Frank Harrell states in one of the answers, "...Note that if the model contains only categorical variables and interactions among the variables are not needed, the model must fit the data and no calibration assessment is needed"
Can someone explain the rationale behind this?
In effect it seems to be saying that if the effects of different variables is additive in a log-odds sense (since "interactions among the variables are not needed", a strong assumption which is why it is usually worth including a priori interactions) then you cannot do better than a simple logistic regression where each variable is an indicator of a category (i.e. taking the value $0$ or $1$) and this will extract the useful information from the data
The rationale is thus that the logistic model describes the situation and all that remains to do is estimating the effects of the different categories on log-odds, i.e. do the logistic regression
