Reconciling Results on Racial Differences in Police Shootings
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Reconciling Results on Racial Differences in Police Shootings
By Roland G. Fryer, Jr.∗
Police use of force – particularly lethal data and that assembled by The Post or
force – is one of the most divisive issues Guardian is that it counts all OIS – fa-
of the twenty-first century. When an offi- tal or non-fatal – whereas The Post and
cer discharges a weapon, it’s a life changing Guardian data count fatal shootings only.
event. A single bullet – which weighs about VICE catalogs 4,400 individuals involved in
.02 pounds and is 10mm long – can end a OIS between 2010 and 2016. Two-thirds
life, erase a pension, or change the image of of the individuals shot by police survived.
those who are sworn to serve and protect. Twenty percent of people involved in OIS
To understand the nexus of race, criminal were unarmed – four times as large as
justice, and police brutality, academics and the rate of unarmed people documented in
journalists have begun to amass impres- The Post/Guardian data of fatal shootings.
sive datasets on Officer-Involved-Shootings Moreover, blacks were 55 percent of shoot-
(OIS). The Washington Post (2015) pro- ing victims, which is more than double the
vided important stylized facts about indi- share of the black population in these ju-
viduals killed at the hands of police officers risdictions. Black subjects shot by police
in one calendar year. Police fatally shoot were more likely to be shot during encoun-
nearly 1,000 individuals a year in the U.S. ters that began with a traffic or pedestrian
As the report argues, the incidents that stop. They were no more likely to be armed
have ignited protests – white police officers with a gun than whites but less likely to be
killing unarmed black men – represent less armed with any weapon (VICE, 2017).
than four percent of fatal police shootings. Among social scientists, the literature is
The Guardian (2015) collected similar surprisingly thin. Ross (2015) estimates
data in their much publicized series, The racial differences in county-specific relative
Counted, and tallied 1,134 deaths of Amer- risks of being shot by police, using crowd-
icans at the hands of police in 2015. sourced data from the US Police-Shooting
And, similar to The Post’s database, The Database. He estimates that the probabil-
Guardian emphasizes that although black ity of being black, unarmed, and shot by
males aged 15-34 make up 2 percent of the police is nearly 3.5 times larger than the
population they comprise 15 percent of all probability of being white, unarmed, and
deaths logged in 2015 – five times the com- shot by police.
parison number for whites. One in every In stark contrast, Fryer (forthcoming)
sixty-five – or 1.5 percent – deaths of a finds that, conditional on a police interac-
black man in the US is a killing by police. tion, there are no racial differences in OIS
Roughly twenty-five percent of blacks that on either the extensive or intensive mar-
were killed were unarmed, eight percentage gins.1 Using data from Houston, Texas –
points higher than whites (Guardian, 2015). where I have both OIS and a randomly cho-
More recently, VICE News collected data sen set of interactions with police where
on OIS from 47 of the 50 largest po- lethal force may have been justified but was
lice departments in the U.S (VICE, 2017). not used – I find, after controlling for sus-
The major difference between the VICE pect demographics, officer demographics,
encounter characteristics, suspect weapon
∗ Fryer: and year fixed effects, that blacks are 27.4
Harvard University and NBER,
1805 Cambridge Street, Cambridge MA 02138,
rfryer@fas.harvard.edu. I would like to thank Tanaya 1 Goff, et al. (2016) reports similar findings on the
Devi, Hannah Ruebeck, and Damini Sharma for application of lethal force, but their data is not available
exceptional research assistance. All errors are my own. for public use and thus cannot be included in this review.
12 PAPERS AND PROCEEDINGS MAY 2018
percent less likely to be shot at by police probability of being involved in an OIS)
relative to non-black, non-Hispanics. Inves- if individual i is black and Yi0 denote the
tigating the intensive margin – who shoots value of Y if individual i is white. Then, the
first in an encounter with police or how “race effect” for individual i is τ = Yi1 − Yi0
many bullets were discharged in the en- – that is, the difference in Y that can be at-
deavor – there are no detectable racial dif- tributed to an individual’s race. This quan-
ferences. tity is the proverbial “holy grail” – the pa-
The key takeaway suggested by The Post rameter that we are all attempting to esti-
(2015), The Guardian (2015), VICE (2017) mate but never quite do.
and Ross (2015) is that there are large Consider the following linear regression
racial differences in OIS that are evidence equation: Yi = β0 +τ Ri +β1 X1 +...+βn Xn +
of racial bias or, at the very least, problem- i . One can trivially demonstrate that this
atic. As the Post writes, “Although black regression recovers the “race effect” if any
men make up only 6 percent of the U.S. difference between the groups that we’d ob-
population, they account for 40 percent of serve in the absence of bias is zero. To see
the unarmed men shot to death by police this formally, write Yi as Yi0 + Ri (Yi1 − Yi0 ).
this year.” The Guardian begins their arti- Note: τ measures E[Yi |Ri = 1] − E[Yi |Ri =
cle remarking, “Young black men were nine 0]. Using these two facts:
times more likely than other Americans to
be killed by police officers in 2015,” and τ = E[Yi1 − Yi0 ] +
| {z }
VICE exclaims, “Black people were shot Race Effect
more often and at higher rates than people E[Y 0 |R = 1] − E[Yi0 |Ri = 0].
of any other races.” Ross (2015) concludes, | i i {z }
Selection Effect
“There is evidence of significant bias in the
killing of unarmed black Americans relative It is straightforward to see that estimat-
to unarmed white Americans.” ing Equation (1) recovers the race effect
Yet, these descriptive statistics, while when the selection effect is zero. This
poignant, do not prove racial bias. In- happens when Yi1 , Yi0 ⊥ Ri |Xi , which en-
deed, when I compare the different data sets sures E[Yi0 |Ri = 1] = E[Yi0 |Ri = 0]. In
described above using a common estimat- words, a linear regression can recover the
ing equation I conclude that the “evidence “race effect” if race is “as good as ran-
for bias” in some studies and not others is domly assigned,” conditional on the covari-
mainly a result of misspecified regression ates (X 0 s). When the selection effect is
equations, not inherent differences in the not zero then one can misinterpret the mis-
datasets. measurement of X 0 s (e.g. omitted contex-
tual variables) as racial bias. One way to
I. The Simple Econometrics of ensure that race is “as good as randomly
Estimating Racial Differences in assigned” is to implement an experiment
Officer-Involved-Shootings (see Bertrand and Mullainathan (2004) for
a classic example). None of the studies
Conceptually, estimating racial differ- mentioned above meet this threshold.
ences in OIS is very similar to the large
and well-established literature in labor eco- II. Analyzing the Literature
nomics designed to understand whether
racial differences in wages is due to discrim- In this section, using the above concep-
ination. tual framework, I analyze the growing liter-
Borrowing extensively from that litera- ature on racial differences in OIS on a set of
ture and simplifying to the extreme, let’s five rubrics: (1) Reliability of Police Data;
suppose there are only two races: black and (2) Representativeness of Data; (3) Endo-
white. Let Ri = 1 if individual i is black geneity of Police-Civilian Contact; (4) Rel-
and Ri = 0 if they are are white. And, sup- evant Counterfactuals; and (5) Quality of
pose that Yi1 denotes the value of Y (e.g. Covariates.VOL. 2 NO. ISSUE RACIAL DIFFERENCES IN POLICE SHOOTINGS 3 The brilliance of the datasets constructed Guardian additionally records the address by the staffs at The Post and Guardian is at which the shooting took place and the that they do not rely heavily on descrip- name of the law enforcement agency re- tions by police. Their goal is to provide sponsible. a more accurate number of American citi- To draw firm conclusions about the “race zens whose lives were ended as a result of effect” one needs to assume that once this a police shooting. Because of their exhaus- sparse set of variables is accounted for, race tive search process – which included news is “as good as randomly distributed” and reports, public records, and original report- that there are no other differences in sus- ing – the data seem representative for what pect behavior or any other potential contex- it is: a count of the fatal shootings by police tual factor that is important. With tremen- in America in a given year. The Post and dous respect to the herculean efforts of The Guardian documented more than twice as Post and Guardian for collecting new data many fatal shootings in 2015 as the annual on one of the most important social issues of average annual reported by the FBI over our time, this assumption is ludicrous and the past decade. one can make no such claims of any “race These data do not deal with the endo- effect” using this data.2 geneity of contact by police, but assuming A typical retort to the lack of rigor when blacks are more likely to be unfairly tar- employing descriptive statistics is to ob- geted than whites, the racial differences ex- serve the size of the racial difference and hibited in their data would only be larger posture that such large disparities are not if one did account for the fact that police- likely to be explained by differences in con- civilian interactions are themselves (poten- text. That type of argument is more about tially) biased. clinging to deeply held beliefs than clinging However, two of the most important to a valid research design. There is no way a rubrics for internal validity are not consis- priori to understand how much of the vari- tent with The Post and Guardian datasets ance in an outcome might be absorbed by – having a relevant counterfactual and the covariates without first analyzing the data. quality of covariates. Theoretically, The The VICE data has similar advantages Post’s data is rich enough to account for and disadvantages, but is more comprehen- several different potential counterfactuals. sive in that it includes both fatal and non- For instance, one could estimate the prob- fatal shootings from 47 of the 50 largest po- ability of being killed if unarmed and black lice departments in America – which is in relative to the probability of being killed itself phenomenal – but less representative if unarmed and white, after accounting for than The Post or Guardian datasets. Their the context of the fatal encounter. Yet, the set of potential covariates, which includes more relevant counterfactual – given it is whether or not the suspect was armed, sus- less likely to be plagued by selection into the pect and officer demographics, the nature sample – is to estimate when police decide of the stop, and number of officers, is quite to pull the trigger. For this, one needs to limited and they do not have any data on catalogue situations in which police would situations in which lethal force would have have been legally justified in using lethal been justified but was not used. force but chose not to – along with the race Recall, Ross (2015) estimates racial dif- of the suspect and other contextual factors. ferences in county-specific relative risks Both The Post and Guardian collect of being shot by police, controlling for manner of death, whether or not the sus- whether or not the civilian was armed, pect was armed, weapon type, demograph- and concludes that the results provide ics, and the city in which the shooting evidence of significant bias. An issue took place. The Post additionally collects whether there were signs of mental illness, 2 Nix et al. (2017), using the data in Post (2015), the threat level to officers, and whether the employ this approach – accounting for a small set of officer was wearing a body camera. The covariates and claiming the residual is “implicit bias.”
4 PAPERS AND PROCEEDINGS MAY 2018
with the crowdsourced U.S. Police-Shooting omit a very important margin.3 One may
Database used in Ross (2015) is that it discover no differences in use of lethal force,
omits large fractions of shootings that oc- conditional on an interaction, even if there
curred and it’s not clear which types are are large racial differences in the probability
recorded. To illustrate this, there are 10 of the types of interactions in which lethal
locations in Ross (2015) that are also in force may be used.
Fryer (forthcoming). Of the 606 officer- Understanding racial differences in the
involved shootings in these locations, be- probability of police interaction is difficult.
tween 2011 and 2014, noted in Fryer (forth- One has to account for differential expo-
coming), only 102 of them (17 percent) were sure to police, race-specific crime participa-
reported in Ross (2015). He cannot account tion rates and perhaps most importantly,
for the endogeneity of police-civilian con- pre-interaction behavior that civilians ex-
tact, but again, this is likely to only increase hibit. Ideally, one might set up a field ex-
estimates of racial differences in the county- periment that randomly assigns similar in-
specific relative risk. A strength of the Ross dividuals (across all physical dimensions ex-
analysis is that the crowdsourced data may cept race) to the vicinity of the same pa-
be less likely to be biased by police depart- trolling officers in a neighborhood and in-
ment reporting practices. struct them to behave identically. Condi-
To assemble the database utilized in tional on random assignment, identical be-
Fryer (forthcoming), I requested police nar- havior, and race-specific crime rates, any
ratives for all OIS from 10 large police differences in the probability of interaction
departments. Houston Police Department could be interpreted as racial bias in po-
also provided a random sample of narra- lice stopping behavior. Without ideal data,
tives for events that did not end in an OIS researchers often compare the racial distri-
but in which justifiable force could have bution of stopped civilians to the racial dis-
been used. Narratives were twice-coded by tribution of various “at risk” civilians that
research assistants for relevant variables. could potentially be stopped. Determining
the probability of an interaction is essen-
This database addresses many of the is- tially a search for the correct “risk set.” The
sues that plague the literature – there is a evidence in Fryer (forthcoming) offers the
relatively comprehensive set of contextual following rough rule of thumb – if one as-
variables (290, for instance, hand-coded sumes that police are non-strategic in their
from Houston arrest data), and the coun- stopping behavior, then there is clear bias.
terfactuals are less likely to be biased by Conversely, if one assumes that police are
selection into sample – but the data are not stopping individuals they are worried will
nearly as representative (only 10 police de- engage in violent crimes, the evidence for
partments) as data collected by The Post, bias is small.
The Guardian, or VICE. Indeed, the analy-
A second concern is the reliability of po-
sis in Fryer (forthcoming) is tantamount to
lice department reports. There are two
trying to estimate discrimination in a labor
types of potential bias. First, police of-
market by analyzing data from firms willing
ficers may bend the truth about the con-
to supply it to researchers. Put differently,
text of a particular interaction so as to
while the estimates in Fryer (forthcoming)
justify their own actions; for instance, in-
are more internally consistent, their exter-
dicating a suspect was threatening when
nal validity is unknown.
they were calmly following an officer’s com-
There are two important fault lines in mands. This type of bias is less of a concern
Fryer (forthcoming): (1) the endogeneity in Fryer (forthcoming) because the quali-
of police-civilian contact and (2) reliabil- tative results are identical whether or not
ity of police department reports. To the one includes contextual factors about the
extent that there are racial differences in
the probability of an interaction with po- 3 This is also true of all other datasets collected, but
lice, the data in Fryer (forthcoming) may is more of an issue when one does not find bias.VOL. 2 NO. ISSUE RACIAL DIFFERENCES IN POLICE SHOOTINGS 5
encounter recounted by police. A second Using the data collected by VICE this num-
type of bias is that officers may be more ber is 9.82, 15.93 when using the data in
likely to charge black suspects with crimes Ross (2015), and 14.87 when using the data
such as resisting arrest or attempted assault in Fryer (forthcoming).
on a public safety officer rather than mis- Perhaps the most striking finding is when
demeanors, relative to whites, for identical one replicates the analysis in Ross (2015)
behavior. This type of bias is an important across all five datasets: calculating the
limitation of Fryer (forthcoming) because it probability of being black, unarmed, and
implies that the counterfactuals coded from shot by police divided by the probability
arrest data may themselves contain bias. It of being white, unarmed, and shot by po-
is unclear how to estimate the extent of such lice. A quantity greater than one is consis-
bias or how to address it statistically. tent with racial bias. Using the data from
Ross (2015), this ratio is 3.28. Using the
III. Reconciling the Literature data from the Post database I get 6.20 and
5.99 if using the data in Fryer (forthcom-
Datasets created by The Post, Guardian, ing). In other words, if I ignore the detail
VICE, or the US Police-Shooting Database available in the Fryer data and simply re-
provide important descriptive statistics port the descriptive statistics reported in
about OIS in America, but have severe Ross (2015), I could conclude that the data
limitations for determining whether those provided evidence of even more racial bias
statistics are due to racial bias. Fryer than that reported in Ross (2015). Yet,
(forthcoming) is better suited for this type when using the simple statistical framework
of analysis, but whether and how much one that economists have used for more than a
should extrapolate from 10 cities willing to half century to analyze racial differences on
supply data is unknown. myriad dimensions – from wages to incar-
I replicate some of the key findings in ceration to teen pregnancy – the evidence
each of the publications discussed through- for bias disappears. The differences in re-
out with the data from the alternative data. sults on police shootings in America seem
That is, I recreate the descriptive statistics to be driven by differences in what qualifies
used by The Post and Guardian, VICE, and for a valid research design and not differ-
Ross (2015) using the data from each pub- ences in datasets.
lication and Fryer (forthcoming). This al-
lows one to better understand whether dif- IV. Conclusion
ferences across datasets are due to different
statistical decisions or inherent differences
The time has come for a national reckon-
in data. For instance, if the data used in
ing on race and policing in America. But,
Fryer (forthcoming) – the most detailed but
the issues are thorny and the conclusions
least representative – shows less bias when
one can draw about racial bias are fraught
one estimates the fraction black unarmed
with difficulty. The most granular data sug-
who are killed by police relative to the frac-
gest that there is no bias in police shootings
tion white unarmed, then this further calls
(Fryer (forthcoming)), but these data are
into question the generality of the data.
far from a representative sample of police
departments and do not contain any exper-
Table 1 presents these estimates. The de- imental variation. We cannot rest. We need
scriptive statistics presented by The Post more and better data. With the advances
(2015), The Guardian (2015), VICE (2017), in natural language processing and the in-
and Ross (2015) are remarkably consis- creased willingness of police departments to
tent across all datasets – including Fryer share sensitive data, we can make progress.
(forthcoming). For instance, The Guardian For those of us who desire a more per-
(2015) shows that the likelihood of black fect union, police use of force has become
men between the ages of 15-34 to be killed our Gettysburg. Of course, black lives mat-
by police, relative to all other races, is 9.82. ter as much as any other lives. Yet, we6 PAPERS AND PROCEEDINGS MAY 2018
Table 1—Reconciling Results on Racial Differences in Police Shootings
Post Guardian Vice Ross Fryer
(1) (2) (3) (4) (5)
Paper: Statistic
Washington Post: Percent of black civilians 39.56 34.25 52.82 38.71 43.84
among unarmed men killed by police
Guardian: Likelihood of black men 9.67 9.82 9.82 15.98 14.87
between the ages of 15-34 to be killed
by police compared to others
Vice: Percent unarmed in sample 9.45 20.42 21.31 28.29 20.50
Percent black civilians in sample 26.80 27.21 54.59 34.95 45.70
Ross (2015): Prob. of being black, unarmed 6.20 3.44 3.21 3.28 5.99
and shot by police relative to prob.
of being white, unarmed, and shot by police
In row (1), I calculate the likelihood as the ((Fraction of 15-34 black men killed by police/Fraction of 15-34 black men in population)/(Fraction of
everyone else killed by police/Fraction of everyone else in population)). In row (2), I calculate the percentage of black civilians among unarmed
men killed by police. In row (3), I calculate the percentage of unarmed civilians and black civilians in the sample. In row (4), I use the companion
program to Ross (2015), available in R, to calculate the county-level risk ratios. To calculate this statistic, I collapse each OIS-level dataset to the
county level (Post and Guardian) or the police department level (VICE and Fryer (forthcoming)). To collapse cities to the county level (Post and
Guardian), I assign cities to the county that they fall the majority in. The VICE and Fryer (forthcoming) datasets provide the police department
responsible for the OIS. I collapse these counts by police departments and assign each police department the population of the city or county that
it has jurisdiction over. I take city and county populations by race from the 2011 ACS 5-year estimates. The Ross (2015) estimates may contain
simulation error but are replicated using the exact program used in Ross (2015). In addition, the exact results are not reproducible because the
program does not set a seed.
do this principle a disservice if we do not Race, Arrests, and Police Use of Force.”
adhere to strict standards of evidence and Center for Policing Equity.
take at face value descriptive statistics that Kindy, Kimberly, Marc Fisher, Julie
are consistent with our preconceived ideas. Tate, and Jennifer Jenkins. 2015.
‘Stay Woke’ – but critically so. “A Year of Reckoning: Police Fatally
Shoot Nearly 1,000.” The Washington
REFERENCES Post, December 26. www.washingtonpost.
com/sf/investigative/2015/12/26/a-year-
of-reckoning-police-fatally-shoot-nearly-
Arthur, Rob, Taylor Dolven, Kee-
1000/
gan Hamilton, Allison McCann, and
Carter Sherman. 2017. “Shot by cops Nix, Justin, Bradley Campbell, Ed-
and forgotten.” VICE News, December ward Byers, and Geoffrey Alpert.
11. news.vice.com/en us/article/xwvv3a/ 2017. “A Bird’s Eye View of Civilians
shot-by-cops Killed by Police in 2015.” Criminology &
Public Policy 16 (1): 309-340.
Bertrand, Marianne and Sendhil Mul-
lainathan. 2004. “Are Emily and Greg Ross, Cody. 2015. “A Multi-Level Bayesian
More Employable Than Lakisha and Ja- Analysis of Racial Bias in Police Shootings
mal? A Field Experiment on Labor Mar- at the County-Level in the United States,
ket Discrimination.” American Economic 2011-2014.” PLoS ONE 10 (11).
Review 94 (4): 991-1013. Swaine, Jon, Oliver Laughland, Jamiles
Fryer, Roland G. Forthcoming. “An Em- Larteyand, and Ciara McCarthy.
pirical Analysis of Racial Differences in 2015. “Young Black Men killed by
Police Use of Force.” Journal of Political US Police at Highest Rate in Year
Economy. of 1,134 Deaths.” The Guardian, De-
cember 31. www.theguardian.com/us-
Goff, Philip A., Tracey Lloyd, Amanda
news/2015/dec/31/the-counted-police-
Geller, Steven Raphael, and Jack
killings-2015-young-black-men
Glaser. 2016. “The Science of Justice:You can also read
