multivariate meta-analysis with fixed variance-covariance matrix

[David West]

Dear Paul,

Thanks for creating, developing and sharing brms!

I was wondering whether it is possible to fit a multivariate random-effects meta-analysis in brms with known variance-covariance matrix like in http://www.metafor-project.org/doku.php/analyses:berkey1998? It looks like cor_fixed isn't currently intended to be used in multivariate models.

Thanks,

David

[Paul Buerkner]

This depends on how you write it down. In the link you provided, the following multivariate model is fitted:

res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="UN", data=dat, method="ML")

This can be translated into brms as follows:

res2 <- brm(yi ~ outcome - 1 + (outcome - 1 | trial), data = dat, autocor = cor_fixed(V))

--
You received this message because you are subscribed to the Google Groups "brms-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to brms-users+unsubscribe@googlegroups.com.
To post to this group, send email to brms-...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/brms-users/03082ca7-6954-4e62-98b8-018229f68102%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[David West]

That works indeed perfect! Thanks a lot! I have one short follow-up question: why do you use "outcome - 1" (no correlation between random slope and intercept?) in the random effect and not just "outcome" (correlation between random slope and intercept)?

Op vrijdag 6 april 2018 16:14:10 UTC+2 schreef Paul Buerkner:

This depends on how you write it down. In the link you provided, the following multivariate model is fitted:

res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="UN", data=dat, method="ML")

This can be translated into brms as follows:

res2 <- brm(yi ~ outcome - 1 + (outcome - 1 | trial), data = dat, autocor = cor_fixed(V))

2018-04-06 16:07 GMT+02:00 David West <davidw...@gmail.com>:

Dear Paul,

Thanks for creating, developing and sharing brms!

I was wondering whether it is possible to fit a multivariate random-effects meta-analysis in brms with known variance-covariance matrix like in http://www.metafor-project.org/doku.php/analyses:berkey1998? It looks like cor_fixed isn't currently intended to be used in multivariate models.

Thanks,

David

--
You received this message because you are subscribed to the Google Groups "brms-users" group.

To unsubscribe from this group and stop receiving emails from it, send an email to brms-users+...@googlegroups.com.

[Paul Buerkner]

outcome - 1  just implies cell-mean coding of the outcome (that is one coefficient per outcome). metafor seem to do this automatically(?) in rma.mv. This has nothing to do with the correlation, which is estimated by default unless you replace | by ||

To unsubscribe from this group and stop receiving emails from it, send an email to brms-users+unsubscribe@googlegroups.com.

To post to this group, send email to brms-...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/brms-users/d056e67e-7931-4dbf-943e-14871b520c44%40googlegroups.com.

[David West]

OK. Thanks!

Op vrijdag 6 april 2018 16:53:30 UTC+2 schreef Paul Buerkner:

outcome - 1  just implies cell-mean coding of the outcome (that is one coefficient per outcome). metafor seem to do this automatically(?) in rma.mv. This has nothing to do with the correlation, which is estimated by default unless you replace | by ||

2018-04-06 16:45 GMT+02:00 David West <davidw...@gmail.com>:

That works indeed perfect! Thanks a lot! I have one short follow-up question: why do you use "outcome - 1" (no correlation between random slope and intercept?) in the random effect and not just "outcome" (correlation between random slope and intercept)?

Op vrijdag 6 april 2018 16:14:10 UTC+2 schreef Paul Buerkner:

This depends on how you write it down. In the link you provided, the following multivariate model is fitted:

res <- rma.mv(yi, V, mods = ~ outcome - 1, random = ~ outcome | trial, struct="UN", data=dat, method="ML")

This can be translated into brms as follows:

res2 <- brm(yi ~ outcome - 1 + (outcome - 1 | trial), data = dat, autocor = cor_fixed(V))

2018-04-06 16:07 GMT+02:00 David West <davidw...@gmail.com>:

Dear Paul,

Thanks for creating, developing and sharing brms!

I was wondering whether it is possible to fit a multivariate random-effects meta-analysis in brms with known variance-covariance matrix like in http://www.metafor-project.org/doku.php/analyses:berkey1998? It looks like cor_fixed isn't currently intended to be used in multivariate models.

Thanks,

David

--
You received this message because you are subscribed to the Google Groups "brms-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to brms-users+...@googlegroups.com.
To post to this group, send email to brms-...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/brms-users/03082ca7-6954-4e62-98b8-018229f68102%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--
You received this message because you are subscribed to the Google Groups "brms-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to brms-users+...@googlegroups.com.
To post to this group, send email to brms-...@googlegroups.com.