Skip to content
Prev 2623 / 5636 Next

[R-meta] Construct the covariance-matrices of different effect sizes

Hi James,

Thank you so much for the response and further explanations around the
merits associated with V-matrices in correlated datasets. I think that both
you and Wolfgang have already provided a thorough insight on this in
previous threads in the mailing list and papers as this recent one
https://osf.io/x8yre/
However, I struggle to be confident with the formulas I use for each effect
index, and since my math background isn't wide, I found it hard to extract
them from papers and I'll be super grateful to get your help here :)

Since I don't have enough information on the correlation between outcomes,
I'm using 'guesstimate' values and back it up with sensitivity analysis (at
first glance, with my current formulas for V-matrices, values range from 0
to 0.9 correlation did not yield changes in the results).

(1) I might start with my Pearson correlation and ICC, because they are
probably the tricky ones, and up to now I don't really have something to
support on besides impute_covariance_matrix formula. For both datasets I
used r to Fisher's z transformation to approximate normal distributed
datasets.

ICC is simply the ICC between test-retest (submaximal heart rate values).
ICCs from the same sample can be derived from the number of test-retests
(i.e. more than one paired tests), different intensities, different test
duration etc., which I investigate  separately as meta-regression. However,
I'm first interested in the average effect.

Pearson correlations are derived from the association between heart rate at
submaximal intensity(ies) and criterion measure of aerobic fitness (let's
say max test). In theory, we expect an inverse relationship?lower heart
rate at submaximal intensity associated with higher result in max test.
Similarly to ICC, these effects are obtained from different (submaximal)
intensities, time-points across the season etc. I want to extract the
average effect, and then continue to moderator analyses.

## these are datasets examples

es.id study.id  n  icc     yi     vi
1     1         25 0.90 1.4722 0.0455
2     1         25 0.84 1.2212 0.0455
3     2         16 0.72 0.9076 0.0769
4     2         16 0.85 1.2562 0.0769
5     2         16 0.83 1.1881 0.0769
6     3         38 0.92 1.5890 0.0286

es.id study.id  n     r      yi           vi
1           1       25 -0.61 -0.7089 0.0455
2           1       25 -0.58 -0.6625 0.0455
3           2       16 -0.58 -0.6625 0.0769
4           2       16 -0.60 -0.6931 0.0769
5           2       16 -0.56 -0.6328 0.0769
6           3       38 -0.55 -0.6184 0.0286

I tried to implement rcalc() with my dataset, but the thing is that I have
two variables that are constant (submax and submax tests for ICC and submax
and max tests for *r*) that were examined repeatedly within samples, while
the formula requires different paired variables within sample/studies as
the example here https://wviechtb.github.io/metafor/reference/rcalc.html Is
it right? I'm quite lost here!

(2) As you mentioned, I construct the correlation between raw mean
difference using impute_covariance_matrix(). However, as I understand, it
is more useful to use with standardised mean difference effects? The same
is true for raw mean difference?

(3) Yes, my effect size index for SDs is the the log transformed SD (plus
bias correction for sample size) using the method proposed by Nakagawa et.al
2015
https://besjournals.onlinelibrary.wiley.com/doi/full/10.1111/2041-210X.12309
("SDLN"
in escal function) whereby the sampling variance is constructed solely
based on sample size 1/2(n-1). To construct the V-matrix I used Wolfgang
response in the mailing list regarding a similar question (see below)
 https://stat.ethz.ch/pipermail/r-sig-meta-analysis/2019-July/001630.html
Is this the right (or a fair) way to approach this? Does the Delta method
work better here? If so, can you refer me to something that I can work with.

## this is a small example of my dataset

es.id study.id  n   SD     yi     vi
1         1         25 1.45 0.3924 0.0208
2         1         25 1.41 0.3644 0.0208
3         2         16 1.05 0.0821 0.0333
4         2         16 1.20 0.2157 0.0333
5         2         16 1.27 0.2724 0.0333
6         3          38 1.24 0.2286 0.0135

I'll be grateful to get any insight from you and others on the mailing
list..

Thanks and Kind regards,

Tzlil Shushan | Sport Scientist, Physical Preparation Coach

BEd Physical Education and Exercise Science
MSc Exercise Science - High Performance Sports: Strength &
Conditioning, CSCS
PhD Candidate Human Performance Science & Sports Analytics



??????? ??? ??, 20 ????? 2021 ?-1:53 ??? ?James Pustejovsky?? <?
jepusto at gmail.com??>:?