Specifying the variance-covariance structure of the random effects in 'mixed'

[Vincent Koppelmans]

Dear stata experts,

When I specify different variance-covariance structures in 'mixed' using the covariance() option, the results do not change.

My data looks as follows:
- one continues normally distributed outcome measure Y
- 10 subjects
- 2 groups of subjects (5 subjects per group)
- all subjects were measured 7 times

I included subject as a random intercept in my model:

mixed Y ib4.time##group || subject:, var reml cov(id)

The random part of my model looks like this:

------------------------------------------------------------------------------
Random-effects Parameters | Estimate Std. Err. [95% Conf. Interval]
-----------------------------+------------------------------------------------
subject: Identity |
var(_cons) | 19.52554 10.16633 7.037354 54.17472
-----------------------------+------------------------------------------------
var(Residual) | 4.015717 1.003929 2.460157 6.554861
------------------------------------------------------------------------------
LR test vs. linear regression: chibar2(01) = 44.89 Prob >= chibar2 = 0.0000

However, when I change to syntax to select 'exchangeable' as the type of variance:

mixed Y ib4.time##group || subject:, var reml cov(exc)

the output does not change at al. In fact, ereturn list still shows 'identity' under e(vartypes):

. ereturn list
macros:
e(cmdline) : "mixed Y ib4.time##group || subject:, var reml cov(exc)"
e(chi2type) : "Wald"
e(rstructlab) : "Independent"
e(rstructure) : "independent"
e(estat_cmd) : "mixed_estat"
e(predict) : "mixed_p"
e(redim) : "1"
e(ivars) : "subject"
e(vartypes) : "Identity"
e(revars) : "_cons"
e(cmd) : "mixed"
e(title) : "Mixed-effects REML regression"
e(technique) : "nr"
e(ml_method) : "d0"
e(opt) : "moptimize"
e(method) : "REML"
e(optmetric) : "matsqrt"
e(depvar) : "Y"
e(properties) : "b V"

I use stata SE 13.1.

Any ideas what is going on here?

Thanks in advance!

- Vincent

[Clyde Schechter]

You only have one random effect at the subject level--so the covariance matrix is a 1x1 matrix and looks the same whether you call it identity or exchangeable (or anything else). I think what you mean to specify is -residuals(exchangeable)-. It is those within-id replicate measurements that will have the exchangeable structure, if I understand your situation.

[Vincent Koppelmans]

Hi Clyde,
Thank you for the explanation. This helps.

Best,

Vincent