How EDI Relates to randomizr, carat, coin, and Other Experiment-Design Packages

R has excellent packages for individual pieces of the randomized-experiment workflow. This page maps EDI (Experimental Design and Inference) onto that landscape honestly: what each neighboring package does well, when you should prefer it, and what EDI does differently. It is written by EDI’s author — corrections from the other packages’ perspectives are welcome on the issue tracker.

EDI’s one distinctive idea is that a single design object carries the assignment mechanism from randomization through analysis: you construct a design (fixed or sequential), it assigns treatment, and the inference classes then compute estimates, intervals, and tests matched to that mechanism — including randomization tests that re-draw assignments from the same algorithm that produced the data. Most neighboring packages cover one stage of that pipeline; many of them do their stage with more options than EDI does.

At a glance

Package	Stage it covers	Relative to EDI
randomizr	Assignment declaration (complete, blocked, clustered, stratified)	Assignment only; pairs with estimatr/ri2 for analysis
blockrand	Permuted-block randomization lists for clinical trials	List generation only; no analysis
Minirand	Pocock-Simon minimization	Assignment only
carat	Covariate-adaptive randomization + some associated tests	Closest neighbor on the sequential-design side
coin	Permutation/conditional inference framework	Inference only; general-purpose, not design-coupled
ri2	Randomization inference (Gerber & Green)	Inference for declared designs; pairs with randomizr
estimatr	Design-based estimators with robust/cluster SEs	Estimation only; frequentist asymptotics
pwr	Closed-form power calculations	Analytic power; EDI’s power is simulation-based
simsurv	Simulating survival data	Data generation only
survival, icenReg	General survival modeling (incl. interval censoring)	General-purpose modeling, not experiment-coupled

Assignment-stage packages

randomizr (and its companions estimatr, ri2)

randomizr declares and draws assignments — complete, simple, blocked, clustered, and blocked-clustered randomization — with a clean, widely used API, and it belongs to the DeclareDesign family where estimatr supplies design-based estimators and ri2 supplies randomization inference. That family’s philosophy is close to EDI’s (the design should inform the analysis).

Prefer that stack when you want to declare a fixed design abstractly, simulate over it with DeclareDesign, or you’re already invested in that ecosystem. EDI differs in bundling the pipeline into one object, adding sequential/covariate-adaptive designs (which randomizr does not cover), and providing per-response-type model-based estimators (GLMs, GLMMs, survival, ordinal) alongside the design-based ones.

blockrand and Minirand

blockrand generates permuted-block randomization lists (varying block sizes, stratification) for clinical trials; Minirand implements Pocock-Simon minimization. Both are focused, dependable generators of assignments — and both stop there.

Prefer them when you only need a randomization list to hand to a trial coordinator. EDI differs in implementing the same schemes (DesignSeqOneByOneRandomBlockSize, DesignSeqOneByOnePocockSimon) as live design objects that subsequently drive analysis matched to the scheme — e.g., randomization tests that replay minimization rather than assuming a coin flip.

carat

carat is the closest neighbor on the sequential side: covariate-adaptive randomization procedures (stratified biased coin, Hu-Hu, Pocock-Simon, and others) together with some hypothesis tests derived for those procedures.

Prefer carat when you want specific covariate-adaptive procedures or their dedicated asymptotic tests from that literature. EDI differs in scope: the matching-on-the-fly (KK) family, matched-pair pooling strategies (inverse-variance weighting and combined likelihood), six response types with per-type estimator menus, resampling/exact inference throughout, and the simulation framework.

Inference-stage packages

coin

coin is a mature, general framework for permutation and conditional inference with a rich class of test statistics. It permutes within strata you specify, but it is not coupled to how treatment was actually assigned.

Prefer coin when you need a permutation test outside the designed-experiment setting or a statistic EDI doesn’t offer. EDI differs in drawing its randomization distributions from the design’s own assignment algorithm (blocked, matched, minimized, rerandomized), which is the difference between a permutation test and a randomization test when the mechanism is not exchangeable-uniform.

Power analysis

pwr

pwr computes closed-form power for standard tests — instant and exact under its assumptions. Prefer it when your planned analysis matches a textbook test. EDI differs in computing power by Monte Carlo over the actual design x estimator you plan to use (SimulationFramework), which is slower but answers the question for analyses with no closed form (GLMMs, covariate-adaptive designs, randomization tests), and additionally reports size and coverage diagnostics.

library(EDI)
# power of a matching-on-the-fly design analyzed by matched OLS,
# versus complete randomization analyzed by plain OLS:
sim = SimulationFramework$new(
  response_type = "continuous",
  design_classes_and_params = list(
    DesignSeqOneByOneKK21 = list(),
    DesignSeqOneByOneBernoulli = list()
  ),
  inference_classes_and_params = list(
    InferenceContinKKOLSIVWC = list(),
    InferenceContinOLS = list()
  ),
  n = 100, p = 3, Nrep_W = 1000L, betaT = 0.5,
  results_filename = "power_comparison.csv.bz2",
  continue_from_last_result_row = FALSE
)
sim$run()
SimulationFrameworkReport$new(sim)$summarize()

Survival modeling and simulation

survival, icenReg, simsurv

survival is R’s canonical survival toolbox and icenReg specializes in interval-censored regression; both are general-purpose. simsurv simulates survival data flexibly.

Prefer them for observational survival analysis, model families EDI lacks, or standalone data simulation. EDI differs in embedding survival inference in the experiment pipeline — Cox, Weibull AFT with exact / left / right / interval censoring in one y/y_L/y_R likelihood, RMST, log-rank/Gehan with randomization p-values, and matched-pair frailty models — and its survival simulations run inside the same power framework as every other response type. (EDI itself Imports survival.)

Using them together

These are complements more often than competitors: generate data with simsurv and analyze the designed part with EDI; sanity-check an EDI randomization p-value against coin on an exchangeable design (they should agree there); use pwr for a quick analytic bound before committing to a long simulation. Where a neighboring package is the better tool for your problem, use it — the point of this page is to make that call easy to get right.

# EDI's end-to-end signature: one object from design to inference
library(EDI)
des = DesignSeqOneByOneKK21$new(n = 40, response_type = "continuous")
for (i in 1 : 40) {
  des$add_one_subject_to_experiment_and_assign(X[i, , drop = FALSE])
}
des$add_all_subject_responses(y)
InferenceSuite$new(des)$run_all_inference(screen = TRUE)$results_table