Can Repeated Cross-Sections Replicate Panels? Optimal Synthetic Cohorts for Complex Survey Econometrics
Year: 2026
Type: article
Abstract: For over four decades since Deaton (1985), empirical econometricians utilizing repeated cross-sections have relied on subjective demographic binning to construct pseudo-panels, inducing severe finite-sample errors-in-variables bias and uncontrolled intra-cohort heterogeneity. This paper introduces Dynamic Optimal Synthetic Cohorts (DOSC), an endogenous statistical learning framework that discovers persistent latent population profiles across mixed structural covariates, preserves longitudinal identity via temperature-scaled Mahalanobis projection, and incorporates complex survey sampling weights. DOSC endogenously optimizes cohort cardinality by minimizing a multi-objective loss function subject to Kish effective sample size constraints. Across an 11-level progressive stress battery (L0 to L10), DOSC recovers the true latent partition with an Adjusted Rand Index of up to 0.9998 and triggers active diagnostic rejection when profiles collapse. In factorial Monte Carlo benchmarks, DOSC reaches near-oracle efficiency in the large-sample environment (IER = 0.913, versus 0.171 for classical Deaton) and dominates independent clustering, which fails completely. Finally, empirical validation on the official Peruvian National Household Survey (ENAHO 2021–2025 balanced panel) demonstrates that DOSC (G* = 28) recovers 85.93% of the true nominal fixed effects parameter (beta = +0.07193 vs. beta* = +0.08371) under double-blind physical identifier purging, successfully passing two one-sided tests for statistical equivalence (p = 0.0494). Decennial application on 262,410 youth observations (2016–2025) reduces standard-error inflation from 141.70x (Deaton) to 63.10x while recovering 84.60% of the microdata parameter, providing a reproducible bridge between machine learning and observational survey econometrics. Keywords: Econometrics, Pseudo-Panels, Synthetic Cohorts, Repeated Cross-Sections, Causal Inference, Deaton Estimator, ENAHO, Microdata. Working Paper / Preprint Version.
Institution: Universidad Tecnológica del Perú
Language: English
Subfield: Statistics and Probability
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235708911
