we-solve (beta)

Econ-164 · economics · Economics

The joint estimation and inference frontier for causal response functions

LLM status Open unverified

Statement

CI-29 |  Econometric theory and causal inference

Research agenda formalization

Definitions and assumptions

Invoke the binary causal baseline on $[0,1]^d$, with bounded outcomes, a covariate density bounded above and below, and fixed overlap. Let $\mathcal P$ impose $\tau\in\mathcal H_d^s(L)$, $e\in\mathcal H_d^\alpha(L)$, $m_w\in\mathcal H_d^\beta(L)$, and conditional outcome variances bounded away from zero. A procedure consists of an estimator $\widehat\tau_n$ and a nonnegative radius function $a_n$, producing a band centered at that same estimator. Define its estimation risk and inference width by \[ r_n=\sup_{P\in\mathcal P}\E_P\|\widehat\tau_n-\tau_P\|_{L^2(P_X)}^2, \qquad w_n=\sup_{P\in\mathcal P}\E_P\|a_n\|_\infty. \] Require the band $\widehat\tau_n\pm a_n$ to cover $\tau_P$ simultaneously with asymptotic probability at least $1-\eta$, uniformly over $\mathcal P$.

Formal question

Characterize the attainable pairs $(r_n,w_n)$, up to specified constants or logarithmic factors. Can a single band center attain the minimax integrated estimation rate while its honest band attains the optimal uniform-width rate? When a construction uses a smoothing scale $h_n$, estimated standard errors, and a Gaussian critical value, derive conditions ensuring that residual smoothing bias and causal nuisance remainder are covered or negligible uniformly in $x$. Determine the unavoidable efficiency or width penalty of undersmoothing, explicit bias bounds, or additional bias correction.

Scope and status

This frontier is an editorial formulation of the causal "bias problem" discussed in the source. An estimator selected to minimize mean squared error need not have negligible standardized bias. Conversely, that observation is not a universal impossibility theorem for joint estimation and inference: an honest band may explicitly cover bias. The question asks for a precise class-dependent tradeoff, and does not presume that nominal Gaussian bands or a particular debiasing recipe are valid.

References for the source of the problem

Carlos Cinelli, Avi Feller, Guido Imbens, Edward Kennedy, Sara Magliacane, and Jose Zubizarreta, Challenges in Statistics: A Dozen Challenges in Causality and Causal Inference, arXiv:2508.17099v1 [stat.ME], 23 August 2025, §3.3.3, "Inference," including its discussion of smoothing bias and nuisance estimation. https://arxiv.org/html/2508.17099v1#S3.SS3.SSS3.

Common conventions: Source mathematical conventions

Well-defined objects. All probability spaces and measurable variables are specified or inherited from a local statistical model. Expectations used as finite targets require the stated integrability. A decision rule or estimator is measurable with respect to its available information; randomized rules may use an auxiliary independent random variable. Norms without a subscript are Euclidean in finite-dimensional spaces. An optimizer is requested only when attainment is assured; otherwise the question concerns the optimal value and $\varepsilon$-optimal admissible rules for every $\varepsilon>0$.

Identification. A structural model $m\in\mathcal M$ implies an observable law $P_m^{\rm obs}$ and target $\theta(m)$. The target is point identified on $\mathcal M$ if \[ P_{m_1}^{\rm obs}=P_{m_2}^{\rm obs} \quad\Longrightarrow\quad\theta(m_1)=\theta(m_2) \qquad(m_1,m_2\in\mathcal M). \] When this fails, the sharp identified set is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M, P_m^{\rm obs}=P\}. \] Specifying an intervention or estimand does not establish identification. Positivity, exclusion, causal sufficiency, measurement restrictions, and transport assumptions are substantive local inputs, never automatic consequences of notation.

Minimax risk and nontrivial inference. For a statistical experiment with data law $P^{(n)}$, target $\theta(P)$, and nonnegative loss $\ell$, define \[ \mathcal R_n(\mathcal P,\ell) =\inf_{\widehat\theta_n}\sup_{P\in\mathcal P} \E_{P^{(n)}}\ell(\widehat\theta_n,\theta(P)). \] The infimum is over measurable estimators. A sharp rate is a sequence $r_n$ for which $0<c\leq C<\infty$, independent of $n$, satisfies $c r_n\leq\mathcal R_n\leq C r_n$ for all sufficiently large $n$. Squared-error parametric risk is of order $n^{-1}$; the corresponding root-mean-squared error is of order $n^{-1/2}$. These different conventions are distinguished locally. A uniform asymptotic confidence region $C_n$ must satisfy \[ \liminf_{n\to\infty}\inf_{P\in\mathcal P} \Pr_{P^{(n)}}\{\theta(P)\in C_n\}\geq 1-\alpha, \qquad 0<\alpha<1, \] together with an informative diameter or expected-width requirement. Coverage alone permits the vacuous whole parameter space. Testing questions similarly impose both size control and power against a declared separated alternative class.

Binary causal baseline. When an entry explicitly invokes this baseline, one observes independent copies of $O=(X,W,Y)$, with $W\in\{0,1\}$ and potential outcomes $Y(0),Y(1)$. Assume consistency $Y=Y(W)$, no interference, conditional exchangeability $(Y(0),Y(1))\perp W\mid X$, and overlap $\epsilon\leq e(X)\leq1-\epsilon$ almost surely for a fixed $0<\epsilon<1/2$, where \[ e(x)=\Pr(W=1\mid X=x),\qquad m_w(x)=\E[Y\mid W=w,X=x],\qquad \tau(x)=m_1(x)-m_0(x). \] These assumptions identify conditional mean effects almost everywhere under the covariate law. Pointwise or uniform effects require appropriate regularity and a specified support. Continuous treatment, longitudinal data, adaptive experiments, and network interference require their own assumptions in place of this baseline. Bounded outcomes, densities, and smoothness are additional assumptions only where stated.

Function classes. On $[0,1]^d$, $\mathcal H_d^s(L)$ is the isotropic Hölder ball of radius $L>0$: put $k=\lceil s\rceil-1$ and $\gamma=s-k\in(0,1]$, bound derivatives through order $k$, and require the order-$k$ derivatives to be $\gamma$-Hölder with constant at most $L$. Thus integer orders use a Lipschitz final derivative convention. The class $\operatorname{BV}_d(L)$ consists of functions $f\in L^1((0,1)^d)$ whose distributional derivative is a finite vector measure and for which $\|f\|_{L^1}+|Df|((0,1)^d)\leq L$. An $L^\infty$ bound or a particular pointwise version is imposed separately when needed. Sparse and neural-network classes require a declared dictionary or architecture and coefficient bounds.

An anisotropic Hölder class with declared block smoothness indices means coordinatewise Hölder regularity in each block, uniformly in the other blocks, with all corresponding derivative and Hölder bounds at most the stated radius. Derivative orders and the integer-order convention in each block are those just given. Mixed-derivative bounds are additional restrictions only when explicitly stated.

Decision and computational questions. Policy regret compares admissible feasible policies under the same law and constraints. In computational entries, finite data and thresholds have rational encodings; running time is measured in their total bit length. Real-valued equilibrium search uses the explicitly declared algebraic or FIXP encoding. An approximation ratio for maximization is written $\operatorname{OPT}/\operatorname{ALG}$ and for minimization as $\operatorname{ALG}/\operatorname{OPT}$, with zero optima and infeasible instances handled locally. Historical approximation bounds are not automatically current bounds.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-164-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-164

One center for estimation and simultaneous inference A projection construction and a sharp smoothness-diagonal benchmark

Problem ID: C21-160. Status: Unresolved; rigorous restricted results.

Question and main distinction

The target asks whether the same center can have minimax integrated risk and an honest band of optimal uniform width. A pointwise Gaussian interval with unaccounted smoothing bias can fail, but that does not prove a universal incompatibility. We give a general projection construction and instantiate it on $s=\beta=t\in(0,1]$. It attains the optimal integrated estimation rate and the optimal simultaneous-band width order at one center, with explicit bias allowance. The general causal nuisance frontier is still unresolved.

The source's bias discussion appears in [challenge, Section 3.3.3]. Our construction combines two smoothing scales and does not assert that an unmodified MSE-optimal histogram itself has the best maximum-error behavior.

A general center-transfer theorem

Let the target $f$ take values in $[-M,M]$ on a measurable domain with probability measure $\mu$. Suppose an estimator $T$ also takes values there. Let $[L(x),U(x)]$ be a nonempty random interval contained in $[-M,M]$ for every $x$, and suppose the simultaneous coverage event $\mathcal E=\{L(x)\leq f(x)\leq U(x)\text{ for every }x\}$ is measurable. All procedures used below are finite histograms, so this measurability is automatic. Define pointwise \[ C(x)=\min\{U(x),\max\{L(x),T(x)\}\},\quad A(x)=\max\{C(x)-L(x),U(x)-C(x)\}. \]
Theorem (Projection preserves risk at a quantifiable cost): If uniformly over a class ${\mathbb{E}}\|T-f\|_{L^2(\mu)}^2\leq r_n$ and $\Pr(\mathcal E^c)\leq\delta_n$, then the band $C\pm A$ has simultaneous coverage at least $1-\delta_n$ and \[ \sup_P{\mathbb{E}}\|C-f\|_{L^2(\mu)}^2\leq r_n+4M^2\delta_n, \qquad \sup_P{\mathbb{E}}\|A\|_\infty\leq\sup_P{\mathbb{E}}\|U-L\|_\infty. \] No independence between $T$ and $[L,U]$ is needed.
Proof: The center $C$ belongs to $[L,U]$, so $[C-A,C+A]$ contains that interval and $A\leq U-L$. On $\mathcal E$, projection onto an interval containing $f(x)$ cannot increase distance to $f(x)$. Explicitly, if $T<L$ then $T\leq L\leq f$; if $T>U$ then $f\leq U\leq T$; inside the interval $C=T$. These cases prove $|C-f|\leq|T-f|$ pointwise on $\mathcal E$. Off that event both functions are in $[-M,M]$, so their integrated squared distance is at most $4M^2$. Taking expectations gives the first bound, and the pointwise radius inequality gives the second. Coverage follows by containment. ∎

A preliminary band with merely fixed failure probability would contribute a constant to this risk argument. The useful choice is a high-coverage band, with $\delta_n$ smaller than the desired integrated risk. Whether that choice has a width penalty is a property of the statistical experiment, rather than an algebraic fact about projection.

An explicit causal construction

Use the binary causal baseline, $|Y|\leq1$, covariate density bounded between fixed positive constants, and fixed overlap $\varepsilon$. Suppose $s=\beta=t\in(0,1]$, both $m_w$ are $t$-Hölder with constant $L$, and $\tau$ satisfies the catalogue's compatible $t$-Hölder bound. The propensity may be unknown and arbitrarily rough within its prescribed overlap and smoothness restrictions.

For a grid with $J=q^d$ cells of side $h=1/q$, estimate each $m_w$ by its cell-arm outcome average, or zero for an empty cell-arm. Let $T_h$ be the difference of these estimates, so $T_h\in[-2,2]$.

Lemma (Integrated risk of the armwise histogram): Uniformly over this class, \[ {\mathbb{E}}\|T_h-\tau\|_{L^2(P_X)}^2 \leq C\{h^{2t}+J/n\}. \]
Proof: For a cell-arm with probability $p$, let $N\sim\operatorname{Bin}(n,p)$. Its conditional mean $\bar m$ differs from $m_w(x)$ throughout the cell by at most $Ld^{t/2}h^t$. Conditional on $N>0$, its sample mean has variance at most $1/N$. The identities and inequalities \[ {\mathbb{E}}\frac1{N+1}=\frac{1-(1-p)^{n+1}}{(n+1)p},\qquad \frac{{\mathbf 1}\{N>0\}}N\leq\frac2{N+1},\qquad (1-p)^n\leq\frac1{(n+1)p} \] follow respectively by integrating $(1-p+pt)^n$ on $[0,1]$, direct comparison for positive integers, and maximizing $(n+1)p(1-p)^n$ over $p$. They bound both sample variance and empty-cell loss by $C/(np)$. Since each arm probability in a cell is at least $\varepsilon P_X(A_j)$, integrating this bound over that cell contributes at most $C/n$. The squared-bias contribution is $Ch^{2t}$ over the entire domain. Sum over cells and arms and use $(u-v)^2\leq2u^2+2v^2$ to obtain the claim. ∎

Choose $h_E$ of order $n^{-1/(2t+d)}$ and let $T=T_{h_E}$. For the band use a second grid $h_B$ of order $(\log n/n)^{1/(2t+d)}$. On that grid, for positive cell-arm counts use center $T_{h_B}$ and radius \[ a_j=2Ld^{t/2}h_B^t+ \sum_{w=0}^1\sqrt{\frac{2\log(4J_B/\delta_n)}{N_{jw}}}, \qquad \delta_n=n^{-2}. \] Intersect the interval with $[-2,2]$, and use $[-2,2]$ when a count is zero. Call the resulting endpoints $L,U$.

Theorem (Joint attainment on the diagonal): Apply Theorem Theorem (Projection preserves risk at a quantifiable cost) with $M=2$ to these $T,L,U$. The resulting common center $C$ and radius $A$ obey \[ \sup_P{\mathbb{E}}\|C-\tau\|_{L^2(P_X)}^2\leq C_1 n^{-2t/(2t+d)}, \qquad \sup_P{\mathbb{E}}\|A\|_\infty\leq C_2(\log n/n)^{t/(2t+d)}, \] and cover $\tau$ simultaneously with probability at least $1-n^{-2}$. The constants depend only on fixed class parameters and $d$.
Proof: Lemma Lemma (Integrated risk of the armwise histogram) gives the stated integrated rate for $T$. Conditional on the cell-arm labels, Hoeffding's inequality for outcomes in $[-1,1]$ bounds each arm-mean error by its displayed square-root term with failure at most $\delta_n/(2J_B)$. A union bound and the two armwise Hölder bias bounds give coverage at least $1-\delta_n$ for $[L,U]$.

Each cell-arm probability is at least $\varepsilon f_*h_B^d$. Except with probability $2J_B\exp(-cnh_B^d)$, every count is at least $c'nh_B^d$, by the binomial lower-tail bound. On this event, maximum full width is at most \[ C\left\{h_B^t+ \sqrt{\frac{\log(4J_Bn^2)}{nh_B^d}}\right\} \leq C'(\log n/n)^{t/(2t+d)}. \] Width is always at most four, and the exceptional probability is smaller than this rate for all large $n$. Thus its expected width has the claimed order. Apply Theorem Theorem (Projection preserves risk at a quantifiable cost); the extra risk term is $16n^{-2}$, smaller than $n^{-2t/(2t+d)}$. The same data may construct both histograms because the transfer theorem does not require independence. ∎

Why the two orders are benchmark-optimal

Proposition (Risk lower bound): If the class constants admit the randomized Bernoulli submodels below, every estimator has worst-case integrated squared risk at least \[c n^{-2t/(2t+d)}.\]
Proof: Let $X$ be uniform and $W$ a fair randomized treatment. Choose a fixed bump $\varphi$ supported in the central half of the unit cube, Lipschitz, nonnegative, and with maximum exactly one. In $J=q^d$ disjoint cells set \[ \tau_\omega(x)=a\sum_{j=1}^J\omega_j\varphi((x-x_j)/h), \qquad \omega_j\in\{-1,1\},\quad a=\kappa h^t, \] with sufficiently small fixed $\kappa$. The separation between supports and the vanishing at support boundaries give a uniform $t$-Hölder constant independent of $J$. Let $m_0=0,m_1=\tau_\omega$, and let $Y\in\{-1,1\}$ have these armwise conditional means. Means have magnitude at most $1/2$ after reducing $\kappa$, so variances are uniformly positive. Potential outcomes can be generated independently conditional on $X$, independently of treatment, verifying the causal assumptions.

For neighboring sign vectors differing only at $j$, squared $L^2$ distance on that cell is $4a^2h^d\int\varphi^2$. The Bernoulli inequality ${\operatorname{KL}}(\operatorname{Ber}(p),\operatorname{Ber}(q))\leq(p-q)^2/[q(1-q)]$ follows from $\log u\leq u-1$. It gives an $n$-observation neighbor divergence at most $Cna^2h^d$. Take $q$ of order $n^{1/(2t+d)}$ and $\kappa$ small enough that every neighbor total variation is at most $1/2$, by Pinsker.

Decode each sign by the nearer of the two local bump functions to an arbitrary estimator in $L^2$ on its cell. An incorrect sign forces local squared error at least $a^2h^d\int\varphi^2$, by the triangle inequality between the two candidates. For every fixed choice of the other signs, the two testing error probabilities sum to at least $1-{\operatorname{TV}}\geq1/2$. Averaging over the uniform sign prior and summing cells gives average risk at least $cJa^2h^d=ca^2$. Maximum risk dominates the average, proving the order. ∎

For completeness, the logarithmic width lower bound follows from a distinct testing construction. Keep the same randomized model but use only $J$ alternatives with one positive bump each and the zero null. At amplitude $a=\kappa h^t$, disjoint supports make the cross likelihood-ratio moments equal to one, whereas a self moment for $n$ observations is $(1+c a^2h^d)^n$. The mixture chi-square distance is therefore \[ \chi^2=J^{-1}\{(1+c a^2h^d)^n-1\}. \] Choose $h$ of order $(\log n/n)^{1/(2t+d)}$ with a small enough constant so that $cna^2h^d\leq\frac12\log J$. Then the distance tends to zero. For a band honest at any fixed $1-\eta$ with $\eta<1/2$, reject the null if it contains any alternative function. Under the mixture this occurs with probability at least $1-\eta-o(1)$, hence also under the null up to $o(1)$. When it contains the null and its full width is less than $a$, it cannot contain an alternative. Null coverage thus implies probability at least $1-2\eta-o(1)$ of width at least $a$, yielding the matching expected-width lower bound. A symmetric band's full width is twice its maximum radius, so the same order applies to the criterion in the target.

Undersmoothing is a property of a construction

There is a concrete variance penalty if a histogram is forced to undersmooth. In the simpler randomized experiment with uniform $X$, $e=1/2$, and centered independent Rademacher outcomes in both arms, the cell-mean difference has zero bias. For a grid with $nh^d\to\infty$, its integrated variance is bounded above and below by constants times $(nh^d)^{-1}$. The upper bound follows from Lemma Lemma (Integrated risk of the armwise histogram). For the lower bound, if $N\sim\operatorname{Bin}(n,h^d/2)$, Chebyshev implies $\Pr(1\leq N\leq nh^d)\to1$, and on that event $1/N\geq1/(nh^d)$. Conditional outcome independence makes the two arm variances add. Integrating over cells proves the lower bound.

Taking $h=b_n n^{-1/(2t+d)}$, $b_n\to0$, with $nh^d\to\infty$, therefore inflates risk in this submodel by $b_n^{-d}$. The usual upper-bound ratio of smoothing bias to a pointwise standard error is proportional to $\sqrt{nh^d}h^t=b_n^{t+d/2}$ and does vanish. This calculation explains the cost for that undersmoothed center. Theorem Theorem (Joint attainment on the diagonal) shows that it is not a lower bound on every honest procedure: a high-coverage, explicitly bias-bounded band and projection avoid that risk penalty.

The unresolved part is the optimal joint pair when effect smoothness exceeds arm-mean smoothness and nuisance estimation constrains the best band. The transfer theorem still applies there if one can build a sufficiently high-coverage band with the right width; establishing that statistical input, including uniform remainder control, is the substantive gap.

References

  • [challenge] C. Cinelli, A. Feller, G. Imbens, E. Kennedy, S. Magliacane and J. Zubizarreta (2025), Challenges in Statistics: A Dozen Challenges in Causality and Causal Inference, arXiv:2508.17099v1, Section 3.3.3, "Inference," inspected 9 October 2026. https://arxiv.org/html/2508.17099v1#S3.SS3.SSS3.

Completion Estimate

55%

Go to contribution →