Skip to contents

mosaic_analysis() tests the association between two categorical columns of a data frame and draws it as a mosaic. It builds their contingency table, drops sparse categories (fewer than min_count rows), runs a Pearson chi-square test (or Fisher's exact test), computes Cramer's V with an effect-size label, and draws a flat ggplot2 mosaic in which tile area is the cell count and the fill is the cell's standardised residual. The per-cell results come as a data frame, one row per cell.

The p-value tests the null hypothesis that the two variables are independent; with large samples almost any departure is significant, so read Cramer's V for the size of the association. The label follows Cohen's (1988) thresholds for the table's degrees of freedom, min(rows, columns) - 1: 0.10 / 0.30 / 0.50 for small / medium / large at 1, 0.07 / 0.21 / 0.35 at 2, 0.06 / 0.17 / 0.29 at 3, and 0.05 / 0.15 / 0.25 above (Cohen's w divided by the square root of the degrees of freedom); a V below the small threshold is labelled "negligible". The residuals are chisq.test()'s standardised residuals, roughly standard normal under independence, so one beyond about +-2 marks a cell that holds more (or fewer) rows than independence predicts; with many cells, some pass that line by chance.

Usage

mosaic_analysis(
  data,
  var1,
  var2,
  min_count = 10L,
  test = c("chisq", "fisher"),
  percentage_base = c("total", "row", "column"),
  tile_label = c("count", "percent", "residual", "category", "none"),
  title = "",
  ...
)

# S3 method for class 'mosaic_analysis'
plot(x, ...)

# S3 method for class 'mosaic_analysis'
print(x, ...)

# S3 method for class 'mosaic_analysis'
summary(object, ...)

Arguments

data

A data frame holding the two variables. Rows missing either of them are left out; a var1 or var2 that is not a column of data, or no complete rows, raises an error.

var1

The name of the first variable, drawn as the mosaic's columns (column width is its marginal share).

var2

The name of the second variable, stacked within each column.

min_count

The smallest number of rows a category must have to be kept. Categories of either variable with fewer rows in the full table are dropped, in one pass, before the test. Default 10. If fewer than two categories of either variable remain, an error is raised.

test

The test reported: "chisq" (default) for Pearson's chi-square test, or "fisher" for Fisher's exact test: exact for a 2 x 2 table, and otherwise with a p-value simulated from 10,000 tables (so it changes with the random seed and cannot fall below 1/10,001). Cramer's V and the residuals always come from the chi-square statistic.

percentage_base

What the pct column and the "percent" tile label are a percentage of: "total" (default) the whole table, "row" each level of var1, "column" each level of var2.

tile_label

What to print inside each tile: "count" (default), "percent", "residual", "category" (the var2 level) or "none".

title

The plot title. Default "" (none).

...

Further styling for the flat mosaic: col_label_side, row_label_side, col_label_angle, row_label_angle, show_legend, legend_position, legend_size, legend_title, label_size, range, palette, min_label_h and min_label_w (and pct_base, which overrides percentage_base in the plot only; the pct column always follows percentage_base). The fill is the ColorBrewer RdBu ramp unless palette is given. Column labels turn vertical when there are more than 6 columns; col_label_angle sets an angle. Any other name raises an error. For the plot() method: the same styling and title, overriding what the fit stored; for the print() and summary() methods: ignored.

x

For the plot() and print() methods: a mosaic_analysis.

object

For the summary() method: a mosaic_analysis.

Value

A mosaic_analysis object. print() shows the sample size, the test, Cramer's V with its label and the dropped categories, and returns the object invisibly; summary() returns the per-cell data frame (the counts table below) with the one-row test summary attached as the attribute "stats"; plot() draws the mosaic on the active device, with new styling if given, and returns the ggplot invisibly. The object is a list with:

plot

The flat mosaic, a ggplot. Building the object does not draw it.

counts

A data frame with one row per (var1, var2) cell: the two variables (named after them), observed, expected, residual (the standardised residual) and pct (on percentage_base); expected and residual are rounded to 3 decimals, pct to 2.

stats

A one-row data frame: test, statistic, df, p_value, cramers_v, effect_size and n, rounded (statistic and cramers_v to 3 decimals, p_value to 4 significant digits). statistic and df are NA for Fisher's test.

test

The htest object of the test reported.

cramers_v, effect_size

Cramer's V, unrounded, and its label.

table

The filtered contingency table.

removed

The dropped categories, as a list with var1 and var2.

n_original, n_filtered

The number of complete rows before and after the sparse categories are dropped.

vars

The two variable names, as c(var1 = , var2 = ).

plot_parts, plot_args

The residual matrix, the table and the styling, kept so plot() can redraw without testing again.

Methods

  • plot.mosaic_analysis() redraws the flat mosaic from the stored contingency table and residuals, so the styling can change without the test being run again. Any flat-mosaic styling argument (tile_label, pct_base, col_label_side, legend_size, ...) can be overridden through ....

References

Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.). Lawrence Erlbaum.

See also

mosaic_plot for the network/table mosaic (which also accepts style = "flat").

Examples

data(group_regulation_long, package = "Nestimate")
res <- mosaic_analysis(group_regulation_long, "Course", "Action",
                       min_count = 20)
res
#> Mosaic analysis: Course x Action
#>   N = 27533 (filtered from 27533); table 3 x 9
#>   Chi-square: X2 = 233.560, df = 16, p < 2.2e-16
#>   Cramer's V = 0.065 (negligible)
head(summary(res))
#>   Course   Action observed expected residual  pct
#> 1      A    adapt      140  249.303   -9.430 0.51
#> 2      B    adapt      261  193.688    6.059 0.95
#> 3      C    adapt      153  111.009    4.502 0.56
#> 4      A cohesion      923  827.560    4.631 3.35
#> 5      B cohesion      590  642.945   -2.680 2.14
#> 6      C cohesion      326  368.495   -2.563 1.18
# \donttest{
plot(res, tile_label = "percent")

# }