Share Baseline Risk Widely, Share Treatment Effects Selectively: Federated Effect Estimation Across Health Contexts

Public health interventions often work in one place and fail in another. Hip protectors cut fracture risk by about 30% in nursing homes but do nothing for seniors living at home. Fact-check labels that reduce misinformation in Western trials did not work for Facebook users in Kenya and Nigeria. Health data, meanwhile, is split across hospitals, clinics and countries that cannot pool records. Federated learning addresses the second problem: each site trains a shared model on its own data and sends only model updates to a central server. But almost every federated method shares one representation across sites, or keeps one private head per site. Neither choice matches what the public health evidence actually says. Who is at risk tends to look similar across sites. Who benefits from an intervention is the part that changes with context.

This note proposes a federated method for estimating individual treatment effects that splits the model into two modules and treats them differently. The first is a prognostic module, which predicts outcomes without treatment; it is averaged across all sites. The second is an effect module, which predicts how much the intervention changes the outcome for a given person. It is shared only among sites whose effect patterns actually agree. That agreement is measured directly: each site scores the other sites' effect modules on its own data, using a loss that does not require the unobserved counterfactual. The output is a set of personalized effect estimates. It is also a map of which contexts can borrow evidence from which, a question that context-specific public health asks but has had no tool to answer.

Core Claim

The claim is that heterogeneity across health sites is asymmetric. Prognosis transfers, and effect modification often does not. A federated estimator that respects this asymmetry should beat two standard choices: sharing everything, and keeping all site-specific structure private. The gains should be largest at small sites, where local effect estimates are too noisy to use. The claim is falsifiable. If, on real multi-site randomized data, the learned between-site effect-sharing pattern is no better than uniform sharing or no sharing, the asymmetry hypothesis fails for that domain. Reporting that outcome is itself useful, because it would tell practitioners that the cheaper method is enough.

The proposal builds on three pieces of prior work.

The method's architecture comes from On Inductive Biases for Heterogeneous Treatment Effect Estimation. Its FlexTENet is a multi-task network with shared and private subspaces. It separates prognostic structure (common to treated and control outcomes) from predictive structure (the treatment contrast), and it outperformed simpler two-network and shared-trunk learners in small samples. That work is centralized, and it treats the two potential outcomes as the tasks. Here the same prognostic/predictive split is moved to the level of sites. The prognostic subspace becomes the federated part, and the predictive subspace becomes the selectively shared part.

The closest federated competitor is Federated Learning for Estimating Heterogeneous Treatment Effects. Its FedTransTEE aggregates a transformer encoder globally, keeps a local prediction head per hospital, and was validated on IHDP, Twins, ACIC-2016 and multi-center stroke and Alzheimer's trials. Its split is between "encoder" and "head", not between prognosis and effect. Each head is either fully local or fully shared, with no sharing among a subset of similar sites.

Selective borrowing is motivated by Federated Adaptive Causal Estimation (FACE) of Target Treatment Effects. FACE learns site weights by penalized regression so that dissimilar sites are shrunk to zero, and it reduced standard errors by 26% to 67% over target-only estimation in a five-site VA vaccine study. FACE estimates one average effect for one target site from one-shot summary statistics. The proposal instead learns individual-level effect functions for every site at once, inside iterative neural training.

The idea of learning a matrix of relationships between sites comes from Federated Multi-Task Learning. Its MOCHA method does this for convex predictive models. Here the relationship matrix is restricted to the effect module, and it is estimated from cross-site evaluation rather than from parameter regularization.

Research Program

Each site $k$ holds data with covariates, a binary treatment and an outcome. It fits a prognostic module $\mu^{(k)}_0$ and an effect module $\tau^{(k)}$. The prognostic module is trained with ordinary federated averaging over all sites. Following the eICU study Federated Learning in Multi-Center Critical Care Research: A Systematic Case Study using the eICU Database, many local epochs and small batches at small sites are permitted. That study found these settings helped on prognostic mortality prediction across 208 US hospitals, which is exactly the kind of quantity hypothesized to transfer.

For the effect module, each site computes doubly robust pseudo-outcomes on a held-out local split. A pseudo-outcome is a per-person quantity whose expectation equals that person's treatment effect. It is built from the observed outcome, the propensity of treatment (known by design in randomized trials) and the current outcome models. The server periodically sends each site the current effect modules of the other sites. Site $k$ then reports a single number for each other site $j$,

$$ A_{kj} = \frac{1}{n_k}\sum_{i \in \text{val}_k} \big(\phi_i - \tau^{(j)}(x_i)\big)^2 - \frac{1}{n_k}\sum_{i \in \text{val}_k} \big(\phi_i - \tau^{(k)}(x_i)\big)^2 . $$

Here $\phi_i$ is the pseudo-outcome for validation person $i$ at site $k$, and $n_k$ is the validation size. $A_{kj}$ measures how much worse site $j$'s effect function explains site $k$'s data than site $k$'s own function does. Under randomization the pseudo-outcome loss differs from the true effect error only by a constant, so differences like $A_{kj}$ are unbiased for differences in effect error. This addresses, at least for trial data, the model-selection problem that FlexTENet's authors flag as unresolved.

Sharing weights are then set as $w_{kj} \propto \exp(-A_{kj}/T)$, truncated to zero when the confidence interval for $A_{kj}$ excludes zero on the harmful side. $T$ is a temperature tuned by the same pseudo-outcome loss. Each site's effect module is then updated toward the $w$-weighted average of its neighbors' modules, which is a personalized, clustered form of averaging applied to the effect module only.

Only model parameters and scalar losses leave a site, never records. Differential privacy is not claimed. As both FedTransTEE and MOCHA note, parameter sharing can leak information. This is stated as a limitation rather than solved.

The analysis output is the matrix $w$ itself. Its block structure can be compared against known context descriptors (institution type, region, income level). This tests whether contextual filters of the kind documented in the hip-protector umbrella review Does hip protector prevent falls and hip fractures? An umbrella review of meta-analyses and in the Kenya/Nigeria misinformation trial Battling the coronavirus ‘infodemic’ among social media users in Kenya and Nigeria are recovered from data alone.

Data

The first stage is controlled semi-synthetic federation. IHDP, ACIC-2016 and Twins covariates are split into synthetic sites by clustering on socioeconomic covariates, so that sites differ in covariate distribution. Outcomes are then generated under three regimes. In the first, sites differ only in the prognostic function. In the second, they differ only in the effect function, with effects grouped into a few latent context clusters. In the third, they differ in both. Site sizes are deliberately unequal, so that a long tail of small sites is present. Because the counterfactuals are known, the root of the precision in estimating heterogeneous effects (PEHE) can be reported per site.

The second stage adds realistic site structure. Synthetic treatments and effects are placed on eICU covariates using the real 208-hospital partition. Effects are made to depend on hospital-level attributes only where the file's description of eICU supports such attributes; where it does not, they depend on latent hospital groupings. This tests the method under real hospital-size imbalance.

The third stage uses real randomized data across contexts. The 63-country climate intervention tournament Addressing climate change with behavioral science: A global intervention tournament in 63 countries randomized the same interventions across countries, so countries can serve as sites. The multi-center stroke trials (ATACH-II, MISTIE-III, ERICH) used by FedTransTEE can serve as sites as well. Access to individual-level data and site identifiers must be confirmed. The CUNY FAFSA trial Machine Learning Who to Nudge: Causal vs Predictive Targeting in a Field Experiment on Student Financial Aid Renewal is a candidate only if campus identifiers are available; that is not confirmed in the file.

On real data, counterfactuals are unknown. Evaluation therefore uses held-out policy value: the inverse-propensity-weighted outcome achieved by treating the top fraction of people ranked by estimated effect. This follows the evaluation logic of the CUNY study, where effect-based targeting raised renewal from 36.5% to 41.2%, about 68% of the attainable gain.

Baselines

The naive reference points are two. The first is purely local estimation: each site fits FlexTENet, or an effect-learning method that regresses pseudo-outcomes, on its own data. The second is "treat everyone", which gives the policy-value floor. Beating local estimation at small sites is the minimum bar. The FedTransTEE, FACE, FeDCM and eICU studies all report that local-only models are worse, but the claim has to be checked here at the small-site end specifically.

Global federated averaging of a single shared-trunk two-head network (TARNet) is the "one model fits all" competitor. It is run with the same architecture size, number of rounds and local-epoch grid. Matching or beating it in the regime with prognostic-only heterogeneity is acceptable. It must be beaten clearly when effects differ by context.

The strongest existing method is a FedTransTEE-style design: a globally averaged encoder with a personalized local head per site. It is re-implemented on tabular inputs with the same backbone capacity, an identical tuning budget and the same pseudo-outcome-based selection rule, so that only the sharing structure differs. No reported result was found on Lacuna for FedTransTEE's numerical PEHE on IHDP, Twins or ACIC-2016; the paper reports only that it beat FedCI and centralized baselines. The comparison is therefore run fresh rather than against a quoted number. Beating it means lower median PEHE across sites on semi-synthetic data, with the largest gain at small sites in the context-clustered regime, and higher held-out policy value on the real trials. Confidence intervals come from site-level bootstrap.

For site-level average effects, FACE is run on the same partitions. A fair test for this proposal is whether its implied site-average effects match FACE's standard-error reduction while additionally providing individual targeting.

A centralized FlexTENet trained on pooled data serves as the upper reference that federation is trying to approach.

Two ablations isolate the key ingredient. One shares the effect module uniformly across all sites, which tests whether selectivity matters. The other keeps the effect module fully local while federating the prognostic module, which tests whether borrowing effects at all matters. A third ablation replaces the cross-evaluation weights with gradient-similarity weights, which tests whether pseudo-outcome evaluation is the right similarity signal.

Why not just…?

Why not just run federated averaging and then fine-tune the whole model locally? Fine-tuning adapts prognosis and effect together. At a small site the effect signal is much weaker than the outcome signal, because effect estimation is underpowered relative to average outcomes, as the CUNY study notes for subgroup analysis. Fine-tuning will therefore mostly re-fit baseline risk and either leave the effect at the global value or overfit noise. This is an empirical claim, and the fine-tuning baseline is included in the eICU and trial experiments to test it. If fine-tuning matches the method in the context-clustered regime, the proposal's extra machinery is not justified.

Why not just keep a personalized head per site, as FedTransTEE does? A private head is the special case of this method in which every site shares only with itself. It cannot pool effect evidence among, say, all nursing-home sites while excluding community clinics. It therefore gains nothing at small sites, which are exactly where pooling is needed. The comparison against the FedTransTEE-style baseline is the main experiment. The prediction is that the two methods tie when effects are truly site-unique, or when there are few large sites, and that the proposal wins when effects cluster by context.

Why not just federate the outcome models and fit the treatment effect locally on top of them? This is the most serious threat. It is the fully-local-effect ablation, and it is simple, private and well grounded in two-stage causal estimation. If effect heterogeneity is mostly site-specific, or if sites are large, it should do as well as the proposal. The proposal's advantage exists only when small sites share effect structure with identifiable peers. The semi-synthetic regimes are designed so that this condition is both present and absent. If the ablation ties the full method on the real trials, the honest conclusion is that federated prognosis plus local effect estimation is the recommended practice, and the note will report it as such.

Why not just group sites by known context descriptors, such as institution type or country income, and federate within groups? When the right descriptor is known this is cheap and interpretable, and it is run as a baseline using whatever descriptors the datasets provide. But the contextual-filter literature shows that the relevant descriptor is often not the obvious one. For example, regional dietary habits and parental education predicted child diet quality better than income Sausenthaler et al.. The test is whether the learned matrix $w$ beats descriptor-based grouping, and whether it recovers that grouping when the grouping is correct.

Decisive Checks

Three results would settle the claim.

First, in the semi-synthetic regime where only prognosis differs across sites, all federated methods should tie, and global sharing of the effect module should be fine. A large win for the proposal there would signal a bug or a tuning advantage, not a real effect.

Second, in the context-clustered effect regime, the learned matrix $w$ should recover the planted clusters. The proposal should also reduce PEHE at the smallest quartile of sites relative to both the personalized-head baseline and the local-effect ablation.

Third, on the multi-country tournament and the multi-center stroke trials, the proposal should give higher held-out policy value than both, and its learned site groupings should be stable across bootstrap resamples.

If the third check fails while the first two pass, the conclusion is that the asymmetry hypothesis holds in simulation but not in these real settings. That would argue for simpler federated prognosis with local effect estimation in practice.