We propose prognostic variables assisted treatment effect detection (PATED), a covariate-adjustment method based on jointly asymptotically normal marginal estimators. The method adjusts a primary treatment-effect estimator by a covariance-weighted combination of baseline-variable contrasts whose population values are zero under randomization. We establish its asymptotic distribution and optimality within the class of linear adjustments, and define a statistic-specific \(R^2\) as the fraction of variance removed. The construction accommodates heterogeneous endpoints and prognostic variables through their influence functions, including censored event histories and pairwise win statistics.
We then address two planning questions: which candidate covariates should be selected, and how much variance reduction should their adjustment provide? For historical data without the treatment comparison of interest, common independent artificial labels define a null-reference version of these questions. Analytic integration over the labels reduces estimation to a projection of marginal influence components. An adaptive lasso selects nonzero additional contributions, with a pointwise oracle property under a joint expansion of the estimated component covariance. We develop inference for the resulting \(R^2\) and independent assessment of achieved variance reduction. Simulations with 5,000–20,000 historical observations, 20 candidate variables, three to five active components, and population \(R^2\) of 8%–15% distinguish support recovery from retained variance reduction. The historical quantities inform trial planning under explicitly stated comparability conditions.
Keywords: covariate adjustment; treatment-effect estimation; historical data; influence function; adaptive lasso; variance reduction.
Baseline characteristics often explain a substantial part of the heterogeneity in clinical outcomes. Randomization protects treatment comparisons against systematic baseline differences, but it does not eliminate the sampling variability associated with that heterogeneity. Prognostic information can therefore improve the precision of an otherwise valid treatment-effect estimator. The FDA’s May 2023 guidance recommends considering prognostic baseline covariates in randomized parallel-group trials, emphasizes prespecification of the analysis, and identifies previous studies as a basis for choosing useful covariates [1]. This role of adjustment differs from its use to address confounding in an observational treatment comparison.
The statistical justification for adjustment need not depend on a correctly specified conditional response model. A working regression can assist estimation while randomization supplies the basis for validity. Ye, Shao, Yi and Zhao [2] examine how the form of adjustment, treatment allocation, and variance estimation determine its performance. Their results distinguish validity from guaranteed efficiency gain. Reluga, Ye and Zhao [3] similarly compare regression adjustments without assuming that the postulated linear conditional-mean model is correct. These findings explain why the presence of prognostic information alone does not determine the benefit of a particular adjusted analysis.
The same distinction is relevant beyond continuous outcomes. Bannick and colleagues [4] study augmented inverse probability weighted estimators using generalized linear, nonparametric, and machine-learning working models. For right-censored outcomes, Ye, Shao and Yi [5] develop covariate adjustment of the log-rank statistic. Together, these developments motivate assessment of prognostic information in relation to the treatment estimator or statistic that will be used in the trial. An association that is useful for predicting the clinical outcome under one loss function need not yield the same precision gain for another treatment comparison.
This raises two related questions at the planning stage. Which of the available baseline variables should be selected for adjustment, and how much reduction in the variance of the intended treatment-effect estimator should that selection provide? The first question becomes especially important when historical trial records or real-world data contain many candidate variables. Claims databases, for example, may contain diagnoses, utilization measures, prior-treatment characteristics, and event histories. A parsimonious adjustment can be easier to prespecify and interpret, but excluding a useful variable may sacrifice precision. The second question is quantitative: a list of prognostic variables does not by itself describe the benefit of including them in the primary analysis.
Historical data provide an opportunity to study both questions before examining outcomes from the future trial. Such data may consist of a previous control arm, a single-arm cohort, or an observational population with relevant baseline and outcome measurements. Li, Li and Luedtke [6] develop methods for identifying and estimating the efficiency gains of future covariate-adjusted analyses from external data. Their work establishes the importance of matching the historical-data estimand to the intended trial analysis and accounting for uncertainty in the estimated gain. The setting also calls for variable selection when the number of potential adjustment variables is substantial.
Conventional regression summaries do not always supply the quantity needed for this purpose. Logistic and Cox regression have several generalized coefficients of determination and explained-variation measures [7,8]. These may summarize likelihood improvement, prediction error, or variation in a survival process. They are not generally equal to the fraction of sampling variance removed from a particular treatment-effect estimator. A variance-reduction measure used for trial planning should therefore be defined by the adjusted and unadjusted procedures being compared.
The types of available prognostic information present a further consideration. A fully observed previous time-to-progression can enter a regression as a numeric covariate. When that event time is censored, however, treating the observed minimum of event and censoring times as the true time changes what is measured. Similar difficulties arise when a clinically meaningful endpoint is a hierarchical comparison of several outcomes. An adjustment framework that permits each variable to retain an appropriate marginal statistical representation can accommodate such information without requiring every original measurement to enter one conditional response family. Its validity must nevertheless retain the observation assumptions required by those representations.
These considerations motivate a unified treatment of covariate adjustment, variable selection, and variance assessment. The adjustment should preserve the primary estimand, its gain should have a direct variance interpretation, and historical analyses should address the same statistical comparison intended for the trial. Historical-to-trial differences in eligibility, measurements, follow-up, and censoring remain relevant to interpreting the resulting planning quantities.
Consider a randomized trial with \(N\) participants, a primary endpoint \(Y\), and baseline variables \(X=(X_1,\ldots,X_p)^\top\). Let \(\hat{\theta}_Y\) estimate a scalar treatment-effect parameter \(\theta_Y\). For each baseline variable, choose a marginal contrast between treatment arms and collect its estimator in \(\hat{\delta}_X\). Because the variables precede randomization and are not affected by treatment assignment, the population vector of these auxiliary contrasts is zero. The contrasts may use different scales and need not be coefficients from a common outcome regression.
Assume that the marginal estimators satisfy
\[ \sqrt N \begin{pmatrix} \hat{\delta}_X\\ \hat{\theta}_Y-\theta_Y \end{pmatrix} \xrightarrow{d} N\left\{0, \begin{pmatrix} \Sigma_{XX}&\Sigma_{XY}\\ \Sigma_{YX}&\Sigma_{YY} \end{pmatrix}\right\}, \tag{1} \]
where \(\Sigma_{XX}\) is positive definite and \(\Sigma_{YY}>0\). Assume also that the joint asymptotic covariance \(\Sigma\) has a consistent estimator \(\hat{\Sigma}\). Joint convergence is essential; separate marginal normal limits do not establish (1). No joint normality assumption is imposed on the original measurements.
We define prognostic variables assisted treatment effect detection (PATED) by
\[ \hat{\theta}_{\mathrm{PATED}} =\hat{\theta}_Y-\hat{c}_*^\top\hat{\delta}_X, \qquad \hat{c}_*=\hat{\Sigma}_{XX}^{-1}\hat{\Sigma}_{XY}, \qquad c_*=\Sigma_{XX}^{-1}\Sigma_{XY}. \tag{2} \]
The auxiliary contrasts correct fluctuations in the primary estimator. They do not estimate treatment effects on baseline characteristics. Their zero population target allows the correction to preserve \(\theta_Y\) asymptotically. A raw baseline mean, whose population value is generally nonzero, cannot serve as such a contrast without appropriate centering.
Theorem 1 (PATED). Suppose (1) holds, \(\hat{\Sigma}\to_p\Sigma\), and
\[ V_*=\Sigma_{YY}-\Sigma_{YX}\Sigma_{XX}^{-1}\Sigma_{XY}>0. \tag{3} \]
Then \(\hat{\theta}_{\mathrm{PATED}}\to_p\theta_Y\) and
\[ \sqrt N(\hat{\theta}_{\mathrm{PATED}}-\theta_Y) \xrightarrow{d}N(0,V_*). \tag{4} \]
Among estimators of the form \(\hat{\theta}_Y-\hat{c}^\top\hat{\delta}_X\) with \(\hat{c}\to_p c\), PATED minimizes the first-order variance. More precisely,
\[ V(c)=\Sigma_{YY}-2c^\top\Sigma_{XY}+c^\top\Sigma_{XX}c =V_*+(c-c_*)^\top\Sigma_{XX}(c-c_*). \tag{5} \]
Proof. Consistency of \(\hat{\Sigma}\) gives \(\hat{c}_*-c_*=o_p(1)\). Since \(\hat{\delta}_X=O_p(N^{-1/2})\), replacing \(\hat{c}_*\) by \(c_*\) changes the estimator by \(o_p(N^{-1/2})\). The normal limit follows from (1) and Slutsky’s theorem. Completing the square gives (5), and positive definiteness establishes the unique variance-minimizing coefficient. \(\square\)
The consistent variance estimator is \(\hat{V}_*=\hat{\Sigma}_{YY}-\hat{\Sigma}_{YX}\hat{\Sigma}_{XX}^{-1}\hat{\Sigma}_{XY}\). Consequently, \((\hat{\theta}_{\mathrm{PATED}}-\theta_0)/\sqrt{\hat{V}_*/N}\) is asymptotically standard normal under \(\theta_Y=\theta_0\). If a marginal statistic is used solely to test a null hypothesis, it must first be placed on the scale of (1); the theorem does not convert every test statistic into an estimator of a clinically interpretable effect.
The optimality is within the stated linear adjustment class. It is not a claim that PATED attains the semiparametric efficiency bound among all possible estimators. Likewise, its validity under a particular allocation design requires a joint limit and covariance estimator valid for that design. Estimating the optimal coefficient consistently does not worsen first-order variance, but the result is not a finite-sample variance guarantee.
Define
\[ R_{\mathrm{PATED}}^2 =\frac{\Sigma_{YX}\Sigma_{XX}^{-1}\Sigma_{XY}}{\Sigma_{YY}}, \qquad \frac{V_*}{\Sigma_{YY}}=1-R_{\mathrm{PATED}}^2. \tag{6} \]
Positive semidefiniteness of \(\Sigma\) gives \(0\le R_{\mathrm{PATED}}^2\le1\). Under the nondegeneracy condition in Theorem 1, the upper bound is strict. A value of \(0.12\) means that optimal linear adjustment removes 12% of the primary estimator’s first-order variance. The corresponding standard-error ratio is \(\sqrt{0.88}\). For a subset of covariates, the same definition uses its covariance submatrix.
This coefficient describes explained sampling variation of a treatment-effect estimator, rather than explained variation of the raw endpoint. It is tied to the chosen primary and auxiliary statistics. Under a nonsingular linear re-expression \(\delta_X^*=B\delta_X\), the span of available adjustment terms is unchanged and so is \(R_{\mathrm{PATED}}^2\). A nonzero rescaling of the primary contrast also leaves it unchanged. Adding a new auxiliary direction cannot increase the optimal population variance, although estimating redundant or weakly informative directions can affect finite-sample performance.
Theorem 1 and (6) give the basis for the two planning questions. Selecting covariates means choosing the auxiliary contrast space. Assessing their benefit means quantifying the reduction in the variance of the intended primary statistic. A goodness-of-fit measure from a different response model generally addresses a different target.
Suppose the marginal estimators have the stronger, joint asymptotically linear representation
\[ \hat{\theta}_Y-\theta_Y =\frac1N\sum_{i=1}^N\phi_Y(Z_i)+o_p(N^{-1/2}), \qquad \hat{\delta}_X =\frac1N\sum_{i=1}^N\phi_X(Z_i)+o_p(N^{-1/2}), \tag{7} \]
where \(Z_i\) is the trial record and the stacked influence function has mean zero and finite second moments. Under independent sampling its covariance is \(\Sigma\), and a multivariate central limit theorem yields (1). PATED then has influence function
\[ \phi_{\mathrm{PATED}}(Z) =\phi_Y(Z)-c_*^\top\phi_X(Z). \tag{8} \]
Thus its variance minimization is precisely the projection \(\min_c E\{(\phi_Y-c^\top\phi_X)^2\}\). This provides a common construction for heterogeneous marginal estimators, without specifying the conditional distribution of \(Y\) given all candidate covariates.
For an estimator defined by \(N^{-1}\sum_i\psi(Z_i;\eta)=0\), write \(D=E\{\partial\psi(Z;\eta_0)/\partial\eta^\top\}\). Its influence function is \(-D^{-1}\psi\), and its covariance is \(D^{-1}E(\psi\psi^\top)D^{-\top}\). For a regular likelihood estimator this reduces to \(I^{-1}s\), with likelihood score \(s\) and information \(I\). The raw score and the complete influence function are therefore distinct objects. Cross-products of the complete influence functions provide the joint covariance needed in (2), including the dependence between marginal analyses of the same participant.
To describe the historical-data construction, consider a common-law reference in which both arms have distribution \(P\) and the arm label \(A\) is independent of the observed record \(O\), with \(P(A=1)=\pi\). The relevant null influence functions have the form
\[ \phi_j(A,O)=W(A)h_j(O), \qquad W(A)=\frac{A-\pi}{\pi(1-\pi)}, \qquad E_P h_j=0. \tag{9} \]
Here \(h_j\) is a label-free influence component. For a difference of regular marginal functionals, it is the influence function of the marginal functional itself. For a pairwise statistic it is an arm-directional Hájek projection. It need not be the derivative of a one-sample functional with the same numerical target. Both primary endpoints and baseline variables can be represented in this way, with the baseline record restricted to information available before randomization.
| Measurement and marginal contrast | Label-free component \(h\) | Regularity and interpretation |
|---|---|---|
| Continuous measurement, difference in means | \(Z-\mu\), \(\mu=E(Z)\) | Finite variance; normality of \(Z\) is unnecessary. |
| Binary measurement, difference in log odds | \((Z-\mu)/\{\mu(1-\mu)\}\) | \(0<\mu<1\); proportional to the risk-difference component at the common-law reference. |
| Count measurement, difference in log means | \((Z-\mu)/\mu\) | \(\mu>0\) and finite variance; equality of the mean and variance is not required. |
| Censored event time, treatment-only Cox coefficient | \(M(\tau)/d_\tau\), where \(d_\tau=E\{N(\tau)\}>0\) | A fixed follow-up horizon and a valid event-time observation model; the displayed expression is the common-law null representation. |
| Censored event time, normalized log-rank statistic | A nonzero scalar multiple of \(M(\tau)\) | Normalize the score sum to the asymptotically linear scale; its variance is meaningful even when no scalar treatment-effect interpretation is asserted. |
| Repeated measurements summarized by a marginal coefficient | The chosen coordinate or contrast of \(-D^{-1}\psi_i\) | One complete contribution per independent participant, with the nuisance terms required by the marginal estimator. |
For event histories, let \(\widetilde{T}=\min(T,C)\), \(\Delta=I(T\le C)\), \(N(t)=I(\widetilde{T}\le t,\Delta=1)\), and \(Y^{\mathrm{risk}}(t)=I(\widetilde{T}\ge t)\). Under independent right censoring in the represented population,
\[ M(t)=N(t)-\int_0^tY^{\mathrm{risk}}(u)\,d\Lambda(u). \tag{10} \]
At the common-law reference, the treatment-only Cox score has first-order contribution \((A-\pi)M(\tau)\) and information per participant \(\pi(1-\pi)d_\tau\), yielding the component in the table. The log-rank score uses the same martingale component up to normalization. These formulas do not impose a Cox model on the conditional relationship between the primary outcome and all baseline variables. They do retain the observation assumptions needed for the chosen marginal event-time analysis.
For a repeated-response vector, a common estimating function is \(\psi_i=\dot{\mu}_i^\top V_i^{-1}(Z_i-\mu_i)\). The corresponding influence function includes the inverse derivative matrix and any required nuisance contributions. A scalar coefficient contrast can be included in PATED once the joint asymptotic representation is established. Separate measurements from one participant must not be treated as independent contributions.
A hierarchical comparison examines prespecified endpoint components in order and records the first decisive win or loss. Components may be continuous, binary, or time-to-event. Let \(w(o,o')\) indicate a win of the first record over the second, \(\ell(o,o')=w(o',o)\), and \(k=w-\ell\). The endpoint priorities, directions, margins and rules for unresolved comparisons are part of the statistical definition.
For independent treatment and control records with distributions \(P_1\) and \(P_0\), let \(p_W=Ew(O_1,O_0)\) and \(p_L=E\ell(O_1,O_0)\). The net benefit is \(p_W-p_L\) and the log win ratio is \(\log(p_W/p_L)\). When \(p_W,p_L>0\), define
\[ b_1(o)= \frac{E_0w(o,O_0)-p_W}{p_W} -\frac{E_0\ell(o,O_0)-p_L}{p_L}, \qquad b_0(o)= \frac{E_1w(O_1,o)-p_W}{p_W} -\frac{E_1\ell(O_1,o)-p_L}{p_L}. \tag{11} \]
The log-win-ratio influence function is \(A b_1(O)/\pi+(1-A)b_0(O)/(1-\pi)\). For net benefit, replace the difference of ratios by the centered conditional expectation of \(k\). Fixed bounded comparison kernels with a nondegenerate Hájek projection satisfy a joint U-statistic central limit theorem. The smooth log-ratio map then supplies joint asymptotic normality with other regular marginal estimators.
At the common-law reference, \(p_W=p_L=q>0\). Antisymmetry gives \(a(o)=E_P k(o,O')\), \(E_Pa=0\), and
\[ h_{\mathrm{NB}}(o)=a(o), \qquad h_{\mathrm{WR}}(o)=a(o)/q, \qquad \phi_{\mathrm{WR}}(A,O)=W(A)h_{\mathrm{WR}}(O). \tag{12} \]
The net-benefit and log-win-ratio components therefore span the same adjustment direction at this reference when based on the same kernel. Their population \(R^2\) values agree, although their nonnull estimands differ. With one continuous endpoint, no ties, zero margin and larger values preferred, \(a(z)=2F_Z(z)-1\) and \(q=1/2\).
The component \(a(O_i)\) can be estimated by the average of \(k(O_i,O_j)\) over \(j\ne i\). These estimated components are dependent, and their joint second-moment expansion must account for estimation of the reference distribution. Section 6 provides that expansion. For censored comparisons, observed ordering may identify a win or loss even when one time is censored; unresolved pairs follow the prespecified comparison rule. The resulting observed-record estimand can depend on the censoring distribution. A latent uncensored interpretation requires additional identification conditions.
A nominal variable with \(K\) categories can be represented by \(K-1\) linearly independent centered indicators. Several prespecified summaries of one clinical variable can likewise form a block. PATED applies to the resulting vector of auxiliary contrasts; its optimal variance reduction depends on their linear span and is invariant to nonsingular recodings within that span. For selection of entire variables rather than individual coordinates, an adaptive block penalty is developed in Appendix A. This distinguishes a change of representation basis from adding genuinely new prognostic information.
Once PATED and its primary marginal statistic have been specified, historical data can be used to study which auxiliary contrasts should be retained and what variance reduction they provide. Let \(P\) be the historical distribution of the baseline variables and a corresponding outcome. It may describe a prior control arm, a single-arm cohort, or a real-world population. Its independent observational units are participants, rather than individual claim records.
An artificial treatment label supplies a reference comparison when the historical observations do not contain the future randomized contrast. This reference sets both arm distributions equal to \(P\). Consequently, it addresses variance reduction at a null treatment comparison, where the variance of an estimator remains meaningful even though the contrast itself is zero. It does not infer a treatment effect from observational association.
Let \(O_1,\ldots,O_n\) be independent observations from a population distribution \(P\). Each candidate variable \(X_j\), \(j=1,\ldots,p\), and the response \(Y\) is represented by a prespecified scalar null-contrast score. For differences of regular marginal functionals \(T_j(P)\) and \(T_Y(P)\) these are their centered influence functions; for pairwise comparisons they are the arm-directional projections in (12). Write
\[ h_j(O;P),\qquad h_Y(O;P),\qquad E_P h_j=E_P h_Y=0. \tag{13} \]
The notation permits observed-data records, such as an event time together with its censoring indicator. The score must be derived for that observation scheme. For fully observed marginal variables, it can depend only on that variable.
Generate the same artificial label \(A_i\sim\operatorname{Bernoulli}(\pi)\) for all variable-specific analyses of subject \(i\), independently of \(O_i\), with fixed \(0<\pi<1\). Let \(\hat{\beta}_j\) be the chosen two-group contrast; differences between groupwise estimators of \(T_j\) are one example. Assume the estimators have compatible, jointly valid asymptotic linear expansions. Because the group distributions coincide, all true contrasts are zero, and
\[ \sqrt n\,\hat{\beta}_j =\frac1{\sqrt n}\sum_{i=1}^n \frac{A_i-\pi}{\pi(1-\pi)}h_j(O_i;P)+o_p(1). \tag{14} \]
This expansion follows by subtracting the two groupwise expansions and using \(n_1/n\to_p\pi\). It also covers other regular contrast estimators when their null influence functions factor in the same way. For win statistics, \(h_j\) is an arm-directional projection, not the influence function of the identically zero diagonal functional \(T(P,P)\); conflating these would incorrectly give a zero score. In the historical-data setting, (19) below equals the null-reference PATED quantity in (6).
Stack \(h=(h_X^\top,h_Y)^\top\), where \(h_X=(h_1,\ldots,h_p)^\top\), and define
\[ G=E_P(hh^\top) =\begin{pmatrix}G_{XX}&g_{XY}\\g_{XY}^\top&G_{YY}\end{pmatrix}. \tag{15} \]
Because the components incorporate the complete marginal influence adjustment, \(G\) already contains the corresponding sandwich factors. Its off-diagonal entries retain dependence among the marginal analyses.
Proposition 1 (artificial-group covariance). If the joint expansion (14) holds and \(E\|h\|^2<\infty\), the asymptotic covariance of \(\sqrt n(\hat{\beta}_X^\top,\hat{\beta}_Y)^\top\) is
\[ \Sigma=\frac{G}{\pi(1-\pi)}. \tag{16} \]
Proof. Independence of \(A\) and \(O\), together with \(E(A-\pi)^2=\pi(1-\pi)\), gives the covariance in (16); apply the multivariate central limit theorem to (14). \(\square\)
The same subject must receive the same artificial label in every variable-specific analysis. Labels may be independent across subjects. To see why sharing matters, let \(W=(A-\pi)/\{\pi(1-\pi)\}\). A common label gives \(E(W^2h_jh_k)=G_{jk}/\{\pi(1-\pi)\}\). Separate labels that are mutually independent and independent of the data instead give
\[ E(W_jW_kh_jh_k)=E(W_j)E(W_k)E(h_jh_k)=0,\qquad j\ne k. \]
Independent labels therefore remove the desired first-order cross-covariance of the contrasts, even when the original variables are associated. They do not remove association from the underlying data.
The equation \(\hat{\beta}_Y=c^\top\hat{\beta}_X\) does not identify a meaningful population coefficient from one grouping. For \(p>1\), it is typically one equation with many solutions. Even with \(p=1\), the ratio \(\hat{\beta}_Y/\hat{\beta}_X\) generally converges in distribution to a ratio of correlated normal variables rather than to a constant, since both contrasts have population value zero.
By Theorem 1 and Proposition 1, the historical PATED coefficient minimizes the residual influence variance:
\[ c_0 =\arg\min_c E\{h_Y-c^\top h_X\}^2 =G_{XX}^{-1}g_{XY} =\Sigma_{XX}^{-1}\Sigma_{XY}. \tag{17} \]
Here and below \(G_{XX}\) is positive definite and \(G_{YY}>0\). Equation (17) makes the target independent of the artificial allocation proportion \(\pi\). If the loss is multiplied by the common covariance scale in (16), its penalty must be multiplied by the same scale to preserve a finite-sample penalized solution.
Let
\[ u=h_Y-c_0^\top h_X,\qquad E(h_Xu)=0, \qquad S_0=\{j:c_{0j}\ne0\}. \tag{18} \]
The orthogonality in (18) holds by projection. Neither \(E(u\mid h_X)=0\) nor homoskedasticity is assumed.
For a subset \(S\) of predictors, write
\[ v(S)=\min_a E(h_Y-a^\top h_S)^2, \qquad R^2=1-\frac{v(\{1,\ldots,p\})}{G_{YY}} =\frac{g_{XY}^\top G_{XX}^{-1}g_{XY}}{G_{YY}}. \tag{19} \]
Define the additional contribution of variable \(j\), conditional on retaining all other candidate variables, as
\[ D_j=\frac{v(-j)-v(\mathrm{all})}{G_{YY}} =\frac{c_{0j}^2G_{j\mid-j}}{G_{YY}}, \qquad G_{j\mid-j}=G_{jj}-G_{j,-j}G_{-j,-j}^{-1}G_{-j,j}. \tag{20} \]
Proposition 2 (interpretation and invariance). Under the stated definiteness conditions, \(0\le R^2\le1\), \(0\le D_j\le1\), and \(D_j=0\) if and only if \(c_{0j}=0\). The support, \(R^2\), and each \(D_j\) are unchanged by separate nonzero rescalings of predictor scores and the response score.
Proof. Residualize \(h_j\) on \(h_{-j}\), obtaining \(r_j\) with variance \(G_{j\mid-j}>0\). Orthogonal projection gives \(v(-j)=v(\mathrm{all})+c_{0j}^2E(r_j^2)\), proving (20) and the bounds. Under \(h_j^*=a_jh_j\) and \(h_Y^*=bh_Y\), the coefficient becomes \(c_{0j}^*=bc_{0j}/a_j\), the residual predictor variance becomes \(a_j^2G_{j\mid-j}\), and response variance becomes \(b^2G_{YY}\). The factors cancel. \(\square\)
The \(D_j\) generally do not add to \(R^2\). They describe deletion of one variable from the full candidate set and therefore depend on that set. They are nonnegative strengths; the sign of \(c_{0j}\) can be reported separately on a prespecified score orientation.
A smooth re-expression of a marginal statistic, \(T_j^*=f_j(T_j)\) with \(f_j'(T_j)\ne0\), rescales its influence function. Proposition 2 therefore gives invariance to such reparameterizations. It does not give invariance to an arbitrary change in the underlying statistic.
For example, \(E(X)\) and \(\log E(X)\), when \(E(X)>0\), have proportional influence functions; \(E(X)\) and \(E(\log X)\) generally do not. These are population invariance statements. Nonlinear refitting and tuning can produce finite-sample differences.
Regular model fitting is not mandatory. Choose a fixed measurable feature \(\phi_j\) and take
\[ T_j(P)=E_P\{\phi_j(X_j)\},\qquad h_j=\phi_j(X_j)-E_P\{\phi_j(X_j)\}. \tag{21} \]
This is a nonparametric marginal functional. Bounded features automatically satisfy the moment conditions needed below. Examples include category indicators, prespecified ordinal thresholds, and fixed bounded transformations of continuous variables. Several features for one original variable can be treated as a block, as developed in Appendix A.
Fix the observed sample and any associated vectors \(\hat{h}_1,\ldots,\hat{h}_n\). They may have been estimated from the pooled data, but are now held fixed as assignments vary. Assign exactly \(n_1\) observations to group 1 uniformly over all such assignments, with \(n_0=n-n_1\). Let
\[ \Delta_A=\overline{\hat{h}}_{A=1}-\overline{\hat{h}}_{A=0}, \quad S_{\hat{h}}=\frac1{n-1}\sum_i(\hat{h}_i-\overline{\hat{h}}) (\hat{h}_i-\overline{\hat{h}})^\top. \]
Proposition 3 (finite-sample integration).
\[ \operatorname{Cov}_A(\Delta_A\mid O_{1:n},\hat{h}_{1:n}) =\left(\frac1{n_1}+\frac1{n_0}\right)S_{\hat{h}}. \tag{22} \]
Proof. Put \(z_i=\hat{h}_i-\overline{\hat{h}}\). Then \(\Delta_A=n(n_1n_0)^{-1}\sum_i A_iz_i\) and \(\sum_i z_i=0\). For complete randomization, \(\operatorname{Var}(A_i)=n_1n_0/n^2\) and \(\operatorname{Cov}(A_i,A_k)=-n_1n_0/\{n^2(n-1)\}\) for \(i\ne k\). Substitution, using \(\sum_{i\ne k}z_iz_k^\top=-\sum_i z_iz_i^\top\), yields (22). \(\square\)
Thus averaging the covariance over artificial groups is available analytically. One can estimate the association parameter from a centered pooled score covariance. Equation (22) is exact for these frozen-score contrasts. For an estimator that is refitted nonlinearly inside every artificial group, it is an influence-function approximation unless a separate exact identity is available.
With \(\pi_n=n_1/n\), the estimator on the asymptotic covariance scale in (16) is \(\hat{\Sigma}_{\mathrm{int}}=S_{\hat{h}}/\{\pi_n(1-\pi_n)\}\). Its consistency requires consistent estimation of the influence-component second moments. The allocation factor cancels from the projection coefficient and \(R^2\), so both can be estimated directly from \(S_{\hat{h}}\), without generating artificial labels. The identity applies to fixed-size uniform allocation and, conditionally on nonempty group sizes, to Bernoulli allocation.
The historical planning target is the variance of the PATED residual component, \(h_Y-c^\top h_X\). Proposition 3 links this target directly to the mean squared residual contrast over artificial assignments. Writing \(k_n=1/n_1+1/n_0\),
\[ E_A\{(\Delta_{Y,A}-c^\top\Delta_{X,A})^2\mid O_{1:n},\hat{h}_{1:n}\} =k_n\{S_{\hat{h},YY}-2c^\top S_{\hat{h},XY}+c^\top S_{\hat{h},XX}c\}. \]
After removing the positive allocation scale and the term independent of \(c\), the objective is quadratic. Thus penalizing this objective selects adjustment terms according to their contribution to the variance of the intended statistic.
Let \(\hat{G}\) estimate (15), with blocks \(\hat{G}_{XX}\) and \(\hat{g}_{XY}\). Take \(\hat{G}\) to be the centered empirical covariance of the estimated pooled components, using divisor \(n\). The unpenalized projection supplies an initial, or pilot, estimator for the adaptive penalty weights:
\[ \tilde{c}=\hat{G}_{XX}^{-1}\hat{g}_{XY},\qquad w_{jn}=(|\tilde{c}_j|\vee\epsilon_n)^{-\gamma},\quad \gamma>0. \tag{23} \]
The penalized estimate is
\[ \hat{c}=\arg\min_c\left\{ \frac12c^\top\hat{G}_{XX}c-c^\top\hat{g}_{XY} +\lambda_n\sum_{j=1}^p w_{jn}|c_j|\right\}. \tag{24} \]
Equivalently, its quadratic term is \(\tfrac12(c-\tilde{c})^\top\hat{G}_{XX}(c-\tilde{c})\) up to a constant. The weights must be based on an estimate of \(c_0\), not on the artificial contrasts \(\hat{\beta}_j\), whose population values are all zero.
This combines the adaptive penalty of Zou [9] with a quadratic objective of the kind studied by Wang and Leng [10]. Here the quadratic matrix is the covariance of predictor components. It is generally not the inverse sampling covariance of \(\tilde{c}\), so the selection result is established directly for the PATED projection.
The theorem treats a fixed number of candidate components as the historical sample size increases.
A1. Sampling and a fixed target. The sampling units are independent and identically distributed. The score system and its population limits define a deterministic \(G\). The true support \(S_0=\operatorname{supp}(c_0)\) is fixed, with \(1\le |S_0|<p\); write \(N_0=S_0^c\). These are population sets, distinct from the random selected set \(\hat{S}=\operatorname{supp}(\hat{c})\).
A2. Identifiability and separated signals. \(G_{XX}\) is positive definite, \(G_{YY}>0\), and \(\min_{j\in S_0}|c_{0j}|>0\). The nonzero coefficients do not approach zero with \(n\) in this theorem.
A3. A root-\(n\) expansion for the estimated Gram matrix. For a symmetric, mean-zero random matrix \(K(O)\) with \(E\|K\|_F^2<\infty\),
\[ \sqrt n(\hat{G}-G)=\frac1{\sqrt n}\sum_{i=1}^n K(O_i)+o_p(1). \tag{25} \]
The remainder is entrywise, equivalently in any matrix norm when \(p\) is fixed. The empirical objective is convex, and \(\hat{G}_{XX}\) is positive definite with probability approaching one.
A4. A root-\(n\) pilot and controlled weights. \(\sqrt n(\tilde{c}-c_0)=O_p(1)\), \(\gamma>0\), and \(0<\epsilon_n=o(n^{-1/2})\). The pilot in (23) satisfies the required rate under A2–A3. Another pilot may be substituted if its rate is established.
A5. Penalty separation. With the normalization in (24),
\[ \sqrt n\lambda_n\longrightarrow0, \qquad n^{(\gamma+1)/2}\lambda_n\longrightarrow\infty. \tag{26} \]
These limits can be in probability for a data-dependent penalty. For example, for \(\gamma=1\), \(\lambda_n=k\hat{G}_{YY}n^{-3/4}\) with fixed \(k>0\) satisfies (26) if \(\hat{G}_{YY}\to_pG_{YY}>0\); \(\epsilon_n=n^{-1}\) is admissible. The constant \(k\) still matters greatly in finite samples.
A6. Nondegenerate active uncertainty. For
\[ z(O)=K_{XY}(O)-K_{XX}(O)c_0, \tag{27} \]
define \(\Omega=\operatorname{Var}\{z(O)\}\). Its active block \(\Omega_{S_0S_0}=\operatorname{Var}\{z_{S_0}(O)\}\) is positive definite when nonsingular normal limits and Wald intervals are desired. A singular normal limit is possible without this last definiteness condition.
Assumption A3 is substantive. Consistency of a covariance estimate for artificial contrasts, by itself, is not enough to prove a root-\(n\) oracle result for a coefficient estimated from that covariance.
Covariance of the unpenalized pilot. On the event that \(\hat{G}_{XX}\) is invertible, the pilot in (23) satisfies the exact identity
\[ \tilde{c}-c_0 =\hat{G}_{XX}^{-1}\{(\hat{g}_{XY}-g_{XY})-(\hat{G}_{XX}-G_{XX})c_0\}. \]
A2–A3 therefore give
\[ \sqrt n(\tilde{c}-c_0) =\frac1{\sqrt n}\sum_iG_{XX}^{-1}z(O_i)+o_p(1) \ \overset{d}{\longrightarrow}\ N(0,V),\qquad V=G_{XX}^{-1}\Omega G_{XX}^{-1}. \]
The first-order covariance approximation for \(\tilde{c}\) is \(V/n\). In the fixed-feature case (35), \(z=h_Xu\) and \(\Omega=E(h_Xh_X^\top u^2)\). For estimated influence functions, use the full \(K\) from a justified expansion, including the contributions described in Section 6. In particular, \(\hat{G}_{XX}\) in (24) is not generally an inverse-covariance weight: if \(\Omega\) is nonsingular, \(V^{-1}=G_{XX}\Omega^{-1}G_{XX}\). Only under additional conditions such as \(E(u^2\mid h_X)=\sigma_u^2\) in the fixed-feature case does \(V=\sigma_u^2G_{XX}^{-1}\).
Theorem 2 (pointwise adaptive oracle property). Under A1–A6, the minimizer of (24) is unique with probability approaching one and satisfies
\[ P\{\operatorname{supp}(\hat{c})=S_0,\quad \operatorname{sign}(\hat{c}_{S_0})=\operatorname{sign}(c_{0,S_0})\}\longrightarrow1, \tag{28} \]
and
\[ \sqrt n(\hat{c}_{S_0}-c_{0,S_0}) \ \overset{d}{\longrightarrow} N\left(0,G_{S_0S_0}^{-1}\Omega_{S_0S_0}G_{S_0S_0}^{-1}\right). \tag{29} \]
Here \(S_0\) is the fixed true support. The limiting distribution in (29) coincides with that of the unpenalized estimator restricted to \(S_0\).
Proof. Put \(N_0=S_0^c\), \(s_0=\operatorname{sign}(c_{0,S_0})\) and \(e_n=\hat{g}_{XY}-\hat{G}_{XX}c_0\). Since \(g_{XY}=G_{XX}c_0\), A3 implies
\[ \sqrt n e_n=\frac1{\sqrt n}\sum_i z(O_i)+o_p(1), \qquad e_n=O_p(n^{-1/2}). \tag{30} \]
For \(j\in S_0\), A4 gives \(w_{jn}\to_p|c_{0j}|^{-\gamma}\), whence \(\lambda_nw_{S_0}=o_p(n^{-1/2})\) by A5. For the inactive coordinates,
\[ \sqrt n\lambda_n\min_{j\in N_0}w_{jn} =\frac{n^{(\gamma+1)/2}\lambda_n} {\{\max_{j\in N_0}(\sqrt n|\tilde{c}_j|\vee\sqrt n\epsilon_n)\}^{\gamma}} \ \longrightarrow_p\infty, \tag{31} \]
because the denominator is \(O_p(1)\) and the numerator diverges.
On the event \(\hat{G}_{XX}\succ0\), whose probability tends to one, define \(c^*_{N_0}=0\) and
\[ c_{S_0}^*=\hat{G}_{S_0S_0}^{-1} \{\hat{g}_{S_0Y}-\lambda_n w_{S_0}\odot s_0\}. \tag{32} \]
Since \(c_{0,N_0}=0\) and \(\hat{G}_{S_0S_0}^{-1}=O_p(1)\),
\[ c_{S_0}^*-c_{0,S_0} =\hat{G}_{S_0S_0}^{-1} \{e_{n,S_0}-\lambda_n w_{S_0}\odot s_0\} =O_p(n^{-1/2}). \tag{33} \]
By A2, \(\operatorname{sign}(c^*_{S_0})=s_0\) with probability tending to one; hence (32) satisfies the active Karush–Kuhn–Tucker equations. The inactive gradient is
\[ \hat{G}_{N_0S_0}c_{S_0}^*-\hat{g}_{N_0Y} =-e_{n,N_0}+\hat{G}_{N_0S_0}\hat{G}_{S_0S_0}^{-1} \{e_{n,S_0}-\lambda_n w_{S_0}\odot s_0\} =O_p(n^{-1/2}). \tag{34} \]
Together, (31) and (34) imply that every inactive gradient is smaller in absolute value than its penalty threshold \(\lambda_nw_{jn}\) with probability tending to one. The candidate therefore satisfies all Karush–Kuhn–Tucker conditions. Strict convexity gives \(\hat{c}=c^*\) with probability tending to one, proving (28).
Multiplying (33) by \(\sqrt n\) and applying (30), A5 and Slutsky’s theorem yields (29). Moreover, \(\hat{G}_{S_0S_0}^{-1}\hat{g}_{S_0Y}-c^*_{S_0} =\lambda_n\hat{G}_{S_0S_0}^{-1}(w_{S_0}\odot s_0)=o_p(n^{-1/2})\), which establishes equivalence to the unpenalized estimator restricted to the true support. \(\square\)
If \(S_0=\varnothing\), the inactive argument with \(c^*=0\) gives \(P(\hat{c}=0)\to1\) under the remaining assumptions; (29) is vacuous.
For fixed features as in (21), with \(E\|h\|^4<\infty\), a centered empirical covariance satisfies
\[ K(O)=h(O)h(O)^\top-G, \qquad z(O)=h_Xu. \tag{35} \]
Estimating the marginal means does not add a first-order term: the centering correction is the outer product of sample mean errors, of order \(O_p(n^{-1})\). Thus A3 follows directly from a central limit theorem for score products.
The active oracle covariance becomes
\[ V_{S_0}=G_{S_0S_0}^{-1}E(h_{S_0}h_{S_0}^\top u^2)G_{S_0S_0}^{-1}. \tag{36} \]
This sandwich form allows nonlinear conditional means and heteroskedastic residuals. Replacing it automatically by a residual variance times \(G_{S_0S_0}^{-1}\) would require additional conditions. The empirical sandwich, evaluated on the selected support, is pointwise consistent when support recovery holds. It does not supply a uniform post-selection confidence guarantee near zero coefficients.
If \(\hat{G}_{XX}\) is diagonal and the stabilizing floor is inactive, the solution is
\[ \hat{c}_j=\operatorname{sign}(\tilde{c}_j) \left(|\tilde{c}_j|-\frac{\lambda_n}{\hat{G}_{jj}|\tilde{c}_j|^\gamma}\right)_+. \tag{37} \]
Therefore a variable is selected exactly when
\[ |\tilde{c}_j|^{\gamma+1}>\lambda_n/\hat{G}_{jj}. \tag{38} \]
With \(\gamma=1\) and \(\lambda_n\asymp n^{-3/4}\), the selection threshold for \(|\tilde{c}_j|\) is of order \(n^{-3/8}\). Relative to root-\(n\) noise, this threshold increases only as \(n^{1/8}\). This slow separation explains why asymptotic support consistency can coexist with many false selections at realistic sample sizes when the penalty constant is small.
For a training fold, estimate its scores, center them using training quantities, estimate its pilot and adaptive weights, and fit a penalty path. Evaluate held-out squared score residuals,
\[ \mathrm{CV}(\lambda)=\frac1n\sum_k\sum_{i\in I_k} \left\{\hat{h}_Y^{(-k)}(O_i) -\hat{c}_{\lambda}^{(-k)\top}\hat{h}_X^{(-k)}(O_i)\right\}^2. \tag{39} \]
Here the superscript means that nuisance quantities and centering values were obtained without the validation observations. A common grid of fractions of a fold-specific maximum penalty is also possible and is used in the simulation. In that case the tuning parameter in (39) is the fraction, and each fold converts it to its own absolute penalty.
This criterion is aligned with the defining projection loss. It estimates out-of-sample residual influence variance. All components, centering constants, pilot estimates and adaptive weights must be estimated within the training fold.
Risk-optimal tuning and exact support recovery are different goals. Neither ordinary CV nor a one-standard-error rule automatically verifies (26). The oracle theorem applies to a data-selected penalty only after its required rates have been established. The simulation therefore reports the fixed-rate rule and CV separately.
The default examined here is \(\gamma=1\), a fixed positive exponent satisfying the theorem. It is not uniquely optimal. Larger \(\gamma\) increases the distinction between small and large pilot coefficients and can make selection more sensitive to a noisy pilot.
Without a floor, the penalty \(\sum_j|c_j|/|\tilde{c}_j|\) for \(\gamma=1\) is invariant to separate predictor rescalings. Under a response rescaling by \(b\), the quadratic loss is multiplied by \(b^2\), so the absolute penalty level must also be multiplied by \(b^2\). A rule proportional to \(\hat{G}_{YY}\) has this property. A stabilizing floor should be transformed with the corresponding coefficient scale if exact finite-sample invariance is required; an unscaled floor is only asymptotically negligible under A4.
Theorem 2 is pointwise at a fixed distribution with fixed nonzero signals. For a local coefficient \(c_{0j,n}=d/\sqrt n\), the pilot remains \(O_p(n^{-1/2})\). Under the strong penalty condition in (26), that local coefficient can be excluded with probability approaching one. There is no contradiction: A2 does not hold along that sequence. This illustrates the nonuniformity analyzed more generally by Pötscher and Schneider [11].
Consequently, a selected-variable confidence interval is not automatically a valid confidence statement for a variable chosen after looking at the data. The oracle theorem also provides neither false-discovery-rate control nor a conditional-independence interpretation.
If \(G_{XX}\) is singular, for example because two candidate scores are identical, individual coefficients and their support need not be identifiable. Combining redundant scores into a meaningful block, removing exact redundancy, or changing the estimand is necessary. Adding a fixed ridge penalty makes the optimization well behaved but changes the target and does not retain Theorem 2 unchanged.
For distribution-dependent scores \(h(O;P)\), the empirical covariance of estimated scores need not have influence function \(hh^\top-G\). Under a valid differentiable plug-in expansion, the \((j,k)\) entry has the form
\[ K_{jk}(o) =h_j(o)h_k(o)-G_{jk} +E\{\dot{h}_j(O;o)h_k(O)+h_j(O)\dot{h}_k(O;o)\}, \tag{40} \]
where \(\dot{h}_j(O;o)\) denotes the first-order change in the score function induced by contaminating \(P\) at \(o\). Equation (40) is a functional derivative formula under conditions permitting differentiation of the expectation. For nonsmooth scores a separate argument may be needed.
Theorem 2 still applies if the full \(K\) satisfies A3, but its covariance must use (27), including the additional terms. Merely observing that the original marginal estimators are asymptotically normal does not prove (25).
Cross-fitting can help control empirical-process dependencies. It does not by itself make the covariance moment insensitive to nuisance estimation. The often-used \(o_p(n^{-1/4})\) nuisance rate is not a generic solution here: it is useful when the relevant moment has the necessary first-order orthogonality and its remainder has a suitable product structure. For an arbitrary covariance of estimated influence functions, that orthogonality must be derived.
There are three routes to a valid covariance expansion: use fixed features for which (35) holds; derive the full plug-in expansion; or construct and justify an orthogonal estimator of the required Gram entries. The latter is an extension requiring a separate derivation for each statistical representation.
Corollary 1 (a constructive smooth-nuisance case). Suppose \(h(O;\eta)\) depends on a fixed-dimensional nuisance parameter, \(E h(O;\eta_0)=0\), and
\[ \sqrt n(\hat{\eta}-\eta_0)=n^{-1/2}\sum_i a(O_i)+o_p(1), \qquad E a=0,\quad E\|a\|^2<\infty. \]
Assume \(h\) is twice continuously differentiable near \(\eta_0\), \(E\|h(O;\eta_0)\|^4<\infty\), and integrable uniform envelopes for the first two derivatives of \(h\) and \(hh^\top\) justify a uniform law of large numbers and the Taylor remainders. Then the centered empirical covariance of \(h(O_i;\hat{\eta})\) satisfies A3 with
\[ K(O)=h(O;\eta_0)h(O;\eta_0)^\top-G +\sum_r D_r a_r(O), \quad D_r=E\left[\frac{\partial}{\partial\eta_r} \{h(O;\eta)h(O;\eta)^\top\}\bigg|_{\eta=\eta_0}\right]. \]
Proof. Taylor-expand the empirical uncentered second moment at \(\eta_0\). Its linear term is the empirical derivative multiplied by \(\hat{\eta}-\eta_0\); the empirical derivative converges to \(D_r\) and the quadratic remainder is \(O_p(n^{-1})\). Substituting the nuisance expansion gives the stated \(K\). The empirical mean of the estimated scores is \(O_p(n^{-1/2})\), so subtracting its outer product also changes the covariance by only \(O_p(n^{-1})\). The finite-moment conditions give the central limit theorem. \(\square\)
This corollary is one verifiable route to the estimated-score oracle theorem. It does not require independence between the nuisance estimate and the observations used in the Gram matrix. The cross-covariance between the two terms of \(K\) is retained automatically.
For fixed bounded antisymmetric kernels, the additional expansion required by the oracle theorem can be made explicit. Suppose coordinate \(j\) has \(a_j(o)=E\{k_j(o,O')\}\) and \(h_j=a_j/q_j\), with \(q_j>0\) either fixed or a win probability. Define \(B_{jk}=E(a_ja_k)\), so \(G_{jk}=B_{jk}/(q_jq_k)\). If \(q_j=E\{w_j(O,O')\}\), its influence function is
\[ r_j(o)=E\{w_j(o,O')+w_j(O',o)\}-2q_j; \tag{41} \]
set \(r_j=0\) for a fixed denominator. Differentiating the outer distribution and the two inner reference distributions in \(B_{jk}\) gives
\[ L_{jk}(o) =a_j(o)a_k(o) +E\{a_k(Z)k_j(Z,o)+a_j(Z)k_k(Z,o)\} -3B_{jk}. \tag{42} \]
Therefore the empirical pooled-score covariance has A3 with
\[ K_{jk}(o)= \frac{L_{jk}(o)}{q_jq_k} -G_{jk}\left\{\frac{r_j(o)}{q_j}+\frac{r_k(o)}{q_k}\right\}. \tag{43} \]
Indeed, the numerator is a bounded statistic of triples of observations; its first-order Hoeffding projection is (42). The denominator is a bounded statistic of pairs, giving (41), and the delta method gives (43). Diagonal and leave-one-out conventions differ by \(O_p(n^{-1})\). Antisymmetry centers the population components; empirical centering changes no first-order term. Thus a joint multivariate central limit theorem supplies A3. If the active variance is nondegenerate and the remaining assumptions hold, Theorem 2 applies to this win-component construction.
Centered bounded fixed features also fit this calculation by taking \(k_j(o,o')=f_j(o)-f_j(o')\) and \(q_j=1\). For comparisons involving unbounded features or estimated survival hazards, the corresponding moment and nuisance-expansion conditions must be verified separately. Joint normality of the marginal estimates is the starting point for PATED; (43) verifies the stronger covariance expansion needed for selection in this particular nonparametric subclass.
For a prespecified subset \(S\), replace the covariance blocks in (6) by those for \(S\). This gives the optimal variance reduction \(R^2(S)\). Variable selection asks for a small useful subset; reporting only the full-candidate \(R^2\) does not answer that question.
For any fitted, frozen coefficient vector \(c\), its actually achieved population variance reduction is
\[ \mathcal{V}(c) =\frac{2c^\top g_{XY}-c^\top G_{XX}c}{G_{YY}} =R^2-\frac{(c-c_0)^\top G_{XX}(c-c_0)}{G_{YY}}, \tag{44} \]
Only the optimal coefficient gives \(R^2\) exactly. A noisy or over-shrunk coefficient can yield less reduction, and \(\mathcal{V}(c)\) can be negative. Apparent in-sample \(R^2\) is therefore insufficient to evaluate a selected adjustment.
Use an independent historical validation sample, or an outer validation scheme enclosing selection, to evaluate \(1-\widehat{\mathrm{Var}}_{\mathrm{val}}(h_Y-c^\top h_X)/ \widehat{\mathrm{Var}}_{\mathrm{val}}(h_Y)\). Keep the fitted score definitions and coefficients fixed when evaluating them. This ratio estimates achieved reduction for that fitted rule; it is not an exactly unbiased estimator of the population ratio. A refit on selected coordinates can be evaluated separately from the penalized coefficients. Outcome records in the validation cohort may be used to compute the validation scores, but not to tune the fitted rule.
For trial planning, the usual output is a selected set, its statistic definitions, and a validated estimate of variance reduction. The eventual trial can re-estimate PATED’s covariance coefficients on this prespecified set. If instead historical coefficients are frozen for use in the trial, (44) must be evaluated under the trial covariance: population shifts can change its sign. Our historical-data oracle result concerns \(c_0\) and selection in the historical population; it is not a proof of transportability.
For a fixed covariate set and the joint Gram expansion in A3, differentiating (19) gives the first-order influence function
\[ \psi_{R^2}(O)= \frac{2c_0^\top K_{XY}(O)-c_0^\top K_{XX}(O)c_0-R^2K_{YY}(O)} {G_{YY}}. \tag{45} \]
Thus the usual asymptotic variance is \(E(\psi_{R^2}^2)/n\), when nondegenerate. To obtain (45), differentiate \(g^\top G_{XX}^{-1}g/G_{YY}\) and use \(d(G_{XX}^{-1})=-G_{XX}^{-1}(dG_{XX})G_{XX}^{-1}\). The same formula applies to a prespecified subset using its corresponding blocks. Estimated-score contributions enter through the complete \(K\), just as they do for coefficient uncertainty. At zero association the first derivative can vanish, so an ordinary nondegenerate normal interval is not justified by this formula alone.
For independent validation of a fitted rule, freeze its scores and coefficients conditional on the development sample. Let \(r_c=h_Y-c^\top h_X\), \(v_r=\operatorname{Var}(r_c)\) and \(v_Y=\operatorname{Var}(h_Y)\). The validation variance-reduction ratio has conditional influence function
\[ -\frac{(r_c-Er_c)^2-v_r}{v_Y} +\frac{v_r}{v_Y^2}\{(h_Y-Eh_Y)^2-v_Y\}. \tag{46} \]
With finite fourth moments, its empirical variance divided by the validation sample size supplies a standard error for that frozen rule’s performance. This conditional interval does not measure variability from repeating model development, nor uncertainty about transport to a new population. These distinctions prevent a selected in-sample \(R^2\) from being reported with an unjustified fixed-model interval.
The experiments represent a large historical development cohort: \(n=5000,10000,20000\), \(p=20\), and three, four, or five active covariates. The three signal configurations have population \(R^2=0.08,0.12,0.15\), respectively. Each of these nine scenarios has 200 independent replications. Two additional scenarios at \(n=10000\), four signals, and \(R^2=0.12\) examine win-ratio scores and a censored baseline covariate. Each development cohort has an independently generated validation cohort of 5,000. There are 2,200 development–validation pairs in total.
For the main mixed-variable design, \(X_1,\ldots,X_6\) are jointly standard normal with correlation \(0.5^{|j-k|}\). Remaining coordinates alternate independent standardized Bernoulli \((1/2)\), Poisson \((2)\), and standard normal measurements. The active set is the first \(s\) indices of \((1,7,8,9,12)\), with coefficients given by the first \(s\) entries of \((0.7,-0.6,0.5,0.4,-0.35)\). Let \(b\) denote this vector and \(G_X=\operatorname{Cov}(X)\). Generate
\[ Y=X^\top b+ a\left\{0.5X_{10}X_{11}+(0.8+0.3|X_1|)\varepsilon\right\}, \qquad \varepsilon\sim t_5\sqrt{3/5},\quad \varepsilon\perp X. \tag{47} \]
The residual term in braces is orthogonal to every coordinate of \(X\) and has variance \(v_e=0.25+0.64+0.48\sqrt{2/\pi}+0.09\). For a requested population value \(r\), choose
\[ a^2=\frac{b^\top G_X b}{v_e}\left(\frac1r-1\right). \tag{48} \]
This makes the population score-projection \(R^2\) exactly \(r\), rather than tuning it after inspecting simulations. Mean-contrast scores are used for the mixed measurements. The nonlinearity and heterogeneous residual variance mean that the linear score projection is not a correctly specified conditional-mean model. The construction isolates the planning problem; it is not intended as a full model of a claims database.
Fivefold CV uses 40 logarithmically spaced fractions from 1 to \(10^{-4}\) of each training fold’s maximum penalty. Score construction, centering, predictor scaling, pilot fitting, and adaptive weights are training-only. Adaptive weights use \(\gamma=1\) and floor \(n_{\mathrm{train}}^{-1}\) on the standardized-predictor coefficient scale. The compared rules are:
Selection is determined by the original penalized fit, using \(10^{-8}\) as the nonzero threshold. For the primary variance assessment, selected coefficients are then refitted without a penalty on that selected set; the refit is frozen before external validation. The oracle is a benchmark, not an implementable method. For the full estimator, exact support is necessarily zero and FDP is large by construction; those entries describe its lack of selection, not a defect in its definition.
Exact support is the fraction of replications with \(\hat{S}=S_0\). TPR is \(|\hat{S}\cap S_0|/|S_0|\). FDP is \(|\hat{S}\setminus S_0|/\max(1,|\hat{S}|)\), averaged across replications. Excess risk is \((c-c_0)^\top G_{XX}(c-c_0)/G_{YY}\) here, normalized to the primary-score variance. By (44), it is exactly the loss in population variance reduction relative to the optimal adjustment.
Population gain is (44) evaluated using the known data-generating covariance and the refitted coefficients. Validation gain is the independent-sample residual-variance ratio in Section 7.1. Apparent \(R^2\) is the refitted development-sample projection value. Percentages in the tables denote variance reductions; differences between them are percentage points. Reported \(\pm\) values are Monte Carlo standard errors of the simulation mean, not confidence intervals for a single historical study. For 200 replications the maximum Monte Carlo SE of an exact-support proportion is 0.035.
The first table shows the primary sparse rule, adaptive 1-SE followed by refitting, across all nine mixed-variable scenarios.
| Development n | Signals | True R² | Exact support | TPR | Mean FDP | Population gain | Validation gain ± MC SE (pp) |
|---|---|---|---|---|---|---|---|
| 5,000 | 3 | 8.00% | 0.870 | 0.948 | 0.000 | 7.64% | 7.55% ± 0.09 |
| 5,000 | 4 | 12.00% | 0.745 | 0.926 | 0.000 | 11.43% | 11.35% ± 0.11 |
| 5,000 | 5 | 15.00% | 0.815 | 0.954 | 0.000 | 14.55% | 14.53% ± 0.12 |
| 10,000 | 3 | 8.00% | 0.980 | 0.992 | 0.000 | 7.92% | 7.98% ± 0.06 |
| 10,000 | 4 | 12.00% | 0.960 | 0.989 | 0.000 | 11.89% | 11.99% ± 0.08 |
| 10,000 | 5 | 15.00% | 0.970 | 0.993 | 0.000 | 14.91% | 14.89% ± 0.08 |
| 20,000 | 3 | 8.00% | 1.000 | 1.000 | 0.000 | 7.98% | 7.95% ± 0.06 |
| 20,000 | 4 | 12.00% | 1.000 | 1.000 | 0.000 | 11.98% | 11.97% ± 0.06 |
| 20,000 | 5 | 15.00% | 1.000 | 1.000 | 0.000 | 14.98% | 14.99% ± 0.07 |
At \(n=5000\), the conservative rule sometimes misses a weaker active coordinate; the resulting population reduction can be below the requested 8%–15% optimum. At \(n=10000\), it usually retains the active set. In these simulations, all 200 replications in each \(n=20000\) scenario recover the support. This last observation is finite Monte Carlo evidence, not a guarantee: with zero observed failures in 200 replications, the approximate one-sided 95% upper bound for the failure probability is 1.5%.
The next table compares all methods at \(n=10000\), four active covariates, and population \(R^2=12\%\).
| Method | Exact support | Mean FDP | Mean size | Population gain | Validation gain ± MC SE (pp) | Apparent R² |
|---|---|---|---|---|---|---|
| Adaptive 1-SE + refit | 0.960 | 0.000 | 3.96 | 11.89% | 11.99% ± 0.08 | 11.89% |
| Adaptive CV + refit | 0.675 | 0.105 | 4.75 | 11.93% | 12.03% ± 0.07 | 11.98% |
| Lasso CV + refit | 0.010 | 0.523 | 9.28 | 11.85% | 11.95% ± 0.07 | 12.06% |
| Adaptive rate + refit | 0.140 | 0.263 | 5.64 | 11.90% | 12.00% ± 0.07 | 12.01% |
| Full | 0.000 | 0.800 | 20.00 | 11.82% | 11.92% ± 0.07 | 12.09% |
| Oracle | 1.000 | 0.000 | 4.00 | 11.96% | 12.07% ± 0.07 | 11.95% |
Here the adaptive 1-SE rule recovers the exact set in 96.0% of replications, compared with 67.5% for adaptive minimum-loss CV and 1.0% for ordinary lasso CV. Yet their refitted population variance reductions are all close to 12%. This distinction is central to the two planning questions: selecting unnecessary covariates and losing useful variance reduction are different errors. A method can estimate almost the same variance gain while selecting many more variables.
The rate-based rule illustrates a limitation of asymptotic tuning prescriptions: an asymptotically admissible penalty rate does not by itself give reliable finite-sample selection. Its exact recovery is 14.0% in this scenario. None of these comparisons establishes universal dominance of an adaptive rule or proves that the 1-SE choice satisfies the oracle theorem’s tuning-rate assumptions.
Figure 1. Adaptive 1-SE with selected-set refitting. Dotted lines show population optimal R²; error bars show 95% Monte Carlo intervals for mean validation gain, not patient-level uncertainty.
The refit matters. In the same scenario, adaptive 1-SE’s unrefitted coefficients achieve a mean population reduction of 10.31%, whereas the selected-set refit achieves 11.89%. Ordinary minimum-loss shrinkage can sometimes lower coefficient risk relative to its own refit, so refitting is not universally preferable. Apparent and validation summaries can differ in either direction in a finite simulation; external validation, rather than the sign of one observed difference, is what separates development from evaluation.
Win-ratio scores. Generate independent \(Z_1,\ldots,Z_{20},E\sim N(0,1)\) and a continuous primary endpoint \(Z_Y=\rho\sum_{j=1}^4Z_j+\sqrt{1-4\rho^2}E\). The representations are the standardized single-endpoint log-win-ratio null scores \(h_j=\sqrt3\{2\Phi(Z_j)-1\}\) and \(h_Y=\sqrt3\{2\Phi(Z_Y)-1\}\). For a bivariate normal pair, their covariance is \((6/\pi)\arcsin(\rho/2)\); consequently, choosing \(\rho=2\sin\{(\pi/6)\sqrt{0.12/4}\}\) gives \(R^2=0.12\). This identity follows by writing \(E\{\Phi(Z_j)\Phi(Z_Y)\}\) as a bivariate normal orthant probability with correlation \(\rho/2\).
Each reference CDF is estimated from the development data, or the training fold inside CV. Validation records are transformed with that frozen empirical CDF. Thus actual validation includes CDF-estimation effects. Population gains in this row are limiting-score benchmarks, while validation gains evaluate the fitted empirical-score rule. The experiment validates the uncensored, one-component win-ratio special case of (12); it does not validate every hierarchical censoring rule.
Censored baseline event history. Generate an independent baseline event time \(T\sim\operatorname{Exp}(1)\) and censoring time \(C\sim\operatorname{Exp}(0.5)\). Use the standardized score \(h_1=\{\Delta-\widetilde{T}\}/\sqrt{2/3}\). This has mean zero and variance one, because the known-hazard martingale residual has variance \(E(\Delta)=2/3\). Use independent remaining covariate scores, four active coordinates \((1,7,8,9)\), and an additive standardized \(t_5\) residual scaled to give \(R^2=0.12\). This is an observed-score experiment with known nuisance quantities. It shows how a censored baseline record enters the projection; it does not establish an interpretation in terms of its unobserved latent event time or evaluate estimated-hazard uncertainty.
| Score construction | Method | Exact support | Mean FDP | Population gain | Validation gain ± MC SE (pp) |
|---|---|---|---|---|---|
| Win ratio | Adaptive 1-SE + refit | 1.000 | 0.000 | 11.96% | 11.94% ± 0.06 |
| Win ratio | Adaptive CV + refit | 0.670 | 0.106 | 11.93% | 11.91% ± 0.06 |
| Win ratio | Lasso CV + refit | 0.005 | 0.542 | 11.85% | 11.82% ± 0.06 |
| Win ratio | Full | 0.000 | 0.800 | 11.82% | 11.79% ± 0.06 |
| Win ratio | Oracle | 1.000 | 0.000 | 11.96% | 11.94% ± 0.06 |
| Censored baseline | Adaptive 1-SE + refit | 0.925 | 0.000 | 11.85% | 11.82% ± 0.07 |
| Censored baseline | Adaptive CV + refit | 0.660 | 0.121 | 11.92% | 11.91% ± 0.07 |
| Censored baseline | Lasso CV + refit | 0.020 | 0.535 | 11.85% | 11.83% ± 0.07 |
| Censored baseline | Full | 0.000 | 0.800 | 11.82% | 11.81% ± 0.07 |
| Censored baseline | Oracle | 1.000 | 0.000 | 11.96% | 11.94% ± 0.07 |
These examples support the score-construction principle while delimiting its validation: no current simulation establishes uniform performance over all endpoint types, missing-data mechanisms, or historical-to-trial shifts. The general estimated-score theorem requires its own Gram expansion for each proposed statistical representation.
PATED and its historical-data extension address complementary parts of covariate adjustment. PATED combines a primary marginal estimator with zero-target baseline contrasts and minimizes variance within the resulting linear adjustment class. Its \(R^2\) measures the fraction of the primary estimator’s sampling variance explained by those contrasts. The historical analysis then uses the same statistical representations to identify useful adjustment variables and assess their expected contribution at a common-law reference.
The linear variance-minimization principle is closely related to control variates. Regularized selection of control variates has been studied by Leluc, Portier and Segers [12] and South and colleagues [13]. Adaptive sparsity and quadratic penalization are likewise established ideas [9,10]. The contribution here is their statistical integration with heterogeneous marginal treatment comparisons, the corresponding variance-reduction estimand, and a historical-data construction with explicit conditions for covariance estimation and selection. The projection identity alone is not claimed as a new mathematical principle.
Assessing future efficiency from external data also has direct precedents [6]. For event-time analyses, the historical martingale-residual strategy of Zhou and colleagues [14], currently a preprint, connects prognostic information with log-rank variance reduction. The present construction permits different primary and auxiliary marginal representations within one covariance framework. Its interpretation remains specific to those representations; replacing them may change both the selected set and the achievable gain.
The simulations distinguish the two planning objectives. At large sample sizes and moderate dimension, several methods preserve nearly the same variance reduction while selecting substantially different covariate sets. Penalized coefficients may also achieve less reduction than a refit on the selected variables. The oracle result provides pointwise justification under separated signals and appropriate tuning, but it neither justifies every cross-validation rule nor gives uniform inference near the selection boundary.
The common-law historical reference is useful even though its treatment contrast is zero: estimation variance remains well defined at the null. Translating the resulting gain to a future trial requires comparable measurement, follow-up and population structure. Under treatment alternatives, arm-specific outcome distributions may produce a different optimal coefficient and a different gain. Re-estimating PATED’s coefficient in the trial on a prespecified historical covariate set preserves the interpretation of Theorem 1, provided the trial’s joint limit and covariance assumptions hold.
Joint asymptotic normality suffices for PATED’s first-order variance result. Joint asymptotic linearity supplies the common influence-function construction, while the stronger covariance expansion is essential for the historical oracle theorem and uncertainty in \(R^2\). Keeping these requirements distinct permits heterogeneous variables without implying assumption-free inference. Further work may examine growing candidate dimension, estimated survival representations and population shifts within this same framework.
For a nominal variable with \(K\) categories, a scalar coding would impose an arbitrary ordering. Instead, represent it by \(K-1\) linearly independent centered indicators. The same principle permits several prespecified thresholds or features for a single variable. A nominal response can likewise be represented by a vector rather than an arbitrary numerical label.
Let \(H_j\in\mathbb R^{d_j}\) represent predictor variable \(j\) and \(V\in\mathbb R^q\) represent the response. Remove exact linear redundancies. For invariance to nonsingular within-variable recodings, whiten each marginal block and the response using their positive-definite covariance matrices. In the resulting coordinates, stack the predictor scores as \(H\) and consider
\[ C_0=\arg\min_C\frac12 E\|V-C^\top H\|_2^2, \tag{49} \]
with adaptive variable-block penalty
\[ \lambda_n\sum_j (\|\widetilde{C}_j\|_F\vee\epsilon_n)^{-\gamma}\|C_j\|_F. \tag{50} \]
Here \(C_j\) contains all predictor-feature rows for variable \(j\) and all response columns. Variable \(j\) is selected when \(C_j\ne0\).
Proposition 4 (fixed-block oracle extension). With a fixed number of blocks and fixed block sizes, positive definiteness, nonzero active block norms bounded away from zero, a joint root-\(n\) expansion for all estimated moments including whitening, a root-\(n\) pilot, and (26), the estimator has consistent block support and the same first-order distribution on active blocks as the known-active-block unpenalized estimator.
Proof. Vectorizing \(C\) gives a quadratic problem with Hessian \(I_q\otimes G_{HH}\), up to a coordinate permutation. On a nonzero block, the Frobenius norm is differentiable and its gradient has norm one, so the active penalty is \(o_p(n^{-1/2})\). For a zero block, the pilot norm is \(O_p(n^{-1/2})\), while its quadratic-loss gradient is \(O_p(n^{-1/2})\). The inactive group optimality condition compares that gradient’s norm with \(\lambda_nw_j\), which dominates by (26). Restricted consistency, the optimality argument, and the central limit theorem proceed as in Theorem 2. \(\square\)
After whitening, invertible recodings within a variable’s feature span induce orthogonal changes of coordinates. Both the squared loss and the Frobenius block penalties are invariant to those changes, provided the pilot and floor conventions are transformed coherently. Thus selection depends on the variable’s represented subspace rather than on an arbitrary basis. Adding genuinely new features changes that subspace and appropriately changes the estimand.
With \(\operatorname{Var}(V)=I_q\), a normalized variable contribution is
\[ D_j^{\mathrm{block}} =\frac1q\operatorname{tr} \{C_{0j}^\top G_{j\mid-j}C_{0j}\}, \tag{51} \]
the reduction in total residual score variance when the block is added to all other blocks, divided by the total response score variance. It is zero exactly when the block coefficient is zero under positive definiteness.
U.S. Food and Drug Administration (2023). Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products: Guidance for Industry. May 2023. Official guidance PDF.
Ye, T., Shao, J., Yi, Y., and Zhao, Q. (2023). Toward better practice of covariate adjustment in analyzing randomized clinical trials. Journal of the American Statistical Association, 118(544), 2370–2382. DOI. Author manuscript.
Reluga, K., Ye, T., and Zhao, Q. (2024). A unified analysis of regression adjustment in randomized experiments. Electronic Journal of Statistics, 18(1), 1436–1454. DOI. Author manuscript.
Bannick, M. S., Shao, J., Liu, J., Du, Y., Yi, Y., and Ye, T. (2025). A general form of covariate adjustment in clinical trials under covariate-adaptive randomization. Biometrika, 112(3), asaf029. DOI. Author manuscript, with earlier title.
Ye, T., Shao, J., and Yi, Y. (2024). Covariate-adjusted log-rank test: guaranteed efficiency gain and universal applicability. Biometrika, 111(2), 691–705; published online in 2023. DOI. Author manuscript.
Li, X., Li, S., and Luedtke, A. (2023). Estimating the efficiency gain of covariate-adjusted analyses in future clinical trials using external data. Journal of the Royal Statistical Society, Series B, 85(2), 356–377. DOI. Author manuscript.
Nagelkerke, N. J. D. (1991). A note on a general definition of the coefficient of determination. Biometrika, 78(3), 691–692. DOI.
Schemper, M., and Henderson, R. (2000). Predictive accuracy and explained variation in Cox regression. Biometrics, 56(1), 249–255. DOI.
Zou, H. (2006). The adaptive lasso and its oracle properties. Journal of the American Statistical Association, 101(476), 1418–1429. DOI. Primary paper PDF.
Wang, H., and Leng, C. (2007). Unified LASSO estimation by least squares approximation. Journal of the American Statistical Association, 102(479), 1039–1048. DOI and journal abstract.
Pötscher, B. M., and Schneider, U. (2009). On the distribution of the adaptive LASSO estimator. Journal of Statistical Planning and Inference, 139(8), 2775–2790. Author manuscript.
Leluc, R., Portier, F., and Segers, J. (2021). Control variate selection for Monte Carlo integration. Statistics and Computing, 31, 50. DOI. Author manuscript.
South, L. F., Oates, C. J., Mira, A., and Drovandi, C. (2023). Regularized zero-variance control variates. Bayesian Analysis, 18(3), 865–888. Author manuscript.
Zhou, J., Liu, Q., Mo, M., and Xia, A. (2026). Improving power in randomized controlled trials with time-to-event endpoints: a risk-free approach. arXiv preprint, 2605.27711, version 1, 26 May 2026. Manuscript.