Fresno, Calif.

The Markup and Gizmodo have obtained and analyzed actual predictions for more than three dozen departments that used PredPol predictive policing software for at least six months between 2018 and 2020. This data sheet provides the findings from our disparate impact analysis and public housing analysis for Fresno, Calif. To learn more about the project read, our investigation. For more details on how we did this analysis, read our methodology.

Findings

Overview

  • Predpol’s algorithm relentlessly targeted the block groups in each jurisdiction that were most heavily populated by people of color and the poor, particularly those containing public housing. The algorithm spared block groups with more White residents the same level of scrutiny.

  • The proportion of each jurisdiction’s Black and Latino residents was higher in the most-targeted block groups and lower in the least-targeted block groups compared to the jurisdiction overall. The opposite was true for the White population: The least-targeted block groups contained a higher proportion of White residents, and the most-targeted block groups contained a lower proportion.

  • For the majority of jurisdictions in our data set (27 jurisdictions), a higher proportion of their low-income households lived in the block groups that were targeted the most. In some jurisdictions, all of their subsidized and public housing was located in block groups PredPol targeted more than the median.

  • These vast disparities were caused by the algorithm relentlessly predicting crime in the block groups in each jurisdiction that contained a higher proportion of the low-income residents and Black and Latino residents. They were the subject of crime predictions every shift, every day, and in multiple locations in the same block group.

  • We also analyzed arrest statistics by race from the FBI’s Uniform Crime Reporting (UCR) Project for 29 of the agencies in our data that were in UCR. In 90 percent of them, per capita arrests were higher for Black people than White people—or any other racial group included in the dataset, mirroring the characteristics of the neighborhoods that the algorithm targeted.

  • We analyzed arrest data provided by 10 law enforcement agencies in our data and the rates of arrest in predicted areas remained the same whether PredPol predicted a crime that day or not.

Race and Ethnicity

Compared to Fresno, Calif., overall, the most-targeted block groups had:

  • A smaller proportion of Asians residents.
  • A greater proportion of Black residents.
  • A greater proportion of Latino residents.
  • A smaller proportion of White residents.

Compared to Fresno, Calif. overall, the least-targeted block groups had:

  • A smaller proportion of Asian residents.
  • A smaller proportion of Black residents.
  • A smaller proportion of Latino residents.
  • A greater proportion of White residents.

Targeting Level Demographic Proportion of Block Group pop.
Most Targeted Block Groups Asian 3.8
Most Targeted Block Groups Black 4.5
Most Targeted Block Groups Latino 45.0
Most Targeted Block Groups White 11.4
Median Targeted Block Groups Asian 10.1
Median Targeted Block Groups Black 2.3
Median Targeted Block Groups Latino 43.4
Median Targeted Block Groups White 22.6
Least Targeted Block Groups Asian 2.9
Least Targeted Block Groups Black 0.0
Least Targeted Block Groups Latino 33.0
Least Targeted Block Groups White 38.9
Jurisdiction Total Asian 7.3
Jurisdiction Total Black 2.1
Jurisdiction Total Latino 39.4
Jurisdiction Total White 24.0

Household Income

Compared to Fresno, Calif. overall, the most-targeted block groups had:

  • A greater proportion of households that made less than $45K a year.
  • A smaller proportion of households that made between between $75K and 100k a year.
  • A smaller proportion of households that made between $120k and 150K a year.
  • A smaller proportion of households that made $200K and above a year.

Compared to the Fresno, Calif. overall, the least-targeted block groups had:

  • A smaller proportion of households that made less than $45K a year.
  • A smaller proportion of households that made between $75K and 100K a year.
  • A greater proportion of households that made between $120K and 150K a year.
  • A greater proportion of households that made $200K and above a year.

Targeting Level Demographic Proportion of Block Group pop.
Most Targeted Block Groups $120k - 150k 0.0
Most Targeted Block Groups $75k - 100k 2.1
Most Targeted Block Groups $200k and above 0.0
Most Targeted Block Groups Less than 45k 50.9
Median Targeted Block Groups $120k - 150k 1.8
Median Targeted Block Groups $75k - 100k 9.3
Median Targeted Block Groups $200k and above 3.4
Median Targeted Block Groups Less than 45k 23.5
Least Targeted Block Groups $120k - 150k 1.3
Least Targeted Block Groups $75k - 100k 3.1
Least Targeted Block Groups $200k and above 5.9
Least Targeted Block Groups Less than 45k 16.9
Jurisdiction Total $120k - 150k 1.3
Jurisdiction Total $75k - 100k 4.9
Jurisdiction Total $200k and above 1.6
Jurisdiction Total Less than 45k 33.9

Public Housing

In Fresno, Calif. 22 percent of public housing was on block groups the software targeted the most, 72 percent of public housing was on block groups the software targeted more than the median.

The table below provides how many predictions each block with public housing received. The final column tells us the percentage of days a block received predictions from PredPol’s software between Feb 15, 2018 and Jun 01, 2020. We confirmed these dates with the Fresno, Calif., police department.

Census GEOID Block Predictions Num. Public Housing Units Pct. days w/ Predictions
060190029052 2000 4730 1 100.0000000
060190034002 2002 5758 16 99.8806683
060190020001 1024 3989 1 99.7613365
060190045052 2006 4155 1 99.7613365
060190014071 1003 4164 1 99.6420048
060190005012 2006 5205 1 99.5226730
060190054091 1001 3450 1 99.5226730
060190038052 2000 3488 1 99.4033413
060190048021 1002 4224 1 97.6133652
060190006004 4013 2756 1 95.8233890
060190053021 1000 2947 2 94.7494033
060190014073 3007 2223 1 90.9307876
060190013043 3001 1359 3 90.3341289
060190047014 4003 1968 1 89.9761337
060190009022 2003 1515 5 89.4988067
060190030031 1007 1308 1 89.1408115
060190053042 2005 1570 1 84.4868735
060190012021 1000 1261 2 83.2935561
060190044041 1018 819 2 81.8615752
060190014071 1001 1407 1 79.9522673
060190038052 2003 1296 1 79.8329356
060190047041 1007 1389 3 77.5656325
060190009011 1000 1248 3 77.0883055
060190044041 1021 852 2 76.0143198
060190002001 1010 1217 11 73.9856802
060190029032 2010 985 8 73.2696897
060190044041 1032 806 1 72.9116945
060190009023 3001 900 4 70.5250597
060190042112 2001 778 1 69.6897375
060190001001 1024 845 3 65.9904535
060190029031 1003 810 1 65.1551313
060190038091 1007 832 2 55.9665871
060190008001 1003 524 2 54.5346062
060190054031 1000 653 1 52.0286396
060190007003 3002 512 1 51.3126492
060190042121 1046 509 4 49.8806683
060190044041 1019 455 1 49.0453461
060190001001 1031 546 2 48.8066826
060190044042 2015 454 1 48.0906921
060190049012 2015 497 1 47.6133652
060190009023 3002 667 2 46.8973747
060190014114 4005 514 1 44.3914081
060190038072 2000 466 1 42.6014320
060190009021 1006 583 2 42.1241050
060190002001 1009 351 3 39.4988067
060190004003 3047 344 1 38.1861575
060190048014 4000 378 2 37.9474940
060190003003 3011 571 39 36.7541766
060190055101 1022 276 1 30.4295943
060190044041 1028 293 1 30.0715990
060190024002 2000 413 11 28.5202864
060190001001 1029 266 1 28.5202864
060190047041 1000 376 1 25.8949881
060190025022 2000 197 1 20.2863962
060190006001 1015 188 1 19.9284010
060190042072 2024 231 8 17.1837709
060190044041 1020 167 2 16.9451074
060190010001 1007 246 1 16.2291169
060190009023 3017 147 1 14.9164678
060190012013 3001 136 1 14.2004773
060190008001 1002 116 17 12.6491647
060190003003 3016 121 1 12.1718377
060190007003 3009 120 3 11.0978520
060190044042 2016 83 2 9.5465394
060190053052 2010 87 3 9.5465394
060190003003 3010 82 5 7.7565632
060190044042 2030 60 2 7.1599045
060190004003 3086 58 1 6.8019093
060190044042 2034 56 1 6.6825776
060190014102 2012 56 1 6.2052506
060190044042 2032 46 1 5.2505967
060190003003 3012 57 3 5.2505967
060190053022 2021 43 1 5.1312649
060190029031 1002 40 2 4.7732697
060190013042 2004 46 1 4.6539379
060190021005 5004 40 2 4.5346062
060190003003 3008 44 1 4.5346062
060190042072 2020 32 1 3.5799523
060190008001 1001 30 8 3.5799523
060190009023 3000 29 34 3.2219570
060190003003 3018 27 1 3.2219570
060190003003 3009 28 2 2.8639618
060190025022 2006 20 1 2.2673031
060190002003 3000 20 28 2.0286396
060190023003 3037 13 1 1.4319809
060190009011 1006 11 1 1.1933174
060190013042 2001 9 1 1.0739857
060190002003 3002 11 8 1.0739857
060190009023 3006 8 1 0.9546539
060190042072 2038 7 2 0.8353222
060190003003 3013 9 4 0.8353222
060190004003 3067 5 1 0.5966587
060190044081 1002 4 1 0.4773270
060190002001 1011 3 4 0.3579952
060190044042 2012 3 3 0.3579952

Maps

Predictions

Density Map

The map below aggregates all the predictions Fresno, Calif., received in our analysis window into a 2D grid. Each square of the grid represents an area approximately 500 ft. x 500 ft., the size of the PredPol prediction box. The color represents the number of predictions that occurred within the square. The more predictions, the darker the square.

Prediction Count Description
0 - 467 Least Predictions
467 - 934
934 - 1400
1400 - 1870
1870 - 2330
2330 - 2800 Most Predictions

Sources: Markup, Predpol

The grid drawn on this map provides an approximate aggregation of the prediction data. The actual prediction box in the reports provided to departments will vary from the ones shown above.

Choropleth

This map shows the predictions aggregated to the level of the Census block group. Aggregating prediction data to the geographic area of a Census block group introduces additional complexity to the analysis, and hence this map should be interpreted with some caution. See the limitations section of the methodology for more details.

Source: Markup, Predpol

Race and Ethnicity

Black

Source: 2018 five-year ACS.

Latino

Source: 2018 five-year ACS.

White

Source: 2018 five-year ACS.

Household Income

Less than $45k

Source: 2018 five-year ACS.

$75k - $100k

Source: 2018 five-year ACS.

$125k - $150k

Source: 2018 five-year ACS.

$200k and above

Source: 2018 five-year ACS.

Methods

We analyzed the distribution of PredPol predictions for Fresno, Calif. at the geographic level of a Census block group, which is a cluster of blocks with a population of between a few hundred to a few thousand people, generally. There are 335 block groups in Fresno, Calif., the smallest block group had a population of approximately 369 and the largest had a population of approximately 8,380.

In Fresno, Calif., we analyzed 751,218 predictions and used there locations to determine the block groups that were targeted the most, the median and the least. This data sheet presents the breakdown of the racial groups and household income ranges of the people who lived in those block groups. We also present the breakdowns for Fresno, Calif. overall for comparison. The predictions we analyzed were between Feb 15, 2018 and Jun 01, 2020 , we received confirmation that Fresno, Calif. department used the software between Apr 01, 2016 and Jun 01, 2020.

For the race/ethnicity and income analyses, we merged 2018 five-Year American Community Survey data and prediction data and observed the makeup of block groups that were targeted above and below the median, those targeted the most and those targeted the least. For the sake of consistency in our analysis we only used demographic groups for which we had reliable population estimates for all the jurisdictions in our data set. These are:

  • Racial Groups
    • Black
    • Asian
    • Latino
    • White
  • Household Income
    • Less than $45K
    • Between $75K-$100K
    • Between $125K-$150K
    • Greater than $200K

Definitions

We used the Census’ “designated place” boundaries as the boundaries for most jurisdictions. For Sheriff’s departments we confirmed the boundaries with the department.

We defined the most-targeted block groups as those in Fresno, Calif. which encompassed the highest five percent of predictions. We defined the median-targeted block groups as the five percent around the median block group for predictions. And we defined the least-targeted block groups as those with the bottom five percent of predictions.

In some of the larger jurisdictions, more than five percent of block groups got zero predictions. In those cases, we chose the most populated block groups with no predictions for the five percent. Learn more about how we did this in our methodology.

We identified public housing through HUD’s online lookup tool available at https://resources.hud.gov

Data

The data used to generate this analysis can be found in our GitHub repository. It also contains the URLs for the rest of the data sheets from our analysis.