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

# Re: Re: st: Re: new to xtmixed - basic question

 From "Clyde Schechter" To statalist@hsphsun2.harvard.edu Subject Re: Re: st: Re: new to xtmixed - basic question Date Fri, 22 Jul 2011 07:36:41 -0700

```Ricardo,

First, let me correct a mistake in my previous post.  Instead of

i.treatment##c.time##c.time2

it should have been

i.treatment##c.time i.treatment##c.time2

(an interaction between time and time2 is not needed unless you want to
introduce a cubic time term.)

After that, before just looking at the terms individually I would do an
omnibus test of all terms that include treatment:

-test 2.treatment 3.treatment 2.treatment#time 3.treatment#time
2.treatment#time2 3 treatment#time2-

If that test _is_ significant, then you can separately look at the effect
of treatment on the constant terms, linear terms and quadratic terms.  It
can happen that none of those is individually significant even when the
omnibus test was--that is often difficult to explain. It means that
overall treatment affects the trajectories as a whole although the
particular effects on the constant, linear and quadratic terms are not
quantified sufficiently precisely by the data to localize the effect or
that in the treated groups there are different linear relationships among
the linear and quadratic terms although the overall marginal values of
those terms are not much changed.

If the omnibus test is not significant, then your data do not provide
evidence of any overall impact of treatment on the trajectory of met over
time (or at least no impact that can be captured by a quadratic model).

When you run a model incorporating a treatment term but no treatment X
time interactions, you are fitting your data to a model in which the time
course of met is constrained to look the same in all three treatment
groups, except that there may be a vertical displacement between them.
Otherwise put, the effect of treatment is to give a "boost" to met levels
that persists unchanged over time. If that describes what theory says the
treatments should do, then that is the model whose results you should rely
on.

The model with interaction terms is more general: it allows treatment to
affect the way in which met varies over time.  But if based on the science
in your situation that is not what is expected, the constrained model
without interaction terms may be more powerful for detecting a persistent
boost effect.

Clyde Schechter
Department of Family & Social Medicine
Albert Einstein College of Medicine
Bronx, NY, USA

-------------Original Message-------------------------------

Thank you Clyde,

In my model I did have -treatment- as a categorical variable, I just
dropped the "i." when I posted it to Statalist. When I use  -
i.treatment##c.time##c.time2- in the model all interaction terms are not
significant
and neither is the treatment effect. When I do not interact treatment and
time,
treatment is significant.

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/
```