Abstract
In the last decade a growing body of research has studied inference on partially identified parameters (e.g., Manski, 1990, 2003). In many cases where the parameter of interest is real-valued, the identification region is an interval whose lower and upper bounds may be estimated from sample data. Confidence intervals may be constructed to take account of the sampling variation in estimates of these bounds. Horowitz and Manski (1998, 2000) proposed and applied interval estimates that asymptotically cover the entire identification region with fixed probability. Here we introduce conceptually different interval estimates that asymptotically cover each element in the identification region with fixed probability (but not necessarily every element simultaneously). We show that these two types of interval estimate are different in practice, the latter in general being shorter. The difference in length (in excess of the length of the identification set itself) can be substantial, and in large samples is comparable to the difference of one- and two-sided confidence intervals. A complication arises from the fact that the simplest version of the proposed interval is discontinuous in the limit case of point identification, leading to coverage rates that are not uniform in important subsets of the parameter space. We develop a modification depending on the width of the identification region that restores uniformity. We show that under some conditions, using the estimated width of the identification region instead of the true width maintains uniformity.
Introduction
In classical econometrics, researchers typically aim for point identification, where a single "true" value for a parameter $\theta$ can be recovered given infinite data. However, real-world data constraints—such as missing values, selection bias, or interval-censored responses—often mean that the parameter can only be narrowed down to a range, or identification region $[\theta_l, \theta_u]$. This state is known as partial identification.
Historically, the standard approach to inference in these settings was to construct confidence intervals (CIs) that cover the entire identification region with a specified probability. While statistically sound, this approach is often overly conservative if the researcher's primary interest is the parameter $\theta$ itself rather than the set containing it. Imbens and Manski (2004) propose a more efficient methodology that targets the coverage of the parameter, resulting in significantly narrower intervals without sacrificing statistical validity.
The Conceptual Shift: Parameter vs. Set Coverage
The fundamental insight of this research is that covering a set is a strictly more demanding task than covering a single point within that set. If a confidence interval covers the entire identification region $[\theta_l, \theta_u]$, it must, by definition, cover the true parameter $\theta$. However, an interval designed specifically to cover $\theta$ can be shorter while still maintaining the desired confidence level (e.g., 95%).
In large samples, the difference in length between these two types of intervals is substantial. The authors demonstrate that the "slack" created by focusing on the parameter rather than the set allows for a reduction in width comparable to the difference between one-sided and two-sided confidence intervals. For a 95% CI, this often means the difference between using a critical value of 1.645 (one-sided) versus 1.96 (two-sided).
Methodology and the Uniformity Challenge
The authors propose constructing CIs based on the estimated lower and upper bounds, $\hat{\theta}_l$ and $\hat{\theta}_u$. A primary technical challenge they address is the "discontinuity trap."
When the identification region is wide, the uncertainty about the parameter's location within the bounds is dominated by the width of the region itself, and the CI behaves like a one-sided interval. However, as the width of the identification region shrinks toward zero (approaching point identification), the CI must transition to a standard two-sided interval to maintain valid coverage.
Without adjustment, a naive parameter-covering CI would be non-uniform; it might paradoxically become wider as data becomes more precise. To solve this, the authors introduce a transition rule for the critical value $C_N$ based on the estimated width of the identification region ($\hat{\Delta} = \hat{\theta}_u - \hat{\theta}_l$). This ensures that the CI maintains the nominal coverage rate across the entire parameter space, regardless of whether the parameter is point-identified or partially identified.
Empirical Application: Workers' Compensation Benefits
To demonstrate the practical efficiency gains of their method, the authors apply it to a dataset from Meyer, Viscusi, and Dubin (1995) concerning workers' compensation benefits in Kentucky and Michigan. The study analyzed how a 1987 benefit increase affected the duration of injury claims. Because the model used is non-parametric and the outcome (weeks on disability) is discrete, the treatment effect is not point-identified.
Using the traditional set-covering approach, the 95% CI for the "Average Effect on Treated" was calculated as [-3.17, 4.42]. Applying the Imbens-Manski adjusted uniform CI, the interval narrowed to [-2.82, 4.08]. This reduction in width provides researchers with a more precise range for policy evaluation while maintaining the same level of statistical confidence.
Conclusion
The methodology introduced by Imbens and Manski represents a significant advancement in the treatment of partial identification. By shifting the focus from set coverage to parameter coverage and solving the associated uniformity problems, they provide a tool that yields tighter, more informative bounds for applied researchers dealing with imperfectly identified economic models.
Summary updated