When a nominal vector has many levels with one observation per level, calling report() on a Kruskal-Wallis test would fail with a data frame construction error because the confidence interval calculation for the effect size would fail when all bootstrapped values are identical.

Problem

The original issue occurred when:

1. Creating a factor with many levels (e.g., as.factor(1:n))
2. Each level has only one observation
3. Running kruskal.test() works fine and is fast
4. But report() on the result would fail with: "arguments imply differing number of rows: 1, 0"

This happened because effectsize::rank_epsilon_squared() uses bootstrap methods to compute confidence intervals, but when there's one observation per group, all bootstrapped effect sizes equal 1, making CI calculation impossible.

Solution

Added error handling in .report_effectsize_kruskal() to:

• Catch CI calculation failures due to degenerate cases
• Fallback to ci = NULL when bootstrap CI fails
• Report the effect size without confidence intervals in these edge cases
• Maintain backward compatibility for normal cases

Before and After

Before (fails):

library("report")
n <- 10
set.seed(123)
df <- data.frame(a = as.factor(1:n), b = rnorm(n))
test <- kruskal.test(df$b, df$a)
report(test)  # Error: arguments imply differing number of rows: 1, 0

After (works):

library("report")
n <- 10
set.seed(123)
df <- data.frame(a = as.factor(1:n), b = rnorm(n))
test <- kruskal.test(df$b, df$a)
result <- report(test, data = df)
print(result)
#> [1] "All values of t are equal to  1 \n Cannot calculate confidence intervals"
#> Effect sizes were labelled following Field's (2013) recommendations.
#> 
#> The Kruskal-Wallis rank sum test testing the difference in ranks between df$b
#> and df$a suggests that the effect is statistically not significant, and large
#> (Kruskal-Wallis chi2 = 9.00, p = 0.437; Epsilon squared (rank) = 1.00)

# Performance test with larger n
n <- 100
set.seed(456)  
df_large <- data.frame(a = as.factor(1:n), b = rnorm(n))
test_large <- kruskal.test(df_large$b, df_large$a)

# This now completes quickly instead of taking a long time
system.time({
    result_large <- report(test_large, data = df_large, verbose = FALSE)
})
#> [1] "All values of t are equal to  1 \n Cannot calculate confidence intervals"
#>    user  system elapsed 
#>   0.161   0.002   0.164

Key Improvements

• ✅ Performance: Degenerate cases now complete in <1 second vs. failing before
• ✅ Reliability: No more crashes with data frame construction errors
• ✅ Graceful degradation: Effect size still reported even when CI can't be calculated
• ✅ Backward compatibility: Normal cases continue to work with CI as before

Note: For optimal results, users should provide data manually report(test, data = df) or use the formula interface kruskal.test(y ~ group, data = df) as suggested by existing warning messages in the report package.

Fixes #454.

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.