Where Do the Returns to Schooling Come From? Educational Transitions and Labor Market Payoffs

Technical Summary: "Where Do the Returns to Schooling Come From? Educational Transitions and Labor Market Payoffs"

Problem Statement
Conventional research on the returns to education typically relies on two approaches: measuring education as a continuous "years of schooling" variable or dichotomizing attainment as a point-in-time treatment (e.g., high school graduation vs. non-graduation). The author argues that these conceptualizations misalign with the sequential nature of educational transitions, where individuals make a series of dependent decisions (e.g., high school graduation $\to$ college attendance $\to$ degree completion $\to$ graduate school). Existing methods fail to disentangle the direct labor market returns of a specific educational level from the indirect returns mediated by subsequent educational transitions. Consequently, policy interventions cannot be effectively evaluated based on whether they boost outcomes by promoting further education or by providing direct labor market advantages.

Methodological Framework
The paper proposes a causal mediation framework specifically designed for monotonic binary mediators in a sequential setting.

• Monotonicity Assumption: The framework relies on the structural feature that educational transitions are monotonic: an individual cannot complete a higher-level transition (e.g., college graduation) without first completing the preceding one (e.g., high school graduation). Formally, if $M_k = 0$, then $M_{k+1} = 0$. This assumption rules out "defiers" and "always-takers" in the context of educational progression.
• Decomposition of the Average Treatment Effect (ATE): The author derives a decomposition of the ATE of an initial educational transition ($A$) on an outcome ($Y$, e.g., earnings) into $K+1$ mutually exclusive Monotonic Path-Specific Effects (MPSEs):
  1. A Direct Effect ($\Delta_0$): The effect of $A$ on $Y$ net of all $K$ subsequent transitions.
  2. $K$ Continuation (or Gross) Effects ($\theta_k$): The effects operating through specific causal chains (e.g., $A \to M_1 \to Y$, $A \to M_1 \to M_2 \to Y$).
• Identification: Unlike conventional mediation analysis which often requires "cross-world" independence assumptions (ruling out all post-treatment confounders), this framework utilizes Sequential Ignorability. This weaker assumption allows for the presence of observed intermediate confounders ($Z_k$) that affect the relationship between a mediator and the outcome, provided these confounders are measured. The monotonicity constraint ensures that the decomposition is algebraically unique and identifies all causal paths, whereas standard decompositions with multiple ordered mediators often fail to identify pure path-specific effects due to "recanting witness" criteria violations.
• Estimation Strategies: The paper introduces two estimation approaches:
  1. Debiased Machine Learning (DML): A semiparametric estimator utilizing Efficient Influence Functions (EIFs) and cross-fitting. This approach is robust to model misspecification and high-dimensional covariates.
  2. Regression-with-Residuals (RWR): A parametric approach using linear models and sequential residualization. While less robust to misspecification, it offers a transparent link between decomposition components and regression coefficients.

Empirical Application and Results
The framework is applied to the NLSY97 cohort ($N=7,305$) to analyze the returns to high school completion on logged annual earnings (ages 32–36). The analysis decomposes the total effect into:

• Direct effect of high school graduation (net of college).
• Indirect effects via: (1) College attendance without a BA, (2) BA completion without graduate school, and (3) Graduate school attendance.

Key Findings:

• Dominance of Direct Effects: The vast majority (approximately 69–75%) of the total return to high school graduation operates directly, net of subsequent college attendance and degree completion. The estimated direct effect corresponds to an earnings premium of roughly 59–60%.
• Limited Mediation: The indirect effects mediated by college attendance, BA completion, and graduate school are relatively small.
  • The continuation effect via college attendance (without a BA) and via BA completion (without graduate school) each accounts for roughly 15% of the total effect.
  • The pathway through graduate school is negligible and statistically insignificant.
• Mechanism of Small Mediation: The paper attributes the small indirect effects not to low returns on college degrees, but to low counterfactual progression rates. Even if high school graduates were to attend college, the probability of them completing a BA or attending graduate school without further intervention is low. Thus, the "continuation" pathway is constrained by the low probability of traversing the full educational chain.
• Robustness: Sensitivity analyses regarding unobserved confounding suggest that the primary finding—that the returns are overwhelmingly direct—remains robust even under strong assumptions about unmeasured confounders.

Significance and Contributions
The paper makes three primary contributions:

1. Education Research: It provides a nonparametric decomposition of schooling effects that distinguishes between direct labor market returns and returns mediated by further education. This allows for a more nuanced understanding of why education pays off, moving beyond simple "years of schooling" estimates.
2. Causal Mediation Analysis: It extends the literature on path-specific effects (PSEs) by exploiting the monotonicity of sequential transitions. This allows for the identification of all causal paths in a sequence under weaker assumptions (sequential ignorability) than required for general multiple-mediator settings, which often cannot identify pure path effects due to recanting witnesses.
3. Methodological Advancement: It introduces and validates semiparametric estimators (DML) for monotonic path-specific effects, demonstrating their superiority over parametric methods (RWR) in the presence of model misspecification.

The author concludes that the framework is applicable to other domains characterized by state-dependent, monotonic transitions, such as family formation (marriage/divorce), health progression, or criminal justice contact, though the paper focuses its empirical demonstration on education.