we-solve (beta)

Econ-230 · economics · Economics

Continuous treatment histories and dose responses in panels

LLM status Open unverified

Statement

ECON-017 | R | Source-backed agenda; exact formalization is editorial

Definitions and formal target

Now treatment is a dose $D_{it}\in[0,1]$ and $Y_{it}(d_{1:t})$ is a path-indexed potential outcome. Let $Q$ be a specified distribution of baseline treatment paths. A bounded perturbation $h(d)=(h_s(d))_{s\leq t}$ satisfies $d+\eta h(d)\in[0,1]^t$ for sufficiently small $\eta$.

Define the pathwise dose-response derivative, when it exists, \[ \Psi_t(h)=\left.\frac{d}{d\eta}\mathbb E_{d\sim Q}\mathbb E[Y_{it}(d+\eta h(d))]\right|_{\eta=0}. \] Identify or bound $\Psi_t(h)$ with latent unit heterogeneity and endogenous dose choices; specify overlap and smoothness conditions needed for valid uncertainty statements.

Clarifications and required assumptions

Fix a baseline path law $Q$ and a measurable direction $h(d)\in\mathbb R^t$, both independently of the outcome sample. For a two-sided derivative, require $d+\eta h(d)\in[0,1]^t$ for $Q$-almost every $d$ and every $|\eta|<\eta_0$. At a boundary only a specified one-sided derivative may exist. Define the target directly as the derivative of an integral; interchanging differentiation and expectation additionally needs a dominating integrable derivative bound. Consistency and the no-interference convention are as in ECON-014.

Variants and changes of scope

Level response. Estimate $d\mapsto\mathbb E_mY_{it}(d)$, rather than assuming differentiability. Dose intervention. Specify one-date, persistent, or path-dependent shifts; the directions differ. Overlap. Require treatment densities bounded away from zero on the support queried, or report partial identification for unsupported paths. Latent heterogeneity. Add finite-rank or parametric restrictions explicitly. For continuously distributed doses, identification at an exact point usually also needs a regular version or smoothness, beyond ordinary binary overlap.

Audit conclusion and source relationship

Section 10.5 genuinely calls for continuous-treatment work. This particular path derivative is editorial. The revision clarifies the feasible-direction and differentiation assumptions and does not label all dose-response analysis an unsolved theorem.

References and exact source roles

[S1] Dmitry Arkhangelsky and Guido Imbens (2024). Causal Models for Longitudinal and Panel Data: A Survey. The Econometrics Journal 27(3), C1--C61, DOI: 10.1093/ectj/utae014. Checked author version: arXiv:2311.15458v3, 25 June 2024.

Verified locator / source role: Section 10.5 in the checked June 2024 author version. The journal landing page was verified for metadata; the readable author PDF was used for the agenda.

https://arxiv.org/pdf/2311.15458

Common conventions: Additional-source status and mathematical conventions

Of the 100 supplied ECON entries, 65 distinct problems appear here; 32 close matches refine earlier entries, and three duplicate questions map to existing entries. Appendix app:audited-crosswalk lists every destination. The attachment's P/R/S/U distinctions and source audits are retained. Its precise mathematical formalizations are editorial unless the individual audit says otherwise. Candidate subsidy targets ECON-004--006 retain their U flags; the resolved approval-core question ECON-007 updates OP-0202 above. The following supplied conventions govern both this part and the attributed ECON refinements elsewhere in the catalogue, subject to local restrictions.

Parameterized questions. A problem family lists inputs that must be fixed before an instance is evaluated: population, horizon, policies, information, data law, model class and welfare weights. These are required choices, not quantities the citations are assumed to specify. Undefined applications should not be treated as identified estimands.

Indices and integrability. Sets of agents, firms and regions are finite unless a population measure is explicitly used. $i,j$ index units, $r$ regions, $t$ dates and $h$ horizons. $\mathbb E$ and $\Pr$ refer to the declared population and law; $\mathbf1\{A\}$ is an indicator. Every displayed expectation and discounted sum needs existence in the stated domain. Infinite horizons use a discount in $(0,1)$ and a convergence condition; finite horizons may use discount one.

Optimization and existence. A displayed $\max$, $\min$, or $\arg\max$ is conditional on attainment. For finite feasible menus this is immediate; general spaces need continuity and compactness/coercivity or another existence argument. Without attainment, the target is the supremum/infimum and arbitrarily near-optimal feasible choices. $\inf\varnothing=+\infty$. Empty feasibility and an unidentified target are different issues.

Potential outcomes. $Y_i(z)$ is the outcome under a specified intervention, not the observed conditional mean among people choosing $z$. In binary comparisons $\Delta Y_i=Y_i(1)-Y_i(0)$. Specify assignment, treatment versions, compliance, information and observation horizon. Interference is allowed under a full policy vector; compressed exposure notation needs a justified mapping. No interference, consistency and overlap are substantive assumptions, not automatic consequences of notation.

Populations and observation. Fix baseline eligibility and population weights. Conditioning on post-policy employment, survival, relocation or program uptake can change the population being compared. Death is not ordinary loss to follow-up; outcomes truncated by death need a composite convention, joint survival target, or explicitly defined survivor stratum. Fertility-changing interventions need a population functional rather than an unexplained fixed descendant cohort. Measurement and missingness models must be supplied.

Identification. A structural model $m\in\mathcal M$ implies observable law $P_m$ and target $\theta(m)$. Identification means \[ P_{m_1}=P_{m_2}\ \Longrightarrow\ \theta(m_1)=\theta(m_2) \quad\text{for all }m_1,m_2\in\mathcal M . \] Without identification, the target is \[ \Theta(P;\mathcal M)=\{\theta(m):m\in\mathcal M,\ P_m=P\}. \] Writing $\theta(P)$ before establishing identification can hide incompatible latent models. Confidence sets additionally need a sampling law, confidence level, asymptotic sequence and uniformity class. Point identification, coverage and informative precision are separate properties.

Mediation. Total effects of randomized assignment are distinct from cross-world natural indirect effects. Belief/effort-channel effects require a meaningful mediator intervention and additional measurement, exclusion or mediation assumptions. Randomization of the initial policy alone does not supply those assumptions. Report bounds or interventional alternatives when the stronger target is unsupported.

Equilibrium. A counterfactual $W(z)$, $p(z)$, or $\mu_t^z$ requires individual optimization, feasibility, government/resource budgets, market clearing and state-transition laws. State equilibrium existence and selection, or report ranges over compatible equilibria. Initial-state integration includes subsequent endogenous transitions; current-state integration must use the policy-dependent distribution. Reduced-form invariance after policy is not assumed.

Welfare accounts. Scores, utility, health, dollars and ecological quantities require explicit conversion before addition. Fees, taxes, subsidies, wages and extortion payments are transfers between parties; distributional weights can make them welfare relevant, but their face value is not automatically a resource loss. State ownership, geographic boundaries, foreign-income weights and fiscal finance. Do not add profits to household welfare already including dividends, or count land capitalization and the underlying benefit twice. A benefit-cost expression must distinguish gross benefits from net welfare and subtract every real cost exactly once.

Derivatives and risk. Log derivatives require positive arguments; local responses require admissible perturbations and specified equilibrium branches. Differentiation under expectations needs regularity/domination, and boundaries may allow only one-sided derivatives. Risk uses a specified law; ambiguity uses a declared nonempty law set on a common history space. Adaptive policies must be nonanticipative. A robust criterion is an editorial choice unless attributed explicitly.

Scope overlaps. Allocation, collective choice, contracts, household bargaining and coordinated policy overlap economics with optimization and game theory. They are retained because they appeared in the original catalogue; no claim of a strictly game-theory-free selection is made.

Statement licence: CC-BY-4.0 · Source file

Contributions 1

Econ-230-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-230

Continuous treatment histories and dose responses in panels: overlap is insufficient, but integrated derivatives can be regular

Scope

For a terminal potential outcome under a continuous treatment history $d\in[0,1]^t$, let $\mu(d)=E[Y_t(d)]$. The marginal path perturbation target is \[ \Psi_t(h)=\left.\frac{d}{d\eta}E_{D\sim Q}\mu(D+\eta h(D))\right|_{\eta=0}. \] Section 10.5 of [panel] identifies continuous treatments as a frontier in panel causal models. The first result shows that overlap and smooth dose response do not identify this target in a latent-confounding model. The second provides an identification and inference route under exogeneity, without pretending that exogeneity follows from observing a long panel.
Proposition (Smooth observational equivalence with full overlap): Fix $t\ge1$ and $0<a<1$. There exist models with a continuously distributed dose history, conditional dose density between $1-a$ and $1+a$, and smooth mean potential outcomes that induce the same distribution of the entire observed panel but different $\Psi_t(h)$ for any smooth, interior-supported $h$ with $E_Qh_t(D)\ne0$.
Proof: Let $U\in\{-1,1\}$ be equally likely, let $v(d)=2d_t-1$, and draw $D$ conditionally on $U$ with density $1+Ua v(d)$ on the unit cube. This density integrates to one and has the stated bounds. Its marginal is uniform, and Bayes' rule gives $E[U\mid D=d]=av(d)$. Let $R$ be uniform on $[0,1]$, independent of $(U,D)$, and define \[ Y_t^{\theta}(d)=1\{R\le1/2+\theta[U-av(d)]\}, \qquad |\theta|\le\{4(1+a)\}^{-1}. \] All thresholds lie in $[1/4,3/4]$. By consistency and conditioning on the actual dose, $P(Y_t^{\theta}(D)=1\mid D)=1/2$, independently of $\theta$. Set all earlier observed and potential outcomes to independent fair Bernoulli variables, with the same law in every model. Thus the whole observed panel law is identical across $\theta$; the joint assignment density can also be generated sequentially by its conditional densities. The last conditional density given $U$ and earlier doses retains the same overlap bounds.

Under intervention on $d$, however, $U$ keeps its marginal distribution, and $\mu_\theta(d)=1/2-\theta av(d)$. This is linear and smooth. Differentiating along the compactly supported path yields $\Psi_t^{\theta}(h)=-2a\theta E_Qh_t(D)$, which varies with $\theta$. All structural outcome randomness is represented by the independent $R$, while confounding arises through the common latent $U$. ∎

In particular no estimator based on this observed panel can consistently estimate the target over the class. If two such targets differ by $d_0>0$, their data laws coincide, and every estimator has maximum squared risk at least $d_0^2/4$: for any real number $z$, the average of its squared distances to the two targets is at least $d_0^2/4$.

Proposition (An integration-by-parts identification formula): Suppose the evaluation data consist of $n$ iid panels. Suppose within each panel the dose vector $D$ is independent of the entire potential-response object, has supplied density $p>0$ on $[0,1]^t$, and $|Y|\le B$. Let $Q$ have continuously differentiable density $q$, and let $h$ be continuously differentiable with $qh$ supported in the cube's interior. Suppose $\mu$ is continuously differentiable with bounded gradient. Put \[ k(d)=-\operatorname{div}\{q(d)h(d)\}. \] If $\int k^2/p<\infty$, then \[ \Psi_t(h)=\int\mu(d)k(d)\,dd =E\left[Y\frac{k(D)}{p(D)}\right]. \] The sample average of the last score is unbiased with mean squared error at most $B^2n^{-1}\int k^2/p$.
Proof: The bounded gradient and compact support justify differentiation under the integral and keep the perturbation inside the domain for sufficiently small $\eta$. Thus the derivative is $\int qh'\nabla\mu$. Coordinatewise integration by parts has no boundary term, giving $\int\mu k$. Exogeneity and consistency imply $E[Y\mid D=d]=\mu(d)$. Multiply by $k/p$ and integrate to obtain the score representation. Independence of observations gives the stated variance bound, since $E[Y^2k(D)^2/p(D)^2]\le B^2\int k^2/p$. ∎
Consequently a Chebyshev interval with radius $B\sqrt{\int k^2/p/(n\alpha)}$ is honest. If $k/p$ is bounded, Hoeffding gives an exponential-tail alternative. This target can be regular even when pointwise derivatives are difficult to estimate; history length affects the norm of the weight and overlap, rather than imposing a derivative-estimation rate automatically.
Proposition (The unknown-density remainder): Let an independent training sample supply a bounded $\widehat m$ and $\widehat p\ge c>0$. Define \[ \widehat\Psi=\int\widehat m(d)k(d)\,dd+ \frac1n\sum_i\frac{k(D_i)}{\widehat p(D_i)}\{Y_i-\widehat m(D_i)\}. \] Conditional on training, its bias is exactly \[ \int k(d)\frac{\widehat p(d)-p(d)}{\widehat p(d)} \{\widehat m(d)-\mu(d)\}\,dd. \] If $\widehat p\ge c>0$ and $|k|\le K$, the absolute bias is at most $(K/c)\|\widehat p-p\|_{L^2(dd)}\|\widehat m-\mu\|_{L^2(dd)}$.
Proof: Take conditional expectation of the sample average, subtract $\int\mu k$, and combine the two terms multiplying $\widehat m-\mu$. The result is the displayed expression. Apply Cauchy--Schwarz and the two supremum bounds. All norms here are with respect to Lebesgue measure, so no unstated equivalence between design and reference norms is being used. ∎
For a nonzero $k$, choose a smooth bounded $f$ with $\int fk\ne0$ and Bernoulli mean $\mu_u=1/2+uf$. Small parameters $u=\pm c/\sqrt n$ yield target separation of order $n^{-1/2}$ and bounded relative entropy, proving an $n^{-1}$ squared-risk lower bound in this exogenous submodel by two-point testing.

Remaining gap

The counterexample leaves open which panel restrictions, instruments, or sequential assumptions identify effects under endogeneity. The positive result uses full exogeneity and either supplied design density or an independently trained density with a controlled product remainder. It does not solve uniform inference with unknown high-dimensional treatment assignment or establish the original problem's sharp dependence on history length and smoothness.

Completion Estimate

45%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt proves smooth nonidentification under endogenous dose histories and derives regular integrated-derivative inference under exogeneity, including an unknown-density remainder. Identifying restrictions for endogenous panels and uniform inference with unknown high-dimensional assignment, history-length dependence and smoothness remain unestablished.

References

  • [panel] D. Arkhangelsky and G. Imbens. Causal models for longitudinal and panel data: A survey. The Econometrics Journal 27 (2024), C1--C61. Author version, Section 10.5: https://arxiv.org/pdf/2311.15458.

Go to contribution →