Skip to contents

Try EDI in Your Browser (webR) — No Installation Needed

Source: vignettes/articles/try-edi-in-your-browser.Rmd
try-edi-in-your-browser.Rmd

You can run EDI (Experimental Design and Inference) entirely in your web browser — no R installation, no compiler, nothing downloaded to your machine beyond the page itself. This works because webR compiles R itself to WebAssembly, and EDI publishes WebAssembly binaries (of the same C++ kernels the native package uses) through R-universe. Every one of EDI’s hard dependencies also has a WebAssembly build, so the whole stack loads in the browser.

Step 1: open the webR REPL

Open https://webr.r-wasm.org/latest/ in a new tab. After a few seconds you get a working R console running locally in your browser — nothing you type leaves your machine.

Step 2: install EDI

Paste this into the webR console (the download is a few tens of MB the first time; give it a minute):

webr::install("EDI",
  repos = c("https://kapelner.r-universe.dev", "https://repo.r-wasm.org"))
library(EDI)

Step 3: run a complete experiment

Design, assign, respond, infer — the full EDI workflow, sized to run comfortably in a browser tab:

library(EDI)

n = 40
X = data.frame(age = rnorm(n, 60, 8), weight = rnorm(n, 80, 12))

# rerandomization: re-draw the assignment until covariate balance passes
des = DesignFixedRerandomization$new(n = n, response_type = "continuous")
des$add_all_subjects_to_experiment(X)
des$assign_w_to_all_subjects()
y = rnorm(n) + 0.5 * des$get_w()          # (simulated outcomes for the demo)
des$add_all_subject_responses(ys = y)

inf = InferenceContinOLS$new(des)
inf$compute_estimate()                    # covariate-adjusted treatment effect
inf$compute_asymp_confidence_interval()   # asymptotic 95% CI
inf$compute_rand_two_sided_pval()         # exact randomization (design-based) test

Or a sequential matching-on-the-fly trial, analyzed by every applicable procedure at once:

seq_des = DesignSeqOneByOneKK21$new(n = 20, response_type = "continuous")
for (i in 1 : 20) {
  # each subject is matched and assigned on arrival
  seq_des$add_one_subject_to_experiment_and_assign(data.frame(x = rnorm(1)))
}
seq_des$add_all_subject_responses(rnorm(20))

suite = InferenceSuite$new(seq_des)
res = suite$run_all_inference(screen = TRUE)
res$results_table

What to expect (honest caveats)

The browser build is for trying EDI, not for production analyses:

  • It is single-threaded. WebAssembly R has no OpenMP, so the parallelized bootstrap/randomization loops run serially.
  • It is a generic build. The wasm binaries carry none of the machine-specific tuning (-march=native, machine-calibrated dispatch policies) that makes native EDI fast. Expect resampling-heavy calls to be one to two orders of magnitude slower than a tuned native install.
  • Memory lives in the tab. Very large simulations can exhaust the browser’s WebAssembly memory; keep n and replication counts modest.

When you’re ready to use EDI for real work, install it natively — and compile from source so the build tunes itself to your CPU:

install.packages("EDI", type = "source")   # builds with -march=native
tune_EDI_for_this_machine()                # calibrates dispatch policies

(Prebuilt native binaries, no toolchain needed, are also available: install.packages("EDI", repos = c("https://kapelner.r-universe.dev", "https://cloud.r-project.org")).)