[Paper Review] Estimating Heterogeneous Effects: Applications to Labor Economics

Stéphane Bonhomme, Angela Denis|arXiv (Cornell University)|Apr 1, 2024

Labor market dynamics and wage inequality13 citations

TL;DR

This paper presents a unified normal random coefficients framework to estimate and analyze heterogeneous effects across units (e.g., neighborhoods, firms, workers) and discusses moment conditions, estimation strategies, and predictors under high-dimensional settings.

ABSTRACT

A growing number of applications involve settings where, in order to infer heterogeneous effects, a researcher compares various units. Examples of research designs include children moving between different neighborhoods, workers moving between firms, patients migrating from one city to another, and banks offering loans to different firms. We present a unified framework for these settings, based on a linear model with normal random coefficients and normal errors. Using the model, we discuss how to recover the mean and dispersion of effects, other features of their distribution, and to construct predictors of the effects. We provide moment conditions on the model's parameters, and outline various estimation strategies. A main objective of the paper is to clarify some of the underlying assumptions by highlighting their economic content, and to discuss and inform some of the key practical choices.

Motivation & Objective

• Clarify the economic content behind assumptions used to identify heterogeneous effects in settings with many unit-specific covariates.
• Propose a normal random coefficients model to recover means, variances, and distributions of unit-specific effects.
• Outline estimation strategies and moment conditions for high-dimensional, networked data such as neighborhoods, firms, and workers.
• Discuss higher-order moments, distributions, and optimal predictors of unit-specific effects under normality.
• Highlight practical choices and potential relaxations of standard assumptions in heterogeneous effects analysis.

Proposed method

• Adopt a linear model with normal random coefficients and normal errors to model heterogeneous effects.
• Derive moment conditions for means, variances, and higher-order moments of the random coefficients given covariates Z.
• Express quantities of interest as linear, quadratic, or nonlinear moments of the random coefficients, and show how to compute them under Assumptions 1 or 2.
• Represent the estimation noise as erences between observed unit-specific estimates and true effects, motivating shrinkage/predictor approaches.
• Show how the OLS estimator of the random coefficients is distributed and how to obtain conditional means and variances of the heterogeneous effects.
• Discuss how to handle high-dimensional fixed-effects like neighborhoods, firms, or workers, including differencing and normalization requirements.

Experimental results

Research questions

• RQ1 What are the means, variances, and distributions of unit-specific heterogeneous effects in labor economics settings?
• RQ2 How can one construct informative predictors of heterogeneous effects from noisy, high-dimensional estimates?
• RQ3 What moment conditions and assumptions enable identification of the random coefficients and their distribution under networked or sparse covariate structures?
• RQ4 How do higher-order moments and nonlinear moments of the effects inform about dispersion and sorting patterns across units?
• RQ5 What practical choices and potential relaxations of normality and exogeneity are feasible in this framework?

Key findings

• A unified normal random coefficients framework enables recovery of means, variances, and distributions of unit-specific effects under high-dimensional covariates.
• The model provides closed-form expressions for moments and nonlinear moments, such as skewness, kurtosis, and distributional features, under normality.
• Estimation noise in high-dimensional settings can bias simple plug-in measures, motivating shrinkage predictors that minimize expected squared errors.
• The framework links to well-known literatures (mixed models, correlated random effects, empirical Bayes) and clarifies identifiability and exogeneity requirements.
• Practical specification choices are discussed, including handling of large x and z matrices, within-group transformations, and the potential relaxation of normality.