Disparities at Points of Prosecutorial Discretion
Terms, Methods, Limitations

• Case: A collection of charges against a defendant arising out of a single incident. In this analysis, cases are classified by their most serious charge.
• Controlling: Taking other factors into account when conducting an analysis. For example, “after controlling for age, gender, criminal history” means that we are taking potential differences in these factors into account when examining the differences between White, Black, and Hispanic defendants.
• Predicted Probability: An estimate of the likelihood of the outcome, based on the defendant’s race/ethnicity, while taking into account the control variables.
• Standard Deviation: A measure of how dispersed the data are in relation to the mean. Low standard deviation means data are clustered around the mean, and high standard deviation indicates data are more spread out.

Data Source

All data was drawn from the Action case management system, stored and shared by the Colorado District Attorneys’ Council (CDAC).

Variable Construction

• Race/Ethnicity: As outlined in the data dashboard, race and ethnicity are determined by law enforcement. For this analysis, we combined the race and ethnicity fields. If an individual was identified as White and Hispanic, we categorized them as Hispanic. If the individual’s race was identified as Black, we categorized them as Black, regardless of their ethnicity. If the individual’s race was identified as Native American, we categorized them as Native American, regardless of their ethnicity. Because we believe that Hispanic individuals were systematically miscategorized as White in the dataset, we used the defendant’s last name to help identify their ethnicity. Based on procedures employed by the Colorado Department of Public Safety in their CLEAR Act reporting, we recategorized any individual as Hispanic who met the following criteria: 1) their race was identified as “White,” “other,” or their race was missing and 2) the 2010 census file (surnames occurring 100 or more times) identified their surname as having 85% or more individuals with that surname as “Hispanic or Latino”. This procedure resulted in the categorization of an additional 1,907 (10.3%) cases with Hispanic defendants.
• Age: We used the following age categories: under 18, 18-25, 26-35, 36-45, and over 45 years old.
• Gender: Law enforcement defines gender in the following categories: male, female, other. Due to the small number of individuals identified as “other,” we limited our analysis to individuals identified as male or female.
• Criminal History: As outlined in the data dashboard, we calculated criminal history based on convictions (since 2007) within the 8 District Attorneys’ Offices participating in the Colorado Prosecutorial Dashboards Project (the 1st, 2nd, 5th, 6th, 7th, 8th, 18th, and 20th). We developed four categories: a) no criminal history, b) prior misdemeanor convictions, c) prior non-violent felony convictions, and d) at least one prior violent felony conviction. The definition of violent was aligned with the definition used in the dashboard.
• Charge Level: Charge represents the most serious filed charge, categorized as a felony or a misdemeanor.
• Charge Class: Charge class represents the most serious filed charge, which we categorized as follows: a) felony 1-4 or drug felony 1-3, b) felony 5-6 or drug felony 4, c) misdemeanor 1-2 or drug misdemeanor 1-2, d) misdemeanor 3, or e) traffic misdemeanor 1-2.
• Charge Type: As defined in the data dashboard, we classified cases by their top charge into the following categories: person or sex, property, drug, driving under the influence (DUI), traffic, weapons, or other.
• Case Length: We calculated the case length as the number of months to case resolution, using the date the case was filed and the date the case was disposed of. We treated values less than zero or more than five years as missing.
• Disposition Quarter: To account for time trends, we constructed a categorical variable representing the calendar year quarter the case was disposed.
• Case Disposition (Dismissal): We identified all cases that had a disposition of “dismissed” for the most serious filed charge. We did not include plea dismissals (a defendant’s case/cases dismissed in exchange for pleading guilty to another case/cases).
• Case Disposition (Deferred Judgment): We identified all cases for which a defendant received a deferred judgment to the most serious filed charge.
• Case Disposition (Plead Guilty): We identified all cases in which the defendant plead guilty to the most serious filed charge.
• Charge Reduction: We developed three ordered categories of charge reduction from filing to disposition: a) no reduction b) within class reduction (for example, a felony 1 to a felony 3 or a misdemeanor 2 to a misdemeanor 4), and c) a charge level reduction (a reduction from a misdemeanor to a petty offense/infraction or a reduction from a felony to a misdemeanor or petty offense/infraction).
• Referral Reduced at Charging: Referral charge reduction represents whether the felony referred was reduced to a misdemeanor at point of filing.
• Incarceration: As defined in the data dashboard, a sentence to incarceration included any of the following sentences: state prison, youth corrections, community corrections, jail, or any sentence of jail and probation, as well as inmate/outmate programs, work release, in-home detention, and weekenders.

The District’s Population was generated from information publicly available through the United States Census Bureau. The 2020 U.S. Census collected race and ethnicity as two questions. In combining the race and ethnicity questions, we used the following logic: If an individual identified as White and Hispanic, we categorized them as Hispanic. If an individual identified as Black, we categorized them as Black, regardless of their ethnicity. If an individual identified as Native American (“American Indian or Alaska Native” or “Native Hawaiian and Other Pacific Islander”) we categorized them as Native American, regardless of their ethnicity. Because data suggest that individuals who identify as Hispanic, often report their race as “some other race,” ( Pew Research Center), we categorized individuals who identified as “Some Other Race” and Hispanic as Hispanic. We categorized individuals that selected two or more races, regardless of their ethnicity, as “Multiracial or Another Race.”

This analysis focused on cases filed and disposed of by the DA’s Office from March 2020 through the end of June 2022. We focused on that timeframe based on feedback from the Office that this was the period for which the most reliable data on race and ethnicity were available. All analyses were focused on the most serious charge.

We focused on defendants identified as Black, White, and Hispanic. Due to small sample size, we excluded individuals identified as Asian (208, 1.1%) or Native American (105, 0.6%). We excluded individuals whose race/ethnicity was unknown (560, 3.0%). To avoid grouping individuals with diverse identities, we also excluded individuals identified as another race/ethnicity (452, 2.4%).

To conduct a complete case analysis, we limited our analyses to individuals with complete data on all of the variables of interest. We excluded individuals with missing data on gender, age, criminal history, any felony referrals declined at filing, open cases, and cases not yet disposed of on July 1, 2022. We limited our analyses to cases that had either a misdemeanor or felony charge at point of filing. We excluded fugitive cases. Lastly we only kept cases with complete data on case length and disposition quarter. This excluded a total of 5,869 cases (34.2%).

We conducted all analyses using Stata 17.0 (Statacorps, 2021). We began by conducting descriptive and bivariate analyses, examining the association between race/ethnicity and all outcomes and covariates.

We used logistic regression to examine the association between the three disposition outcomes (dismissed, deferred judgment, and plead guilty) and incarceration and race/ethnicity. We included gender, age, criminal history, case length, disposition quarter, charge type, and charge class as covariates.

We used a generalized ordinal logistic regression model to examine the association between charge reduction and race/ethnicity. We selected a generalized original logistic model because the likelihood-ratio test suggested that the proportional odds assumption was violated in the ordinal model. We included gender, age, criminal history, case length, disposition quarter, charge class, charge type, and whether the referral charge was reduced at filing as covariates.

For all models, we conducted stratified analysis, looking for potential differential impacts by charge class, charge type, the top five most frequent charges, gender, and age. When presenting results for stratified analysis, we considered categories that had the highest proportion of cases within that outcome (e.g., types of cases most frequently resolved by plea agreement), while aiming to avoid small cell sizes (<50).

To support interpretation, we used the margins command to calculate mean predicted probabilities using the sample values of the other predictor variables. Given sample sizes and equity considerations, we have chosen to report p-value in the technical appendix.

This analysis has a number of limitations:

• We do not have any information on why cases were dismissed or received a deferred judgment. Likewise, it is not possible to tease out whether cases were dismissed because they were referred to or successfully completed a diversion program.
• We are using a proxy measure of criminal history. As noted on the dashboard, we do not have information on convictions outside the 8 pilot DA Offices participating in the Colorado Prosecutorial Dashboard Project or cases from outside the state of Colorado. Likewise, we only have data since 2007. For these reasons, criminal history may be underestimated. However, when benchmarked against data from the Bureau of Justice Statistics, criminal history was found to be of similar magnitude.
• Race and ethnicity is reported to the DA’s Office by law enforcement agencies. Law enforcement currently captures this data through various mechanisms: (1) by linking to prior criminal history records, (2) by scanning a Colorado ID or driver’s license, (3) through fingerprint technology, or (4) based on the officer’s “perceived demographic information of the person contacted” (as required by HB21-1250). Officer assumptions have the potential to lead to inaccurate or inconsistent data. We have attempted to correct potential under-identification of Hispanic individuals using census data. While we were able to test this correction using a small sample of individuals who self-reported race to Jefferson County Pretrial Services, we have no way to assess to what extent this correction is producing accurate results in the full dataset.
• We were not able to examine outcomes for all racial/ethnic groups. We excluded race/ethnicities which represented less than 2% of the overall defendant population. Likewise, in order to calculate reliable percentages and predicted probabilities, we limited our analyses to groups where there were more than 50 individuals of a particular race/ethnicity.
• We were not able to tease apart guilty pleas in relation to other cases; rather, all cases resulting in a guilty plea are counted individually using the top sentence on each unique case. This likely overstates the actual use of different sentences and could impact our results if having multiple cases is not evenly distributed across racial/ethnic groups.
• While we have considered a number of individual and case-level factors in our analyses, it is not possible to consider all unique aspects of the case. For example, examining charge reduction, we have not examined the types of sentences imposed, which may impact whether a defendant received a charge reduction.
• This analysis was limited in its examination of discretion points. For example, due to lack of data, we have not examined felony declination (felony referrals not filed) and bond decisions, which prosecutors provide input on.