When a hospital system emulates a clinical trial from its own electronic health records, it only sees the part of each patient's life that happened inside its walls. A patient whose diabetes was diagnosed at a clinic across town looks newly diagnosed. A prescription filled elsewhere looks like no treatment. A stroke treated at another hospital looks like a patient who simply stopped visiting. Target trial emulation, the practice of writing down the protocol of the randomized trial one wishes one had and then mapping observational records onto it, has become the standard way to get causal effects from such data. Each of these gaps corrupts a different element of that protocol: who is eligible, who is treated, and whether the outcome is seen. The gaps are also not random. Insurance churn and access to care follow race, income, and age, so the damage is likely larger for some patient groups than for others.
We propose to measure this damage and to correct it one protocol stage at a time. The test bed is a regional health information exchange, where a patient's records from many health systems are already linked. That linkage lets us build a fragmented view, the records from a single system only, and a defragmented view, the linked records, for the same patients and the same emulated trials. The core claim is twofold. First, the bias that fragmentation induces in single-system emulations is spread across eligibility, exposure, and follow-up, and it differs systematically across demographic subgroups. Second, a correction that targets each stage, using a learned model of whether a patient's record is complete at a given time, removes more of that bias, and removes it more evenly across subgroups, than any single-stage fix. Here "multi-stage bias mitigation" is borrowed from the fairness literature as a lens: bias enters at several points in a pipeline, and fixing only one of them leaves the rest in place. "Fair" means equal estimation error across groups, not equal predictions.
What exists and what is missing
The closest work is FL-TTE. It runs a federated propensity model, inverse probability of treatment weighting (reweighting patients so treated and untreated groups look alike on measured confounders), and a federated Cox model across 192 eICU and MIMIC-IV hospitals. Its corticosteroid hazard ratio of 1.08 sits close to the pooled 1.10, compared with 1.16 for fixed-effects and 1.01 for random-effects meta-analysis. FL-TTE assumes horizontal partitioning: each patient lives entirely at one site. Our problem is the opposite case, where one patient's record is split across sites. Federation cannot reassemble that record, and FL-TTE measures no subgroup error. Two other papers handle related fragmentation by transport. RPM-CI/RPM-AB builds valid confidence intervals when only summary covariate means of a target population are available. The control-variate estimator combines a small unbiased sample with a large outcome-selected one and reports about 77% variance reduction. We reuse RPM-AB as a method inside one of our stages. The control-variate approach serves as a constraint and a baseline: its assumption that selection depends only on the outcome is exactly what migration violates, since leaving the network depends on insurance and covariates.
Two papers motivate the problem. The immune-related adverse event emulation shows how much a single protocol error matters: aligning time zero and using cloning with inverse-probability-of-censoring weights moved a progression-free survival hazard ratio from 0.37 to 1.02. The nonstationary Gaussian process study shows that missing lab tests outnumber redundant ones by an order of magnitude. That deficit means baseline covariates from a partial record are systematically thinner, not just noisier. The INPC diabetes-complication graph model shows that the exchange supports longitudinal prediction across systems for more than 800,000 patients, though it does no causal estimation.
The fairness review describes pre-, in-, and post-processing mitigation and notes that longitudinal settings are largely unstudied. The reconstruction-fairness study shows that fixing an upstream stage does not reliably repair downstream disparity. We take that structural lesson and apply it to the stages of a trial protocol rather than to a classifier pipeline.
Core Claim
Write $\hat\tau_g^{F}$ for the effect estimate in subgroup $g$ from the fragmented single-system view and $\hat\tau_g^{L}$ for the estimate from the linked exchange view. Both estimates use the same protocol, the same patients, and the same analysis code. The fragmentation bias in subgroup $g$ is
$$ B_g = \hat\tau_g^{F} - \hat\tau_g^{L}. $$
We hypothesize three things. The bias $B_g$ is materially nonzero for long-horizon chronic-disease emulations. Its spread across subgroups, $\max_g |B_g| - \min_g |B_g|$, is larger for groups with higher insurance churn. Most of it is removed only when eligibility, exposure, and censoring are all corrected.
The linked estimate is still observational, so it is not the true causal effect. It is the right reference for the fragmentation component specifically, because everything except record completeness is held fixed.
Research Program
The shared ingredient is a continuity model. It predicts, for each patient and each month, the probability that the single system's record is complete: no diagnoses, fills, or encounters are happening elsewhere. The model is trained on the exchange, where the answer is known. It uses only features visible inside a single system, such as visit regularity, payer, distance, and lab-ordering density. That restriction makes it deployable at hospitals that have no exchange access. We check its calibration within each subgroup, because a continuity model that is well calibrated on average but poorly calibrated for one group would itself reintroduce uneven bias.
The continuity model drives three stage corrections. At eligibility and baseline, we restrict the cohort to patients with high predicted look-back completeness, so that "new user" and "no prior event" mean what the protocol says. This restriction shifts the population toward stably insured patients. We therefore transport the estimate back to the full eligible population using RPM-AB, with summary covariate means as the target; the method was built for exactly the case where only target summaries are available. At treatment assignment, we treat exposure as misclassified when continuity is low. The propensity model, a regularized logistic regression following the high-throughput Alzheimer's emulation finding that LSTMs did not beat it on covariate balance, is fit with continuity as an input and with exposure ascertained only in well-observed windows. During follow-up, we define a predicted drop in continuity as an informative censoring event and model it with inverse-probability-of-censoring weights, alongside the usual cloning-and-censoring machinery used for time-varying strategies.
Each correction can be switched on alone, which gives the ablation structure.
Baselines
The naive reference point is a single-system emulation using the FL-TTE analysis stack. Concretely, this means a logistic propensity model, weighting on that propensity score, and a Cox model, run on the single-system view with the protocol's ordinary administrative censoring. This is what a careful hospital team would do today. It shares our protocol, cohort, covariate dictionary, and tuning budget for the propensity model.
The strongest existing method for this question is that same stack plus censoring weights at the last observed visit, the device used in the urothelial emulation. It already treats disappearing patients as informatively censored, which is the most obvious fix for migration. On the eICU and MIMIC-IV corticosteroid emulation, the comparable federated pipeline reports a hazard ratio of 1.08 against a pooled 1.10 (FL-TTE). For fragmentation bias on INPC, we found no reported result on Lacuna for target trial emulation, so our baseline numbers will be the first.
We also run the control-variate estimator. In that setup, a small linked subsample serves as the unbiased source and the single-system data as the selected source. This tests whether its outcome-only selection assumption is tolerable in practice.
The ablations apply each stage correction alone and in every pair, against the full stack. A further ablation replaces the continuity model with a crude rule: the patient had at least one visit per year.
Beating the strongest baseline means three things. Overall $|B|$ must be lower. The cross-subgroup spread of $|B_g|$ must be lower. Confidence intervals must cover $\hat\tau^{L}$ at close to nominal rates. All three must hold across the majority of emulated trials, not on average with a few large wins.
Why not just…?
Why not just restrict to patients with continuous records? This is the real threat. It may remove most of the overall bias by itself. Its cost is that it changes whom the trial is about and drops the very patients whose care is most fragmented. The eligibility-only ablation, with and without RPM-AB transport, measures this directly. If restriction plus transport matches the full stack on both overall and subgroup bias, we will report that the simpler recipe suffices and that the other stages add nothing.
Why not just federate across the systems, as FL-TTE does? FL-TTE assumes each patient sits at one site. It combines site-level gradients and never joins one patient's records across sites, so a stroke recorded at system B still goes unseen by system A's cohort. Where linkage exists, linking is the reference condition and not a competitor. Where linkage is absent, federation does not address the split record.
Why not just add censoring weights when patients stop visiting? This baseline fixes only the follow-up stage. It leaves baseline confounders and prior exposures misread at eligibility, which the lab undersampling evidence in the Gaussian process study suggests is substantial. The follow-up-only ablation quantifies how much bias remains.
Why not just impose fairness constraints on the propensity or outcome model? Forcing the propensity model to be independent of race, or equalizing predicted risks across groups, would bias the causal estimate whenever race-linked factors are genuine confounders or effect modifiers. The fairness tradeoffs described in the review and the Pareto-front work concern prediction parity, which is the wrong target here. We will run a group-adversarial propensity model once, to show this degradation empirically, and otherwise define equity as equal estimation error across groups.
Proposed Experiments
The primary benchmark is INPC. We emulate a battery of new-user trials comparing diabetes drug classes and common repurposing candidates on the six complications studied in the graph model. Each trial is run once on every single-system view and once on the linked view, and $B_g$ is reported by race, sex, age band, and payer. The battery is drawn in the high-throughput style of the Alzheimer's emulation so that conclusions rest on many trials rather than a chosen few. Balance is reported as standardized mean differences.
A falsification check uses the eICU and MIMIC-IV corticosteroid emulation, where 28-day in-hospital follow-up leaves little room for migration. There, the stage corrections should move the estimate only negligibly from the published 1.10 and 1.08. A large shift would mean our corrections inject bias rather than remove it.
The decisive result has three parts. The full stack must beat censoring-weights-only on subgroup bias spread. No single-stage ablation may match it. The continuity model trained in one system must transfer to held-out systems without losing subgroup calibration. If restriction plus transport alone achieves all three, the paper's conclusion becomes that simpler recipe, which would still be a useful and new finding for emulation practice.
References
- Federated target trial emulation using distributed observational data for treatment effect estimation
- Confidence Interval Construction for Causally Generalized Estimates With Target Sample Summary Information
- Multi-Source Causal Inference Using Control Variates under Outcome Selection Bias
- Target trial emulation to evaluate the effect of immune-related adverse events on outcomes in metastatic urothelial cancer
- Nonstationary Gaussian Process Regression for Evaluating Repeated Clinical Laboratory Tests
- Predicting onset of complications from diabetes: a graph based approach
- AI-driven healthcare: A review on ensuring fairness and mitigating bias
- Evaluating the Impact of Medical Image Reconstruction on Downstream AI Fairness and Performance
- High-throughput target trial emulation for Alzheimer’s disease drug repurposing with real-world data
- Efficient Fairness-Performance Pareto Front Computation