Do we teach regression correctly?

Abstract

Many interesting social phenomena are innately multidimensional and require suitable data modelling tools. Regression modelling (which includes log linear modelling for contingency tables as a special case) is often the „go-to'“ tool. However, much of the math theory was developed for designed experiments (where explanatory variables X are orthogonal and fixed). Conversely, societal data is often observational with random non-orthogonal X. The pedagogic route to data modelling usually starts with linear models before the introduction of the generalised linear models that can address contingency tables. Whilst good textbooks do feature caveats, the search for a parsimonious model is often carried out in a manner that may promote unsafe interpretation of observational data. This poster tries to present a case for a reform in the teaching of regression for observational data.