is estimated on city averages, thus removing information due to intracity variation over time, this is preferred over within-effects for longer time periods.⁵
Additionally, I used an implementation of k-means clustering for longitudinal data that clusters trajectories⁶ to create a variable that would account for the effects of different growth rates using average total per capita spend for the cities. The following table details the 4 clusters with their growth rates from 1985–2014, average spend, and cities contained within the cluster. Cluster 4 consists of only one city, Washington D.C., which more than doubled its spend.
The model in figure 14 uses per capita spending on education, welfare, policing, and environment/housing as regressors for violent crime rate. Interestingly, there is a positive association for education, policing, and environment/housing, though the strength of the police spending variable is much higher. On average, for every $1 increase in police spending per capita there is an associated increase of 1.89 violent crimes per 100,000 people. Also, welfare spending has a negative association where every $1 increase in welfare spending per capita there is an associated decrease of 0.86 violent crimes per 100,000 people. The created cluster variable was not included in this model as it would introduce multicollinearity because the independent variables are per capita spending instead of a proportion of total spend.
The next model uses spending rates instead of per capita spending, so now the clusters can be included. Education spending did not prove to be statistically significant at any level. We see a negative association for crime rate with welfare and transportation, and, once again, police spending has a positive relationship with crime rates. On average, if a city spends 1 percentage point more of its budget on police then there is an associated increase of 69.49 violent crimes per 100,000 people. The clusters indicate that the cities with more average spend and higher growth rates are associated with higher violent crime rates.
Now killings are used as the dependent variable instead of crime rate. The only per capita spend item that has statistical significance at any confidence level is police spending. This model can be interpreted as, on average, every $10 per capita increase in police spending there is an associated 0.1 increase in police killings. Initially, this may not seem like a practically large effect, but when it is weighed in human lives it is significant. Also, because police spending is the only statistically significant raw expenditure, I believe this model to be useful.
Applying the natural log transformation to both variables from the previous model finds that, on average, a 1% increase in police spending per capita is associated with a 1.20% increase in police killings. To apply the transformation, a constant of 0.5 was added to the killings so that cities with no killings in a year would still be included in the model.
Our final model uses percentages of total expenditures to predict the natural log of killings. The results show that education is negatively associated with the log of killings, while welfare and police have a positive association. On average, a 1 percentage point increase in police spending as a percentage of all expenditures is associated with a 7.0% increase in killings. While welfare has a positive association, it is possible that this is related to a discussion on policing around class and race, which is outside the scope of this work.
For brevity I did not discuss the full model-building process, if you’re interested in how they were built please reach out. The primary takeaways of the empirical analysis are thus:
- Police spending, both per capita and as a percentage of expenditures, has a significant positive relationship with both police killings and violent crime rate
- Cities that have grown more quickly have higher crime rates
- Welfare spending has a negative relationship with violent crime rates and a positive one with police killings
Conclusion
This work sought to explore the effect of defunding the police in favor of other government services. Exploratory data analysis as well as between fixed effects models were used to explore the relationships between local expenditures, police killings, and violent crime. Police funding was found to have positive relationships with both violent crimes and police killings, and even when other services have a positive relationship, the effect of police spending outweighs them. The results of this analysis argue that defunding the police may be a viable solution to reducing crime as well as police killings.
Limitations include lack of police killing data and the voluntary nature of UCR. If all police departments were mandated to submit data on all crimes as well as their instances of killings and excessive force then more solutions could be explored. Another limitation is that this work does not include data from county departments, further research should include a county level analysis. Additionally, future work should involve introducing variables such as diversity index and population densities.
This is a general spending analysis, a more focused examination comparing cities such as Portland, where they have just implemented a mental health first response team⁷, may provide a strong argument for defunding the police in favor of such alternatives.
Citations
[1]: Reaves, B. (2009). State and local law enforcement training academies, 2006 (USA, DoJ, Office of Justice Programs). Washington, DC: U.S. Dept. of Justice, Office of Justice Programs, Bureau of Justice Statistics. https://www.bjs.gov/content/pub/pdf/slleta06.pdf
[2]: Levitt, Steven. (1997), Using Electoral Cycles in Police Hiring to Estimate the Effect of Police on Crime, American Economic Review , 87, issue 3, p. 270-90, https://EconPapers.repec.org/RePEc:aea:aecrev:v:87:y:1997:i:3:p:270-90 .
[3]: Di Tella, Rafael, and Schargrodsky, Ernesto. (2004), Do Police Reduce Crime? Estimates Using the Allocation of Police Forces after a Terrorist Attack, American Economic Review, Vol. 94, №1, pp. 115–133
[4]: Levitt, S. D. (1998). The relationship between crime reporting and police: Implications for the use of uniform crime reports. Journal of Quantitative Criminology, 14 (1), 61–81. https://doi.org/10.1023/A:1023096425367
[5]: Croissant, Y., & Millo, G. (2008). Panel Data Econometrics in R: The plm Package. Journal of Statistical Software, 27 (2), 1–43. doi: http://dx.doi.org/10.18637/jss.v027.i02
[6]: Genolini C, Ecochard R, Benghezal M, Driss T, Andrieu S, Subtil F (2016) kmlShape: An Efficient Method to Cluster Longitudinal Data (Time-Series) According to Their Shapes. PLoS ONE 11(6): e0150738. https://doi.org/10.1371/journal.pone.0150738
