Data transformation
Selection was performed based on relative amounts of acetates. However,
describing relationships among relative amounts is problematic, because
they sum to 100%, thereby mathematically constraining the (co)variation
in the pheromone and biasing the analysis. We therefore transformed
pheromone measurements to log-contrasts for all down-stream analyses in
this study. This approach breaks the interdependency and normalizes the
data. Since the divisor used in the contrast of the variable can no
longer be part of any downstream analyses, we chose 14:Ald as the
divisor, because this component has a small but clearly detectable peak
in the chromatogram, while it is irrelevant for male response behavior
(Heath et al., 1990). Prior to downstream analyses, samples with a
χ2-distributed Mahalanobis distance score (calculated
using the ‘mahalanobis’ function in the ‘stats’ package) that exceeded a
threshold value corresponding to a Bonferroni-corrected P-value
< 0.05 were removed, which resulted in removing 64 out of a
total of 2,861 samples. These samples showed abnormal pheromone ratios
and were present across all selection and control lines and generations.
We are therefore confident that these samples represent outliers, e.g.
due to extraction or measurement errors, and are not representative of
any relevant biological phenomena.