TL;DR
This paper introduces a new type of confidence interval (CI) for "partially identified" parameters, which are common in econometrics when data is incomplete. Unlike existing methods that aim to cover the entire range of possible parameter values (the "identification region"), these new CIs focus on covering the true parameter value with a specified probability. The authors also address a critical issue where their initial approach can lead to unreliable coverage rates when the parameter is nearly point-identified, proposing a modification to ensure uniform coverage.
What Problem Does This Paper Address?
In many real-world scenarios, especially in econometrics, we can't precisely pinpoint a single value for a parameter of interest. Instead, we can only determine a range of plausible values, known as the "identification region." This often arises from missing data or measurement errors. Traditional statistical inference methods, like confidence intervals, are designed for precisely estimated parameters. When dealing with partially identified parameters, a key challenge is how to construct confidence intervals that reliably reflect the uncertainty in our estimates. Existing methods often construct CIs that aim to cover the entire identification region with a certain probability. However, this can lead to overly wide intervals. The authors propose a more focused approach: CIs that cover the true parameter value itself, which can be shorter and more informative. A significant hurdle they identify is that their initial proposed CI can behave erratically when the identification region shrinks to a single point (i.e., when the parameter becomes "point-identified"), leading to coverage rates that aren't uniform across all possible scenarios.
Methods
The core contribution is the development of a new class of confidence intervals. The authors contrast two types of CIs:
- Identification Region CIs: These aim to cover the entire interval of plausible parameter values with a fixed probability. This is the approach taken by previous work (e.g., Horowitz and Manski, 1998, 2000).
- Parameter CIs (Proposed): These aim to cover the true parameter value with a fixed probability. The authors show that this is generally a less demanding objective than covering the entire identification region, and thus these CIs can be shorter.
A key technical challenge arises with the proposed parameter CIs: their coverage probabilities might not be uniform across all possible values of the unknown parameters (like the width of the identification region). This means the CI might not perform as expected in all situations, particularly when the parameter is nearly point-identified. To address this, they propose a modified parameter CI that ensures uniform coverage probabilities. This modification involves adjusting the interval based on the estimated width of the identification region, ensuring more reliable performance across different scenarios.
The paper uses asymptotic theory (large sample properties) to derive the behavior of these CIs. They also provide exact results for a specific case (normally distributed outcomes) to motivate their modified CI.
Datasets
Not discussed in the paper. The paper focuses on theoretical development and uses simulated data for illustration in the empirical section.
Key Insights and Ideas
- Focusing on the Parameter, Not Just the Region: The fundamental insight is that aiming to cover the true parameter value is often more efficient and informative than aiming to cover the entire range of possible values (the identification region). This can lead to shorter, more precise confidence intervals.
- The Uniformity Problem: A crucial finding is that even if a CI covers the true parameter value asymptotically, its coverage might not be uniform across all possible data-generating processes or parameter values. This "non-uniformity" is particularly problematic when the parameter is nearly point-identified, leading to potentially misleadingly short intervals.
- The Solution: Adaptive Intervals: The authors demonstrate that by making the CI's construction "adaptive" – specifically, by incorporating an estimate of the identification region's width – uniform coverage can be achieved. This is a significant technical achievement that makes the proposed CIs more robust.
- Trade-offs in CI Construction: The paper highlights that there's a choice for researchers: CIs that cover the identification region versus CIs that cover the parameter. While the latter can be shorter, ensuring their reliability requires careful construction to avoid uniformity issues.
Key Results
- Lemma 2.1: Shows that the coverage probability of a parameter CI is always at least as large as the coverage probability of an identification region CI constructed with the same bounds. This implies parameter CIs can be shorter.
- Empirical Illustration (Table 1): Demonstrates that the different types of CIs can yield substantially different results in practice. For instance, the "uniform" CI for the parameter is often narrower than the CI for the identification region.
- Modified CI Guarantees Uniformity: The paper successfully develops a modified CI that achieves uniform asymptotic coverage probabilities across a range of nuisance parameters, including the width of the identification region.
Limitations and Future Work
- Asymptotic Results: Much of the analysis relies on large-sample (asymptotic) properties. While the paper provides some exact results for a specific case, the general results are asymptotic. Finite-sample performance might still differ.
- Assumptions: The theoretical results rely on several technical assumptions (e.g., uniform estimation of bounds, convergence of interval width). The applicability of the proposed methods depends on these assumptions holding in practice.
- Computational Complexity: While not explicitly detailed, constructing these adaptive CIs might involve more complex computations than simpler methods.
- Future Work: The paper suggests that the modified CI can be implemented more generally and provides conditions for uniform asymptotic coverage by substituting estimates for unknown nuisance parameters. The empirical illustration hints at the practical importance of these theoretical developments.
Summary updated