Different mlogit outcome for separate and*single models

[Valeriia Popova]

I run three separate models looking into a probability of party identification (Republican, Democrat, Independent) given feeling thermometer toward blacks, Muslims, and transgenders (independent variables). As the practice teaches and Stata confirms, for all the categories, higher feeling thermometer increases a probability of being a democrat.
An example:

However, when I run a single model including all three independent variables, feeling thermometer toward blacks goes completely the opposite way of being predicted before: rrr is below 1, which makes the relationship negative and nonsensical: in general, republicans do not favor blacks higher than do democrats.

I honestly have no idea how to interpret this. At first, I though of multicollinearity as blacks are overwhelmingly democrats. But it's not the case since I limited the sample to only whites. Where do I start to interpret and fix this? Please give me some ideas.

[Joseph Coveney]

It's fairly common phenomenon with a list of usual suspects. You could try Googling sign reversal simpson's paradox and taking a look at some of the links that it shows. With logistic regression, you also have noncollapsibility and omitted variable problems.

[Valeriia Popova]

Thank you, Joseph!

[Richard Williams]

These are pretty trivial changes too. If you drop the RRR, you'll see that the coefficients switch from being slightly positive to being slightly negative.