fabricatr

fabricatr makes the data you do not have yet: you describe the variables and how they relate, and fabricate() builds a data frame that looks like what you expect to collect, so a research design can be simulated and diagnosed before there is anything to analyze. It handles hierarchical and cross-classified data, potential outcomes, and draws with a target intra-cluster correlation.

library(fabricatr)
fabricate(
  villages = add_level(N = 5, wealth = rnorm(N)),
  citizens = nest_level(N = 4, income = wealth + rnorm(N))
)

Installation

fabricatr 2.0 is not yet on CRAN. Until it is, install it from the rewrite branch:

# install.packages("remotes")
remotes::install_github("DeclareDesign/fabricatr@rewrite", build_vignettes = TRUE)

The branch installs under the released name, so it replaces the CRAN fabricatr in your library. To keep 1.0.2 for comparison, put it in a library of its own:

lib102 <- file.path(tools::R_user_dir("fabricatr", "cache"), "lib102")
dir.create(lib102, recursive = TRUE)
install.packages("fabricatr", lib = lib102)
library(fabricatr, lib.loc = lib102)

lib.loc says where to look, not which version to use. In a session that has already attached fabricatr, that library() call is a silent no-op: no error, no warning, and you keep the version you had. Run the two versions in separate R processes.

What changed in 2.0

2.0 is a rewrite of the package on dplyr, tibble, purrr, and rlang. Every function 1.0.2 exported is here, with one addition, declare_level(). A script written for 1.x runs, and each old spelling warns once with the call to write instead. vignette("fabricatr2.0") has the full account; the short list:

• add_level(nest = FALSE) becomes declare_level().
• by = join_using(A, B) becomes .by = c("A", "B") in cross_levels() and link_levels(), and rho is an ordinary argument of link_levels().
• modify_level(by =) becomes modify_level(.by =), and modify_level() named after a level works at that level while keeping the levels nested inside it.
• break_labels and category_labels become labels.
• N is supplied by name: fabricate(N = 100, ...), never fabricate(100, ...).
• total_sd in draw_normal_icc() is a parameter of the draw, not a rescaling, so the same call gives different numbers.
• The correlated draw in link_levels() takes one code path on every machine, with or without mvnfast.
• n() is the number of rows the level is building, wherever N works, and nothing can rebind it.
• Levels are about twice as fast to build, and a 20,000-row three-level design closer to three times, which is where design diagnosis spends its time.

Where to read next

vignette("getting-started") is a first tour for someone new to the package. The book at https://book.declaredesign.org puts fabricated data to work in complete research designs.