we-solve (beta)

Econ-198 · economics · Economics

High-dimensional causal inference

LLM status Open unverified

Statement

Original identifier: E4 (atlas item 104). Field: Econometrics and causal inference. Classification: editorial research family with established restricted solutions; current open status not certified.

Definitions and assumptions

Let independent blocks $B_1,\ldots,B_m$ contain possibly dependent observations $(Y_i,D_i,X_i)$, with $X_i\in\mathbb R^{p_m}$ and $p_m$ potentially larger than $m$. Specify block-size growth and weighting before estimation. In a partially linear model $Y=\theta D+g(X)+\epsilon$ assume $\mathbb E[\epsilon\mid D,X]=0$, and let $l(X)=\mathbb E[Y\mid X]$, $q(X)=\mathbb E[D\mid X]$. Require $\mathbb E[(D-q(X))^2]>0$. The score \[ \psi(W;\theta,l,q)=(D-q(X))\{Y-l(X)-\theta(D-q(X))\}, \qquad W=(Y,D,X), \] has zero expectation at the true parameter. A causal interpretation requires structural invariance, for example $Y(d)=\theta d+g(X)+\epsilon$, and suitable assignment restrictions.

Formal research question

Characterize model classes permitting uniform inference for $\theta$ with block-level cross-fitting, growing covariate dimension and distribution drift. A baseline sufficient rate condition for independent effective observations is \[ \|\widehat q-q\|_2 \left(\|\widehat l-l\|_2+\|\widehat q-q\|_2\right) =o_P(m^{-1/2}), \] along with suitable moments, nondegenerate score variance and consistent variance estimation; $\|f\|_2=(\mathbb E[f(X)^2])^{1/2}$. Determine appropriate replacements when block dependence reduces effective information or nuisance functions differ by environment. Seek confidence sets $C_m$ satisfying uniform $1-\alpha$ coverage, for $\alpha\in(0,1)$, over a stated approximately sparse or otherwise learnable class, and derive lower bounds when nuisance complexity, weak residual variation or drift prevents the required rates.

Interpretation and scope

Folds must keep dependent blocks intact; splitting individual observations across folds need not make training and evaluation independent. The number of predictors alone is not the difficulty: unrestricted confounding cannot be learned from finite data without complexity restrictions. Approximate sparsity needs a quantified approximation error; prediction accuracy and causal identification are separate requirements. The partially linear coefficient is a constant causal effect only under its stated structural model. If treatment effects are heterogeneous, specify another estimand and an appropriate orthogonal score instead of relabeling this slope as an ATE.

Source audit and status

[S1] establishes post-selection inference under approximate sparsity; [S2] gives double/debiased ML theory under explicit regularity conditions. The undated Athey--Imbens--Wager citation is completed with the verified 2018 approximate-residual-balancing article [S3]; intended-source certainty is limited because the input supplied no title. Standard independent-data high-dimensional inference has established solutions. The dependent and drifting regime above is a scoped editorial frontier, not a published conjecture certified open today.

References

    },leftmargin=*]

  1. Alexandre Belloni, Victor Chernozhukov and Christian Hansen (2014). Inference on Treatment Effects after Selection among High-Dimensional Controls. Review of Economic Studies 81(2), 608–650. https://doi.org/10.1093/restud/rdt044. Locator: Publisher or institutional abstract and bibliographic record. Support and limits: Post-double-selection inference under approximate sparsity and regularity; not validity for arbitrary high-dimensional confounding.

  2. Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey and James Robins (2018). Double/debiased machine learning for treatment and structural parameters. Econometrics Journal 21(1), C1–C68. https://doi.org/10.1111/ectj.12097. Locator: Publisher or institutional abstract and bibliographic record. Support and limits: Orthogonal scores and cross-fitting under nuisance-rate and sampling restrictions; not automatic protection against dependence or misspecification.

  3. Susan Athey, Guido W. Imbens and Stefan Wager (2018). Approximate Residual Balancing: Debiased Inference of Average Treatment Effects in High Dimensions. Journal of the Royal Statistical Society Series B 80(4), 597–623. https://doi.org/10.1111/rssb.12268. Locator: Publisher or institutional abstract and bibliographic record. Support and limits: Verified three-author candidate; the input omitted year and title. Selected explicitly as a relevant completion, not proof of intended citation.

Common conventions: Source mathematical conventions

Unless an entry states otherwise, agent and firm sets are finite. State and action spaces are measurable, strategies and outcome maps are measurable, probability laws are specified, and expectations exist for the targets under discussion. Infinite-horizon discount factors lie in $(0,1)$ and discounted objectives are finite. Norms, tolerances and complexity input sizes must be specified for computational comparisons. Constants or primitives that appear locally are inputs, not empirical facts asserted by this catalogue.

Identification. A statistical model $m\in\mathcal M$ implies an observable law $P_m$ and target $\theta(m)$. At law $P$, its identified set is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M,\ P_m=P\}. \] Point identification holds when this set is a singleton, or equivalently when $P_{m_1}=P_{m_2}$ implies $\theta(m_1)=\theta(m_2)$ for all compatible models. Sharp bounds mean bounds attained, or approached when the boundary is not attained, by compatible models. Writing this set is a definition, not a solution: the substantive task is to establish informative restrictions and derive or estimate the set. An unrestricted-confounding model may give only trivial bounds.

Inference. Identification, estimability and valid uncertainty quantification are separate questions. Fix $\alpha\in(0,1)$. A confidence procedure $C_n$ over a declared law class $\mathcal P$ has asymptotic uniform coverage at level $1-\alpha$ when \[ \liminf_{n\to\infty}\inf_{P\in\mathcal P}\Pr_P\{\theta(P)\in C_n\}\geq1-\alpha. \] For set-identified targets the coverage event must be defined separately, for example coverage of the full identified set. If an entry uses identification-indexed models instead of uniquely indexed laws, the same coverage convention is applied over those models. Pointwise limits do not establish uniform validity, and asymptotic validity does not guarantee finite-sample precision. Sample sizes, dependence structures and weak-identification sequences are part of each application.

Counterfactuals. $Y_i(a)$ is a potential outcome under a specified intervention, including its timing, implementation and versions. It is not $Y_i$ conditional on voluntarily choosing $a$. A full intervention vector permits interference; shorthand scalar treatment notation does not silently impose no interference. Assignment, selection, attrition, anticipation and measurement must be specified. A claim about an instrument's local effect is not automatically a population policy effect.

Economic equilibrium. A counterfactual model includes best responses, constraints, feasibility, market clearing, state transitions and expectations consistent with the stated information. When several equilibria exist, either a declared selector is an input or the target is set-valued. Comparisons across policy regimes must use the stated selection convention. Reduced-form relationships are not presumed invariant after an equilibrium-changing intervention.

Optimization and welfare. Optimization is over the stated feasible set. If attainment is not established, $\sup$ or $\inf$ and $\varepsilon$-optimal choices replace an asserted maximizer or minimizer. For example, with value $V=\sup_{a\in\mathcal A}W(a)<\infty$, an $\varepsilon$-optimal set is $\{a\in\mathcal A:W(a)\geq V-\varepsilon\}$ for $\varepsilon>0$. Benefits, costs, risk and health or environmental outcomes can be aggregated only using declared units and conversion weights. Interpersonal utility weights, the treatment of fiscal transfers and foreign profits, discounting, and the behavioral welfare criterion are explicit inputs. A comparative-static derivative additionally requires existence and appropriate differentiability of the selected counterfactual.

Sources. Local labels [S1], [S2], and so forth restart in every question. Locators identify the relevant abstract, model, section or result at the level actually checked; a bibliographic or abstract-level verification is not represented as inspection of an entire proof. Working papers are distinguished from later publications and changing versions. Source summaries are paraphrased. All URL checks and editorial formulations refer to the audit date stated above.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-198-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-198

Orthogonal inference with independent blocks Weighted information, environment drift, and a weak-signal lower bound

Problem ID: C21-194. Status: Unresolved; rigorous restricted results.

Scope and statistical model

The target concerns high-dimensional nuisance learning with dependence and changing environments. We give a uniform inference theorem for independent blocks with arbitrary dependence inside each block, explicit deterministic weighting, and environment-specific nuisance functions. The theorem specifies prediction conditions rather than asserting they hold for every high-dimensional learner. We also give drift and weak-information counterexamples. It leaves the optimal learnability and lower-bound frontier unresolved.

For each $m$, blocks $B_b=(O_{bi}:1\leq i\leq N_b)$, $1\leq b\leq m$, are independent. Their sizes may grow without restriction. Use deterministic within-block weights $a_{bi}\geq0$, $\sum_i a_{bi}=1$, and outer weights $w_b>0$, $\sum_b w_b=1$. Each marginal observation satisfies \[ Y_{bi}=\theta D_{bi}+g_b(X_{bi})+\epsilon_{bi},\qquad {\mathbb{E}}[\epsilon_{bi}\mid D_{bi},X_{bi}]=0. \] Within each block the nuisance functions $q_b(x)={\mathbb{E}}[D_{bi}\mid X_{bi}=x]$ and $l_b(x)={\mathbb{E}}[Y_{bi}\mid X_{bi}=x]$ are common across its coordinates; the marginal covariate distributions may differ. The coefficient $\theta$ is common across blocks. Assume $|Y|,|D|,|\theta|,|q_b|,|l_b|\leq K$ for a fixed $K$. A causal interpretation additionally requires the structural invariance in the catalogue.

Write $v_{bi}=D_{bi}-q_b(X_{bi})$, and define the ideal block quantities \[ S_b=\sum_i a_{bi}v_{bi}\epsilon_{bi},\qquad T_b=\sum_i a_{bi}v_{bi}^2,\quad J_m=\sum_b w_b{\mathbb{E}} T_b,\quad \sigma_m^2=\sum_b w_b^2{\mathbb{E}} S_b^2. \] Because ${\mathbb{E}} S_b=0$, $\sigma_m^2$ is the variance of the weighted ideal score. It incorporates all within-block dependence.

Independent-data orthogonality is developed in [dml, Section 4.1]; dependence-aware cross-fitting has established extensions, including [multiway, Theorems 1--2]. Our proof addresses the simpler independent-block sampling geometry directly.

Prediction conditions that preserve block independence

Partition the block indices into a fixed number $F$ of folds. For $b$ in fold $f$, fit bounded functions $\widehat q_b,\widehat l_b$ using only blocks outside $f$. An environment label may index these predictions; the assumption does not make unobserved environments automatically learnable. Let $\mathcal A_f$ denote the training data for fold $f$. Define conditional prediction errors \[ e_{qb}^2=\sum_i a_{bi}{\mathbb{E}}[(\widehat q_b-q_b)^2(X_{bi})\mid\mathcal A_f], \qquad e_{lb}^2=\sum_i a_{bi}{\mathbb{E}}[(\widehat l_b-l_b)^2(X_{bi})\mid\mathcal A_f]. \] The expectations integrate a fresh marginal from block $b$. In a sequence of law classes assume, uniformly in the law, \begin{align} &J_m\geq j_*>0,\qquad c_*\sum_b w_b^2\leq\sigma_m^2\leq C_*\sum_b w_b^2, \qquad \frac{\max_b w_b}{(\sum_b w_b^2)^{1/2}}\longrightarrow0, \label{wsref-0-69-6e-66-6f}\\ &\max_b(e_{qb}+e_{lb})=o_P(1),\qquad \sum_b w_b e_{qb}(e_{lb}+e_{qb})=o_P(\sigma_m). \label{wsref-0-72-61-74-65-73} \end{align} Constants do not depend on $m$ or the law. Uniform $o_P$ means the supremum of the associated failure probabilities tends to zero. The variance lower bound is substantive: the theorem does not pretend that every choice of block averaging has this amount of information.

Put $\widehat v_{bi}=D_{bi}-\widehat q_b(X_{bi})$ and \[ \widehat J=\sum_b w_b\sum_i a_{bi}\widehat v_{bi}^2, \qquad \widehat\theta=\widehat J^{-1} \sum_b w_b\sum_i a_{bi}\widehat v_{bi}(Y_{bi}-\widehat l_b(X_{bi})). \] Use an arbitrary bounded value if $\widehat J<j_*/2$; that event will have probability tending uniformly to zero. Define \[ \widehat S_b=\sum_i a_{bi}\widehat v_{bi} \{Y_{bi}-\widehat l_b(X_{bi})-\widehat\theta\widehat v_{bi}\}, \qquad \widehat\sigma_m^2=\sum_b w_b^2\widehat S_b^2. \]

Theorem (Uniform block inference): Under (info)--(rates) and the boundedness assumptions, \[ \frac{J_m(\widehat\theta-\theta)}{\sigma_m} =\frac{\sum_b w_bS_b}{\sigma_m}+o_P(1) \ \Longrightarrow\ N(0,1) \] uniformly over the declared law classes. Moreover, $\widehat J-J_m=o_P(1)$ and $\widehat\sigma_m^2/\sigma_m^2\to_P1$, uniformly. Consequently \[ \left[\widehat\theta\ \pm\ z_{1-\alpha/2} \frac{\widehat\sigma_m}{\widehat J}\right] \] has asymptotic uniform coverage $1-\alpha$ for every fixed $0<\alpha<1$.
Proof: All constants below depend only on the displayed fixed bounds. Set $\delta_{qb}=\widehat q_b-q_b$ and $\delta_{lb}=\widehat l_b-l_b$. Since $Y-l_b=\theta v+\epsilon$, the fitted score at the true coefficient expands, observation by observation, as \begin{align} &(v-\delta_q)\{Y-l_b-\delta_l-\theta(v-\delta_q)\}-v\epsilon\nonumber\\ &\hspace{10mm}=-\delta_q\epsilon-v\delta_l+\theta v\delta_q +\delta_q\delta_l-\theta\delta_q^2.\label{wsref-0-65-78-70-61-6e-73-69-6f-6e} \end{align} Conditional on a fold's training data, its evaluation blocks are independent. The first three terms have conditional marginal expectation zero, using ${\mathbb{E}}[v\mid X]=0$ and ${\mathbb{E}}[\epsilon\mid D,X]=0$. Thus the absolute conditional mean of the block score error is at most $e_{qb}e_{lb}+K e_{qb}^2$. Its conditional second moment is at most $C(e_{qb}^2+e_{lb}^2)$: boundedness first gives this bound for each summand, and Jensen's inequality for the $a_{bi}$ gives it for the block average. No independence inside the block is used.

For a given fold, the variance of its centered weighted score error is therefore at most \[ C\sum_{b\text{ in fold}}w_b^2(e_{qb}^2+e_{lb}^2) =o_P(\sigma_m^2). \] Conditional Chebyshev, first on training events where the errors are small and then off those events, makes the centered error $o_P(\sigma_m)$. The number of folds is fixed, so the sum over folds remains $o_P(\sigma_m)$ even though the fitted scores across folds are dependent. The total conditional-mean bounds are $o_P(\sigma_m)$ by (rates). We have proved \[ \sum_b w_b\widehat S_b(\theta)=\sum_b w_bS_b+o_P(\sigma_m). \tag{4} \]

For the denominator, the ideal weighted average $\sum_b w_bT_b$ has variance at most $C\sum_b w_b^2=o(1)$. The last convergence follows from (info): writing $s=(\sum w_b^2)^{1/2}$, one has $s^2\leq\max w_b=o(s)$ and hence $s\to0$. Also $|(v-\delta_q)^2-v^2|\leq C|\delta_q|$ by boundedness. Its weighted expected magnitude, conditional fold by fold, is at most $C\max_b e_{qb}=o_P(1)$. Conditional Markov gives $\widehat J-J_m=o_P(1)$. In particular the denominator exception is negligible. Since the score is affine in $\theta$, (4) gives \[ \widehat\theta-\theta=\widehat J^{-1} \{\sum_b w_bS_b+o_P(\sigma_m)\}=O_P(\sigma_m). \] The bounded independent variables $w_bS_b$ satisfy the Berry--Esseen bound \[ C\frac{\sum_b w_b^3}{\sigma_m^3} \leq C'\frac{\max_b w_b}{(\sum_b w_b^2)^{1/2}}\longrightarrow0. \] This proves the uniform normal approximation and the linear representation.

It remains to verify variance estimation. Independence and boundedness give \[ {\operatorname{Var}}\left(\sum_b w_b^2S_b^2\right)\leq C\sum_b w_b^4 =o(\sigma_m^4), \] since $\sum w_b^4\leq(\max w_b)^2\sum w_b^2$ and (info) holds. Its expectation is $\sigma_m^2$. The conditional second-moment bound already proved, now weighted by $w_b^2$, yields \[ \sum_b w_b^2\{\widehat S_b(\theta)-S_b\}^2=o_P(\sigma_m^2) \] by conditional Markov and a finite fold sum. Finally $|\widehat S_b-\widehat S_b(\theta)|\leq C|\widehat\theta-\theta|$, so its weighted square sum is $O_P(\sigma_m^2)\sum_b w_b^2=o_P(\sigma_m^2)$. Cauchy--Schwarz applied to the difference of the two sums of squares proves $\widehat\sigma_m^2/\sigma_m^2\to_P1$. Combining these uniform probability bounds with the uniform normal approximation proves coverage. ∎

For equal block weights, the theorem requires $\sum_bm^{-1}e_{qb}(e_{lb}+e_{qb})=o_P(m^{-1/2})$ and yields width of order $m^{-1/2}$ when $J_m$ is fixed away from zero. More generally the effective number for this variance-comparable weighting scheme is $(\sum w_b^2)^{-1}$. This is a property of the chosen scores and weights, not a universal formula based solely on the total number of rows. If a prediction theorem supplies errors $e_q\lesssim\sqrt{s_q\log p/m}+a_q$ and $e_l\lesssim\sqrt{s_l\log p/m}+a_l$, the exact sufficient restriction is obtained by inserting those expressions into (rates); the approximation errors $a_q,a_l$ cannot be omitted.

Drift and weak variation cannot be hidden in prediction notation

Proposition (Pooling distinct environments can create a false slope): There is a bounded independent-observation two-environment model with true structural slope zero such that using the pooled nuisance functions in the orthogonal score converges to a nonzero coefficient.
Proof: Take an even number of observations, half in each supplied environment $s_b\in\{-1,1\}$. Let $v_b,\epsilon_b$ be independent centered bounded random variables, independent across $b$, with ${\mathbb{E}} v_b^2=t^2>0$. Set \[ D_b=s_ba+v_b,\qquad Y_b=s_ba+\epsilon_b, \] where $a>0$. With the environment included among the controls, $\theta=0$, $q_b=l_b=s_ba$, and the conditional mean restriction holds. Omitting the environment gives pooled functions $q=l=0$. The resulting sample slope has numerator tending to $a^2$ and denominator tending to $a^2+t^2$, by bounded independent-variable variance bounds. Its limit is $a^2/(a^2+t^2)>0$. Equivalently, the last two terms of (expansion) have nonvanishing mean. Orthogonality is local to the correct nuisance functions and does not excuse this drift misspecification. ∎
Proposition (Perfect replication and weak residual variation): Consider $m$ independent blocks, each containing any number of exact copies of one pair \[ D_b\in\{-\sqrt\kappa,\sqrt\kappa\}\text{ equiprobably},\qquad Y_b=\theta D_b+\epsilon_b,\quad\epsilon_b\sim N(0,1), \] with $0<\kappa\leq1$ and $|\theta|\leq T$ for fixed $T>0$. All signs and errors are independent. For a universal $c>0$, \[ \inf_{\widehat\theta}\sup_{|\theta|\leq T} {\mathbb{E}}(\widehat\theta-\theta)^2 \geq c\min\{T^2,(m\kappa)^{-1}\}. \] Thus unbounded replication inside a block does not create more information, and a vanishing $\kappa$ can defeat consistency.
Proof: The copies generate the same sigma field as the $m$ distinct pairs. For two coefficients differing by $\Delta$, the Gaussian conditional likelihood gives ${\operatorname{KL}}(P_0,P_1)=m\kappa\Delta^2/2$. Choose two coefficients in $[-T,T]$ at separation $\Delta=c_0\min\{T,(m\kappa)^{-1/2}\}$ with a sufficiently small fixed $c_0$. Pinsker's inequality makes their total variation at most $1/2$. Classify any estimator by the nearer coefficient. On a classification error its squared loss is at least $\Delta^2/4$, and the sum of the two testing error probabilities is at least $1-{\operatorname{TV}}\geq1/2$. The larger risk is therefore at least $\Delta^2/16$, proving the result. This lower-bound submodel uses Gaussian, rather than bounded, outcomes; it belongs to the catalogue's moment-based model, not the bounded sufficient class of Theorem Theorem (Uniform block inference). ∎

Remaining frontier

The proof isolates the correct objects: block score variance, residual treatment information, environment-specific prediction errors, and intact evaluation blocks. It does not determine how many distinct drifting nuisance functions can be learned, optimal weighting for arbitrary internal covariance, or the minimax nuisance threshold near weak information. Those tasks require a declared complexity and drift class. In particular, counting predictors or quoting independent-row rates cannot establish (rates) in a dependent experiment.

References

  • [dml] V. Chernozhukov et al. (2018), Double/debiased machine learning for treatment and structural parameters, Econometrics Journal 21, C1--C68. Author-posted version arXiv:1608.00060v7, Section 4.1, Assumption 4.1 and Theorem 4.1 inspected 9 October 2026. https://arxiv.org/pdf/1608.00060.
  • [multiway] H. D. Chiang, K. Kato, Y. Ma and Y. Sasaki, Multiway Cluster Robust Double/Debiased Machine Learning, arXiv:1909.03489. Theorems 1--2 and the abstract inspected 9 October 2026. Its multiway array design is different from the independent-block theorem proved here. https://arxiv.org/pdf/1909.03489.

Completion Estimate

45%

Go to contribution →