Running vertex-wise federated analyses and meta-analyses

Serena Defina

2026-04-08

Quite commonly, the data we want to analyze is annoyingly scattered around. For example, multiple sites/hospitals collect sensitive individual-level data that cannot be shared publicly due to privacy concerns.

We generally strive towards integrating such distributed data because it would increase statistical power and generalizability, and therefore lead to more robust inferences.

The core verywise linear mixed model functionality allows you to account for “multi-site” settings quite flexibly (e.g. using random intercepts (1 | site) and even slopes (age | site)). However this “mega-analysis” approach is only possible when you have direct access to the entire dataset (all sites).

If you do not have access to the entire dataset, there are 2 options to consider.

1. (traditional) meta-analysis

verywise implements vertex-wise meta-analysis via the metafor package.

# Run a meta-analysis
run_vw_meta(
  term = "age", # Which "term" / predictor / effect to pool
  hemi = "lh", # (default) or "rh": which hemisphere to run
  measure = "area", # (default) or any available FreeSurfer metric.
  res_dirs = c("/path/to/study1/results", "/path/to/study2/results"),
  study_names = c("Study1", "Study2"),
  n_cores = 4  # parallel processing
)

[TODO: expand with more details]

2. Federated / distributed analysis

verywise implements a lossless algorithm (i.e. results are identical to running the model on the entire dataset, as in a mega-analysis, but privacy preserving).

[TODO: expand with more details]

STEP 1: at each local site

s1_res = run_vw_fed_local( 
  site_name = "site1",
  formula = vw_area ~ sex + age,
  pheno = pheno_site1,
  subj_dir = "/path/to/site1/local/data",
  outp_dir = "/path/to/site1/partial/results",
  hemi = "lh",
  fs_template = "fsaverage", 
  n_cores = 1)

# [...] Run other models, e.g. rh, thickness... once done: 
compress_local("/path/to/site1/partial/results", "site1")

# Send the "site1.tar.gz" the the aggregating center

STEP 2: at the aggregating center

tot_res <- run_vw_fed_aggr(
  site_names = c('site1', 'site2', 'site3'),  
  formula = vw_area ~ sex + age,
  inpt_dir = "/path/to/tarred/partial/results",
  outp_dir = "/path/to/final/results",
  hemi = "lh",
  fs_template = "fsaverage"
  n_cores = 1)