Inference for best linear approximations to set identified functions

Arun G. Chandrasekhar, Victor Chernozhukov, Francesca Molinari, Paul Schrimpf

Introduction

In empirical economics and social sciences, researchers frequently encounter situations where a parameter of interest cannot be uniquely identified. Instead, due to missing data, censoring, or interval reporting, the data only suggests that the true function $f(x)$ lies within a specific band defined by a lower bound $\theta_0(x)$ and an upper bound $\theta_1(x)$. This phenomenon is known as partial identification.

This paper provides a robust framework for performing statistical inference on the "best linear approximations" to these set-identified functions. By extending the partial identification literature, the authors allow these bounds to be any functions—including those estimated non-parametrically or carrying an index—making the framework applicable to a wide array of problems including quantile regression, sample selection, and treatment effects.

Methodological Framework

The authors bridge the gap between abstract set theory and practical econometrics through several key innovations designed to make inference tractable and reliable.

Support Functions and Gaussian Processes

To describe the identification region (the set of all possible linear approximations), the paper utilizes support functions. In geometry, a support function describes a convex set by measuring the distance from the origin to its supporting hyperplanes. This approach effectively transforms a set-estimation problem into a function-estimation problem.

The core theoretical contribution is proving that the estimated support function converges to a Gaussian process. This allows researchers to utilize the well-understood machinery of normal distributions to conduct hypothesis tests and construct confidence intervals for the set of parameters.

Practical Implementation: Bootstrap and Jittering

Because the theoretical covariance of these Gaussian processes is often too complex to calculate directly, the authors establish the validity of the Bayesian bootstrap. This provides practitioners with a straightforward, simulation-based method to estimate the distribution of their parameters.

Additionally, the paper addresses the challenge of discrete regressors (such as "years of schooling"). Standard set-identification math often requires continuous variables; the authors introduce a "jittering" technique—adding controlled, infinitesimal noise to discrete variables—to ensure the mathematical properties required for valid inference are maintained.

Versatility and Applications

One of the primary strengths of this framework is its versatility. By allowing the bounds to carry an index ($\alpha$), the methodology extends far beyond simple mean regression:

  • Quantile and Distribution Regression: Researchers can analyze how specific percentiles (e.g., the 10th or 90th) of an outcome change, even when data is interval-valued.
  • Sample Selection: The framework can account for individuals who select out of a sample (e.g., people not in the labor force) without requiring the restrictive functional form assumptions of traditional Heckman-style models.
  • Instrumental Variables: The framework can incorporate instruments to sharpen the bounds of the identification region.

Empirical Application: The Gender Wage Gap

The authors apply their methodology to the study of female labor force participation and the gender wage gap, using Current Population Survey (CPS) data from 1975–2001. This application replicates and extends the influential work of Mulligan and Rubinstein (2008).

Key Findings

  • Standard OLS: Suggests a clear narrowing of the gender wage gap from -0.41 to -0.26 log points between the 1970s and 1990s.
  • Heckman Selection Models: Often suggest the gap did not actually shrink, but rather that the composition of the female workforce changed as more high-ability women entered the labor market.
  • Set-Identified Results: Using the new methodology, the authors find that while they cannot definitively reject the idea that the gap declined, the bounds are remarkably wide. For instance, in the 1970s, the median wage gap could have been anywhere from -1.24 to +0.58.

These results highlight a fundamental trade-off: without making strong, potentially unverifiable parametric assumptions about how individuals select into the workforce, the data alone may not contain enough information to definitively characterize the evolution of the wage gap.

Limitations and Future Directions

While the framework is powerful, it carries certain caveats:

  1. Conservative Inference: The data-jittering technique used for discrete variables results in confidence sets that are slightly wider (more conservative) than theoretically necessary.
  2. Instrument Sensitivity: The precision of the results (the width of the bounds) is highly dependent on the strength of the instruments used. Weak instruments can lead to bounds so wide they offer little practical insight.
  3. Computational Demand: While the Bayesian bootstrap is a practical tool, it requires repeated re-estimation of non-parametric models, which can be computationally intensive for very large-scale datasets.

In conclusion, this paper provides a rigorous yet flexible toolkit for researchers to summarize complex, partially identified relationships using interpretable linear models, offering a clearer view of what the data truly reveals—and what remains unknown—under minimal assumptions.

Summary updated

Abstract

This paper provides inference methods for best linear approximations to functions which are known to lie within a band. It extends the partial identification literature by allowing the upper and lower functions defining the band to be any functions, including ones carrying an index, which can be estimated parametrically or non-parametrically. The identification region of the parameters of the best linear approximation is characterized via its support function, and limit theory is developed for the latter. We prove that the support function approximately converges to a Gaussian process and establish validity of the Bayesian bootstrap. The paper nests as special cases the canonical examples in the literature: mean regression with interval valued outcome data and interval valued regressor data. Because the bounds may carry an index, the paper covers problems beyond mean regression; the framework is extremely versatile. Applications include quantile and distribution regression with interval valued data, sample selection problems, as well as mean, quantile, and distribution treatment effects. Moreover, the framework can account for the availability of instruments. An application is carried out, studying female labor force participation along the lines of Mulligan and Rubinstein (2008).
BibTeX
@misc{chandrasekhar2012inference,
  author = {Arun G. Chandrasekhar and Victor Chernozhukov and Francesca Molinari and Paul Schrimpf},
  title = {{Inference for best linear approximations to set identified functions}},
  year = {2012},
  eprint = {1212.5627},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1212.5627},
}
Figure