Statistical Visualizations with {ggstatsplot}: A Biography
Indrajeet Patil
Life in the trenches (c. 2017, Harvard)
External Stimulus
- Reporting errors:
“half of all published psychology papers contained at least one p-value that was inconsistent”1
- Interpretation errors:
“in 72% of cases, nonsignificant results were misinterpreted [to mean] that effect was absent”2
- Replication crisis:
“39% of effects were subjectively rated to have replicated the original result”3
and more…
Internal Response
How to:
- avoid reporting errors?
- improve quality of statistical reporting?
- emphasize the importance of the effect?
- interpret null results?
- easily assess validity of model assumptions?
- increase replicability?
A visualization with statistical summary
💡 Visualizations reveal problems not discernible from model summaries!
Ready-made plots with one-line syntax
The grammar of graphics framework can prepare any visualization! But building plots from scratch can be time-consuming.
💡 Using ready-made plots lowers the effort needed for visualizing data!
And there is more!
Show, don’t tell
Without ggstatsplot
Pearson’s correlation test revealed that, across 142 participants, variable x was negatively correlated with variable y: \(t(140)=-0.76, p=0.446\). The effect size \((r=-0.06, 95\% CI [-0.23,0.10])\) was small, as per Cohen’s (1988) conventions. The Bayes Factor for the same analysis revealed that the data were 5.81 times more probable under the null hypothesis as compared to the alternative hypothesis. This can be considered moderate evidence (Jeffreys, 1961) in favor of the null hypothesis (absence of any correlation between x and y).
With ggstatsplot
✅ No need to worry about reporting or interpretation errors!
Thoughtful Defaults
Data Visualization
Statistical Reporting
✅ Follows best practices in data visualization and statistical reporting!
User Love
Total downloads > 500K (97 percentile)
Total citations > 1000
From publications across a wide range of fields:
biology, medicine, psychology, economics, etc.
Collaborative Solutions
While re-architecting ggstatsplot, I started contributing upstream.
As part of easystats core team
- leadership skills to steer the project
- long-term vision for the project
- API design
- CI infrastructure
- code review
- documentation
- scouting for new talent
- developer advocacy
- community engagement
Making it a habit
- co-maintainer of ggsignif
- contributor to WRS2, ggcorrplot
Quality Assurance
“The only way to go fast, is to go well.”
- Robert C. Martin
CI Checks (GitHub Actions)
- Unit tests (random-order)
- Code coverage (100%)
- Linting (0 lints)
- Formatting (0 issues)
- Documentation (website, no link rot, plenty examples)
- Pre-commit hooks (0 issues)
- Zero user-facing warnings
- Portability (Linux, macOS, Windows)
- Robustness (dependencies, language versions)
- CRAN checks (0 notes, 0 warnings, 0 errors)
Healthy and active code base
Communication
Training material on best practices in software/package development to support community contributions keeping in mind the diverse backgrounds of contributors.
ggwithinstats()
Hypothesis about group differences: repeated measures design
ggwithinstats(
data = WRS2::WineTasting,
x = Wine,
y = Taste
)
Important
✏️ Defaults
- raw data + distributions
- descriptive statistics
- inferential statistics
- effect size + uncertainty
- pairwise comparisons
- Bayesian hypothesis-testing
- Bayesian estimation
Statistical approaches available
- parametric
- parametric
- robust
- Bayesian
gghistostats()
Distribution of a numeric variable
gghistostats(
data = movies_long,
x = budget,
test.value = 30
)
Important
✏️ Defaults
- counts + proportion for bins
- descriptive statistics
- inferential statistics
- effect size + uncertainty
- pairwise comparisons
- Bayesian hypothesis-testing
- Bayesian estimation
Statistical approaches available
- parametric
- parametric
- robust
- Bayesian
ggdotplotstats()
Labeled numeric variable
ggdotplotstats(
data = movies_long,
x = budget,
y = genre,
test.value = 30
)
Important
✏️ Defaults
- descriptive statistics
- inferential statistics
- effect size + uncertainty
- pairwise comparisons
- Bayesian hypothesis-testing
- Bayesian estimation
Statistical approaches available
- parametric
- parametric
- robust
- Bayesian
ggscatterstats()
Hypothesis about correlation: Two numeric variables
ggscatterstats(
data = movies_long,
x = budget,
y = rating
)
Important
✏️ Defaults
- joint distribution
- marginal distribution
- effect size + uncertainty
- pairwise comparisons
- Bayesian hypothesis-testing
- Bayesian estimation
Statistical approaches available
- parametric
- parametric
- robust
- Bayesian
ggcorrmat()
Hypothesis about correlation: Multiple numeric variables
ggcorrmat(dplyr::starwars)
Important
✏️ Defaults
- inferential statistics
- effect size + uncertainty
- careful handling of
NAs - partial correlations
Statistical approaches available
- parametric
- parametric
- robust
- Bayesian
ggpiestats()
Hypothesis about composition of categorical variables
ggpiestats(
data = mtcars,
x = am,
y = cyl
)
Important
✏️ Defaults
- descriptive statistics
- inferential statistics
- effect size + uncertainty
- goodness-of-fit tests
- Bayesian hypothesis-testing
- Bayesian estimation
ggbarstats()
Hypothesis about composition of categorical variables
ggbarstats(
data = mtcars,
x = am,
y = cyl
)
Important
✏️ Defaults
- descriptive statistics
- inferential statistics
- effect size + uncertainty
- goodness-of-fit tests
- Bayesian hypothesis-testing
- Bayesian estimation
ggcoefstats()
Hypothesis about regression coefficients
mod <- lm(
formula = rating ~ mpaa,
data = movies_long
)
ggcoefstats(mod)
Important
✏️ Defaults
- estimate + uncertainty
- inferential statistics (\(t\), \(z\), \(F\), \(\chi^2\))
- model fit indices (AIC + BIC)
Supports all regression models supported in {easystats} ecosystem.
Meta-analysis is also supported!
grouped_ functions
grouped_ggpiestats(
data = mtcars,
x = cyl,
grouping.var = am
)
Available grouped_ variants:
grouped_ggbetweenstats()grouped_ggwithinstats()grouped_gghistostats()grouped_ggdotplotstats()grouped_ggscatterstats()grouped_ggcorrmat()grouped_ggpiestats()grouped_ggbarstats()
Modify the look 🎨
By using ggplot2 functions
Too much information 🙈
Alternative: Pure Pain
Hunting for packages
📦 for inferential statistics ({stats})
📦 computing effect size + CIs (effectsize)
📦 for descriptive statistics (skimr)
📦 pairwise comparisons (multcomp)
📦 Bayesian hypothesis testing (BayesFactor)
📦 Bayesian estimation (bayestestR)
📦 …
Inconsistent APIs
🤔 accepts data frame, vector, matrix?
🤔 long/wide format data?
🤔 works with NAs?
🤔 returns data frame, vector, matrix?
🤔 works with tibbles?
🤔 has all necessary details?
🤔 …
Simplified data analysis workflow
✅ Quick insight into data by combining visualization and modeling!
A grain of salt
The “Golem of Prague” problem
❌ Promotes mindless application of statistical tests.
