Hospitals increasingly attach neighborhood statistics to patient records. These statistics, called social determinants of health (SDoH), include median income, transit dependence and deprivation scores, and they are typically looked up from a patient's ZIP code. They are added to risk models in the hope of capturing the life outside the hospital that the chart never records. The best evidence so far is puzzling. Adding these features to intensive care unit (ICU) models leaves discrimination unchanged. They do not help the model rank sicker patients above healthier ones. They do sharply improve calibration, meaning whether a predicted 20% risk really corresponds to 20% of patients dying, and they reduce false alarms for underserved subgroups (Evaluating the Impact of Social Determinants on Health Prediction). A feature that fixes calibration without improving ranking behaves less like new individual risk information and more like a correction to the baseline rate for groups of patients. That is the same quantity, "mean discrepancy," that a separate line of work identified as the main driver of subgroup unfairness when ICU data are pooled across hospitals (Investigating Data Interventions for Subgroup Fairness: An ICU Case Study).
This note proposes to test which reading is correct. The first possibility is that area-level SDoH features carry social content that changes individual risk. The second is that they act as a regional offset: a label for where a patient came from, which corrects baseline rates distorted by referral geography and local practice. A model whose social features are really place labels will tell clinicians the wrong story when they look at its explanations. It will also transfer badly to a hospital with a different catchment area. We combine three ingredients. The first is a glass-box additive model that separates an area offset from SDoH shape effects. The second is a suppressor-variable correction for those shape effects. The third is a set of permutation controls that remove the meaning of the SDoH values while keeping the grouping structure they induce.
Core Claim
Our hypothesis, which could fail, is that in a single-center ICU cohort a large share of the calibration and false-positive-rate gains from area-level SDoH can be reproduced by any feature that groups patients by area. That includes a random relabeling of the SDoH values across ZIP codes. We expect the gains to track distance from the hospital and transfer status, which reflect regional referral patterns, rather than deprivation as such. We further expect that the SDoH risk curves a standard additive model displays will often have the wrong sign. The reason is that these indices are strongly correlated with each other and with comorbidity, so some of them act as "suppressor variables." A suppressor is a feature that receives a large weight only because it cancels noise in a correlated feature, not because it is associated with the outcome. If both expectations hold, the field's main argument for adding SDoH to acute-care models needs rewording: the benefit would come from a regional calibration patch, not from modeling social risk. If neither holds, we will have the first clinically auditable evidence that area-level social features carry individual risk signal beyond place.
What Prior Work Establishes and Leaves Open
The motivating results come from two papers. Evaluating the Impact of Social Determinants on Health Prediction and its ICU companion ( Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit) built MIMIC-IV-SDOH. They linked Boston ICU stays by ZIP and county to County Health Rankings, the CDC Social Vulnerability Index and the AHRQ SDOH Database. With logistic regression, random forests and XGBoost they found SDoH to be "accuracy-neutral" but "fairness-active": expected calibration error (ECE) for in-hospital mortality fell from 0.11 to 0.03, and false-positive rates were elevated in low-income, high-transit areas. They did not ask whether the calibration gain needs the values of the SDoH features or only the partition of patients that those values induce.
Investigating Data Interventions for Subgroup Fairness treated the 12 eICU hospitals as data sources. It showed that the gap between predicted and observed outcome rates per subgroup, rather than data quantity, drives cross-site disparity, and that post-hoc isotonic recalibration helps. That paper used site and race indices only, with no social content. Our proposal is the bridge between the two papers: we test whether "SDoH helps calibration" and "fixing group mean discrepancy helps calibration" are the same phenomenon.
The epidemiological literature sets a sign expectation for any SDoH curve. In Medicare dementia patients, neighborhood deprivation raised mortality odds by adjusted odds ratios of 1.05 to 1.09 across quartiles, but only in urban areas; the association was not significant in rural ones (Neighborhood deprivation and mortality among people with Alzheimer's disease and related dementia). We use this directional prior, including its urban limit, as an external check on whether a learned curve has the right sign. We do not treat it as ground truth for ICU mortality.
The method comes from two sources. Correcting misinterpretations of additive models showed that standard additive models, SHAP and Integrated Gradients all assign misleading shape to suppressor variables. On MIMIC-IV mortality, a standard additive model gave diastolic blood pressure an inverted risk curve. The paper introduced PatternGAM, which uses univariate generative surrogates to restore the property that a feature's displayed effect matches its actual statistical association. PatternGAM has only been run on vital signs, never on the highly collinear block of area indices, so we reuse its released code (PatternGAM repository). Beyond Black-Box Predictions introduced NAMformer, an additive path plus a Transformer path for interactions, and matched XGBoost on generic tabular benchmarks. We use it as the accuracy-preserving interpretable reference, since it has never been tested on clinical or spatial data.
Research Program
The central object is an additive mortality model with an explicit area term:
$$\text{logit}\, p_i = \beta_0 + \sum_j f_j(x_{ij}) + u_{a(i)} + \sum_k g_k(s_{a(i)k})$$
In this equation, $p_i$ is patient $i$'s predicted risk and $x_{ij}$ are the patient's clinical features with shape functions $f_j$. The term $u_{a(i)}$ is a shrunken random intercept for the patient's area $a(i)$, at either ZIP or county level, which absorbs place identity without using any social content. The terms $s_{a(i)k}$ are area-level SDoH indices with shape functions $g_k$. Because SDoH are constant within an area, the $g_k$ terms can only explain between-area variation that the intercept has not already absorbed. Fitting with and without $u$ therefore splits the SDoH effect into a part that a place label can explain and a part that only social content can explain. We keep the intercept shrinkage strong so that sparse ZIP codes default to zero. This avoids the obvious overfitting of a free per-area parameter.
We fit the $g_k$ in two ways. The first is a standard neural additive model. The second applies the PatternGAM correction, so that inverted or inflated SDoH curves caused by suppression can be counted. We then add two regional covariates that the original MIMIC-IV-SDOH analysis did not separate from social content: distance from home ZIP to the hospital, and transfer-in status. Both are proxies for referral geography, a form of the unobserved regional confounding the research file flags. The test is whether the area intercept and the SDoH effects shrink once these two covariates are included.
The program has three analyses, run on in-hospital mortality, 30-day readmission and one-year mortality. The first is a content-versus-partition comparison. We compare the full model against the same model with SDoH vectors permuted across areas, which preserves the grouping and destroys the meaning, and against a model with area intercepts only. The second is a suppressor audit. For each SDoH curve we record whether its sign agrees with the feature's univariate association with the outcome and with the deprivation direction reported in the Medicare study, under the standard and the PatternGAM fits. The third is a transfer test. If the offset reading is right, its gains should not survive a change of catchment. We fit on MIMIC-IV-SDOH and evaluate under the geographic shift that Yang et al. used with All of Us, if access permits. On eICU we test whether an explicit per-hospital offset term in the additive model reproduces the mean-discrepancy gains that Tan et al. obtained through data selection and isotonic recalibration.
Baselines
The naive reference is the clinical-only model from Yang et al., an XGBoost model and a logistic regression on MIMIC-IV clinical features without SDoH. It defines the gap that everything else must close.
The strongest existing method on this exact question is Yang et al.'s XGBoost model trained on clinical plus SDoH features. They report an ECE of 0.03 for in-hospital mortality, down from 0.11 without SDoH (Evaluating the Impact of Social Determinants on Health Prediction). We reimplement it on the same MIMIC-IV-SDOH patient splits with the same hyperparameter search budget as our additive models. For the comparison to be fair we must reproduce roughly 0.03 before drawing any conclusion. If our reproduction misses badly, we report that first. For this baseline, "beating it" does not mean a lower ECE. It means matching its ECE and AUROC within confidence intervals while also providing curves that pass the sign audit. Separately, it means showing how much of its 0.11 to 0.03 gain survives SDoH permutation.
The second fairness baseline is Tan et al.'s data selection targeting mean discrepancy plus isotonic regression, applied to LightGBM. No reported result found on Lacuna for its numerical subgroup metrics, so we rerun it ourselves on eICU and on MIMIC-IV-SDOH with the same model class and tuning budget.
The interpretable accuracy reference is NAMformer from Beyond Black-Box Predictions. It is the strongest model that still gives per-feature curves, and it tests whether our additive restriction costs discrimination. No reported clinical result found on Lacuna for NAMformer.
Three ablations isolate the key ingredient: the area-intercept-only model, the permuted-SDoH model, and our model without the PatternGAM correction. If the permuted-SDoH model recovers most of the ECE gain, the content hypothesis loses.
Why not just…?
Why not just run XGBoost with SHAP, the standard practice? Our decomposition requires asking whether a feature's contribution is a place offset or a shaped risk effect. TreeSHAP spreads the area effect across every correlated SDoH column. The PatternGAM paper shows that SHAP also assigns importance to suppressors that have no association with the outcome (Correcting misinterpretations of additive models). We still run TreeSHAP on the Yang baseline and apply the same sign audit to its dependence plots. If SHAP passes the audit as well as PatternGAM does, the case for the glass-box model rests only on the offset decomposition, and we will say so.
Why not just recalibrate per area stratum and skip SDoH features altogether? This is a real threat, and in a sense it is our hypothesis. If isotonic recalibration within income or area strata matches the ECE and false-positive-rate gains of SDoH features, as Tan et al.'s results suggest it might, then SDoH in acute-care models amounts to a recalibration index. We include per-stratum isotonic recalibration as an arm. A tie would be reported as the main finding, not explained away. The additive model is still worth fitting in that case, because it shows clinicians which areas are being corrected and whether distance and transfer status explain the correction.
Why not just fit a mixed-effects logistic regression with area random intercepts, as epidemiologists do? That model is our area-intercept ablation. It is the right tool for the partition question, and we expect it to do well. What it cannot do is display nonlinear SDoH shapes or catch suppressor inversions among a dozen collinear indices. RAITR's evidence that a linear baseline with penalized nonlinear additions avoids spurious curves (A reluctant additive model framework) leads us to start the $g_k$ as linear terms and admit nonlinearity only when it improves held-out concordance.
Why not just use individual-level SDoH extracted from clinical notes? Note-derived SDoH avoid the ecological fallacy, but current extractors are fragile. Bio-ClinicalBERT failed 44.5% of negation tests and 70.9% of education attribution tests on MIMIC-SBDH (MIMIC-SBDH). Individual features are also not what the calibration result was measured on. We add note-extracted Community and Economics labels as a secondary arm, restricted to the MIMIC-III overlap, and only to check whether they shrink the area intercept.
Proposed Experiments
The first decisive check is semi-synthetic. On real MIMIC-IV-SDOH covariates we simulate outcomes with three regimes: a known area offset and zero SDoH effect, a known monotone SDoH effect, and both together. The pipeline must recover each regime, meaning it attributes the effect to the intercept or to $g_k$ correctly, and PatternGAM must flatten curves for SDoH columns that have no effect. The audit is only credible if it passes this test. We score curve recovery with the earth-mover-style attribution accuracy used in the PatternGAM paper.
The second check is the main real-data comparison on MIMIC-IV-SDOH. We report AUROC, ECE, worst-subgroup ECE, and the false-positive-rate gap for the subgroups Yang et al. highlighted (Black diabetic and female diabetic patients, and low-income or high-transit areas). Every quantity gets bootstrap intervals, and we compare the full, permuted-SDoH, intercept-only, stratified-isotonic and clinical-only arms. The outcome that settles the question is the fraction of the ECE gain from clinical-only to clinical-plus-SDoH that the permuted arm retains. A retained fraction above one half supports the offset reading. A fraction near zero supports social content. We also report how that fraction changes once distance and transfer status are added.
The third check is the sign audit. For each SDoH curve under the standard additive model, PatternGAM and TreeSHAP, we report the rate of disagreement with the univariate association and with the deprivation direction from the Medicare study. Because the Medicare effect was urban-only, we also stratify the audit by urban and rural ZIP codes.
The final check is transfer. We apply the MIMIC-trained models under geographic shift (All of Us if access permits) and fit eICU models with explicit hospital offsets. We test whether the arms whose gains came from place identity lose calibration while any genuinely content-driven SDoH curves keep their sign and size. Calibration that collapses while correctly signed curves survive would be the clearest evidence for distinguishing social signal from regional offset.
References
- Evaluating the Impact of Social Determinants on Health Prediction
- Investigating Data Interventions for Subgroup Fairness: An ICU Case Study
- Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit
- Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit
- Neighborhood deprivation and mortality among people with Alzheimer's disease and related dementia
- Correcting misinterpretations of additive models
- Beyond Black-Box Predictions: Identifying Marginal Feature Effects in Tabular Transformer Networks
- A reluctant additive model framework for interpretable nonlinear individualized treatment rules
- MIMIC-SBDH: A Dataset for Social and Behavioral Determinants of Health