Despite the numerous studies conducted on different types of neuroimaging modalities applied in the diagnostic classification of autism spectrum disorders (ASD) by use of machine learning algorithms, none have assessed whether one modality may be more informative compared to others. In this study, a combined multimodal imaging technique using conditional random forest (CRF) was applied to structural (anatomical) MRI, diffusion tensor imaging (DTI), and functional connectivity MRI (fcMRI) to assess whether different modalities may be more or less informative. In-house data from the Brain Development Imaging Laboratory (BDIL) were used, which included 47 typically developing (TD) and 46 ASD participants (N=93), matched on age, motion, and non-verbal IQ. The fcMRI data consisted of a matrix of 220x220 functionally defined regions of interest, yielding 24090 variables. The anatomical data had 397 variables, and the DTI data consisted of 192 variables. When combined, the top 100 variables from each modality, based on mean decrease in accuracy (MDA), reached an accuracy of 88.2%. Of these combined top 300 variables, 93% of the top 100 belonged to the fcMRI modality. An accuracy rate of 92.5% was achieved when the top 19 variables from each modality were combined, and all 19 variables from the fcMRI modality had the highest MDAs. Furthermore, the two most informative networks within the fcMRI modality were the visual and default mode networks (DMN), which was not surprising given previous findings.