Certifying Rare-Subgroup Evaluation with Synthetic Test Data

Teams that must report how a model performs on small intersectional groups, such as Black women in a census dataset or a rare patient cohort, increasingly fill in the missing test data with samples from a generative model and score the model on those. It works surprisingly often. On intersectional subgroups, SYNG4ME: Model Evaluation using Synthetic Test Data reports a mean absolute error of 11.90 for synthetic-test estimates of performance, against 20.29 when using the few real samples alone. What is missing is any statement of when that estimate can be trusted. The uncertainty intervals it reports come from disagreement among several generators. If all the generators miss the same structure, those intervals are narrow and wrong.

This proposal starts from a simple observation. The generator was trained on the same real data, so it contains no new information about the rare subgroup. Any gain it provides must therefore come from an implicit assumption: that the generator's mistakes vary smoothly across related groups. We make that assumption explicit and give it a size. We then compute two-sided certified intervals on subgroup performance by propagating interval bounds up a lattice of subgroups, and we test the assumption directly on groups large enough to check it. The practical outcome is a rule for when a synthetic test set is actually worth using for a rare group, and an interval whose validity rests on a stated, checkable quantity rather than on generator agreement.

Core Claim

Fix a classifier $h$, a real distribution $P$, and a generator distribution $Q$. For a subgroup $g$, defined as an intersection of attribute values, write $a_P(g)$ and $a_Q(g)$ for the accuracy of $h$ on $g$ under the real and synthetic data. The quantity that matters for evaluation is the generator's evaluation bias on that subgroup, $\Delta(g) = a_P(g) - a_Q(g)$.

Our first claim is a negative one, and we will write the proof in full. Consider the nonparametric setting, where the real distribution restricted to $g$ can vary independently of the real distribution outside $g$. In that setting, no procedure that uses the real data plus a generator fit to it can produce a valid interval for $a_P(g)$ narrower than order $1/\sqrt{n_g}$, where $n_g$ is the number of real samples in $g$. The reason is that samples outside $g$, and anything computed from them, carry no information about $P$ inside $g$. This sharpens the warning in On the Hardness of Approximating Distributions with Tractable Probabilistic Models, which proves that bounded global error gives no guarantee on conditional probabilities. Subgroup accuracy is exactly such a conditional query, so global fidelity scores cannot certify it. The same argument explains the result in 3S Testing that intervals break down below about 10 real samples.

Our second claim is constructive. Decompose the bias over the attribute lattice, in the style of an analysis-of-variance expansion:

$$\Delta(g) = \sum_{S \subseteq A(g),\, |S| \le k} \beta_S + r_k(g), \qquad |r_k(g)| \le \varepsilon_k.$$

Here $A(g)$ is the set of attribute values defining $g$, and $\beta_S$ is the bias contribution of the lower-order cell $S$ (for example "female" or "Black × female"). The residual $r_k(g)$ collects all interactions above order $k$, and $\varepsilon_k$ bounds its size.

Low-order cells are large, so their $\beta_S$ can be estimated with narrow confidence intervals. The unestimable high-order remainder is replaced by the whole interval $[-\varepsilon_k, \varepsilon_k]$. This is the same move Bounds on marginal probability distributions makes when it replaces the uncontrolled part of a graph with the full probability simplex. Propagating these boxes, by solving a small linear program over the lattice, gives a two-sided interval for $\Delta(g)$ and hence for $a_P(g)$. That interval is valid whenever the stated bound on $\varepsilon_k$ holds.

The empirical bet is that $\Delta$ is much lower-order than accuracy itself. In other words, a decent generator absorbs most of the subgroup-specific structure, so what remains is smooth across the lattice. If that bet is false, synthetic data adds nothing beyond pooling real data, and the experiments are designed to reveal this.

Research Program

The work has three parts.

Theory. We will prove the minimax lower bound. We will then derive the certified interval under the bounded-interaction assumption, including its finite-sample confidence level, by applying a union bound over the estimated low-order cells. We will also state the regime in which it beats real-only evaluation. Roughly, the certified width is about $\varepsilon_k$ plus the estimation error of the order-$k$ cells, and the real-only width is about $1/\sqrt{n_g}$. Synthetic evaluation is therefore worthwhile exactly when

$$\varepsilon_k + \sum_{|S|\le k} \frac{c}{\sqrt{n_S}} < \frac{c'}{\sqrt{n_g}},$$

where $n_S$ is the real count in lower-order cell $S$ and $c, c'$ are confidence constants. This inequality is our analytical stand-in for the "exchange rate" that On the Equivalency, Substitutability, and Flexibility of Synthetic Data could only measure empirically, as 30 synthetic frames per real frame for tracking.

The bias estimator for each cell. Estimating $\beta_S$ requires real labeled samples in $S$, and we give it the lowest-variance unbiased form available. For each real sample, we compare its correctness indicator under $h$ with the generator's predicted correctness at that point. This is the control-variate construction from Accelerating Unbiased LLM Evaluation via Synthetic Feedback, which that paper applied to a single scalar win rate. We apply it separately in each cell of the lattice.

Assumption calibration. The bound $\varepsilon_k$ cannot be verified on the rare group itself. That would contradict the lower bound. We instead estimate the distribution of $|r_k|$ on held-out intersectional cells large enough to measure it. We then report certified intervals as a curve over $\varepsilon_k$, so users see exactly how much trust they are placing in the generator.

The generator matters here. Structured Evaluation of Synthetic Tabular Data found that synthesizers which model dependencies explicitly (synthpop, probabilistic cross-categorization) beat GANs on its substructure hierarchy. We will test whether that ranking carries over to smaller $\varepsilon_k$. That paper evaluates fidelity at each level of the hierarchy but never connects fidelity to downstream error. Our lattice links the two directly.

Baselines

The naive reference point is real-only evaluation: the subgroup accuracy computed from its $n_g$ real samples, with an exact binomial (Clopper–Pearson) interval. This is valid by construction, so every other method must be judged by whether it is narrower at matched coverage.

The strongest existing method is 3S Testing from SYNG4ME. It uses a conditional GAN trained on the real test set and a deep generative ensemble of $K$ generators to form intervals. Its reported subgroup MAE is 11.90, versus 20.29 for real data, which makes it the result to match on point accuracy. We will rerun it with the same real data, the same $K$, and the same tuning budget given to our generators. Beating it means achieving at least 90% empirical coverage at a nominal 90% level on subgroups where 3S falls well short, while keeping MAE within the 3S range.

A strong second competitor is the per-subgroup debiased estimator. It takes the synthetic accuracy and corrects it with the mean residual on the $n_g$ real samples, which is the control-variate approach applied to one subgroup. It is unbiased and valid, but its width is bounded below by the minimax rate. It is the right test of whether lattice pooling, rather than debiasing alone, supplies the gain.

The key ablation replaces $\Delta$-pooling with pooling of accuracy itself over the same lattice using real data only, with no generator. If this matches our method, the generator contributes nothing. We also report the two extreme settings of the assumption: $\varepsilon_k = 0$, which trusts the lattice fully, and a large $\varepsilon_k$, which should recover real-only width. Finally, we compare interaction orders $k=1$ versus $k=2$.

Why not just…?

Why not just use generator-ensemble or bootstrap intervals? Validated Variational Inference via Practical Posterior Error Bounds shows how approximations can look self-consistent while being badly wrong: its certified error bound is 983 for the centered Eight Schools model, versus 15 for the non-centered one. Ensembles of generators with the same architecture share blind spots in the same way. We will measure 3S coverage directly as a function of $n_g$ and of injected mode dropping, deleting a subgroup's rows before training the generator. We expect coverage to collapse in exactly the cases where our $\varepsilon_k$ curve widens honestly.

Why not just debias each subgroup with its few real labels, as control variates do? This is valid, and we include it as a baseline. Its limitation is the lower bound itself: it cannot beat $1/\sqrt{n_g}$. Its reported savings of 12% to 24.8% of annotations on a scalar target show that per-target debiasing alone leaves most of the cost in place. Any gain beyond it must come from sharing information across the lattice, and our ablation isolates exactly that.

Why not just fit a hierarchical model to the real data and skip the generator? This is the real threat. Pooling accuracy across related groups is a standard, strong approach. Our method only wins if the generator's bias $\Delta$ has smaller high-order interactions than accuracy does. We test this directly by comparing the estimated $|r_k|$ for $\Delta$ against the same quantity for accuracy on large held-out cells. If they match, we will report that synthetic test data offers no certified advantage over real-data pooling, which is a useful negative result for practitioners.

Why not just use a disagreement-discrepancy bound? (Almost) Provable Error Bounds Under Distribution Shift via Disagreement Discrepancy reaches about 99% coverage, versus 18–30% for confidence-based estimators such as average thresholded confidence (ATC). However, it gives only an upper bound, it requires unlabeled samples from the target, and it bounds global error rather than subgroup error. Applied with synthetic data as source and real data as target, its bound on a subgroup scales like global discrepancy divided by subgroup frequency. That is precisely the conditional collapse our lower bound predicts. We will include it and plot its width against subgroup frequency.

Proposed Experiments

We need ground truth for subgroup accuracy, so we use the large census version of Adult (299,285 rows) from Structured Evaluation of Synthetic Tabular Data. We compute "true" subgroup accuracies on a large held-out split, then repeatedly subsample small test sets of the sizes 3S Testing uses. We repeat this on Covid-19 (Brazil), Support, Bank, and Drug, using larger pooled splits as the reference. For each setting we train CTGAN, synthpop, and probabilistic cross-categorization generators, run all methods, and record three metrics across hundreds of resamples, stratified by $n_g$ from 2 to 200: empirical coverage, interval width, and point MAE.

We run two stress tests. The first is the SEER-to-CUTRACT cross-national cancer shift, where the generator is fit on one country and evaluated on the other. The second deletes a subgroup from the generator's training data to simulate a shared blind spot.

Three checks decide the outcome. First, our intervals must reach nominal coverage at the calibrated $\varepsilon_k$ wherever 3S intervals do not. Second, at $n_g < 10$ our intervals must be strictly narrower than real-only binomial intervals and than the per-subgroup debiased intervals. Third, the estimated high-order residual of $\Delta$ must be clearly smaller than that of accuracy. That third comparison against real-data hierarchical pooling determines whether synthetic test data earns its place at all. No results have been reported for certified coverage on these subgroup benchmarks, so these measurements would be the first.