we-solve (beta)

Econ-192 · economics · Economics

Hybrid panel models combining selection and latent factors

LLM status Open unverified

Statement

ECON-016 | R | Source-backed agenda; exact formalization is editorial

Definitions and formal target

For binary treatment $d$, one candidate outcome model is \[ Y_{it}(d)=m_d(X_{it})+\ell_i^{\mathsf T}f_{dt}+\varepsilon_{it}(d). \] Specify structural classes $\mathcal M_{X,n}$ and $\mathcal M_{F,n}$: the first imposes a precise observed-covariate selection restriction; the second a precise latent-factor and treatment-design restriction. Define the potential-outcome target $\tau(m)$, sample $O_n$, and asymptotic sequence. Conditional on identification across their union, seek \[ \sup_{m\in\mathcal M_{X,n}\cup\mathcal M_{F,n}} P_m\{|\widehat\tau_n(O_n)-\tau(m)|>\varepsilon\}\longrightarrow0 \] for every $\varepsilon>0$, with honest confidence sets under declared factor-rank, overlap and dimension bounds. The classes are required inputs, not assumed identified by the displayed factor equation.

Clarifications and required assumptions

Let $\mathcal M_X$ and $\mathcal M_F$ be structural-model classes, rather than classes of observable laws alone. The factor representation uses a fixed or explicitly varying rank $r$, $\ell_i\in\mathbb R^r$, and $f_{dt}\in\mathbb R^r$. Rotation normalizations do not identify treatment effects by themselves. Before seeking an estimator robust on the union, require $P_{m_1}=P_{m_2}\Rightarrow\tau(m_1)=\tau(m_2)$ for all $m_1,m_2\in\mathcal M_X\cup\mathcal M_F$. Identification separately within each class is insufficient when cross-class observationally equivalent models disagree.

Variants and changes of scope

Union consistency. Require $\sup_{m\in\mathcal M_{X,n}\cup\mathcal M_{F,n}}P_m(|\widehat\tau_n-\tau(m)|>\varepsilon)\to0$ for each $\varepsilon>0$. Honest inference. Also specify uniform overlap, moment, approximation-error and rank bounds. Intersection efficiency. Seek efficiency when both restrictions hold, separately from union robustness. Approximate models. Declare misspecification tolerances and the resulting sensitivity set. A completely unspecified union gives no automatic doubly robust theorem.

Audit conclusion and source relationship

Section 10.4 motivates bridging unconfoundedness and two-way fixed effects. Replacing fixed effects by general factor models and imposing uniform union robustness are editorial extensions. The revision removes the circular use of a single-valued $\tau(P)$ before union identification is established.

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 10.4 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-192-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-192

Hybrid panel models combining selection and latent factors Cross-model incompatibility and a compatible temporal-bridge union

Problem ID: C23-188. Status: Unresolved; rigorous restricted results.

Short recap and contribution

The target is a uniformly consistent estimator and honest inference on a union of observed-selection and latent-factor models. Identification inside each class does not establish identification across their union. We give an explicit two-period incompatibility example, even with overlap given the observed lagged outcome. We then specify a smaller compatible union, prove its common identifying formula, give finite-sample confidence sets, and quantify departures from its assumptions. The general noisy-factor and unknown-design problem remains open.

Arkhangelsky and Imbens [survey], Section 10.4 of the June 2024 author version, motivate methods combining unconfoundedness and fixed effects. That section and the survey's Section 7.3.2 on augmentation were checked. The broader arbitrary-rank union is the catalogue's editorial extension. The augmented weighting algebra used here is standard; the point is to state structural classes on which its target is actually common.

An explicit failure of union identification

Observe independent copies of $(D,Y_0,Y_1)$, with treatment only in the second period, consistency $Y_1=DY_1(1)+(1-D)Y_1(0)$, no interference, and target $\tau={\mathbb{E}}[Y_1(1)-Y_1(0)\mid D=1]$. The following two structural classes are each identified for this target. Class $\mathcal X$ assumes conditional mean exchangeability of $Y_1(0)$ given $Y_0$ and overlap given $Y_0$. Class $\mathcal F$ assumes \[ Y_t(0)=\ell+g_t+\eta_t,\quad g_0=0,\quad {\mathbb{E}}[\eta_1-\eta_0\mid D=1] ={\mathbb{E}}[\eta_1-\eta_0\mid D=0]. \tag{1} \] This is a rank-one fixed-effect component with a stated group-mean restriction on the disturbances; it does not require conditional mean independence given $\ell$. All expectations are finite. The untreated trend in (1) has the same group mean, so the DID formula identifies $\tau$ within $\mathcal F$.
Proposition (Separate identification does not survive union): There are models $m_X\in\mathcal X$ and $m_F\in\mathcal F$ with the same observable law, with $\Pr(D=1\mid Y_0)$ taking only values $1/4,3/4$, but with $\tau(m_X)=1$ and $\tau(m_F)=1/2$. No estimator is consistent at both models. Any confidence set honest at level $1-\alpha$ at both has diameter at least $1/2$ with probability at least $1-2\alpha$ under their common observable law.
Proof: Generate a fair binary latent variable $A$, then a binary assignment type $T$ with $\Pr(T=1\mid A)=1/4+A/2$. Set $D=T$, $Y_0=A$, and let $Y_1=D$ in the observable law. Here $\Pr(D=1)=1/2$, ${\mathbb{E}}[A\mid D=1]=3/4$, and ${\mathbb{E}}[A\mid D=0]=1/4$.

For $m_X$, define $Y_1(0)=0$, $Y_1(1)=1$. Conditional mean exchangeability given $A=Y_0$ holds, both potential outcomes being constants, and the stated overlap holds. Consistency gives the stipulated observed law and $\tau(m_X)=1$.

For $m_F$, define $Y_1(0)=T(A-1/4)$ and $Y_1(1)=1$. The assignment type is a common cause of treatment and potential outcomes; it is held fixed under treatment intervention, so this definition is a valid structural joint law and not a replacement of treatment by its intervention value inside a potential outcome. Consistency again gives $Y_1=D$. Set $\ell=A$, $g_1=-1/4$, $\eta_0=0$, and $\eta_1=T(A-1/4)-A+1/4$. For $D=1$ this error is zero; for $D=0$ it is $1/4-A$ with conditional mean zero. Thus (1) holds. The treated untreated mean is $3/4-1/4=1/2$, giving $\tau(m_F)=1/2$.

For any estimator, the events that it lies within distance less than $1/4$ of the two targets are disjoint. Their probabilities under the same observable sample law cannot both tend to one. For confidence sets, the two coverage events each have probability at least $1-\alpha$ under that law; their intersection has probability at least $1-2\alpha$. On the intersection the set contains both $1/2$ and $1$, proving the claimed diameter restriction. ∎

To check separate identification explicitly, $\mathcal X$ uses \[ {\mathbb{E}}[Y_1\mid D=1]- {\mathbb{E}}\{{\mathbb{E}}[Y_1\mid D=0,Y_0]\mid D=1\}, \] whereas $\mathcal F$ uses ${\mathbb{E}}[Y_1-Y_0\mid D=1]-{\mathbb{E}}[Y_1-Y_0\mid D=0]$. In the displayed observed law these give $1$ and $1/2$, respectively. This counterexample is not offered against more restrictive factor models that require conditional restrictions absent from (1). It establishes why a proposed union must be audited across classes, even after checking rank and observed overlap.

A compatible union through a supplied temporal bridge

Now specify a different experiment. Observe $n$ independent copies of $O=(U,D,Y)$, where $U=(X,P)$ contains baseline covariates and a vector $P\in{\mathbb{R}}^T$ of pretreatment outcomes. Let $Y(d)$ be the post-treatment potential outcomes, with $|Y(d)|\leq M$ and consistency. The target is again the ATT, $\tau={\mathbb{E}}[Y(1)-Y(0)\mid D=1]$. Assume $p=\Pr(D=1)\geq p_0>0$ and actual propensity $e(U)\in[\epsilon,1-\epsilon]$. Fix before seeing these observations measurable working functions $e_*(U)\in[\epsilon,1-\epsilon]$ and $h_*(U)\in[-M,M]$. They are supplied inputs, not estimates implicitly assumed known to be correct.

Define the following structural classes. In $\mathcal M_X$, require $e=e_*$ and \[ {\mathbb{E}}[Y(0)\mid U,D=1]={\mathbb{E}}[Y(0)\mid U,D=0]. \tag{2} \] In $\mathcal M_F$, require the stronger outcome bridge \[ {\mathbb{E}}[Y(0)\mid U,D=d]=h_*(U),\qquad d=0,1, \tag{3} \] while the actual propensity need not equal $e_*$. Thus one class supplies a correct treatment design with selection adjustment, and the other a correct untreated-outcome bridge. Neither class restricts selection on $Y(1)$, which is unnecessary for the ATT.

Here is an explicit fixed-rank factor model giving (3). Let \[ P=g(X)+F\ell,\qquad Y(0)=g_0(X)+f^\mathsf T\ell+\eta, \qquad{\mathbb{E}}[\eta\mid U,D]=0, \] where $F$ is a supplied $T\times r$ matrix and $f=F^\mathsf Tb$ for a supplied $b\in{\mathbb{R}}^T$. Then \[ h_*(U)=g_0(X)+b^\mathsf T[P-g(X)] \tag{4} \] is the bridge. Assume its absolute value is at most $M$. This construction permits latent loadings and their selection, but requires noiseless pretreatment factors and a supplied temporal span restriction. Rotation of factors changes neither (4) nor identification. It does not claim that short noisy histories automatically reveal the loadings.

Theorem (A common identified functional): On $\mathcal M_X\cup\mathcal M_F$, define the observed score \[ A(O)=D[Y-h_*(U)]- (1-D)\frac{e_*(U)}{1-e_*(U)}[Y-h_*(U)]. \tag{5} \] Then ${\mathbb{E}} A=p\tau$. Consequently the union is point identified by the single observable functional ${\mathbb{E}} A/{\mathbb{E}} D$, including observationally equivalent models belonging to different classes.
Proof: Write $\mu_d(U)={\mathbb{E}}[Y(0)\mid U,D=d]$ and $\mu_{11}(U)={\mathbb{E}}[Y(1)\mid U,D=1]$. Conditioning on $U$ gives \[ {\mathbb{E}} A={\mathbb{E}}\left[e(\mu_{11}-h_*) -(1-e)\frac{e_*}{1-e_*}(\mu_0-h_*)\right],\qquad p\tau={\mathbb{E}}[e(\mu_{11}-\mu_1)]. \] Subtracting yields the exact identity \[ {\mathbb{E}} A-p\tau= {\mathbb{E}}[e(\mu_1-\mu_0)]+ {\mathbb{E}}\left[\frac{e-e_*}{1-e_*}(\mu_0-h_*)\right]. \tag{6} \] In $\mathcal M_X$, the first term is zero by (2) and the second by $e=e_*$. In $\mathcal M_F$, (3) makes both terms zero. This proves the functional identity on the entire union. If two such models have the same observed law, they have the same two expectations in the ratio, whose denominator is at least $p_0$, so their targets coincide. For the factor example, substituting $P-g(X)=F\ell$ into (4) gives $h_*=g_0+f^\mathsf T\ell$, and the stated error condition proves (3). ∎

Finite-sample inference on the compatible union

Let $\overline A,\overline D$ be the sample means. If $\overline D\geq p_0/2$, set $\widehat\tau$ equal to $\overline A/\overline D$ clipped to $[-2M,2M]$; otherwise use zero. Put $H=2M/\epsilon$ and, for $\alpha\in(0,1)$, define \[ r_A=H\sqrt{\frac{2\log(4/\alpha)}n},\qquad r_D=\sqrt{\frac{\log(4/\alpha)}{2n}},\qquad r=\frac{2(r_A+2Mr_D)}{p_0}. \] If $\overline D\geq p_0/2$, use $C=[\widehat\tau-r,\widehat\tau+r]\cap[-2M,2M]$; otherwise let $C=[-2M,2M]$.
Theorem (Uniform confidence and consistency): For every $n$ and every model in the declared union, $\Pr(\tau\in C)\geq1-\alpha$. For fixed $M,\epsilon,p_0$, $\widehat\tau$ is uniformly consistent, and ${\mathbb{E}}{\operatorname{diam}} C=O(n^{-1/2})$ for fixed $\alpha$.
Proof: The score in (5) satisfies $|A|\leq H$ because $|Y-h_*|\leq2M$ and $e_* /(1-e_*)\leq1/\epsilon$. Hoeffding's inequality, proved by the bounded tilted-variance argument, and a union bound give probability at least $1-\alpha$ to \[ |\overline A-{\mathbb{E}} A|\leq r_A,\qquad |\overline D-p|\leq r_D. \] On this event and when $\overline D\geq p_0/2$, the common functional identity yields \[ \left|\frac{\overline A}{\overline D}-\tau\right| \leq\frac{|\overline A-p\tau|+|\tau|\,|p-\overline D|} {\overline D} \leq\frac{2(r_A+2Mr_D)}{p_0}. \] Clipping cannot increase distance to $\tau\in[-2M,2M]$. On the other branch the whole target range is reported, proving coverage. Moreover $\Pr(\overline D<p_0/2)\leq\exp(-np_0^2/2)$, since $p\geq p_0$. The diameter is at most $2r$ on the first branch and $4M$ on the second, so its expectation is at most $2r+4M e^{-np_0^2/2}$. Uniform consistency follows from the same concentration calculation with any fixed error tolerance: sample-mean deviations of order a fixed positive tolerance have exponentially small probabilities uniformly in the stated bounded class. ∎

This is deliberately an independent-unit sampling theorem, with $T,r$ fixed or supplied and the bridge and working design treated as inputs. It is not a central limit theorem for a dependent array with estimated factor rotations or an unknown factor rank.

Proposition (Explicit sensitivity enlargement): Allow departures from both models while preserving the boundedness and overlap assumptions. If \[ \|\mu_1-\mu_0\|_\infty\leq d_0,\quad \|e-e_*\|_{L^2(P_U)}\leq d_e,\quad \|\mu_0-h_*\|_{L^2(P_U)}\leq d_h, \] then the same procedure, with radius $r+2[d_0+d_ed_h/\epsilon]/p_0$ on its first branch, has coverage at least $1-\alpha$ for $\tau$.
Proof: Equation (6), the first sup-norm bound, and Cauchy--Schwarz imply $|{\mathbb{E}} A-p\tau|\leq p d_0+d_ed_h/\epsilon$. On the concentration event used above, the numerator in the ratio-error bound is now at most $r_A+2Mr_D+pd_0+d_ed_h/\epsilon$. Since $p\leq1$ and the first branch has $\overline D\geq p_0/2$, the extra ratio error is at most $2[d_0+d_ed_h/\epsilon]/p_0$. On the second branch the full target range is reported, so the same concentration event proves coverage in all cases. ∎

The first unresolved step

The negative example and positive theorem address different declared classes; the latter does not repair the incompatible union by notation. It imposes a common temporal bridge or a correct observed selection design, which supplies the shared observable functional. To extend the result to estimated bridges, independent training and explicit uniform bounds on bridge and propensity errors would be needed; (6) identifies the relevant bias terms. General noisy factor histories may fail (3), and approximation of the conditional confounding term is then a substantive identification restriction rather than a routine nuisance-estimation rate. Neither intersection efficiency nor optimal inference with estimated rank is established here.

References

  • [survey] D. Arkhangelsky and G. Imbens (2024), Causal Models for Longitudinal and Panel Data: A Survey, arXiv:2311.15458v3, 25 June 2024. Section 10.4, Bridging Unconfoundedness and the TWFE Approach, printed p. 71; Section 7.3.2, Augmented Synthetic Control. https://arxiv.org/pdf/2311.15458v3. Published in The Econometrics Journal 27(3), C1--C61, https://doi.org/10.1093/ectj/utae014.

Completion Estimate

40%

Go to contribution →