Bio Data Visualization Distribution Plots

Version Compatibility

Reference examples tested with: ggplot2 3.5+, ggbeeswarm 0.7+, ggdist 3.3+, gghalves 0.1.4+, seaborn 0.13+, matplotlib 3.8+, ptitprince 0.3+ (Python raincloud).

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• R: packageVersion('<pkg>') then ?function_name to verify parameters

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Distribution Plots

"Plot the distribution per group" -> Render boxplot, violin, beeswarm, or raincloud calibrated to N per group, the underlying distribution shape, and the audience's ability to read each encoding. The default geom_bar(stat='summary') is the canonical misleading choice — Weissgerber 2015 PLOS Biol documented that 703 top physiology papers use bar-of-mean despite multiple distinct distributions producing identical bars.

• R: ggplot2::geom_boxplot, ggplot2::geom_violin, ggbeeswarm::geom_quasirandom, ggdist::stat_halfeye, gghalves::geom_half_violin
• Python: seaborn.boxplot/violinplot/swarmplot/stripplot, ptitprince.RainCloud

The Single Most Important Modern Insight -- Bars of Means Lie

Weissgerber, Milic, Winham & Garovic 2015 PLOS Biol 13:e1002128 surveyed 703 papers in top physiology journals and found that bar-and-line graphs of means dominate, despite many distinct distributions producing identical bar plots. Bimodal data, skewed data, and data with outliers all collapse to the same bar height and error bar. The bar plot is a hypothesis test result rendered as visualization; the visualization should show the data.

The modern alternative is to show every point for n < 30, layer summary on top, and reserve summary-only plots for large N where points would overplot.

Decision Tree by N per Group

Always annotate N somewhere on the plot (caption, x-axis tick label, or stratum count).

Box, Violin, Beeswarm, Raincloud -- The Four Standard Encodings

Boxplot (Tukey 1977) -- summary only

ggplot(df, aes(group, value, fill = group)) +
    geom_boxplot(outlier.shape = NA, alpha = 0.7, width = 0.5) +
    geom_jitter(width = 0.2, alpha = 0.5, size = 1) +
    scale_fill_manual(values = c('#0072B2', '#D55E00')) +
    labs(x = NULL, y = 'Expression') +
    theme_classic()

Box shows: median, IQR, 1.5×IQR whiskers, outliers. Hides: bimodality, sample size, density.

Notched boxplot (notch = TRUE): notches show 95% CI for median (±1.58·IQR/√n); non-overlapping notches roughly indicate distinct medians. Use with N ≥ 15.

Violin -- density + summary

ggplot(df, aes(group, value, fill = group)) +
    geom_violin(alpha = 0.7, trim = FALSE,
                bw = 'SJ') +                            # Sheather-Jones bandwidth
    geom_boxplot(width = 0.1, fill = 'white', outlier.shape = NA) +
    scale_fill_manual(values = c('#0072B2', '#D55E00'))

KDE bandwidth pitfall: ggplot's default is Silverman's rule of thumb, which oversmooths bimodal data into a single mode. Use bw = 'SJ' (Sheather-Jones plug-in) for honest representation of multimodality.

trim = TRUE (default) cuts the violin at the data range — visually misleading because the violin's tails imply density extending beyond the data. trim = FALSE lets the KDE extend.

Beeswarm / quasirandom -- every point shown deterministically

library(ggbeeswarm)
ggplot(df, aes(group, value, color = group)) +
    geom_quasirandom(method = 'quasirandom', width = 0.3, alpha = 0.7) +
    scale_color_manual(values = c('#0072B2', '#D55E00')) +
    stat_summary(fun = median, geom = 'crossbar', width = 0.5, color = 'black')

Quasirandom (van der Corput sequence; Bostock implementation) gives reproducible jitter that fills space without random scatter. Beeswarm is similar but with collision avoidance. Both are deterministic — reruns produce identical layouts.

Raincloud (Allen 2019) -- distribution + summary + raw

Goal: Show distribution (half-violin), summary (boxplot), and raw observations (jittered points) in a single per-group panel without occlusion.

Approach: Place a half-violin on one side, a thin boxplot in the middle, and jittered points on the other side via gghalves::geom_half_violin + geom_boxplot + geom_half_point with position_nudge offsets; flip to horizontal so the visual reads as a literal raincloud.

library(gghalves)
ggplot(df, aes(group, value, fill = group, color = group)) +
    geom_half_violin(side = 'r', alpha = 0.7, position = position_nudge(x = 0.15)) +
    geom_boxplot(width = 0.15, outlier.shape = NA, alpha = 0.7,
                 position = position_nudge(x = -0.05)) +
    geom_half_point(side = 'l', alpha = 0.5, size = 1.5, range_scale = 0.4,
                    position = position_nudge(x = -0.2)) +
    scale_fill_manual(values = c('#0072B2', '#D55E00')) +
    scale_color_manual(values = c('#0072B2', '#D55E00')) +
    coord_flip()                                          # horizontal "raincloud"

import ptitprince as pt
import seaborn as sns
pt.RainCloud(x='group', y='value', data=df,
             palette=['#0072B2', '#D55E00'],
             bw='scott', cut=0,                          # bandwidth + trim
             width_viol=0.6, orient='h')

Raincloud = half-violin (distribution) + boxplot (summary) + jittered raw points. Allen 2019 Wellcome Open Res 4:63 — modern publication default for N 30-200.

Letter-value plot (Hofmann-Wickham 2017)

library(lvplot)
ggplot(df, aes(group, value, fill = group)) +
    geom_lv(k = 5, alpha = 0.7) +
    scale_fill_manual(values = c('#0072B2', '#D55E00'))

Extends Tukey's boxplot via additional letter-value quantiles (Hofmann, Wickham, Kafadar 2017 J Comput Graph Stat 26:469). For large N, the standard boxplot collapses tail structure; letter-value preserves it.

sns.boxenplot(x='group', y='value', data=df,
              palette=['#0072B2', '#D55E00'])             # seaborn calls it boxenplot

Stacked / split violin (paired comparisons)

library(introdataviz)               # split-violin geom
ggplot(df, aes(group, value, fill = condition)) +
    geom_split_violin(alpha = 0.7) +
    geom_boxplot(width = 0.15, position = position_dodge(0.5), outlier.shape = NA)

For 2-condition comparison within each group, split-violin shows both densities back-to-back. More compact than dodged violins.

Per-Method Failure Modes

Bar of mean with SEM

Trigger: geom_bar(stat = 'summary') + geom_errorbar(stat = 'summary', fun.data = mean_se).

Mechanism: Mean ± SEM collapses all distributional information; reader cannot assess bimodality, skew, or N.

Symptom: Reviewer asks to "show the data"; the figure must be redone.

Fix: Replace with raincloud, beeswarm, or boxplot+jitter. Show points for N < 30.

Violin with default Silverman bandwidth oversmooths bimodality

Trigger: geom_violin() without specifying bw.

Mechanism: Silverman's rule of thumb assumes unimodal Gaussian; oversmooths bimodal data into a single peak.

Symptom: Single-cell expression bimodality (off / on) renders as a unimodal violin; biologically false.

Fix: bw = 'SJ' (Sheather-Jones plug-in) for honest bimodality. Note: nrd0 IS Silverman; nrd (Scott) oversmooths less than Silverman but Sheather-Jones is preferred.

Notched boxplot with too-small N

Trigger: notch = TRUE with N < 15 per group.

Mechanism: Notch can extend beyond Q1/Q3, producing visually-misleading "inside-out" notches.

Symptom: ggplot warning ("notch went outside hinges"); notches look weird.

Fix: Use notches only with N ≥ 15. For smaller N, show raw points instead.

Boxplot hides outliers when jittered points are overlaid

Trigger: geom_boxplot() + geom_jitter() with default outlier.shape = 19.

Mechanism: Outliers render twice — once from boxplot (large dots), once from jitter (smaller dots) — visually duplicated.

Symptom: Some points appear bigger than others without reason.

Fix: geom_boxplot(outlier.shape = NA) when overlaying raw points.