Hi, thank you for the answer.
Scaling the covariates does not change the situation.
For example, in the following table I'm showing the coefficients of the complete model (8 covariates):
Bb: estimated coefficients of the binomial model using the counts of category 2 as "y" and category 1 as reference.
B1: estimated coefficients of the Poisson model for category 1.
B2: estimated coeffcients of the Poisson model for category 2.
B3: estimated coeffcients of the Poisson model for category 3.
B1.B2: difference between B2 and B1 (B2 - B1).
Bb B1 B2 B3 B1.B2
Intercept -1.94195 5.43225 5.43225 5.43225 0.00000
1 0.00015 -0.00087 0.00008 0.00079 0.00095
2 -0.02148 0.00508 -0.02062 0.01555 -0.02569
3 -0.00104 0.00561 -0.00117 -0.00444 -0.00678
4 -0.00393 0.00820 -0.00034 -0.00786 -0.00854
5 -0.00222 0.00236 -0.00016 -0.00220 -0.00252
6 0.04084 0.06314 0.05919 -0.12228 -0.00396
7 0.12816 0.04282 -0.08897 0.04610 -0.13179
8 -0.00618 -0.00034 0.00495 -0.00460 0.00529
ft 0.18417 694.29214 157.51790 109.32593 0.22688
In the Poisson model, the intercept is a random variable, so I'm showing the one for the firts observation. I'm also showing the fitted values (ft), calculated "manually", for this firts observation.
Using the Binomial model, the probability for the first observation is 0.18. Using the Poisson model, the expected value for the condition 2 is 157.52 and the real sample size is 783, so the proportion is 0.2 . The expected value for category 3 is also accurate, but is not the case for category 1, however, the predicted value acoording to the model is 516.47, which meets with the restriction, so I suppose that a correction is applied after. Then I have the probability according to B2-B1 wich is not bad either (0.23). Actually, this is a spatial model, and the resulting spatial patterns looks very similar, no matter what model I use. My problem is when I tried to explain each effect. For example, the number 7 is negative for B2.B1 (which has not sense at all) but positive for Bb. The thing is that I'm getting very similar results (fitted values) with both methods, but with very differents coefficients and intercepts. Clearly, a magnitude difference has a lot of sense (logits vs logs, counts vs propabilities), but how a difference in sign can be explained? Is there a way to fix this? Or simply the coefficients can't be interpretated properly if you use the Multinomial-Poisson transformation?