Adversarial Random Forests for Synthetic Epidemiological Data
A Problem Older Than the Computer Age
Long before anyone typed “generative AI” into a search bar, medical researchers faced a stubborn arithmetic problem: the most useful health data is often the least accessible. Privacy law is not an inconvenience bolted onto epidemiology late in its history — it is a structural feature of the field. The Health Insurance Portability and Accountability Act in the United States and the General Data Protection Regulation in the European Union exist precisely because health records describe people, not abstractions. The cost of that protection is measured in months: access requests can take several months from initial application to final data access, and every one of those months is a month in which a study sits idle.
Synthetic data has been proposed as the escape hatch from this arithmetic for years. Generate artificial rows that behave statistically like real ones, hand them out freely, and let researchers rehearse their analyses before the real data arrives. The pitch is elegant. The execution, historically, has been messy. Many current synthesis methods suffer from limited quality, demand heavy computation, and require expertise that most epidemiologists do not have and should not need. Worse, the standard ways of judging synthetic data often fail to reflect what researchers actually care about — whether the numbers hold up — and they rarely measure privacy risk with any rigor.
A new replication study, published in the International Journal of Epidemiology and also available as a preprint, takes this old problem and asks a sharper question than usual. [1] Not “can synthetic data look realistic?” but “can synthetic data reproduce the actual published findings of epidemiological research while keeping the people behind those findings private?” The answer, across six real studies, is a qualified and surprisingly concrete yes.
What Was Actually Done
The method under test is adversarial random forests, or ARF. It belongs to the family of generative AI, but it is built on random forests rather than neural networks — a deliberate choice for tabular data, where rows are individual observations and columns are variables. Tabular data is awkward for deep learning: variables mix continuous and categorical types, columns have no natural positional order, and distributions are irregular. Tree-based methods sidestep much of that awkwardness, and ARF in particular can be invoked with a single line of R code, requires no GPU, and handles missing values natively.
The evaluation was not a toy exercise. The authors replicated statistical analyses from six published epidemiological papers spanning blood pressure, anthropometry, myocardial infarction, accelerometry, loneliness, and type 1 diabetes. The underlying data came from three sources: the German National Cohort, Germany’s largest cohort study with over 200,000 participants; the Bremen STEMI Registry U45 Study, which focuses on patients aged 45 or younger who experienced myocardial infarction; and the Guelph Family Health Study, a Canadian cohort of more than 246 families investigating early-life risk factors for obesity and chronic disease.
For each original dataset, the team trained 100 ARF models and drew 20 samples from each, producing 2,000 synthetic datasets per original. [2] They then ran the original studies’ own statistical analyses on those synthetic datasets and compared the results to the published findings, using metrics designed to be interpretable in applied research: the mean absolute standardized difference and the mean confidence interval overlap, alongside bootstrap confidence intervals for the originals.
Six Replications, One Pattern
The results across the six studies share a shape. In the blood pressure analysis, based on 98,332 participants and seven variables with no missing values, the synthetic medians closely resembled the originals and the confidence intervals aligned well, with a mean absolute standardized difference of 0.85 and a confidence interval overlap of 0.823. [2] Every conclusion the original authors drew from their data was equally supported by the synthetic version.
The anthropometry study tested the method under harder conditions: 101,557 participants, 12 variables, and more than a third of values missing. The synthetic histograms recreated the original distributions closely enough that the Wasserstein distance stayed at or below 0.005 for every subgroup. Missingness at that scale is not a corner case in cohort research — it is the norm — and the method held.
The myocardial infarction case-control study, with 1,713 participants and 12 variables, produced a synthetic version of a typical “Table 1” and logistic regression results that tracked the originals closely. The largest relative deviation appeared in a hypertension subgroup of only six control participants — a small-cell problem that would trouble any method — yet the synthetic confidence interval still covered the original value.
The accelerometry study posed a different challenge: 262 participants and 28 variables, three of them categorical with 34 categories in total. That combination of small sample and high dimensionality is precisely where generative methods tend to fail. The synthetic results still matched the original findings well, though standard deviations and subgroup sizes showed lower precision, likely because of the low participant-to-dimensionality ratio and the presence of derived variables.
The loneliness analysis, drawing on 113,527 participants and 47 variables, produced the study’s clearest demonstration of a practical lever. Under general synthesis — all variables included, no derivations — the reproduction was shaky, especially for sex and relationship status. Under task-specific synthesis, where the dataset was restricted to the nine variables actually used in the regression after aggregating questionnaire items into sum scores, the results moved consistently closer to the originals. The same pattern appeared in the diabetes study, where a complex identification algorithm spanning five variables was reproduced far more faithfully once the data was reduced to the eight derived variables the analysis required. In the subgroup restricted to type 1 diabetes cases, the improvement was stark.
The Trade-Off Nobody Escapes
Reproducing published findings is only half of what makes synthetic data useful. The other half is privacy, and the two pull against each other. The authors confronted this directly rather than sidestepping it. They ran membership inference attacks — attempts to determine whether a specific person’s record was in the training data — and attribute inference attacks, which try to infer a sensitive variable from supposedly public ones. They also tested generalization: whether the synthetic data resembles the underlying population rather than merely memorizing the training sample.
Against five competing synthesizers, ARF consistently ranked among the highest in utility, and it matched or exceeded all alternatives on the utility-privacy trade-off in four of the six datasets, ranking second in the others. The minimum node size, a hyperparameter controlling how finely the trees partition the data, turned out to govern that trade-off directly: smaller values improved utility but raised privacy risk. The authors suggest that slightly larger values than the default of 2 — for example 5 or 10 — may provide a more balanced default in privacy-sensitive applications. The runtime comparison was not close. For the largest dataset, ARF completed training and sampling within 35 seconds, while alternative approaches needed between six and 36 minutes. [2] The generalization tests mattered because a synthesizer that merely memorizes its training data would look excellent on replication and terrible on privacy. ARF did not merely reproduce the training distribution; it generalized competitively to held-out data while often keeping privacy risk lower than its competitors. The one clear exception was the loneliness dataset, which is almost entirely categorical — a limited combination space that lets any method reproduce feature combinations too closely — and where a competitor achieved the better overall balance.
Where the Gain Actually Lives
Strip away the metrics and the practical gain is specific. An epidemiologist who needs to develop analysis code, test a regression specification, or train junior researchers can now generate a working dataset from a real cohort in under a minute, without a GPU and without filing an access request. The synthetic data will not be identical to the real data, but the study shows it will support the same descriptive and inferential conclusions across a wide range of analysis types — which is exactly the property that makes rehearsal meaningful.
The task-specific finding sharpens this further. Restricting synthesis to the variables an analysis actually uses, and performing derivations before rather than after, measurably improves fidelity. That is a workflow recommendation with real consequences: it tells researchers where to spend their effort. The authors are candid about its limit, though. In scenarios where data must be shared broadly, preprocessing has to happen at the provider’s end, and the analyst may not control which variables get synthesized.
The open question is whether this holds at the edges. The study’s own data shows degradation in small samples with many categories, in datasets dominated by categorical variables, and in subgroups with only a handful of cases. The authors point toward a differentially private variant of ARF and an extension for extremely high-dimensional data such as omics datasets as future directions. Until those exist, the honest summary is that ARF has moved synthetic data from a promising idea to a tool that reproduces published epidemiology — but the boundary of that threshold is drawn by the data itself, and the study maps only part of it.