Covariances set to 0 in pool.compare

[CaroHaensch]

Dear Stef, dear Gerko,

Jörg Drechsler and I noticed quite different results in pool.compare() between mice >3 and mice <3 for the Wald test. I think we pinned it down to the variance-covariance matrix which is not stored completely anymore.

In line 32 of pool.compare(), the variance vector is transformed again into a matrix with diag()

Ubar <- Q %% diag(est1$pooled$ubar) %% (t(Q))

and the covariances are of course then all set to zero.

In the old mice version it was

Ubar <- Q %% est1$ubar %% (t(Q))

The difference in covariance values will lead to different results and we think in case of the newer version to incorrect ones. Li et al 1991 use the variance-covariance matrix too.

Cheers
Carolina

[stefvanbuuren]

Carolina, thanks for alerting. You are right: pool.compare() gives the wrong results.

pool.compare() should have been deprecated in mice 3.0 in favour of D1() and D3(). I noted the problem in https://github.com/stefvanbuuren/mice/blob/master/tests/testthat/test-D1.R, but evidently forgot to deprecate. I will do so in the next release, and remove the function in due time.

My suggestion is to use D1() instead, which does have the full variance-covariance matrix.

Related:
#121

https://stackoverflow.com/questions/51224252/mice-package-in-r-mipo-object-does-not-return-variance-covariance-matrix-anymor

[stefvanbuuren]

mice 3.8.0 deprecates the pool.compare() function.