Methodology.

Project architecture

This project proceeded in two analytical phases. The first phase was a state-level spatial audit testing whether public health infrastructure density predicts immigration enforcement intensity - the original ICEberg hypothesis. The second phase, which became the primary finding, was an individual-level audit of 713,464 arrest records examining whether the administrative data pipeline through which ICE locates individuals constitutes a source of measurable bias in the dataset itself. The two phases are analytically connected: the null result from the spatial audit redirected the inquiry toward the enforcement dataset's internal structure, where the bias was ultimately located.

Phase one - Spatial audit (state-level analysis)

Research question

Are states with higher public health administrative visibility - measured through federally qualified health center (FQHC) infrastructure density - associated with higher immigration enforcement intensity, after controlling for enforcement policy agreements, sanctuary status, border geography, and demographic composition?

The theoretical premise, drawn from James Scott's concept of legibility in Seeing Like a State (1998) and Foucault's analysis of administrative enumeration, is that data production through public service systems makes populations more visible to state institutions - and that this visibility may translate into heightened enforcement exposure.

Data sources

1. ICE Enforcement Data - Deportation Data Project. Source: deportationdata.org (FOIA-obtained from ICE). Coverage: October 2022 through early March 2026. Records: 713,000 arrests, 610,000 detainer requests. Unit of analysis for Phase One: U.S. state (aggregated from individual records). Key variables: apprehension_state, facility_state.
2. HRSA Health Center Service Delivery Sites. Source: data.hrsa.gov (updated daily, downloaded April 2026). Coverage: all federally-funded health center sites in the United States. Unit: individual site, aggregated to state level. Key variable: site count per state, normalized to rate per 100,000 residents.
3. American Community Survey (ACS) 2022 5-Year Estimates. Source: U.S. Census Bureau. Variables: total population, foreign-born population, poverty count, uninsured population (ages 19-64, male and female). Derived controls: foreign_born_share, poverty_rate, uninsured_share_proxy.
4. Policy controls (coded from ILRC and ICE public records): 287(g) agreement status (binary, state level), sanctuary policy designation (binary), border state designation (binary: TX, CA, AZ, NM).

Measurement: Administrative Accessibility Index (AAI)

The AAI is constructed as the z-score of log-transformed FQHC service delivery sites per 100,000 residents:

The log transformation is applied to account for right skew in site density distributions (California alone has over 3,000 sites). The z-score standardization allows interpretation in standard deviation units across states.

Critically, demographic variables (foreign-born share, uninsured rate) are excluded from the index and reserved as controls only. Including them in the index would create a construct validity problem - placing the same variable on both sides of the regression - which is a standard methodological error in structural bias audits.

Regression specification

Outcome variables:

• ICE arrests per 100,000 residents (continuous)
• ICE detainers per 100,000 residents (continuous)

Main specification (OLS with HC3 heteroskedasticity-robust standard errors, N=51 states):

HC3 robust standard errors (MacKinnon-White) are used throughout to account for heteroskedasticity in small samples (N=51). The interaction term (AAI × foreign_born_share) tests whether the visibility-enforcement relationship is stronger in states with larger immigrant populations - a key theoretical prediction of the legibility hypothesis.

Additional specifications were estimated without the interaction term for stability checks. Poisson models with population offset were estimated as a supplementary specification (results consistent with OLS for policy coefficients; Poisson dropped from main reporting due to overdispersion concerns at N=51).

Robustness: permutation placebo test

To evaluate whether the AAI-enforcement relationship is stronger than would be expected under random spatial assignment, we conduct a state-stratified permutation test:

1. Compute the observed OLS coefficient on AAI in the full model.
2. Randomly permute AAI values across all 51 states 999 times.
3. For each permutation, re-estimate the model and record the AAI coefficient.
4. Compute the empirical p-value as the share of permuted coefficients with absolute value ≥ the observed coefficient.

State-level (rather than unconstrained) permutation is used to preserve spatial clustering in the null distribution - a methodologically important choice, as naive shuffling produces an artificially narrow null distribution for spatially autocorrelated data.

Results (Phase One)

The original hypothesis is rejected. Administrative visibility through public health infrastructure does not amplify enforcement - and for detainers, the relationship runs in the opposite direction. States with denser FQHC networks tend to be states that also adopt protective immigration policies; the two reflect the same underlying political orientation rather than a data-exposure mechanism.

The 287(g) finding - the strongest predictor across all models and specifications - pointed toward law enforcement data infrastructure as the operative legibility mechanism. This became the organizing insight for Phase Two.

Phase two - Pipeline bias audit (individual-level analysis)

Research question

Does the administrative data pipeline through which ICE locates individuals - criminal justice databases, police systems, or community enforcement - independently predict deportation outcomes after controlling for stated criminality classification, and does pipeline assignment vary systematically by nationality in ways that constitute a measurable bias in the dataset itself?

Theoretical frame: construct validity failure

The analysis is grounded in the measurement validity literature. Friedler et al. (2019) define construct validity failure as the condition where a measured variable does not correspond to the latent construct it claims to represent - producing disparate treatment of equivalent cases when the variable is used in downstream classification or decision-making.

The ICE dataset's apprehension_criminality variable is presented as a measure of legal status. We show it is confounded by apprehension_method - the data pipeline - because different pipelines systematically surface different populations regardless of underlying legal situation. This is not a claim about individual officer bias but about the structural properties of the dataset as a measurement instrument.

Kate Crawford (2021, Atlas of AI) and Virginia Eubanks (2018, Automating Inequality) provide the broader frame: technical categories that appear neutral can launder discrimination through the infrastructure of data collection. The bias is not declared - it accumulates across a chain of administrative decisions, each individually defensible, collectively producing systematic disparate outcomes.

Data

Dataset: Deportation Data Project - ICE Arrests. Source: deportationdata.org (FOIA-obtained). Records: 713,464 individual arrest events. Period: October 2022 through early March 2026. Format: individual-level records with apprehension date, method, criminality classification, threat level, citizenship country, gender, final program, case status, and departure information.

Key variables:

• apprehension_method - how ICE located the individual (28 distinct values)
• apprehension_criminality - stated legal category at time of arrest (3 values)
• case_threat_level - enforcement priority score (1=highest, 3=lowest)
• case_criminality - stated legal category at case resolution (3 values)
• final_program - enforcement program assignment (14 values)
• case_status - outcome including deportation/removal status
• citizenship_country - country of citizenship (206 distinct values)
• gender - Male, Female, Unknown

Pipeline taxonomy

The 28 distinct values of apprehension_method are consolidated into five analytical pipeline categories based on the administrative system ICE queried to locate the individual:

The primary analytical comparison is CAP vs Community, representing the two largest pipelines and the maximal contrast between database-mediated and database-absent enforcement.

Outcome variable

The primary outcome is deportation, coded as a binary indicator from case_status:

• Deported = 1 if case_status contains "Deported", "Removed", or "Excluded"
• Deported = 0 otherwise (active cases, voluntary departure, alternatives to detention)

This coding captures formal removal orders executed, excluding cases still pending. Among the 713,464 records, 427,264 (59.9%) result in deportation under this definition.

Analytical strategy

The analysis tests four sequential claims that together establish the pipeline bias:

Mediation analysis (partial): to estimate what fraction of the nationality-outcome gap operates through pipeline assignment, we compute deportation rates within the same-criminality, same-pipeline cells. The reduction in the nationality gap when controlling for pipeline provides an estimate of pipeline-mediated disparity. This is a descriptive mediation (no causal identification), appropriate for dataset bias characterization.

No regression model is estimated in Phase Two. The analysis is deliberately descriptive - the claim is about dataset structure, not causal effect identification. A regression controlling for pipeline would be circular: pipeline is part of the bias mechanism, not a confounder to be removed.

Limitations

Selection concern: ICE may deliberately use CAP for higher-priority cases, making pipeline correlated with unmeasured case severity beyond what criminality captures. We control for stated criminality (the only available severity proxy), and the gap persists within every category. However, we cannot rule out that CAP cases have systematically more severe underlying legal situations not captured in the data.

Missing threat scores: 63.4% of records have null case_threat_level values. The threat level analysis is conducted on the available subset (261,476 records). If missingness is systematic - e.g., threat levels are disproportionately recorded for certain pipelines or nationalities - the threat level findings may not generalize to the full dataset.

Temporal non-stationarity: the pipeline mix changed substantially across the 2022-2026 period (CAP share: 30% in 2022, 61% in 2024, 21% in 2025). Point estimates are unstable over time and should not be interpreted as reflecting a structural constant. The 2025 drop in CAP share likely reflects a policy shift in enforcement priorities.

Observational design: all findings are associational. The analysis establishes that pipeline and outcome are correlated, and that pipeline and nationality are correlated, but does not identify a causal pathway from nationality → pipeline → outcome in a statistical sense. The causal argument is theoretical, grounded in the known operational logic of the CAP program.

Nationality proxy: citizenship country is used as a proxy for the communities ICE targets with each enforcement method. This conflates citizenship with national origin community membership and does not capture mixed-status households or long-term residents with non-citizen status. The analysis is descriptive of the dataset's internal structure, not a claim about individual-level targeting.

Figure-by-figure analytical notes

Figure 1 - Pipeline deportation gap within identical criminality labels

This figure presents the core bias finding. Each panel shows deportation rates by pipeline within a single criminality category - holding the legal label constant and varying only the data source through which ICE found the person. The key comparison is CAP (Criminal Database) vs Community Enforcement.

All comparisons are statistically significant at p < 2e-308. The "Other Immigration Violator" category is analytically the most important. By definition, these are people with no criminal record - their criminality label is the minimum possible. The 1.83× ratio in this category cannot be explained by differential criminality. The gap is entirely a function of which database found the person.

Figure 2 - Nationality determines pipeline assignment

This figure shows the structural targeting layer: which pipeline ICE uses varies systematically by citizenship country, independent of where the person was actually located or what they were doing.

The CAP rate ranges from 10.8% (Peru) to 40.7% (Mexico) - a 3.8× spread across nationalities. This spread is not explained by geographic proximity to detention facilities or border location: Honduran, Salvadoran, and Guatemalan nationals are arrested via CAP at much higher rates than Mexican nationals despite similar Central American origin.

Critically, the criminality composition of arrests also differs substantially. For Venezuela, 65.6% of arrests are coded as "Immigration Violator" and only 10.7% as "Convicted Criminal." For Mexico, 54.4% are "Convicted Criminal." This means Mexico is both more likely to be found via database AND more likely to receive the highest-severity criminality label - both effects compound in the direction of higher enforcement intensity.

Figure 3 - Pipeline determines enforcement track (final program)

This figure shows that pipeline assignment does not merely affect deportation probability - it determines which institutional track the person enters, which then structures all subsequent processing.

The ERO Criminal Alien Program is the highest-priority deportation track - the program responsible for the vast majority of formal removals. A CAP arrest is almost certain (90.4%) to land in this program. A community enforcement arrest has a 32.9% probability of entering ERO-CAP, with the remaining 67.1% distributed across non-detained dockets and alternatives to detention - substantially less intensive enforcement tracks that often involve no detention and extended docket timelines. The pipeline locks in the enforcement path.

Figure 4 - Threat score contamination within same criminality category

This figure examines the case_threat_level variable - a score of 1 (highest) to 3 (lowest) assigned by ICE to prioritize enforcement resources. This score is supposed to reflect threat to public safety or national security. We show it also reflects which pipeline found the person. Available for 261,476 records (36.6% of total; missing pattern not fully explained).

Within the Convicted Criminal category - where all individuals carry the same stated legal label - a CAP arrest generates a threat level 1 score 50.7% of the time. The same label generated through 287(g) produces a threat level 1 score only 26.2% of the time: less than half the rate. The threat score is supposed to measure the individual. It partly measures the database. This is the compounding mechanism: pipeline → label → threat score → program → outcome.

Figure 5 - The enforcement shift 2022-2025

This figure shows that the pipeline mix is not static - it changed substantially during the analysis period, and the direction of change amplifies the biases documented above.

From 2022 to 2024, CAP's share of all arrests more than doubled (30% to 61%) while community enforcement's share collapsed by two-thirds (53% to 17%). This was not a gradual drift but a structural policy shift - a deliberate reorientation of ICE enforcement toward database-mediated operations. The overall deportation rate tracked almost perfectly with the CAP share, rising from 26.9% in 2022 to 62.5% in 2024 (an increase of 35.6 percentage points). The 2025 data shows a reversal - CAP share drops to 21.5% and the deportation rate falls to 55.4% - suggesting the shift was policy-responsive rather than permanently structural.

Data & code availability

All data and code are available in the project repository. The core analysis runs via python3 analysis.py; the Phase One spatial pipeline runs via ./run_pipeline.sh.

Raw data sources:

• ICE arrests: deportationdata.org/data.html
• HRSA sites: data.hrsa.gov/data/download
• ACS 2022: census.gov/programs-surveys/acs

Cite enforcement data as: "Government data provided by ICE in response to a FOIA request, processed by the Deportation Data Project."