TL;DR
This paper introduces a novel statistical method for analyzing distributional effects in complex network and panel data models. It provides a way to construct reliable confidence bands for quantile functions and effects, even in the presence of unobserved two-way effects and discrete outcomes, which are common challenges in real-world data. This advancement is crucial for understanding the full impact of policies or interventions beyond just average effects.
What Problem Does This Paper Address?
In many fields, understanding the impact of a policy or intervention isn't just about its average effect on an outcome (like income or test scores). It's equally important to know how it affects different parts of the distribution – for example, does a new education program help low-achievers more than high-achievers? This requires analyzing quantile effects. However, traditional methods struggle when dealing with:
- Panel Data: Data collected over time for the same individuals or entities. This introduces unobserved heterogeneity (individual-specific traits) and time-specific shocks.
- Network Data: Data where relationships between entities matter (e.g., trade between countries, social connections). This introduces sender and receiver effects.
- Unobserved Two-Way Effects: In both panel and network data, there are often unobserved factors that influence outcomes and are correlated with observed variables, leading to endogeneity and omitted variable bias.
- Discrete Outcomes: When the outcome variable can only take on a limited number of values (e.g., yes/no, purchase/no purchase).
Existing methods often fail to provide reliable inference (like confidence intervals) in these complex settings, especially when trying to understand effects across the entire distribution. This paper aims to fill that gap by providing a robust framework for distributional and quantile effect analysis in such models.
Methods
The core of the paper's contribution lies in developing a method to construct simultaneous confidence bands for quantile functions and effects. This is achieved through a multi-step process:
- Distribution Regression (DR) with Fixed Effects: They extend the distribution regression framework to handle unobserved two-way effects (individual and time in panels, or sender/receiver in networks). This is done by treating these unobserved effects as fixed parameters to be estimated.
- Addressing the Incidental Parameter Problem: In nonlinear models with fixed effects, estimating these effects can lead to biased estimates, especially when the number of time periods or individuals grows. The authors adapt analytical bias correction techniques to debias these fixed-effects DR estimators.
- Constructing Confidence Bands for Distributions: They build simultaneous confidence bands for the estimated counterfactual distributions.
- Inverting Bands for Quantiles: Using a generic method, they "invert" the confidence bands for the distributions to obtain confidence bands for the quantile functions and, subsequently, for the quantile effects. This is a key step that allows for inference on quantiles.
- Multiplier Bootstrap for Inference: To obtain the critical values for these confidence bands, they employ a multiplier bootstrap scheme. This is computationally convenient and can handle complex data dependencies like pairwise clustering in network data.
Datasets
The paper uses a dataset on gravity equations for international trade. This dataset includes bilateral trade flows for 157 countries in 1986, along with covariates like distance, common colonial ties, currency union, and other country-specific characteristics. This dataset is used for an empirical application to illustrate the method's applicability to network data.
Key Insights and Ideas
- Beyond Averages: The paper emphasizes the importance of understanding distributional effects, not just average effects, for policy analysis.
- Unified Framework: The distribution regression approach provides a flexible way to model outcomes that can be continuous, discrete, or mixed, and to incorporate unobserved heterogeneity.
- Bias Correction is Crucial: For fixed-effects estimators in nonlinear models, bias correction is essential for obtaining reliable inference, especially for coefficients.
- Confidence Bands for Entire Functions: The method provides simultaneous confidence bands, meaning they cover the entire quantile function or effect function over a range of quantiles, not just at a single point. This is more informative than pointwise intervals.
- Network Data Applicability: The framework is shown to be applicable to network data, a growing area of research, by accounting for sender and receiver effects.
Key Results
- The proposed method successfully constructs simultaneous confidence bands for quantile functions and effects in nonlinear network and panel models with unobserved two-way effects.
- The bias-corrected estimators show improved coverage probabilities for confidence bands compared to uncorrected estimators, especially for coefficients.
- The empirical application to gravity models of trade reveals significant heterogeneity in the effects of distance and common legal systems across the distribution of trade volume, which traditional mean-based methods would miss.
- The Poisson model, commonly used in trade literature, is shown to miss important features of the trade volume distribution (heavy tails), leading to poor approximations.
- The multiplier bootstrap method provides a computationally convenient way to obtain valid critical values for inference.
Limitations and Future Work
- Assumptions: The asymptotic theory relies on strict exogeneity assumptions for covariates. The authors suggest extending the theory to predetermined or weakly exogenous covariates as future work.
- Computational Cost: While the multiplier bootstrap is computationally convenient, estimating high-dimensional fixed effects models can still be computationally intensive.
- Extension to Other Models: The authors note that their method can be applied to more complicated models (e.g., with interactive unobserved effects or factor structures) provided that confidence bands for distributions in those models are available. Extending the central limit theorem to such cases is a direction for future research.
- Missing Data: The current framework allows for a finite number of missing observations. Handling more complex missing data patterns might require further development.
Summary updated