Package 'colleyRstats'

Description

This function calculates the Pareto front using emoa for a given set of objectives in a data frame and adds a new column, PARETO_EMOA, which indicates whether each row in the data frame belongs to the Pareto front.

Usage

add_pareto_emoa_column(data, objectives, maximise = FALSE)

Arguments

Value

A data frame with the same columns as data, along with an additional column, PARETO_EMOA, which is TRUE for rows that are on the Pareto front and FALSE otherwise. Identical rows share one verdict (a copy of a non-dominated point is non-dominated too), as in add_pareto_moocore_column(). Rows with a missing objective value get NA, with a warning, and the front is computed from the complete rows.

See Also

add_pareto_moocore_column(), which answers the same question via moocore and accepts the same maximise argument.

Examples

# Define objective columns
objectives <- c("trust", "predictability", "perceivedSafety", "Comfort")

# Example data frame
main_df <- data.frame(
  trust = runif(10),
  predictability = runif(10),
  perceivedSafety = runif(10),
  Comfort = runif(10)
)

# Add the Pareto front column (minimising, the default)
main_df <- add_pareto_emoa_column(data = main_df, objectives)
head(main_df)

# All four objectives are ratings where higher is better
main_df <- add_pareto_emoa_column(main_df, objectives, maximise = TRUE)

Description

This function calculates the Pareto front using moocore for a given set of objectives in a data frame and adds a new column, PARETO_MOOCORE, which indicates whether each row in the data frame belongs to the Pareto front.

Usage

add_pareto_moocore_column(data, objectives, maximise = FALSE)

Arguments

Value

A data frame with the same columns as data, along with an additional column, PARETO_MOOCORE, which is TRUE for rows that are on the Pareto front and FALSE otherwise. Identical rows share one verdict: every copy of a non-dominated point is kept (keep_weakly = TRUE), as in add_pareto_emoa_column(), where moocore::is_nondominated() on its own would mark only the first copy. Rows with a missing objective value get NA, with a warning, and the front is computed from the complete rows.

See Also

add_pareto_emoa_column(), which answers the same question via emoa and accepts the same maximise argument.

Examples

# Define objective columns
objectives <- c("trust", "predictability", "perceivedSafety", "Comfort")

# Example data frame
main_df <- data.frame(
  trust = runif(10),
  predictability = runif(10),
  perceivedSafety = runif(10),
  Comfort = runif(10)
)

# Add the Pareto front column (minimising, the default)
main_df <- add_pareto_moocore_column(data = main_df, objectives)
head(main_df)

# All four objectives are ratings where higher is better
main_df <- add_pareto_moocore_column(main_df, objectives, maximise = TRUE)

# Mixed: maximise the ratings, minimise a workload score
main_df$workload <- runif(10)
main_df <- add_pareto_moocore_column(
  main_df,
  c(objectives, "workload"),
  maximise = c(TRUE, TRUE, TRUE, TRUE, FALSE)
)

Description

One-call pipeline for a single dependent variable: checks the assumptions (producing a ready-made methods sentence via assumption_methods_text()), builds the matching ggstatsplot figure with automatic parametric/non-parametric selection, reports the omnibus test via reportggstatsplot(), and – for more than two groups – reports the significant post-hoc comparisons via reportggstatsplotPostHoc().

Usage

analyze_and_report(
  data,
  dv,
  iv,
  design = c("between", "within"),
  ylab = dv,
  xlabels = NULL,
  plotType = "boxviolin",
  sink_to = NULL,
  subject = NULL
)

Arguments

Details

Every part of the output is computed from one data set: the rows with both dv and iv observed and, for design = "within", only the participants measured in every condition (the others are dropped with a message naming them; more than one row per participant and condition is an error). The methods sentence, the test, the figure and the means quoted in the post-hoc sentences therefore all describe the same participants.

Value

Invisibly returns a list with components plot (the ggplot), methods (assumption-check sentence), text (omnibus result), posthoc (post-hoc sentences, or NULL for two groups), and sentences (all text combined, in manuscript order).

Examples

result <- analyze_and_report(mtcars, dv = "mpg", iv = "cyl")
result$plot

# Repeated measures: name the participant column
set.seed(1)
rm_df <- data.frame(
  participant = rep(1:15, times = 3),
  condition   = rep(c("A", "B", "C"), each = 15)
)
rm_df$score <- rnorm(45, mean = c(A = 5, B = 6, C = 7)[rm_df$condition])
analyze_and_report(rm_df, dv = "score", iv = "condition",
                   design = "within", subject = "participant")

Description

Renders generateMoboPlot2() one iteration at a time and encodes the frames into a video, so a talk or a supplement can show an optimisation run building up rather than only its end state. Frame i is the same plot restricted to the data up to iteration i: the mean points, their bootstrapped intervals, the fitted line and its equation all move as the run proceeds.

Usage

animate_mobo2(
  data,
  x = "Iteration",
  y,
  filename,
  ...,
  fps = 5,
  end_pause = 2,
  width = 8,
  height = 5,
  dpi = 150,
  label_iterations = TRUE
)

Arguments

Details

The plot is built once from the complete data, and a frame only ever hides rows. That is what keeps the axes, the sampling/optimisation guides and the legend still while the data grows – and it is also the only way the early frames can be drawn at all, since generateMoboPlot2() requires both phases to be present and the first iterations are all sampling.

Value

Invisibly returns filename.

See Also

generateMoboPlot2() for the static plot.

Examples

# Kept deliberately tiny: every frame is a full ggplot render, so the cost
# of the example is set by the number of iterations, not by the frame size.
if (requireNamespace("av", quietly = TRUE)) {
  set.seed(42)
  df <- data.frame(
    Iteration   = rep(1:3, each = 4),
    ConditionID = rep(rep(c("A", "B"), each = 2), 3),
    Phase       = rep(c("Sampling", "Optimization"), times = c(2 * 4, 4))
  )
  df$score <- 0.3 + 0.05 * df$Iteration + stats::rnorm(nrow(df), sd = 0.05)

  animate_mobo2(
    df,
    y = "score", ytext = "Score",
    filename = file.path(tempdir(), "mobo.mp4"),
    width = 4, height = 2.5, dpi = 96, end_pause = 0
  )
}

Description

Runs check_normality_by_group() (and optionally a test for homogeneity of variances) and turns the outcome into a ready-made methods-section sentence that says what was tested and reports the statistic the decision rests on. This is the justification reviewers expect next to the choice of a parametric or non-parametric test.

Usage

assumption_methods_text(
  data,
  x,
  y,
  include_homogeneity = FALSE,
  subject = NULL
)

Arguments

Details

The sentence follows the check exactly:

• Between subjects: one Shapiro–Wilk test per group, Holm-corrected across the groups; the adjusted p-values are labelled $p_{Holm}$. A rejection names the group(s) with their $W$ and p; otherwise the group with the smallest p is reported.

• Within subjects (subject given): one test on the paired differences (two conditions) or on the residuals of the additive participant + condition model (more conditions). Participants excluded for lacking a condition are counted in the text.

• A group that could not be tested (fewer than three values, or all values identical – e.g. everyone ticked the top of a rating scale) is named and stated to have been treated as non-normal, which sends the analysis to the non-parametric branch.

p-values are printed so that rounding never moves them across .05, .01 or .10, and all statistics are plain LaTeX math (no custom macros needed).

Value

Invisibly returns the sentence(s) as a single string; the text is also emitted via message().

Examples

set.seed(1)
d <- data.frame(g = rep(c("A", "B"), each = 20), v = rnorm(40))
assumption_methods_text(d, x = "g", y = "v")

# within subjects: the paired differences are tested
d$id <- rep(1:20, times = 2)
assumption_methods_text(d, x = "g", y = "v", subject = "id")

Description

Runs the Brown-Forsythe test: Levene's test computed on the absolute deviations from each group's median rather than its mean (rstatix::levene_test()'s default, center = median). The median-centred version keeps close to its nominal error rate when the data are skewed or heavy-tailed, as rating-scale data often are, where Levene's mean-centred original rejects too often (Brown & Forsythe, 1974). Report it under that name; the "method" attribute carries it, e.g. for assumption_methods_text().

Usage

check_homogeneity_by_group(data, x, y)

Arguments

Details

The grouping column is treated as a factor whatever its type, so numeric condition codes (1, 2, 3) define groups rather than a covariate, and non-syntactic column names such as "Mental Demand" work.

Value

TRUE if the test is non-significant (p >= .05), FALSE otherwise. Attributes: "test", the test result (columns df1, df2, statistic, p); "method", the name of the test that was run ("Brown-Forsythe test (median-centred Levene's test)").

References

Brown, M. B., & Forsythe, A. B. (1974). Robust tests for the equality of variances. Journal of the American Statistical Association, 69(346), 364–367. doi:10.1080/01621459.1974.10482955

Description

Decides between a parametric and a non-parametric analysis from Shapiro-Wilk tests, testing the quantity the parametric test actually assumes to be normal.

Usage

check_normality_by_group(data, x, y, subject = NULL, p_adjust = "holm")

Arguments

Details

• Between subjects (subject = NULL): one test per group, with the p-values corrected for the number of groups (p_adjust, Holm by default). Without the correction, six perfectly normal groups would send about one analysis in four to the non-parametric branch by chance alone.

• Within subjects (subject given): a paired t-test assumes the differences are normal, and a repeated-measures ANOVA the residuals after removing participant and condition effects – not the raw scores per condition, which also carry the between-participant spread. With two conditions the per-participant differences are tested; with more, the residuals of the additive model y ~ x + subject. Participants lacking any condition are left out (they cannot enter a paired analysis), and more than one row per participant and condition is an error.

A group that cannot be tested – fewer than three values, or no variance at all, as in a rating scale where everyone ticked the top box – counts as not normal, so the decision errs towards the non-parametric test rather than letting an untestable group pass silently.

Testing assumptions with significance tests has well-known limits (low power in small samples, trivial deviations flagged in large ones); report the check, and treat it as one input to the choice rather than the whole of it.

Value

TRUE if no test rejects normality and every group could be tested, FALSE otherwise. Attributes:

tests

data frame with columns group (the condition, or "differences" / "residuals" within subjects), n, W, p_value, p_adjusted and testable, e.g. for assumption_methods_text().

method

"groupwise", "differences" or "residuals".

p_adjust

the correction applied ("none" for a single test).

untestable

the groups that could not be tested.

dropped_subjects

within subjects: the participants left out for lacking a condition.

For a group with more than 5000 values, Shapiro-Wilk is computed on a random sample of 5000 (a warning is emitted), so the result is only reproducible with a seed set beforehand.

Examples

set.seed(1)
d <- data.frame(id = rep(1:20, 2), cond = rep(c("A", "B"), each = 20))
d$score <- rnorm(40, mean = ifelse(d$cond == "A", 5, 6))
check_normality_by_group(d, "cond", "score")                  # between
check_normality_by_group(d, "cond", "score", subject = "id")  # within

Description

Prints the mapping score_questionnaire() would use – which column supplies which item, which subscale it loads on, whether it is reverse-coded – along with the observed range of each column and the instrument's scoring notes. Run it once per study: it is the cheapest available check against the failure mode in which a shifted survey export produces perfectly plausible, wrong scores.

Usage

check_questionnaire(
  data,
  instrument,
  items = NULL,
  prefix = NULL,
  scale = NULL,
  reverse_items = NULL,
  unreverse_items = NULL
)

Arguments

Details

Problems that score_questionnaire() would stop on (a column that is not numeric, responses outside the assumed range) are reported here as warnings instead, together with the warnings scoring would raise (a coding that looks shifted, an item named in reverse_items that the key already reverses), so the check can explain them rather than abort.

Value

Invisibly, a data frame with one row per item: the item number and code, the column it maps to, its subscale, whether it is reverse-coded, and the observed minimum, maximum and number of missing values of that column.

Caution – verify the mapping against your own survey

Item numbers, item order and item polarity are properties of the sheet a study actually administered, not of the instrument in the abstract. Survey tools renumber items, translations reorder them, short forms drop them from the middle, and semantic differentials get printed with the poles the other way round.

This package applies each instrument's published scoring key, which is the right default and is still only a default. If the sheet your participants saw differed, the scores will be wrong – and wrong quietly, because a mismatched mapping raises no error and produces entirely plausible numbers.

So: run check_questionnaire() once per instrument per study and read the mapping it prints, and double-check any figure before it goes into a paper. The mapping used is also attached to the result as the "mapping" attribute, and summarised in a console note the first time each distinct mapping is scored in a session (silence it with options(colleyRstats.quiet_questionnaires = TRUE)).

A named items argument (items = c(mental = "tlx_md", ...)) removes the positional assumption altogether and is the safer choice for an export you did not lay out yourself. A prefix maps columns by what their names say – an item code or label, or one item number per column running 1 to the number of items – never by sort order, and stops when the names do not identify the items. Note that the attrakdiff key is blocked by dimension with the negative pole first; data stored as answered on the official AttrakDiff sheet belong to attrakdiff_official.

See Also

score_questionnaire(), questionnaire_items()

Examples

set.seed(1)
d <- as.data.frame(matrix(sample(1:5, 10 * 4, TRUE), nrow = 4))
names(d) <- paste0("sus_", 1:10)
check_questionnaire(d, "sus", prefix = "sus_")

Description

Checks the normality and variance assumptions of a factorial ANOVA and recommends the parametric or the non-parametric analysis. Every factor is treated as categorical – numeric condition codes (1, 2, 3) included, which stats::lm() would otherwise fit as a single linear covariate – and the message names the tests that were run and their results.

Usage

checkAssumptionsForAnova(data, y, factors, subject = NULL)

check_assumptions_anova(data, y, factors, subject = NULL)

Arguments

Details

• Between subjects (subject = NULL, the default): a Shapiro-Wilk test on the residuals of the linear model y ~ A * B * ...; a Shapiro-Wilk test within every cell of the design, Holm-corrected across the cells (as in check_normality_by_group()); and the Brown-Forsythe test (median-centred Levene's test) across the cells. This assumes one row per participant: a within-subject factor analysed this way is tested on residuals that still contain each participant's overall level, so pass subject for repeated measures.

• Within subjects or mixed (subject given): a Shapiro-Wilk test on the residuals of y ~ A * B * ... + subject, with the participant as a fixed factor – the residuals after removing each participant's overall level, which is what a repeated-measures ANOVA assumes to be normal. The raw scores per cell are not tested, as they also carry the between-participant spread. Variance homogeneity is checked only for between-subject factors (those constant within a participant), with the Brown-Forsythe test across their groups at each combination of the within-subject factors (Holm-corrected). For within-subject factors the corresponding assumption is sphericity, which rstatix::anova_test() tests (Mauchly) and corrects (Greenhouse-Geisser) itself. More than one row per participant and cell is an error: aggregate repeated trials first.

A residual set or cell that cannot be tested – fewer than three values, or no variance – counts as not normal, so the advice errs towards the non-parametric analysis. Rows with a missing value in y, factors or subject are left out.

Value

The guidance text (also emitted as a message), invisibly, with attributes:

parametric

TRUE if every check passed.

design

"between", "within" or "mixed".

method

named character vector naming the tests run (residuals, groupwise, homogeneity).

residuals

the Shapiro-Wilk result on the model residuals (columns as in check_normality_by_group()'s tests).

groupwise

between subjects: the per-cell Shapiro-Wilk results with Holm-adjusted p-values; NULL otherwise.

homogeneity

the Brown-Forsythe result(s) (df1, df2, statistic, p, p_adjusted), or NULL when no between-subject factor exists.

Naming

check_assumptions_anova() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

checkAssumptionsForAnova() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

set.seed(123)

main_df <- data.frame(
  tlx_mental      = rnorm(40),
  Video           = factor(rep(c("A", "B"), each = 20)),
  DriverPosition  = factor(rep(c("Left", "Right"), times = 20))
)

checkAssumptionsForAnova(
  data    = main_df,
  y       = "tlx_mental",
  factors = c("Video", "DriverPosition")
)

# The same two factors measured within each of 10 participants
within_df <- expand.grid(
  id             = 1:10,
  Video          = c("A", "B"),
  DriverPosition = c("Left", "Right")
)
within_df$tlx_mental <- rnorm(10)[within_df$id] + rnorm(40)
checkAssumptionsForAnova(
  data    = within_df,
  y       = "tlx_mental",
  factors = c("Video", "DriverPosition"),
  subject = "id"
)

Description

Prints a ready-made methods phrase plus the BibTeX entries for the R packages behind the requested analysis methods, so a manuscript's methods section and bibliography can be filled in one step.

Usage

cite_methods(methods = c("ggstatsplot", "effectsize"), bibtex = TRUE)

Arguments

Value

Invisibly returns the generated lines as a character vector; the text is also emitted via message(). Methods whose package is not installed are skipped with a message. Every BibTeX entry gets a citation key – the package name, with -2, -3, ... for a package's further entries – because R's citation entries carry none, and BibTeX keeps only the first of several keyless entries ("Repeated entry"). The methods phrase then cites those keys (\cite{ARTool,ARTool-2}).

Examples

cite_methods("ggstatsplot", bibtex = FALSE)

Description

Decides how a dependent variable should be modelled by inspecting its type and distribution of values. The measurement scale is the first branch of a principled model choice: it dictates the family (Gaussian, binomial, Poisson, cumulative-link) before any distributional assumption is checked.

Usage

classify_outcome(y, ordinal_max_levels = 7L)

Arguments

Details

The rules are deliberately simple and transparent:

• ordered factor $\rightarrow$ "ordinal";

• logical, a factor with two observed levels, a character vector with two distinct values, or a numeric coded 0/1 $\rightarrow$ "binary";

• unordered factor/character with more than two levels $\rightarrow$ "nominal";

• a numeric with at most two distinct values that are not 0/1 – a 1–7 item on which only 6 and 7 were ticked, or a yes/no item coded 1/2 – is not taken for binary: it is classified as "ordinal" (whole numbers) or "continuous", with a warning;

• integer-valued numeric with at most ordinal_max_levels distinct values (a Likert-type item) $\rightarrow$ "ordinal";

• non-negative whole numbers that are all multiples of 5 within $[0, 100]$ (the shape of a raw NASA-TLX subscale or a 0–100 slider in steps of 5) $\rightarrow$ "continuous", with a message;

• any other non-negative whole numbers $\rightarrow$ "count". This is always announced with a message that says how to override it, and the message becomes a warning when the values lie in a closed range without zeros ($[1, 100]$) – the typical shape of a summed questionnaire score or a 1–10 rating, which a Poisson model would misdescribe;

• any other numeric $\rightarrow$ "continuous".

The heuristics can never be perfect (a 1–7 Likert item and a small count are genuinely ambiguous); pass an explicit outcome_type to recommend_test() or fit_recommended() when you want to override them.

Value

A single string, one of "continuous", "ordinal", "binary", "count", or "nominal".

Examples

classify_outcome(rnorm(50)) # "continuous"
classify_outcome(factor(sample(1:5, 50, TRUE), ordered = TRUE)) # "ordinal"
classify_outcome(sample(0:1, 50, TRUE)) # "binary"
classify_outcome(rpois(50, 3)) # "count", with a message

Description

see::theme_lucid() with the sizing and legend conventions used across colleyRstats figures. Every text element is expressed as a multiple of base_size rather than in absolute points, which is what lets the same theme serve a 3.33 in journal column and a full-width figure: only base_size changes.

Usage

colley_theme(base_size = 17, base_family = "")

Arguments

Details

Sizing text in absolute points is the usual reason a figure comes out unreadable. Point sizes do not shrink when the canvas does, so a theme tuned on a large canvas puts 20 pt axis titles on a 3.33 in figure, where they overrun the panel and the axis labels collide. Choose base_size from the size the figure is finally PLACED at – roughly the body-text size of the document, or a point or two below it. save_paper_figure() does this for you.

The multipliers are, relative to base_size: axis titles 1.15, axis text 1.0, plot title 1.65, subtitle 1.0, legend text 0.9, caption 0.8, strip text 1.3.

Value

A ggplot2 theme object.

Examples

if (requireNamespace("see", quietly = TRUE)) {
  ggplot2::ggplot(mtcars, ggplot2::aes(mpg, wt)) +
    ggplot2::geom_point() +
    colley_theme(base_size = 8)   # sized for a single journal column
}

Description

Sets the ggplot2 theme, and on request the conflicted preferences, to match the standards used in the colleyRstats workflow.

Usage

colleyRstats_setup(
  set_options = FALSE,
  set_theme = TRUE,
  base_size = 17,
  set_conflicts = FALSE,
  print_citation = TRUE,
  verbose = TRUE
)

Arguments

Details

set_conflicts changes what library() does, so call it last. conflicted::conflict_prefer() does not merely record a preference: it activates conflicted, which attaches a .conflicts environment carrying its own library() and require() shims. Every later attach in the session goes through them, and they forward quietly and verbose into base::library() as unevaluated symbols. A package whose .onAttach inspects the calling library() frame – meta-packages that attach their own constituents do this to decide whether to print a banner – then evaluates those symbols in a frame that does not bind them, and the attach fails outright:

colleyRstats_setup(set_conflicts = TRUE)
library(easystats)
#> Error: .onAttach failed in attachNamespace() for 'easystats':
#>   object 'quietly' not found

The two mechanisms do not compose, and colleyRstats_setup() cannot know whether more library() calls are coming. The preferences are therefore opt-in, and the contract when you opt in is that colleyRstats_setup() runs after everything else is attached:

library(colleyRstats)
library(easystats)
library(dplyr)
colleyRstats_setup(set_conflicts = TRUE)   # last

That is also the ordering in which the preferences are worth most: conflicted resolves only those names that are ambiguous among the packages attached at the time, so a call made before the rest of the script's library() calls has less to work with.

Value

Invisibly returns NULL.

Examples

# Runs everywhere, no extra packages, no session side effects
colleyRstats::colleyRstats_setup(
  set_options = FALSE,
  set_theme = FALSE,
  set_conflicts = FALSE,
  print_citation = FALSE,
  verbose = FALSE
)


# Full setup (requires suggested packages; changes session defaults)
if (requireNamespace("ggplot2", quietly = TRUE) &&
    requireNamespace("see", quietly = TRUE)) {
  local({
    old_theme <- ggplot2::theme_get()
    on.exit(ggplot2::theme_set(old_theme), add = TRUE)

    colleyRstats::colleyRstats_setup(
      set_options = FALSE,
      set_conflicts = FALSE,   # the default; see Details
      print_citation = FALSE,
      verbose = TRUE
    )

    ggplot2::ggplot(mtcars, ggplot2::aes(mpg, wt)) +
      ggplot2::geom_point()
  })
}

Description

Debug contrast errors in ANOVA-like models

Usage

debug_contr_error(dat, subset_vec = NULL)

Arguments

Value

A list with two elements:

nlevels

Integer vector giving the number of levels for each factor variable in dat.

levels

List of factor level labels for each factor variable in dat.

Examples

dat <- data.frame(
  group = factor(rep(letters[1:3], each = 3)),
  score = rnorm(9)
)

debug_contr_error(dat = dat)

Description

Adds an instrument to the registry so that score_questionnaire(), check_questionnaire() and score_reliability() handle it exactly like a built-in one. Use it for a lab-specific scale, a translated or shortened form, or a published instrument this package does not ship – and put the call in a project's setup script so every analysis in that project scores it the same way.

Usage

define_questionnaire(
  key,
  name,
  scale,
  subscale,
  code = NULL,
  label = NULL,
  reverse = integer(0),
  recode = c("none", "center", "zero_base"),
  total = NULL,
  total_name = "Total",
  reference = NA_character_,
  higher = NA_character_,
  notes = character(),
  integer_responses = TRUE
)

Arguments

Value

Invisibly, the instrument definition.

Caution – verify the mapping against your own survey

Item numbers, item order and item polarity are properties of the sheet a study actually administered, not of the instrument in the abstract. Survey tools renumber items, translations reorder them, short forms drop them from the middle, and semantic differentials get printed with the poles the other way round.

This package applies each instrument's published scoring key, which is the right default and is still only a default. If the sheet your participants saw differed, the scores will be wrong – and wrong quietly, because a mismatched mapping raises no error and produces entirely plausible numbers.

So: run check_questionnaire() once per instrument per study and read the mapping it prints, and double-check any figure before it goes into a paper. The mapping used is also attached to the result as the "mapping" attribute, and summarised in a console note the first time each distinct mapping is scored in a session (silence it with options(colleyRstats.quiet_questionnaires = TRUE)).

A named items argument (items = c(mental = "tlx_md", ...)) removes the positional assumption altogether and is the safer choice for an export you did not lay out yourself. A prefix maps columns by what their names say – an item code or label, or one item number per column running 1 to the number of items – never by sort order, and stops when the names do not identify the items. Note that the attrakdiff key is blocked by dimension with the negative pole first; data stored as answered on the official AttrakDiff sheet belong to attrakdiff_official.

See Also

list_questionnaires(), score_questionnaire()

Examples

define_questionnaire(
  key = "acceptance",
  name = "Van der Laan acceptance scale",
  reference = "Van der Laan, Heino & De Waard (1997), Transp. Res. C 5(1)",
  scale = c(-2, 2),
  subscale = c(
    "Usefulness", "Satisfying", "Usefulness", "Satisfying", "Usefulness",
    "Satisfying", "Usefulness", "Satisfying", "Usefulness"
  ),
  label = c(
    "useful - useless", "pleasant - unpleasant", "bad - good",
    "nice - annoying", "effective - superfluous", "irritating - likeable",
    "assisting - worthless", "undesirable - desirable", "raising alertness - sleep-inducing"
  ),
  reverse = c(1, 2, 4, 5, 7, 9),
  higher = "better"
)
list_questionnaires()[1, ]

Description

Emits \newcommand{\<name>}{<value>} so you can write \<name> in your prose and have it always reflect the latest analysis – re-run the R code and the number updates everywhere it is referenced, the gold standard for reproducible manuscripts. The name is sanitised to a letters-only LaTeX command name: other characters start a new camel-case word and digits are spelled out, so "tlx_1" and "tlx_2" become \tlxOne and \tlxTwo rather than colliding. A name that would redefine an existing LaTeX or colleyRstats command (e.g. "p", which is the package's own \p, or "time") is refused with an error that suggests an alternative.

Usage

define_result_macro(name, value, path = NULL)

Arguments

Value

Invisibly, a named character scalar: the \newcommand line, named by the generated command. Also emitted via message().

Examples

define_result_macro("tlx_mental_omnibus", "F(2, 57) = 4.50, p = .02")

Description

The report functions can emit variable and factor-level names as LaTeX commands (e.g. \Video) so their typography is controlled centrally. This writes the matching \newcommand definitions so those commands are never undefined – the classic "Undefined control sequence" that stops an Overleaf build. A name gets a macro only when it can safely become a new command: letters only, and not already a LaTeX command. time, L or small, for instance, are TeX/LaTeX commands already, so \newcommand{\time} would stop the build ("Command \time already defined") and redefining them would break LaTeX itself; names starting with end are refused by LaTeX too. Such names are skipped with a warning, and the reporters emit them as escaped plain text instead.

Usage

emit_name_macros(vars, path = NULL, labels = NULL)

Arguments

Value

Invisibly, the \newcommand lines as a character vector (one per distinct name); also emitted via message().

Examples

emit_name_macros(c("Video", "DriverPosition"))
emit_name_macros(c(tlxMental = "TLX Mental Demand"))

Description

Writes everything a manuscript needs into one directory you can drag into Overleaf and compile immediately: a main.tex that already \inputs the results, one .tex per result section, the figures, a references.bib, and – unless macros are expanded inline – colleyRstats.sty. This is the one-call end of the "R analysis to compiled PDF" pipeline.

Usage

emit_overleaf(
  x,
  dir,
  figures = TRUE,
  methods = c("ggstatsplot", "effectsize"),
  title = "Results",
  plain = NULL,
  columns = 1,
  overwrite = FALSE
)

Arguments

Details

Section keys become file names with only letters, digits and hyphens kept ("tlx_1" is written to sections/tlx-1.tex); keys that would share a file name – including names differing only in case, which Windows and macOS treat as one file – get a numbered suffix, so no section can overwrite another. The keys themselves are escaped with latex_escape() wherever they are typeset (headings, captions). Name macros in the text (\Video) get \providecommand stubs in names.tex, and a {} where a space follows them, because TeX would otherwise swallow that space ("\cyl on" typesets as "cylon").

Value

Invisibly, a list with the paths written (dir, main, results, sections, figures, bib, sty, names); bib, sty and names are NULL when the file was not needed.

Examples

out <- report_all(mtcars, dvs = c("mpg", "disp"), iv = "cyl")
emit_overleaf(out, dir = file.path(tempdir(), "paper"), overwrite = TRUE)

Description

The report functions normally emit compact custom macros (\F, \p, \m, \sd, \df, \chisq, \padj, \padjminor, \pminor, \rankbiserial, \effectsize) that require latex_preamble() definitions. This expands them into equivalent plain math (e.g. \F{2}{57}{4.50} becomes $F(2, 57) = 4.50$) so the text compiles in any document with no custom preamble – the "zero-setup Overleaf" path. It is applied automatically by the sink_to/emit_overleaf() writers when options(colleyRstats.macros = FALSE).

Usage

expand_latex_macros(x)

Arguments

Value

The text with the macros expanded to standard LaTeX math.

Examples

expand_latex_macros("A significant effect (\\F{2}{57}{4.50}, \\p{0.012}).")

Description

The rule save_paper_figure() uses to pick a type size from a figure width. Text in a figure should read at roughly the body-text size of the document the figure is placed in; since a figure is usually placed at 100\ means the type size has to follow the physical width. The rule is calibrated so that the two standard widths land on sensible values: 3.33 in (a two-column journal column) gives 7 pt and 7 in (full text width) gives 9 pt, with linear interpolation between and clamping outside.

Usage

figure_base_size(width, min_size = 6, max_size = 12)

Arguments

Value

A single number: the base font size in points.

Examples

figure_base_size(3.33)  # 7
figure_base_size(7)     # 9

Description

Runs recommend_test() to choose an analysis, then actually fits it: coerces the outcome and predictors into the classes that model family needs, builds the random-effect structure for a repeated-measures design, fits, computes post-hoc contrasts with the matching machinery, and produces the APA/LaTeX sentence via this package's reporter for that model family.

Usage

fit_recommended(
  data,
  outcome,
  predictors = NULL,
  cluster = NULL,
  design = c("auto", "between", "within"),
  outcome_type = c("auto", "continuous", "ordinal", "binary", "count", "nominal"),
  ordinal_max_levels = 7L,
  interaction = TRUE,
  contrasts = TRUE,
  adjust = "holm",
  sink_to = NULL,
  verbose = TRUE,
  factors = NULL,
  random_slopes = NULL
)

Arguments

Details

The point is that the test you justify in the methods section and the model you actually ran are produced by the same call, from the same data, and therefore cannot drift apart – the failure mode of a pipeline where recommend_test() prints advice that someone then re-types by hand.

Value

An object of class "colley_fit": a list with recommendation (the recommend_test() object, updated to record the function, call and random effects actually used), formula, model (the fitted object), data (the final analyzed data, with model coercions and any incomplete-participant exclusions applied), engine, model_function (the function actually called, e.g. "lmerTest::lmer"), random (the random-effects term kept), singular and converged (flags), anova (the omnibus table: term, df1, df2, statistic, stat_name, p), emmeans (a named list of the estimated marginal means grids) and contrasts (a data frame with one row per comparison: term, by, contrast, estimate, scale, SE, df, confidence limits, statistic, p.value, adjust; or NULL), methods (plain text), methods_tex (LaTeX-escaped), text (result sentences, LaTeX), and sentences (methods_tex and text, in manuscript order – what sink_to writes). A print method summarises it.

What it can fit

Cumulative link models with and without random effects (ordinal), generalized and linear mixed models (lme4, lmerTest, glmmTMB for negative-binomial counts), Poisson, negative-binomial (MASS) and logistic GLMs, Type III factorial ANOVA (afex, or car), heteroscedasticity-robust (HC3) factorial ANOVA (car), the aligned rank transform (ARTool), rank-based repeated measures (nparLD), multinomial regression (nnet) and its mixed counterpart (mclogit), and the classical one-way ANOVA / Welch / Kruskal-Wallis / Wilcoxon tests. Model packages live in Suggests: a branch you use needs its package installed, and says which one if it is not.

Random effects

A repeated-measures model starts from the maximal random-effects structure the design supports (Barr et al., 2013): by-cluster random slopes for every within-cluster factor whose levels are replicated within a cluster, numeric covariates with enough repeated observations to estimate a slope, and the interaction slope when there are several trials per cluster and cell. When that fit is singular or does not converge, the structure is simplified step by step down to random intercepts; what was dropped, and why, is announced and written into $methods. A singular final fit is flagged in $singular.

Post-hoc contrasts

Each categorical predictor gets its own family of pairwise comparisons of its marginal means (averaged over the other factors), adjusted with adjust within that family. Simple effects – one factor compared within each level of the others – are added only for interactions that are in the model and significant in the Type III omnibus test.

References

Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68(3), 255–278.

Matuschek, H., Kliegl, R., Vasishth, S., Baayen, H., & Bates, D. (2017). Balancing Type I error and power in linear mixed models. Journal of Memory and Language, 94, 305–315.

See Also

recommend_test() for the choice alone, plot_fit() for a figure using this model's tests, analyze_and_report() for the figure-plus-sentence path through ggstatsplot.

Examples

set.seed(1)
d <- data.frame(
  id = factor(rep(1:20, each = 3)),
  cond = factor(rep(c("A", "B", "C"), times = 20)),
  score = rnorm(60)
)
if (requireNamespace("lme4", quietly = TRUE)) {
  fit <- fit_recommended(d, outcome = "score", predictors = "cond", cluster = "id")
  fit
}

Description

Function to define a plot, either showing the main or interaction effect in bold.

Usage

generateEffectPlot(
  data,
  x,
  y,
  fillColourGroup,
  ytext = "testylab",
  xtext = "testxlab",
  legendPos = c(0.1, 0.23),
  legendHeading = NULL,
  shownEffect = "main",
  effectLegend = FALSE,
  effectDescription = NULL,
  xLabelsOverwrite = NULL,
  useLatexMarkup = FALSE,
  numberColors = lifecycle::deprecated(),
  subject = NULL
)

plot_effect(
  data,
  x,
  y,
  fillColourGroup,
  ytext = "testylab",
  xtext = "testxlab",
  legendPos = c(0.1, 0.23),
  legendHeading = NULL,
  shownEffect = "main",
  effectLegend = FALSE,
  effectDescription = NULL,
  xLabelsOverwrite = NULL,
  useLatexMarkup = FALSE,
  numberColors = lifecycle::deprecated(),
  subject = NULL
)

Arguments

Details

Points and lines. The small points and the dashed (main) or bold (interaction) per-group lines are the cell means. The bold main-effect line and its large points are the unweighted marginal means of x: for each level of x, the mean of the cell means across fillColourGroup. That is the quantity a main effect in a factorial ANOVA is about; the mean of the raw observations would weight each cell by its size and, in an unbalanced design, follow whichever group happens to be larger.

Error bars are 95\

• subject = NULL (default): nonparametric bootstrap intervals (ggplot2::mean_cl_boot(), requires 'Hmisc'). These treat every observation as independent, i.e. they are between-subject intervals. That is right for between-subject factors, but for a factor that varies within participants they include the between-participant spread a within-subject comparison does not depend on, and look misleadingly wide.

• subject given: within-subject intervals after Cousineau (2005), with the Morey (2008) correction. Each participant's scores are centred on their own mean, the variance of the normalised scores is inflated by $M/(M-1)$ ($M$ = number of within-subject conditions), and a factor that does not vary within participants is respected by normalising within its groups. Participants without a value in every within-subject condition are dropped (with a message), and more than one row per participant and cell is an error.

Either way, the intervals describe single means, not differences between them; overlapping bars do not by themselves imply a non-significant difference.

Value

a plot

Naming

plot_effect() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

generateEffectPlot() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

References

Cousineau, D. (2005). Confidence intervals in within-subject designs: A simpler solution to Loftus and Masson's method. Tutorials in Quantitative Methods for Psychology, 1(1), 42–45.

Morey, R. D. (2008). Confidence intervals from normalized data: A correction to Cousineau (2005). Tutorials in Quantitative Methods for Psychology, 4(2), 61–64.

Examples

set.seed(123)
main_df <- data.frame(
  strategy    = factor(rep(c("A", "B"), each = 20)),
  Emotion     = factor(rep(c("Happy", "Sad"), times = 20)),
  trust_mean  = rnorm(40, mean = 5, sd = 1)
)

generateEffectPlot(
  data = main_df,
  x = "strategy",
  y = "trust_mean",
  fillColourGroup = "Emotion",
  ytext = "Trust",
  xtext = "Strategy",
  legendPos = c(0.1, 0.23)
)

Description

This function generates a multi-objective optimization plot using ggplot2. The plot visualizes the relationship between the x and y variables, grouping and coloring by a fill variable, with the option to customize legend position, labels, and annotation of sampling and optimization phases.

Usage

generateMoboPlot(
  data,
  x,
  y,
  fillColourGroup = "ConditionID",
  ytext,
  legendPos = c(0.65, 0.85),
  numberSamplingSteps = 5,
  labelPosFormulaY = "top",
  verticalLinePosY = 0.75
)

plot_mobo(
  data,
  x,
  y,
  fillColourGroup = "ConditionID",
  ytext,
  legendPos = c(0.65, 0.85),
  numberSamplingSteps = 5,
  labelPosFormulaY = "top",
  verticalLinePosY = 0.75
)

Arguments

Value

A ggplot object representing the multi-objective optimization plot, ready to be rendered.

Naming

plot_mobo() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

generateMoboPlot() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

library(ggplot2)
library(ggpmisc)

# Example with numeric x-axis
df <- data.frame(
  x = 1:20,
  y = rnorm(20),
  ConditionID = rep(c("A", "B"), 10)
)
generateMoboPlot(df, x = "x", y = "y")


# Example with factor x-axis
df <- data.frame(
  x = factor(rep(1:5, each = 4)),
  y = rnorm(20),
  ConditionID = rep(c("A", "B"), 10)
)
generateMoboPlot(df, x = "x", y = "y", numberSamplingSteps = 3)

Description

This function generates a multi-objective optimization plot using ggplot2. The plot visualizes the relationship between the x and y variables, grouping and coloring by a fill variable, with the option to customize legend position, labels, and annotation of sampling and optimization phases. Appropriate if you use https://github.com/Pascal-Jansen/Bayesian-Optimization-for-Unity in version 1.1.0 or higher.

Usage

generateMoboPlot2(
  data,
  x = "Iteration",
  y,
  phaseCol = "Phase",
  fillColourGroup = "ConditionID",
  ytext,
  legendPos = c(0.65, 0.85),
  labelPosFormulaY = "top",
  labelPosFormulaX = "left",
  horizontalLinePosY = 0.75,
  horizontalLineDistToText = 0.3,
  fillLabels = NULL,
  annotationTextSize = 5
)

plot_mobo2(
  data,
  x = "Iteration",
  y,
  phaseCol = "Phase",
  fillColourGroup = "ConditionID",
  ytext,
  legendPos = c(0.65, 0.85),
  labelPosFormulaY = "top",
  labelPosFormulaX = "left",
  horizontalLinePosY = 0.75,
  horizontalLineDistToText = 0.3,
  fillLabels = NULL,
  annotationTextSize = 5
)

Arguments

Value

A ggplot object representing the multi-objective optimization plot, ready to be rendered.

Naming

plot_mobo2() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

generateMoboPlot2() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

library(ggplot2)
library(ggpmisc)

# Example with numeric x-axis
df <- data.frame(
  x = 1:20,
  y = rnorm(20),
  ConditionID = rep(c("A", "B"), 10),
  Phase = rep(c("Sampling", "Optimization"), 10)
)
generateMoboPlot2(data = df, x = "x", y = "y")

Description

For independent groups: one row per participant. If the group-wise normality check (check_normality_by_group()) passes, ggstatsplot runs Welch's t-test (two groups) or Welch's one-way ANOVA with Games-Howell post-hoc tests and labels the means; none of these assume equal variances, which is why no Levene test is run. Otherwise it runs a Wilcoxon rank-sum (Mann-Whitney) or Kruskal-Wallis test with Dunn post-hoc tests and labels the medians. Post-hoc p-values are Holm-adjusted. Rows without a group or an outcome are dropped, and the grouping column is treated as a factor, so numeric condition codes are categories.

Usage

ggbetweenstatsWithPriorNormalityCheck(
  data,
  x,
  y,
  ylab,
  xlabels = NULL,
  showPairwiseComp = TRUE,
  plotType = "boxviolin"
)

plot_between_stats(
  data,
  x,
  y,
  ylab,
  xlabels = NULL,
  showPairwiseComp = TRUE,
  plotType = "boxviolin"
)

Arguments

Value

A ggplot object produced by ggstatsplot::ggbetweenstats, which can be printed or further modified with +.

Naming

plot_between_stats() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

ggbetweenstatsWithPriorNormalityCheck() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

set.seed(123)

# Toy between-subject data: each participant sees one condition
main_df <- data.frame(
  CondID      = factor(rep(c("A", "B", "C"), each = 20)),
  tlx_mental  = rnorm(60, mean = 50, sd = 10)
)

# Custom x-axis labels
labels_xlab <- c("Condition A", "Condition B", "Condition C")


ggbetweenstatsWithPriorNormalityCheck(
  data = main_df,
  x = "CondID",
  y = "tlx_mental", ylab = "Mental Demand",
  xlabels = labels_xlab,
  showPairwiseComp = TRUE
)

Description

As ggbetweenstatsWithPriorNormalityCheck(), but the significant comparisons are drawn as asterisk brackets (*** p < .001, ** p < .01,

• p < .05) instead of p-values. The brackets come from the same test as the figure: with two groups, the omnibus test in the subtitle (ggstatsplot runs no post-hoc test then, and a second test of the same pair could disagree with it); with more, the Holm-adjusted Games-Howell (parametric) or Dunn (non-parametric) tests ggstatsplot itself would compute.

Usage

ggbetweenstatsWithPriorNormalityCheckAsterisk(
  data,
  x,
  y,
  ylab,
  xlabels,
  plotType = "boxviolin"
)

plot_between_stats_asterisk(data, x, y, ylab, xlabels, plotType = "boxviolin")

Arguments

Value

A ggplot object produced by ggstatsplot::ggbetweenstats with additional significance annotations, which can be printed or modified.

Naming

plot_between_stats_asterisk() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

ggbetweenstatsWithPriorNormalityCheckAsterisk() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

set.seed(123)

# Toy between-subject data: each participant sees one condition
main_df <- data.frame(
  CondID      = factor(rep(c("A", "B", "C"), each = 20)),
  tlx_mental  = rnorm(60, mean = 50, sd = 10)
)

# Custom x-axis labels
labels_xlab <- c("Condition A", "Condition B", "Condition C")


ggbetweenstatsWithPriorNormalityCheckAsterisk(
  data = main_df,
  x = "CondID", y = "tlx_mental", ylab = "Mental Demand", xlabels = labels_xlab
)

Description

Observations are paired by the participant ID in subject, never by row order. Before anything is computed, rows without a condition, outcome or ID are dropped, participants lacking any condition are left out (with a message saying how many and which), and more than one row per participant and condition is an error – aggregate repeated trials first. The normality check (check_normality_by_group() with subject, i.e. on the paired differences or on the residuals of the repeated-measures model), the figure and the pairwise tests all use this same data. The condition column is treated as a factor, so numeric condition codes are categories.

Usage

ggwithinstatsWithPriorNormalityCheck(
  data,
  x,
  y,
  ylab,
  xlabels = NULL,
  showPairwiseComp = TRUE,
  plotType = "boxviolin",
  subject
)

plot_within_stats(
  data,
  x,
  y,
  ylab,
  xlabels = NULL,
  showPairwiseComp = TRUE,
  plotType = "boxviolin",
  subject
)

Arguments

Details

If the check passes, ggstatsplot runs a paired t-test (two conditions) or a repeated-measures ANOVA with paired Student's t post-hoc tests, and labels the means; otherwise a Wilcoxon signed-rank or Friedman test with Durbin-Conover post-hoc tests, and labels the medians. Post-hoc p-values are Holm-adjusted.

Value

A ggplot object produced by ggstatsplot::ggwithinstats with additional significance annotations, which can be printed or modified.

Naming

plot_within_stats() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

ggwithinstatsWithPriorNormalityCheck() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

set.seed(123)

# Toy within-subject data: every participant sees every condition
main_df <- data.frame(
  Participant = factor(rep(1:20, each = 3)),
  CondID      = factor(rep(c("A", "B", "C"), times = 20)),
  tlx_mental  = rnorm(60, mean = 50, sd = 10)
)

# Custom x-axis labels
labels_xlab <- c("Condition A", "Condition B", "Condition C")


ggwithinstatsWithPriorNormalityCheck(
  data = main_df,
  x = "CondID", y = "tlx_mental",
  ylab = "Mental Demand",
  xlabels = labels_xlab,
  showPairwiseComp = TRUE,
  subject = "Participant"
)

Description

As ggwithinstatsWithPriorNormalityCheck() – including pairing by subject and dropping participants who lack a condition – but the significant comparisons are drawn as asterisk brackets (*** p < .001, ** p < .01, * p < .05) instead of p-values. The brackets come from the same test as the figure: with two conditions, the omnibus paired test in the subtitle; with more, the Holm-adjusted paired Student's t (parametric) or Durbin-Conover (non-parametric) tests ggstatsplot itself would compute, paired by participant.

Usage

ggwithinstatsWithPriorNormalityCheckAsterisk(
  data,
  x,
  y,
  ylab,
  xlabels,
  plotType = "boxviolin",
  subject
)

plot_within_stats_asterisk(
  data,
  x,
  y,
  ylab,
  xlabels,
  plotType = "boxviolin",
  subject
)

Arguments

Value

A ggplot object produced by ggstatsplot::ggwithinstats with additional significance annotations, which can be printed or modified.

Naming

plot_within_stats_asterisk() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

ggwithinstatsWithPriorNormalityCheckAsterisk() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

set.seed(123)

# Toy within-subject data: every participant sees every condition
main_df <- data.frame(
  Participant = factor(rep(1:20, each = 3)),
  CondID      = factor(rep(c("A", "B", "C"), times = 20)),
  tlx_mental  = rnorm(60, mean = 50, sd = 10)
)

# Custom x-axis labels
labels_xlab <- c("Condition A", "Condition B", "Condition C")


ggwithinstatsWithPriorNormalityCheckAsterisk(
  data = main_df,
  x = "CondID", y = "tlx_mental",
  ylab = "Mental Demand", xlabels = labels_xlab,
  subject = "Participant"
)

Description

Makes an arbitrary string safe to drop into a LaTeX document by escaping the characters that would otherwise be interpreted as markup (\ { } $ & # _ % ~ ^ < >). Use it on variable names, factor-level labels, captions – anything user-supplied that reaches the .tex. This is what prevents a dependent variable called tlx_mental from producing an un-compilable tlx_mental (a subscript error) in Overleaf.

Usage

latex_escape(x)

Arguments

Value

A character vector with LaTeX specials escaped; NA is preserved.

Examples

latex_escape("tlx_mental")
latex_escape("cost (%) & margin")

Description

All report functions emit LaTeX text that relies on a small set of custom commands. This helper prints the complete set, ready to paste into a manuscript preamble, or writes it to a file that can be included with \input{} (or renamed to .sty and loaded via \usepackage).

Usage

latex_preamble(path = NULL)

Arguments

Value

Invisibly returns the macro definitions as a character vector; the text is also emitted via message().

Examples

latex_preamble()

Description

This function transforms the text output from report::report() by performing several substitutions to prepare the text for LaTeX typesetting. Every LaTeX special character (\ { } $ & # _ % ^) is escaped – report() keeps variable and level names such as tlx_mental verbatim – ~, < and > become $\sim$, $<$ and $>$, and R2 and Rhat become $R^2$ and $\hat{R}$. Additionally, it provides options to:

• Omit bullet items marked as "non-significant" (when only_sig = TRUE).

• Remove a concluding note about standardized parameters (when remove_std = TRUE).

• Wrap bullet items in a LaTeX itemize environment or leave them as plain text (controlled by itemize).

Usage

latexify_report(
  x,
  print_result = TRUE,
  only_sig = FALSE,
  remove_std = FALSE,
  itemize = TRUE
)

Arguments

Value

A single string with the LaTeX-friendly formatted report text.

Examples

if (requireNamespace("report", quietly = TRUE)) {
  # Simple linear model on the iris dataset
  model <- stats::lm(
    Sepal.Length ~ Sepal.Width + Petal.Length,
    data = datasets::iris
  )

  # Format the report output, showing only significant items, removing the
  # standard note, and wrapping bullet items in an itemize environment.
  report_text <- try(report::report(model), silent = TRUE)
  if (!inherits(report_text, "try-error")) {
    latexify_report(
      report_text,
      only_sig = TRUE,
      remove_std = TRUE,
      itemize = TRUE
    )
  }
}

Description

List the questionnaires this package can score

Usage

list_questionnaires()

Value

A data frame with one row per instrument: its key (the value to pass as instrument), full name, number of items, number of subscales, the response range it assumes, the direction a high score means, and its reference.

See Also

questionnaire_items(), score_questionnaire(), define_questionnaire()

Examples

list_questionnaires()

Description

Build a median/size label for plot annotations

Usage

n_fun(x)

Arguments

Value

A data frame with the median and label.

Description

Replace NA values with zero

Usage

na.zero(x)

Arguments

Value

A vector with NAs replaced by zeros.

Examples

na.zero(c(NA, 1, NA, 2))

Description

This function normalizes the values in a vector to the range [new_min, new_max] based on their original range [old_min, old_max].

Usage

normalize(x_vector, old_min, old_max, new_min, new_max)

Arguments

Value

A numeric vector with the normalized values.

Examples

normalize(c(1, 2, 3, 4, 5), 1, 5, 0, 1)

Description

Stops execution if x is NULL, empty, contains only NAs, or is a data frame without rows.

Usage

not_empty(x, msg = NULL)

Arguments

Value

Invisible TRUE if valid.

Description

Negate %in% membership

Usage

not_in(x, y)

x %!in% y

Arguments

Value

Logical vector indicating non-membership.

Description

Convert Windows paths to R-friendly format

Usage

pathPrep(path = "clipboard", read_fn = NULL, write_fn = NULL)

Arguments

Value

A normalized path string.

Description

Draws the analyzed observations and annotates them with the model and omnibus tests stored by fit_recommended(). No model, assumption check or significance test is run while building the figure. The annotations therefore describe the same fit as the manuscript sentences.

Usage

plot_fit(fit, x = NULL, ylab = NULL, xlabels = NULL, show_test = TRUE)

Arguments

Details

For a categorical predictor and a numeric continuous/count outcome, repeated observations are averaged per cluster and predictor level for display. These are descriptive means, not model predictions or confidence intervals. Ordinal, binary and nominal responses are shown as counts of observed responses without averaging categories. Numeric predictors are shown as scatterplots. Other predictors remain in the fitted model whose tests are printed in the subtitle.

Value

A ggplot object. It can be saved with save_paper_figure() or combined with other plots using patchwork.

See Also

fit_recommended()

Examples

fit <- fit_recommended(iris, "Sepal.Length", "Species",
                       outcome_type = "continuous", verbose = FALSE)
plot_fit(fit)

Description

The items of one questionnaire

Usage

questionnaire_items(instrument, notes = TRUE)

Arguments

Value

Invisibly, a data frame with one row per item: number, code, wording, subscale, and whether the scoring key reverses it.

Caution – verify the mapping against your own survey

Item numbers, item order and item polarity are properties of the sheet a study actually administered, not of the instrument in the abstract. Survey tools renumber items, translations reorder them, short forms drop them from the middle, and semantic differentials get printed with the poles the other way round.

This package applies each instrument's published scoring key, which is the right default and is still only a default. If the sheet your participants saw differed, the scores will be wrong – and wrong quietly, because a mismatched mapping raises no error and produces entirely plausible numbers.

So: run check_questionnaire() once per instrument per study and read the mapping it prints, and double-check any figure before it goes into a paper. The mapping used is also attached to the result as the "mapping" attribute, and summarised in a console note the first time each distinct mapping is scored in a session (silence it with options(colleyRstats.quiet_questionnaires = TRUE)).

A named items argument (items = c(mental = "tlx_md", ...)) removes the positional assumption altogether and is the safer choice for an export you did not lay out yourself. A prefix maps columns by what their names say – an item code or label, or one item number per column running 1 to the number of items – never by sort order, and stops when the names do not identify the items. Note that the attrakdiff key is blocked by dimension with the negative pole first; data stored as answered on the official AttrakDiff sheet belong to attrakdiff_official.

See Also

check_questionnaire(), score_questionnaire()

Examples

questionnaire_items("ueq_s")

Description

Works out, from the data alone, which statistical model is appropriate for a given outcome and set of predictors, and – crucially – why. The decision follows a transparent three-question tree:

1. What is the outcome's measurement scale? (via classify_outcome(): continuous, ordinal, binary, count, nominal.) This fixes the model family. A count is checked for over-dispersion (performance::check_overdispersion(), Pearson $\chi^2$/df); an over-dispersed count is modelled as negative binomial (glmmTMB::glmmTMB(family = nbinom2) when clustered, MASS::glm.nb() otherwise).

2. Are the observations independent or clustered? A cluster column in which some cluster contributes several rows makes the design clustered, whatever design says: the cluster then needs a random effect, i.e. a mixed model. For a categorical predictor that varies within clusters and whose levels are replicated within a cluster, or a numeric covariate with enough repeated observations, by-cluster random slopes are added (Barr et al., 2013); fit_recommended() simplifies the structure when the fit is singular.

3. For a continuous outcome, do the parametric assumptions hold? A Shapiro–Wilk test on the residuals of the model that would be fitted (all predictors, their interactions and the random effects) and, for between-subjects designs, a Brown–Forsythe test (Levene's test centred on the median) across all design cells decide between a parametric model and a rank-based alternative.

Usage

recommend_test(
  data,
  outcome,
  predictors = NULL,
  cluster = NULL,
  design = c("auto", "between", "within"),
  outcome_type = c("auto", "continuous", "ordinal", "binary", "count", "nominal"),
  ordinal_max_levels = 7L,
  factors = NULL,
  random_slopes = NULL
)

recommend_analysis(
  data,
  outcome,
  predictors = NULL,
  cluster = NULL,
  design = c("auto", "between", "within"),
  outcome_type = c("auto", "continuous", "ordinal", "binary", "count", "nominal"),
  ordinal_max_levels = 7L,
  factors = NULL,
  random_slopes = NULL
)

Arguments

Details

The recommendation ranges over ordinary ANOVA (Type III sums of squares for factorial designs, via afex), heteroscedasticity-robust (HC3) ANOVA, Welch tests, rank-based methods (Kruskal–Wallis + Dunn, Wilcoxon, the Aligned Rank Transform, nparLD), generalized linear models, cumulative link models, multinomial models, and their mixed-model counterparts – linear mixed models (LMM), generalized linear mixed models (GLMM, lme4/glmmTMB), cumulative link mixed models (CLMM, ordinal) and multinomial mixed models (mclogit::mblogit). Only models that can actually be fitted to the design are recommended: continuous covariates take the (mixed) regression route, because ART and nparLD accept only categorical predictors; ART is used only for all-factor full-factorial models; nparLD only for a single within-subject factor.

Value

An object of class "colley_recommendation" (a list) with components including outcome_type, clustered, design ("between", "within", "mixed", or "between (clustered)" for between-cluster predictors with repeated rows per cluster), n_obs (complete rows), categorical (which predictors are categorical), within (which predictors vary within clusters), random (the random-effect structure), family, assumptions (normality and homogeneity results with their test statistics), dispersion (for counts), recommendation (human-readable label), model_function (the R function to call, e.g. "ordinal::clmm"), reporter (the matching colleyRstats reporter), fit_call (a ready-to-edit call as a string), alternatives, rationale, methods_text (an APA-style sentence) and methods_tex (the same, escaped for LaTeX). A print method summarises it.

Naming

recommend_test() is the canonical name. recommend_analysis() is an alias for the same function object, kept for backward compatibility.

References

Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language, 68(3), 255–278.

Examples

set.seed(1)
d <- data.frame(
  id    = factor(rep(1:20, each = 3)),
  cond  = factor(rep(c("A", "B", "C"), times = 20)),
  score = rnorm(60),
  rating = factor(sample(1:5, 60, TRUE), ordered = TRUE)
)
# Ordinal outcome measured repeatedly within subject -> CLMM
recommend_test(d, outcome = "rating", predictors = "cond", cluster = "id")
# Continuous, between-subjects -> ANOVA or its rank-based fallback
recommend_test(d, outcome = "score", predictors = "cond")

Description

Computes each respondent's Response Entropy Index (Tawa, 2021) and flags unusually low or high values. Note that no rows are removed; entries are only flagged via the Suspicious column.

Usage

remove_outliers_REI(df, header = FALSE, variables = "", range = c(1, 5))

Arguments

Details

The REI is the Shannon entropy (base 10) of a respondent's distribution of answers over the response options, $REI_i = -\sum_k p_{ki} \log_{10} p_{ki}$, where $p_{ki}$ is the proportion of the respondent's answered items that received option $k$. Low values indicate overly consistent answering (e.g. straight-lining), high values overly scattered answering; either can indicate careless responding. Because the index ignores item content, compute it on the items as answered, before reverse-coding (Tawa, 2021).

Missing responses are left out: proportions are taken over the items a respondent answered, so the same answer pattern yields the same REI however many items were skipped (dividing by the total number of items, as versions before 0.3.0 did, lowered the REI of anyone who skipped items). A respondent who answered nothing gets NA. Responses outside the declared Likert range trigger a warning (they often indicate mis-coded data, or a non-item column) but are still included in the REI computation.

Flags are relative to the sample: each REI is converted to a percentile of a normal distribution with the sample's mean and standard deviation. "Maybe" marks the outer 10% on either side (the preliminary guideline of Tawa, 2021), "Yes" the outer 5%. When all respondents have the same REI the percentiles are undefined: Percentile is NA, no row is flagged, and a warning says so.

Value

A data frame with the calculated REI, the item columns used, Percentile, and a Suspicious flag ("No", "Maybe", "Yes"; NA for a respondent without answers).

References

Tawa, J. (2021). The Response Entropy Index: Comparative assessment of performance and cultural bias across indices of careless responding. Survey Research Methods, 15(3), 299–325. doi:10.18148/srm/2021.v15i3.7832

Examples

df <- data.frame(
  id = 1:6,
  q1 = c(1, 5, 3, 3, 2, 4), q2 = c(1, 1, 4, 3, 2, 5),
  q3 = c(1, 4, 3, 3, 5, 4), q4 = c(1, 2, 4, 3, 1, 5)
)
# select the item columns; the ID column is no response
result <- remove_outliers_REI(df, TRUE, c("q1", "q2", "q3", "q4"), c(1, 5))
result

Description

Replace all occurrences of given values in all columns of a data frame. Factor levels are preserved (and extended by the replacement values), and numeric/logical columns are only touched where a value actually matches, so unrelated entries keep their exact binary representation.

Usage

replace_values(data, to_replace, replace_with)

Arguments

Value

Modified data frame with specified values replaced.

Examples

data <- data.frame(
  q1 = c("neg2", "neg1", "0"),
  q2 = c("1", "neg2", "neg1")
)

replace_values(
  data,
  to_replace = c("neg2", "neg1"),
  replace_with = c("-2", "-1")
)

Description

Runs analyze_and_report() for each dependent variable (e.g., all questionnaire scales of a study) and additionally returns a summary table of the omnibus tests with Holm-adjusted p-values across the dependent variables, plus – when patchwork is installed – a combined figure.

Usage

report_all(
  data,
  dvs,
  iv,
  design = c("between", "within"),
  labels = NULL,
  xlabels = NULL,
  plotType = "boxviolin",
  sink_dir = NULL,
  subject = NULL
)

Arguments

Details

Each dependent variable is analysed on its own complete cases (see analyze_and_report()), so in a within-subjects design a participant missing one rating is left out of that dependent variable only. The p.holm column corrects the omnibus p-values for the number of dependent variables; the per-dv sentences report the uncorrected ones.

Value

Invisibly returns a list with components results (named list of analyze_and_report() results), summary (data frame with one row per dv: method, statistic, p.value, and Holm-adjusted p.holm), and combined_plot (a patchwork figure, or NULL when patchwork is not installed).

Examples

out <- report_all(mtcars, dvs = c("mpg", "disp"), iv = "cyl")
out$summary

Generate the Latex-text based on the ARTool (see https://github.com/mjskay/ARTool). The ART result must be piped into an anova(). Only significant main and interaction effects are reported. P-values are rounded for the third digit. Each effect is accompanied by its partial eta squared, derived from the F statistic and its degrees of freedom with effectsize::F_to_eta2(), and a two-sided 95% confidence interval. A term is reported as an interaction when its name contains ":"; each factor name is rendered as a \Name macro only when that is a safe new LaTeX command, and as escaped text otherwise. Attention: the independent variables of the formula and the term specifying the participant must be factors (i.e., use as.factor()).

Description

To easily copy and paste the results to your manuscript, the following commands must be defined in Latex: \newcommand{\F}[3]{$F({#1},{#2})={#3}$} \newcommand{\p}{\textit{p=}} \newcommand{\pminor}{\textit{p$<$}}

Usage

reportART(
  model,
  dv = "Testdependentvariable",
  write_to_clipboard = FALSE,
  sink_to = NULL
)

report_art(
  model,
  dv = "Testdependentvariable",
  write_to_clipboard = FALSE,
  sink_to = NULL
)

Arguments

Value

Invisibly returns the reported sentence(s) as a character vector; the text is also emitted via message().

Naming

report_art() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportART() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

if (requireNamespace("ARTool", quietly = TRUE)) {
  set.seed(123)

  main_df <- data.frame(
    tlx_mental = stats::rnorm(80),
    Video      = factor(rep(c("A", "B"), each = 40)),
    gesture    = factor(rep(c("G1", "G2"), times = 40)),
    eHMI       = factor(rep(c("On", "Off"), times = 40)),
    UserID     = factor(rep(1:20, each = 4))
  )

  art_model <- ARTool::art(
    tlx_mental ~ Video * gesture * eHMI +
      Error(UserID / (gesture * eHMI)),
    data = main_df
  )

  model_anova <- stats::anova(art_model)
  reportART(model_anova, dv = "mental demand")
}

Description

Companion to reportDunnTest() for aligned-rank-transform (ART) models. It extracts the significant pairwise comparisons produced by ARTool::art.con() (an emmeans contrast grid), describes the groups involved from the raw data, and prints LaTeX-formatted sentences such as "An ART-C post-hoc test (Holm-adjusted) found that ...".

Usage

reportArtCon(
  ac,
  data,
  iv = "testiv",
  dv = "testdv",
  paired = FALSE,
  id = NULL,
  sink_to = NULL,
  descriptives = c("auto", "mean", "median", "trimmed")
)

report_art_con(
  ac,
  data,
  iv = "testiv",
  dv = "testdv",
  paired = FALSE,
  id = NULL,
  sink_to = NULL,
  descriptives = c("auto", "mean", "median", "trimmed")
)

Arguments

Details

The p-values are taken as-is from the contrast object, i.e. they are already adjusted by whatever adjust was passed to art.con() (e.g. "holm"), and that correction – read from the contrast summary – is named in the sentence. With adjust = "none" the p-values are emitted as \p{}/\pminor{} rather than as $p_{adj}$. The effect size is the rank-biserial correlation computed from the raw data. ART is most often used for within-subjects designs; pass paired = TRUE together with id (the subject column) to obtain the paired rank-biserial effect size.

Which level is "higher" follows the sign of the contrast estimate (first minus second level, on the aligned-rank scale). Until 0.3.0 it followed the raw means, which can disagree with the test. By default (descriptives = "auto") the levels are described by their median and IQR, which match a rank-based test; "mean" restores M/SD and warns where the means order two levels against the contrast.

The two levels of each contrast are read from the contrast coefficients, so level names that emmeans rewrites in its labels – numbers ("mode1 - mode2") or names containing -, +, * or / ("Both - (Hand-only)") – are handled.

Attention: ac must be a pairwise contrast over a single factor iv (e.g. art.con(model, ~ interaction_mode, adjust = "holm")).

Required commands in LaTeX (all part of latex_preamble()): \padj, \padjminor, \p, \pminor, \mdn, \iqr, \m, \sd and \newcommand{\rankbiserial}[1]{$r_{rb} = #1$}.

Value

Invisibly returns the reported sentence(s) as a character vector; the text is also emitted via message().

Naming

report_art_con() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportArtCon() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

if (requireNamespace("ARTool", quietly = TRUE) &&
  requireNamespace("emmeans", quietly = TRUE)) {
  set.seed(123)
  n <- 20
  df <- data.frame(
    UserID = factor(rep(seq_len(n), times = 3)),
    mode   = factor(rep(c("Hand", "Eye", "Both"), each = n)),
    prime  = factor(rep(rep(c("A", "B"), each = n / 2), times = 3))
  )
  df$score <- as.numeric(df$mode) * 2 + stats::rnorm(nrow(df))

  m  <- ARTool::art(score ~ mode * prime + Error(UserID / mode), data = df)
  ac <- ARTool::art.con(m, ~ mode, adjust = "holm")
  reportArtCon(ac, data = df, iv = "mode", dv = "score", paired = TRUE, id = "UserID")
}

Report ART contrasts (art.con) as a LaTeX table. Customizable with sensible defaults. Companion to reportDunnTestTable().

Description

The table lists the significant contrasts with their test statistic – a column headed t, or z when emmeans reports z ratios – its degrees of freedom (fractional Kenward-Roger or Satterthwaite dfs are shown to two decimals, not rounded to integers; the column is omitted for z ratios), the p-value and the rank-biserial correlation. The p-value column is headed "p-adjusted" and the caption names the correction read from the contrast summary; with adjust = "none" the header is "p" and the caption says the p-values are not adjusted. A positive statistic means the first-named level is higher on the aligned-rank scale.

Usage

reportArtConTable(
  ac,
  data,
  iv = "testiv",
  dv = "testdv",
  paired = FALSE,
  id = NULL,
  orderByP = FALSE,
  numberDigitsForPValue = 4,
  latexSize = "small",
  orderText = TRUE,
  style = c("hline", "booktabs"),
  sink_to = NULL
)

report_art_con_table(
  ac,
  data,
  iv = "testiv",
  dv = "testdv",
  paired = FALSE,
  id = NULL,
  orderByP = FALSE,
  numberDigitsForPValue = 4,
  latexSize = "small",
  orderText = TRUE,
  style = c("hline", "booktabs"),
  sink_to = NULL
)

Arguments

Value

Invisibly returns the rendered LaTeX table as a string (or NULL when xtable is unavailable); the table is also printed.

Naming

report_art_con_table() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportArtConTable() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

if (requireNamespace("ARTool", quietly = TRUE) &&
  requireNamespace("emmeans", quietly = TRUE)) {
  set.seed(123)
  n <- 20
  df <- data.frame(
    UserID = factor(rep(seq_len(n), times = 3)),
    mode   = factor(rep(c("Hand", "Eye", "Both"), each = n)),
    prime  = factor(rep(rep(c("A", "B"), each = n / 2), times = 3))
  )
  df$score <- as.numeric(df$mode) * 2 + stats::rnorm(nrow(df))

  m  <- ARTool::art(score ~ mode * prime + Error(UserID / mode), data = df)
  ac <- ARTool::art.con(m, ~ mode, adjust = "holm")
  reportArtConTable(ac, data = df, iv = "mode", dv = "score", paired = TRUE, id = "UserID")
}

Description

Reporter for ordinal proportional-odds models fitted with ordinal: cumulative link mixed models (ordinal::clmm) and their fixed-effects counterpart (ordinal::clm). Each model term is first tested with a Type III Wald $\chi^2$ test (emmeans::joint_tests(), independent of the contrast coding); each location coefficient is then reported – as an odds ratio for the logit link (the multiplicative change in the odds of being in a higher outcome category), as the raw coefficient with the link named otherwise – with its confidence interval, z statistic and p-value, labelled as the contrast it is (see reportGLMM()).

Usage

reportCLMM(
  model,
  dv = "Testdependentvariable",
  exponentiate = "auto",
  conf_level = 0.95,
  write_to_clipboard = FALSE,
  sink_to = NULL,
  omnibus = TRUE
)

report_clmm(
  model,
  dv = "Testdependentvariable",
  exponentiate = "auto",
  conf_level = 0.95,
  write_to_clipboard = FALSE,
  sink_to = NULL,
  omnibus = TRUE
)

Arguments

Details

Only the location coefficients are reported: the thresholds (cut-points, including the threshold.1/spacing parameters of equidistant thresholds), scale effects and nominal effects are not predictor effects on the location, so unlike reportGLMM() this reporter has no include_intercept argument.

Value

Invisibly returns the reported sentence(s) as a character vector; the text is also emitted via message().

Naming

report_clmm() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportCLMM() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

if (requireNamespace("ordinal", quietly = TRUE) &&
  requireNamespace("parameters", quietly = TRUE)) {
  m <- ordinal::clmm(rating ~ temp + contact + (1 | judge), data = ordinal::wine)
  reportCLMM(m, dv = "wine rating")
}

Description

Reports the significant comparisons of an FSA::dunnTest() result as sentences such as "A Dunn post-hoc test (Holm-adjusted) found that ... for the Species virginica was significantly higher (Mdn=..., IQR=...) than for setosa (Mdn=..., IQR=...; p_adj<.001, r_rb=...)". The multiplicity correction is read from the test object; with method = "none" the p-values are uncorrected and are emitted as \p{}/\pminor{}, not as $p_{adj}$.

Usage

reportDunnTest(
  d,
  data,
  iv = "testiv",
  dv = "testdv",
  sink_to = NULL,
  descriptives = c("auto", "mean", "median", "trimmed")
)

report_dunn_test(
  d,
  data,
  iv = "testiv",
  dv = "testdv",
  sink_to = NULL,
  descriptives = c("auto", "mean", "median", "trimmed")
)

Arguments

Details

Which level is "higher" follows the sign of the test's $Z$ (FSA labels a comparison "A - B"; $Z > 0$ means A has the higher mean rank). Until 0.3.0 it followed the raw means, which can point the other way: with a few large outliers a level can have the larger mean but the significantly lower mean rank. As a consistency check, the mean ranks are recomputed from data and a warning is given when they disagree with $Z$ (i.e. when data is not the data the test was computed on).

By default (descriptives = "auto") each level is described by its median and IQR, the location measure that matches a rank-based test; "mean" restores M/SD and warns where the means order two levels against the test.

Required commands in LaTeX (all part of latex_preamble()): \padj, \padjminor, \p, \pminor, \mdn, \iqr, \m, \sd and \newcommand{\rankbiserial}[1]{$r_{rb} = #1$}.

Value

Invisibly returns the reported sentence(s) as a character vector; the text is also emitted via message().

Naming

report_dunn_test() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportDunnTest() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

if (requireNamespace("FSA", quietly = TRUE)) {
  # Use built-in iris data
  data(iris)

  # Dunn test on Sepal.Length by Species
  d <- FSA::dunnTest(Sepal.Length ~ Species,
    data   = iris,
    method = "holm"
  )

  # Report the Dunn test
  reportDunnTest(d,
    data = iris,
    iv   = "Species",
    dv   = "Sepal.Length"
  )
}

Description

The table lists the significant comparisons with their $Z$, p-value and rank-biserial correlation. FSA's $Z$ for "A - B" is positive when A has the higher mean rank, which the caption states. The p-value column is headed "p-adjusted" and the caption names the correction read from the test object; a test run with method = "none" gets the header "p" and a caption saying the p-values are not adjusted. P-values are never rounded across .05 or .01.

Usage

reportDunnTestTable(
  d = NULL,
  data,
  iv = "testiv",
  dv = "testdv",
  orderByP = FALSE,
  numberDigitsForPValue = 4,
  latexSize = "small",
  orderText = TRUE,
  style = c("hline", "booktabs"),
  sink_to = NULL
)

report_dunn_test_table(
  d = NULL,
  data,
  iv = "testiv",
  dv = "testdv",
  orderByP = FALSE,
  numberDigitsForPValue = 4,
  latexSize = "small",
  orderText = TRUE,
  style = c("hline", "booktabs"),
  sink_to = NULL
)

Arguments

Value

Invisibly returns the rendered LaTeX table as a string (or NULL when xtable is unavailable); the table is also printed.

Naming

report_dunn_test_table() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportDunnTestTable() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

if (requireNamespace("FSA", quietly = TRUE)) {
  # Use built-in iris data
  data(iris)

  # Dunn test on Sepal.Length by Species
  d <- FSA::dunnTest(Sepal.Length ~ Species,
    data   = iris,
    method = "holm"
  )

  # Report the Dunn test
  reportDunnTestTable(d,
    data = iris,
    iv   = "Species",
    dv   = "Sepal.Length"
  )
}

Description

Writes the omnibus test of a ggbetweenstats()/ggwithinstats() plot as a sentence, naming the test statsExpressions ran and the effect size it produced. Rank-based tests are reported with their $\chi^2$, V or W statistic, t tests (including Yuen's trimmed-means test) as t(df), and ANOVA-type tests as F(df1, df2). A plot made with type = "bayes" carries a Bayes factor instead of a p-value; it is reported as $BF_{10}$ together with the evidence category of Jeffreys (1961), as classified by effectsize::interpret_bf().

Usage

reportggstatsplot(
  p,
  iv = "independent",
  dv = "Testdependentvariable",
  write_to_clipboard = FALSE,
  sink_to = NULL
)

report_ggstatsplot(
  p,
  iv = "independent",
  dv = "Testdependentvariable",
  write_to_clipboard = FALSE,
  sink_to = NULL
)

Arguments

Value

Invisibly returns the reported sentence(s) as a character vector; the text is also emitted via message().

Naming

report_ggstatsplot() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportggstatsplot() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

library(ggstatsplot)
library(dplyr)

# Generate a plot
plt <- ggbetweenstats(mtcars, am, mpg)

reportggstatsplot(plt, iv = "am", dv = "mpg")

Description

This function extracts significant pairwise comparisons from a ggstatsplot object, describes the groups involved from the raw data, and prints LaTeX-formatted sentences reporting the results.

Usage

reportggstatsplotPostHoc(
  data,
  p,
  iv = "testiv",
  dv = "testdv",
  label_mappings = NULL,
  sink_to = NULL,
  descriptives = c("auto", "mean", "median", "trimmed"),
  subject = NULL
)

report_ggstatsplot_posthoc(
  data,
  p,
  iv = "testiv",
  dv = "testdv",
  label_mappings = NULL,
  sink_to = NULL,
  descriptives = c("auto", "mean", "median", "trimmed"),
  subject = NULL
)

Arguments

Details

Each sentence names the post-hoc test that produced the comparison and the multiplicity correction applied to its p-value, both read from the pairwise table ggstatsplot attaches to the plot (e.g. "A Games-Howell post-hoc test (Holm-adjusted) found that ..."). Which test that is depends on the type of the plot and on whether it is between- or within-subjects – Games-Howell, Dunn, Durbin-Conover, Student's t or Yuen's trimmed means – so it is worth reporting rather than assuming. When the plot carries no such information (an older ggstatsplot, or a hand-built table) the sentence falls back to a plain "A post-hoc test ...".

Value

Invisibly returns the reported sentence(s) as a character vector; the text is also emitted via message().

Direction and descriptives

Which level "was significantly higher" is decided by the quantity the test compares, computed from data: the means for Student's t and Games-Howell, the mean ranks of the pooled ranking for Dunn, the within-participant rank sums for Durbin-Conover, and the 20% trimmed means for Yuen's test (taken from the sign of the estimate column of the pairwise table when it is present). Until 0.3.0 the direction always came from the raw means, which contradicts a rank-based test whenever outliers or skew pull a mean – the situation in which such a test is chosen.

The descriptives printed beside each level match the test by default (descriptives = "auto"): \m{}/\sd{} for the mean-based tests, \mdn{}/\iqr{} for the rank-based ones, and the 20% trimmed mean $M_t$ with the 20% winsorized standard deviation $SD_w$ for Yuen's test. Any of these can be forced; a warning is given when the forced descriptives order two levels against the test's direction.

For a within-subjects plot pass subject: ggwithinstats() tests only the participants observed in every level, and with subject the descriptives (and the Durbin-Conover rank sums) are computed from exactly those participants. Without it, all rows are described, and a Durbin-Conover table pairs observations by row order within each level – what ggwithinstats() itself does without subject.id – with a warning.

LaTeX Requirements

To easily copy and paste the results to your manuscript, the commands of latex_preamble() (or the shipped colleyRstats.sty) must be defined: \m, \sd, \mdn, \iqr, \padj, \padjminor, \p and \pminor.

\p/\pminor are used only when the plot reports p.adjust.method = "None": those p-values are uncorrected and must not be labelled $p_{adj}$.

Naming

report_ggstatsplot_posthoc() is the spelling used throughout the documentation and the one to prefer in new code: the ⁠report_*⁠ / ⁠plot_*⁠ / ⁠check_*⁠ prefixes make the API discoverable through autocomplete.

reportggstatsplotPostHoc() is the original name. Both names refer to the same function object, so they are entirely interchangeable; the original remains fully supported and is not scheduled for removal, and existing scripts keep working unchanged.

Examples

library(ggstatsplot)
library(dplyr)

# Generate a plot (a factor with three levels, so there are pairwise tests)
plt <- ggbetweenstats(mtcars, cyl, mpg)

# Report stats
reportggstatsplotPostHoc(
  data = mtcars,
  p = plt,
  iv = "cyl",
  dv = "mpg",
  label_mappings = list("4" = "four cylinders")
)

Description

Turns a fitted mixed model into ready-to-paste manuscript sentences: first a Type III omnibus test per model term (main effects and interactions), then one sentence per fixed-effect coefficient with its estimate (or odds / incidence-rate ratio), confidence interval, test statistic and p-value. Works with linear mixed models (lme4::lmer, lmerTest::lmer), generalized linear mixed models (lme4::glmer, glmmTMB::glmmTMB) and, for convenience, ordinary lm, glm and MASS::glm.nb fits.

Usage

reportGLMM(
  model,
  dv = "Testdependentvariable",
  exponentiate = "auto",
  include_intercept = FALSE,
  conf_level = 0.95,
  write_to_clipboard = FALSE,
  sink_to = NULL,
  omnibus = TRUE
)

report_glmm(
  model,
  dv = "Testdependentvariable",
  exponentiate = "auto",
  include_intercept = FALSE,
  conf_level = 0.95,
  write_to_clipboard = FALSE,
  sink_to = NULL,
  omnibus = TRUE
)

Arguments

Details

Omnibus tests. With treatment (dummy) coding – R's default – a coefficient such as aa2 in y ~ a * b is the a2 - a1 difference at the reference level of b, not "the effect of a", and a factor with three or more levels has no single coefficient at all. The omnibus tests answer the questions a results section asks: for linear mixed models, lmerTest's Type III $F$-tests with Satterthwaite's degrees of freedom; for all other models, emmeans::joint_tests() (Type III Wald $F$ for lm, Wald $\chi^2$ otherwise). Both are independent of the contrast coding. The coefficient sentences that follow are labelled as what they are, e.g. "the contrast a2 vs. a1 of a (at b = b1)", "the slope of age (at cond = A)".