Voxelwise Multivariate Analysis of Multimodality Magnetic Resonance Imaging (P06.026)

OBJECTIVE: Simulation studies and a data analysis are performed to explore the relative merits of various methods of controlling for the number of tests per voxel in imaging studies with multiple measurements per voxel. BACKGROUND: Most brain magnetic resonance imaging (MRI) studies concentrate on a single MRI contrast or modality, frequently structural MRI. By performing an integrated analysis of several modalities, such as structural, perfusion-weighted, and diffusion-weighted MRI, new insights may be attained to better understand the underlying processes of brain diseases. DESIGN/METHODS: We compare two voxelwise approaches: (1) fitting multiple univariate models, one for each outcome and then adjusting for multiple comparisons among the outcomes and (2) fitting a multivariate model. In both cases, adjustment for multiple comparisons is performed over all voxels jointly to account for the search over the brain. The multivariate model is able to account for the multiple comparisons over outcomes without assuming independence because the covariance structure between modalities is estimated. To illustrate the power of each approach, we simulate data with multiple outcomes under four different covariance assumptions and analyze a case control study of Alzheimer9s disease, in which data from three MRI modalities are available. RESULTS: Simulations show that the multivariate approach is more powerful when the outcomes are correlated and, even when the outcomes are independent, the multivariate approach is just as powerful or more powerful when at least two outcomes are dependent on predictors in the model. However, multiple univariate regressions with Bonferroni correction remains a desirable alternative in some circumstances. CONCLUSIONS: When choosing a method to account for multiple comparisons, researchers must consider several factors, including the primary hypothesis, whether the correlations between modalities are of interest, and computational resources. Supported by: This work was supported in part by the National Institute of Health grants T32 MH017119, U01 HL089856, R01 MH081862, R01 MH087590, and R01 CA157528. This work was also partially supported by the following National Institutes of Health Grants administered by the Northern California Institute for Research and Education: R03EB8136, P41RR023953, P50AG023501, and P01AG19724, and with resources of the Veterans Affairs Medical Center, San Francisco, California. Disclosure: Dr. Naylor has nothing to disclose. Dr. Cardenas has nothing to disclose. Dr. Tosun has nothing to disclose. Dr. Schuff has nothing to disclose. Dr Weiner is engaged in consulting activities for Astra Zeneca, Araclon, Medivation/Pfizer, Ipsen, TauRx Therapeutics LTD, Bayer Healthcare, Biogen Idec, Exonhit Therapeutics, Servier, Synarc, Pfizer, Janssen, Harvard University and KLJ Associates. Dr. We....Dr. Weiner owns stock options in Synarc and Elan.Dr. Weiner has received research support from Merck & Co., Inc. and Avid Radiopharmaceuticals. Dr. Schwartzman has nothing to disclose.