[R] questions about "metafor" package

[Viechtbauer Wolfgang (STAT)]

Dear Emilie,

Regarding your questions:

1) It's not the weighting that is the main issue when you do not have the SDs. The problem is that you need the SDs to calculate the sampling variances of the mean differences (I assume that this is your outcome measure for the meta-analysis). Those are needed to calculate the standard errors of the model coefficients.

There are two possible routes to take. The first would be to try your hardest to get your hands on as many of the missing SDs as possible. Whatever is left missing could be imputed, using a sensible range of values and checking for the robustness of the findings.

The other approach would be to choose some other weights (e.g., sample size weights), then fit the model by WLS, and then estimate the standard errors of the model coefficients using a robust method (e.g., using a "sandwich" estimator).

2) Difficult to say. I haven’t had a chance to read this article, but this will probably tell you more:

Ishak, K. J., Platt, R. W., Joseph, L., & Hanley, J. A. (2008). Impact of approximating or ignoring within-study covariances in multivariate meta-analyses. Statistics in Medicine, 27(5), 670-686.

Best,

-- 
Wolfgang Viechtbauer 
Department of Psychiatry and Neuropsychology 
School for Mental Health and Neuroscience 
Maastricht University, P.O. Box 616 
6200 MD Maastricht, The Netherlands 
Tel: +31 (43) 368-5248 
Fax: +31 (43) 368-8689 
Web: http://www.wvbauer.com 


> -----Original Message-----
> From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org]
> On Behalf Of Emilie MAILLARD
> Sent: Wednesday, August 17, 2011 17:21
> To: r-help at r-project.org
> Subject: [R] questions about "metafor" package
> 
> Hello,
> 
> I would like to do a meta-analysis with the package « metafor ». Ideally I
> would like to use a mixed model because I’m interested to see the effect
> of some moderators. But the data set I managed to collect from literature
> presents two limits.
> 
> -         Firstly, for each observation, I have means for a treatment and
> for a control, but I don’t always have corresponding standard deviations
> (52 of a total of 93 observations don’t have standard deviations).
> Nevertheless I have the sample sizes for all observations so I wonder if
> it was possible to weight observations by sample size in the package
> « metafor ».
> -         Secondly, some observations are probably not independent as I
> have sometimes several relevant observations for a same design. More
> precisely, for these cases, the control mean is identical but treatment
> means varied. Ideally, I would not like to do a weighted average for these
> non-independent observations because these observations represent levels
> of a moderator. I know that the package « metafor » is not designed for
> the analysis of correlated outcomes. What are the dangers of using the
> package even if observations are not really independent ?
> 
> Thank you for your help,
> 
> Émilie.