[R-meta] logarithmic scale in forest plot

[Antonia Sudkaemper]

Thanks a lot, Wolfgang! That makes sense, but we have a lot more studies in
our meta-analysis and used the random effects model.

All the best, Antonia

On Wed, 11 Mar 2020 at 11:06, Viechtbauer, Wolfgang (SP) <
wolfgang.viechtbauer using maastrichtuniversity.nl> wrote:

> Yes. (I used method="FE" because there were only 2 studies in the toy
> example, so it would be asking a bit much to fit a random-effects model
> IMHO.)
>
> Best,
> Wolfgang
>
> -----Original Message-----
> From: Antonia Sudkaemper [mailto:a.sudkaemper using gmail.com]
> Sent: Wednesday, 11 March, 2020 11:29
> To: Viechtbauer, Wolfgang (SP)
> Cc: r-sig-meta-analysis using r-project.org
> Subject: Re: [R-meta] logarithmic scale in forest plot
>
> Thank you - now it loos right! :)
>
> Just to double check - if I wanted to run the analysis with a random
> effects model, rather than a fixed effects model, would I just delete the
> "method="FE""? This is what I understand from the metafor package CRAN, but
> I might have misunderstood.
>
> Thank you for clarifying.
>
> All the best, Antonia
>
> On Mon, 9 Mar 2020 at 16:15, Viechtbauer, Wolfgang (SP) <
> wolfgang.viechtbauer using maastrichtuniversity.nl> wrote:
> Ah, sorry, I meant to write:
>
> addpoly(res, row=0, atransf=exp)
>
> Then it's correct (i.e., everything in the plot is then shown as odds
> ratios). Apologies for the confusion.
>
> Best,
> Wolfgang
>
> -----Original Message-----
> From: Antonia Sudkaemper [mailto:a.sudkaemper using gmail.com]
> Sent: Monday, 09 March, 2020 17:11
> To: Viechtbauer, Wolfgang (SP)
> Cc: r-sig-meta-analysis using r-project.org
> Subject: Re: [R-meta] logarithmic scale in forest plot
>
> Hello Wolfgang,
>
> thank you, and apologies - mailing list is now cc'ed in.
>
> What confuses me, however, it that for the average the numbers on the
> right seem to indicate the log odds ratio (i.e. 0), but the poly in the
> graph seems to show the odds ratio (i.e. 1) - I would have expected these
> to be congruent?
>
> Also, for "Study 1" and "Study 2",  do the numbers and the graph display
> the odds ratio or log odds ratio? Based on the numbers and the back
> transformation I am assuming the odds ratio?
>
> Thank you very much!
>
> All the best, Antonia
>
> On Mon, 9 Mar 2020 at 15:31, Viechtbauer, Wolfgang (SP) <
> wolfgang.viechtbauer using maastrichtuniversity.nl> wrote:
> Please always cc the mailing list.
>
> For log odds ratios, the CI can be negative. After back-transformation
> (i.e., exponentiation), the CI will be >= 0.
>
> Similarly, the estimated average log odds ratio is 0 in the example.
> Hence, after back-transformation, the estimate will be 1 (exp(0) = 1).
>
> Best,
> Wolfgang
>
> -----Original Message-----
> From: Antonia Sudkaemper [mailto:a.sudkaemper using gmail.com]
> Sent: Monday, 09 March, 2020 15:05
> To: Viechtbauer, Wolfgang (SP)
> Subject: Re: [R-meta] logarithmic scale in forest plot
>
> Hello Wolfgang,
>
> I just ran this with our data and it seems to work - thank you so much!
>
> I just noticed that both in your example as well as with our data the mean
> weighted effect size lower CI is below 0 - does that make sense?
>
> Also, in your example and similar with our data, it says 0.00 as mean
> weighted effect size, but shows up in the graph at 1. Is that correct?
>
> Thanks again for your help.
>
> All the best, Antonia
>
> On Wed, 4 Mar 2020 at 15:56, Viechtbauer, Wolfgang (SP) <
> wolfgang.viechtbauer using maastrichtuniversity.nl> wrote:
> Dear Antonia,
>
> You should use log odds ratios as input. In fact, we never meta-analyze
> the odds ratios directly, because they are not 'symmetric' around 1. For
> example, an OR of 2 in favor of the treatment group corresponds to an OR of
> 1/2 = 0.5 in favor of the control group, so if one were to combine these
> two ORs (assuming equal weights), one would end up with a combined OR of
> 1.25 in favor of the treatment group, but this is not logical if the two
> ORs are exact opposites of each other. On the other hand, after log
> transformation, we get log(2) = 0.6931472 and log(0.5) = -0.6931472, which
> are now symmetric around 0. Hence, the combined log OR is 0 and, after
> back-transforming (i.e., exponentiation), we get exp(0) = 1, which makes
> sense.
>
> Also, how were those standard errors (sei) computed? Are they really SEs
> of the ORs? In most cases, the SEs we can extract from papers are for the
> log odds ratios.
>
> But to answer your question, here is how it's done. Plot the log odds
> ratios and then use atransf=exp, which in essence puts the the x-axis on a
> log scale. An example:
>
> yi  <- log(c(2, 0.5))
> sei <- c(.20, .20)
> forest(yi, sei=sei, atransf=exp, ylim=c(-0.5,5))
> res <- rma(yi, sei=sei, method="FE")
> addpoly(res, row=0)
>
> (And I just saw that Michael has also answered with the same concern about
> the use of the ORs.)
>
> Best,
> Wolfgang
>


-- 
Dr Antonia Sudkämper
Researcher, OCR, Cambridge Assessment

www.antoniasudkaemper.com
a.sudkaemper using gmail.com

	[[alternative HTML version deleted]]