In my current project, I'm using PLS regression on infrared spectra (FTIR). For this I'm using R and the pls function from the
pls always mean centers both the input data and the infrared spectra. When predicting using a fitted PLS model, the means fitted during the estimation process will be subtracted from the data.
In one particular situation this is not a desirable effect for me, as the mean for each of the IR wavenumbers is not constant between measurements. This is caused by the fact that the path length of the beam which goes through the sample can change, either because a particular machine has a different path length than the machine used for fitting the model, or because the path length slowly changes over time because of wear and tear in the measuring cell.
By not mean centering the data, the means end up the first PLS factor. This allows a linear scaling of the mean, taking into account linear effects such as the cell path length. This makes the model more robust in particular situations. I am aware that we could correct the IR spectra first, and then use the fitted means. But, we would like to put this inside the fitted model.
I sent an e-mail to the maintainer of the PLSR package about why the package did not support switching off mean centering. The reply was:
No, that is not possible. Theoretically, if one does not center, it is not PLSR.
Other tools used in the field of spectroscopy (Grams, Unscrambler) do allow switching off mean centering. And in the situation above I feel that disabling mean centering has a big advantage.
Now for my concrete question:
Is PLS without mean centering still PLS, are there any theoretical or practical reasons not to do this?