Exploring Racial Disparities in NYS Patrol Stop Data
ARiefer
10/6/2021
Introduction
Around 500,000 people get pulled over by the police every single day in the United States. In the last few years, videos of traffic stops have created a national debate about the way law enforcement treats minorities (Vignesh 2016). Cases have spurred calls for more reliable information; everything from police video to data logged every time someone is pulled over, because traffic stops are one of the most common ways members of the public interact with the police (Vignesh 2016). The Fourth Amendment requires that before stopping the suspect, the police must have a reasonable suspicion that a crime has been, is being, or is about to be committed by the suspect. If the police reasonably suspect that the suspect is armed and dangerous, the police may stop the suspect. A reasonable stop is one “in which a reasonably prudent officer is warranted in the circumstances of a given case in believing that his safety or that of others is endangered, he may make a reasonable search for weapons of the person believed by him to be armed and dangerous.” Stops fall under criminal law, as opposed to civil law (Cornell Law School 1992). We believe that within these stops, there lies some biases or disparities, so we have run many tests on the data to explore this hypothesis.
Methods
We used a popularly used statistical methods, i.e., benchmarking, to analyze the data for detecting disparities in policing decisions, specifically, a benchmark comparison for stop rates. Benchmark analysis is the most common statistical method for assessing racial bias in police stops. Traditional benchmarks include the residential population, licensed drivers, arrestees, and reported crime suspects (Engel and Calnon 2004). McConnell and Scheidegger (see McConnell and Scheidegger 2001) have looked at stops initiated by aerial patrols and those based on radar and cameras (Lange, Blackman, and Johnson 2001), arguing that such stops are less prone to potential bias, and thus more likely to reflect the true population of traffic violators. Antonovics and Knight use officer-level demographics in a variation of the standard benchmark test: they argue that search rates that are higher when the officers race differs from that of the suspect is evidence of discrimination (Antonovic and Knight 2009). Finally, “internal benchmarks” have been used to flag potentially biased officers by comparing each officers stop decisions to those made by others patrolling the same area at the same time (Ridgeway and MacDonald 2009).
We also use descriptive and inferential statistics to explore and analyze the New York State Patrol data, along with quantitative observations about the proportion of stops for each race. We define proportion of stops to be the number of stops within a specific race divided by the total number of stops. We are using the population in NYS of each race as our baseline, defining stop rate to be the proportion of stopped drivers of a specific race. Furthermore, we made quantitative claims about disparities in stop rates by dividing the white stop rate by each of the other stop rates for each specific race. We also computed Cohen’s H value which measures the distance between the proportions of stops for drivers of each specific race compared to proportion of stops of White drivers (Wikipedia, n.d.). Cohen’s H value allows us to determine if the difference between the two proportions is statistically significant (Wikipedia, n.d.). Lastly, Grogger and Ridgeway construct benchmarks by considering stops at night, when a “veil of darkness” masks race (Grogger and Ridgeway 2006). We conducted a Veil of Darkness test the hypothesis that if stops made after dark had a smaller proportion of black drivers stopped than stops made during daylight, this would provide evidence of the presence of racial profiling.
The Data
We consider a comprehensive dataset of 7,960,103 traffic stops conducted in New York State between January 2010 and December 2017 that was obtained through the Stanford Open Policing Project website (Pierson 2020). Our data, The Stanford Open Policing Project, is a unique partnership between the Stanford Computational Journalism Lab and the Stanford Computational Policy Lab. Starting in 2015, the Open Policing Project began requesting such data from state after state. To date, the project has collected and standardized over 200 million records of traffic stop and search data from across the country (Pierson 2020). Even when states do provide the information, the way they track and then process the data varies widely across the country, creating challenges for standardizing the information. Data from 21 state patrol agencies and 29 municipal police departments, comprising nearly 100 million traffic stops, are sufficiently detailed to facilitate rigorous statistical analysis (Pierson 2020). The project has found significant racial disparities in policing. These disparities can occur for many reasons: differences in driving behavior, to name one. But, in some cases, we find evidence that bias also plays a role. Several variables are recorded for each stop, including the location of the stop, race of the driver (White, Black, Hispanic, and Other), subject sex (Male or Female), subject age, and the county the stop happened in. We have acquired race population data of New York State and Monroe County from the Census Bureau website (New York State Census Bureau, n.d.).
The Results
First, here are the number of stops per year from 2010 to 2017. The year 2010 consisted of the highest number of stops with.
Next we have the number of stops by sex.
We then visually explore the distribution of stops in NYS by race.
The below table depicts the number of stops by race and the proportion of stops per race (number of stops per race divided by total number of stops).
| Race | Number of Stops | Proportion of Stops |
|---|---|---|
| black | 888472 | 0.112 |
| hispanic | 553336 | 0.070 |
| other | 593420 | 0.075 |
| white | 5924875 | 0.744 |
Among this set of stops, we find that 74.4% of drivers are white, 11.2% are black, 7% are Hispanic, and 7.5% are other. Notice how the highest percentage of drivers being stopped are of White drivers and the second highest percentage of drivers being stopped is of Black drivers.
We then want to graphically look at the number of stops per year subjected by race to see the trend of stops over time for each race.
Graphically, we can see that the number of stops of White drivers is extremely higher than stops of any other race throughout the years. From this plot we see that, at least for Black and White drivers, the annual trends are very different by race. All races experienced a spike in stops in 2014, but thereafter, there were fewer White drivers stopped from 2015-2017, whereas there continued to be an increase in the number of Black and Hispanic drivers stopped over those two years.
We investigated further to see if there was a difference between male drivers and female drivers within each specific race over time.
Both of these figures for male and female drivers depict the same trends for both male and female drivers for each race over the years from 2010-2017. Regardless of gender, White drivers seem to be stopped at a much higher rate than compared to any other race. However, notice how in the figure for male drivers, the number of male drivers being stopped is drastically higher than the number of female drivers being stopped, regardless of race. This could be due to there being more males in NYS than females, more males driving than females, etc.
We then look at some bivariate relationships as shown below. First we look at the subject’s race versus the subject’s sex.
Notice how there tends to be more male drivers than female drivers, regardless of race. Also, note how White drivers consist of the most male and female drivers in our data. Hence, this shows there are more White drivers than any other race, regardless of gender.
We then explore the relationship between a subject’s race and a subject’s age.
The graph shows the relationship between race and mean age of the drivers. First notice how the average age of drivers seems to be the same around 34-37, regardless of race. Also, note how the mean age of White drivers is the highest with an average age of 37. Lastly, note that this graph shows that the lowest average age of drivers is amongst Black and Hispanic drivers with an age of 34.
Lastly, we look at the relationship between a subject’s sex and a subject’s age.
The graph shows the association between sex and age of the drivers. Notice how the average age of drivers is about the same, around 35-37, regardless of gender. Also, note how the highest mean age of drivers is among male drivers with a mean age of 37 while female drivers have a slightly lower mean age of 35.
We then distributed the number of stops in NYS across a heat map. The following heat map is a map of New York State with different shades of blue representing different number of stops per county in NYS. The lighter shade of blue, the more stops occurred in those respective counties. We also recreated the same map except with different colors to make the map more colorful.
We now will use specific statistical methods for assessing racial bias including comparing Cohen’s H values, benchmark analysis, and the Veil of Darkness test using police stops in NYS in 2017 only.
We look at the number of stops within NYS in 2017 only.
| Race | Number of Stops | Proportion of Stops |
|---|---|---|
| black | 139151 | 0.141 |
| hispanic | 91226 | 0.092 |
| other | 93919 | 0.095 |
| white | 662493 | 0.671 |
In 2017 stops, we find that 67.1% of drivers are White, 14.1% are Black, 9.2% are Hispanic, and 9.5% are of Other races. Notice how the highest percentage of drivers being stopped are of White drivers and the second highest percentage of drivers being stopped is of Black drivers.
Next, we computed Cohen’s H value which measures the distance between the proportions of stops for drivers of each specific race compared to proportion of stops of White drivers.
| Cohen’s H Value | |
|---|---|
| White vs. Non-White | 0.70 |
| White vs. Black | 1.15 |
| White vs. Hispanic | 1.30 |
| White vs. Other | 1.29 |
From the table above, we find that when each race is compared to the White race, the Cohen’s H value is above 1, except for when White is compared to non-White which gives us a value of 0.70. For each comparison that is larger than 1, we can conclude that each difference in proportions of stops between these comparisons is large. Thus, each of these differences have a large effect size. However, when White is compared to non-White, we have a value of 0.70, thus we can conclude this difference in proportions of stops has a medium effect size.
In order to conclude any racial discrimination of police stops, we need to conduct a benchmark analysis. With a benchmark analysis, we are able to compare the number of stops of each race with the population of that race as a whole. We are now able to determine if any specific race is over-represented in the number of stops compared to its respective population. First, we retrieve the population numbers of each specific race within NYS.
| Race | Population in NYS | Population Proportion in NYS |
|---|---|---|
| black | 2763709 | 0.142 |
| hispanic | 3749257 | 0.193 |
| other | 2230161 | 0.115 |
| white | 10710524 | 0.551 |
Among the population data for each race in NYS in 2017, we find that Whites make up 55.1% of the population. We find that NYS population consists of 14.2% Black, 19.3% Hispanic, and 11.5% of other races. Notice how NYS population consists of mostly people who are White and Hispanic being the second largest race that is represented.
We will visually represent the population race distribution in NYS in 2017 below.
We then want to investigate the stop rates for each race to determine if there is any racial bias within police stops. We compute stop rates for each race by dividing the number of stops in 2017 by population of that race in 2017 within NYS. Below shows the table of number of stops, population, stop rate, and population proportion for each race in 2017 in NYS.
| Race | Number of Stops | Population | Stop Rate |
|---|---|---|---|
| black | 139151 | 2763709 | 0.050 |
| hispanic | 91226 | 3749257 | 0.024 |
| other | 93919 | 2230161 | 0.042 |
| white | 662493 | 10710524 | 0.062 |
Note that the White stop rate is larger than any other races stop rate. This provides evidence that White drivers are stopped more than drivers of any other race, relative to their share of the city’s population. Note that White drivers are stopped at a rate 1.24 times higher than Black drivers, and White drivers are stopped at a rate 2.58 times higher than Hispanic drivers. By conducting a benchmark analysis and comparing the number of stops for each race with the population of each race in NYS, we can determine if a certain race is over-represented. Therefore, proving the existence of racial bias within police stops. However, from the results above, we see that White drivers are stopped the most in NYS in 2017.
Next, we will conduct the Veil of Darkness test to further explore any racial profiling among police stops within NYS.
| term | estimate | std.error | statistic | p.value |
|---|---|---|---|---|
| (Intercept) | -1.3414672 | 0.0253991 | -52.8154434 | 0.0000000 |
| darkTRUE | -0.2118589 | 0.0123820 | -17.1102946 | 0.0000000 |
| ns(time, df = 6)1 | 0.0434678 | 0.0259303 | 1.6763305 | 0.0936735 |
| ns(time, df = 6)2 | 0.1055094 | 0.0338065 | 3.1209802 | 0.0018025 |
| ns(time, df = 6)3 | 0.4629234 | 0.0356582 | 12.9822250 | 0.0000000 |
| ns(time, df = 6)4 | 0.2018509 | 0.0290239 | 6.9546522 | 0.0000000 |
| ns(time, df = 6)5 | 0.3066085 | 0.0513551 | 5.9703644 | 0.0000000 |
| ns(time, df = 6)6 | 0.3275128 | 0.0247104 | 13.2540297 | 0.0000000 |
| weekdayMon | -0.0874229 | 0.0164112 | -5.3270283 | 0.0000001 |
| weekdayTue | -0.0787830 | 0.0166021 | -4.7453530 | 0.0000021 |
| weekdayWed | -0.0606694 | 0.0169233 | -3.5849677 | 0.0003371 |
| weekdayThu | -0.0562594 | 0.0169036 | -3.3282523 | 0.0008739 |
| weekdayFri | -0.1398241 | 0.0154763 | -9.0347328 | 0.0000000 |
| weekdaySat | 0.0304680 | 0.0152133 | 2.0027227 | 0.0452071 |
| seasonFall | 0.0957696 | 0.0097204 | 9.8524716 | 0.0000000 |
| factor(year(date1))2011 | -0.0457921 | 0.0174656 | -2.6218479 | 0.0087454 |
| factor(year(date1))2012 | -0.0141841 | 0.0178602 | -0.7941751 | 0.4270935 |
| factor(year(date1))2013 | 0.0060312 | 0.0176345 | 0.3420140 | 0.7323404 |
| factor(year(date1))2014 | 0.0000176 | 0.0175406 | 0.0010062 | 0.9991972 |
| factor(year(date1))2015 | 0.0608611 | 0.0172270 | 3.5328982 | 0.0004110 |
| factor(year(date1))2016 | 0.1649086 | 0.0171069 | 9.6399040 | 0.0000000 |
| factor(year(date1))2017 | 0.4517588 | 0.0164284 | 27.4985825 | 0.0000000 |
## Estimate
## -0.21185886470634951450620064861141145229339599609375000000000000000000000
## Std. Error
## 0.01238195309020868536964421480206510750576853752136230468750000000000000
## z value
## -17.11029456846204510611642035655677318572998046875000000000000000000000000
## Pr(>|z|)
## 0.00000000000000000000000000000000000000000000000000000000000000001243689Here we used logistic regression extract the Veil of Darkness coefficient or the estimate to determine the likelihood that a stopped driver will be nonwhite, while adjusting for time of day. You can see that our coefficient is negative (-0.2118589), which means the darkness decreases the likelihood that a stopped driver is nonwhite, after adjusting for night times. Also, because the standard error is so small (0.0123820), this indicates that this test outputted a statistically significant finding, which is that the veil of darkness does affect the likelihood of nonwhite drivers being pulled over at night.
Now we will produce similar results for stops in Monroe County only.
Below, we look at the number of stops by sex.
The below table depicts the number of stops by race and the proportion of stops per race (number of stops per race divided by total number of stops) within Monroe County.
| Race | Number of Stops | Proportion of Stops |
|---|---|---|
| black | 3397 | 0.294 |
| hispanic | 1255 | 0.108 |
| other | 457 | 0.039 |
| white | 6464 | 0.559 |
Among this set of stops, we find that 61.7% of drivers are White, 26.1% are Black, 8.1% are Hispanic, and 4.0% are of Other races. Notice how the highest percentage of drivers being stopped are of White drivers and the second highest percentage of drivers being stopped is of Black drivers.
We then visually explore the distribution of stops in Monroe County by race.
We then want to graphically look at the number of stops per year subjected by race to see the trend of stops over time for each race in Monroe County.
Graphically, we can see that the number of stops of White drivers is extremely higher than stops of any other race throughout the years. From this plot we see that, at least for Black and White drivers, the annual trends are very different by race. All races experienced a drop in stops after 2014.
We now will use specific statistical methods for assessing racial bias including comparing Cohen’s H values, benchmark analysis, and the Veil of Darkness test using police stops in Monroe County in 2017 only.
We look at the number of stops within Monroe County in 2017 only.
| Race | Number of Stops | Proportion of Stops |
|---|---|---|
| black | 317 | 0.316 |
| hispanic | 180 | 0.180 |
| other | 48 | 0.048 |
| white | 457 | 0.456 |
In 2017 stops in Monroe County, we find that 53.6% of drivers are White, 31.4% are Black, 10.5% are Hispanic, and 4.5% are of Other races. Notice how the highest percentage of drivers being stopped are of White drivers and the second highest percentage of drivers being stopped is of Black drivers.
Next, we computed Cohen’s H value which measures the distance between the proportions of stops for drivers of each specific race compared to proportion of stops of White drivers.
| Cohen’s H Value | |
|---|---|
| White vs. Non-White | 0.14 |
| White vs. Black | 0.45 |
| White vs. Hispanic | 0.98 |
| White vs. Other | 1.22 |
From the table above, we find that when each race is compared to the White race, the Cohen’s H value is below 1, except for when White is compared to Other races which gives us a value of 1.22. For this comparison we can conclude that this difference in proportions of stops has a large effect size. The difference in proportions of stops between White drivers and Hispanic drivers has a value of 0.98 which is also considered to be a large effect size. The last two comparisons between White and Black drivers and White versus non-White drivers, both have a small value, leading us to conclude that this difference has a small effect size.
In order to conclude any racial discrimination of police stops, we need to conduct a benchmark analysis. With a benchmark analysis, we are able to compare the number of stops of each race with the population of that race as a whole. We are now able to determine if any specific race is over-represented in the number of stops compared to its respective population. First, we retrieve the population numbers of each specific race within Monroe County.
| Race | Population in Monroe County | Population Proportion in Monroe County |
|---|---|---|
| black | 114292 | 0.154 |
| hispanic | 68242 | 0.092 |
| other | 39256 | 0.053 |
| white | 519980 | 0.701 |
Among the population data for each race in Monroe County in 2017, we find that Whites make up 70.1% of the population. We find that Monroe County population consists of 15.4% Black, 9.2% Hispanic, and 5.3% of other races. Notice how Monroe County population consists of mostly people who are White and Black being the second largest race that is represented.
We will visually represent the population race distribution in Monroe County in 2017 below.
We then want to investigate the stop rates for each race to determine if there is any racial bias within police stops. We compute stop rates for each race by dividing the number of stops in 2017 by population of that race in 2017 within Monroe County Below shows the table of number of stops, population, stop rate, and population proportion for each race in 2017 in Monroe County.
| Race | Number of Stops | Population | Stop Rate |
|---|---|---|---|
| black | 317 | 114292 | 0.003 |
| hispanic | 180 | 68242 | 0.003 |
| other | 48 | 39256 | 0.001 |
| white | 457 | 519980 | 0.001 |
Note that the White stop rate is not larger than any other races stop rate. This provides evidence that drivers of any other race are stopped more than White drivers, relative to their share of the city’s population. Note that Black drivers are stopped at a rate 2.62 times higher than White drivers, and Hispanic drivers are stopped at a rate 1.47 times higher than White drivers. By conducting a benchmark analysis and comparing the number of stops for each race with the population of each race in NYS, we can determine if a certain race is over-represented. Therefore, proving the existence of racial bias within police stops. From the results above, we can clearly conclude that Black and Hispanic drivers are stopped much more often than compared to White drivers in Monroe County in 2017.
Discussion:
Graphically, we saw that in NYS the number of stops of White drivers is extremely higher than stops of any other race throughout the years. We also saw how there tends to be more male drivers than female drivers, regardless of race and that there were more White drivers than any other race, regardless of gender. The average age of drivers that were stopped was around 34-37 years, regardless of race or sex. Computing Cohen’s H values led us to conclude that the difference in proportion of stops between races was statistically significant for each comparison besides White versus non-White. Lastly, due while these baseline stats give us a sense that there are racial disparities in policing practices in NYS, but they are not evidence of discrimination due to the results of the benchmark analysis. The argument against the benchmark test is that we have not identified the correct baseline to compare to.
We received similar trends for stops only in Monroe County. However, these same baseline stats showed us that there are racial disparities in policing practices in Monroe County. Due to the results of the benchmark analysis, and these stats are evidence of racial discrimination. In Monroe County, Black and Hispanic drivers were over-represented and disproportionately stopped relative to their population numbers. Specifically, the population of Whites in Monroe County in 2017 was 4.5 times higher than Blacks and 7.6 times higher than Hispanic drivers, making it obvious that with stop rates being almost triple for Black drivers and almost double for Hispanic drivers than compared to White drivers, racial bias is present in Monroe County.
Acknowledgements
I would like to say a special thanks to Dr. Zaihra for the guidance and completion of this project.
References: