improving summary output for BFBayesFactor objects

[IndrajeetPatil]

We should include the name of the frequentist effect size estimate in the Parameter column:

For example:

current behavior

library(BayesFactor)
library(parameters)

ttestBF(mtcars$wt, mu = 3) |> parameters() |> as.data.frame()
#>     CI     CI_low   CI_high  Parameter     Median     pd ROPE_Percentage
#> 1 0.95 -0.5512066 0.1682628 Difference -0.2039748 0.8745       0.2365167
#> 2 0.95 -0.1729238 0.5331445       <NA>         NA     NA              NA
#>   Prior_Distribution Prior_Location Prior_Scale        BF  Cohens_d
#> 1             cauchy              0   0.7071068 0.3868518        NA
#> 2               <NA>             NA          NA        NA 0.1977558
#>            Method
#> 1 Bayesian t-test
#> 2 Bayesian t-test

^{Created on 2021-10-20 by the reprex package (v2.0.1)}

expected behavior

library(BayesFactor)
library(parameters)

ttestBF(mtcars$wt, mu = 3) |> parameters() |> as.data.frame()
#>     CI     CI_low   CI_high  Parameter     Median     pd ROPE_Percentage
#> 1 0.95 -0.5512066 0.1682628 Difference -0.2039748 0.8745       0.2365167
#> 2 0.95 -0.1729238 0.5331445 Cohens_d  0.1977558     NA              NA
#>   Prior_Distribution Prior_Location Prior_Scale        BF  
#> 1             cauchy              0   0.7071068 0.3868518       
#> 2               <NA>             NA          NA        NA 
#>            Method
#> 1 Bayesian t-test
#> 2 Bayesian t-test

The same for output from contingency table analysis where we return Cramer's V, Phi, etc...

[IndrajeetPatil]

I would personally vote for One row for estimate + effect size, if and only if we also make the behavior of model_parameters() for BFBayesFactor objects similar to that for htest: return frequentist effect sizes only when the user asks for them.

So, for example, the following would return only Median + CI columns:

BayesFactor::ttestBF(mtcars$wt, mu = 3) |> 
  parameters::model_parameters()

While, this would return Median + CI and Cohen's d + CI columns:

BayesFactor::ttestBF(mtcars$wt, mu = 3) |> 
  parameters::model_parameters(standardized_d = TRUE)

[strengejacke]

Currently, we merge the results from describe_posterior() and effectsize(), so new rows are added. Instead, we could just cbind(), or not?

[IndrajeetPatil]

What remains to be decided:
Should we bind the frequentist effect size plus CIs as additional rows or columns?

[strengejacke]

I think we decided on columns, as we have for htests?

[IndrajeetPatil]

Yes, that will work for t-tests, but the output from Contingency table analysis is an exception since it's not a one-row return:

library(BayesFactor)
library(parameters)
data(raceDolls)

contingencyTableBF(raceDolls, sampleType = "indepMulti", fixedMargin = "cols") |>
  parameters() |>
  as.data.frame()
#> Warning: Could not estimate a good default ROPE range. Using 'c(-0.1, 0.1)'.
#>     CI     CI_low   CI_high Parameter    Median pd ROPE_Percentage
#> 1 0.95 0.23724792 0.3778516 cell[1,1] 0.3088749  1      0.00000000
#> 2 0.95 0.08060299 0.1836630 cell[2,1] 0.1331295  1      0.09129177
#> 3 0.95 0.23433029 0.3741541 cell[1,2] 0.2979956  1      0.00000000
#> 4 0.95 0.19290301 0.3234083 cell[2,2] 0.2563099  1      0.00000000
#> 5   NA         NA        NA     Ratio        NA NA              NA
#> 6 0.95 0.02164193 0.3011174      <NA>        NA NA              NA
#>        Prior_Distribution Prior_Location Prior_Scale       BF Cramers_v
#> 1                    <NA>             NA          NA 1.814856        NA
#> 2                    <NA>             NA          NA 1.814856        NA
#> 3                    <NA>             NA          NA 1.814856        NA
#> 4                    <NA>             NA          NA 1.814856        NA
#> 5 independent multinomial              0           1 1.814856        NA
#> 6                    <NA>             NA          NA       NA 0.1621184
#>                                Method
#> 1 Bayesian contingency table analysis
#> 2 Bayesian contingency table analysis
#> 3 Bayesian contingency table analysis
#> 4 Bayesian contingency table analysis
#> 5 Bayesian contingency table analysis
#> 6 Bayesian contingency table analysis

Compare it to htest return:

library(parameters)

chisq.test(table(mtcars$am, mtcars$cyl)) |> 
  parameters(cramers_v = TRUE) |> 
  as.data.frame()
#> Warning in chisq.test(table(mtcars$am, mtcars$cyl)): Chi-squared approximation
#> may be incorrect
#>       Chi2 df Cramers_v   CI Cramers_CI_low Cramers_CI_high          p
#> 1 8.740733  2 0.5226355 0.95       0.180431               1 0.01264661
#>                       Method
#> 1 Pearson's Chi-squared test

^{Created on 2021-10-24 by the reprex package (v2.0.1)}

I am wondering if we should be returning only Ratio parameter for Bayesian contingency table?

JASP also does the same:

[mattansb]

I think it's redundant to return the posterior cell proportions/counts from the BFBayesFactor model. The analysis is intended to estimate and evaluate the dependency - which is captured not by the individual cells, but by the effect size(s).

[mattansb]

(@IndrajeetPatil why do you as.data.frame() everything?? @strengejacke worked so hard on those beautify printing methods!)

[IndrajeetPatil]

Sorry 🙈
As a user, I appreciate the pretty-printing, but as a developer, I want to see everything that exists in the dataframe.

I think it's redundant to return the posterior cell proportions/count

I think so, too! So if we decide to not return those rows, the output will be a one-row output, and it will be consistent with {htest} summary outputs!

[strengejacke]

So from your example, that would be rows 5 and 6 that will be merged into one row, right?

[IndrajeetPatil]

Exactly!

When cramers_v = NULL: only BF values
When TRUE: also frequentist effect sizes in the same row

[mattansb]

_{(The effect size isn't frequentists - it is based on the posteriors!)}

[IndrajeetPatil]

Can we please fix this soon? I haven't been able to work on my package for these two weeks because of this issue.

The GitHub {effectsize} breaks {statsExpressions}, so the next update will also need to inform CRAN about that 🙈

I am really looking forward to the 1.0.0 release 😅

[IndrajeetPatil]

Since the outputs have changed since then and this issue has got too murky, I am closing it in favor of the cleaner #630 issue.