This chapter describes Mixed Effects (MFX) analysis of fMRI data. The algorithm on which this is based is described fully in \cite{karl_mixed}.
Before doing an MFX analysis you will need to have previously implemented a within-subject analysis for each subject.
See, for example, chapter~\ref{Chap:data:faces} for how to do this.
If you have 12 subjects you should have 12 separate within-subject SPM analyses.
The results of these within-subject analyses can then be used in a summary statistic approach to random effects inference. This entails using contrasts from the group of subjects as data in a 'second-level' design as described in the previous chapter.
Alternatively you can implemented a Mixed Effects (MFX) analysis.
There are five steps to a MFX analysis.
\begin{enumerate}
\item \textbf{Specify FFX}
This step will create a Fixed Effects model with data from all subjects, and a single design matrix comprising partitions for each subject.
In the SPM batch editor (press the Batch button in the SPM top left (command) window) go to {\em SPM, Stats, Mixed Effects Analysis, FFX Specification}. Select the directory where you want the FFX model to be saved - we will refer to this as DIR.
Then select the SPM.mat files that contain the analyses for each individual subject. If you have 12 subjects you should select 12 SPM.mat files.
It is essential that the design matrices contained in the SPM.mat files have the same number of columns. More specifically it is required that, over subjects, there be the same number of sessions, conditions per session, and columns per condition (eg parametric/time modulators if any). This information is written out to the command window so, in the event of an error, you can see which subjects are the odd ones out.
\item \textbf{Estimate FFX}
This is a very simple step. Press the Estimate button in the top-left (command) SPM window, select the DIR/SPM.mat file created in the previous step, and then press the green play button. SPM will now estimate the group FFX model.
\item \textbf{Specify MFX} In the SPM batch editor go to {\em SPM, Stats, Mixed Effects Analysis, MFX Specification}. Select the DIR/SPM.mat file that was estimated in the previous step, and then press the green play button.
SPM will now specify a second level model that is equivalent to a two-level hierarchical model. This equivalence is derived in equation four of \cite{karl_mixed}. The second level model comprises (a) a second level design matrix, (b) data, which are the regression coefficient images from the estimated FFX model and (c) an error covariance matrix, whose variance components are computed using the Restricted Maximum Likelihood (ReML) algorithm.
It is the structure of this covariance matrix that makes an MFX analysis different from the alternative summary statistic implementation of RFX. The difference is that the MFX error covariance structure also contains a contribution from within-subject errors from the first level that have been projected through (the pseudo-inverse of) the first level design matrix.
SPM will create a subdirectory DIR/mfx and place an SPM.mat file there. This is the MFX model.
\item \textbf{Estimate MFX}
This is a very simple step. Press the Estimate button in the top-left SPM window, select the DIR/mfx/SPM.mat file created in the previous step, and then press the green play button. SPM will now estimate the MFX model.
\item \textbf{Results}
The estimated MFX model can be interrogated in the usual way. Press the Results button in the SPM command window. This will bring up the SPM contrasts manager, where you can specify effects to test as described in previous chapters.