The one you mentioned (assuming you are talking about expected prob. of pain for each level of treatment):
(1) in freq = column percent for each level of trt
(2) in logistic = binomial(p), p =(#of 1's)/(#of both 0 and 1's) for each treatment
(3) in glm =cell means = (sum of all 1 or 0's)/# of obs for each treatment
For binary 0/1 variable, all the three cases are the same for sure.
I don't see any reason to use general linear regression to model 0/1 data. Over dispersed and non-normal errors, etc. makes GLM not realistic at all.
至于为什么GLM模型得到的调整概率发生了变化,而且很有可能更加接近真实想要的结果,需要进一步的研究。
你不是说三个结果都相同吗?
JingJu