Federal criminal sentencing: race-based disparate impact and differential treatment in judicial districts

The study investigates why racial sentencing inequities persist in the U.S. federal system despite guidelines intended to ensure similar sentences for similar conduct. Prior research and government reports document substantial disparities, notably longer sentences for Black and Hispanic men compared to similarly situated white men. Potential sources include structural disparate impact arising from criminal history and offense types embedded in guidelines, district-level practices in applying guidelines, and differential treatment by courtroom actors (judges, prosecutors, defense attorneys), as characterized by the focal concerns framework. The authors aim to quantify both structural disparate impact and potential differential treatment nationally and across individual federal judicial districts, particularly in the advisory-guidelines era post-Booker.

Empirical work consistently finds racial disparities in sentencing outcomes in federal and state courts, with especially robust evidence for Black-white gaps and additional disparities affecting Hispanic and Native defendants. The post-Booker era has mixed findings: some studies suggest increased Black-white disparity, while others do not; overall, several reports indicate declining disparities over time. The focal concerns framework posits that courtroom actors infer blameworthiness, community protection, and practical constraints using stereotypes, potentially leading to biased outcomes favoring white defendants. Structural contributors include guideline designs rooted in past (possibly biased) practices and statutory regimes like drug mandatory minimums that have disproportionately impacted Black defendants. Interdistrict and interjudge variability is documented, yet analysis at the judge level is limited due to data redaction. Multi-level studies show district-level practice differences (e.g., downward departures), and USSC has reported interdistrict variation. Prior work cautions against aggregation bias when pooling jurisdictions, motivating district-level analyses.

Data were drawn from U.S. Sentencing Commission Individual Offender Files (FY2006–FY2020). To avoid Booker-era confounding and focus on comparable cases, the authors excluded immigration cases (OFFGUIDE=17, OFFTYPE2=27, or OFFTYPSB=27, by year) and noncitizens (NEWCIT=1), following prior practice. This removed 479,347 records, leaving 548,629. The outcome is total prison sentence in months (SENTTOT). For 82,101 missing SENTTOT records, 75,338 involved non-incarcerative sanctions (probation/fines/home confinement) and were imputed as 0 months; 6,763 remaining missing outcomes were dropped (final 541,866). Life sentences coded as 470 months (N=2,295) were retained; sentences >470 months (N=1,570) were dropped to avoid nonmonotonic coding, yielding 540,296 records. Records missing key variables were removed: race (NEWRACE), age (AGE), sex (MONSEX), education (NEWEDUC), offense-level and guideline variables (CHAP2, COADJLEV, XFOLSOR, XCRHISSR, XMIN-SOR/XMINSOR, GLMIN), plea (NEWCNVTN), and government departure indicator (BOOKERCD or SENTRNGE), dropping 21,477 cases (4.0%), resulting in 518,819. The Northern Mariana Islands (N=98) were excluded for insufficient sample size, yielding the final analytic dataset N=518,721. Variables except AGE and SENTTOT were treated categorically, including offense levels and criminal history, to avoid imposing linearity. Derived variables included: MANDMIN (true if GLMIN > XMINSOR), GOVDEPART (true if government-sponsored downward departure present via BOOKERCD/SENTRNGE), and GRID, the categorical pairing of final offense level (XFOLSOR) and criminal history group (XCRHISSR) representing the presumptive guideline cell. National analyses employed 12 linear regression models progressively adding controls: demographics (AGE, MONSEX, NEWEDUC), year (SENTYR), plea (NEWCNVTN), criminal history (XCRHISSR), offense level adjustments (CHAP2, COADJLEV, XFOLSOR), GRID, MANDMIN, GOVDEPART, and district (CIRCDIST). Heteroskedasticity-robust standard errors were used; racial disparities were estimated as differences in race coefficients versus white. Bonferroni adjustments were applied across multiple estimates. District-level analyses used two specifications: District Model I included national coefficients for controls but interacted district with race (NEWRACE:CIRCDIST) to estimate district-specific racial gaps under a uniform structural application; District Model II fit Model 11 separately within each district, allowing district-specific impacts of demographics, grid application, and other factors. Robust SEs and Bonferroni adjustments were again used. Model diagnostics and data are archived in the referenced repository.

National level (N=518,721): Unconditional mean sentence differences versus white defendants: Black +18.5 months (±0.5), Hispanic +5.3 (±0.5), ARI −9.0 (±0.9). Conditioning sequentially reduced disparities. Key steps: demographics reduced Black gap to 12.9 and Hispanic to 1.0 months; adding plea reduced Black to 10.3; adding criminal history dropped Black to 2.2 but increased Hispanic to 4.6; incorporating offense-level components and grid cell substantially reduced gaps (Black 5.4; Hispanic nonsignificant); adding mandatory minimums and government departures further reduced Black to 1.9 and yielded a slight Hispanic advantage (−1.1 months), which became 0.0 after adding district fixed effects. Final fully adjusted Model (12): Black-white +1.9 months (significant), Hispanic-white 0.0 (ns), ARI-white +2.8 months (significant). These patterns indicate strong structural disparate impact via demographics, criminal history, guideline grid, mandatory minimums, and departures, accounting for nearly 17 of the 19 months for Black and all of the 5 months for Hispanic unconditional disparities. District level: District Model I (uniform structural application) found significant unexplained disparities in 22 districts (17 Black, 3 Hispanic, 4 ARI). Examples include Black-white disparities up to +13.0 months in E.D. Virginia and Hispanic-white up to +9.3 months in W.D. North Carolina; ARI-white up to +12.5 months in Arizona. District Model II (district-specific structures) found 14 districts with significant unexplained disparities, each for one minoritized group: 11 Black-white disparities (e.g., S.D. Iowa +7.8 ± 2.8; E.D. Missouri +3.8 ± 1.1; W.D. Missouri +3.8 ± 1.5) and 3 ARI-white disparities (Arizona +14.2 ± 1.6; New Mexico +7.8 ± 2.7; Wisconsin Eastern +7.9 ± 4.0 in Model I; ARI in Model II for AZ and NM noted). Many disparities observed in Model I were reduced or eliminated in Model II, indicating substantial district-specific structural disparate impacts; however, the 14 remaining significant gaps suggest possible differential treatment by courtroom actors.

The study demonstrates that much of the raw national racial disparity in federal sentencing is explained by structural factors—defendant demographics, criminal history, guideline grid placement, mandatory minimums, and government-sponsored departures—consistent with prior literature. Nonetheless, after accounting for these, a small but significant Black-white disparity remains nationally, and an ARI-white disparity appears, though interpretation is limited due to the heterogeneous ARI category. Importantly, analyses at the district level reveal that national aggregation can mask meaningful local disparities. District Model I highlights numerous districts where racial gaps persist under uniform structural assumptions, potentially reflecting both structural disparate impact and/or differential treatment. District Model II shows that allowing district-specific structures reduces many of these disparities, evidencing structural differences in how guidelines and other factors are applied within districts. However, 14 districts retain significant racial gaps even after district-specific conditioning, which the authors interpret as evidence of possible differential treatment by judges, prosecutors, and defense attorneys. The geography of the remaining gaps suggests concentrations for Black-white disparities in certain circuits and regions (plains, mid-Atlantic, deep South) and ARI-white disparities in states with higher Native American populations, though the ARI category and jurisdictional complexities limit precise interpretation.

The paper integrates sources and loci of racial inequity in federal sentencing by jointly quantifying structural disparate impact and potential differential treatment at national and district levels. Structural factors explain most of the national disparities, yet residual gaps persist for Black defendants, and multiple districts exhibit significant unexplained differences after district-specific controls, indicating possible bias in courtroom decision-making. The work underscores the importance of district-level analyses to avoid aggregation bias and identifies specific districts where scrutiny and policy interventions may be warranted. Future research directions include disaggregating by offense type, examining the imprisonment decision in addition to sentence length, and improving transparency of courtroom actor identities (e.g., unredacting judge information) to enable more precise, causal assessments of differential treatment.

1. Focus on sentence length, not the incarceration versus non-incarceration decision, which itself can drive disparities. 2) No disaggregation by offense type; district-specific disparities may vary by offense. 3) Exclusion of noncitizens and immigration cases limits generalizability, particularly for understanding disparities affecting Hispanic defendants. 4) Evidence of differential treatment is statistical and contingent on modeling assumptions; it is not proof of bias. 5) The analysis cannot identify which specific courtroom actors may contribute to disparities. 6) Potential biases in other districts may be masked or counterbalanced by opposing effects, leading to non-detection.