Modern causal inference routinely uses flexible machine learning to model the "nuisance" parts of a problem: how likely each patient was to be treated, and what outcome to expect under each treatment. The standard rule says these models are good enough only if their errors shrink faster than $n^{-1/4}$, where $n$ is the sample size. Given that, a correction based on the influence function (a first-order formula for how much one data point moves the estimate) removes the leading bias and gives the usual $\sqrt{n}$ confidence intervals. That rule was derived for targets that are themselves estimable at the $\sqrt{n}$ rate. Many clinically important targets are not. The coefficients of an optimal linear treatment rule, of the form "treat if $\theta^T Z > 0$", converge only at the cube-root rate $n^{-1/3}$ because the rule jumps from "treat" to "don't treat" at a boundary. Kernel-smoothed individualized thresholds converge at $(n\delta)^{1/2}$, where $\delta$ is the smoothing bandwidth.
This note argues that the $n^{-1/4}$ barrier comes from $\sqrt{n}$ targets and does not apply to these slower ones. When the target converges at rate $r_n$, the bias-corrected (doubly robust) estimator should only need the product of the two nuisance errors, measured near the decision boundary, to be smaller than $r_n$. For cube-root regime parameters that means each nuisance may be as slow as roughly $n^{-1/6}$. A slow target therefore buys tolerance to slow machine learning. We propose to prove this "rate-matched orthogonality," to locate the predicted breakdown points by simulation, and to test whether cross-fitted machine-learning nuisances can then be combined with an existing cube-root bootstrap to give valid intervals in observational health data.
Core Claim
Take the doubly robust objective used by Inference for Optimal Linear Treatment Regimes in Personalized Decision-making. It estimates the value of the rule "treat if $\theta^T Z>0$" as the sample average of an augmented inverse-probability-weighted pseudo-outcome, built from a propensity estimate $\hat e$ and outcome regressions $\hat\mu_0,\hat\mu_1$. The regime parameter $\hat\theta$ maximizes this objective. Using estimated rather than true nuisances adds a deterministic remainder $R(\theta)$ to the population objective. This remainder is a product of propensity and outcome errors integrated over the region where the rule treats.
Because the population value is smooth in $\theta$ even though the rule is not, the maximizer shifts by approximately $H^{-1}\nabla R(\theta_0)$. Here $H$ is the curvature of the true value at the optimal parameter $\theta_0$. The gradient $\nabla R$ is an integral of the error product over the decision boundary $\{\theta_0^T Z=0\}$. Cube-root asymptotics are unaffected when this shift is $o(n^{-1/3})$. This gives the conjectured sufficient condition
$$ \big\|\hat e-e\big\|_{\partial}\cdot\max_a\big\|\hat\mu_a-\mu_a\big\|_{\partial}=o_P(n^{-1/3}), $$
where $\|\cdot\|_{\partial}$ is an error norm on a shrinking neighbourhood of the decision boundary. Cross-fitting, meaning nuisances fit on other folds, should keep the random part of the nuisance error negligible: it contributes variance of order $n^{-4/3}\|\hat\eta-\eta\|^2$ inside the $n^{-1/3}$ window, which is smaller than the $n^{-2/3}$ scale of the leading fluctuation.
The same argument applied to a single nuisance, with no correction, makes a contrasting prediction. Plugging in an outcome regression alone requires that regression to have error $o(n^{-1/3})$. Weighting by the propensity alone requires the same of the propensity. The general template is that the estimator needs nuisance product error $o(r_n)$, localized wherever the estimand is identified. For the smoothed threshold score of Test of Significance for High-Dimensional Thresholds with Application to Individualized Minimal Clinically Important Difference, this conjecturally becomes $o((n\delta)^{-1/2})$ in a $\delta$-localized norm.
Relation to Prior Work
Three papers motivate the gap. Inference for Optimal Linear Treatment Regimes in Personalized Decision-making shows cube-root, non-Gaussian limits for $\hat\theta$. It obtains valid intervals by adapting the Cattaneo et al. numerical bootstrap, which perturbs the objective with a tuning constant $e_n$. Its theory, however, assumes parametric or fast nuisances, so it never states what rate its doubly robust objective actually requires. We keep its estimator and bootstrap and replace that assumption with a rate condition that machine learning can meet.
Demystifying Statistical Learning Based on Efficient Influence Functions lays out the von Mises expansion and the $n^{-1/4}$ requirement for pathwise-differentiable, $\sqrt{n}$ targets. Our claim is that the requirement scales with the target's rate rather than being universal. Locally Robust Semiparametric Estimation supplies the method we reuse: adding the first-step influence function to a moment makes it insensitive to first-order nuisance error. We apply that idea to an objective whose maximizer is not regular.
Two other lines constrain the design. The decorrelated, bias-corrected smoothed score of Test of Significance for High-Dimensional Thresholds with Application to Individualized Minimal Clinically Important Difference and the rate theory of Nonregular and minimax estimation of individualized thresholds in high dimension with binary responses handle nuisances by sparse linear decorrelation. They do not use influence functions and do not address flexible nuisance models; we test whether rate-matching extends to that $(n\delta)^{1/2}$ setting. New $\sqrt{n}$-consistent, numerically stable higher-order influence function estimators removes second-order bias for $\sqrt{n}$ targets under minimal smoothness. Our analysis says how many correction orders a slower target actually needs, and we predict first order usually suffices for cube-root regimes.
Finally, Debiasing Synthetic Data Generated by Deep Generative Models reports that a first-order correction restores $\sqrt{n}$ behaviour even when generators converge with exponents near 0.01. The research file flags this as an apparent contradiction of the $n^{-1/4}$ rule. Our framework offers a reading: that paper's targets are means and linear regressions, whose second-order remainder is exactly or nearly zero, so the rule's premise never binds. We note this as an interpretive consequence and do not pursue synthetic data here.
Research Program
The theoretical part proves the rate-matched condition for single-stage linear regimes with cross-fitted nuisances. It has two parts. First, it shows that the localized remainder shifts the argmax by $O(\|\hat e-e\|_\partial\|\hat\mu-\mu\|_\partial)$. Second, it shows that the Cattaneo-style numerical bootstrap stays consistent when nuisances are held fixed within each fold. We also derive the single-nuisance conditions above. Then we attempt the analogous statement for the smoothed threshold score, where the localization scale is $\delta$ rather than $n^{-1/3}$.
Part of the work is establishing whether the boundary-localized norm can be bounded by ordinary $L_2$ rates under a margin-type condition. Such a condition limits how much probability mass sits near the boundary. If it fails, the condition is a sup-norm condition and is less practical. That outcome would itself be a useful negative result.
The empirical part tests the theory directly. We use the "oracle plus controlled error" design: nuisances are set to the truth plus a smooth perturbation of size $n^{-a}$. Varying $a$ traces coverage as a function of nuisance rate. The theory predicts sharp breakpoints. The doubly robust regime estimator should lose coverage when $a$ falls below about $1/6$ for both nuisances together. Single-nuisance plug-ins should break near $a=1/3$. The value functional itself, a $\sqrt{n}$ target, should break near $a=1/4$. Breakpoints that are visible, shift with $n$ as predicted, and differ across these three estimators would confirm the claim. Breakpoints at $1/4$ for the regime parameters would refute it.
Baselines
The naive reference point is the doubly robust regime estimator with flexible nuisances fit on the full sample, no cross-fitting, and Efron's nonparametric bootstrap. Cheng and Yang state that this bootstrap fails for cube-root estimators, and we expect undercoverage.
The strongest existing method is Cheng and Yang's procedure as published: parametric logistic propensity and linear outcome models, plugged into the doubly robust objective, with the numerical bootstrap. On their simulation they report only that "an optimal en of approximately 0.5 resulted in good coverage and short confidence intervals". No reported coverage percentage was found on Lacuna for this method, so we will re-run it ourselves. We use their $n=20{,}000$ linear-regime simulation design and their $e_n$ search heuristic, with identical folds and the same $e_n$ grid for every method. Beating it means two things. Under nuisance misspecification, our cross-fitted machine-learning version should hold 95% coverage within Monte Carlo error where theirs does not. Under correct specification, it should be no worse in interval length.
The single-nuisance plug-ins are outcome regression only, and inverse propensity weighting only. Both use the same learners and the same tuning budget, a fixed number of cross-validated gradient-boosting and random-forest configurations per nuisance.
Two ablations isolate the key ingredients. Removing cross-fitting, while keeping everything else, isolates sample splitting. Swapping the doubly robust objective for the outcome-only objective, with nuisances held fixed, isolates the correction term.
For the threshold extension, the baseline is the published smoothed decorrelated score with sparse linear nuisances. That paper reports that it "consistently maintained a 5% Type I error rate" in its simulations. Beating it means preserving that error rate when the nuisance structure is non-linear, where linear decorrelation is misspecified.
Why not just…?
Why not just keep parametric nuisances, as Cheng and Yang do? In randomized data the known propensity makes this fine. In observational data like eICU, both nuisances can be misspecified, and the remainder is then of order one rather than vanishing. That makes $\hat\theta$ inconsistent, and no bootstrap repairs that. The misspecified arm of our simulation shows this directly. If parametric nuisances happen to be nearly correct in eICU, the two procedures will agree, which is an acceptable and informative outcome.
Why not just plug in a good outcome model, since the target is slow anyway? This is a real threat. Our own theory says single-nuisance plug-ins only need $o(n^{-1/3})$, and a strong learner in low dimension might reach that. The controlled-error experiment settles it. If outcome-only coverage holds across the range of $a$ achieved by real learners on eICU-like covariates, the correction term is unnecessary and we will say so. The double robustness argument wins only in the band $1/6<a<1/3$, and we will report how much of realistic practice falls in that band.
Why not just use m-out-of-n subsampling over the whole pipeline, refitting nuisances in each subsample? Subsampling is valid for cube-root estimators in principle. It still inherits whatever nuisance bias exists at the subsample size, so it needs the same rate condition, evaluated at the smaller size $m$, which is stricter. It also multiplies the cost of nuisance fitting. We include it as a secondary comparison at matched compute, expecting comparable coverage only when nuisances are fast.
Why not just smooth the indicator and treat the problem as regular? Replacing the jump with a logistic surrogate restores $\sqrt{n}$ asymptotics, but it changes the estimand. Nonregular and minimax estimation of individualized thresholds in high dimension with binary responses shows such surrogates underperform the smoothed-bandwidth approach. Shrinking-bandwidth smoothing keeps the estimand but yields the non-standard $(n\delta)^{1/2}$ rate, which is exactly our second test case.
Proposed Experiments
The core simulation reuses Cheng and Yang's design with $n$ from 2,000 to 20,000. We add a misspecified arm with non-linear propensity and outcome surfaces, plus the controlled-error perturbations. We report coverage, interval length, and bias of each coordinate of $\hat\theta$ against $a$ and $n$. A second simulation follows the heteroskedastic designs of the threshold test paper with non-linear nuisance structure, reporting Type I error.
On real data, we apply all methods to the eICU fluid-balance cohort ($n=9{,}697$). We check whether Cheng and Yang's sensitivity to $e_n$ shrinks under cross-fitted flexible nuisances: they found temperature significant at $e_n=0.3$ and white blood cell count at $e_n=0.5$. On ChAMP ($n=138$, $d=160$), we test whether the three reported predictors survive non-linear nuisance modeling.
The decisive check is the location of the coverage breakpoints in the controlled-error experiment. Regime parameters breaking near $a\approx 1/6$ while the value breaks near $1/4$ would confirm rate-matched orthogonality. A common breakpoint at $1/4$ would show that the $n^{-1/4}$ rule is not just a $\sqrt{n}$ artifact after all.
References
- Inference for Optimal Linear Treatment Regimes in Personalized Decision-making
- Test of Significance for High-Dimensional Thresholds with Application to Individualized Minimal Clinically Important Difference
- Demystifying Statistical Learning Based on Efficient Influence Functions
- Locally Robust Semiparametric Estimation
- Nonregular and minimax estimation of individualized thresholds in high dimension with binary responses
- New $\sqrt{n}$-consistent, numerically stable higher-order influence function estimators
- Debiasing Synthetic Data Generated by Deep Generative Models