Biscuit dough piece data

The experiment involved varying the composition of 
biscuit dough pieces.  Two sets of dough pieces were
measured, a calibration set and a prediction
set. They were created and measured as
two distinct sets, on separate occasions, and do not
result from a random (or any other) split of a larger set.

The files labc.asc (c for cal) and labp.asc (p for pred)
contain the reference data on the composition of the doughs.
The files nirc.asc and nirp.asc contain 700 point nir spectra
for the same doughs. They are listed case by case,
with the 700 points in 87 rows of 8 and 1 row of 4.
The spectral range is 1100-2498nm in steps of 2nm.

The file labc.asc is 40 rows (cases) by 4 columns, the columns being
fat, sucrose, flour, water all in percents.
The percents do not quite add up to 100, since there are other minor
ingredients present, but they add up to nearly 100.

We recommend excluding observation 23 in the calibration
set. It appears as an outlier in most analyses, we suspect
because of an error in the compositional (lab) data. 

The background is explained very briefly in the Stone
& Brooks Continuum Regression paper in JRSS, and in
more detail in the JSFA paper referenced in Stone and
Brooks.

The files labp.asc and nirp.asc contain data for a
further 32 samples, intended for use as a validation
set.  Sample number 21 in this set shows up as an
outlier in at least some analyses and might be
considered for exclusion.  
In Brown, Fearn and Vannucci, JASA (2001), we report 
prediction results on 39, not 31, samples.  In fact we
accidentally used the first 8 samples in the set
of 32 twice, which is the explanation for the
discrepancy.


P.J. Brown, T. Fearn and M. Vannucci
(tom@stats.ucl.ac.uk)
January 2003