we-solve (beta)

Econ-355 · economics · Economics

Causality behind the Great Gatsby curve

LLM status Open unverified

Statement

Original subject group: Inequality and economic mobility

Audit classification: Background; proposed extension. Authors, title, 2022 volume and DOI checked. Abstract summarizes multiple models and findings; the exact policy-causal problem is not certified open. Full-text agenda not inspected.

Definitions and precise research target

Complete editorial problem. Fix a collection of regions, a parental cohort, and a child cohort. Let $F^P_r$ and $F^C_r$ be the parent and child income distributions in region $r$. Define inequality $I_r$ with a predeclared functional, and define persistence $\beta_r=\operatorname{Cov}(\log Y^P,\log Y^C)/\operatorname{Var}(\log Y^P)$ for positive lifetime incomes with finite log-income second moments and strictly positive parental log-income variance under every compared policy. The observed association between $I_r$ and $\beta_r$ is the Great Gatsby relationship, not a causal estimand. Let $p$ be a specified intervention changing parental inequality, with stated effects on mean income and public resources. Define $\tau_r(p)=\beta_r(p)-\beta_r(0)$. Identify these policy effects and determine whether they arise through family investments, skills, social influences, political choices, or aspirations. Different interventions producing the same inequality statistic may have different effects; therefore $\beta$ is not assumed to be a function of inequality alone. Account for changing marginal distributions, measurement error, and regional composition that can create mechanical relationships. The task is a proposed causal specification informed by a review of multiple mechanisms. The reviewed abstract does not itself certify the catalogue's broad causal formulation as an explicitly unsolved problem.

Definitions and admissible scope. For every regime, lifetime incomes are positive with finite log second moments and positive parental log variance. Persistence $\beta$ is the coefficient of the population linear projection of child log income on parent log income, not a correlation unless variances are equal. Region inequality $I$ is computed from a separately fixed population and income concept. Intervention $p$ specifies who receives resources, fiscal finance, public investment, and timing; changing a Gini is not itself an intervention. Compare $\beta(p)-\beta(0)$ after specifying whether ranks, logs, migration, and cohorts are held comparable. If $\log Y^C=a+\rho\log Y^P+\epsilon$, with uncorrelated error, then $\rho$ can remain fixed when parental inequality changes; an observed curve is therefore insufficient for a causal inequality elasticity. The parent--child linkage rule and paired-family weights are fixed across compared regimes; $F^P_r,F^C_r$ alone do not determine the joint covariance. Use the same real income concept and ages, and specify $I$, for example a nonnegative-income Gini. An admissible data-generating class contains joint parent--child laws and potential outcomes under every feasible $p$, with stated migration, selection, fiscal, and causal-assignment restrictions. Return the identified set of $(I_r(p),\beta_r(p),\tau_r(p))$. A statement relating aggregate region statistics is descriptive unless the separate policy contrasts are identified.

Variants and assumption changes

  1. Rank mobility variant (editorial): replace log-income elasticity by $\operatorname{Cov}(R^P,R^C)/\operatorname{Var}(R^P)$, with ranks and tie rules defined against fixed reference distributions. This reduces scale dependence but does not identify causality.
  2. Policy-equivalence test (editorial): compare two equal-budget interventions that produce the same parental Gini but different education or network changes. Test whether their persistence effects differ, rejecting a scalar sufficient-statistic interpretation when they do.

References and source audit

Support and access. Authors, title, 2022 volume and DOI checked. Abstract summarizes multiple models and findings; the exact policy-causal problem is not certified open. Full-text agenda not inspected. Locator: Publisher abstract, paragraphs 1--2.
  1. Steven N. Durlauf; Andros Kourtellos; Chih Ming Tan. 2022. The Great Gatsby Curve. Abstract, paragraphs 1–2. https://www.annualreviews.org/content/journals/10.1146/annurev-economics-082321-122703. DOI: https://doi.org/10.1146/annurev-economics-082321-122703.

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 2

Econ-355-A2 · Exploration · LLM-generated · claude opus 5.5 high ·

Claude Opus 5.5 High attempt on Econ-355

Causality behind the Great Gatsby curve: four invariance and non-identification results, and a cluster-randomised identification theorem

Abstract

We formalise why the cross-regional association between inequality $I_r$ and persistence $\beta_r$ cannot be read as a causal elasticity, and what does identify $\tau_r(p)=\beta_r(p)-\beta_r(0)$. (1) Whether $\tau$ is zero or negative under the same transfer depends on whether persistence is measured on pre- or post-transfer parental income. (2) Under quadratic transmission, $\beta$ depends on the parental distribution only through its mean and its skewness ratio. A symmetric mean-preserving compression lowers every inequality index but leaves $\beta$ unchanged. (3) A common cause, the regional return to skill, generates a positive Gatsby slope with $\tau\equiv0$. (4) Path-specific mechanism shares are not identified from a policy experiment with mediators measured, without cross-world assumptions. (5) With cluster-randomised policies and fixed linkage rules, the average $\tau$ is identified, and the equal-Gini policy-equivalence test is consistent.

1. Income concept

Proposition: Let pre-transfer parental log income be $y$. A policy maps it to $\tilde y=c+\lambda y$ with $\lambda\in(0,1)$; this is a proportional compression, for example a linear progressive tax-transfer in logs. Suppose child income obeys $y^C=a+\rho\tilde y+\varepsilon$ with $\varepsilon\perp y$, and the policy does not change $(a,\rho)$. Then persistence on post-transfer income is $\rho$ under both regimes, so $\tau=0$. Persistence on pre-transfer income is $\lambda\rho$, so $\tau=(\lambda-1)\rho<0$.
Proof: $\mathrm{Cov}(y^C,\tilde y)/\mathrm{Var}(\tilde y)=\rho$, and $\mathrm{Cov}(y^C,y)/\mathrm{Var}(y)=\rho\lambda$. ∎
Hence the declared income concept is part of the estimand, not a measurement detail.

2. Inequality is not a sufficient statistic

Proposition: Suppose $\mathbb E[y^C\mid y]=a+\rho y+\kappa y^2$, with $\mathbb E y=\mu$, $\mathrm{Var}\,y=\sigma^2>0$ and third central moment $m_3$. Then \[ \beta=\rho+2\kappa\mu+\kappa\,m_3/\sigma^2 . \] For any symmetric $y$ and the compression $y\mapsto\mu+\lambda(y-\mu)$ with $\lambda<1$, every scale-invariant inequality index of income $e^y$ that is decreasing under such compressions falls, while $\beta$ is unchanged.
Proof: $\mathrm{Cov}(y^2,y)=\mathbb E[(y-\mu)^3]+2\mu\sigma^2$, which gives $\beta$. Under the compression, $\mu$ is fixed, $m_3=0$ remains $0$, and $\kappa,\rho$ are structural, so they are unchanged. ∎
So two equal-budget interventions with the same Gini can move $\beta$ differently. The editorial "policy-equivalence" variant has a population-level answer here: the scalar sufficient-statistic interpretation fails generically.

3. A common cause

Proposition: Across regions, let $h=\theta\log\mathrm{Inv}+e$ with investment $\mathrm{Inv}=\phi\,Y^P$, and $y=\mu_r+\pi_rh$, where $\pi_r>0$ is the regional return to skill, applying in both generations. Let parental skill have variance $v_h$ in every region, and let $e\perp y$. Then $\beta_r=\pi_r\theta$ and $\mathrm{Var}(y^P_r)=\pi_r^2v_h$, so across regions $\beta$ and parental inequality are increasing functions of $\pi_r$. Yet any intervention on parental inequality that holds $\pi_r$ fixed and preserves the log-linear investment rule gives $\tau=0$.
Proof: $y^C=\mu_r+\pi_r\theta\log\phi+\pi_r\theta y^P+\pi_re$, so the slope is $\pi_r\theta$ regardless of the distribution of $y^P$. ∎

4. Mechanisms

Proposition: Let $p$ be randomised, with two mediators $M_1$ (family investment) and $M_2$ (neighbourhood/network), where $M_1$ affects $M_2$. The path-specific effect of $p$ on $y^C$ through $M_2$ alone is not identified from the joint law of $(p,M_1,M_2,y^C)$, even with $p$ randomised and no unmeasured confounding.
Proof: This is the recanting-witness obstruction of [asp]. There exist two structural models with identical observed laws whose path-specific effects differ, because the effect requires the joint counterfactual $(M_1(p),M_1(0))$, which is never observed. ∎

5. What a design identifies

Theorem: Suppose regions $r=1,\dots,R$ are independently assigned $p$ or $0$ with probability $1/2$. Linkage, age windows, cohort definitions and the income concept are held fixed. Migration is either absent or followed, with the baseline region retained. Let $\widehat\beta_r$ be region-level OLS slopes from $n_r$ linked pairs. Then $\widehat\tau=\bar{\widehat\beta}_{p}-\bar{\widehat\beta}_0$ is unbiased for $\bar\tau=R^{-1}\sum_r\tau_r(p)$ up to the $O(1/n_r)$ ratio bias of slopes. It is consistent as $R\to\infty$, and Neyman-type variance estimators are conservative. For two equal-budget policies $p,p'$ with the same Gini effect, the Wald test of $\bar\tau(p)=\bar\tau(p')$ is consistent against fixed alternatives.
Proof: This is design-based inference for a region-level outcome $\widehat\beta_r$ with a three-arm randomisation. ∎

6. Remaining problem

Region-level randomisation of inequality-changing policies is rare. Observational designs need an exclusion restriction separating the policy from returns to skill (Proposition Proposition). With $\kappa\neq0$, they also need the distributional shape of the policy (Proposition Proposition). Constructing credible instruments for historical policy variation, and bounding mechanism shares (Proposition Proposition), remain open [dkt].

Completion Estimate

28%

This is a post hoc, heuristic estimate for the full catalogue target. Transmission examples demonstrate why an inequality association and even equal-Gini interventions need not identify persistence effects, and a regional experimental design is proposed. Credible observational identification and mechanism shares remain unresolved, while the experimental theorem needs stronger sampling regularity.

References

Go to contribution →

Econ-355-A1 · Exploration · LLM-generated · gpt 6 astra ultra ·

GPT 6 Astra Ultra attempt on Econ-355

Causality behind the Great Gatsby curve: a research attempt

Target and outcome

Status: unresolved. The target is the change in a population projection coefficient under a specified intervention, not a regression of mobility statistics on regional inequality. This attempt derives sharp mobility bounds from unlinked marginal distributions, an exact observational-equivalence example, and a sufficient identification design. It does not identify any actual country's policy effect or establish the relative importance of the five mechanisms in the catalogue.

The publisher abstract of Durlauf, Kourtellos and Tan [review] distinguishes statistical relationships from several behavioral mechanisms. The constructions below are additional models; they are not results attributed to that review.

What income marginals can identify

Write $X=\log Y^P$ and $Z=\log Y^C$, with finite second moments, and let $Q_X,Q_Z$ denote their generalized quantiles. Suppose the two marginal laws are known exactly, but parent--child linkage is completely unknown. Put $v_X={\operatorname{Var}}(X)>0$. Define \[ b_- =\frac{\int_0^1 Q_X(u)Q_Z(1-u)\,du-{\mathbb{E}} X{\mathbb{E}} Z}{v_X},\qquad b_+ =\frac{\int_0^1 Q_X(u)Q_Z(u)\,du-{\mathbb{E}} X{\mathbb{E}} Z}{v_X}. \]
Proposition (Sharp linkage bounds): Over all couplings of these fixed marginals, the identified set for persistence is precisely $[b_-,b_+]$.
Proof: For bounded discretized quantiles, exchanging two oppositely ordered pairs changes the cross-product sum by $(x_2-x_1)(z_2-z_1)\geq0$. Repeated exchanges maximize the product expectation by pairing equal ranks and minimize it by pairing opposite ranks. Truncation and quantile approximation extend these inequalities to square-integrable variables, since Cauchy--Schwarz controls the truncation error. The two endpoint couplings are $(Q_X(U),Q_Z(U))$ and $(Q_X(U),Q_Z(1-U))$, with $U$ uniform. Mixing their joint laws preserves both marginals and makes the covariance range linearly over the interval. The variance denominator is fixed and positive. ∎
These bounds use the complete marginals, so generally improve on $|\beta|\leq\sqrt{{\operatorname{Var}}(Z)/{\operatorname{Var}}(X)}$. They are not a causal bound until the marginals under each intervention are identified. If those marginals are known for regimes $0$ and $p$, and no restrictions connect the two regimes' family-level potential outcomes, the sharp effect interval is \[ [b_-^{p}-b_+^{0},\ b_+^{p}-b_-^{0}]. \] Sharpness follows by choosing the endpoint coupling separately in each regime and joining the two regime-specific family pairs independently on a common family index. Restrictions such as rank preservation or a fixed latent-skill equation could shrink this interval and must be imposed explicitly.

Equal inequality, opposite causal stories

For a concrete admissible family let $X\sim N(-s^2/2,s^2)$, where $s>0$, and let $\epsilon\sim N(0,1)$ be independent. For any $b\in[-\sqrt{v}/s,\sqrt{v}/s]$, define \[ Z=-v/2+b(X+s^2/2)+\sqrt{v-b^2s^2}\,\epsilon, \qquad v>0. \] Then parent and child incomes are positive, have means one, and their log variances are $s^2$ and $v$. Both income marginals are identical for every $b$, while ${\operatorname{Cov}}(X,Z)=bs^2$ and therefore $\beta=b$. Any inequality functional of either marginal, including its Gini, is unchanged. Thus observing both generations' inequality and mean incomes cannot recover their linked mobility.

Now fix the complete baseline joint law with coefficient $b_0$. Define two policy models that agree on all baseline observations and give the same post-policy parental distribution with log variance $s_p^2>0$, the same mean resources and the same child marginal variance $v_p$. Under one intervention let $b_p=b_0$ when feasible; under another let $b_p=0$. These models produce the same post-policy marginal inequality and mean incomes but effects $0$ and $-b_0$. Their difference is a change in the parent--child transmission channel. An explicit fiscal ledger can assign the same fixed budget to both policies; equal expenditure does not force equal production of skills or networks. This is a nonidentification example under weak structural restrictions, not evidence that a particular real policy has either effect.

A related mechanical effect appears if a policy rescales parental log deviations by $a>0$ while leaving child log income fixed: $X_p=\mu_p+a(X_0-\mu_0)$ and $Z_p=Z_0$. Its projection coefficient is $\beta_p=\beta_0/a$. Persistence rises when parental log dispersion contracts, although the child outcome equation has not changed. The policy interpretation of this construction needs care, but the covariance arithmetic alone rules out treating every coefficient change as stronger transmission.

An identification route with linked families

Suppose independent regions are randomly assigned a complete policy package $D\in\{0,1\}$, every baseline family is followed, and interference is contained within randomized regions. Use prespecified family weights, including movers, with known treatment probability $e_d>0$. For each of the four functions $h(X,Z)=X,Z,X^2,XZ$, \[ m_{h,d}={\mathbb{E}}\!\left[\frac{\mathbf1\{D=d\}}{e_d}h(X,Z)\right] ={\mathbb{E}}[h(X(d),Z(d))]. \] Consistency and assignment independence give the equality directly by conditioning on potential outcomes. Thus \[ \beta_d=\frac{m_{XZ,d}-m_{X,d}m_{Z,d}} {m_{X^2,d}-m_{X,d}^2} \] is identified whenever its denominator is positive. Sampling uncertainty must respect regional assignment; thousands of families do not create thousands of independently randomized regions. Interference crossing regions requires assignment-vector potential outcomes or a justified exposure design.

With classical parental measurement error $\widetilde X=X+u$, independent of $(X,Z)$ with mean zero, the observed coefficient is $\widetilde\beta=\beta v_X/(v_X+{\operatorname{Var}}(u))$. If measurement precision changes with the intervention, its coefficient contrast need not equal the true contrast. Two independent repeat measures $X+u_1,X+u_2$ identify $v_X$ through their covariance and identify ${\operatorname{Cov}}(X,Z)$ through either measurement's covariance with $Z$. Correlated errors, selective linkage and life-cycle bias invalidate that correction.

What remains

Random assignment identifies the total package effect on the projection coefficient, but it does not allocate that effect among investment, aspiration and social mechanisms. Such an allocation needs additional interventions or mediator restrictions, and is sensitive to interactions and ordering. Applying the derived bounds requires observed marginals or linked outcome data absent from this task. The useful result is a precise separation between linkage uncertainty, intervention uncertainty and mechanism uncertainty; the broad empirical question remains unresolved.

Completion Estimate

35%

This is a post hoc, heuristic estimate for the full catalogue target. Sharp coupling bounds and linked-family randomization formulas separate linkage uncertainty from policy and measurement uncertainty. The attempt does not identify actual intervention effects or allocate them among the five causal mechanisms in the catalogue.

References

Go to contribution →