Statistics Methodologies for Neuroconnectivity Analysis Using fMRI Data in Autism

Previous research has identified a series of impairments in neuroconnectivity in autism, leading to the under-connectivity theory. However, several recent studies provided evidence that neural network was actually over-connected, but not under-connected as previously thought, in autism. These apparent discrepancies can be attributed to the drawbacks in study design and statistical methods. First of all, most previous studies suffered from small sample size, a problem that rendered these studies to be underpowered. Secondly, these studies failed to account for two critical sources of variation: difference among subjects and variation among sites. As a result, the estimated difference between patient and control and its standard error were not accurate. This not only decreased the efficiency of analysis but also lead to biased inferences. In this work we took advantage of the largest autism database, the Autism Brain Imaging Data Exchange(ABIDE). Our analyses included 361 subjects from 8 medical centers. We believed that the sample size in this work avoided the 'underpower' issue in most previous studies. By implementing EM algorithm, which iterated between E step that used empirical Bayesian to estimate random effect and M step which evaluated fixed effect via MLE, we were able to obtain accurate estimate of difference between patient and control and its standard error. After applying false discovery rate control, we identified 12 links with significant difference between patient and control. The second statistical methodology that we implemented in this work was Bayesian hierarchical modeling. In the work, we fitted this model to ABIDE data. The FDR control approach in the above-mentioned methods failed to account for the correlation among statistical tests. As fMRI data was obtained from different regions of the same brain, it was highly unlikely that the statistical tests were independent. Principal factor approximation was recently proposed to implement FDR control in the presence of dependence structure. We applied PFA to the ABIDE dataset and identified links with significant difference in autism. Lastly, the three above-mentioned methods did not take into account site effect, which could be an important source of variation. We performed meta-analysis by conducting separate analysis by site and then combined analysis with weight equal to the inverse of standard error. Meta-analysis identified additional significant links.