TL;DR
This paper introduces the "group Shapley value," a novel method for interpreting counterfactual simulations in structural economic models. It quantifies the importance of different parameter groups, providing an interpretable "importance table" that sums to one, akin to regression tables. This approach offers a robust way to understand how changes in model components drive outcomes, even when data is incomplete.
What Problem Does This Paper Address?
Structural economic models are powerful tools for understanding complex systems, but interpreting the results of counterfactual simulations can be challenging. When parameters are changed to explore "what-if" scenarios, it's difficult to disentangle the individual contributions of different groups of parameters to the overall outcome. Traditional methods like ceteris paribus analysis (holding all but one factor constant) can be problematic because contributions might not sum to one, and they don't directly compare different sets of parameters. This paper addresses the need for a principled, interpretable, and axiomatically sound method to attribute changes in model outcomes to specific groups of parameters in counterfactual simulations.
Methods
The core contribution is the group Shapley value, a variant of the Shapley value from cooperative game theory. Instead of attributing value to individual players, it attributes value to groups of parameters. The authors show that this group Shapley value can be characterized as the solution to a constrained weighted least squares (CWLS) problem. This formulation is particularly useful because it allows for robust decomposition methods when some inputs (i.e., specific counterfactual outcomes) are missing. They also extend this by proposing Shapley bounds (SLB, SUB) and Shapley Minimum Norm Solutions (SMNS) to handle situations where not all necessary counterfactual simulations can be computed.
Datasets
The paper uses data from two published studies for its empirical applications:
- David and Venkateswaran (2019): Data on capital misallocation in China and the U.S.
- Coşar et al. (2016): Establishment-level data from Colombia for analyzing trade barriers, tariffs, and firing costs.
No new datasets are introduced.
Key Insights and Ideas
- Interpretable Counterfactuals: The group Shapley value provides a way to create "importance tables" for counterfactual simulations that are as easy to understand as regression tables, with contributions summing to 1.
- Axiomatic Foundation: By preserving the core axioms of the Shapley value (Linearity, Dummy, Symmetry, Efficiency), the group Shapley value offers a theoretically sound way to attribute contributions.
- Robustness to Missing Data: The formulation as a constrained least squares problem allows for the development of methods (bounds and minimum norm solutions) to handle scenarios where some counterfactual outcomes are unobservable or too costly to compute.
- Beyond Local Sensitivity: Unlike traditional derivatives, the Shapley value can handle non-differentiable functions and provides a global sensitivity analysis, making it suitable for complex, nonlinear structural models.
Key Results
The paper demonstrates the utility of the group Shapley value through two main applications:
- Simple Roy Model: They apply the method to a Roy model to quantify the importance of different parameter groups (coefficients for human capital, sector-specific benefits, and variances) in explaining changes in income inequality. The results show that specific parameter groups, like those related to sector-specific returns to wages, can be major drivers of inequality changes.
- Capital Misallocation (David & Venkateswaran, 2019): By comparing parameter estimates between China and the U.S., they decompose the change in capital misallocation. The analysis highlights that parameters related to distortion (specifically, the parameter $\gamma$) and the variance of permanent shocks are the most significant contributors to the increase in misallocation when moving from U.S. to Chinese parameters.
- Globalization (Coşar et al., 2016): They apply the method to analyze the impact of trade reforms on various economic aggregates. Notably, they demonstrate how to use Shapley bounds and minimum norm solutions when some counterfactual outcomes are missing. For real income growth, the reduction in iceberg trade cost is identified as the most significant factor.
Limitations and Future Work
- Comparing Two Sets of Parameters: The current framework primarily focuses on comparing two sets of parameters. The authors suggest extending the approach to cases involving more than two sets of parameter values (e.g., comparing multiple countries simultaneously) would be valuable.
- Computational Complexity: The numerical examples are limited to small-scale problems. For larger models, computational challenges associated with solving the constrained least squares problem will become more significant. Developing efficient sampling techniques and accounting for sampling errors will be crucial for scaling up the methodology.
- Causal Interpretation: While the paper focuses on counterfactual simulations, the authors acknowledge ongoing debates about the causal interpretation of Shapley values, particularly when covariates are dependent. This debate is noted as irrelevant for their specific application to counterfactual simulations but remains an area of broader discussion.
Summary updated