Statistical approaches to temporal and spatial autocorrelation in resting-state functional connectivity in mice measured with optical intrinsic signal imaging

Abstract

. Significance: Resting-state functional connectivity imaging in mice with optical intrinsic signal (OIS) imaging could provide a powerful translational tool for developing imaging biomarkers in preclinical disease models. However, statistical interpretation of correlation coefficients is hampered by autocorrelations in the data. Aim: We sought to better understand temporal and spatial autocorrelations in optical resting-state data. We then adapted statistical methods from functional magnetic resonance imaging to improve statistical inference. Approach: Resting-state data were obtained from mice using a custom-built OSI system. The autocorrelation time was calculated at each pixel, and scores for correlation coefficients were calculated using Fisher transforms and variance derived from either Bartlett’s method or xDF. The significance of each correlation coefficient was determined through control of the false discovery rate (FDR). Results: Autocorrelation was generally even across the cortex and parcellation reduced variance. Correcting variance with Bartlett’s method resulted in a uniform reduction in scores, with xDF preserving high scores for highly correlated data. Control of the FDR resulted in reasonable thresholding of the correlation coefficient matrices. The use of Bartlett’s method compared with xDF results in more conservative thresholding and fewer false positives under null hypothesis conditions. Conclusions: We developed streamlined methods for control of autocorrelation in OIS functional connectivity data in mice, and Bartlett’s method is a reasonable compromise and simplification that allows for accurate autocorrelation correction. These results improve the rigor and reproducibility of functional neuroimaging in mice.

Controlling the familywise error rate in widefield optical neuroimaging of functional connectivity in mice

“…Although it is possible to process optical data using existing fMRI analysis programs, such methods should be used with caution. The statistical properties of optical data can differ substantially from fMRI data, 15 and it is unclear if results from fMRI-specific programs would be valid, absent an analysis comparable to the present paper. Thus, we applaud the efforts of Brier and Culver 7 for developing the first freely available analysis software for widefield optical imaging in mice.…”

“…Each seed trace was constructed by averaging the signal across a 5-pixel-radius ( -radius) circle centered on the seed location. Correlation coefficients with each seed were converted to -scores using Fisher’s transformation and a variance that accounted for temporal autocorrelation using “Bartlett’s method.” 15 …”

“…If is an image, then (assuming unit pixel dimensions and total pixels) the partial spatial derivatives are estimated as The variance and covariance are then estimated as where is the number of degrees of temporal freedom in the data, which we estimate using Bartlett’s method. 15 …”

“…Correlation coefficients with each seed were converted to z-scores using Fisher's transformation and a variance that accounted for temporal autocorrelation using "Bartlett's method." 15…”

“…where η is the number of degrees of temporal freedom in the data, which we estimate using Bartlett's method. 15 The estimated variance-covariance matrix is E Q -T A R G E T ; t e m p : i n t r a l i n k -; e 0 1 9 ; 1 1 6 ; 4 0 8…”