Multiresolution spatial generalized linear mixed model for integrating multi-fidelity spatial count data without common identifiers between data sources

Abstract

A motivating example for this paper is a large human location information system that collects two types of information on mobile device locations: 1) large amounts of low-accuracy cell tower triangulation (CTT) calculated location data and 2) small amounts of high-accuracy assisted global positioning system (AGPS) pinpointed location data. Integrating the CTT and AGPS data and extracting more complete and accurate location information is important to achieve better estimation of the true spatial density. However, the problem is challenging because there is no direct link between the CTT and APGS data. In this paper, we propose a multiresolution spatial generalized linear mixed model to integrate low-accuracy and high-accuracy spatial count data given no direct link between two data sources. The relationship between the high-accuracy data and low-accuracy data is estimated at a low-resolution level, where the relationship between the two types can be better captured, and then the estimated relationship is propagated to the high-resolution level. Using the high-accuracy data, the location information of the low-accuracy data is flexibly adjusted via spatial random effects that are modeled using a Gaussian process. The proposed method is validated using simulated and real data examples.