Estimation of circular statistics in the presence of measurement bias

TL;DR

This paper introduces a novel method for accurately estimating circular statistics and their significance in cyclic event analysis, even with biased or incomplete data sampling, improving reliability over traditional methods.

Contribution

It presents a new approach combining simulation-based measurement estimation and linear distribution parametrization to correct for measurement bias in circular statistics.

Findings

Low estimation error demonstrated in toy and real-world examples

Corrected moments showed residual RMS less than 0.007 in real data

Numerical significance estimation outperforms traditional p-value calculations

Abstract

Background and objective. Circular statistics and Rayleigh tests are important tools for analyzing the occurrence of cyclic events. However, current methods fail in the presence of measurement bias, such as incomplete or otherwise non-uniform sampling. Consider, for example, studying 24-cyclicity but having data not recorded uniformly over the full 24-hour cycle. The objective of this paper is to present a method to estimate circular statistics and their statistical significance even in this circumstance. Methods. We present our objective as a special case of a more general problem: estimating probability distributions in the context of imperfect measurements, a highly studied problem in high energy physics. Our solution combines 1) existing approaches that estimate the measurement process via numeric simulation and 2) innovative use of linear parametrizations of the underlying distributions. We compute the estimation error for several toy examples as well as a real-world example: analyzing the 24-hour cyclicity of an electrographic biomarker of epileptic tissue controlled for state of vigilance. Results. Our method shows low estimation error. In a real-world example, we observed the corrected moments had a root mean square residual less than 0.007. We additionally found that, even with unfolding, Rayleigh test statistics still often underestimate the p-values (and thus overestimate statistical significance) in the presence of non-uniform sampling. Numerical estimation of statistical significance, as described herein, is thus preferable. Conclusions. The presented methods provide a robust solution to addressing incomplete or otherwise non-uniform sampling. The general method presented is also applicable to a wider set of analyses involving estimation of the true probability distribution adjusted for imperfect measurement processes.