Statistical analyses
We tested differences in the duration, frequency of push-ups and lateral compressions, as well as the distance between the lizards when these behavioral traits were displayed using non-parametric Mann-WhitneyU -test for independent groups. The difference in frequency of aggressive versus submissive behaviors after each of these focal events was tested with a χ2. In addition, to analyze whether self’s (“SELF”) versus rival’s (“RIVAL”) traits were more informative to explain the behavioral traits summarized with the factor analysis, we used a corrected information criterion (AICc) to compare two general mixed models fitted to a Gaussian distribution, where the trial number was included as random factor. Lower AICc is indicative of a more informative model (Kletting et al. 2009).
The ‘SELF’ model included SVL, body condition, spectral chroma and area of blue patch –residual value on SVL–, spectral chroma of yellow patch in the forelimbs, spectral luminance of the back, spectral hue of the throat, ectoparasites (presence of mites and number of ticks), and presence of both hematic and intestinal coccidians of focal individuals.
The ‘RIVAL’ model was built with predictors including the behavioral traits displayed by the rival lizard (factors 1 to 5), and also the spectral chroma of the rival’s yellow patch, which correlates with the body length and condition of the males (i.e. Megía-Palma et al. 2018b). We also included the rival’s spectral luminance and chroma of the blue patch because the first correlates with body length and the second is lower in the males infected by intestinal coccidians (Megía-Palma et al. 2018b). These two spectral variables of the blue patch show low auto-correlation (r2 = 0.02, P = 0.31). Both models were controlled by the time of the day when the staged contest took place.
We based our analyses on information theory to select the best set of likely models, which is a recommended practice in behavioral ecology (Hegyi & Garamszegi 2011; Symonds & Moussalli 2011). The multimodel inference approach uses a corrected criterion for small sample sizes (i.e. AICc) to estimate the relative importance of each predictor (Bedrick & Tsai 1994). We considered sufficiently informative all the models with ΔAICc ≤ 4 in relation to the best model (the one with the lowest AICc). Model selection was performed using the R-package MuMIn (Burnham & Anderson 2004; Barton 2018). We summed the AICc weights of all the models where the predictor appeared (i.e. conditional average) to calculate the relative importance of each predictor. We calculated the significance of each effect based on the z-standardized ßcoefficient and standard error. All the analyses were performed in R version 3.4.3 (R Core Team 2017).