[Rich Goldstein]

in the first of your two logistics, there is a constant and category 5 is omitted (becomes the base level); in the second, there is no constant but, of course, the coefficient for category 5 of rep78 is the same as the constant for the first model - so, yes, the LLs and predicted values would be the same for the two models

[Richard Williams]

The question, I think, is how do you want the model parameterized. If you want a coefficient for each level, then use ibn and nocons. Personally, though, I rarely have any great desire for that parameterization. And I am not even sure how to interpret it if you want odds ratios rather than coefficients.

[Rich Goldstein]

I agree

[Jorge L. Guzman]

Thank you, Rich and Richard. I was asking about -nocons- because I am working on a project that finds the probability of running for office based on the human capital of the candidate. I used a logit to estimate the probability. I have the human capital of the candidate divided into four categories: 1. went to graduate school, 2 went to college, 3 finished high school, 4 no info. available. After I created this four dummies, I ran my model and if I put all of these dummies (and several other discrete and continuous variables), it either drops them due to multicollinearity or Stata tells me that the model is perfectly predicted (I think) by these variables. So, do you think that by using -nocons- i would be able to solve this problem? I really want to introduce the human capital variables for a more complete analysis, but I keep getting these errors. Thanks

[Rich Goldstein]

perfect prediction is different; it mean, e.g., that for one variable there is perfect prediction; e.g., if all males run for office and no females run for office; in the future, please show exactly what Stata responds to your command as well as showing your exact command - this is discussed in the FAQ

nocons might cover the issue of including all 4 dummies - remember, however, that in doing so you have changed the null hypothesis being tested; when you have 3 dummies you are comparing each to the fourth (the reference group); when you include all 4 and exclude the constant you are asking if each indicator variable has an effect different from 0 (i.e., if the OR is different from 1)

also, the constant has a role to play (it is the odds when each and every predictor is 0) and this can be useful (if you have coded your variables so that zero is meaningful for each variable)

[Richard Williams]

As Rich G. basically points out and as shown before there are different and equivalent ways to parametrize the same model. A lot just depends on how you prefer to see the results presented. I don't see how using nocons is going to give you new abilities that you didn't have before. Personally I just find nocons confusing.