[R-meta] Dear Wolfgang

Dear Ju,

Another (maybe simplistic) solution could be to use only one (e.g., 
always the first) time point for those studies that report repeated 
measurements. This can be justified because you wrote that the "large 
majority of studies measure this over short term experiment and thus on 
a single time point" - so, as I understand, you it is only a minority of 
studies that causes the multiplicity issue.

A similar simplification is often done when including a small number of 
cross-over trials into a meta-analysis of RCTs with parallel-group design.

Best,

Gerta

Am 30.03.2020 um 20:36 schrieb Nicky Welton:

If you are interested in modelling the time-course relationship then 
there is a new package in R to do this within a network meta-analysis 
framework (although it can also be used if there are only 2 
interventions):
https://cran.r-project.org/web/packages/MBNMAtime/index.html

Best wishes,

Nicky

-----Original Message-----
From: R-sig-meta-analysis <r-sig-meta-analysis-bounces at r-project.org> 
On Behalf Of Viechtbauer, Wolfgang (SP)
Sent: 30 March 2020 19:00
To: Ju Lee <juhyung2 at stanford.edu>; r-sig-meta-analysis at r-project.org
Subject: Re: [R-meta] Dear Wolfgang

Thanks for the clarification.

Computing a time-averaged d (or g) value is tricky because the values 
are not independent. So, if you meta-analyze them, the standard error 
of the pooled estimate would not be correct unless you take the 
dependency into consideration. And to answer one of your questions, 
squaring the standard error from the model would give you the sampling 
variance, but again, that value would not be correct.

Basically what you have is the 'multiple-endpoint' case described here:

http://www.metafor-project.org/doku.php/analyses:gleser2009#multiple-endpoint_studies 


You would need an estimate of the correlation between the repeated 
measurements (and if you have more than two time points, then an 
entire correlation matrix) to construct the V matrix for each study 
before meta-analyzing the values. Then you could use:

res <- rma.mv(yi, V, data=dat)

to pool the estimates into a time-averaged estimate, coef(res), and 
vcov(res) would give you the sampling variance. But the difficult part 
is constructing V.

Maybe you can make some reasonable assumptions about the size of 
correlation (which probably should be lower the further time points 
are separated, although if there are seasonal effects, then 
measurements taken during similar seasons - even if they are further 
apart - may tend to be more correlated again). Based on equations 
(19.26) and (19.27) from the Gleser and Olkin? (2009) chapter, you can 
then construct the V matrix.

Best,
Wolfgang

-----Original Message-----
From: Ju Lee [mailto:juhyung2 at stanford.edu]
Sent: Monday, 30 March, 2020 18:16
To: Michael Dewey; Viechtbauer, Wolfgang (SP); 
r-sig-meta-analysis at r-project.org
Subject: Re: [R-meta] Dear Wolfgang

Dear Wolfgang, Michael
These questions are important, and thank you for pointing them out.
In answers to your questions:
1. Studies are measuring predation or herbivory rate on experimental 
prey in control and treatment conditions in the field. Large majority 
of studies measure this over short term experiment and thus on a 
single time point (ex. After 24 hr of field exposure).

2. However, some studies will monitor these responses over long-term 
period over multiple seasons to understand the seasonal dynamics. The 
issue here is that responses show seasonal fluctuation, which is what 
they were looking for. But the reviewer of our study has warned 
against using single representative timepoint within these multiple 
measures (ex. Peak season) but rather time-average these multiple 
measurements to reduce time-related bias.

3. So back to the point: In all of studies, in multiple time point, 
same treatment vs. control group are being compared using the same 
sampling method. We have mean and SD for all of these data point, 
separately for control and treatment groups for each time point of 
multiple measurements.

4. I am using Hedges' d as effect sizes.
Thank you and I hope this clarifies the question better!
Best,
JU