[R-meta] Possible bug in metafor::weights()

[Stefanou Revesz]

Dear Meta-analysis Experts,

I just came across the following post
**https://stat.ethz.ch/pipermail/r-sig-meta-analysis/2022-February/003851.html**
about the weights in rma.mv() models. I have two questions.

1- It looks like the weights() function in metafor doesn't work if a
model has used 'sparse=TRUE' (reproducible example below)?

2- The above-linked post provides a method to detect negative weights
for effect estimates as a way to discover the inadequacy of using a
single r in the V matrix. But I tried that method with Dr.
Pustejovsky's own demo and I still see negative weights (below). So,
are negative weights sometimes OK?

Thanks,
Stefanou

TSL15 <- readRDS(url("https://osf.io/xpmjk/download")) %>%
  filter(!is.na(percoll), !is.na(attrition_all), !is.na(permale))

TSL15_cent <-
  TSL15 %>%
  mutate(
    postwks_c = pmin(postwks, 26) - 12,
    postwks_long = as.numeric(postwks > 26),
    percoll_c = percoll - 1,
    permale_c = permale - 0.5,
    attrition_c = attrition_all - median(attrition_all),
    study_dv = paste(studyid, dv_cat, sep = "-"),
    study_ctype = paste(studyid, Ctype, sep = "-"))


V_mat <- impute_covariance_matrix(TSL15_cent$V,
                                  cluster = TSL15_cent$studyid,
                                  r = .6, smooth_vi = TRUE)

##-- change to sparse=TRUE to see weights() bug
dv_multilevel <- rma.mv(es ~ 0 + dv_cat + postwks_c + postwks_long +
percoll_c + permale_c + attrition_c,
                        V = V_mat, random = ~ 1 | studyid / esid,
                        data = TSL15_cent, sparse = FALSE)

W <- weights(dv_multilevel, type = "matrix")
X <- model.matrix(dv_multilevel)
WX <- W %*% X
B <- solve(t(X) %*% WX)
weighting_mat <- WX %*% B
any(weighting_mat < 0) # <--- TRUE