TL;DR
This paper introduces a new method for estimating structural economic models where the relationship between variables is complex and non-linear, particularly when dealing with unobserved disturbances. They develop a nonparametric instrumental variable approach that allows for these nonseparable disturbances, offering a way to identify and estimate causal effects in situations where traditional linear models fall short.
What Problem Does This Paper Address?
In econometrics, understanding causal relationships is crucial. Often, economic models aim to capture a "structural relationship" – how one variable directly influences another. However, real-world economic phenomena are rarely simple. Unobserved factors, or "disturbances," can influence outcomes in ways that are not easily separated from the main relationship (nonseparable disturbances). Furthermore, these disturbances might be correlated with the variables we observe, a problem known as endogeneity.
Traditional methods often assume linear relationships and separable disturbances, which can lead to biased estimates when these assumptions are violated. This paper tackles the challenge of estimating these structural models when the disturbance is nonseparable and potentially endogenous, and when we don't want to impose strong functional form assumptions (nonparametric). Specifically, they focus on models where the structural function is monotonic in the nonseparable disturbance, which leads to conditional quantile restrictions. The core problem is how to identify and estimate these models reliably in the presence of these complexities.
Methods
The authors propose a nonparametric instrumental variable (IV) estimation approach. The key innovation lies in how they handle the nonseparable disturbance and endogeneity.
- Conditional Quantile Restrictions: They leverage the monotonicity assumption to derive conditional quantile restrictions. These restrictions relate the quantiles of the outcome variable to the instruments, providing a way to link observed data to the underlying structural model.
- Local Identification Conditions: They develop conditions for "local identification," meaning that the model parameters can be uniquely determined in a neighborhood around the true values. This is a crucial step before estimation can be considered. They introduce a "modified completeness condition" as a sufficient condition for this local identification.
- Nonparametric Minimum Distance Estimator: For estimation, they propose a nonparametric minimum distance estimator. This estimator works by minimizing the sum of squares of predicted values obtained from a nonparametric regression of the "quantile residual" on the instruments. This can be thought of as a nonlinear two-stage least squares approach, but adapted for residuals that are not continuously differentiable with respect to the parameters.
- Series Approximations: To make the estimator operational, they use series approximations for both the unknown structural function and the conditional expectation estimator. This allows them to handle the nonparametric nature of the problem.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Monotonicity is Key: The assumption that the structural function is monotonic in the nonseparable disturbance is critical. It allows for the derivation of conditional quantile restrictions, which are the foundation for both identification and estimation.
- Completeness for Identification: The paper highlights the importance of "completeness" conditions for identifying the model parameters. A modified completeness condition is shown to be sufficient for local identification, extending existing work.
- Nonlinear IV is Tractable: They demonstrate that even with complex, nonseparable disturbances and nonlinear relationships, a nonparametric IV approach can be developed and shown to be consistent. This broadens the applicability of IV methods.
- Bridging Parametric and Nonparametric: The proposed estimator can be seen as an extension of conditional quantile estimation to handle endogeneity, and it also relates to nonlinear parametric IV estimation, offering a bridge between these approaches.
Key Results
- Sufficient Conditions for Local Identification: The paper provides theoretical conditions under which the structural model is locally identified. These conditions involve aspects of completeness of the conditional distribution of the data given the instruments.
- Consistency of the Estimator: They prove that the proposed nonparametric minimum distance estimator is consistent. This means that as the sample size grows, the estimator converges to the true, unknown parameters of the model.
- Weaker Conditions: The consistency results are shown to hold under conditions that are weaker than those required by some previous work, particularly regarding the smoothness of the unknown functions.
Limitations and Future Work
- Local vs. Global Identification: The paper focuses on local identification. While important, global identification (uniqueness of parameters over the entire parameter space) is a stronger and often more desirable property, and the paper notes that global identification conditions are harder to check.
- Convergence Rates and Asymptotic Normality: The paper establishes consistency but does not derive convergence rates or asymptotic normality for the estimator. The authors suggest that these results would be of interest and might be affected by the "ill-posed inverse problem" nature of the estimation.
- Further Restrictions: The paper mentions that faster convergence rates might be achieved with further restrictions on the model or distributions, which could be an avenue for future research.
- Other Restrictions: The authors acknowledge that other restrictions implied by the model might be useful for estimation and suggest this as a topic for future work.
Summary updated