--- title: "Getting started with masque" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{Getting started with masque} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.width = 7, fig.height = 4.5, dpi = 150 ) old_options <- options(width = 100) set.seed(1) ``` ```{r library} library(masque) ``` ## Why A field agronomist has a confidential 72-plot trial and wants an outside statistician to build a spatial yield model on it, without the statistician seeing the real plot layout or genotype identities. This vignette shows the agronomist's side of that handoff on the John and Williams (1995) alpha-design trial shipped with the package. The synthetic table `masque` produces is a development surrogate for one named collaborator. It is not an anonymised dataset and is not safe for public release. *Confidentiality and the threat model* sets out what is and is not protected; read it before sharing any output. ## What `masque()` is the main function. Point it at a table, or a folder or workbook of related tables, and it reads the data, proposes a masking plan and masks. In an interactive session it pauses first so the custodian can review the plan. The output of `masque()` is always a pair: the *synthetic* clone, the only part that crosses the trust boundary, and a private `masque_recipe` (see `recipe()`) that never leaves the custodian's machine. The recipe is what lets a pipeline written against the synthetic re-target to the original later with no code change; its internals and the round-trip verbs are the subject of *Recipe anatomy and the round-trip*. Every column in the masking plan has a `role` and an `action`, proposed by `propose_roles()` and changed with `set_role()`. The role says what the column is: `design`, `treatment`, `outcome`, `covariate`, `date`, `id`, `text` or `other`. The action says what happens to it: * `keep` passes it through unchanged; * `scramble` re-simulates numerics through a Gaussian copula and shuffles categories and dates between rows. For a treatment it swaps the labels among the levels, so the allocation is unchanged; * `alias` replaces the labels with opaque codes (a categorical covariate is shuffled between rows first); * `drop` leaves it out. `role_options()` lists every role and action pair the package accepts. The `mode` argument sets the defaults for both axes at once: * `local` is for the custodian's own development and keeps labels intact. * `collaborate` is for handing the synthetic to someone else. It aliases treatment and categorical labels, jitters numerics, drops identifiers and free text (a character column in which more than half the values are distinct), and runs the leakage audit. A per-column `action` always overrides the mode's default. ## Example ### The one-call path ```{r one-call} f <- system.file("extdata", "john_alpha.csv", package = "masque") df <- read.csv(f, stringsAsFactors = TRUE) knitr::kable( head(df), caption = "The first six plots of the alpha-design trial." ) m <- masque(df, mode = "collaborate", seed = 1L, ask = FALSE) ``` `ask = FALSE` skips the interactive review, which is what a vignette needs; in an interactive console, `masque(df)` shows the proposed plan and a prompt to proceed, edit, or stop. The message above is the summary `masque()` prints for every guided run: how many columns were masked, in which mode, and the leakage-audit tally. The result carries the synthetic data and the private recipe: ```{r one-call-extract} synth <- synthetic(m) knitr::kable( head(synth), caption = paste0( "The first six plots of the synthetic clone: `gen` is aliased, ", "`yield` is re-simulated." ) ) ``` ### The masking plan: roles and actions Under the one-call path, `masque()` built the roles table without asking. Building and reviewing it directly gives full control: ```{r plan} plan <- propose_roles(df, mode = "collaborate") knitr::kable( plan[, c("col", "role", "action", "kind")], caption = paste0( "The proposed masking plan for the alpha-design trial ", "(collaborate mode)." ) ) ``` `propose_roles()` fills in a sensible action for each column given the mode, so the table reviewed here is the plan that will run. `yield` was proposed as the trial's `outcome` from its name. The role records intent, and the *conditional* clone (see *Confidentiality and the threat model*) reads it to find the response. Any role can be changed with `set_role()`. Here the plot number is re-roled as an identifier, which collaborate mode drops: ```{r set-role} plan_id <- set_role(plan, "plot", role = "id") knitr::kable( plan_id[, c("col", "role", "action")], caption = "The plan after re-roling `plot` as an identifier." ) ``` Re-assigning a role re-resolves the default action; passing an explicit `action` pins the column instead. There is no requirement to name an outcome: with none marked, every scrambled numeric is re-simulated jointly. ### Editing the plan as code The plan is a data.table, so you can edit it in code. The spreadsheet the guided prompt opens when the custodian chooses `e` edits the same table. `set_role()` is vectorised over columns: ```{r edit-vectorised} r2 <- set_role(plan, c("row", "col"), action = "drop") ``` Direct edits work too. A direct `role` edit leaves `action` untouched, so setting the action back to `NA` tells `mask()` to re-resolve the default for the new role: ```{r edit-direct} r2[col == "rep", `:=`(role = "covariate", action = NA_character_)] knitr::kable( r2[, c("col", "role", "action")], caption = paste0( "A hand-edited plan: `row`/`col` dropped, `rep` re-roled to a ", "covariate." ) ) ``` The `kind` column comes from the column's class and cannot be edited. If it is wrong, convert the column in the data and run `propose_roles()` again. `role_options()` lists the pairs the package accepts, here for a factor column: ```{r role-options} knitr::kable( role_options(kind = "factor"), caption = paste0( "Every (role, action) pair the validator accepts for a factor ", "column." ) ) ``` Finally, `pii_suspected`. `propose_roles()` sets it from the column *name* (`email`, `phone`, `owner`, ...), and the leakage audit treats a flagged column that survives into the synthetic as a HIGH finding. The scan reads names, not content, so when a harmlessly named column holds sensitive values, flagging it by hand is the correct fix: ```{r pii-by-hand, eval = FALSE} plan[col == "comments", pii_suspected := TRUE] ``` Passing the edited table back to `mask()` (or to `masque(df, roles = plan)`) runs the plan: ```{r mask} m <- mask(df, plan, mode = "collaborate", seed = 1L) synth <- synthetic(m) byte_identical_plot <- identical(synth$plot, df$plot) gen_levels_changed <- !setequal(levels(synth$gen), levels(df$gen)) byte_identical_plot gen_levels_changed head(levels(synth$gen)) ``` ### A plan the package refuses `plot` is a `design` column. Scrambling it would break the trial's layout, and `mask()` refuses: ```{r fail-closed, error = TRUE} bad_plan <- set_role(plan, "plot", role = "design", action = "scramble") mask(df, bad_plan, mode = "collaborate", seed = 1L) ``` The call stops before any masking, names the column and the combination at fault, and lists the three actions that do work for a design column (`keep`, `alias`, `drop`). The same check runs whether the plan came from `propose_roles()`, a hand edit or the guided spreadsheet. ### More than one table A confidential dataset often arrives as several related files or a multi-sheet workbook. `masque()` handles those too. Point it at a folder, an `.xlsx` file, or a named list of tables and it masks every table at once, aliasing any shared key (a site code, a genotype name) *identically across tables* so a join of the synthetic tables still resolves. ```{r set} set_dir <- system.file("extdata", "met_set", package = "masque") ms <- masque(set_dir, mode = "collaborate", seed = 1L, ask = FALSE) ms ``` The genotype column `gen` appears in both tables and is masked to the same codes in each, so the field and laboratory tables still join. The warning is about `year`. It is an environment column, but it is numeric, so collaborate mode keeps its real values. To alias it, read the tables, convert `year` to a factor, and mask the list: ```{r set-year} tabs <- read_set(set_dir) tabs$agronomy[, year := factor(year)] ms <- masque(tabs, mode = "collaborate", seed = 1L, ask = FALSE, quiet = TRUE) knitr::kable( head(synthetic(ms)$agronomy[, c("env", "loc", "year", "gen")], 3L), caption = "The first rows of the clone, with `year` now aliased." ) ``` For a workbook, `sheets =` picks which sheets to read, and `info =` names documentation sheets, such as a data dictionary, to set aside: they are not masked and not written. See *Confidentiality and the threat model* for the set-level controls. ### Describing the clone: the data card A collaborator planning an analysis needs to know the shape of the data before any code is written: how many rows, which columns are the design, which is the outcome, how much is missing. `data_card()` writes those facts down without a single value of any column. It reads only the clone and its role table, which `roles()` returns in the clone's own column names: ```{r data-card-roles} knitr::kable( roles(m), caption = "The role table of the clone, as the data card reads it." ) ``` The role table says which columns are design columns, but not which of them give a plot's position or identify the plot. The custodian names those when writing the card: ```{r data-card} card_file <- tempfile(fileext = ".yaml") card <- data_card( m, file = card_file, name = "alpha_trial", spatial = c("row", "col"), unit = "plot" ) writeLines(head(readLines(card_file), 30)) ``` Every count carries a provenance class. `structure` marks a count that is exact in the clone and so equal on the real table, such as the number of missing values or the number of genotypes. `clone_estimate` marks one that was re-simulated, such as the number of distinct yields. `analyst` marks what was named in the call. Level labels never appear: the card says there are 24 genotypes and does not list them. When column names are aliased, the card uses the aliases. For a set, `data_card(ms)` returns one card per table, and `data_card(ms, file = dir)` writes one file per table. ### Multi-environment structure For a multi-environment trial, `detect_design()` answers three questions separately: 1. Which rows belong to each environment? 2. Do the treatments connect the environments, so that they can be compared? 3. What randomisation structure can be recovered within each environment? A block or field layout seen in the data suggests the original randomisation but does not confirm it. A small toy with two environments and three genotypes: ```{r met-scope} met <- expand.grid( env = factor(c("E1", "E2")), rep = factor(seq_len(2L)), gen = factor(c("G1", "G2", "G3")) ) met$yield <- seq_len(nrow(met)) + rep(c(0, 2), each = 6L) ds <- detect_design(met) ds@scope_label ds@env_cols ds@connectivity$status ds@within_design_label ``` Automatic detection is conservative. A column named `env` or `environment`, or a common site-year pattern, is taken as the environment only when no other column explains the structure as well. A site column on its own is taken only when treatments are replicated across sites, so a nested block is not mistaken for an environment. When you know the environment columns, name them: ```{r met-explicit} ds_explicit <- detect_design(met, env = "env") ``` ### Figure: the multi-environment coverage plot `plot()` on a design summary, or `plot_design_summary()`, draws a compact environment overview in base graphics. With `engine = "ggplot2"` it returns a `ggplot2` object you can save with `ggplot2::ggsave()`; printed, it is the figure below. It shows whether coverage is even across environments, which the connectivity flag alone does not. ```{r met-plot, eval = requireNamespace("ggplot2", quietly = TRUE), fig.cap = "Genotypes observed per environment (E1, E2) for the toy multi-environment trial; the subtitle carries connectivity status, component count, and the recovered within-environment design."} plot(ds_explicit, df = met, engine = "ggplot2") ``` ```{r met-plot-skip, eval = !requireNamespace("ggplot2", quietly = TRUE), echo = FALSE, results = "asis"} cat("ggplot2 is not installed on this build, so the coverage figure is", "skipped; `plot(ds_explicit, df = met)` still draws the base-graphics", "equivalent.\n") ``` Both bars are at three treatments, since every genotype in this toy trial appears in both environments. A shorter bar, or a subtitle reporting more than one component, marks environments that cannot go into one connected analysis until you have checked whether the missing coverage is real or an artefact of how the table was put together. High-confidence environment columns feed the masking plan directly: * Local mode keeps environment values byte-identical. * Collaborate mode aliases categorical environment labels in place, so each row stays in its environment and missing values stay where they were. The recipe turns the aliases back into the real labels. A numeric environment, such as year, stays at `keep` and raises a disclosure warning because its values are still visible. The toy's `env` column is categorical: ```{r met-roles} met_roles <- propose_roles(met, mode = "collaborate") knitr::kable( met_roles[met_roles$col == "env", c("col", "role", "action")], caption = "How the masking plan treats the `env` column." ) ``` This safeguard protects the allocation a pipeline reads. It does not imply that the synthesised outcomes preserve genotype-by-environment effects: sparse treatment-by-environment cells may fall back to pooled synthesis, and the clone must not be used as a substitute for the original trial in scientific inference. ## Results ```{r audit-tally, include = FALSE} audit <- audit_mask(m, print = FALSE) grades <- c(audit$leakage_class, attr(audit, "joint")$leakage_class) n_high <- sum(grades == "high") n_medium <- sum(grades == "medium") ``` On this trial the audit, counting its column and table-level checks together, gave `r n_high` HIGH and `r n_medium` medium findings. A HIGH finding would stop the package writing the synthetic to disk. A design column left at `keep` comes back unchanged and an aliased treatment comes back as `trt_001` to `trt_024`; a plan that asks for something the package cannot do is refused before it runs. Related tables are masked together and still join. The agronomist can therefore hand over a clone that keeps the plot layout and the links between tables, with the genotype labels aliased, and the statistician can build a full pipeline on it without seeing the real genotype names. How the finished pipeline runs on the real data is covered in *Recipe anatomy and the round-trip*. ## Limits This vignette shows the custodian's side of the handoff, on one small public trial with a clean, complete design; it does not exercise the leakage audit's failure modes, the geographic-coordinate controls, or the conditional clone that preserves a treatment effect rather than only a marginal distribution, all of which live in *Confidentiality and the threat model*. A sensitive column with an ordinary name is caught only if someone flags it by hand. Multi-environment detection can miss a real environment when the names and replication are ambiguous; prefer an explicit `env` whenever you know better. Only one numeric column is re-simulated here, so the vignette does not show what the Gaussian copula keeps of the relationships between columns; that is also in *Confidentiality and the threat model*. ## What to read next *Confidentiality and the threat model* sets out exactly what a synthetic clone protects and what it does not, the depth controls beyond role and action, the leakage audit, and the conditional clone that preserves a treatment-to-outcome relationship. *Recipe anatomy and the round-trip* is the analyst's side of this handoff: what the recipe carries, and how a pipeline built on the synthetic re-targets to the original. *What a conditional clone keeps of the design* tests what a model fitted on a conditional clone still finds, on eight field designs. ## Reproduce `set.seed(1)` is set once for the document, and every `masque()` or `mask()` call passes `seed = 1L`, so each is reproducible on its own. Package versions follow. ```{r session-info} sessionInfo() ``` ```{r reset-options, include = FALSE} options(old_options) ```