Abstract: Regression analysis often involves models that are linear in the parameter of interest. A least-squares analysis of data using a linear model yields an estimate of the parameter, which will be normally distributed if the response variable is similarly distributed. For successively larger numbers of observations, there is reasonable confidence that the true value is being approached. For a non-linear model, things become a bit strange. The true parameter value becomes elusive, even when the data are normally distributed. For a small number of observations, things get stranger still. A geometric representation of regression analysis helps clarify what is happening.