Feature request for new function ggbarstats about ggstatsplot

[ibecav]

Yes I remember the concept of minimizing the number of new functions. Problem is three fold:

1. I need the functionality now for my day job so it's either invest time in a one off or move the overall package along.
2. Piecharts really are different than bars or mosaics since the base functionality is univariate whereas bars and mosaics are best for bivariate cases (then for both grouped_ functionality allows you to go farther
3. Related ggcatstats would be a misnomer since bars and mosaics deal nicely with ordinals and other levels of variables.

Anyway I will certainly endeavor to minimize the amount of new or overlapping code. When I get to a good stable point (soon) I'll formally submit a PR and you can take a look.

[ibecav]

Let's see there's this usual issue that I know you have a method for...

checking R code for possible problems (11.8s)
ggbarstats: no visible binding for global variable ‘N’
Undefined global functions or variables: N

The minor grid fix on y is very nice I had removed the x grid

Can we make the outline color selecatable. I don't care about the default

Ditto sample size. I don't much care about the defaults as long as I can adjust

[IndrajeetPatil]

Completely agree with this! The plan is actually to write a more general version of the function ggpiestats and call it ggcatstats (for categorical data) and provide a plot.type argument that can work with "pie", "bar", "slopegraph", and "alluvial" plots.

This is something that should happen by 1.0.0, so still a few months away. But I can already start adding some basic functionality/setup needed for this.

[ibecav]

Understand and as long as you're thinking about a plot.type argument please consider mosaic using ggmosaic. If you're unfamiliar it's like a barchart only the x axis factors are proportional...

library(ggplot2)
library(dplyr)
library(ggstatsplot)
library(ggmosaic)
ddd <- group_by(Titanic_full, Class, Sex, Age, Survived) %>% count()
ggplot(data=ddd) +
  geom_mosaic(aes(weight=n, x=product(Class), fill=Survived)) +
  ggtitle("test", subtitle = ggstatsplot::subtitle_contigency_tab(Titanic_full,Class,Survived))
#> Note: 95% CI for Cramer's V was computed with 25 bootstrap samples.
#> 

^{Created on 2018-10-30 by the reprex package (v0.2.1)}

[ibecav]

I'm about 65% done with this... have not created a pull request yet but working.

[IndrajeetPatil]

Hmm, but the plan was not to create a new function just for the bar plots, but rather to create a general function ggcatstats (for categorical data) that has plot.type argument and deprecate ggpiestats. The problem with writing a separate function for each type of plot is that there is a lot of duplication of code across plotting functions. This means each modification will simultaneously need to be made to all functions since only the plotting portion of the function will change across functions.

[IndrajeetPatil]

Fair enough!

Let's proceed with separate functions and we can later figure out how to refactor them to avoid code duplication.

[ibecav]

Getting there. An example using one of the common datasets.

library(ggstatsplot)
# using the current ggpiestats
# pies are especially inefficent when you have many levels of a factor
ggpiestats(movies_long,
           mpaa,
           genre,
           bf.message = TRUE,
           sampling.plan = "jointMulti",
           title = "MPAA Ratings by Genre",
           caption = "As of January 23, 2019",
           nboot = 5,
           perc.k = 1,
           facet.proptest = FALSE,
           palette = "Set2")
#> Note: 95% CI for effect size estimate was computed with 5 bootstrap samples.
#> 

# using nthe still nascent ggbarstats
ggbarstats(movies_long,
           mpaa,
           genre,
           bf.message = TRUE,
           sampling.plan = "jointMulti",
           title = "MPAA Ratings by Genre",
           caption = "As of January 23, 2019",
           nboot = 5,
           perc.k = 1,
           facet.proptest = FALSE,
           palette = "Set2")
#> Note: 95% CI for effect size estimate was computed with 5 bootstrap samples.
#> 

^{Created on 2019-01-23 by the reprex package (v0.2.1)}

[IndrajeetPatil]

Thanks, Chuck. This looks awesome!

In case you already don't have these changes in mind, here are some minor comments to have aeshetic consistency across the plots from these two similar functions-

1. The sample size should be denoted using n = (in ggpiestats) and not N= (in ggbarstats). Also, they should not be bold (in ggbarstats).
2. The percentages inside the bars should have text box around them with a white color background.
3. The y-axis label should be "percent" or "percentage" (or even "proportion", like in gghistostats)? I don't feel strongly either way.
4. The default legend position should be at the bottom (in ggpiestats) and not on the right hand side (in ggbarstats).

[ibecav]

Laugh out loud. We have such different aesthetic tastes. But okay using your numbering system

1. Okay n instead of N but I am worried that on wider plots the spaces around the equals sign take up a lot of room.
2. Okay i'll use geom_label instead of geom_text. But since I detest the white background I'll allow the user to select color and alpha
3. Technically percent or percentage is more correct
4. Legend at the bottom I also detest so I'll make bottom the default and allow user to move it amongst "top", "bottom", "left", and "right".

Should be able to do these and cleanup tomorrow.

[ibecav]

Okay all these are fixed as well as adding a few features. PR follows. Need some more doco cleanup and some testing but it's done.

library(jmv)
#> 
#> Attaching package: 'jmv'
#> The following object is masked from 'package:stats':
#> 
#>     anova
library(ggstatsplot)
# for reproducibility
set.seed(123)

# simple function call with the defaults (with condition)
ggstatsplot::ggbarstats(
  data = datasets::mtcars,
  main = vs,
  condition = cyl,
  bf.message = TRUE,
  nboot = 10,
  factor.levels = c("0 = V-shaped", "1 = straight"),
  legend.title = "Engine"
)
#> Warning in stats::chisq.test(x = data$main, y = data$condition, correct =
#> FALSE, : Chi-squared approximation may be incorrect
#> Note: 95% CI for effect size estimate was computed with 10 bootstrap samples.
#> 

# simple function call with count data

ggstatsplot::ggbarstats(
  data = as.data.frame(HairEyeColor),
  main = Eye,
  condition = Hair,
  counts = Freq
)
#> Note: 95% CI for effect size estimate was computed with 25 bootstrap samples.
#> 

ggbarstats(movies_long,
  mpaa,
  genre,
  bf.message = TRUE,
  sampling.plan = "jointMulti",
  title = "MPAA Ratings by Genre",
  caption = "As of January 23, 2019",
  nboot = 5,
  perc.k = 1,
  x.axis.orientation = "slant",
  facet.proptest = FALSE,
  ggplot.component = ggplot2::theme(axis.text.x = ggplot2::element_text( face = "italic")),
  palette = "Set2"
)
#> Note: 95% CI for effect size estimate was computed with 5 bootstrap samples.
#> 

^{Created on 2019-01-24 by the reprex package (v0.2.1)}

[IndrajeetPatil]

Thanks, Chuck. This all looks good to me. I will make minor modifications after the PR is merged-

1. Add ( to sample sizes to be consistent with ggpiestats.
2. The size of the sample size label text seems a bit small compared to the rest of the label text in the plot. I'll toy around with different values to see if a bit bigger text size looks better.
3. Will also add me to the author field (#' @author Chuck Powell -> #' @author Chuck Powell, Indrajeet Patil) since the entire non-plotting related body of the function comes from ggpiestats which I had written.

Would also like to add grouped_ggbarstats and tests for this function or do you want me to take care of it?
I am fine either way.

[IndrajeetPatil]

Here is what the output looks like with the modifications I have introduced to make this aesthetically as close to ggpiestats as possible. Lemme know what you think before I push these changes to master.

Reproducing the same examples you used above-

• Example 1

set.seed(123)

ggstatsplot::ggbarstats(
  data = datasets::mtcars,
  main = vs,
  condition = cyl,
  bf.message = TRUE,
  nboot = 10,
  factor.levels = c("0 = V-shaped", "1 = straight"),
  legend.title = "Engine"
)
#> Note: Results from one-sample proportion tests for each
#>       level of the condition variable testing for equal
#>       proportions of the main variable.
#> 
#> # A tibble: 3 x 7
#>   condition `0`     `1`    `Chi-squared`    df `p-value` significance
#>   <fct>     <chr>   <chr>          <dbl> <dbl>     <dbl> <chr>       
#> 1 4         9.09%   90.91%         7.36      1     0.007 **          
#> 2 6         42.86%  57.14%         0.143     1     0.705 ns          
#> 3 8         100.00% 0.00%         14         1     0     ***
#> Warning in stats::chisq.test(x = data$main, y = data$condition, correct =
#> FALSE, : Chi-squared approximation may be incorrect
#> Note: 95% CI for effect size estimate was computed with 10 bootstrap samples.
#> 

• Example 2

set.seed(123)

ggstatsplot::ggbarstats(
  data = as.data.frame(HairEyeColor),
  main = Eye,
  condition = Hair,
  counts = Freq
)
#> Note: Results from one-sample proportion tests for each
#>       level of the condition variable testing for equal
#>       proportions of the main variable.
#> 
#> # A tibble: 4 x 9
#>   condition Brown Blue  Hazel Green `Chi-squared`    df `p-value`
#>   <fct>     <chr> <chr> <chr> <chr>         <dbl> <dbl>     <dbl>
#> 1 Black     62.9~ 18.5~ 13.8~ 4.63%         87.3      3     0    
#> 2 Brown     41.6~ 29.3~ 18.8~ 10.1~         63.3      3     0    
#> 3 Red       36.6~ 23.9~ 19.7~ 19.7~          5.45     3     0.142
#> 4 Blond     5.51% 74.0~ 7.87% 12.6~        164.       3     0    
#> # ... with 1 more variable: significance <chr>
#> Note: 95% CI for effect size estimate was computed with 25 bootstrap samples.
#> 

• Example 3

set.seed(123)

ggstatsplot::ggbarstats(
  ggstatsplot::movies_long,
  mpaa,
  genre,
  bf.message = TRUE,
  sampling.plan = "jointMulti",
  title = "MPAA Ratings by Genre",
  caption = "As of January 23, 2019",
  nboot = 5,
  perc.k = 1,
  x.axis.orientation = "slant",
  facet.proptest = FALSE,
  ggplot.component = ggplot2::theme(axis.text.x = ggplot2::element_text(face = "italic")),
  palette = "Set2"
)
#> Note: 95% CI for effect size estimate was computed with 5 bootstrap samples.
#> 

^{Created on 2019-01-24 by the reprex package (v0.2.1)}

[ibecav]

So the biggest difference I can see visually is that you want the significance testing results from the proportions test up top rather than down below? I can certainly live with that.

Hard for me to see anything else you've done.

[IndrajeetPatil]

Yes, additional changes that are not conspicuous-

• minor grid for y-axis is restricted to [0,1]
• outline color for bars has been changed from grey to black
• the sample size label has bigger text size

And that's about it.

If you are okay with these changes, I will push these changes to master and close this issue.

[IndrajeetPatil]

@ibecav I have added tests for this function.

Do you also want to make a PR for the grouped_ version of this function?

[ibecav]

Yes I will happily do so. I've been thinking about the other issue (how to consolidate top level functions effectively) as well. One strategy that comes to my mind is to group more by the number of variables involved (univariate, bivariate & multivariate) and less by plot type. For example in my mind pie charts are more like histograms and belong in that grouping whereas chi square association tests are bivariate and not much different in many ways than ggbetweenstats.

Just early thinking. More when I have a chance to think it out.