We have compared various statistical methods to estimate the number of comp
onents that contribute to a set of spectra. The methods are tested both on
simulated and on experimental data. No assumptions are made about noise lev
el, since this in most experimental situations is unknown. For tests that f
ormally require such information we have devised novel criteria for their p
redictions. The criteria have been integrated with the NIPALS algorithm to
create a routine that in an automated way predicts the number of components
. We find that the methods almost always predict the correct number of comp
onents when the quality of data is high. Also for multi-component samples a
nd at high-noise levels most of these methods make satisfactory predictions
. Those that gave the overall best results were the factor indicator functi
on (IND) and the imbedded error function (IE). The F-test also worked well,
but it has the disadvantage that a significance level must be chosen rathe
r arbitrarily. The residual standard deviation (RSD), the root mean square
(RMS), the chi-squared and the residual percentage variance (RPV) tests als
o gave satisfactory results. Less good were the eigenvalue (EV) and the red
uced eigenvalue (REV). The ability of all indicators to predict the number
of components was significantly improved when the degree of digitalization
of the spectra was increased. (C) 1999 Elsevier Science B.V. All rights res
erved.