we-solve (beta)

Econ-573 · economics · Economics

Transporting causal effects

LLM status Open unverified

Statement

Original subject group: Econometrics, causal inference and measurement

Audit classification: Background; proposed extension. The cited paper supplies methods rather than the exact stated sensitivity conjecture. The sharp interval under a supplied delta is immediate; the research component is obtaining defensible delta and efficient validation.

Definitions and precise research target

Let $S=1$ denote a source population with a randomized binary treatment $A$, and $S=0$ a target population with covariates $X$ observed but no randomized outcomes. Potential outcomes $Y(0),Y(1)\in[0,1]$ obey consistency. Source randomization identifies $\tau_1(x)=E[Y(1)-Y(0)\mid X=x,S=1]$, and the target estimand is $\tau_0=E[Y(1)-Y(0)\mid S=0]$. Require the target covariate law to be absolutely continuous with respect to the source covariate law, with an explicit lower-overlap bound if sampling inference is sought. Unobserved moderators and institutional differences may invalidate equality of conditional effects. For a stated nonnegative sensitivity function $\delta(x)$, impose \[ |E[Y(1)-Y(0)\mid X=x,S=0]-\tau_1(x)|\leq\delta(x). \] Together with bounded outcomes, this yields an identified set by integrating admissible conditional effects over the target covariate distribution. Determine which additional target experiments, moderator measurements, institutional restrictions or calibration data justify smaller $\delta(x)$, and construct sharp bounds and confidence regions under the declared restrictions. If target support differs, include the corresponding unconstrained outcome bounds rather than silently extrapolating. Treatment versions and outcome definitions must agree across populations. The problem is to obtain informative transport under credible added information; unrestricted transport is not identifiable, and known sufficient assumptions are not being presented as unresolved discoveries.

Definitions, assumptions and mathematical qualifications. Source randomization means $A\perp(Y(0),Y(1))\mid X,S=1$ with probabilities bounded away from zero and one. Let $Q$ be the target covariate law. On source support, define $l(x)=\max\{-1,\tau_1(x)-\delta(x)\}$ and $u(x)=\min\{1,\tau_1(x)+\delta(x)\}$. With no other constraints coupling conditional effects, the sharp set is $[\int l\,dQ,\int u\,dQ]$. Off support replace $l,u$ by $-1,1$. This elementary bounding result is not an open theorem; meaningful work concerns credible calibration of $\delta$, target validation design and inference. Treatment versions, measurement scale and sampling law are explicit input assumptions. Here overlap means absolute continuity of the target covariate law with respect to the source law, not merely containment of topological supports: source conditional means are defined up to source-null sets. For an allowed target mass on source-null sets, use the unconstrained outcome bounds there.

Variants and assumption changes

  1. Editorial, exact transport benchmark. Set $\delta=0$ and require target support overlap, identifying $\tau_0=\int\tau_1(x)dQ(x)$; this is a sufficient assumption, not a finding from covariates alone.
  2. Editorial, target validation. Allocate a limited target experimental sample across covariate strata and seek bounds on conditional drift with stated coverage, then propagate those bounds into $\tau_0$.

References and source audit

Checked source. NBER Working Paper 30398 (August 2022), revision June 2023, indexed abstract; Research in Economics 78(3) (2024), article 100964. The paper reviews transport methods and supplies a CATE method. Primary metadata/abstract checked; full article returned HTTP 403. https://www.nber.org/papers/w30398.
  1. Sebastian Galiani; Brian Quistorff. 2022. Assessing External Validity in Practice. NBER Working Paper 30398 (August 2022), revision June 2023, indexed abstract; Research in Economics 78(3) (2024), article 100964. https://www.nber.org/papers/w30398. DOI: https://doi.org/10.3386/w30398.

Common conventions: Source mathematical conventions

These conventions apply to each entry unless explicitly replaced.

Models and identification. A primitive model $m$ specifies all probability laws, feasible choices, information, equilibrium and measurement rules used by its target. The admissible class $\mathcal M$ is fixed independently of outcomes. If $P_O(m)$ is its observable probability law and $\tau(m)$ the target, its identified set at observable law $P$ is \[ \mathcal I_\tau(P)=\{\tau(m):m\in\mathcal M,\ P_O(m)=P\}. \] Point identification means that this set is a singleton. An empty set signals incompatibility of data and assumptions. Coordinate bounds need not describe the joint set. Bounds are sharp only if every retained target is attainable or arbitrarily approachable by compatible models. Finite bounds require explicit restrictions on unobserved quantities; unrestricted confounding, hidden wealth, extrapolation or arbitrary equilibrium selection can yield uninformative sets.

Causal interventions. A potential outcome $Y_i(p)$ is the outcome under a complete policy regime $p$, including timing, financing, information and market spillovers. The target population and units are fixed before treatment. With interference, use the complete assignment vector or a justified exposure mapping; individual no-interference notation alone cannot identify an economy-wide intervention. A difference of mean potential outcomes does not require the cross-world joint distribution. Conditioning on post-treatment migration, survival, enrollment or employment changes the estimand and needs separate justification.

Statistical statements. An estimator, confidence set, forecast or test specifies a sampling law, model class and loss/coverage criterion. Uniform $(1-\alpha)$-coverage means $\inf_{m\in\mathcal M}\Pr_m\{\tau(m)\in C_n\}\geq1-\alpha$, or an explicitly asymptotic version. The whole parameter space trivially has coverage, so informativeness requires a stated width, risk or power objective. A forecasting improvement is relative to a named benchmark, information set and evaluation population.

Equilibrium and optimization. An equilibrium comprises feasible plans, optimality, beliefs where needed, budget constraints and all market/resource clearing. The primitives or a stated benchmark define each correspondence $\mathcal E(p;m)$. Nonemptiness is assumed or proved, never inferred just from compact policy sets. A selected-equilibrium target uses a declared selection rule; a robust target quantifies over all permitted equilibria. Use suprema or infima unless attainment is established. An $\varepsilon$-optimal policy is feasible and within $\varepsilon$ of the finite optimum value. Report an empty feasible set separately.

Welfare and accounting. Normative utility, interpersonal normalization, social weights, numeraire, discounting and the use of public receipts are primitives. Behavioral utility need not coincide with the welfare criterion. Household welfare includes ownership income and rents once; adding profits again to compensating variation generally double-counts them. A monetary equivalent is defined by an explicit transfer and price convention, with monotonicity and range conditions. Average effects, mechanism contributions, transfers and welfare losses are different targets. Infinite sums require convergence and dynamic feasible plans require terminal or no-Ponzi conditions.

Source versions. A working-paper date, revision date and journal publication year are distinguished. A DOI for a journal version is not automatically the DOI of an earlier manuscript. Locators refer to the stated version and printed pages where specified. A metadata-only check does not verify a claim about a full-text conclusion. Every entry's source subsection narrows the provenance claim accordingly.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-573-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-573

Transporting causal effects: target validation under a spatial drift restriction

Target and scope

With source conditional effect $\tau_1(x)$ and target covariate law $Q$, write \[ d(x)=\tau_0(x)-\tau_1(x),\qquad \tau_0=\int\{\tau_1(x)+d(x)\}\,dQ(x). \] A supplied bound on $d$ gives an immediate interval, but does not explain how to calibrate it. This note studies target experiments at selected covariate locations under an explicit Lipschitz restriction on drift. It gives sharp coupled bounds and an exact design result for noiseless validation. The review and CATE transport method in [transport] motivate external-validity assessment; the Lipschitz restriction and design theorem below are additional maintained assumptions.

A finite-stratum calibration problem

There are $J$ covariate strata with supplied target weights $q_j\ge0$, $\sum_jq_j=1$, and metric distances $d_{jk}$. Source treatment is randomized with overlap, consistency and comparable outcome measurement. To isolate target calibration, suppose its stratum effects $t_j\in[-1,1]$ are supplied. Historical support failure is handled by using a direct target-effect coordinate without relying on an arbitrary source conditional mean.

Let $v_j$ denote target-minus-source drift, fix a finite $L\ge0$, and impose \[ a_j\le v_j\le b_j,\qquad |v_j-v_k|\le Ld_{jk}. \] The intervals incorporate $-1-t_j\le v_j\le1-t_j$, substantive drift restrictions, and any target validation confidence intervals. No claim is made that nearby observed covariates ensure nearby unobserved effect drift.

Theorem (Sharp coupled transport bounds): Define \[ A_j=\max_k(a_k-Ld_{jk}),\qquad B_j=\min_k(b_k+Ld_{jk}). \] The restriction class is nonempty if and only if $A_j\le B_j$ for every $j$. When nonempty, the sharp target-effect interval is \[ \left[\sum_jq_j(t_j+A_j),\ \sum_jq_j(t_j+B_j)\right]. \] Both endpoint drift vectors are admissible, and every intermediate target value is attained.
Proof: Every admissible $v$ obeys $v_j\ge v_k-Ld_{jk}\ge a_k-Ld_{jk}$ and the analogous upper bound, so $A_j\le v_j\le B_j$. Each function $j\mapsto a_k-Ld_{jk}$ is $L$-Lipschitz by the triangle inequality; their maximum is also $L$-Lipschitz. Similarly $B$ is $L$-Lipschitz. Taking $k=j$ gives $A_j\ge a_j$ and $B_j\le b_j$. If $A\le B$, it follows that $A$ and $B$ both satisfy the local intervals and Lipschitz constraints. This proves sufficiency and simultaneous endpoint attainment; necessity was already shown. The class is convex, so convex combinations fill the interval. Every effect in $[-1,1]$ can be realized by bounded potential outcomes, for example using means $(1+t_j+v_j)/2$ and $(1-t_j-v_j)/2$. Independent uniforms and randomized assignments realize the conditional experiments, so the endpoints are causal-model compatible in this declared class. ∎
The coupling can tighten strata with no direct validation, but can also refute an overly small $L$. Empty feasibility must be reported as incompatibility, not converted into a point estimate.
Proposition (Coverage from a stratified target experiment): In selected strata collect independent randomized target samples with fixed arm counts $n_{j1},n_{j0}>0$, independent across strata, consistency, no interference and outcomes in $[0,1]$. Let $\widehat t_j^0$ be the difference of arm means. For $K$ validated strata, put \[ e_j=\sqrt{\frac{\log(4K/\alpha)}{2n_{j1}}} +\sqrt{\frac{\log(4K/\alpha)}{2n_{j0}}}. \] Intersect the prior drift interval with $[\widehat t_j^0-t_j-e_j,\widehat t_j^0-t_j+e_j]$ at validated strata and apply the theorem. If the maintained restrictions are true, the resulting interval covers the true target effect with probability at least $1-\alpha$.
Proof: Hoeffding's inequality bounds the error of each arm mean with two-sided failure probability at most $\alpha/(2K)$. A union bound gives simultaneous effect coverage. On that event the true drift vector remains in every intersected interval and obeys the maintained metric restriction. It is therefore feasible in the theorem's optimization, so its weighted target lies between the computed endpoints. Independence is a sufficient sampling condition for each arm-mean bound; arbitrary observational selection into validation would not justify it. ∎
An independent source confidence region can be included by optimizing jointly over source effects and target drifts, with a union-bound allocation of coverage error. Simply plugging noisy source estimates into the metric constraints would omit uncertainty in the drift definition.

An exact validation-location result

Consider a continuous covariate $x\in[0,1]$ with uniform target law, source effect zero, and $0<L\le1$. Observe the true drift at $K\ge1$ chosen locations, all equal to zero. The only other restrictions are $|v|\le1$ and Lipschitz constant at most $L$.
Theorem (Optimal noiseless anchors): For an anchor set $S$ of cardinality $K$, the target-effect interval is \[ [-L\int_0^1\operatorname{dist}(x,S)dx, L\int_0^1\operatorname{dist}(x,S)dx]. \] Its minimum possible width is $L/(2K)$, attained by the equally spaced midpoints $(2j-1)/(2K)$.
Proof: The anchor equations imply $|v(x)|\le L\operatorname{dist}(x,S)$. The positive and negative distance functions are Lipschitz and vanish at all anchors; because $L\le1$, they satisfy the outcome-effect bound and attain both integrals. Partition the interval into the nearest-anchor Voronoi cells, with lengths $\ell_1,\ldots,\ell_K$. On any interval of length $\ell$, the integral of distance to a point is minimized at its midpoint, with value $\ell^2/4$; if the point lies outside that interval its integral is larger. Therefore $\int\operatorname{dist}(x,S)dx\ge\sum_j\ell_j^2/4\ge1/(4K)$ by Cauchy--Schwarz. Equal cells and their midpoints attain both inequalities. Multiplying by $2L$ proves the width assertion. ∎
With noisy anchor intervals of radius $e_j$ centered at zero, the pointwise magnitude is bounded by $\min_j\{e_j+L|x-x_j|\}$, capped at one. This displays the competing costs of covering more locations and measuring each precisely. It does not establish that anchor designs are minimax for the single average effect: direct random sampling throughout the target can estimate that average at a parametric rate without learning a uniform drift surface.

Remaining gap

The results specify exactly what a credible smooth-drift assumption buys and how validation uncertainty propagates. They do not make that assumption credible from covariates alone, determine its constant from finitely many anchors, or solve adaptive noisy validation in high dimension. Unmeasured institutional discontinuities can violate the restriction even when measured covariates are close. Those calibration and design issues remain central to the original target.

Completion Estimate

56%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt proves sharp coupled transport bounds under Lipschitz drift, propagates stratified validation uncertainty, and identifies optimal noiseless anchor locations in a one-dimensional benchmark. Credible calibration of the drift modulus, adaptive noisy validation in higher dimensions, and informative transport across institutional discontinuities remain unresolved.

References

  • [transport] S. Galiani and B. Quistorff. Assessing External Validity in Practice. NBER Working Paper 30398 (2022), revised June 2023. Primary abstract and version metadata. https://www.nber.org/papers/w30398.

Go to contribution →