Fig. 3 in Iridescent plumage in a juvenile dromaeosaurid theropod dinosaur

Fig. 3. Melanosome length against diameter for each colour category. The fifth panel shows the measurements for the dromaeosaurid dinosaur Wulong bohaiensis Poust, Gao, Varricchio, Wu, and Zhang, 2020 (DNHM D2933), from Shangheshou, Chaoyang, Liaoning, China, Early Cretaceous Jiufotang Formation with a minimum age of 120.3 Ma. Note 15c, which outlies all extant measurements. This figure is a good visual representation of why the models conflict in their prediction for 15c. cluded in SOM: table S2. Sample 15 is from the chest region Colour prediction.—In this study, the two prevailing preof the abdomen. Multiple unique populations of melano- dictive modelling approaches (QDA and MLR) for analysing somes were observed with differing morphologies. These fossil melanosome shape were applied to each of the datawere treated as separate samples (15a, 15b, and 15c) and sets (see Table 1), using length, diameter and aspect ratio as assessed for their colour. Fig. 3 plots length against diam- predictor variables for the QDA. For MLR, diameter, aspect eter for the melanosome data, with the right-most panel in- ratio, hollowness (categorical) and flatness (categorical) precluding the measurements from Wulong bohaiensis DNHM dictor variables were used. The analyses were conducted in D2933 facilitating visual comparison of the colour category Stata-16 (StataCorp 2019a), see SOM for commands to exedistributions in the first four panels. cute the models and justifications for the variable selection. Accounting for melanosome shrinkage.—Only melanosome The first dataset "LiNord" incorporates the Nordén et al. imprints were available for study, which have been argued (2019) modifications of the original Li et al. (2012) dataset to preserve the original morphology better than organically where, to avoid systematic bias from different sampling preserved melanosomes (Vinther 2020). Organically pre- methods and sample sizes, coefficient of variation and skew served melanosomes appear to shrink isometrically up to variables were excluded, as well as all samples with a sam- ~20% (McNamara et al. 2013; Colleary et al. 2015). The ple size less than 10. It does not include any of the new addiassumption is that if the rock matrix formed earlier than the tional samples from Hu et al. (2018) or Nordén et al. (2019). diagenetic shrinkage taking place, then imprints are a better The second dataset "NordSC" is a revised version of the proxy for the original size. While aspect ratio is unaffected expanded Nordén et al. (2019) dataset. First, a minor correcby shrinkage, the length and diameter would be affected tion reassigned colour categories to four samples that were and could affect predictions. To inspect this effect, three mislabelled (see SOM). Nordén et al. (2019) also excluded the scaling compensations of 10%, 20%, and 30% for length species sampled by Hu et al. (2018) because the data was not and diameter have been applied to investigate any possible phylogenetically broad, instead creating their own, broader shift in prediction. iridescent dataset. Additionally, here all hollow and flat sam-