Analysis of Racial and Gender Bias in SLCPD's Use of Force and Street Checks, 2014-2017

Slides for this talk are available at https://rpubs.com/kaince/SLCPDIthaca.

• Nationally, racial bias in police use of force exploded into the national conversation in 2014 with the killings of Trayvon Martin, Michael Brown, and Eric Garner, all unarmed black men killed by police.
• One year later, the Washington Post found that unarmed Black men were seven times more likely than whites to die by police gunfire.
• On August 13, 2017, the Salt Lake City Police Department (SLCPD) shot and killed a Black man, Patrick Harmon, from behind while Harmon was fleeing police.
• On May 23, 2020, SLCPD shot the fleeing Bernardo Palacios Carbajal, a Latinx man, 34 times in the back. Two days later, in Minneapolis, George Floyd was killed by officer Derek Chauvin, who was convicted of second-degree murder, third-degree murder and second-degree manslaughter in April 2021. These killings sparked anti-police brutality protests in the Salt Lake City area, as well as initiating the largest nationwide civil rights protest movement in decades.

• Our study is the first independent study of SLCPD's use of racial profiling since 2000 (Deseret News). That study found no evidence of racial bias in ticketing.
• SLCPD provides aggregate use of force data by race on their website (SLC Police Use of Force, 2021), but the data is not cleaned, lists 25% of the subjects' races as ''unknown'', and is limited to the last 24 months.
• Moreover, this data is not compared to the SLC population to test for statistically significant evidence of bias.
• SLCPD shared internal spreadsheets with us in order to determine whether there is evidence of bias in their use of force.

• Statistical tests start by assuming there is no bias (the skeptical viewpoint)
• SLC's population was \( 3.3\% \) Black or African-American in \( 2014 \) (American Community Survey, 2014), so we would expect the residents against which SLCPD used force in that year to be \( 3.3\% \) Black as well (if no bias existed).
• Pearson's \( \chi^{2} \) test quantifies how unlikely observed use-of-force demographics are given SLC's demographics (the expected use-of-force demographics under the skeptic's view).

• Expected proportions: American Community Survey SLC estimate for \( 2014 \), \( 2015 \) Selected Population Tables for \( 2015 \)-\( 2017 \)
  • Conducted by Census Bureau
  • Best representation of SLC's growing population
• Observed proportions: Use of Force and Street Checks data
  • Collected by the SLCPD, encompasses all of SLC
  • Analyzed primarily by race and gender

• We detail our cleaning process for the 2014 use of force data. Since the process for our analysis of \( 2015 \)-\( 2017 \) is similar, we will only discuss the results of that analysis.
• We sought out and removed exact duplicates in the dataset (e.g., a subject charged with multiple crimes).
• We removed entries that said “no force” or “race unknown”.
• We then plotted the race distributions in our observed use of force data, which will later be compared to the American Community Survey.

We create a geographic heatmap of the number of use-of-force incidents in various neighborhoods of Salt Lake City between \( 2014 \)-\( 2017 \).

From the heatmap, the largest number of police use-of-force incidents in Salt Lake City occurred in the area of Pioneer Park, bounded roughly by the streets 100 South and 400 South, as well as by 600 West and 300 West. In the timeframe considered for this study, this area included The Road Home shelter, which along with the park was considered a high-crime area (Salt Lake Tribune, \( 2014 \))

• “Asian” population in use of force dataset includes Pacific Islanders, but the ACS considers both groups separately
  • we added together \( \text{AA}+\text{PI} \) in the ACS data in order to follow the police department's data.
• ACS allows reporting more than one race, but SLCPD only has \( 4 \) monoracial categories
  • unclear how to handle this
  • we assigned multiracial folks who were white and a member of a racialized group to the corresponding racialized group
  • unclear how to match multiracial folks of \( 2+ \) racialized groups to SLCPD dataset (we removed them from the dataset)

• The null hypothesis for our dataset is that the proportion of races in our 2014 Use of Force dataset is equal to the proportion of races in the 2014 ACS. In symbols: \[ H_{0}: p_{\text{A}} = 0.079, p_{\text{B}} = 0.033, p_{\text{I}} = 0.019, p_{\text{W}} = 0.74 \]
• The alternative hypothesis states that the proportion of races in our Use of Force dataset is not equal to the proportion of races in the 2014 ACS. This could mean that if any of our expected proportions are wrong, then we reject the null in favor of our alternative hypothesis.

We used a \( \chi^{2} \) model with \( 4-1=3 \) degrees of freedom. The three conditions that must be checked are:

1. Independence: SLCPD's use of force against one individual is (usually) not correlated with their use of force against another individual. (Or is it..?)
2. Sample size / distribution: the expected counts (the number of individuals of each of the four races in the ACS) is greater than \( 5 \) (by far).
3. Degrees of freedom: we have more than two degrees of freedom.

# A tibble: 1 x 4
  statistic p.value parameter method                                  
      <dbl>   <dbl>     <dbl> <chr>                                   
1     2237.       0         3 Chi-squared test for given probabilities

The value of our 2014 \( \chi^{2} \) is 2237.4214612 with a \( p \)-value of less than \( 0.0001 \). We reject the null hypothesis of no difference in racial demographics. There is statistically significant evidence that the distribution of races against whom SLCPD used force is different from the distribution of races in Salt Lake City.

• The residuals show the difference between our observed and expected counts. Negative numbers show underrepresentation, while positive numbers represent overrepresentation.

• The farther the number is from 0, the more over/underrepresented a group is in our data. Residuals show a relative distance between the observed and expected values.

• African Americans and Indigenous folks (coded as American Indians/Alaska Natives in the dataset) were overrepresented compared to their proportion in the population (residuals of around thirty-four and twenty-seven, respectively). Asian and Pacific Islanders were fairly underrepresented, with a residual of around \( -16 \).

race_2014
Asian or Pacific Islander                     Black                Indigenous 
               -16.033295                 34.054482                 27.916891 
                    White 
                -6.426069 

# A tibble: 1 x 4
  statistic p.value parameter method                                  
      <dbl>   <dbl>     <dbl> <chr>                                   
1     3196.       0         3 Chi-squared test for given probabilities

The value of our \( \chi^{2} \) is 3196.2268752. The \( p \)-value is less than \( 0.001 \). We reject the null hypothesis of no discrimination.

We may have made a Type I error, rejecting the null hypothesis when the null hypothesis was in fact true. This would mean that there really was no difference in the proportion of races, and that the observed overrepresentation of Black and Indigenous people in the use of force dataset was due to chance.

• “Typically a 'street check' or 'police stop' is a practice where police stop a person in public, question them, and record their personal information in a police database. A vehicle check stop is not a street check.” (British Columbia Civil Liberties Association, 2018).
• A freedom of information request in \( 2017 \) showed that Edmonton, Alberta, police use street checks disproportionately often against Indigenous and Black people (CBC News, 2017).
• Street check data is not made public by SLCPD, but the Department shared their data with us in a collaboration to better understand evidence of potential bias.
• As before, we will analyze the street check data from \( 2014 \) separately from the data from \( 2015 \)-\( 2017 \).

The value of our \( \chi^{2} \) is 46.7187277. The \( p \)-value is 3.9889362 × 10^-10.

The \( p \)-value calculated is exceptionally small, meaning that we reject the null in favor of the alternative hypothesis. This means that there is significant evidence that the proportions of races in our data is not equal to the model proportions in the American Community Survey.

The value of our \( \chi^{2} \) is 857.9825016. The \( p \)-value is 1.1498775 × 10^-185.

The p-value calculated is exceptionally small, meaning that we reject the null in favor of the alternative hypothesis. This means that there is significant evidence that the proportions of races in our data is not equal to the model proportions in the American Community Survey.

• There is statistically significant evidence of discrimination against Black and Indigenous people in SLCPD use of force (residuals \( +46.7 \) and \( +24.4 \), respectively, for 2015-2017), while Asians and Pacific Islanders are underrepresented in SLCPD use of force (residual \( -19.3 \))
• There is statistically significant evidence of discrimination against Black people in SLCPD street checks (residual \( +24.5 \) for 2015-2017), while Asians and Pacific Islanders are underrepresented in street checks (residual \( -5.2 \) for 2015-2017).
• This study should be viewed as preliminary and highlights the need for more research into SLCPD's use of force and street checks.

• Limited time frame
• attempt to simplify multiracial people (assigning them to their racialized identity to match SLCPD records)
• large proportion of individuals with “unknown” race in SLCPD datasets
• not enough data to analyze ethnicity (e.g. Latinx folks)
• no analysis on age

• no control for/analysis of crime rates by race or gender
• no control for/analysis of locations of various races in SLC
• need for more SLCPD context
  • police demographics on particular beats could influence force used or if someone was checked
  • no analysis of police presence in a given location

• Since the original 2017 analysis, which has been published as a white paper and been shared in multiple research talks in Utah and Virginia as well as virtually, the need for more robust data on potential racial bias by local police forces has only grown, fueled by the continued murder of unarmed Black and Indigenous people by police and the 2020 civil rights movement and Black Lives Matter protests.
• I plan to work with two undergraduate students at Westminster College this summer to implement better controls for lurking variables like police presence in a region, crime rates and police interactions by race, and others.

• The SLCPD also shared data on ''persons linked to cases'', arrests, and traffic citations with us in \( 2017 \). Using this data, we will attempt to control for crime rates by race or gender, the locations of people of various racial groups in Salt Lake City, the number of police beats and police presence in a given region, and potentially the amount of police-public interactions in various ZIP codes.

Any questions?