we-solve (beta)

Econ-518 · economics · Economics

Returns to education: mechanisms and heterogeneity

LLM status Open unverified

Statement

Original atlas ID: L1 (entry 21). Classification: Editorial causal-mechanism identification and transport problem.

Original atlas question: How much of education’s return reflects skills, signaling, networks or selection, and for whom?

Definitions and assumptions

Fix a population, a schooling margin, and an observation horizon. Let $D\in\{0,1\}$ denote completing the specified educational program, $X$ denote pretreatment characteristics, and $Y_i(d)$ denote individual $i$'s discounted lifetime earnings under schooling status $d$, with prices and other people's schooling held fixed. Assume integrability and consistency. Program identity includes institution, field, duration, and credential; one additional year is not automatically the same treatment in different programs. Define the conditional earnings effect $\tau(x)=\mathbb E[Y_i(1)-Y_i(0)\mid X_i=x]$.

For a mechanism decomposition, specify a modular structural causal model with channels $K=\{h,c,n,r\}$: skills, credential information, networks, and remaining direct pathways. Let $Y_i(S)$ be earnings when exactly the channels in $S\subseteq K$ transmit the schooling change, according to explicitly defined channel interventions. Require $Y_i(\varnothing)=Y_i(0)$ and $Y_i(K)=Y_i(1)$. Such interventions are additional model assumptions, not quantities defined by observing a mediator.

Formal research question

For each channel $k$, identify or sharply bound \[ \phi_k(x)=\sum_{S\subseteq K\setminus\{k\}} \frac{|S|!(|K|-|S|-1)!}{|K|!} \mathbb E[Y_i(S\cup\{k\})-Y_i(S)\mid X_i=x]. \] These Shapley allocations satisfy $\sum_{k\in K}\phi_k(x)=\tau(x)$ and allocate interaction effects by an explicit convention. Given an observed law $P$ and an announced model class $\mathcal M$, the identification target is the set of $(\tau,\phi)$ generated by models in $\mathcal M$ consistent with $P$. Determine which additional experiments or restrictions reduce this set and permit transport to a second population.

Interpretation and scope

Selection is a difference in who takes schooling, not a causal transmission channel. An instrumental variable identifies a complier effect only under independence, relevance, exclusion, and monotonicity, and generally does not identify the displayed mechanism allocation. Mediator measurements need intervention or structural restrictions. A causal earnings effect is also different from an internal rate of return: tuition, foregone earnings, taxes, uncertainty, and discounting must enter an investment calculation. General-equilibrium credential dilution requires a separate enrollment-policy model.

Source audit and status

Sources [S1]--[S3] verify the cited education literature and already provide substantial causal and conceptual results. The joint mechanism-and-transport target above is an editorial formulation. Neither the bibliography nor a local schooling estimate establishes a universally unresolved decomposition, or certifies its status throughout 2026.

References

    },leftmargin=*]
  1. Joshua D. Angrist and Alan B. Krueger (1991). Does Compulsory School Attendance Affect Schooling and Earnings?. The Quarterly Journal of Economics 106(4), 979--1014. https://doi.org/10.2307/2937954. Locator: Primary publisher, working-paper, or author record: bibliographic information and abstract; full-text theorem verification is not claimed. Role: Compulsory-schooling instrumental-variable evidence; motivates the distinction between local returns and mechanisms.
  2. David Card (1999). The Causal Effect of Education on Earnings. Handbook of Labor Economics 3A, chapter 30, 1801--1863. https://doi.org/10.1016/S1573-4463(99)03011-4. Locator: Primary publisher, working-paper, or author record: bibliographic information and abstract; full-text theorem verification is not claimed. Role: Survey of causal schooling effects and heterogeneous returns; background for extrapolation.
  3. James J. Heckman, Lance J. Lochner, and Petra E. Todd (2006). Earnings Functions, Rates of Return and Treatment Effects: The Mincer Equation and Beyond. Handbook of the Economics of Education 1, chapter 7, 307--458. https://doi.org/10.1016/S1574-0692(06)01007-5. Locator: Primary publisher, working-paper, or author record: bibliographic information and abstract; full-text theorem verification is not claimed. Role: Distinguishes earnings effects from investment rates of return; conceptual source for the mathematical correction.

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-518-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-518

Education returns: exact interaction allocations and identification failures

Problem ID: I26-514. Status: Unresolved. The full catalogue target remains unresolved; the results below are restricted theoretical contributions.

Target and modeling commitments

Fix the catalogue's educational program, baseline population and covariates $X$, partial-equilibrium prices, and discounted lifetime earnings. The total effect is $\tau(x)=E[Y(1)-Y(0)\mid X=x]$. The mechanism target additionally defines sixteen modular interventions $Y(S)$, $S\subseteq K=\{h,c,n,r\}$, transmitting skills, credential information, networks, and remaining pathways. Their existence is a substantive structural assumption: setting a recorded skill score is not automatically an intervention on the skill-generating mechanism.

Write $v_x(S)=E[Y(S)\mid X=x]$. The catalogue allocates changes using the Shapley formula. This note derives its interaction content, characterizes an exact nonidentification example even under randomized schooling, and proves a transport sensitivity bound. The primary education literature cited below motivates heterogeneous causal returns and distinctions from investment returns. The algebra here is self-contained; it neither attributes a new theorem to those references nor claims an empirical decomposition.

Interactions and the allocation convention

Suppress $x$. Define the interaction coefficients by \[ \theta_T=\sum_{R\subseteq T}(-1)^{|T|-|R|}v(R),\quad T\subseteq K. \]
Proposition: One has $v(S)=\sum_{T\subseteq S}\theta_T$, and the Shapley mechanism contribution is \[ \phi_k=\sum_{T\ni k}\frac{\theta_T}{|T|}. \] Consequently $\sum_k\phi_k=v(K)-v(\varnothing)$, including when interactions or contributions are negative.
Proof: In the sum $\sum_{T\subseteq S}\theta_T$, the coefficient of $v(R)$ is $\sum_{R\subseteq T\subseteq S}(-1)^{|T|-|R|}=(1-1)^{|S\setminus R|}$, which is one only for $R=S$. To establish the contribution formula, represent the Shapley weights by a uniform random permutation of the four channels: the probability that exactly $S$ precedes $k$ is $|S|!(3-|S|)!/4!$. Adding $k$ changes an interaction term $\theta_T$ precisely when $k\in T$ and every other member of $T$ precedes $k$. Among the $|T|$ equally likely last members of $T$, this occurs with probability $1/|T|$. Linearity of expectation proves the formula. Summing over $k$ counts each nonempty interaction once, proving efficiency. ∎
For example, let an earnings premium of $8$ arise only when both skill and a credential are present. Then $v(S)=v(\varnothing)+8{\mathbf 1}\{h,c\subseteq S\}$, so $\phi_h=\phi_c=4$ and the other two contributions are zero. It would be incorrect to call the return "entirely skills" simply because removing skills destroys the premium: removing the credential also destroys it. A pure four-channel complementarity of $8$ allocates $2$ to each channel. These numerical allocations are consequences of a specified convention, not separately observed dollar flows.

A sharp mechanism failure under randomized schooling

Suppose schooling $D$ is randomized independently of all potential outcomes, and the observed experiment has deterministic earnings $Y=y_0+\tau D$, where $\tau>0$. The total effect is therefore identified exactly. Consider a modular model class allowing any bounded, monotone set function $v$ between $y_0$ and $y_0+\tau$, with those fixed endpoint values. Realize its intervention outcomes deterministically. No intermediate intervention is observed.
Proposition: The identified set of the mechanism vector is exactly \[ \{\phi\in{\mathbb{R}}^4:\phi_k\ge0,\ \sum_k\phi_k=\tau\}. \]
Proof: Monotonicity makes every marginal increment in the Shapley formula nonnegative, and efficiency gives the sum constraint. Conversely take any vector $b\ge0$ summing to $\tau$ and define $v(S)=y_0+\sum_{k\in S}b_k$. This is an admissible monotone modular response, its Shapley vector is $b$, and its endpoint experiment is exactly the same observed law. Every simplex point is therefore attained. ∎
Thus even a perfect schooling experiment leaves open everything from an entirely credential-driven premium to an entirely skill-driven premium within this simple class. This result is stronger than merely observing that an observational earnings regression may suffer selection bias. Randomization removes that bias and still does not expose the channel interventions. The example fixes $X$; repeating it separately in each covariate stratum preserves the conclusion for heterogeneous returns.

Without monotonicity or bounds on intermediate outcomes, the set becomes the entire efficiency hyperplane $\sum_k\phi_k=\tau$. Indeed the same additive construction works with arbitrary real $b_k$ of that sum. Negative components can offset arbitrarily large positive components while leaving both observed endpoints fixed. Finite mechanism bounds consequently require assumptions on unobserved interventions, not just finite observed earnings.

Which experiments actually identify the allocation?

If all sixteen interventions are ethically and operationally well-defined, randomized assignment with positive probabilities identifies each $v_x(S)$ under consistency, conditional randomization and appropriate interference restrictions. The displayed finite linear map then identifies every $\phi_k(x)$. This is an identification sufficiency statement, not a recommendation that one can literally remove a child's networks or skills.

There is also a randomized-permutation estimator. Draw a channel permutation uniformly, sample a predecessor set $S_k$, and randomize between the adjacent interventions $S_k$ and $S_k\cup\{k\}$. Average their mean earnings difference over permutations. The expectation is precisely $\phi_k$. Its feasibility still depends on defining those channel interventions, and its variance depends on earnings dispersion and assignment probabilities. Observing employers' responses to credential disclosure identifies an information intervention holding the studied skill and network distribution fixed. It need not identify the all-population credential Shapley term because that term averages over eight distinct predecessor sets.

An instrument for completing schooling generally changes the population of compliers and the joint bundle of mechanisms. Even a valid exclusion restriction at the schooling level does not create exclusions for individual channels. Local total effects and modular mechanism effects therefore require separate equations and distinct assumptions.

Transport to a second population

Let source and target covariate laws be $F_0,F_1$, with $F_1\ll F_0$. Suppose source experiments identify $v_{0,x}(S)$. If all conditional intervention means are invariant, the target mechanism mean is \[ \Phi_{1,k}=\int\phi_{0,k}(x)\,dF_1(x) =E_0\!\left[\frac{dF_1}{dF_0}(X)\phi_{0,k}(X)\right]. \] This follows directly by change of measure and linearity. It requires population overlap and mechanism invariance, not merely similar observed earnings distributions.

For an explicit sensitivity calculation assume $|v_{1,x}(S)-v_{0,x}(S)|\le\delta_S(x)$ for every subset, with integrable bounds. Let $w_S=|S|!(3-|S|)!/4!$. The triangle inequality gives \[ |\Phi_{1,k}-\int\phi_{0,k}\,dF_1| \le\int\sum_{S\subseteq K\setminus\{k\}}w_S [\delta_{S\cup\{k\}}(x)+\delta_S(x)]\,dF_1(x). \] For a common bound $\delta$, the right side is $2\delta$ because the weights sum to one. This is a valid coordinate bound, not a characterization of the joint identified set: the same sixteen shifts enter multiple coordinates, and their efficiency relation couples the contributions. Regions with no source overlap require new restrictions or target experiments.

Remaining gap

The catalogue asks for a credible joint mechanism and heterogeneity analysis in a substantive schooling setting. That requires actual modular intervention definitions, causal evidence about intermediate channel configurations, and defensible population transport assumptions. This note establishes what endpoint schooling evidence cannot do and what stronger experiments would identify. It supplies no earnings estimates, tuition-cost accounting, general-equilibrium credential response, or empirically justified mechanism shares, so the broad target remains unresolved.

Completion Estimate

50%

This is a post hoc, heuristic estimate for the full catalogue target. The attempt proves the interaction content of education Shapley allocations, sharp endpoint-only mechanism nonidentification, and identification and transport sensitivity under modular experiments. Full heterogeneous returns still require feasible channel interventions, observed intermediate configurations, justified target-population overlap, and empirically validated mechanism invariance.

References

  • [ak] J. D. Angrist and A. B. Krueger (1991), Does Compulsory School Attendance Affect Schooling and Earnings?, QJE 106(4), 979--1014. Working-paper record and abstract: https://www.nber.org/papers/w3572. Journal DOI: https://doi.org/10.2307/2937954.
  • [hlt] J. J. Heckman, L. J. Lochner and P. E. Todd (2006), Earnings Functions, Rates of Return and Treatment Effects: The Mincer Equation and Beyond, Handbook of the Economics of Education 1, 307--458. Primary working-paper record: https://www.nber.org/papers/w11544; published chapter: https://doi.org/10.1016/S1574-0692(06)01007-5. These references support the distinctions among causal earnings effects, heterogeneity and investment returns; no mechanism-share estimate is imported here.

Go to contribution →