Le Cam optimal tests for (Reciprocal) Inverse Gaussian sub-models within Generalized Inverse Gaussian models

Angelo Efoévi Koudou, Christophe Ley

Abstract

The Generalized Inverse Gaussian (GIG) distribution is a versatile tool for modeling complex phenomena in hydrology, finance, and neuroscience. However, its reliance on modified Bessel functions makes Maximum Likelihood (ML) estimation computationally intensive. This paper introduces a new class of statistical tests that achieve the efficiency of Likelihood Ratio Tests (LRT) with the computational simplicity of the Method of Moments (MM). By leveraging Le Cam’s theory of Uniform Local Asymptotic Normality (ULAN), the authors provide a closed-form testing procedure that outperforms classical methods in small samples and avoids complex numerical optimization.

The Computational Dilemma in GIG Models

Working with GIG models has historically presented a frustrating trade-off for researchers. Because the probability density function involves modified Bessel functions, there is no "neat" closed-form formula for ML estimators. Consequently, practitioners have had to choose between:

  1. Likelihood Ratio Tests (LRT): These are statistically efficient—making the most of the available data—but require complex numerical approximations to find parameter estimates.
  2. Method of Moments (MM): These are easy to calculate using simple arithmetic but are statistically inefficient, often requiring much larger sample sizes to reach the same level of certainty as an LRT.

This research bridges this gap, creating a test that is as powerful as the LRT but as computationally efficient as the MM.

Methodology: The Le Cam Framework

The researchers utilize the Le Cam methodology, a sophisticated framework in asymptotic statistics, to derive their testing procedure. The approach follows three main steps:

  1. Proving ULAN: They first prove that the GIG family satisfies the Uniform Local Asymptotic Normality (ULAN) property. This essentially means that, for large samples, the complex GIG model behaves locally like a much simpler Gaussian (Normal) shift model.
  2. The Efficient Central Sequence: They derive a "decorrelated" version of the score function (the gradient of the log-likelihood). This allows them to test a specific parameter (such as the shape parameter $p$) while treating others (scale and concentration) as "nuisance" parameters.
  3. The MM Plug-in: Because of the ULAN property, the authors demonstrate that they can "plug in" simple Method of Moments estimators for the nuisance parameters without losing any statistical power. This bypasses the need for the iterative numerical solvers required by ML estimation.

Experimental Validation

The authors validated their theoretical work using both simulated and real-world data:

  • Monte Carlo Simulations: 10,000 replications across various sample sizes ($n = 30, 50, 100, 200$) were used to compare the new test against the classical LRT. In these simulations, the new test matched or slightly exceeded the power of the LRT across nearly all parameter combinations. Notably, the new test maintained its "size" (the false positive rate) better than the LRT when the sample size was very small ($n=30$).
  • Traffic and Repair Data: The test was applied to a dataset of 128 vehicle passing intervals and a dataset of 46 airborne transceiver repair times. In both instances, the authors formally confirmed that the Inverse Gaussian (IG) model was a statistically sound fit, while the Reciprocal Inverse Gaussian (RIG) model was rejected.

Key Insights and Results

The most striking insight of this work is that Maximum Likelihood Estimators are not strictly necessary to build an "optimal" test. As long as an initial estimator is "root-n consistent" (like the Method of Moments), Le Cam’s framework can refine it into a test statistic that is asymptotically identical to the best possible likelihood-based test.

Furthermore, by projecting the test statistic onto a subspace orthogonal to the nuisance parameters, the authors created a test that is robust to uncertainty in the scale and concentration of the data. Because the test uses closed-form expressions with no loops or iterations, it is significantly faster to implement in software like R or Mathematica than traditional likelihood-based approaches.

Conclusion

While the current paper focuses on hypothesis testing, it lays the groundwork for the future development of "one-step" efficient estimators—formulas that could provide ML-quality parameter estimates in a single calculation. While the test is mathematically optimal in an asymptotic sense, future research may further explore its performance in extremely heavy-tailed scenarios or with data containing significant outliers.

Summary updated

BibTeX
@inproceedings{koudou2014le,
  author = {Angelo Efoévi Koudou and Christophe Ley},
  title = {{Le Cam optimal tests for (Reciprocal) Inverse Gaussian sub-models within Generalized Inverse Gaussian models}},
  booktitle = {Test},
  year = {2014},
  doi = {10.1007/s11749-014-0378-2},
}
Figure