Among individual (co) variation
The multivariate model provides statistical support for individual-level covariance among the traits tested (LRT comparison of full model to one in which all COVI terms fixed to zero; X228=104.87, P<0.001). Pairwise correlations between traits in ID are generally strong (Figure 3a) with a median absolute magnitude of 0.736 (range -0.95 to 0.963). Furthermore, correlations among traits within- and across- assays are of similar magnitude. However, contrary to our prediction of a simple shy-bold axis of variation, the among-individual correlations between (transformed) trait pairs are not uniformly positive (Figure 3a). This result is also reflected in the eigen decomposition whereidmax captures 61% (95% CI, 46.82%-72.23%) of the among-individual (co)variance, consistent with a strong axis of personality underpinning the observed behaviours, but loads antagonistically on some traits (Figure 4, blue points). These loadings show that individuals appearing bolder than average as measured by most traits (e.g., OFT Track Length , OFT -(Freezings) ), tend to appear less bold than average as measured by OFT Wall Distance , FST -(Food Latency) , and FST Time in Open . Of these traits, only FST Time in Open clearly loads significantly onidmax (based on 95% CI not overlapping zero), but this is particularly notable as it indicates shrimp we might view as bolder based on most other criteria actually spend more time than average in the shelter during FST. We present the full IDvariance-covariance matrix from which the correlations are derived in Table S3.