Hausman-Type Overidentification Tests

• Hausman-type overidentification tests are statistical procedures that assess overidentifying restrictions in semiparametric models using doubly robust estimators.
• They compare estimators derived from distinct nuisance models, ensuring that differences under the null hypothesis are asymptotically negligible.
• Recent advances integrate geometric and influence-function methods to establish necessary conditions for DR estimator existence and robust inference.

Hausman-type overidentification tests are a class of statistical procedures designed to assess the validity of overidentifying restrictions in semiparametric and nonparametric estimation frameworks, particularly in the presence of nuisance parameters. These tests play a central role in determining whether a doubly robust (DR) estimator can be constructed and whether it is meaningful to exploit multiple models for inference. The recent geometric and influence-function-based literature provides a foundational characterization of the existence, necessity, and sufficiency for these overidentification tests, extending the classical Hausman test paradigm to semiparametric models with double robustness.

1. Conceptual Foundations

Hausman-type overidentification tests arise in the context of models with more restrictions (moment or otherwise) than necessary to identify the parameter of interest. In semiparametric settings, this commonly manifests as the availability of two or more candidate models for nuisance functions, leading to "overidentified" estimating equations for the target parameter. The classical logic is to compare estimators arising from different models: if both models are correct under the null, the estimators should agree asymptotically; a significant discrepancy signals model misspecification.

In the context of double robustness, the existence of a DR estimator is predicated on the orthogonality of certain estimating functions to nuisance tangent spaces relative to each overidentification equation. The geometric perspective provided by recent work reveals that the global invariance (across whole nuisance contours) of influence curves corresponds to precisely the property tested by generalized Hausman-type procedures (Ying, 2024).

2. Formal Statistical Framework

Let $M$ be a statistical model (possibly infinite dimensional), $p \in M$, and consider a finite-dimensional parameter of interest $\theta(p) \in \Theta \subset \mathbb{R}^k$, together with two nuisance functionals $\gamma(p) = (\gamma_1(p), \gamma_2(p))$. The DR estimating function $\phi(X; \theta, \gamma_1, \gamma_2)$ is designed so that, for all $p$ and for either $\gamma_1$ or $\gamma_2$ at their true values,

$\E_p [\phi(X; \theta(p), \gamma_1(p), \gamma_2)] = 0,\quad\text{for all }\gamma_2,$

and symmetrically for $\gamma_1$. This leads immediately to a system of overidentifying restrictions: the estimator $p \in M$0 is identified by more than one equation arising from different parametric or semiparametric submodels for the nuisances.

The tangent space $p \in M$1 is the space of all regular parametric score functions. The influence curve $p \in M$2 is an element of this tangent space satisfying the score equation for $p \in M$3. The existence of a Hausman-type test in this context reduces to asking whether, given two (or more) estimating equations constructed via distinct nuisance specifications, the difference between resulting estimators $p \in M$4 and $p \in M$5 is asymptotically negligible under the null that both nuisance models are correct.

3. Geometric Characterization and Necessary/Sufficient Conditions

Recent advances establish necessary and sufficient conditions for the existence of DR estimators (hence, of meaningful Hausman-type overidentification tests) in terms of global orthogonality relative to all nuisance directions. The fundamental criteria are:

1. Variation independence: For any fixed $p \in M$6, there exists $p \in M$7 realizing these values.
2. $p \in M$8-connectedness: Each nuisance-parameter contour set (holding $p \in M$9 and one $\theta(p) \in \Theta \subset \mathbb{R}^k$0 fixed) must be smoothly path-connected, so scores along these submodels are well-defined.

Given these, the DR estimator (and associated test statistic) exists if and only if there is an influence-curve-valued field on $\theta(p) \in \Theta \subset \mathbb{R}^k$1 orthogonal to the tangent space of each nuisance contour at every point—not just at $\theta(p) \in \Theta \subset \mathbb{R}^k$2, but uniformly across the entire contour submanifolds. In other words, only when this stringent global condition holds does a valid Hausman-type test exist for DR estimation (Ying, 2024).

When the nuisance contours are convex subsets of $\theta(p) \in \Theta \subset \mathbb{R}^k$3 (e.g. in many canonical semiparametric models), the orthogonality condition "lifts" from a local to a global property automatically, so any influence curve fulfills the requirements for double robustness “for free.” Here, standard Hausman tests (comparisons between moment-based or efficient estimators) remain valid, and their difference will have mean zero under the null.

4. Information-Geometric Interpretation

The geometric approach to double robustness extends naturally to the design and interpretation of Hausman-type tests. Utilizing information geometry, e-parallel and m-parallel transport describe how influence curves and tangent vectors translate along the statistical manifold. The DR property (and thus the exact equality under the null of Hausman-type estimator differences) is equivalent to invariance of the influence function under e-parallel transport along each nuisance contour.

Specifically, the DR estimator is said to be m-flat along the nuisance contours if the family of efficient influence functions is stable under m-parallel transport, which ensures global orthogonality and hence the validity of overidentification tests for DR estimation. If m-flatness holds, any influence curve is globally DR; otherwise, only a restricted set of DR estimators (if any) exist. A stronger notion, m-curvature freeness, is sufficient for the efficient IC to be DR, which then underpins exact Hausman testing in nonconvex scenarios.

5. Construction and Implementation of Hausman-Type Tests

In practice, construction of a Hausman-type overidentification test for DR estimation proceeds by:

• Specifying two (or more) candidate models for each nuisance functional, ensuring variation independence and connectedness of their parameter spaces.
• Computing the DR estimator under each model and forming a test statistic based on their scaled difference.
• Verifying (either analytically or by computation) that the orthogonality conditions required for the DR property are satisfied globally (particularly in nonconvex frameworks).
• Under classical asymptotics, under the null hypothesis (both nuisance models correct), the difference between the estimators is asymptotically normal with mean zero and a covariance estimable from the joint influence functions.

In convex settings, this approach is standard and underpins a vast literature on specification and overidentification tests in econometrics and semiparametric statistics. When the global geometric conditions fail, as in certain non-convex parameterizations, traditional overidentification tests may yield non-null limiting distributions under the global null; thus, their application requires careful theoretical justification (Ying, 2024).

6. Implications for Semiparametric Inference

Hausman-type overidentification tests in the DR framework provide a principled method to detect model misspecification and to validate the structuring of DR estimators. Their existence, validity, and power are tied to deep geometric features of the statistical model, notably variation independence and the (possibly convex) geometry of nuisance parameter contours.

A practical implication is that in canonical settings—such as partially linear models, mean difference functionals, or ATE estimation—the classical conditions for DR and Hausman-type testing are satisfied automatically (by convexity). In such settings, robust specification testing and model comparison via Hausman-type approaches are fully justified.

In more complex semiparametric structures (e.g., models with hierarchical, manifold, or non-convex constraints on the nuisance space), researchers must ensure global geometric conditions are met before applying overidentification testing; otherwise, the logical foundations of the test—and associated DR estimators—may be invalid.

7. Summary Table: Overidentification and DR Existence Criteria

Setting	DR Estimator Exists	Hausman-Type Test Valid	Key Condition
Nuisance contours convex	Yes	Yes	Convexity (m-flatness)
Nonconvex, but m-curvature-free	Efficient IC only	Partial	m-curvature-free
Nonconvex, not m-curvature-free	Typically not	No	Global orthogonality

This formalizes the relationship between overidentification testing, the geometric structure of the semiparametric model, and the existence of double-robust estimators as established in (Ying, 2024). The Hausman-type test is thus both a methodological tool for practical model assessment and a lens for understanding the deep structure of semiparametric and robust estimation.