4

I have following data and code:

> mat <- matrix(c(120,230,84,70,130,83,13,26,18),3)
> dimnames(mat) <- list(c("good","fair","poor"),c("a","b","c"))
> mat

> mat
       a   b  c
good 120  70 13
fair 230 130 26
poor  84  83 18
> 
> chisq.test(mat)

        Pearson's Chi-squared test

data:  mat
X-squared = 11.411, df = 4, p-value = 0.02231

> library(vcd)
> mosaic(mat, shade=T)

enter image description here

The chi-squared test shows significant difference but no color is seen since residuals are small. With higher residuals, colors are seen:

enter image description here

?mosaic shows this information: 

highlighting_fill: color vector or palette function used for a highlighted variable, if any.

How can I change fill color for above plot?

rnso
  • 23,686
  • 25
  • 112
  • 234

2 Answers2

3

No really an elegant solution but you can manually define your colours using gpar

mosaic(mat, shade=T, colorize = T, gp = gpar(fill=matrix(c("red","blue","light green", "green", "black","red","brown","white","blue"), 3, 3)))

The next step is to populate the matrix in gpar with sensible colours. There is a pretty good post here about how to assign a colour ramp based on values. https://stackoverflow.com/a/6320940/1502898

Community
  • 1
  • 1
jprockbelly
  • 1,533
  • 15
  • 30
3

By default, the shading of mosaic plots in vcd uses the cut-offs 2 and 4 for the absolute value of the residuals associated with each cell as suggested by Friendly (1994, Journal of the American Statistical Association, 89, 190-200). The idea behind this is that these cut-offs should be crossed with approximately 5% and 0.01% if the model fits well (due to the approximate normality of the residuals). However, these cut-offs corresponds to individual significance of the cells while the chi-squared test assesses overall significance.

There are two strategies to deal with this: (1) As the Friendly (1994) shading is intended to bring out the pattern of deviation rather than to correspond directly to significance, one can simply modify the cut-offs to lower values that work well for a given table (e.g., 1 and 1.8 here). (2) Switch to a test of the maximum absolute residuals (rather than their sum of squares) so that all cells that are significant at certain significance levels (e.g., 10% and 1%) can be shaded. Both strategies are implemented in vcd and illustrated in the examples of ?shading_hcl (using an example with very similar properties to your mat). The underlying ideas are explained in Zeileis, Meyer, Hornik (2007). "Residual-Based Shadings for Visualizing (Conditional) Independence", Journal of Computational and Graphical Statistics, 16(3), 507-525. doi:10.1198/106186007X237856

Strategy (1) can be applied to your data by:

mosaic(mat, gp = shading_hcl, gp_args = list(interpolate = c(1, 1.8)))

mosaic-hcl Strategy (2) is implemented as:

set.seed(403)
mosaic(mat, gp = shading_max)

mosaic-max

Achim Zeileis
  • 15,710
  • 1
  • 39
  • 49
  • Thanks for your answer. Using both shade=T and gp=shading_max options produces even better color. Can there be some automatic variables instead of c(1,1.8) for gp_args ? – rnso Apr 24 '15 at 07:01
  • Setting both `shade = TRUE` and `gp = shading_max` is equivalent to just setting `gp = shading_max`. Note, though, that the maximum test is permutation-based and hence results may change if (as in this case) the results can switch between significant and non-significant (by default judged at 5% level). – Achim Zeileis Apr 24 '15 at 08:17
  • Automatic selection of cut-offs is difficult. The inference-based way is to use the critical values of the maximum statistic as in `shading_max`. The other cut-offs are only heuristics and you could either use other quantiles of the normal distribution, e.g., `qnorm(c(0.85, 0.96))`, or quantiles of the empirical residuals, e.g., `quantile(abs(residuals(chisq.test(mat))), c(0.7, 0.9))`. However, these may or may not lead to useful patterns in the visualization of the mosaic plot. In particular, the latter will _always_ have colored cells, no matter whether there is significance or not. – Achim Zeileis Apr 24 '15 at 08:20