Development of hamster tracheal epithelium: III. Illustration of statistical methods for proportional data in biology

The data in the two companion papers (McDowell et al., 1985a,b) are of a commonly occurring type in which the dependent variable takes on one of a small number of distinct values; we wish to compare the relative proportions in which these values occur. Student's test is inapplicable. We explain here in detail the linear logistic model that is appropriate, using our data as an example. Several reasonable models must first be fit. We did this by the maximum likelihood method, using a computer. We then used a chi 2 test to decide statistically between models. This process amounts to deciding directly which of the independent variables are important in determining the outcome, and thus it is the center of physiologic interest. The outcome depends on the initial choice of parametrization of the model. Generalizations are required in more complicated, but still commonly occurring, types of data.