dummy encoding in metafor
At 22:06 23/01/2013, Alma Wilflinger wrote:
Hi Michael, The supervisor for my Master's Thesis told me that my means are the effect size and cause of this I have to take figure 1 for all standard deviations. So I hope that was the right information.
Alma There is a fairly comprehensive list of all the things which might be an effect size on http://en.wikipedia.org/wiki/Effect_size Is what you call Mean one of them?
From: Michael Dewey <info at aghmed.fsnet.co.uk> To: Alma Wilflinger <alma_anima at yahoo.com>; Viechtbauer Wolfgang (STAT) <wolfgang.viechtbauer at maastrichtuniversity.nl>; Michael Dewey <info at aghmed.fsnet.co.uk>; "r-help at r-project.org" <r-help at r-project.org> Sent: Wednesday, January 23, 2013 10:22 AM Subject: Re: [R] dummy encoding in metafor At 08:30 23/01/2013, Alma Wilflinger wrote:
Dear Wolfgang and Michael, thank you very much for your help! Concerning the Variance: I took the variance I used for CMA
(which is always 1), so I think it should be the right one. It seems unlikely to me that the variance from each study would be the same although I suppose it could be possible. Are you sure you are supplying the right values to CMA?
Thank you for noticing and mentioning though :) I really appreciate how helpful you both are. best, Alma From: Viechtbauer Wolfgang (STAT)
<<mailto:wolfgang.viechtbauer at maastrichtuniversity.nl>wolfgang.viechtbauer at maastrichtuniversity.nl>
To: Michael Dewey
<<mailto:info at aghmed.fsnet.co.uk>info at aghmed.fsnet.co.uk>; Alma Wilflinger <<mailto:alma_anima at yahoo.com>alma_anima at yahoo.com>; "<mailto:r-help at r-project.org>r-help at r-project.org" <<mailto:r-help at r-project.org>r-help at r-project.org>
Sent: Monday, January 21, 2013 11:10 AM Subject: RE: [R] dummy encoding in metafor As Michael already mentioned, the error: Error in qr.solve(wX, diag(k)) : singular matrix 'a' in solve indeed indicates that your design matrix is not of full rank
(i.e., there are linear dependencies among your predictors). With this many factors in the same model, this is not surprising if k is "only" 94 (which is actually quite large for a meta-analysis). One options is to leave out some of the predictors. You can also try collapsing some of the levels of the factors. Of course, you lose some "details" that way, but apparently you don't have enough data in the first place to carry out such a detailed analysis.
One other thing I noticed. You wrote: rma(yi=Mean, vi=Variance, ni=N.1, ...) I suspect that your variable "Variance" is actually the variance
of the raw scores. However, the vi argument is used to pass the sampling variances of the yi values to the function -- not the variance of raw scores. The (estimated) sampling variance of a mean is s^2 / n, so if I am not mistaken, you really want to use:
rma(yi=Mean, vi=Variance/N.1, ...) Best, Wolfgang -- Wolfgang Viechtbauer, Ph.D., Statistician Department of Psychiatry and Psychology School for Mental Health and Neuroscience Faculty of Health, Medicine, and Life Sciences Maastricht University, P.O. Box 616 (VIJV1) 6200 MD Maastricht, The Netherlands +31 (43) 388-4170 | http://www.wvbauer.com
-----Original Message----- From:
<mailto:r-help-bounces at r-project.org><mailto:r-help-bounces at r-project.org>r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org]
On Behalf Of Michael Dewey Sent: Monday, January 21, 2013 10:40 To: Alma Wilflinger; Michael Dewey;
<mailto:r-help at r-project.org><mailto:r-help at r-project.org>r-help at r-project.org
Subject: Re: [R] dummy encoding in metafor At 14:48 20/01/2013, Alma Wilflinger wrote:
Hi, thank you very much for your kind answer.
If you look a bit further down the manual page you will see ### using a model formula to specify the same model rma(yi, vi, mods=~factor(alloc)+year+ablat, data=dat, method="REML", btt=c(2,3))
which is much easier.
I have seen the possibility of using a model formula for dummy encoding and you are right it is much easier than doing it by hand. Thing is that if I include some moderator variables into the parameters I get the error: Error in qr.solve(wX, diag(k)) : singular matrix 'a' in solve
I suspect that you have a linear dependence between your moderator variables. Depending on how many levels there are for country, sample, and so on you do have a lot of predictors (you presumably know that a factor counts as levels-1 for this purpose?)
For example this call works: result = rma(yi=Mean, vi=Variance, ni=N.1, mods=~factor(Country) + relevel(factor(Sample), ref="Students") + Gender + Age + factor(Category) + relevel(factor(Block), ref="c")+ relevel(factor(order), ref="x"), data=csvDataCmaAll, method="REML") If I add the trials which is of type INT: result = rma(yi=Mean, vi=Variance, ni=N.1, mods=~factor(Country) + relevel(factor(Sample), ref="Students") + Gender + Age + factor(Category) + relevel(factor(Block), ref="c")+ relevel(factor(order), ref="x") + trials, data=csvDataCmaAll,
method="REML")
I get the error and I was not able to find a definite reason for this error or how to solve it I wanted to try it by doing it manually. I think I have found out that it somehow relates to the
If you code them yourself R does not know. You know.
Regarding this I think my question was not clear enough. If R does the dummy encoding automatically via a model formula it leaves out one of the factors and uses it as a baseline automatically. If I do it by hand R is still able to execute the function but the baseline is missing because I do not define it via a parameter.
You perhaps would benefit from rereading some of the introductory material about formulas. Also look for anything about the model matrix (also called the design matrix)
I simply want to know how R is handling this and what I have to do by hand to get the correct results. Sorry, this may be a beginners question, but as stated I am new to this field.
You say you have seven moderator variables. Unless you have a shed load of studies you will not be able to look at them simultaneously. Apologies if you already knew that.
No I have not known that. In total I have about 94 studies and want to test different sets of moderators. Do you think this is sufficient or do you suggest another approach?
The truthful but perhaps unhelpful answer is that you need to collect more data or use fewer moderators.
I started in CMA (comprehensive meta analysis) but one of the benefits of R is that I am able to test multiple moderators at once - at least as I was told. kind regards, Alma From: Michael Dewey
<<mailto:info at aghmed.fsnet.co.uk><mailto:info at aghmed.fsnet.co.uk>info at aghmed.fsnet.co.uk>
To: Alma Wilflinger
<<mailto:alma_anima at yahoo.com><mailto:alma_anima at yahoo.com>alma_anima at yahoo.com>; "<mailto:r-help at r-project.org><mailto:r-help at r-project.org>r-help at r-project.org"
<<mailto:r-help at r-project.org><mailto:r-help at r-project.org>r-he
lp at r-project.org>
Sent: Sunday, January 20, 2013 12:52 PM Subject: Re: [R] dummy encoding in metafor At 17:14 19/01/2013, Alma Wilflinger wrote:
Hi, I am quite new to R and in need of some advice. I am trying to conduct a meta regression over a some studies with about 7 mod variables which I have to dummy encode.
Alma, although you can generate your own dummy variables by hand you do not have to as R will do it for you. See below for more comments.
I have found the following piece of code in the manual for the metafor library: ### manual dummy coding of the allocation factor alloc.random <- ifelse(dat$alloc == "random", 1, 0) alloc.alternate <- ifelse(dat$alloc == "alternate", 1, 0) alloc.systematic <- ifelse(dat$alloc == "systematic", 1, 0)
If you look a bit further down the manual page you will see ### using a model formula to specify the same model rma(yi, vi, mods=~factor(alloc)+year+ablat, data=dat, method="REML", btt=c(2,3)) which is much easier.
### test the allocation factor (in the presence of the other
moderators)
### note: "alternate" is the reference level of the allocation factor ### note: the intercept is the first coefficient, so btt=c(2,3) rma(yi, vi, mods=cbind(alloc.random, alloc.systematic, year, ablat), data=dat, method="REML", btt=c(2,3)) What I do not understand is the following: How does R know which columns in my data.frame are related to the dummy encoded variables?
If you code them yourself R does not know. You know.
It is clear that in the call of cbind I just do not use the reference variable as a parameter but I do not get it how R knows that alloc.random and alloc.systematic refer to the column alloc in the data frame. Thank you very much in advance for your help,
You say you have seven moderator variables. Unless you have a shed load of studies you will not be able to look at them simultaneously. Apologies if you already knew that.
kind regards,
Alma
[[alternative HTML version deleted]]
Michael Dewey <mailto:info at aghmed.fsnet.co.uk><mailto:info at aghmed.fsnet.co.uk
>i <mailto:nfo at aghmed.fsnet.co.uk>nfo at aghmed.fsnet.co.uk
Michael Dewey
<mailto:info at aghmed.fsnet.co.uk><mailto:info at aghmed.fsnet.co.uk>info at aghmed.fsnet.co.uk
______________________________________________
<mailto:R-help at r-project.org><mailto:R-help at r-project.org>R-help at r-project.org mailing list
PLEASE do read the posting guide
and provide commented, minimal, self-contained, reproducible code.
Michael Dewey <mailto:info at aghmed.fsnet.co.uk>info at aghmed.fsnet.co.uk http://www.aghmed.fsnet.co.uk/home.html
Michael Dewey info at aghmed.fsnet.co.uk http://www.aghmed.fsnet.co.uk/home.html