Improving Mouse Behaviour Assessment with Generalized Additive Modelling (GAM): The Case of the Home Cage Social Interaction Test

The Home Cage Social Interaction (HCSI) test is a standard tool for assessing locomotor activity and social behaviour in numerous mouse models of psychiatric disorders. Data obtained from this test are commonly analysed with linear modelling. However, linear modelling is unfit for HCSI data because the relation between HCSI response variables and time is cyclic rather than linear. Moreover, the response variables are commonly assumed to follow a Gaussian (normal) distribution, which is pretty much never the case with HCSI data. In brief, the statistical model currently applied to HCSI data is substantially incorrect.We thus propose to employ nonlinear modelling techniques such as General Additive Modelling (GAM) for HCSI data. GAM utilizes smoothing functions that allow for nonlinear relationships between the response variable and the covariates. Moreover, GAM enables the researcher to employ smoothing functions designed for analysing cyclic data. Finally, GAM also implements non-Gaussian exponential family distributions to meet the model’s statistical assumptions. We compare linear modelling and GAM on 27 HCSI statistically independent experiments involving wild-type and mutant mice over seven days. In all the cases, the GAM models outperform the linear models in both explained deviance and inference reliability. Most notably, the linear models tend towards Type I error with regard to the group effect and Type II error with respect to the time-by-group interaction. We thus recommend researchers to adopt the present statistical model to analyse HCSI data.