The rate of homicides caused by police in America is around
3.3 homicides per million people annually. I calculated this rate by taking the
total number of police-induced deaths (including via Taser and physical force)
I’ve cataloged in the Lethal Force Database and dividing by the U.S.
population, then dividing by the number of years of data I’ve collected (1.83
at this point).
3.3 people killed by police per million Americans is not a
geographically constant rate. There are maximums and minimums. In the north and
east, the rate is low, and the rate tends to increase as one travels southward
and westward. Just one person was killed by police in the state of Rhode Island
between January 2014 and October 2015, so Rhode Island had the lowest
police-induced homicide rate in the nation (0.5 homicides per million
residents), while New Mexico, which had double the population of Rhode Island
but 35 times more police-induced homicides, had the highest rate (9.2 homicides
per million residents).
Breaking it down into America’s 382 metropolitan statistical
areas reveals even more heterogeneity in the police-induced homicide rate. The
deadliest location in the United States was Grants Pass, Oregon, with a rate of
19.7 people killed by police per million residents per year, six times the
national rate. At the other end, 95 metro areas enjoyed 22 months without a
single officer-involved fatality; the largest of these metro areas was Buffalo,
New York.
What are the characteristics of metro areas with high rates
of fatal officer-involved shootings? Why are some places prone to repeated
tragedies, like Oklahoma City, Albuquerque, and Salinas, California, while
others rarely see a fatal police shooting, like Buffalo, Toledo, or
Fayetteville, Arkansas?
I looked at 35 measures of metropolitan statistical areas
that I thought might have some kind of relationship with the police-induced
homicide rate. Then I downloaded a statistics program, JASP, to determine if
there was any correlation between the rate at which people are killed by police
and any other characteristic of a metro area. (My knowledge of statistics comes
from one undergraduate level course in engineering statistics, plus what I read
on the internet, so I will probably be making many errors in my descriptions of
statistics in this piece.)
The set of 35 characteristics of metro areas that I was able
to look at include deaths by homicide, suicide, drugs and alcohol from the
Centers for Disease Control and Prevention’s WONDER database, averaged over a
four year period between 2012 and 2015. I also split out the suicide and
homicide deaths by whether or not the death occurred through use of a firearm,
and I calculated the ratio of firearm deaths to all deaths for suicides and
homicides. I also included the murder rate and the violent crime rate from the
FBI’s Uniform Crime Reports for 2014, and I included 19 separate
characteristics from the U.S. Census American Community Survey, taken either
from 2015 alone or averaged from 2011-2015. These U.S. Census variables include
median household income, poverty rate, racial demographics, median age, fertility
rate, household size, housing characteristics, education, foreign language,
foreign-born population, and veteran status. In addition to these 2015
measurements, I also included the 2013 median household income and calculated a
median household income growth rate over two years for each metro area. And
finally, I also calculated the ratio of median home price to median household
income as a measurement of housing affordability in each metro area.
I used JASP to derive a Bayesian correlation matrix for each
of the 35 variables. I looked at both Pearson’s correlation coefficient, which
looks at a linear correlation, and Kendall’s rank correlation coefficient,
which looks at an ordinal association. JASP used a system to flag correlations
that were significant, with one star showing positive evidence of a
correlation, two stars showing strong evidence, and three stars showing very
strong evidence of correlation. One star was equivalent to a Bayes Factor BF10
of 10, two stars indicated a value for BF10 equal to 30, and three
stars indicated that BF10 equaled 100 or greater. The hypothesis
was that the variables were correlated, not that they were correlated
positively or correlated negatively.
I restricted my dataset to only the 104 metro areas with a
population of 500,000 or greater. Smaller metro area populations get a bit
noisy, and if an officer-involved shooting just happens to occur in the
22-month window of my dataset in a metro area like Cumberland, Maryland
(population 101,225), it can artificially send my calculated police-induced
homicide rate soaring (in the case of Cumberland, Maryland, to 5.4 deaths per
million, 70% higher than the national average. The chance that 5.4 deaths per
million is even within 50% of whatever the true rate is is not very likely, or
at least, not nearly as likely as a one-death rate in a metro area like
Syracuse, New York, (population 661,934) where I’m pretty sure that if the true
rate isn’t 0.8 deaths per million, it is very likely also very low.
Results
The strongest evidence for a Pearson’s rho linear correlation
was between the police-induced homicide
rate and the non-firearm homicide rate. Alcohol death rate and firearm suicide
rate were also very strongly correlated with rates of people killed by police.
|
Metro area statistic |
Pearson’s rho |
BF10 |
Evidence for correlation |
|
Non-firearm homicide rate |
0.437 |
5037.524 |
Very strong |
|
Alcohol death rate |
0.403 |
857.406 |
Very strong |
|
Firearm suicide rate |
0.399 |
706.087 |
Very strong |
|
All homicide rate (CDC) |
0.359 |
123.355 |
Very strong |
|
All suicide rate |
0.356 |
107.957 |
Very strong |
|
Violent crime rate |
0.350 |
85.432 |
Strong |
|
Firearm homicide rate |
0.324 |
31.477 |
Strong |
|
Murder rate (UCR 2014) |
0.316 |
23.994 |
Positive |
|
Ratio: firearm suicides to all suicides |
0.314 |
21.741 |
Positive |
|
College graduation rate |
-0.297 |
12.554 |
Positive |
The strongest evidence for a Kendall’s tau-b ordinal rank
correlation was between the police-induced homicide rate and the firearm
suicide rate, though the overall suicide rate and the non-firearm homicide rate
also showed very strong evidence of a correlation with the rate of
police-caused killings.
|
Metro area statistic |
Kendall’s tau-b |
BF10 |
Evidence for correlation |
|
Firearm suicide rate |
0.295 |
2184.977 |
Very strong |
|
Non-firearm homicide rate |
0.279 |
791.556 |
Very strong |
|
All suicide rate |
0.253 |
165.468 |
Very strong |
|
Ratio: firearm suicides to all suicides |
0.248 |
141.734 |
Very strong |
|
Violent crime rate |
0.248 |
126.251 |
Very strong |
|
Alcohol death rate |
0.236 |
63.603 |
Strong |
|
Median age |
-0.232 |
54.047 |
Strong |
|
Geographical mobility |
0.232 |
51.833 |
Strong |
|
All homicide rate (CDC) |
0.225 |
37.064 |
Strong |
|
Murder rate (UCR 2014) |
0.221 |
31.053 |
Strong |
|
Fertility rate |
0.212 |
19.235 |
Positive |
|
Firearm homicide rate |
0.205 |
13.948 |
Positive |
Suicide and Homicide Rates
The firearm suicide rate and the non-firearm homicide rate
both appear to be very important to the rate of people killed by police (KBP
rate).
Their counterparts, non-firearm suicide rate and firearm
homicide rate, also showed some evidence of correlation with the KBP rate, but
the evidence for a correlation was weaker.
What does this mean? Well, correlation does not equal
causation, so I can’t say for sure. But it would make sense that suicides would
correlate with the KBP rate. For the ten months of 2015 that I’ve analyzed so
far, I was able to classify 213 of the 908 incidents (23%) as a suicide-by-cop
situation, where the decedent had been feeling suicidal prior to the
intervention of police. (In 102 of these incidents, the police knew the person
was suicidal before arriving on the scene.)
And it would also make sense that specifically the rate of suicide by
firearm would correlate with the KBP rate because it is an indication of places
where guns in the home are more prevalent. This is borne out by the fact that
the ratio of firearm suicides to all suicides also showed a very strong
correlation with the KBP rate with regard to Kendall’s rank correlation
coefficient, and a weaker but positive correlation with regard to Pearson’s
correlation coefficient for linear fitness. The firearm suicide to all suicide ratio is
used in research as a proxy for firearm ownership due to the fact that publicly-funded
research into measuring the number of firearms in a community is not allowed by
congress.
However, I am not sure why the rate of homicides caused by
implements other than firearms would show such a strong correlation with the
police-induced homicide rate. It seems real. Of the top nine metro areas with a
population greater than 500,000 ranked by highest KBP rate, four of them also
appear in the top eight of metro areas ranked by non-firearm homicide rate (#1
Oklahoma City is #7, #2 Bakersfield is #1, #7 New Orleans is #6, and #9 Las
Vegas is #8). All of the top 10 metro areas with the highest KBP rate appear in
the top third of metro areas ranked by non-firearm homicide rate. Evidence for
a correlation with firearm homicides is positive to strong, though weaker than
the correlation with non-firearm homicides.
Violent Crime and Murder Rates
Samuel Sinyangwe of Mapping Police Violence made an
infographic that argued that the rate of police-induced homicides was not
related to violent crime rate in a city.
 |
| from mappingpoliceviolence.org |
Sinyangwe made a claim that “levels of violent crime in US
cities do not make it any more or less likely for police to kill people”. I
found that actually a metro area’s violent crime rate showed either a strong or
very strong correlation with a metro area’s rate of people killed by police.
It’s not so obvious when shown on a graph like the one Sinyangwe made, but
unless the correlation is close to perfect, that kind of graph is going to
appear to show randomly scattered data to the human eye. One key difference between
my investigation and Sinyangwe’s is that Sinyangwe only counts cities and not
metropolitan areas, which I believe allow for a fairer comparison between
cities with incorporated areas that only encompass dense urban places like St.
Louis and cities whose incorporated areas include both a dense inner city part
and more sprawling suburban parts, like Oklahoma City.
The murder rate as derived by the UCR for the year 2014
showed weaker evidence for correlation with the KBP rate for 2014-2015 than did
the homicide rate as derived by the CDC for the years 2012-2015. The violent
crime rate appears to be a better predictor of the KBP rate than the murder
rate.
Alcohol and Drug Rates
30% of the people killed by police officers from January
through October 2015 were known to be intoxicated with either drugs or alcohol
(or both) at the time of their death (272 out of 908). Many more probably were
intoxicated, but we will never know because officials don’t always release this
information. It is often the case that the actions taken by the intoxicated
person led police officers to react in a way that ended with the intoxicated
person’s death. It stands to reason that the rate of death by drugs or alcohol
would be correlated with the rate of death by police, if we can assume a
correlation between the death rate by alcohol and drugs to the usage rate of
alcohol and drugs. However, the alcohol death rate showed much stronger
evidence for a correlation with the KBP rate than did the drug death rate.
Race and Other Census Data
One thing striking to me about this analysis was that
evidence for a correlation between race and the rate of police-induced
homicides was very weak.
The percentage of a metro area composed of black people mostly
showed no correlation with the KBP rate, though if there was a correlation, it
showed a slightly decreasing tendency in the KBP rate (the Pearson’s rho was
-0.069, Bayes Factor of 0.155). This was surprising to me. Black people get
killed by police at a rate three times higher than white people (7.1 deaths per
million black people versus 2.5 deaths per million white people), so I expected
metro areas with higher numbers of black people to generally also be high in
the rate of people killed by police, but this wasn’t the case.
It is doubly
interesting when one considers the fact that the KBP rate was correlated with the violent crime rate. Sam Sinyangwe probably
made that graph at mappingpoliceviolence.org in response to a police argument
that the reason for the disproportionate rate of black people getting killed by
police was that black people tended to be the perpetrators of violent crime,
and officers have more deadly interactions with violent criminals. (I found
that the percentage of black people in a metro area was very strongly
correlated with the violent crime rate (Pearson’s rho 0.361, Bayes Factor
130.3)). Yet the evidence, at least for the largest 104 metro areas, shows that
the KBP rate is correlated with the violent crime rate and is not correlated with
the percentage of black people living in the metro area.
The correlation with the percentage of Hispanic people was
slightly more significant, but not significant enough to state with confidence
that it was not due to randomness (Pearson’s rho of 0.254, Bayes Factor of
3.477). The non-white percentage was slightly more correlated (Pearson’s rho
0.261, Bayes Factor 4.197) but not enough to claim that it was a significant
predictor of the KBP rate.
In fact, most of the census-based data showed no significant
evidence of correlation with the KBP rate. I expected that certain variables,
like travel time to work, household size, percentage of single unit housing,
and the percentage of rental occupied housing units would probably not show a
correlation with the KBP rate. But I was a bit surprised that the data did not
suggest that poverty rate, median household income, housing affordability,
foreign born population, foreign language use, or high school graduation rate
was significantly correlated with the KBP rate either. In fact, the KBP rate
was the least responsive to variability in the housing affordability variable;
the linear fit of the data is just a horizontal line.
The census variables that did show at least positive
evidence of a correlation were the college graduation rate, median age,
geographical mobility, and fertility rate; the first of these showed
significance in Pearson’s rho but not in Kendall’s tau-b, and the final three
showed significance in Kendall’s tau-b but not in Pearson’s rho.
What can we conclude from this? The rate at which people are
killed by police in a given metro area is generally affected by the metro
area’s guns, booze and crime, and not necessarily by what type of people live
in the metro area. But even after controlling for these metro area
characteristics, a lot of noise exists within the dataset. This is to be
expected. Police-induced homicides occur due to actions by individual police
officers and individual decedents, no matter what the alcohol-caused death rate
or whatever is like in their city. But this data does suggest that if somehow
we could reduce the amount of crime or the number of guns or abuse of alcohol
(but not necessarily drugs), there might also be an associated positive reduction in the
number of people killed by police officers.
 |
| Relationship between population and number of people killed by police |
