insight (version 0.18.6)

get_datagrid: Create a reference grid

Description

Create a reference matrix, useful for visualisation, with evenly spread and combined values. Usually used to make generate predictions using get_predicted(). See this vignette for a tutorial on how to create a visualisation matrix using this function.

Usage

get_datagrid(x, ...)
# S3 method for data.frame
get_datagrid(
  x,
  at = "all",
  factors = "reference",
  numerics = "mean",
  preserve_range = FALSE,
  reference = x,
  length = 10,
  range = "range",
  ...
)
# S3 method for numeric
get_datagrid(x, length = 10, range = "range", ...)
# S3 method for factor
get_datagrid(x, ...)
# S3 method for default
get_datagrid(
  x,
  at = "all",
  factors = "reference",
  numerics = "mean",
  preserve_range = TRUE,
  reference = x,
  include_smooth = TRUE,
  include_random = FALSE,
  include_response = FALSE,
  data = NULL,
  verbose = TRUE,
  ...
)

Value

Reference grid data frame.

Arguments

x

An object from which to construct the reference grid.

...

Arguments passed to or from other methods (for instance, length or range to control the spread of numeric variables.).

at

Indicates the focal predictors (variables) for the reference grid and at which values focal predictors should be represented. If not specified otherwise, representative values for numeric variables or predictors are evenly distributed from the minimum to the maximum, with a total number of length values covering that range (see 'Examples'). Possible options for at are:

• "all", which will include all variables or predictors.

• a character vector of one or more variable or predictor names, like c("Species", "Sepal.Width"), which will create a grid of all combinations of unique values. For factors, will use all levels, for numeric variables, will use a range of length length (evenly spread from minimum to maximum) and for character vectors, will use all unique values.

• a list of named elements, indicating focal predictors and their representative values, e.g. at = list(c(Sepal.Length = c(2, 4), Species = "setosa")).

• a string with assignments, e.g. at = "Sepal.Length = 2" or at = c("Sepal.Length = 2", "Species = 'setosa'") - note the usage of single and double quotes to assign strings within strings.

There is a special handling of assignments with brackets, i.e. values defined inside [ and ].For numeric variables, the value(s) inside the brackets should either be

• two values, indicating minimum and maximum (e.g. at = "Sepal.Length = [0, 5]"), for which a range of length length (evenly spread from given minimum to maximum) is created.

• more than two numeric values at = "Sepal.Length = [2,3,4,5]", in which case these values are used as representative values.

• a "token" that creates pre-defined representative values:

  • for mean and -/+ 1 SD around the mean: "x = [sd]"

  • for median and -/+ 1 MAD around the median: "x = [mad]"

  • for Tukey's five number summary (minimum, lower-hinge, median, upper-hinge, maximum): "x = [fivenum]"

  • for terciles, including minimum and maximum: "x = [terciles]"

  • for terciles, excluding minimum and maximum: "x = [terciles2]"

  • for quartiles, including minimum and maximum: "x = [quartiles]"

  • for quartiles, excluding minimum and maximum: "x = [quartiles2]"

  • for minimum and maximum value: "x = [minmax]"

  • for 0 and the maximum value: "x = [zeromax]"

For factor variables, the value(s) inside the brackets should indicate one or more factor levels, like at = "Species = [setosa, versicolor]". Note: the length argument will be ignored when using brackets-tokens.

The remaining variables not specified in at will be fixed (see also arguments factors and numerics).

factors

Type of summary for factors. Can be "reference" (set at the reference level), "mode" (set at the most common level) or "all" to keep all levels.

numerics

Type of summary for numeric values. Can be "all" (will duplicate the grid for all unique values), any function ("mean", "median", ...) or a value (e.g., numerics = 0).

preserve_range

In the case of combinations between numeric variables and factors, setting preserve_range = TRUE will drop the observations where the value of the numeric variable is originally not present in the range of its factor level. This leads to an unbalanced grid. Also, if you want the minimum and the maximum to closely match the actual ranges, you should increase the length argument.

reference

The reference vector from which to compute the mean and SD. Used when standardizing or unstandardizing the grid using effectsize::standardize.

length

Length of numeric target variables selected in "at". This arguments controls the number of (equally spread) values that will be taken to represent the continuous variables. A longer length will increase precision, but can also substantially increase the size of the datagrid (especially in case of interactions). If NA, will return all the unique values. In case of multiple continuous target variables, length can also be a vector of different values (see examples).

range

Option to control the representative values given in at, if no specific values were provided. range can be one of the following:

• "range" (default), will use the minimum and maximum of the original data vector as end-points (min and max).

• if an interval type is specified, such as "iqr", "ci", "hdi" or "eti", it will spread the values within that range (the default CI width is 95% but this can be changed by adding for instance ci = 0.90.) See IQR() and bayestestR::ci(). This can be useful to have more robust change and skipping extreme values.

• if "sd" or "mad", it will spread by this dispersion index around the mean or the median, respectively. If the length argument is an even number (e.g., 4), it will have one more step on the positive side (i.e., -1, 0, +1, +2). The result is a named vector. See 'Examples.'

• "grid" will create a reference grid that is useful when plotting predictions, by choosing representative values for numeric variables based on their position in the reference grid. If a numeric variable is the first predictor in at, values from minimum to maximum of the same length as indicated in length are generated. For numeric predictors not specified at first in at, mean and -1/+1 SD around the mean are returned. For factors, all levels are returned.

include_smooth

If x is a model object, decide whether smooth terms should be included in the data grid or not.

include_random

If x is a mixed model object, decide whether random effect terms should be included in the data grid or not. If include_random is FALSE, but x is a mixed model with random effects, these will still be included in the returned grid, but set to their "population level" value (e.g., NA for glmmTMB or 0 for merMod). This ensures that common predict() methods work properly, as these usually need data with all variables in the model included.

include_response

If x is a model object, decide whether the response variable should be included in the data grid or not.

data

Optional, the data frame that was used to fit the model. Usually, the data is retrieved via get_data().

verbose

Toggle warnings.

See Also

get_predicted()

Examples

# Datagrids of variables and dataframes =====================================
if (require("bayestestR", quietly = TRUE) & require("datawizard", quietly = TRUE)) {
  # Single variable is of interest; all others are "fixed" ------------------
  # Factors
  get_datagrid(iris, at = "Species") # Returns all the levels
  get_datagrid(iris, at = "Species = c('setosa', 'versicolor')") # Specify an expression

  # Numeric variables
  get_datagrid(iris, at = "Sepal.Length") # default spread length = 10
  get_datagrid(iris, at = "Sepal.Length", length = 3) # change length
  get_datagrid(iris[2:150, ],
    at = "Sepal.Length",
    factors = "mode", numerics = "median"
  ) # change non-targets fixing
  get_datagrid(iris, at = "Sepal.Length", range = "ci", ci = 0.90) # change min/max of target
  get_datagrid(iris, at = "Sepal.Length = [0, 1]") # Manually change min/max
  get_datagrid(iris, at = "Sepal.Length = [sd]") # -1 SD, mean and +1 SD
  get_datagrid(iris, at = "Sepal.Length = [quartiles]") # quartiles

  # Numeric and categorical variables, generating a grid for plots
  # default spread length = 10
  get_datagrid(iris, at = c("Sepal.Length", "Species"), range = "grid")
  # default spread length = 3 (-1 SD, mean and +1 SD)
  get_datagrid(iris, at = c("Species", "Sepal.Length"), range = "grid")


  # Standardization and unstandardization
  data <- get_datagrid(iris, at = "Sepal.Length", range = "sd", length = 3)
  data$Sepal.Length # It is a named vector (extract names with `names(out$Sepal.Length)`)
  datawizard::standardize(data, select = "Sepal.Length")
  data <- get_datagrid(iris, at = "Sepal.Length = c(-2, 0, 2)") # Manually specify values
  data
  datawizard::unstandardize(data, select = "Sepal.Length")


  # Multiple variables are of interest, creating a combination --------------
  get_datagrid(iris, at = c("Sepal.Length", "Species"), length = 3)
  get_datagrid(iris, at = c("Sepal.Length", "Petal.Length"), length = c(3, 2))
  get_datagrid(iris, at = c(1, 3), length = 3)
  get_datagrid(iris, at = c("Sepal.Length", "Species"), preserve_range = TRUE)
  get_datagrid(iris, at = c("Sepal.Length", "Species"), numerics = 0)
  get_datagrid(iris, at = c("Sepal.Length = 3", "Species"))
  get_datagrid(iris, at = c("Sepal.Length = c(3, 1)", "Species = 'setosa'"))

  # With list-style at-argument
  get_datagrid(iris, at = list(Sepal.Length = c(1, 3), Species = "setosa"))
}

# With models ===============================================================
# Fit a linear regression
model <- lm(Sepal.Length ~ Sepal.Width * Petal.Length, data = iris)
# Get datagrid of predictors
data <- get_datagrid(model, length = c(20, 3), range = c("range", "sd"))
# same as: get_datagrid(model, range = "grid", length = 20)
# Add predictions
data$Sepal.Length <- get_predicted(model, data = data)
# Visualize relationships (each color is at -1 SD, Mean, and + 1 SD of Petal.Length)
plot(data$Sepal.Width, data$Sepal.Length,
  col = data$Petal.Length,
  main = "Relationship at -1 SD, Mean, and + 1 SD of Petal.Length"
)