The Bias of combining variables on fish's aggressive behavior studies

Abstract

Quantifying animal aggressive behavior by behavioral units, either displays or attacks, is a common practice in animal behavior studies. However, this practice can generate a bias in data analysis, especially when the variables have different temporal patterns. This study aims to use Bayesian Hierarchical Linear Models (B-HLMs) to analyze the feasibility of pooling the aggressive behavior variables of four cichlids species. Additionally, this paper discusses the feasibility of combining variables by examining the usage of different sample sizes and family distributions to aggressive behaviour variables. The subject species were: the angelfish (Pterophyllum scalare), the tiger oscar (Astronotus ocellatus), the Cichlasoma paranaense and the Nile tilapia (Oreochromis niloticus). For each species, 15 groups of 3 individuals were assigned to daily observations (10-min recordings) for 5 days. Aggressive behavior data was labeled according to its aggressive intensity. The variables chase (C), tail beating (TB), push (P), lateral attack (LA) and bite (B) were classified as high intensity. The variables undulation (U), lateral threat (LT) and frontal displays (FD) were classified as low intensity. These behaviors, however, were not present in all species. Model parameters were estimated by Monte Carlo Markov chains using non-informative priors. B-HLMs were performed to assess the impact probability of each variable in the analysis. Results revealed that when combining variables, the resulting distribution is strongly influenced by only one variable in each category. Moreover, in some cases the aggregate values altered the results, which changed the probabilities of the main variables. Species with low aggressive behavior frequencies, such as A. ocellatus, are more sensitive to this bias. LT was the main low intensity variable for all species, while B was the main high intensity variable for the P. scalare and the O. niloticus. LA was the high intensity category variable that was the most relevant for the C. paranaense and A. ocellatus. Moreover, combining the variables did not impact the feasibility of reducing the sample size when compared to using the most quantitative variable. For all species a sample size of 12 did not change the study conclusions. With respect to family distribution, based on DIC values the Gaussian model is more suitable for most of the studied species. However, caution should be taken, because the Gaussian posterior probability distribution overlapped 0 in some cases, which is biologically impossible in aggressive behaviors. The only exception is the A. ocellatus, which, based on DIC values, was the only species better modeled by a Poisson distribution. Bayesian analysis can be therefore considered a strong tool for analyzing aggressive behavior.

Keywords: Aggressive behaviour; Bayesian analysis; Pooled data.