Customising the models to compare • metaUI

library(metaUI)
#> Note re metaUI: This package is still under active development - so please report any issues and desired features on https://github.com/LukasWallrich/metaUI.

Adding, editing and removing models

You might wish to show different models. For now, the easiest way to do that is to save the app into a folder, as shown in the vignette(“getting_started”). Then open models.R to edit the models.

Removing models

To remove models, you can add a comment (#) in front of the line where it is first defined in upper part of the file where models_to_run are defined … and ignore the longer part in the models_code below. So to remove the Trim-and-fill model, you simply need to add the # shown below - or alternatively to delete that line.

   "Robust Variance Estimation",                      FALSE,       "as.numeric(mod$reg_table$b.r)",      ... 
#   "Trim-and-fill",                                   TRUE,        "mod$TE.random",                     ...
   "P-uniform star",                                  TRUE,        "mod$est",                            ...

Add or edit models

To add or edit models, you need to understand a bit more about the structure of the file. If contains two parts - firstly, the models_to_run which has the names of the models and the details on how to extract summary data, and then the models_code that contains the code to estimate them.

For models_to_run, you need to provide the following details for each model:

name of the model to be displayed
aggregated TRUE/FALSE value whether it needs to be estimated with the dataset aggregated by sample (i.e. without dependent effect sizes)

And then code within “” that allows you to export the following from the model object, called mod:

es: the effect size
LCL & UCL: the lower and upper bounds of the 95% confidence intervals for the effect size
k: the model-specific count (state whether effects, studies, or imputed effects)

For models_code, you need to provide code, again in ““, that estimates the model based on fields available in the dataset produced by prepare_data() (for models where aggregated = FALSE) or aggregated by metaUI_aggregate() in the generated analysis.R. Aggregation retains the first row’s metadata but does not infer aggregate p-values or N; do not interpret retained metadata as a new scientific aggregation rule.

If your model code is complicated, it might be worthwhile to create helper functions to call from within the models_code entry. Add them above the final models_to_run expression in models.R, e.g., my_meta_fun <- function(...). The app sources model code in its own environment. Declare and install any extra dependencies yourself; the default manifest does not infer them. Custom models must return the declared fitting scale. NA intervals are allowed only when explicitly specified; failed fits are reported.

The default multilevel and RVE specifications call helpers in the generated analysis.R. Study-level aggregation also happens there through metaUI_aggregate; df_agg is exposed by the server but does not infer N or p-values. Editing those helpers requires restarting the app. generate_shiny(overwrite=TRUE) replaces all generated files, including analysis.R; build into a fresh folder and compare changes to preserve hand edits. It never clears unrelated files from the folder.

If you add an author model that is your primary analysis (such as the Dannheim example’s source-reference row), pass its name as generate_shiny(primary_model = ...). Renaming that model later requires the same change in labels_and_options.R; the app reports a declared primary model it cannot find instead of picking another.

Custom models default to options=list(fit_cache_entries=0). Only explicitly opt deterministic code into a 1–3-entry session cache. Random models or code that reads external state must leave it disabled. A cache hit preserves the original fit times, warnings, and status, and is disclosed as cache_hit.