When hospitals pool statistics without pooling patient records, the same variable is often missing at wildly different rates from site to site. In the Georgia stroke registry, the "EMS notice" field is missing in under 2% of records at some hospitals and 89% at others (Federated Multiple Imputation for Variables that Are Missing Not At Random in Distributed Electronic Health Records). Existing federated imputation methods treat this spread as a nuisance. They either average it away or, in the Heckman-based method, drop the extreme sites because the model becomes numerically unstable there. This proposal treats the spread as evidence.
Suppose the relationship between a variable and its predictors is roughly the same across hospitals, while the process that causes values to go unrecorded differs. Then the way observed values drift as the missingness rate rises reveals whether the missing values are systematically different from the observed ones. A missing-not-at-random process (MNAR, where a value's absence depends on the value itself, as when the sickest patients' timestamps go unrecorded) should produce this drift. It also indicates how far the missing values should be shifted when they are filled in. Sites that record the variable almost completely become anchors rather than outliers. The idea is to build a one-round protocol on this: a test for MNAR, an estimate of its strength, and multiple imputation adjusted for it. Multiple imputation means filling each gap several times with plausible values and combining the analyses with Rubin's rules, the standard formulas for pooling the per-copy estimates and uncertainties.
Core Claim
In a federation, the between-site variation in missingness rates is an identifying signal for MNAR. The signal becomes usable once a "site" is recognized as a candidate instrument: something that shifts the chance a value is recorded without shifting the value itself, given the covariates. Every federation already has one, whereas the instruments demanded by Heckman-style selection models are, by the file's account, "extraordinarily difficult" to justify in clinical data.
Concretely, let $Y$ be an incomplete variable, $X$ its predictors, $R$ the indicator that $Y$ is observed, and $k$ the site. Two assumptions carry the method. The first is conditional transportability: $p(y \mid x, k) = p(y \mid x)$. This allows each site to have its own patient mix but rules out site-specific differences in how $Y$ depends on $X$. The second specifies how the missing values differ from the observed ones, using an exponential tilt, a one-parameter reweighting of the observed distribution toward higher or lower values:
$$p(y \mid x, R=0, k) \propto p(y \mid x, R=1, k)\, e^{\gamma y},$$
Here $\gamma$ measures how strongly missingness depends on the unseen value, and $\gamma = 0$ is missing at random (MAR). If the observed residuals are approximately Gaussian with variance $\sigma^2$, the tilt shifts the missing values' mean by $\gamma\sigma^2$. Transportability then forces the observed conditional mean at site $k$ to satisfy
$$\mu^{\text{obs}}_k(x) = \mu(x) - \big(1 - \pi_k(x)\big)\,\gamma\sigma^2,$$
where $\mu(x)$ is the shared full-data conditional mean and $\pi_k(x)$ is the probability that $Y$ is observed at site $k$. The relation follows from writing the full mean as a mixture of the observed and tilted-missing parts.
Under MAR, observed conditional means should not track missingness rates across sites. Under MNAR, they drift linearly in $1-\pi_k$, and the slope gives the imputation shift directly. A site with $\pi_k \approx 1$ pins down $\mu(x)$. Every quantity involved is a low-dimensional per-site summary: an observed-case regression, a residual variance, and a missingness-rate model. All of it can be exchanged in a single round.
The contrast with the closest work is sharp. HDMI (Federated Multiple Imputation for Variables that Are Missing Not At Random in Distributed Electronic Health Records) fits a Heckman selection model at each site and averages the parameters. Within one site the site label is constant, so it cannot serve as the instrument, and HDMI still needs a domain-supplied exclusion variable. It also discards the near-0% and near-100% sites that this method depends on. Fed-MIWAE (Fed-MIWAE) assumes MAR outright, and its authors name MNAR as future work. not-MIWAE (not-MIWAE) models the missingness process with a neural network but is centralized, and its authors stress that MNAR models are poorly identified without structural assumptions. Here, the site structure supplies that missing assumption.
Research Program
The work has four parts.
The first part is a distributed MNAR diagnostic. Each site sends one set of summaries: coefficients from regressing $Y$ on $X$ among observed cases, the residual variance, and coefficients of a logistic model for $R$ given $X$. The coordinator then tests whether the site-level deviations of observed conditional means are explained by $1-\pi_k$. This is a federated analogue of the comparison-of-conditional-means diagnostic in Diagnosing missing always at random in multivariate data. That paper checks whether conditional means agree across missingness patterns, but it requires the pooled data matrix and fully observed anchor variables. The proposed diagnostic replaces the missingness patterns with sites and needs only summaries.
The second part is estimation. The model is fit by a one-round weighted meta-regression, borrowing the communication pattern of ODAL, which reconstructs a pooled logistic regression from a single exchange of gradients but assumes complete data. The design differs from ODAL in one respect: no lead site holds individual-level data. There are two variants. One uses a shared $\gamma$. The other lets $\gamma_k$ vary smoothly with site-level descriptors such as size or academic status. The second variant partly addresses the concern (Gap 4 in the file) that different sites follow different missingness processes.
The third part is imputation. Each site imputes locally with chained equations, filling one variable at a time from a regression on the others, and adds the estimated shift $\gamma\sigma^2$ plus draws from the uncertainty in $\gamma$. Every site thus imputes under the federation's estimate instead of an assumed MAR. Downstream estimates are pooled with Rubin's rules, and the between-imputation variance absorbs the uncertainty about $\gamma$.
The fourth part is falsification. Transportability is the assumption most likely to fail, because rural hospitals may simply be slower as well as worse at recording. Among sites with low missingness, observed means approximate full-data means. The protocol therefore tests transportability on that subset, before and independently of the MNAR estimate, and reports how large a violation would be needed to erase the estimated $\gamma$. Because the exchanged summaries are small, the protocol can add Gaussian noise that satisfies differential privacy, a formal limit on what can be inferred about any single record, to these summaries at modest cost. That claim is tested, not assumed.
Baselines
The naive reference points are complete-case analysis and local imputation within each silo. The file gives reported values for both. On GCASR, complete-case analysis gives a 0.24 increase in log CT time against HDMI's 0.15. In HDMI's MNAR simulations, independent local imputation (HIMI) has 5% to 15% bias (HDMI). Both will be run on identical site splits and masks.
The strongest existing federated MNAR method is HDMI, reported at "less than 5% bias" in most simulated MNAR cases. It will be run with its recommended handling of extreme sites and the same number of imputations. Beating it means lower bias and nominal 95% confidence-interval coverage in the regimes where HDMI discards sites or lacks a valid exclusion variable. It also means matching HDMI where HDMI's assumptions hold.
The strongest MAR federated baseline is Fed-MIWAE with FedProx. It reports a normalized MSE of 0.4299 on the non-IID ADNI split under MAR, where non-IID means each site holds a differently distributed slice of patients (Fed-MIWAE). It gets the same data and hyperparameter search budget, and its total communication rounds are reported against our single round.
The centralized upper reference is not-MIWAE on pooled data, which reports a Banknote RMSE of 0.57 against MIWAE's ~1.19 under self-masking (not-MIWAE). The target is to approach it without pooling.
The ablation that isolates the key ingredient is the same pipeline with $\gamma$ fixed at 0. This is federated MAR chained equations with identical summaries. Any gain over this ablation is attributable to exploiting cross-site variation.
Why not just…?
Why not just train not-MIWAE with federated averaging, adding a site indicator to the missingness network? This is the most serious competitor, and it may work when sites are large. However, its identification comes only implicitly, through a neural network whose flat directions not-MIWAE's authors already flag (not-MIWAE). It also needs many communication rounds, and it offers no test of the assumption it relies on. It will be implemented as a baseline. If it matches our bias and coverage in the small-site and extreme-rate regimes, the contribution narrows to the diagnostic and the one-round protocol, and the note will say so.
Why not just add site indicators to a pooled Heckman selection equation? In substance this uses the same instrument. What prevents it in practice is HDMI's local-fit-then-average protocol, not the model itself. It will be run as an ablation, implemented with one-round summaries. If a Gaussian Heckman model does as well, then the value lies in recognizing site-as-instrument and in the anchor-site test, not in the tilt estimator. The tilt should matter mainly for skewed variables such as times and lab values, where the Gaussian error assumption fails.
Why not just run a standard delta-adjusted sensitivity analysis and report a range of $\gamma$? This is honest but uninformative for decisions. On GCASR, the gap between 0.24 and 0.15 is the scientific answer, and a range spanning both settles nothing. The proposal keeps the sensitivity analysis as its falsification step. What it adds is a data-driven point estimate and interval for $\gamma$, whose width we report against a conventional prior range.
Why not just impute zeros with a missingness mask, which needs no communication? Zero imputation acts as ridge regularization only for linear models under MCAR, where missingness is independent of everything (Naive imputation implicitly regularizes high-dimensional linear models). In neural networks it makes output magnitude scale with the number of observed features (Why Not to Use Zero Imputation?). It also cannot produce the imputed values needed for downstream coefficient estimation. It is included only in the prediction-accuracy experiments.
Proposed Experiments
Synthetic federations extend HDMI's 3- and 6-site designs to 5 to 40 sites. Site observation rates range from 2% to 90%, and $\gamma$ is either shared or varies by site. The generator crosses violations of transportability of known size with Gaussian and skewed outcomes. Metrics are relative bias of a downstream regression coefficient, 95% coverage under Rubin's rules, and the size and power of the diagnostic.
For the semi-synthetic experiments, the UCI datasets used by not-MIWAE (Banknote, Concrete, Red and White Wine, Yeast, Breast Cancer) are partitioned into pseudo-sites. Each pseudo-site applies self-masking of different intensity, so ground truth is known. This evaluation reports imputation RMSE against pooled not-MIWAE and federated baselines.
ADNI's four natural centers give a small-federation stress test. Four sites should leave the diagnostic underpowered, and demonstrating that failure honestly is part of the result. GCASR, if access is granted, is the motivating real case. The question there is whether the protocol recovers an estimate near HDMI's 0.15 without an analyst-chosen exclusion variable, and whether the anchor-site test supports transportability.
The decisive checks are three. The diagnostic must hold its nominal false-positive rate under MAR with case-mix differences across sites. It must detect MNAR at realistic effect sizes with 10 to 20 sites. And on the skewed-outcome synthetic settings, the method must beat both the $\gamma = 0$ ablation and pooled-Heckman-with-site-indicators on coverage. If the first check fails because case mix masquerades as MNAR, the method is not safe to deploy, and that is itself a reportable finding.
References
- Federated Multiple Imputation for Variables that Are Missing Not At Random in Distributed Electronic Health Records
- Fed-MIWAE: Federated Imputation of Incomplete Data via Deep Generative Models
- not-MIWAE: Deep Generative Modelling with Missing not at Random Data
- Diagnosing missing always at random in multivariate data
- ODAL: A one-shot distributed algorithm to perform logistic regressions on electronic health records data from multiple clinical sites
- Naive imputation implicitly regularizes high-dimensional linear models
- Why Not to Use Zero Imputation? Correcting Sparsity Bias in Training Neural Networks