Message-ID: <bb04492b-b4c4-1d8d-aa7b-c56f61cd7716@dewey.myzen.co.uk>
Date: 2021-12-03T17:11:27Z
From: Michael Dewey
Subject: [R-meta] including uninformative levels
In-Reply-To: <CACw+TffnGJG01ezJ9mE2+8TTipqYVgsC-R4xL503TxLMfxE=JA@mail.gmail.com>
Dear C?tia
In general, in regression, if you have a level in one of the factors
which is represented by a single observation then that observation will
be perfectly fitted and that observation will have high leverage.
In a random-effects model since it should contribute to the between
study variation then it would affect all the estimates since the weights
are dependent on the estimate of tau^2.
Michael
On 03/12/2021 13:29, C?tia Ferreira De Oliveira wrote:
> Hello,
>
> I hope you are well.
> Is there any benefit/disadvantage of including levels in your model that
> only have information coming from a single study? I always remove the
> intercept and do not include it in the follow-up comparisons between levels
> but wonder if there's a consequence to doing it this way since this one
> study will potentially introduce more heterogeneity and does not add much
> empirical value for this particular analysis.
>
> Best wishes,
>
> Catia
>
--
Michael
http://www.dewey.myzen.co.uk/home.html