Table
See the “Discussion” tab for more information on how race is
ascertained and why the number of missing cases is so high.
tab <- homicides %>%
mutate(race = case_match(race,
"API" ~ "Asian/Pacific Islander",
"BAA" ~ "Black/African American",
"HL" ~ "Hispanic/Latinx",
"NAA" ~ "Native American/Indigenous",
"WEA" ~ "White/European American",
.default = race),
race = fct_relevel(race, "Unknown", after = Inf)
) %>%
group_by(race) %>%
summarize(Number = n(),
Percent = round(100*Number/nrow(homicides), 1)
) %>%
bind_rows(data.frame(race="Total",
Number = sum(.$Number),
Percent = sum(.$Percent))) %>%
rename(Race = race)
tab %>%
kable(caption = "Breakdown by Race") %>%
kable_styling(bootstrap_options = c("striped","hover")) %>%
row_spec(row=dim(tab)[1], bold = T) %>%
add_footnote(label = "Percents may not sum to 100 due to rounding",
notation = "symbol")
Breakdown by Race
|
Race
|
Number
|
Percent
|
|
Asian/Pacific Islander
|
29
|
6.3
|
|
Black/African American
|
56
|
12.3
|
|
Hispanic/Latinx
|
66
|
14.4
|
|
Native American/Indigenous
|
20
|
4.4
|
|
White/European American
|
223
|
48.8
|
|
Unknown
|
63
|
13.8
|
|
Total
|
457
|
100.0
|
|
* Percents may not sum to 100 due to rounding
|
Plots
homicides %>%
mutate(race = case_match(race,
"API" ~ "Asian/Pacific Islander",
"BAA" ~ "Black/African American",
"HL" ~ "Hispanic/Latinx",
"NAA" ~ "Native American/Indigenous",
"WEA" ~ "White/European American",
.default = race),
race = fct_relevel(race, "Unknown", after = Inf)
) %>%
count(race) %>%
mutate(perc = n / nrow(homicides)) %>%
ggplot(aes(x=race,
y = perc,
label = n)) +
geom_bar(stat="identity", fill="blue", alpha=.5) +
geom_text(aes(y = perc), size = 3, nudge_y = .025) +
theme(axis.text.x = element_text(size=5.5)) +
scale_y_continuous(labels = scales::percent_format(acc=1)) +
labs(title = "Fatalities by Race",
caption = paste0("WA State since ",
params$from,
"; y-axis=pct, bar label=count")) +
xlab("Reported Race") +
ylab("Percent of Total")
homicides %>%
mutate(raceb = case_when(race == "WEA" ~ "White",
race == "Unknown" ~ "Unknown",
TRUE ~ "BIPOC"),
raceb = fct_relevel(raceb, "Unknown",
after = Inf)) %>%
count(raceb) %>%
mutate(perc = n / nrow(homicides)) %>%
ggplot(aes(x=raceb,
y = perc,
label = n)) +
geom_text(aes(y = perc), size = 3, nudge_y = .025) +
geom_bar(stat="identity", fill="blue", alpha=.5) +
scale_y_continuous(labels = scales::percent_format(acc=1)) +
labs(title = "Fatalities by Race",
caption = paste0("WA State since ",
params$from,
"; y-axis=pct, bar label=count")) +
xlab("Reported Race") +
ylab("Percent of Total")
Discussion
Racial disparities in the risk of being killed by police are one of
the most important factors driving the public demand for police
accountability and reform. For that reason it is important to understand
how these numbers can, and cannot be used.
TL;DR While there are many uncertainties in the data
that make it difficult/impossible to know the exact numbers, there are
still some conclusions we can draw with confidence.
How is race identified in our data?
All crowd-sourced datasets rely primarily on media reports to find
cases, and these media reports rarely include information on the race of
the person killed. Each of the projects that provide data for this
report has made additional efforts made to identify race.
The MPV project makes extensive additional efforts to identify the
race of the person killed. From their data
methodology page:
Our data has been meticulously sourced from official police use of
force data collection programs in states like California, Texas and
Virginia, combined with nationwide data from The Gun Violence Archive
and the Fatal Encounters database, two impartial crowdsourced databases.
We’ve also done extensive original research to further improve the
quality and completeness of the data; searching social media,
obituaries, criminal records databases, police reports and other sources
to identify the race of 90 percent of all victims in the database.
The FE project used a statistical imputation model to predict race
for the missing cases. A brief description of the methodology is online
here.
They were able to impute just over half of the missing cases with
reasonable confidence, and we use these imputations for the FE data in
this report.
The IN project do not provide a description of the methods they use
for coding race, but their records typically include information on the
race of the person killed, as well as a photo.
The local project conducts a search for coroner/ME reports,
obituaries, multiple media reports and social media channels for
information that can be used to assign the race of the person
killed.
How often is race missing in our data?
We use a hierarchical assignment of race for the data in this report:
MPV data when available, IN data when available, WA local data for cases
not included in MPV or IN, and Fatal Encounters (FE) with imputation for
cases uniquely contributed by FE.
tab <- homicides %>%
mutate(race = factor(race),
Race = fct_relevel(race, "Unknown", after = Inf)) %>%
group_by(Race, Source=source) %>%
count() %>%
pivot_wider(names_from = Race, values_from = n) %>%
ungroup() %>%
mutate(across(where(is.numeric), ~replace_na(., 0))) %>%
add_row(Source = "Total", summarise(., across(where(is.numeric), sum))) %>%
rowwise() %>%
mutate(Total = sum(across(where(is.numeric))),
"Unknown Pct" = round(100*Unknown/Total, 1)) %>%
arrange(Total)
tab %>%
kable(caption = "Breakdown by Source of Information") %>%
kable_styling(bootstrap_options = c("striped","hover")) %>%
row_spec(which(tab$source == "Total"), bold=T) %>%
column_spec(which(names(tab)=="Total"), bold=T) %>%
column_spec(which(names(tab)=="Unknown Pct"), italic=T) %>%
add_footnote(label = "API = Asian and Pacific Islander; BAA = Black/African American; HL = Hispanic/Latino; NAA = Native American/Alaskan native; WEA = White/European American",
notation = "symbol")
Breakdown by Source of Information
|
Source
|
API
|
BAA
|
HL
|
NAA
|
WEA
|
Unknown
|
Total
|
Unknown Pct
|
|
IN
|
0
|
4
|
3
|
2
|
4
|
3
|
16
|
18.8
|
|
FE
|
5
|
8
|
4
|
2
|
36
|
13
|
68
|
19.1
|
|
WA
|
11
|
18
|
29
|
5
|
85
|
23
|
171
|
13.5
|
|
MPV
|
13
|
26
|
30
|
11
|
98
|
24
|
202
|
11.9
|
|
Total
|
29
|
56
|
66
|
20
|
223
|
63
|
457
|
13.8
|
|
* API = Asian and Pacific Islander; BAA = Black/African
American; HL = Hispanic/Latino; NAA = Native American/Alaskan native;
WEA = White/European American
|
Bottom line: The unknown race cases make it impossible to
say exactly how many people killed by police are in each racial
group.
Reported vs. self-identified race
The race classification in our data represent the subject’s race as
reported by by some other person – law enforcement, an ME/Coroner, a
journalist, or a coder for one of the open-source projects that we draw
from. It may be that person’s perception, or it may be informed by
interactions with the subject’s friends or family, we have no way of
knowing.
Bottom line: The accuracy of the race classification in the
data we have is unknown.
We report the raw counts in this report, not per capita rates
Breaking the total count down by race, the largest single group of
persons killed by police are identified as White/European-American:
49%
So, does this mean that we can say: “There aren’t any racial
disparities in persons killed by police?”.
No. This raw count can not be used to assess racial disparities,
because the word disparity implies the risk is “disproportionate” – that
is, higher (or lower) than proportional. Proportional is a comparative
term: it compares the proportion of fatalities by race, to the
proportion of the population by race. If the two proportions are the
same, then we can say there are no disparities. In this case, if 49% of
the WA state population is White, only then would we be able to
say that their risk of being killed by police is proportional to their
share of the population.
Bottom line: The raw counts can not be used to identify
racial disparities.
Why don’t we calculate standardized per capita rates?
We don’t calculate standardized per capita rates because the race
categories our dataset are not consistent with the race categories used
by the US Census. This is currently a limitation shared by all of the
open-source datasets we rely on.
Per capita rates are calculated by taking the ratio of the fatality
count (in the numerator) to the population count (in the denominator).
This requires a consistent measure of race to use for the cases in the
numerator and denominator - and we don’t have that. Census race
categories are different for three reasons:
The Census allows for multiple race coding: About 5% of people in
WA state report two or more races in the Census. This type of
multiple-race classification does not exist in the data from any of our
source datasets, and it is almost never used in reports on persons
killed by police, law enforcement statistics, or official medical
examiner reports. The likely reason for this is because the Census
categories are self-reported – by the persons themselves – while race
classification in the other settings is assigned by someone to another
(deceased) person.
The Census categories for single race are different: The
standardized Census categories for race are American Indian/Alaska
Native, Asian, Black, Native Hawaiian/Pacific Islander, White. Slightly
different versions are used in each of our source datasets.
The Census separates the coding of race from ethnicity:
Hispanic/Latinx is an ethnicity that crosses several racial groups,
primarily White, Black and Native American. In the US Census data,
ethnicity is measured separately from race, so one can observe both
values for each person. In all of our source datasets, “Hispanic” is
coded as one of the racial classifications, rather than as a separate
ethnicity classification, so the race of Hispanic persons can not be
observed.
For more information on how the Census codes race and
ethnicity:
U.S.
Decennial Census Measurement of Race and Ethnicity Across the Decades:
1790–2020
Bottom line: The mismatch between racial categories coded in
our data, and the categories used in the US Census makes it impossible
to calculate per capita rates accurately for all groups.
What can we say about racial disparities in the police use of deadly
force in WA State?
We can say that, while we can not estimate the disparities with
precision, or for all groups, the the data do suggest there are racial
disparities in the police use of deadly force in WA State.
The population of Washington State overwhelmingly identifies as White
– 73% in the last US Census – while the fraction who identify as Black
is 4% (2020, source:
WA State OFM for this and all other population estimates in this
section). The fraction (of all races) who report Hispanic ethnicity is
14%. If we restrict the focus to non-Hispanics, these fractions drop to
64% and 4% for those who identify as White and Black respectively.
For these three groups – non-Hispanic White, non-Hispanic Black and
Hispanic – our source data are coded in a way that we can compare our
data to the Census populations, if we make three assumptions:
First, that that cases classified as “White” or “Black” in the
our datasets are not Hispanic – because they would be classified as
Hispanic otherwise – and that cases classified as “Hispanic” in our
dataset include all races.
Second, that the coding of race and Hispanic origin in our
dataset are accurate, and
Third, that all race/Hispanic groups are equally likely to be
coded as “unknown” race in our data.
Under these assumptions (or, to the extent that they hold) we can
compare the fraction of each group in the fatality data to their
fraction in WA population. The table below shows this, along with a
summary based on the “risk ratio”.
The “risk ratio” is the ratio of the fatality share to the population
share. When the two shares are the same, the risk ratio equals 1. When
the risk ratio is less than one, this means the share of fatalities is
lower than the population share; the risks are disproportionately low.
When the risk ratio is greater than one it means the share of fatalities
is larger than the population share; the risks are disproportionately
high.
This can be translated into a relative risk percentage: 100*(risk
ratio-1), representing the percent lower (or higher) the risk ratio is
than expected if the risk was proportional to the share of
population.
tab.all <- homicides %>%
group_by(race) %>%
summarize(all.cases = n()) %>%
mutate(pct.all = round(100*all.cases/sum(all.cases), 1)) %>%
filter(race %in% c("BAA", "HL", "WEA", "Unknown"))
tab.known <- homicides %>%
filter(race != "Unknown") %>%
group_by(race) %>%
summarize(known.cases = n()) %>%
mutate(pct.known = round(100*known.cases/sum(known.cases), 1)) %>%
filter(race %in% c("BAA", "HL", "WEA")) %>%
select(-known.cases)
tab.pop <- data.frame(race = c("BAA", "HL", "WEA"),
pct.pop = c(round(100*prop.nhb.census,1), round(100*prop.hisp.census, 1),
round(100*prop.nhw.census,1))
)
tab <- tab.all %>%
left_join(tab.known) %>%
left_join(tab.pop) %>%
mutate(#rr.all = round((pct.all/pct.pop -1), 2),
rr.known = round((pct.known/pct.pop - 1), 2)) %>%
rename(Race=race, Fatalities=all.cases, "Percent (all cases)"=pct.all,
"Percent (known race)"=pct.known, "Percent (WA pop)"=pct.pop, "Relative Risk" =rr.known) %>%
mutate(Race = case_match(
Race, "BAA" ~ "Black/African-American",
"HL" ~ "Hispanic/Latinx",
"WEA" ~ "White/European-American",
.default = Race)
)
tab %>%
kable(caption = "Police use of deadly force by race: Relative risks") %>%
kable_styling(bootstrap_options = c("striped","hover")) %>%
add_footnote(label = "Relative risks are calculated based on percent of fatalities with known race",
notation = "symbol")
Police use of deadly force by race: Relative risks
|
Race
|
Fatalities
|
Percent (all cases)
|
Percent (known race)
|
Percent (WA pop)
|
Relative Risk
|
|
Black/African-American
|
56
|
12.3
|
14.2
|
3.9
|
2.64
|
|
Hispanic/Latinx
|
66
|
14.4
|
16.8
|
14.1
|
0.19
|
|
White/European-American
|
223
|
48.8
|
56.6
|
64.1
|
-0.12
|
|
Unknown
|
63
|
13.8
|
NA
|
NA
|
NA
|
|
* Relative risks are calculated based on percent of
fatalities with known race
|
In WA State:
For non-Hispanic Whites/European Americans this value is 100*(
57% / 64% -1) = -0.12 – their risk of being killed by police is 12%
lower than expected.
For non-Hispanic Blacks/African Americans this value is 100*( 14%
/ 4% -1) = 2.66 – their risk of being killed by police is 266% higher
than expected.
For Hispanics this value is 100*( 17% / 14% -1) = 0.19 – their
risk of being killed by police is 19% higher than expected.
Note the direction of these disparities is robust to one of our
assumptions: that each group is equally likely to be reported as “race
unknown”. Even if we assume that all of the “Unknown” race cases are
non-Hispanic White, their share of incidents, 63%, would still be below
their share in the population. And if all of the unknown cases are
assumed to be either non-Hispanic Black or Hispanic, that would actually
increase the relative risks for these groups.
So, even though we can’t be certain of the exact value of the risk
ratio, we can say with some confidence that there are racial
disparities in the risk of being killed by police, and that these
disparities indicate that non-Hispanic Whites are less likely
than other racial groups to be killed by law enforcement officers, while
non-Hispanic Blacks and Hispanics are more likely to be
killed.
Bottom line: while the largest single group of persons
killed by police in WA is identified as White, the risk ratio for this
group, which adjusts for their share of the population, shows they are
disproportionately less likely to be killed than those
identified as Black or Hispanic.
Can we say anything about implicit bias?
Probably not. To use these data as evidence for (or against) the
effects of systemic implicit bias in the police use of deadly force, we
would need to know the officer’s perception of the subject’s
race, and control for other variables that may influence these patterns,
like the perceived and actual weapon status of the person killed, the
circumstances that precipitated the incidents, the presence or absence
of other persons at risk in the incidents, as well as the demographic
composition of the jurisdiction.
Bottom line: We can’t use these data to answer the question
of the officer’s subjective intention or implicit bias. But the data do
support the argument that objective per capita disparities in risk by
race exist.
