If you decide to perform a PCA model you'll find
that most standard procedures such as the MATLAB built-in SVD or the PCA
m-file of the PLS Toolbox do not handle missing data. There are at least
two ways to circumvent this. All standard chemometric software such as
Unscrambler, SIMCA, etc. have sub-optimal algorithms for handling missing
data but the N-way Toolbox does provide means for estimating an optimal
two-way PCA model with missing entries. Simply rearrange, the data to a two-way
array and fit a PARAFAC model to these data using
orthogonality on the first and second mode. This will produce a PCA model.
PARAFAC as such, is not made to fit two-way data. For example, you can not have
orthogonality in all modes in PARAFAC and we need that here. Instead of
specifying the constraints as [1 1] becaus of the two modes we specify the
constraints as [1 1 0] and that should do the trick, even though we only input a
two-way array.
Another possibility is to use the m-file MDPCA of the PLS_Toolbox which
directly handles missing data.