Suppose we had a categorical predictor that had more than k > 2 levels. If we coded the different levels using the integers 0, 1, 2,...,k - 1, we would be back to familiar territory with a single quantitative predictor, and we could fit the standard linear model. This is not, however, the usual way to handle a categorical predictor. Why not? That is, what constraints does this model impose that we often do not want?