Normalising the data is sometimes deemed necessary after
baseline correction, depending on the degree of visible variation,
and usually amounts to standardising the area under each spectrum
by multiplying the intensity scale by an appropriate constant.
One particular limitation of this approach, as discussed in Arneberg
et al. (2007), is its vulnerability to heteroscedastic noise in peak
heights (i.e. the tendency for absolute measurement error to increase
as peak intensity increases, leading normalisation to ascribe
more importance to aligning the heights of larger peaks than smaller
peaks). They propose applying an nth root or logarithmic transformation
(see Kvalheim, Brakstad, & Liang, 1994) to the full
spectrum prior to normalisation. Such monotonic transformations
will tend to reduce the range of peak heights across the spectrum
and may therefore lead to improved normalisation (though at the
cost of the relationship between increasing peak height and
increasing component concentration in the sample).