MELD: Mixed Effects for Large Datasets

Abstract

Mixed effects models provide significant advantages in sensitivity and flexibility over typical statistical approaches to neural data analysis, but mass univariate application of mixed effects models to large neural datasets is computationally intensive. Threshold free cluster enhancement also provides a significant increase in sensitivity, but requires computationally-intensive permutation-based significance testing. Not surprisingly, the combination of mixed effects models with threshold free cluster enhancement and nonparametric permutation-based significance testing is currently completely impractical. With mixed effects for large datasets (MELD) we circumvent this impasse by means of a singular value decomposition to reduce the dimensionality of neural data while maximizing signal. Singular value decompositions become unstable when there are large numbers of noise features, so we precede it with a bootstrap-based feature selection step employing threshold free cluster enhancement to identify stable features across subjects. By projecting the dependent data into the reduced space of the singular value decomposition we gain the power of a multivariate approach and we can greatly reduce the number of mixed effects models that need to be run, making it feasible to use permutation testing to determine feature level significance. Due to these innovations, MELD is much faster than an element-wise mixed effects analysis, and on simulated data MELD was more sensitive than standard techniques, such as element-wise t-tests combined with threshold-free cluster enhancement. When evaluated on an EEG dataset, MELD identified more significant features than the t-tests with threshold free cluster enhancement in a comparable amount of time.