The warning you are getting is due to ties in the data. This is typically due to ties in the sampling variances (the values specified via the
V argument or, if it is an entire var-cov matrix, its diagonal elements). This is unrelated to whether you fitted the model via
rma.mv() and in fact is unrelated to any code in the
metafor package itself; the warning is generated by the
cor.test() function from the
stats package that is used by the
ranktest() function in
metafor. It just properly warns you that computing an exact p-value for the rank correlation coefficient is problematic when some values in the data are tied. Instead,
cor.test() then uses a normal approximation for the significance test. You can read more about Kendall's tau and hypothesis tests thereof on wikipedia.
The rank correlation test (as implemented in the
ranktest() function) by Begg and Mazumdar (1994) is a simple test for detecting an association between the observed outcomes and the corresponding sampling variances (which may be indicative of funnel plot asymmetry and, in a larger sense and under certain assumptions, may be an indication for publication bias). Note that the test does not account for any moderators in the model and assumes independence between the data points. The
rma.mv() function becomes relevant when fitting multivariate/multilevel meta-analytic models where we are typically trying to account for various forms of non-independence in the data. Therefore, in the best case, the results from the rank correlation test should be treated with caution when the data are not independent, and more typically simply cannot be trusted and should be ignored.
If you are trying to detect an association between the outcomes and sampling variances (or some other measure of the precision of the studies, such as their standard errors or sample sizes), then my suggestion would be to use some version of the regression test (often called Egger's test). There is the convenience function
regtest() that carries out this test, but it currently will not do so for objects fitted with the
rma.mv() function. However, it is easy to carry out a test in the same spirit by adding the sampling variances (or some other measure of the precision of the studies) as a moderator to the model (that is in essence all that the
regtest() function does anyway). This way, you can obtain a test that accounts for any other moderators in the model and also properly accounts for non-independence (assuming that this is what the model fitted with
rma.mv() is doing in the first place).