Secondary Analysis - The ICEberg Effect

The Original Hypothesis

Our initial research question was whether Federally Qualified Health Centers - public health infrastructure specifically designed to serve uninsured and undocumented populations - might inadvertently create administrative visibility that increased ICE enforcement exposure. The theoretical frame, drawn from James Scott's concept of legibility and Foucault's analysis of administrative enumeration, predicted that communities more deeply embedded in public administrative systems would face higher enforcement intensity.

We tested this with a state-level computational audit combining ICE enforcement data from the Deportation Data Project, HRSA health center data, and ACS demographic controls.

Original Research Question

Are states with higher public health administrative visibility - measured through FQHC infrastructure density - associated with higher immigration enforcement intensity, after controlling for 287(g) agreements, sanctuary policy, border geography, and demographic composition?

Data & Method

We built a state-level panel combining three public datasets:

U.S. states analyzed

713K

ICE arrests aggregated to state level

610K

ICE detainer requests

999

Permutation iterations for robustness

Administrative Accessibility Index (AAI)

The AAI captures public health administrative exposure as the z-score of log-transformed FQHC service delivery sites per 100,000 residents. Demographic variables (foreign-born share, uninsured rate) are used only as controls, not as index components, to avoid placing the same construct on both sides of the regression.

Administrative Accessibility Index

\mathrm{AAI}_{s} = z\!\left(\, \log\!\bigl(1 + \tfrac{\mathrm{FQHC}_{s}}{\mathrm{Pop}_{s}/10^{5}}\bigr) \,\right)

Regression Specification

OLS models with HC3 robust standard errors. Controls include foreign-born share, poverty rate, uninsured share, 287(g) status, sanctuary policy, and border state designation. Robustness is checked via 999-iteration state-level permutation tests.

Estimating Equation

Y_{s} = \beta_{0} + \beta_{1}\,\mathrm{AAI}_{s} + \beta_{2}\,(\mathrm{AAI}_{s}\!\times\!\mathrm{FB}_{s}) + \boldsymbol{\gamma}^{\!\top}\!\mathbf{X}_{s} + \varepsilon_{s}

where $Y_{s}$ is arrests or detainers per 100k in state $s$ , $\mathbf{X}_{s}$ the control vector, and errors are HC3-robust.

Results

Finding 1 - Null: Health Visibility Does Not Predict Arrests

The AAI coefficient on arrests per 100k is -12.8 (SE=19.4, p=0.511). The permutation placebo confirms this: empirical p=0.570. FQHC density is not associated with higher ICE arrest rates after controlling for enforcement policy variables.

Figure A1 · AAI vs ICE arrests per 100k (51 states)

Fig. A1The relationship between Administrative Accessibility (AAI) and arrest intensity is weakly negative and statistically indistinguishable from zero (slope −28.7, p=0.511). Hover any point to identify the state.

Figure A1b · Top 15 states by ICE arrests / 100k

Fig. A1bTop 15 states by ICE arrest rate. Above-median FQHC density (red) is interleaved with below-median (blue) - visual confirmation that AAI alone does not structure arrest intensity. Texas dominates regardless.

Finding 2 - Negative Association with Detainers

For ICE detainers, the AAI coefficient is -50.3 (SE=13.9, p<0.001), confirmed by permutation (p=0.008). States with denser FQHC infrastructure have fewer detainers per capita - likely because these states also adopt protective immigration policies.

Figure A2 · AAI vs ICE detainers per 100k

Fig. A2States with higher AAI issue fewer detainers per capita. 287(g) participants (red) cluster at higher detainer levels regardless of AAI - a preview of the dominant policy variable. Toggle by hovering points.

Finding 3 - 287(g) Is the Dominant Predictor

287(g) agreements are the strongest predictor across all specifications: +46.9 arrests per 100k (p=0.020) and +77.6 detainers per 100k (p=0.006). This result led us toward the pipeline analysis in the main analysis.

Figure A3 · OLS coefficients (HC3 robust SEs, N=51)

Fig. A3Across both outcomes, state_287g is the dominant policy predictor; AAI is null for arrests and significantly negative for detainers. Hover bars for exact β and significance.

Full Regression Table (OLS with HC3 robust SEs, N=51)

Term	Arrests/100k β	SE	p	Detainers/100k β	SE	p
AAI (FQHC density)	-12.8	19.4	0.511	-50.3	13.9	<0.001
AAI × Foreign-born	-279.5	236.0	0.236	+172.4	156.3	0.270
Foreign-born share	+62.6	257.8	0.808	+294.4	206.9	0.155
Poverty rate	+1,722	892.8	0.054	+22.5	464.3	0.961
Uninsured share	-503	1,412	0.722	+1,657	888.2	0.062
287(g) agreement	+46.9	20.2	0.020	+77.6	28.5	0.006
Border state	+94.5	90.3	0.295	+69.2	67.3	0.303
Sanctuary policy	+23.4	22.8	0.303	+31.3	25.8	0.225

Permutation Test Results

Outcome	Observed AAI Coef	Empirical p	N permutations	Interpretation
Arrests per 100k	-12.8	0.570	999	Null confirmed
Detainers per 100k	-50.3	0.008	999	Robust negative

Interpretation

The original ICEberg hypothesis - public health data visibility amplifies enforcement - is not supported at the state level. The direction for detainers is reversed: states with denser FQHC networks have fewer detainers after controls, likely because they also adopt sanctuary policies. Investment in public health safety nets and protective immigration policy reflect the same political orientation.

The state-level null result is not a failure of the project - it is scientifically valuable. It rules out one channel of the legibility mechanism and redirects attention to where the bias actually operates: inside the enforcement dataset itself, at the individual record level, through the pipeline-classification entanglement documented in the main analysis.

Connection to the Main Analysis

The 287(g) finding here - the strongest predictor at the state level - pointed us toward law enforcement data infrastructure as the operative legibility mechanism. That insight became the organizing principle of the main analysis: not health databases, but criminal justice databases. The ICEberg was real. We found it in the right channel.

Limitations of the Secondary Analysis

State-level aggregation (N=51) is coarse and loses within-state variation. The AAI uses FQHC site count rather than patient volume, which would more directly measure administrative record production. ACS controls are 2022 cross-sectional estimates applied to a 2022-2026 outcome window. The design is observational and cannot establish causation.

Caveat - downstream interpretation

The 287(g) coefficient is the bridge between this state-level finding and the individual-level pipeline analysis on the main page. Readers should carry the main-analysis caveats with them: the pipeline-outcome gap there could partly reflect severity selection, geographic and gender composition, voluntary departure asymmetry, and the policy regime in force at each point in time. The state-level evidence here is consistent with a law-enforcement-data legibility mechanism but does not by itself establish individual causation.

Flagged through healthcare data.