[R-meta] Facing some issues on Meta - regression

[Dr. Gerta Rücker]

Thank you Guido!

There are always two ways for backtransforming:
First, backtransform to a linear y-axis with equidistant proportions (0, 0.1, 0.2, ..., 1). This is what you have now implemented for sm = "PLOGIT", and of course it results in a non-linear regression curve (which is nice).

The other possibility would be to keep the figure as is, i.e., the transformed y-axis and a linear line, and only change the marks on the y-axis. This keeps the regression line linear and is the default for bubble plots following metabin:

m1 <- metabin(ev.exp, n.exp, ev.cont, n.cont, sm = "OR",
              data = Olkin1995, studlab = paste(author, year))
mr1 <- metareg(m1, year)
bubble(mr1)
bubble(mr1, backtransf = FALSE) 

This would be the apparent alternative also for metaprop().

Best,
Gerta


-----Ursprüngliche Nachricht-----
Von: Dr. Guido Schwarzer <guido.schwarzer using uniklinik-freiburg.de> 

Hi all,

The argument 'backtransf = TRUE' was not recognized by bubble.metareg() for the logit-transformation (and other transformations) as this results in a non-linear regression line. The argument worked for relative effect measures like the odds or risk ratio as a logarithmic scale was be used on the y-axis (in this case the regression line is a straight line).

The development version of meta on GitHub allows now to show the back-transformed proportions on the y-axis (with exception of the Freeman-Tukey transformation which needs an average sample size in the back-transformation). For example, metaprop() with logit-transformed proportions:

remotes::install_github("guido-s/meta")
library("meta")
data(Olkin1995)
m <- metaprop(ev.exp, n.exp, data = Olkin1995)
bubble(metareg(m, year)) 
bubble(metareg(m, year), backtransf = FALSE)

Best,
Guido