On the background of the recent discussion about the appropriateness of linear difference and ratio models for testing interaction effects of social support and stress, it is argued that interaction effects are basically unstable when dichotomized continuous variables are analysed. A data set with known interrelationships of the continuous variables 'stress', 'social support', and 'depression' was simulated and analysed in tabular format. It is shown that the choice of cut-points for the independent variables crucially determines the size and presence of interaction effects in ratio analysis models.