TL;DR
This paper introduces a new statistical framework for performing inference on "best linear approximations" of functions that are only partially identified (known to lie within a band). This method is highly versatile, extending existing partial identification techniques to handle more complex scenarios, including cases where the bounding functions themselves need to be estimated.
What Problem Does This Paper Address?
In many real-world scenarios, especially in economics and social sciences, we don't have exact knowledge of a function of interest. Instead, we might only know that its values lie within a certain range or "band." This is known as partial identification. While previous work has developed methods to estimate parameters in such partially identified models, they often struggle with:
- Estimating the Bounds: The functions defining the upper and lower bounds of the band might not be known and may need to be estimated from data, adding a layer of complexity.
- Flexibility: Existing methods are often tailored to specific problems (like mean regression) and may not easily extend to other areas like quantile regression or sample selection problems.
- Complexity of the Identified Set: The set of possible parameter values can be complex, making direct interpretation and inference difficult.
This paper tackles these challenges by developing a unified approach to perform inference on "best linear approximations" of these partially identified functions, offering a more flexible and tractable way to analyze such data.
Methods
The core of the paper's contribution lies in its novel approach to inference using the support function of the identified set of parameters. Here's a breakdown of the key techniques:
- Support Function: For a convex set (like the set of possible parameter values), its support function provides a way to characterize the set by its extreme points. This is a standard tool in convex geometry, but its application to statistical inference in this context is novel.
- Plug-in Estimation: The authors propose estimating the support function by replacing population moments with their sample analogs and using estimators for the bounding functions (which can be parametric or non-parametric series estimators).
- Asymptotic Theory: They develop limit theory for the estimated support function, showing it approximately converges to a Gaussian process. This is a crucial step for constructing confidence intervals and performing hypothesis tests.
- Bayesian Bootstrap: Due to the complexity of the covariance functions involved in the Gaussian process approximation, the paper advocates for and establishes the validity of a Bayesian bootstrap procedure for practical inference. This allows for robust confidence interval construction.
- Handling Discrete Covariates: A technical contribution is the development of a method to handle discrete covariates, which previously complicated inference in this area.
Datasets
The paper uses the Current Population Survey (CPS) data for an empirical application, specifically revisiting the analysis of the gender wage gap by Mulligan and Rubinstein (2008). The authors do not introduce new datasets or benchmarks.
Key Insights and Ideas
- Support Function as a Unifying Tool: The support function elegantly transforms a complex, potentially high-dimensional identified set into a more manageable object for statistical inference. This allows for a unified treatment of various partially identified models.
- Flexibility in Estimating Bounds: The ability to estimate the bounding functions non-parametrically (using series estimators) significantly broadens the applicability of the framework to situations where the exact form of the bounds is unknown.
- Bayesian Bootstrap for Practicality: The reliance on the Bayesian bootstrap for inference is a practical advantage, as it bypasses the need to estimate complex covariance functions directly, making the method more accessible to practitioners.
- Generalizability: The framework is shown to be applicable to a wide range of problems, including quantile regression, distribution regression, sample selection, and treatment effects, highlighting its versatility.
Key Results
- The paper provides a theoretical framework for conducting inference on best linear approximations to set-identified functions.
- It establishes the approximate convergence of the estimated support function process to a Gaussian process.
- The validity of the Bayesian bootstrap for constructing confidence intervals and performing hypothesis tests is proven.
- The empirical application to the gender wage gap suggests that, without strong functional form assumptions or instruments, it's difficult to draw definitive conclusions about changes in the gap over time when accounting for selection. The bounds are often wide, indicating limited information in the data for certain scenarios.
Limitations and Future Work
- Conservatism with Discrete Covariates: While the paper provides a method for discrete covariates, it acknowledges that the resulting confidence sets might be slightly conservative.
- Instrument Strength: The empirical example highlights that the informativeness of the bounds is heavily dependent on the strength of the instruments used. Weak instruments lead to wide bounds and inconclusive results.
- Future Work: The paper suggests that the framework can be extended to handle more complex scenarios, such as non-linear approximations and more general identification regions. The authors also point to the need for further research on improving the power of tests in certain situations.
Summary updated