Meta-analysis in brms -- Combining se() and mi()

Interfaces brms
, , ,
1

Hi there, I am using brms to estimate my fairly conventional meta-analysis model as follows:

yi | se(sei, sigma = TRUE) ~ 1 + mods

In my “new to brms” precociousness, I am now wondering if it’s possible to extend this simple model to include additional observations for which sei are missing. For observations with missing sei, I’d like to impute them based on a model, something along the following lines

yi | se(sei, sigma = TRUE) ~ 1 + mods etc
sei ~ 1 + mods (only when sei is missing)

Does anyone have any pointers on how to set-up the formula for such a model? Apologies if the answer is obvious, I’m fairly new to the wonderful world of brms. I did read this vignette on missing values yet remain unclear on the best way to proceed:
https://cran.csiro.au/web/packages/brms/vignettes/brms_missings.html

2

I am not aware of a method for this purpose. At least no such method is implemented in brms.

1 Like
3

thanks Paul useful to know either way. My workaround at the moment is to:
– Estimate the sei model using brms
– Predict values for missing sei using this model
– And then estimate the meta-analysis model

Of course, this approach treats the predicted sei values as deterministic. Beyond running brms multiple times, I’d appreciate any ideas people have …

4

Hi, I was hoping to find a solution here for my data that has a similar missing data structure.

My workaround was doing multiple imputation (with the package mice) on the missing standard errors, but I thought there should be a more elegant way to do this in a bayesian framework (?)

In the vignette cited by @stu.donovan I read

Missing value terms in brms cannot only handle missing values but also measurement error, or arbitrary combinations of the two.

so I expected something like

bf(yi|mi(sei) ~ 1 + mods) + bf(sei ~ 1 + mods)

would work – but it throws an error.

Thank you in advance for any advice!