Standardizing the spectra (Shootout 2002)

We saw in the previous post that it was necessary to adjust the bias in Instrument 2, to get similar results to Instrument 1. Bias adjustment is the easiest way to transfer a model to other instrument if we see clearly the bias in the plots and if there is an improvement in the standard errors of predictions corrected by the bias (SEP).

But you know that this is not the better way to do the transfer. It is better to standardize the instruments being "Instrument 1" the "Master" and "Instrument 2" the "Host". For that reason I select a group of samples from the Test file scanned at Instrument 1 and the same samples scanned at Instrument 2, and calculate a correction Matrix to apply to all the spectra in  Instrument 2, to seem like they were scanned at Instrument 1. This procedure is described in the book "Chemometrics with R" - Ron Wehrens.

The results are a big improvement in the transferability.

These are the statistics monitoring the calibration samples of "Instrument 2" versus the model developed with the calibration samples of "Instrument 1", with and without "std".

# (without std..RMSEP: 3.642)       (with std..RMSEP: 2.913)
# (without std..Bias :-2.249)       (with std..Bias :-0.159)
# (without std..SEP  : 2.875)       (with std..SEP  : 2.918)

These are the statistics monitoring the Test samples of Instrument 2 versus the model developed with the calibration samples of "Instrument 1", with and without "std".

# (without std..RMSEP: 3.358)        (with std..RMSEP: 2.936)
# (without std..Bias :-1.712)        (with std..Bias : 0.390)
# (without std..SEP  : 2.892)        (with std..SEP  : 2.913)

These are the statistics monitoring the Validation samples of "Instrument 2" versus the model developed with the calibration samples of "Instrument 1", with and without "std".

# (without std..RMSEP: 5.635)        (with std..RMSEP: 2.961)
# (without std..Bias :-4.688)        (with std..Bias :-1.268)
# (without std..SEP  : 3.168)        (with std..SEP  : 2.71 )