Safe Dose Finding That Does Not Trust Borrowed Strength for Rare Patient Groups

Adaptive dose-finding trials try to find, for each kind of patient, a dose inside the therapeutic window, meaning high enough to work and low enough not to poison. Modern designs model how toxicity and efficacy change with dose and share information across patient subgroups, so that a subgroup with few enrolled patients can borrow from the others. That sharing is what makes rare groups tractable. It is also what endangers them. When a rare combination of attributes truly responds differently, for example patients who are both Black and Latino on warfarin, a shared model shrinks that group toward everyone else. It then reports a narrow, confident toxicity estimate for the patients it knows least about.

This note proposes a safe dose-allocation method for trials that stratify patients by several attributes at once. Its central feature is that the uncertainty the model cannot resolve for a rare intersection is kept explicit instead of being averaged away. Each intersectional cell carries a separate "interaction allowance": the amount by which that cell could still differ from what its marginal attributes predict. This allowance widens the cell's toxicity bound until the cell's own data shrink it. Rare cells are therefore neither excluded from the trial nor dosed as if they were average. They are dosed conservatively inside the window, and the enrollment budget is spent partly on resolving whichever cell's allowance is currently blocking a safe recommendation.

Core Claim

Existing safe dose-finding designs either treat subgroups as a few isolated outputs or pool them through learned similarity. SAFE-T models toxicity and efficacy jointly with a multi-output Gaussian process, which is a probabilistic curve-fitter that returns a mean and an uncertainty band, with one output per predefined subgroup. Its authors state that it does not yet handle intersectionality without over-splitting the data. The natural extension feeds attribute vectors into a single Gaussian process, or places hierarchical priors over cells in the spirit of hierarchical Beta calibration. That extension would fix the fragmentation.

Our claim is that it introduces a different failure. The model's hyperparameters, including how much each cell may deviate from the additive pattern, are usually fit by marginal likelihood (choosing the settings under which the observed data are most probable). When cells are sparse, the data barely constrain those deviations, so the fitted interaction variance collapses toward zero. The resulting safe sets are overconfident exactly in rare cells that carry real interactions. We will show this collapse empirically. We then fix it with a floor on interaction variance calibrated from external data, and a budget rule that spends patients on resolving cell-level allowances instead of abandoning cells.

Concretely, write the toxicity probability for dose $d$ and attribute vector $a$ as an additive decomposition over attribute subsets,

$$ g(d,a) = \sum_{S \subseteq \{1,\dots,q\},\ |S|\le k} g_S(d, a_S) + r_{c(a)}(d), $$

where $g_S$ is the effect of the attribute subset $S$ (main effects when $|S|=1$, low-order interactions up to order $k$), and $r_{c}$ is a residual specific to the full intersectional cell $c$. The residual has a Gaussian-process prior whose variance $v_c$ is not fit by marginal likelihood. It is set to a floor $\underline{v}$ elicited from a held-out transfer residual, for example how badly a model fit on one pharmacogenetic cohort misses on another. The decomposition follows the logic of Molina and Loiseau. They show that intersectional unfairness equals the sum of marginal unfairnesses when attributes are conditionally independent, and that the gap is bounded by total correlation otherwise. We use the marginal terms to share strength and the residual to represent the unbounded remainder honestly.

A dose is admitted for cell $c$ only if

$$ \mu_t(d,c) + \beta_t\,\sqrt{\sigma^2_{t}(d,c) + s^2_{c,t}(d)} \le \tau, $$

Here $\mu_t$ and $\sigma_t$ are the posterior mean and standard deviation of the pooled additive part after $t$ patients, $s_{c,t}$ is the posterior standard deviation of the cell residual, $\beta_t$ is the confidence multiplier from safe Bayesian optimization, and $\tau$ is the toxicity limit. The residual's standard deviation $s_{c,t}$ shrinks only with the cell's own observations. It behaves roughly like $\sqrt{\underline{v}/(1+n_c\underline{v}/\sigma^2_\epsilon)}$, the posterior of a random effect with $n_c$ observations and noise variance $\sigma_\epsilon^2$. With no cell data the safe set falls back to the low end of what the marginals allow. It does not fall to an empty set.

We do not claim to verify optimal dosing in every cell. Paes et al. prove that auditing whether personalization helps every group becomes no better than guessing once the number of groups exceeds the per-group sample size, and that about 19 binary attributes would need more samples than there are people. That result constrains the design: at trial scale, guarantees for every cell are impossible. What is achievable, and what we target, is that no cell is exposed to toxicity beyond $\tau$ because of borrowed confidence, and no cell is left without a recommendation.

Research Program

The work has three parts.

The first is a diagnostic study of the collapse. On synthetic trials we vary the number of binary attributes $q$ from 2 to 8, giving 4 to 256 cells, and skew enrollment so that the rarest cells receive single-digit patients. We plant a toxicity interaction in a few rare cells. We then measure how the marginal-likelihood estimate of interaction variance, and the per-cell coverage of the toxicity upper bound, degrade as cells become sparser. This is the empirical premise. If pooled models with fitted hyperparameters keep nominal coverage in planted-interaction cells, the proposal's motivation fails and we would say so.

The second part is the method. It uses an additive kernel with main effects and pairwise interactions, which is the structured-kernel direction flagged as unexplored for continuous dose finding. It adds a floored cell-residual kernel and inherits the two-stage safe exploration of SAFE-T: first expand the safe set, then maximize Thall utility, a clinical score trading efficacy against toxicity, within it.

The allocation rule is adapted from the budgeted linear program of C3T, which decides whether to treat or skip each arriving patient given the remaining budget. We change its objective. C3T-Budget-E drops poorly performing subgroups so that at least one majority group reaches a significant result. We instead add a term rewarding reduction in the largest $s_{c,t}$ among cells whose safe set currently excludes a dose the pooled model believes is good. That is the point where the allowance is actually costing patients benefit.

The floor $\underline{v}$ is calibrated once, before any trial, from observational pharmacogenetic data, as described under experiments. We also report sensitivity to it, because the whole method rests on this one elicited number.

The third part is evaluation in two regimes. One is the synthetic suites of SAFE-T and C3T extended to many attributes. The other is a warfarin simulator built from real cohorts, where the rare intersection is a documented one: Villanueva et al. report that dosing models persistently underperform for Black or African American patients within the US Latino cohort.

Baselines

The naive reference points are two. One is the linear IWPC dosing formula applied to every patient with no adaptation. This is the right floor because Villanueva et al. found that nonlinear learners (support vector regression, adaptive regression splines, Bayesian additive regression trees) using IWPC variables did not significantly beat it in any cohort. The other is a 3+3 escalation design on the synthetic trials. For IWPC on warfarin, the file reports 46.7% of patients within 20% of stable dose on the merged cohort, rising to 47.8% for an expanded linear model, and 47.9% for IWPC alone on the Latino cohort, with mean absolute error of 8.11 to 8.45 mg/week across models (Villanueva et al.).

The strongest existing method is SAFE-T. We run it in two configurations. The first uses the coarse subgroups it was designed for, such as sex alone. The second makes every intersectional cell a separate output, which exposes the over-splitting its authors warn about. SAFE-T reported the lowest final recommended dose error and the fewest safety violations in 16 of 18 of its own scenarios; the file gives no per-scenario numeric dose error. We therefore rerun it ourselves on its scenarios and on our extensions, using the same safety threshold, the same number of patients, and the same hyperparameter-tuning budget as our method.

The strongest budgeted competitor is C3T-Budget. It reported a dose recommendation error rate of 0.047 against 0.247 for contextual upper-confidence-bound bandits on its three-subgroup suite, and a safe-dose estimation error of 0.0212. We first reproduce it on that suite as a sanity check. We then run it on the many-cell extensions with an identical budget.

Two further baselines isolate the key ingredient. The first is a pooled contextual Gaussian process: the identical additive kernel plus cell residual, but with all variances, including the residual's, fit by marginal likelihood. This is the version a careful practitioner would build, and it differs from ours only in the floor. The second ablation keeps the floor but uses C3T's original allocation rule, to separate the contribution of the safe-set bound from that of the budget rule.

Beating the baselines means the following. Worst-cell toxicity violation rate and worst-cell recommended-dose error should be lower than the pooled Gaussian process with a gap exceeding seed-to-seed variation. Average utility and average dose error should be within a pre-stated tolerance of the pooled model. No cell should end the trial without a recommendation, unlike C3T-Budget-E.

Why not just…?

Why not just put the attributes into one Gaussian process as inputs? This is the most serious alternative, and it is exactly our main ablation. It solves fragmentation and would likely match us on average metrics. Our claim is that its fitted interaction variance collapses in sparse cells. That produces confident but wrong safety bounds where interactions exist. The diagnostic study and the per-cell coverage comparison settle this directly. If fitted pooling keeps coverage in planted-interaction cells, the floor is unnecessary and we would report that.

Why not just use an additive linear model with demographic covariates, as the IWPC formula does? The Villanueva results favor this on average. The same paper shows it still fails the Black Latino intersection, and an additive model by construction cannot represent a cell-specific deviation. Paes et al. further show that population-average training can make small groups worse off than a non-personalized model. We include it as the floor and report its per-cell error, not only the pooled 46.7% figure.

Why not just stratify, or drop cells below a size threshold as DRAF does? Stratified estimation hits the minimax limits of Paes et al. and the resolution limits described by Ferrara et al.. Thresholding, the "active subgroup" rule of DRAF, avoids noise by leaving the smallest cells unprotected. In a trial, that is the same abandonment C3T-Budget-E performs. The per-cell SAFE-T configuration tests the stratification route. A thresholded variant that merges sub-threshold cells into their parent marginal is added as an extra arm. That variant is a real competitor if planted interactions are small, and the sweep over interaction size shows where it breaks.

Why not just apply hierarchical partial pooling to cell-level parameters? Hierarchical pooling, as in Ji et al., is the right idea for estimating a metric: it cut error by 3 to 5 times with ten labeled examples per group. Our residual is itself partial pooling. The difference is only whether the pooling variance is learned or floored. Learning it is precisely the collapse we test. In a dose-escalation setting, an underestimated variance means administered toxicity, not just a mis-reported number.

Proposed Experiments

The synthetic study extends the SAFE-T 18-scenario suite, which varies efficacy shape (plateau or parabolic), toxicity, and enrollment skew. We replace its single binary attribute with $q\in\{2,4,6,8\}$ attributes and a long-tailed enrollment distribution. We plant multiplicative toxicity shifts of three sizes in 1 to 5 percent of cells, chosen among the rarest. The C3T suite is extended the same way while keeping its budget of 400 treated patients over 1,200 arrivals.

We report worst-cell and cell-size-stratified toxicity violation rate, recommended-dose error, and Thall utility. We also report the empirical coverage of each method's toxicity upper bound within planted-interaction cells, and the number of cells left without a recommendation. The decisive check is a plot of coverage in planted cells against cell sample size. We expect the pooled Gaussian process to fall well below nominal at small $n_c$ while our method stays near nominal, at a modest average-utility cost that we quantify rather than assume.

The warfarin study builds a sequential simulator from IWPC data, following the use of that cohort in KAL and the Stiefel-manifold dose model. Each patient's recorded stable dose defines the target. A proposed dose counts as efficacious if within 20% of it and as toxic, a proxy for over-anticoagulation, if more than 20% above it. The proxy is stylized, and we state it as such. Cells are formed by self-reported race, ethnicity, sex, age band, and the VKORC1 and CYP2C9 genotype categories.

The floor $\underline{v}$ is calibrated from the cell-level residual of an IWPC-fit model evaluated on the Latino cohort of Villanueva et al., if those data can be obtained. Otherwise we use a leave-one-site-out residual within IWPC. The decisive real-data check is the Black Latino cell, together with the other rarest cells. We ask whether our method reduces over-dosing there relative to the pooled Gaussian process and SAFE-T, while keeping the whole-cohort within-20% rate no worse than IWPC's reported 46.7 to 47.9% range. A sensitivity sweep over $\underline{v}$, from one tenth to ten times the calibrated value, shows how much of any gain depends on getting that single number right.