Likelihood-based Spacings Goodness-of-Fit Statistics for Univariate Shape-constrained Densities

TL;DR

This paper introduces a likelihood ratio goodness-of-fit test for univariate shape-constrained densities, demonstrating its asymptotic properties, bootstrap calibration, and effectiveness through numerical studies.

Contribution

It develops a novel likelihood-based test for nonparametric shape-constrained densities, extending existing spacing-based methods with theoretical guarantees and practical validation.

Findings

Test is asymptotically normal and distribution-free under null.

Test is consistent under fixed alternatives.

Numerical studies show good error control and power.

Abstract

A variety of statistics based on sample spacings has been studied in the literature for testing goodness-of-fit to parametric distributions. To test the goodness-of-fit to a nonparametric class of univariate shape-constrained densities, including widely studied classes such as k-monotone and log-concave densities, a likelihood ratio test with a working alternative density estimate based on the spacings of the observations is considered, and is shown to be asymptotically normal and distribution-free under the null, consistent under fixed alternatives, and admits bootstrap calibration. The distribution-freeness under the null comes from the fact that the asymptotic dominant term depends only on a function of the spacings of transformed outcomes that are uniformly distributed. Applications and extensions of theoretical results in the literature of shape-constrained estimation are required to show that the average log-density ratio converges to zero at a faster rate than the sample spacing term under the null, and diverges under the alternatives. Numerical studies are conducted to demonstrate that the test is applicable to various classes of shape-constrained densities and has a good balance between type-I error control under the null and power under alternative distributions.