Data imputation in multilevel meta analysis brms

Please also provide the following information in addition to your question:

  • Operating System: R 4.0.0
  • brms Version: 2.12.0

Dear all,

I am conducting a multilevel meta-analysis in brms. Some of my predictor variables I was unable to obtain in the studies and in requests to authors. So, somehow I have to impute this data. I used the “mice” package according to the brms vignette. However, the model releases the following message: "Error: Argument 'data' must be coercible to a data.frame"

My question is whether there is a way to impute this data in this model of analysis. Below is a reproducible example of my analysis.

Thank you for your attention!
Best

    #covariance matrix of outcomes
        cor_matrix <- function(x, r, v = rep(1, length(x)), na.rm = FALSE) {
          mat <- diag(v)
          se <- sqrt(v)
          se[is.na(se)] <- 0
          if (length(x) > 1L) {
        for (i in 2:nrow(mat)) {
          for (j in 1:(i-1)) {
            if (x[i] == x[j]) {
              mat[i, j] <- mat[j, i] <- r * se[i] * se[j] 
            }
          }
        }
          }
          dimnames(mat) <- list(1:nrow(mat), 1:ncol(mat))
          if (na.rm) {
        keep <- !is.na(diag(mat))
        mat <- mat[keep, keep]
          }
          mat
        }

    #data
        dat <- data.frame(
          yi = rnorm(20),
          level = factor(c(1, 1, "NA", "NA", "NA", 1, "NA", "NA", 1, 1, 1, 1, 2, 2, 2, 3, 3, 3, "NA", 4)),
          log_time = rnorm(20),
          exer = rnorm(20),
          study_ID = 1:20,
          vi = 0.25
        )

    #data imputation
        library(mice)
        imp <- mice(dat, m = 5, print = FALSE)

    #prior
        prior <-c(prior(normal(0, 100), class = Intercept),
              prior(normal(0, 10), class = b),
              prior(cauchy(0, 10), class = sd))

    #matrix
        v.m1 <- cor_matrix(dat$study_ID, r = 0.9, v = dat$vi)
        cor_obs <- cor_matrix(dat$study_ID, r = 0.9)

    #model
        m1 <- brm(bf(
                yi ~ 1 + log_time + level + exer + (1|study_ID) + fcor(v.m1),
                sigma = 1
              ),
              family = gaussian(), 
              prior = prior, chains = 1,
              data2 = list(v.m1 = v.m1),
              control = list(adapt_delta = 0.99), data = imp)

Sorry, it seems your question fell through - maybe @torkar can answer?

2 Likes

Hi,

not sure I can provide much here except,

  1. brms has direct support for using mice. See Paul’s excellent guide here.
  2. If it’s any help you can check out this which I’ve written.

If you only have missingness in predictors (independent variables/covariates) then I would recommend you to model that directly, i.e., what Paul refers to as: Imputation during model fitting, in the link above.

1 Like

Hi,

Thank you for the answers @torkar and @martinmodrak.

I found the problem, instead of using m1 <- brm(bf... the folowing code
m1 <- brm_multiple(bf... must be used.

However, the imputation during model fitting seems to be preferable from the point of view of computation time.

1 Like

Indeed, I didn’t spot that. And you’re right, brms will re-samples x times. In the old days x=5, but now they say one should set x to the number of % missing in the data. So if you have 20% missingness, then x=20, so brms would need to sample 20 times… :)

1 Like