[R-meta] Non-positive definite variance-covariance matrix

Dear James (and others),

Thank you for your assistance. Especially using the matrixcalc package, it
resolved the problem so far that the problem appears to be related to the
proces I used to formulate by matrices, and probably not content wise. This
because all matrices with 6 dimensions were off, and none of the others.
However, when I check these matrices I simply can't see what is wrong with
them. The data is correctly sorted by ID. I tried using sapply, but this
did not change much. Under here is an example matrix with 6 dimensions,
does anybody else see why it is not positive definite?


.00825680 .00994727 .00959348 .01321687 .01290479 .01133872
.00994727 .00825568 .00937481 .01966662 .01270540 .01118369
.00959348 .00937481 .00824400 .01269867 .01964920 .01392237
.01321687 .01966662 .01269867 .00826429 .00938956 .03499702
.01290479 .01270540 .01964920 .00938956 .00826134 .01393701
.01133872 .01118369 .01392237 .03499702 .01393701 .00818451

Thanks again!

Kind regards,

Erik
------------------------------
*Van:* James Pustejovsky [jepusto at gmail.com]
*Verzonden:* maandag 19 augustus 2019 3:39
*Aan:* E. van der Meulen
*CC:* r-sig-meta-analysis at r-project.org
*Onderwerp:* Re: [R-meta] Non-positive definite variance-covariance matrix

Erik,

A few possibilities occur to me:
1) Is the data frame sorted by ID? If not, then the split-calculate-bldiag
calculations will return a matrix that is not in the same order as the
original data frame.
2) Why do you use sapply rather than lapply in creating the
remove_zero_mat object? Perhaps it makes no difference.
3) To further isolate the problem, it might be useful to check for
positive-definiteness of the component covariance matrices. The matrixcalc
package provides a handy function for doing so:
library(matrixcalc)
lapply(remove_zero_mat, is.positive.definite)

James

On Sun, Aug 18, 2019 at 12:42 PM E. van der Meulen <E.vdrMeulen_1 at uvt.nl>
wrote:

Dear all,

I am trying to run a multivariate meta-analysis for a review. For this
review I included multiple effect sizes from single studies into my
analysis. The number of effect sizes from a single study range from 1 to
36. To account for covariance between effect sizes extracted from the same
sample, I created a variance-covariance matrix for each study with multiple
effect sizes (which is the majority). I am using a syntax I have used
before, in the previous attempt it worked perfectly. However, in this new
study I am continuously ending up with the same error message:

Error in .ll.rma.mv(opt.res$par, reml = reml, Y = Y, M = V, A = A, X.fit
= X,  :
  Final variance-covariance matrix not positive definite.
In addition: Warning message:
In rma.mv(dat$ESP, V, mods = ~1, random = list(~1 | id/nummer),  :
  'V' appears to be not positive definite.

In which V is the variance-covariance matrix I made. As far as I know an
error due to 'non-positive definite matrices' can occur in cases in which
negative or exactly zero eigenvalues appear anywhere in any of the
matrices. As far as I can determine this is not the case. What could be the
problem? If it helps this is full the syntax:


library(metafor) # For meta-analysis
library(clubSandwich) # For cluster-robust variance-covariance matrix
library(foreign)

dat<- read.spss("TEST.sav", to.data.frame= TRUE)

list_mat<- split(dat[ ,c("v1p", "v2p", "v3p", "v4p", "v5p", "v6p")],
dat$id)

remove_zero<- lapply(list_mat, function(x) x[ ,colSums(x) != 0])

remove_zero_mat<- sapply(remove_zero, as.matrix)

V<- bldiag(remove_zero_mat)

PTSD<- rma.mv(dat$ESP, V, mods= ~ 1, random= list(~ 1| id/nummer),
data=dat)

summary (PTSD,digits=3)

In which:
V = the covariance-variance matrices
ESP = the effect size
v1p to v6p = dimensions of the variance-covariance matrices

Thanks in advance.

Kind regards,

Erik van der Meulen





        [[alternative HTML version deleted]]