Who Pays to Act on Social Risk? Split-Incentive Utility Evaluation of SDoH Features

Adding social determinants of health (SDoH) to a clinical risk model barely changes its accuracy. SDoH are the non-medical conditions that drive health, such as neighborhood income, transit access, housing, and education. When researchers linked intensive-care records to county-level social indices, the area under the ROC curve did not move. Calibration, meaning whether predicted probabilities match observed rates, improved sharply: in-hospital mortality expected calibration error fell from 0.11 to 0.03 (Evaluating the Impact of Social Determinants on Health Prediction). A separate line of work shows that AUC is the wrong yardstick when a model decides who receives a costly intervention. On half a million claims, a cost-predicting model with AUC 0.63 produced more economic value than a readmission classifier with AUC 0.74 (Improving Hospital Readmission Prediction using Individualized Utility Analysis). No one has asked the obvious combined question: do SDoH features, judged "accuracy-neutral," change who gets treated and how much money that saves?

The question has a second half, which comes from sustainability economics. There, green investments look unprofitable when the investing firm does not capture the returns. In a three-tier supply chain, coordinated investment beats uncoordinated leader-follower play on both profit and emissions (Non-Cooperative Game and Cooperative Operation of Multi-Level Supply Chain Under Background of Carbon Emission Reduction). Health care has the same split incentive. A hospital that pays for transport or post-discharge social support recovers only part of the savings; much of the rest goes to payers and post-acute facilities. We propose to measure, on public benchmarks, how much value SDoH-informed targeting creates for each payer arrangement. The comparison runs from a hospital acting alone to a coalition that pools savings. We also estimate the transfer needed to make the coalition allocation worthwhile for the hospital.

Core Claim

SDoH features add value mainly at the budget cutoff, the margin where a care manager decides between the last patient treated and the first patient left out. We expect that value to show up in expected net savings, not in AUC. Under a payer arrangement where the decision-maker captures only a fraction of downstream savings, much of that value disappears. The disappearance should fall hardest on patients from deprived areas, because their savings accrue later and to other parties. Our claim therefore has two parts that can each be falsified. First, at a fixed treatment budget, an SDoH-augmented model raises system-wide net benefit relative to a clinical-only model. Second, the gap between hospital-only and coalition allocation is larger for high-deprivation patients than for others.

From the sustainability literature we borrow the comparison between coordinated and uncoordinated decisions and the idea of a cost-sharing transfer. We do not borrow its differential-game machinery. The economics are mostly accounting, and the note keeps them that way.

Research Program

We first build the targeting models. Following the individualized utility analysis setup (Improving Hospital Readmission Prediction using Individualized Utility Analysis), we train gradient-boosted trees in three forms: a readmission-risk model, a downstream-cost model, and their product. Each is trained with and without the community-level SDoH features of MIMIC-IV-SDOH, which draw on County Health Rankings, the Social Vulnerability Index, and the AHRQ SDOH database (Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit).

MIMIC does not record charges, so we build a cost proxy from length of stay and diagnosis-group weights. We treat the choice of proxy as a sensitivity axis, not a fact.

Next we define value for each party. Let $\Delta p_i$ be the reduction in readmission probability that the intervention gives patient $i$, $C_i$ the expected cost of that patient's readmission, $c$ the per-patient intervention cost, and $s_k$ the share of averted cost that party $k$ captures. The net benefit to party $k$ of treating the set $T$ is

$$U_k(T) = \sum_{i \in T} \left( s_k \, \Delta p_i \, C_i - c_k \right),$$

where $c_k$ is party $k$'s share of the intervention cost. We study two allocation rules. In the hospital-alone rule, the hospital picks $T$ to maximize its own $U_{\text{hosp}}$, with $s_{\text{hosp}} < 1$. The value of $s_{\text{hosp}}$ is swept over a plausible range, because the file gives no empirical capture share. In the coalition rule, $T$ maximizes total benefit, with all shares summing to one. This is the analogue of the centralized cooperative mode in the supply-chain work. The cooperative dividend is the difference in total benefit between the two rules. The required transfer is the smallest payment from the other parties that makes the coalition's choice of $T$ at least as good for the hospital as acting alone.

Following that supply-chain work, a third component is the timing of returns. Savings that land in a 90-day or one-year window are invisible to a 30-day accounting horizon. That mirrors the early-drag pattern reported for green credit, where profitability falls before it recovers (The Impact of Green Finance on the Profitability of Commercial Banks). MIMIC-IV-SDOH has both 30-day readmission and one-year mortality outcomes, so we can evaluate both horizons.

The fourth component is honesty about intervention effect. The individualized utility analysis assumed a flat 10% risk reduction at $500. Our data cannot identify a different effect for each kind of social barrier. We therefore treat that difference in effect as a swept parameter: the effect is either uniform, larger for deprived patients, or smaller for deprived patients. We report the parameter region in which each conclusion holds, and claim nothing outside it.

Baselines

The naive reference point is to treat no one, which gives zero net benefit to everyone, alongside treating a random subset of the same size.

The strongest existing method is the cost-target model from individualized utility analysis, trained on clinical features only and run on the same MIMIC-IV-SDOH splits and cost proxy. It is the strongest competitor because it already corrects AUC's blindness to how costly each readmission is. On the Optum cohort it reports maximum economic utility of −$12,472 per patient at AUC 0.63, against −$12,645 at AUC 0.74 for the readmission classifier. That is $173 saved per patient, with peak utility at treating 33.6% versus 24.5% of patients ([Improving Hospital Readmission Prediction using Individualized Utility Analysis](/work/improving-hospital-readmission-prediction)). We found no reported utility result on Lacuna for any model on MIMIC-IV-SDOH, so the clinical-only number on this dataset will be our own. To validate our pipeline, we first rerun the clinical-only version on Optum, if access is available, and check that we roughly recover the $173 gap.

The calibration-oriented baseline is the clinical-plus-SDoH model of Yang et al. (Evaluating the Impact of Social Determinants on Health Prediction), judged as that paper judges it: by AUROC, ECE, and false-positive-rate parity.

The ablation that isolates our key ingredient is a clinical-only model recalibrated after training within each deprivation tier. If recalibration recovers most of the SDoH gain, then the features add nothing beyond fixing calibration.

All models share the same features apart from the SDoH block, the same splits, and the same hyperparameter search budget, with 50 configurations per model. SDoH counts as beating the strongest baseline only if the bootstrap 95% interval for the difference in per-patient total benefit at matched budget excludes zero and the result holds across the cost-proxy and efficacy sweeps.

Why not just…?

Why not just recalibrate the clinical model by area deprivation? This is the most serious threat. The known SDoH benefit is mostly a calibration benefit, and per-group recalibration is cheap. If it closes the utility gap, the honest conclusion is that deployment needs a deprivation index for recalibration, not an SDoH-augmented model. The recalibration ablation tests this directly, and we will report that outcome if it occurs.

Why not just treat everyone above a deprivation threshold? A rule that treats patients in the top Social Vulnerability Index tier is transparent and needs no model. The evidence against it is indirect. In pediatric data, 75% of the highest-utilizing children had no documented chronic condition (Beaney et al.), which suggests that single-axis rules misplace demand. Linked SDoH features also raised false-positive rates for low-income communities (Evaluating the Impact of Social Determinants on Health Prediction in the Intensive Care Unit), which suggests that deprivation alone over-selects. We include the rule at matched budget. If it matches the model, that is a useful finding in its own right.

Why not just compute total-system benefit and skip the split-incentive analysis? Total benefit is our coalition rule, so it is included. Computing it alone would hide the policy question, which is whether the party that actually makes the decision would choose that allocation. The supply-chain results show that uncoordinated optimizers settle below the coordinated optimum (Non-Cooperative Game and Cooperative Operation of Multi-Level Supply Chain Under Background of Carbon Emission Reduction). Whether that happens here is an empirical question about who captures savings. We deliberately avoid multi-agent reinforcement learning of the kind used in RICE-N (AI for Global Climate Cooperation: Modeling Global Climate Negotiations, Agreements, and Long-Term Cooperation in RICE-N). With a single static allocation, the equilibria can be computed in closed form, and a learned negotiation would add cost without adding insight.

Proposed Experiments

The primary experiment runs on MIMIC-IV-SDOH, with All of Us as the external validation cohort used in the original SDoH study. It traces benefit-versus-budget curves for each model under both the coalition and the hospital-alone rule, across a sweep of the hospital's capture share and of the effect-heterogeneity patterns. Alongside the curves we report AUROC, ECE, and false-positive-rate parity, so the accuracy-neutral result can be confirmed side by side.

Three checks decide the outcome. First, does the SDoH model beat both the clinical-only cost model and the recalibration ablation on total benefit at matched budget? Second, does the share of high-deprivation patients in the hospital-alone treatment set fall faster than the overall share as the capture share decreases, and how large is the transfer needed to restore the coalition's set? Third, is the SDoH advantage larger under a one-year horizon than under a 30-day one, which would be the healthcare version of delayed return?

A null result on the first check, with recalibration matching the SDoH features, would redirect the claim toward recalibration. A null result on the second would show that split incentives are not an equity problem in this setting. Either outcome is informative.