Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

st: Re: st: test the coefficients in multinominal logit model‏

From   Maarten buis <>
Subject   st: Re: st: test the coefficients in multinominal logit model‏
Date   Wed, 24 Feb 2010 17:50:12 -0800 (PST)

--- On Thu, 25/2/10, atsu june wrote:
> I would like to run multinomial logit model for two
> sub-populations which are distinct and test whether a1 for
> X1 in group A differs b1 for X1 in group B.
> for example,  I would like to run the following
> model.
> mlogit educ age inc if men==1
> mlogit educ age inc if men==0
> then test whether the coefficient for age in group of men
> differs that in group of women.

Others have sugested -suest-. An alternative is to use 
interaction terms. However, both -susest- and interaction
terms used in this way test one specific way of thinking
about interaction terms in non-linear models: a ratio
change in effects as measured in relative risk ratios. So
in the example below the baseline rrr is .24 meaning that
we expect to find .24 domestic cars of average price in 
repair status 4 for every domestic car of average price in
repair status 3. This ratio will decrease 5% for every 
1000$ increase in price. This is the effect of price for
domestic cars. The effect of price for foreign cars is
1.79 times larger, i.e. 1.79*.95= 1.70, which means that
the ratio of foreign cars in repair status 4 per foreign 
car in repair status 3 increases by 70% for every 1000$
increase in price. The test statistic next to the 
interaction term will give you the test of the null 
hypothesis that this ratio of effects is 1.

*-------------- begin example ------------------------
sysuse auto, clear
recode rep78 1/2=3
gen byte baseline = 1
sum price , meanonly
gen c_price = (price - r(mean))/1000
mlogit rep78 i.foreign##c.c_price baseline, rrr nocons
*---------------- end example ------------------------
( For more on how to use examples I sent to statalist see: )

Some disciplines like to think of effects in terms of 
marginal effects, in that case testing for interactions 
becomes a lot harder. Actually, the two test slightly 
different hypotheses, see:

Hope this helps,

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen


*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index