STATISTICAL ANALYSIS OF LIFE HISTORY TRAIT DATA
We argue that alternatives to popular analytical approaches in life history theory should be considered more often [47], and judiciously applied. Tractable generalisations of life history variation frequently require analyses to collapse multivariate trait datasets and associated covariates into a smaller set of latent variables. Common statistical means of reducing dimensionality have both strengths and drawbacks (Box 3). The most popular approach in this context is the Principal Components Analysis (PCA). Although PCA is a well-established exploratory tool, it should not be used naïvely [48].
First, researchers should be aware of their own interpretations and biases when interpreting dimension reduction analyses. Prior to analysis, careful consideration should be given to the aim of the study: for example, PCA informs on correlative rather than causative relationships between variables. Interpreting results of multivariate analyses can be challenging, as composite axes are combinations of underlying traits. Second, careful consideration should be given to how the chosen data may affect outcomes. Indeed, sets of life history traits that are derived algebraically from vital rate schedules may be correlated due to mathematical relationships as much as natural covariation [49]. If a trait set includes measures that are a function of multiple vital rates (e.g ., generation time), or that are measured in different units (e.g ., energy for metabolic rate and time for mortality), structuring axes from PCAs (principal components) may be challenging to determine. Even if the multivariate trait set is carefully chosen, it is likely to be difficult to draw fair comparisons between studies that use different trait sets and therefore have component axes with different orientations. Third, the statistical shortcomings of analyses must be carefully considered. In particular, a weakness of PCA is that, in its simplest form, it does not consider measurement errors (but see [50]). Fourth, PCAs do not account for non-linearities among life history traits [45]. Finally, and perhaps most importantly, inquiry should be hypothesis-driven: PCA is designed to inform on how variation in multivariate data is partitioned, rather than to test associations between specific traits (Box 3).