How to choose a reference category in multinomial logistic regression?

[Manar Serageldin]

Hello,

I have 2 questions if anyone can help:

1) I have 2 groups (treatment and control), and the outcome is categorical (8 categories). I used chi2 square test and found significant difference between treatment and control groups. But now I want to know which one of the categories was affected the most by the intervention (treatment). How can I know this statistically?

2) Now that I thought of multinomial logistic regression to answer the above-mentioned question 1, if I'm right, on what basis shall I choose the reference category? There is no baseline value or category to compare others to, because the main aim is to compare treatment to control. I found it hard to interpret.

[Clyde Schechter]

If you follow-up your -mlogit- with

margins, dydx(name_of_my_group_variable)

Stata will show you the marginal effect of group_variable on each of the 8 outcome probabilities. The choice of the reference category in -mlogit- itself will make no difference to these results as you will get a value for each of the 8 outcome categories.

[Manar Serageldin]

Clyde Schechter
Thank you so much. Would you please help me how to interpret the results of this marginal effect command. I got significant p-values, but I don't know what does dy/dx refers to ?
Can I change it with relative risk for example?

[Clyde Schechter]

The -dydx()- option in -margins- tells Stata to compute a marginal effect. In other words, you are getting, for each of the 8 outcomes, the expected difference between your two groups, in the probability of that outcome. In #1 you said you wanted to know which outcome levels were most different between the groups, so there you have it.