Algorithmic & Data Pipeline Bias in U.S. Immigration Enforcement Data

Overview

A two-phase computational audit of U.S. immigration enforcement data.

Phase 1 (Spatial Audit): Tests whether public health infrastructure density predicts state-level ICE enforcement intensity. Result: null. The hypothesis is rejected.

Phase 2 (Pipeline Bias Audit): Audits 713,464 individual ICE arrest records for evidence that the administrative data pipeline through which ICE locates individuals constitutes a source of measurable bias. Result: yes, substantially and robustly. Community vs CAP odds ratios range 0.357–0.592 across 7 specifications. All p < 1e-100.

Data Sources

Phase 1: ICE Enforcement (DDP/FOIA) + HRSA FQHC Sites + ACS 2022

Phase 2: ICE Arrests individual-level, 713,464 records, Oct 2022–Mar 2026

Setup & Run

pip install -r requirements.txt

# Phase 1
bash phase1_spatial/run_phase1.sh

# Phase 2 (requires arrests_core.csv from Colab notebook first)
bash phase2_pipeline/run_phase2.sh

Key Results

Phase 1 - Null Result

• AAI → Arrests: β=−12.8, p=0.511 - null
• AAI → Detainers: β=−50.3, p<0.001 - negative (reversed)
• 287(g) → Arrests: +46.9 (p=0.020) - dominant predictor

Phase 2 - Pipeline Bias (Community vs CAP Odds Ratios)

Model	OR	p
B: Threat + Pipeline	0.376	<2e-308
D: + Year FE + Criminality	0.474	<2e-308
E: + AOR Fixed Effects	0.448	<2e-308
Sensitivity (resolved only)	0.592	<1e-100
Sensitivity (+ vol. departure)	0.357	<2e-308

Figure Index (Phase 2)

Fig	Content	Addresses
1	Core bias + ACTIVE upper bound	Right-censoring concern
2	Nationality → pipeline	Structural targeting
3	Within-AOR gap (8 field offices)	Geographic confound
4	Pipeline → enforcement track	Mechanism
5	Threat score contamination	Compounding
6	All 7 models OR forest plot	All debunks
7	Year × pipeline ratio	Policy responsiveness
8	Records integrity	Professor feedback
9	Extended outcome	Voluntary departure

Limitations

1. No causal identification - pipeline selection is not random
2. Records falsification concern - DDP flags reliability issues; 98.3% label consistency is unusual
3. Policy confounders - 2022–2024 CAP expansion; year FEs partially control
4. Unobserved confounders - prior orders, legal representation, court backlog
5. Selection - only arrested individuals in dataset