Analysis of Longitudinal Neuroimaging Data with Time-varying Covariates Measured with Error

Longitudinal neuroimaging investigations, where participants are scanned repeatedly over time and imaging data are obtained at numerous time-points, are essential to understanding structural and functional changes in healthy and pathological brains. From a data analysis perspective, the complexity of such datasets is immense. Longitudinal imaging data is highly correlated, both temporally (because subjects participate in multiple scanning sessions) and in the spatial domain (because each image of the brain collected at each scanning session contains hundreds of thousands of spatially-related measurements). Further, the accuracy of collected covariates in a neuroimaging study is often far from perfect in that measurement error may arise from different sources such as flawed measurement systems and self-reports. Research that considers techniques to model neuroimaging data is vast and ongoing, slowly expanding to include longitudinal modeling approaches, but rarely taking into account the spatial nature inherent in brain images. Additionally, methods that reduce bias caused by measurement error in covariates have been discussed extensively in literature but consideration of how to integrate error-prone covariates into models fitted to longitudinal imaging data has not been addressed. In order to address these issues, we develop a spatial modeling method for analysis of imaging datasets collected at repeated occasions that includes timevarying covariates measured with error. Our method provides some advantage over currently-used imaging analysis techniques in that it allows researcher to (a) analyze longitudinal imaging data, (b) incorporate the spatial features of data into analysis, and (c) reduce the bias caused by covariates measured with error.