xtreg or xtmixed advice?

[Ted Bartholomew]

Hi, all,

I'm new to Stata, transitioning mostly from point and click analyses...so I apologize in advance for what may seem like ignorant questioning. I'm working with some psychotherapy data collected from 72 clients who vary in number of sessions attended (as few as 2 and roughly as many as 20). I am collecting a symptom distress measure as the outcome variable at every session. In addition, I have data for clients' resilience and perception of working alliance at every session and hope every third session. I'm wanting to enter resilience, working alliance (with therapist), and hope as predictors of symptom distress in a longitudinal model, but I'm a bit lost as to whether I should be using xtreg or xtmixed. I am thinking xtmixed because the responses are nested by clients (and clients nested by therapists, but I have too few therapists for now in the dataset). Any insight or comments would be very much appreciated!

[Clyde Schechter]

If you want a 3-level model, with visits nested within patients nested within therapists, then you cannot use -xtreg-. -xtreg- only handles one level of nesting. If however, you have only a handful of therapists, and you use therapist as a fixed effect, then you can use either -xtreg- or -mixed- (-xtmixed- is an old name going back to version 12; Stata still accepts that command name for now, but -mixed- is the current correct name.) When we are talking about just two levels and a random intercept, they are equivalent. Equivalent, that is, in that they estimate the same model. The numerical methods used for estimation are different so the results, although very close, may not be exactly the same.

If you add any complications, such as random slopes, or more than 2 levels, then -xtreg-'s capabilities are exceeded and you must use -mixed-.

[Ted Bartholomew]

Thanks! So, this is tentatively what I've put together for the command:

mixed Distress total_resilience total_alliance total_hope || clientid:

I'm guessing this is the most simple way to produce the model? Also, because this is then longitudinal, are the coefficients (which are all significant and negative when I run the above model) indicate that as these variables increase over the course of treatment, distress tends to decrease?

Thanks again for your help and patience!

[Clyde Schechter]

Yes, a negative coefficient in a regression model means that the expected value of the outcome variable is lower when the predictor is higher.

Added: You may want to add therapist to the model as well. Having only a small number of them, it is probably best done as an additional fixed effect.