A sample article title

Nicholas Woolsey, Xianzheng Huang · 2025

Read A sample article title by Nicholas Woolsey, Xianzheng Huang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-15962.txt.

Introduction

Some outcomes live on a circle: a wind direction, a phase, a time of day. Nicholas Woolsey and Xianzheng Huang study how to regress such a circular response on a predictor that is itself measured with error. An existing estimator assumes the predictor is exact, and the pair work out how to repair it.

They propose three revised estimators, all built around the same deconvolution idea, the mathematical undoing of the measurement blur, but exploiting it in different ways. They also add a faster method for choosing the smoothing bandwidth, a known bottleneck when errors are present. The team works out the estimators' large-sample behavior and tests them in simulations. Readers get a comparison of the three options' strengths, and tools for a data type that shows up in biology, geology, and movement studies.

Catalog record

CREATORNicholas Woolsey, Xianzheng Huang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE178 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-15962

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A sample article title Nicholas Woolsey, Xianzheng Huang Abstract This study considers regression analysis of a circular response with an error-prone linear covariate. Starting with an existing estimator of the circular regression function that assumes error-free covariate, three approaches are proposed to revise this estimator, leading to three nonparametric estimators for the circular regression function accounting for measurement error. The proposed estimators are intrinsically connected through some deconvoluting operator that is exploited differently in different estimators. Moreover, a new bandwidth selection method is developed that is more computationally efficient than an existing method well-received in the context of tuning parameter selection in the presence of measurement error. The efficacy of these new estimators and their relative strengths are demonstrated through a thorough investigation of their asymptotic properties and extensive empirical study of their finite-sample performance. keywords [class=MSC] Primary 62G08 ; secondary 62H11 keywords Bandwidth selection , complex error , deconvolution , measurement error † † orcid: 0000-0001-7077-0869 , ?? Statistics, University of South Carolina, ?? ?? Statistics, University of South Carolina, ?? Contents 1 Introduction Circular data are data of a periodic nature that arise in a host of applications. Examples of circular data include data on the direction of animal movements that are of interest in zoology and ecology [ 2 , 35 ] , and wind direction data as a subject of investigation in meteorology [ 24 ] . Times of certain events, such as hospital visits and crimes, as observed on the 24-hour clock can be transformed to circular data, so are dates of certain events on a calendar, such as births and migration. In these applications, researchers are often interested in studying the association between a circular response and linear covariates. For instance, Birkett et al., 2012 considered regressing elephant’s movements on rainfall; Hodel and Fieberg, 2022 analyzed the association between a moving animal’s turn-angle and step-length; García-Portugués et al., 2013 formulated a regression model for wind direction and \text{SO}_{2} concentration; Gill and Hangartner, 2010 developed a circular regression model for the timing of domestic terrorism events, with linear covariates such as the number of attackers and death count. Early developments of circular-linear regression analysis led to parametric models that usually assume a von Mises distribution [ 28 ] for the circular response given covariates. Frequently referenced regression models of this type include those proposed by Gould, 1969 , Johnson and Wehrly, 1978 , and Fisher and Lee, 1992 . More recent parametric models that adopt different distributional assumptions on the circular response include those considered in [ 43 ] , [ 36 ] , and [ 33 ] , among many others. A complication in formulating such regression models is the choice of a link function that relates predictors in the Euclidean space to the mean direction in a circular space. There is also the concern that the actual circular response may violate parametric assumptions implied by the chosen circular distribution. Nonparametric circular-linear regression models have been proposed to avoid choosing a link function or imposing stringent assumptions on the circular response. One of the first notable developments along this line is the local polynomial regression of a circular response introduced by Di Marzio et al., 2013 . Meilán-Vila et al., 2020 generalized this work by allowing a multivariate linear covariate. Oftentimes covariates of scientific interest cannot be measured precisely, either due to human error or imprecise measuring instruments. For example, rainfall measurements from a rain gauge are affected by the exposure conditions of the rain gauge; the measurement accuracy of \text{SO}_{2} concentration is also highly device-dependent. This motivates our study of circular regression models with a linear covariate prone to measurement error. The topic of regression analysis with error-prone covariates has been studied extensively. Comprehensive reviews of existing methodolgies for dealing with error-in-covariate can be found in [ 5 ] , [ 17 ] , [ 4 ] , [ 46 ] , and [ 45 ] . Despite the existing extensive research on regression models with error-in-covariate, parallel problems in the context of regression models for circular responses have received little attention. Our study contributes in this underinvestigated area. Concurrently, [ 11 ] considered regression models for a response that can be circular or linear, with error-prone covariates that also can be circular or linear. Focusing on circular-linear regression models, we propose in this study three methods based on local linear estimation adapted for a circular response while accounting for covariate measurement error. To prepare for the methodology development, Section 2 briefly reviews local polynomial regression of a circular response with error-free covariates proposed by [ 12 ] . In Section 3 we develop three strategies to revise their method to account for measurement error in the linear covariate. Section 4 summarizes asymptotic properties of our proposed estimators. Bandwidth selection required for local polynomial regression is especially challenging when covariates are prone to measurement error, which is an important issue not addressed in [ 11 ] . We present in Section 5 two bandwidth selection methods utilized in our proposed estimation methods. Section 6 reports simulation studies where the proposed estimators’ finite-sample performance is inspected and compared with a naive estimator and a benchmark estimator. We implement these methods in a real-life application in Section 7 . Lastly, Section 8 summarizes contributions of our study and discuss extensions of the study to different relevant settings. 2 Local polynomial circular-linear regression 2.1 The circular mean function Consider n independent bivariate observations, \{(\Theta_{j},X_{j})\}_{j=1}^{n} , identically distributed according to the joint distribution of (\Theta,X) supported on [-\pi,\pi)\times\mathbb{R} . Given the linear covariate X=x , the distribution of the circular response \Theta is specified by \displaystyle\Theta \displaystyle=\{m(x)+\epsilon\}(\text{mod }2\pi), (1) where \epsilon is an angular error following an unspecified distribution with a circular mean of zero, and m(x) is the circular mean of \Theta , i.e., the circular regression function that maps from a linear space to a circular space. The circular mean of a circular random variable \Theta is defined as the angle \alpha\in[-\pi,\pi) that minimizes the mean cosine dissimilarity between \Theta and \alpha , D(\alpha)={\mathbb{E}}\{1-\cos(\Theta-\alpha)\} , which can be shown to be \argmin D(\alpha)=\text{atan2}[{\mathbb{E}}(\sin\Theta),{\mathbb{E}}(\cos\Theta)] [ 28 , Chapter 2] . Here, \text{atan2}[y,x] is the angle between the x -axis and the vector from the origin to the point (x,y) in the Cartesian coordinate system. Because \text{atan2}[y,x] depends on (x,y) only via the ratio y/x , one may infer the circular mean function m(x) via inferring m_{1}(x)/m_{2}(x) , where m_{1}(x)={\mathbb{E}}(\sin\Theta|X=x) and m_{2}(x)={\mathbb{E}}(\cos\Theta|X=x) , both functions mapping to a linear space as in the traditional regression setting. 2.2 Local polynomial estimators Given the response data \boldsymbol{\Theta}=(\Theta_{1},\ldots,\Theta_{n})^{\top} and covariate data \boldsymbol{X}=(X_{1},\ldots, X_{n})^{\top} , [ 12 ] proposed kernel-based estimators for m(x) in the form of \tilde{m}(x)=\text{atan}2[\tilde{g}_{1}(x),\tilde{g}_{2}(x)], (2) where \tilde{g}_{1}(x)=\frac{1}{n}\sum_{j=1}^{n}\sin(\Theta_{j})\mathcal{W}(X_{j}-x),\,\,\tilde{g}_{2}(x)=\frac{1}{n}\sum_{j=1}^{n}\cos(\Theta_{j})\mathcal{W}(X_{j}-x), (3) in which \mathcal{W}(t) is a weight function. For example, if \mathcal{W}(t)=K_{h}(t) , where K_{h}(t)=K(t/h)/h , K(t) is a kernel function, and h is the bandwidth, then \displaystyle\tilde{g}_{1}(x) \displaystyle=\frac{1}{n}\sum_{j=1}^{n}\sin(\Theta_{j})K_{h}(X_{j}-x) \displaystyle=\frac{\sum_{j=1}^{n}\sin(\Theta_{j})K_{h}(X_{j}-x)}{\sum_{j=1}^{n}K_{h}(X_{j}-x)}\times\frac{1}{n}\sum_{j=1}^{n}K_{h}(X_{j}-x). Now one can see that the first factor in the preceding expression is a local constant estimator [ 30 ] for m_{1}(x) , and the second factor is a traditional kernel density estimator for the probability density function of X evaluated at x , denoted by f_{X}(x) . Hence, with

Read the complete text