logistic模型是针对分类变量的，拟合模型时，原始数据是不需要计算率也就是你说的p，而是各分类的个数，只要你的样本量足够大，并不需要考虑p是否接近0或1吧，都是可以拟合模型的

3# zhougeyao
有没有文献提及这个问题的？

If your logit model is built upon individual data, then p is not observed. You only observe events which is usually as {0,1}

0 --- no event, good, buy, sucess, etc.
1 --- event,      bad, not buy, failure, etc.

If your logit model is built upon grouped data, then you need to enlarge group size.

1楼回答的不对，对于逻辑模型，p不能太小，这是一个基本的条件；
如p很小，可以使用poisson回归

Many believe that:
Logistic model, theoretically, can fit to any probability between o and 1 but exclusively. However, extremely small probability or sample size will lead to unreliable estimates when the infinity is approaching in the process of calculation. In terms of SAS or other software, you may think of using exact logistic instead.
If n is sufficiently large and p is sufficiently small thus n*p is fixed, Poisson distribution is a very good approximate to Binomial.