[R-meta] clubSandwich: Computing covariance matrix for autoregressive effects

Dear James,

Thank you for your response. I must clarify what I mean by "time" here.

In our dataset below (NEW_data), we have 3 senses of time.

(1) The column titled "time_wk" represents the time interval between each
post-test and the last treatment in each study in weeks.
(2) The column titled "time_cat" represents the categorized (1, 2, 3, 4)
version of "time_wk".
(3) The column titled "post_id"   represents the index of each post test
(1, 2, 3, 4). [note: this column doesn't necessarily match "time_cat"]

We don't have a "pre-test indicator" because the effect sizes we are
computing are obtained by obtaining the pre-post standardized mean change
for each specific time interval (e.g., pre to post 1) in each treatment
group and then subtracting pre-post standardized mean change for the
corresponding interval in the control/comparison group from that.

Indeed, the time interval between each post-test and the last treatment in
used each study is so variable across the studies that focusing on effects
at each time point seemed not very reasonable.

I hope I have described our effect sizes clearly, but basically "pre-test"
is involved in all our effect sizes. So, can we, in any way, benefit from
an "ar1" structure in ***impute_covariance_matrix()*** given our coding of
"time"?


Indeed, can we generally think of "time" in the traditional longitudinal
sense or we should use "post_id" when modeling our meta-analysis using
metafor::rma.mv()?

Sorry for the long message, please see our datset below.
Fred
---------- HERE IS OUR DATASET with 3 senses of "time":
NEW_data <- read.csv("https://raw.githubusercontent.com/ilzl/i/master/z2.csv
")

Hi Fred,

The matrices are NPD because you have studies where the same time-point is
repeated across multiple rows. With the AR1 structure, this leads to
perfect correlation between effect sizes that have the same time-point.

Currently, the AR1 structure implemented in impute_covariance_matrix() is
designed only for data structures where the dependence arises from multiple
time points, but not also from multiple outcomes. If you (or others on the
listserv) have ideas for other structures that can accommodate multiple
time points and multiple outcomes, I am open to implementing them in
clubSandwich.

James

On Apr 30, 2021, at 8:22 PM, Farzad Keyhan <f.keyhaniha at gmail.com>

wrote:

?Dear List Members,

I'm trying to form an autoregressive variance-covariance matrix for my
studies using the clubSandwich package. But for most of my studies, the
resultant matrices seem to be non-positive definite as suggested by the
warning message. My reproducible R code is below.

Am I missing something?

Thank you all, Fred
-------------
library(clubSandwich)

dat = read.csv("https://raw.githubusercontent.com/ilzl/i/master/z.csv")

impute_covariance_matrix(vi = dat$vi, cluster = dat$studyID,
                             ti = dat$time, ar1 = .7)

Warning message:
In check_PD(vcov_list) :
 The following clusters have non-positive definite covariance matrices:
A1
B1
B2
B3
B4
B6
D1
D2
D3

   [[alternative HTML version deleted]]

_______________________________________________
R-sig-meta-analysis mailing list
R-sig-meta-analysis at r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-meta-analysis