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]

Re: st: Advice on xtmixed specification,pre/post two-group design

From   Brandon Olszewski <>
Subject   Re: st: Advice on xtmixed specification,pre/post two-group design
Date   Wed, 27 Jul 2011 08:36:26 -0700

Hi Clyde:

Thanks for replying. The mention of "time" in the "score_change" model
was indeed a typo on my part - no "time" var is included in the change

What is most odd about the situation is how different the results are
between regressing change on inde vars vs. post scores on inde vars
(including the pre score) vs. including time as a dummy. Besides a
recommendation on "which is the preferred method, generally," that
answer is what I'm searching for.


On Wed, Jul 27, 2011 at 7:10 AM, Clyde Schechter
<> wrote:
> I can suggest a couple of things to look into:
> xtmixed score_change female urm english control time#control || tch_id:
> control, cov(unstruct) || id: , mle
> seems mis-specified.  In particular, if your dependent variable is
> score_change, I don't think there should be any reference to time in the
> independent variables.  In fact, with score_change as the dependent
> variable, in your data there should be only a single observation
> corresponding to the pre-post pair of observations in the original data.
> So either you are analyzing largely duplicated observations, or if you
> have reduced to one observation per pre-post pair (you don't say whether
> or how)  whatever value time takes on in these observations is
> meaningless.  Either way, you wouldn't expect to see a time#control
> effect.
> I think if you take time#control out of the model and just focus on the
> control coefficient you will get what you are looking for.
> Other things to bear in mind: most of the time a model with change score
> as dependent variable will give you more or less the same results as a
> model with separate observations at each time and within-pair clustering
> accounted for.  But, if there is a substantial amount of missing data on
> the score variable, you may find that the sample analyzed in the change
> score variable is a noticeably smaller, and probably biased, subset of the
> cases included in the analysis relying on separate pre and post
> observations.
> Hope this helps.
> Clyde Schechter
> Department of Family & Social Medicine
> Albert Einstein College of Medicine
> Bronx, NY, USA
> Please note new e-mail address:
> *
> *   For searches and help try:
> *
> *
> *

*   For searches and help try:

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