Skip to content

[R-meta] Moderation analysis in IPD meta-analysis

6 messages · Wolfgang Viechtbauer, Michael Dewey, Gerta Ruecker +1 more

#
Dear Metafor?s users,

This is the first time that I post on this mailing list.

Our team is currently planning an individual patient data meta-analysis of
prospective cohorts.

We adopt an IPD approach because no prospective study has yet reported this
association while many should have the data to assess it (i.e., they have
access to the targeted variables but did not report the association).
Unfortunately, we anticipate that most of the included studies will not
have an ethics committee that gives the right to share their data with us.

To overcome this, we plan to ask the authors of the included studies to
perform the analyses on their own data and to share only the results of the
analyses with us. Our dependent variable is binary (DV: yes/no) and our
independent variable is an ordered variable (IV: a scale variable in 12
points [from 1 to 12]) treated as a continuous variable.

We ask authors to perform a logistic regression. Based on their results
(log odds ratio and associated variance), we adopt a classic two-stage
approach. I show the R code for a particular study to highlight our
approach.

#R code for study 1

study1<-glm(DV~IV, data=datastudy1)

yi1<- summary(study1)$coefficients[2,1] #extract the log odds ratio

vi1<- summary(study1)$coefficients[2,2]^2 #extract the squared standard
error

# then, we repeat the same process for each included study

# Once all the effect sizes and their variance are collected, we can store
them within a dataset and run a standard two stage meta-analysis

dat<-data.frame(
yi=c(yi1, yi2, yi3?),
vi=c(vi1, vi2, vi3?),
study=c(1, 2, 3?))

model<-rma(yi,vi, dat)

This is the code for our primary analysis. In an exploratory analysis, we
would like to perform a moderation analysis with a patient-level moderator.
We understand how to perform a moderation analysis for a study-level
moderator but we are not sure on how to implement it with a patient-level
moderator. The aim of this moderation analysis will be to obtain the
estimated average effects for each level of a moderator. I describe here
the approach we have envisaged:



# example of R code for study 1

#let VM denote a participant-level moderator with 3 categories (a,b,c)

study1<-glm(DV~IV*VM, data=datastudy1)

EM1<-emmeans::emtrends(study1, ~VM, var=" IV")

yi1<- as.data.frame(EM1$emtrends)[,2] #extract log odds ratio for each
level of the VM for study 1 (contains 3 values)

V1<-vcov(EM1) # extract the variance/covariance matrix for study 1 (a 3x3
matrix)

# then, we can build a dataset which will look like this...

dat<-data.frame(
yi=c(yi1, yi2, yi3?),
VM=c(a,b,c,a,b,c,a,b,c?),
study=c(1,1,1, 2,2,2, 3,3,3?))

# ...and a variance-covariance matrix using the bldiag function

V<-bldiag(list(V1,V2,V3?))

# Last, we plan to perform a multivariate model in which we leave out the
model intercept and in which we use an unstructured variance structure (if
the model converges).

model<-rma.mv(yi, V, mods = ~ VM-1, random=~VM|study, struct="UN", data=dat)



We were wondering if you could give us some feedback on the correctness of
our approach. We have read in several textbooks that two-stage
meta-analyses are not designed to assess patient-level moderator but, given
that asking for raw data would probably decrease the likelihood of getting
return from authors of primary studies, we would prefer staying at a
two-stage approach.



Thank you very much for your help and for this amazing mailing list!



Corentin J Gosling
Charlotte Pinabiaux
Serge Caparos
Richard Delorme
Samuele Cortese
#
Dear Corentin,

Overall, your approach seems sound. But a few notes:

1) study1<-glm(DV~IV, data=datastudy1) is not logistic regression. You need:

study1 <- glm(DV ~ IV, data=datastudy1, family=binomial)

2) I've only played around with emmeans a little bit, so can't comment on that part. But I don't think you even need it. You can just fit the model in such a way that you directly get the three log odds ratios for the three levels of IV. In fact, the estimates of the three log odds ratios are independent, so one could even just fit three simple logistic regression models that will give you the same results. An example:

dat <- data.frame(DV = c(1,0,0,1,0,1,1,1,0,1,1,1),
                  IV = c(1,3,2,3,5,3,7,7,4,9,6,3),
                  VM = rep(c("a","b","c"),each=4))

# parameterize logistic regression model so we get the three log odds ratios directly
res <- glm(DV ~ VM + IV:VM - 1, data=dat, family=binomial)
summary(res)

# the covariance between the three estimates is 0
round(vcov(res), 5)

# show that the simple logistic regression model for a subset gives the same results
res.a <- glm(DV ~ IV, data=dat, family=binomial, subset=VM=="a")
summary(res.a)

So actually the V matrix corresponding to the three log odds ratios is diagonal. But you still would want to account for potential dependency in the underlying true log odds ratios, so the model

model <- rma.mv(yi, V, mods = ~ VM-1, random=~VM|study, struct="UN", data=dat)

is still appropriate (with V being diagonal, so you can also just pass a vector with the sampling variances to rma.mv).

3) The statement that 2-stage approaches cannot be used to analyze patient-level moderators isn't quite true. If one actually analyzes the patient-level moderator in stage 1 (as you describe), then the 2-stage approach definitely allows you to examine such a patient-level moderator.

Best,
Wolfgang
#
Dear Pr Viechtbauer,

Thank you very much for your answer!

1) Sorry for the family argument, I suppressed it when I copy/paste the
code.
2) I was not aware of this solution in the glm function. I have compared it
with the initial solution using emmeans and it gives similar results. Since
your solution is definitively more elegant, we are going to use it.
3) Great, it is very reassuring to have your confirmation! We had the
feeling that this was feasible but we were afraid to miss the reason
preventing us from applying our approach to a patient-level moderator.

Thank you so much for your help!
Best
Corentin J Gosling


Le ven. 7 ao?t 2020 ? 20:34, Viechtbauer, Wolfgang (SP) <
wolfgang.viechtbauer at maastrichtuniversity.nl> a ?crit :

  
  
#
Dear Corentin

I have not investigated this in detail but there are two packages on 
CRAN you might want to look at.

https://cran.r-project.org/package=multinma

claims to be able to integrate IPD and aggregate data in one analysis

https://cran.r-project.org/package=metagam

which claims to be able to get the other researchers to run the analysis 
and share it with you when they are not allowed to share the data.

As I say I have not looked at any of them in detail so this may be wide 
of the mark but worth a brief look.

Michael
On 08/08/2020 09:14, GOSLING Corentin wrote:

  
    
  
#
Dear Corentin,

you may also be interested in the DATAShield project, see their website 
https://www.datashield.ac.uk/

Best,

Gerta

Am 08.08.2020 um 12:21 schrieb Michael Dewey:
#
Hi everyone,

Thank you so much much for taking the time to reply to us.

@Pr Dewey

We have never heard of the ?multinma? package. I will take a closer look to
the package in the next few weeks. If it allows us to fit both a one-stage
and two-stage meta-analysis in our situation, I will refresh this post so
that every member of the list could have the information.

I have already heard from metagam (thanks to this mailing list if I
remember well). However, when I have read the R code underlying the
critical function handling individual patient data at github, I found (if I
understood it well) that it extracts only parts of the model containing
aggregated data. Therefore, I think that our approach is almost identical
to metagam except that we use a ?home-made? function to extract critical
information and that we target glm and not gam objects. Again, if I
misunderstood something, do not hesitate to correct me.

Thank you very much for your advice!

@ Pr R?cker

Thank you very much for sharing this initiative, none of us knew it.

I am going to present this approach to the DPO of my university, but it
looks extremely promising for our study: the authors of the primary studies
would not have to do the analyses themselves, and this could encourage them
to participate in our project.



Thank you all for your help!

Best

Corentin

Le sam. 8 ao?t 2020 ? 12:42, Dr. Gerta R?cker <ruecker at imbi.uni-freiburg.de>
a ?crit :