Nonparametric estimation of the pair correlation function of replicated inhomogeneous point processes (Q2209829)

Pair correlation functions (PCF) are a useful tool to explore and study spatial-temporal point pattern data. PCF determines the second-order characteristics of a spatial or temporal process and is often estimated by some nonparametric approach ranging from the most popular one the kernel estimation (see [\textit{D. Stoyan} and \textit{H. Stoyan}, Fractals, random shapes and point fields. Methods of geometrical statistics. Chichester: John Wiley \& Sons Ltd. (1994; Zbl 0828.62085); \textit{J. Illian} et al., Statistical analysis and modelling of spatial point patterns. Chichester: John Wiley \& Sons (2008; Zbl 1197.62135)], the Bayesian estimator [\textit{Y. R. Yue} and \textit{J. M. Loh}, J. Nonparametric Stat. 25, No. 2, 463--474 (2013; Zbl 1307.62218)] and the orthogonal series estimator [\textit{A. Jalilian} et al., Stat. Sin. 29, No. 2, 769--787 (2019; Zbl 1421.62069)]. However, the previous mentioned methods are highly dependent on the estimation of the first-order intensity function and they are widely studied only in single point pattern scenario. In this paper, motivated by a temporal point process dataset obtained from Sina Weibo, the largest Twitter-type social media platform in China, i.e., from a typical replicated point patterns data, the authors propose two nonparametric methods to estimate the PCF of inhomogeneous point process when replicates are available based on the estimation equation technique. More specifically, they propose the local polynomial estimator and the orthogonal series estimator. Both estimators are built by utilizing estimation equation technique, where the estimating equations are constructed by the score function of the weighted likelihood function in [\textit{Y. Guan}, J. Am. Stat. Assoc. 101, No. 476, 1502--1512 (2006; Zbl 1171.62348)] and [\textit{R. P. Waagepetersen}, Biometrics 63, No. 1, 252--258 (2007; Zbl 1122.62073)]. In this frame, the paper consists of six sections, three appendices and an online supplementary material. In the sequel, the subject of each one of them is briefly presented. Section 1 (Introduction) explains the theoretical and practical motivation of the current work and its main contribution. In Section 2 (Methodology), based on carefully designed estimating equations for the PCF, the two types of nonparametric estimators, namely the local polynomial estimator and the orthogonal series estimator, are presented and proposed. Since the methods, as nonparametric ones, depend on tuning parameters, Appendix A describes a data driven method for the selection of the tuning parameters. In Section 3 (Asymptotic properties), the asymptotic properties of both estimators are investigated, while valid-confidence bands are also derived. In Appendix B, some details on the empirical variance estimator needed to construct valid confidence bands are given, while the sketches of all technical proofs are given in Appendix C. The finite sample performance of the proposed PCF estimators is evaluated in Section 4 (Simulation study), while the proposed method is applied to the motivated data set, i.e., to the Weibo data set, in Section 5 (Sina Weibo data analysis). Section 6 (Discussion) discusses the the achieved properties for the proposed estimators and some further directions and possible subjects for future research work. Note that the relevant R code and detailed technical proofs can be found in an online supplementary material available on Prof. Rasmus Waagepetersen's home page (\url{https://people.math.aau.dk/~rw/Papers/EJSsupplementary/}) Summarizing, this is a well written paper which contributes to the analysis of replicated point processes (in contrast with the existing methods which mainly are focused on single observed point patterns) by introducing two types of nonparametric estimators of the PCF that do not require the estimation of the intensity function (a common condition of the existing methods). Finally, the authors recommend the use of a local influence estimator in practice, although the orthogonal series estimator potentially has better performances in some case scenarios.