BUCHI NIRCal - Manual (Page 135)

Chemometrics
NIRCal 5.5 Manual, Version A 135
Spectrum No. 16 in the calibration set has the max. residual value (0.00449) and defines the limit for
the calibration and application:
Max. allowed residual = Max. C-Set residual * 2
2: default Residual Blow up, can be adjusted by the user.
In this example, these spectra are not real residual outliers for NIRCal. In case these spectra would be
real outliers, they can be selected as C-set spectra to entlarge the limits or eliminated from the
calibration (U-Set = unused set).
There are two different types of outliers:
1. False measurement:
If for example air bubbles in liquid are measured or particles pollute the sample just in front of
the optics of a probe, an outlier spectrum is measured. As three spectra are collected from
every sample, it is easy to recognize these spectra. The spectra of one and the same sample
should be similar. Nevertheless variations due to the production process will also manifest in
the NIR spectra. Spectra originating from false measurements should be deleted or eliminated
from the selection (unused).
2. Samples "out of specification":
If all spectra of one sample are different to spectra from other batches of the same material,
the sample itself may be considered as an outlier. The difference can be caused by several
parameters, for example changes in the production process, the type and/or the composition
of single substances in a product were changed and/or varied, or the sample is polluted. The
user has to verify what caused the difference and if the particular sample can still be used for
its purpose.
Before outliers are deleted, a careful clarification of the reason should be made for the appearance of
the outlier.
For quantitative calibration to find out if the reference value or the measured spectra must be
regarded as an outlier, the score plots should be reviewed (Graphics / Scores / 2D-Scatter). Spectra
breaking ranks, show clearly deviating scores and residuals (Graphics / Spectra / Residuals). Is that
not the case, the reference value can be considered as false.
If there are big differences between the reference values and the predicted values, but the scores do
not have particular deviations, with high probability, the outliers appear because of false reference
values.
Groups of samples with systematically deviations
This effect can be seen from time to time when samples are evaluated their reference values have
been determined in laboratories not using exactly the same reference methods. Here only an
alignment of the reference methods can help.
Significantly different results depending on the chosen classification of the samples into the C-
and V-Set
The number of used samples is too small, for instance because of not considered, hidden properties.
Remedy: selective completion of the master data set that all possible variations flow into the
calculation.