The principle of obtaining a measure of dependence based on the notion of information gain is examined for normal models and for several bivariate exponential models. It is shown that, under normal models, the principle usually gives good measures. It is also shown that it gives interesting measures for some exponential models.