Multiple linear regression modelling is commonly used to investigate how hormones and body composition interact, but for valid interpretation a sound methodological approach must be used. It is particularly important that the assumptions for regression are met so that spurious associations are not generated. In this article we show how different approaches to building a multiple linear regression model can influence perceived associations, using examples from the literature and our own data related to predicting fasting insulin and leptin levels from total body fat and fat distribution in children.