In Model 4, an ‘offest’ term was introduced to take into account the total number of conflicts per dyad (Appendix Table A1). We also included various interactions between our different variables, especially between different conflict characteristics, between the role of the opponent and conflict characteristics, between the role of the opponent and conflict characteristics or between dyadic characteristics. None of the interactions tested contributed significantly to our models (likelihood ratio tests (LRT), full versus reduced model, all Ps>0.05); hence we excluded them all from the analyses. We transformed continuous variables to improve normality whenever necessary and standardized them to a mean of 0 and a standard deviation of 1 to make estimates comparable. GLMMs, with Gaussian, binomial or Poisson error structures, were implemented with the function ‘lmer’ from the package ‘lme4’ (Bates, Maechler, & Bolker, 2011). For all models, we checked the assumption that residuals were normally distributed and/or homogeneous by visually inspecting plots of the residuals and of the residuals against fitted vales (Field, Miles,& Field, 2012). We also checked for model stability by excluding data points one by one and compared the resulting estimates with those obtained from the full model (Field et al., 2012). Variance inflation factors were derived using the function ‘vif.mer’ and were considered acceptable because all were below 4 (Field et al .,2012). Since we aimed to test general hypotheses about the influence of a set of predictor variables, rather than the influence of single