Friends Don't Let Friends Make Bad Graphs: A Data Visualization Guide

Data visualization is a critical skill for researchers, analysts, and anyone who communicates with data. Yet common mistakes persist across publications, presentations, and dashboards. This guide, inspired by Chenxin Li's popular resource "Friends Don't Let Friends Make Bad Graphs," examines 14 common visualization pitfalls and provides actionable guidance for creating effective, honest, and accessible graphics.

The Core Philosophy

Good data visualization should:

1. Accurately represent the data without distortion
2. Highlight meaningful patterns while acknowledging uncertainty
3. Be accessible to diverse audiences including colorblind viewers
4. Match the visualization type to the data structure and research question

Let's examine specific pitfalls and their solutions.

Pitfall 1: Bar Plots for Distributions

The Problem

Bar plots with error bars show only the mean and standard error, hiding crucial information about the underlying distribution:

• Sample size
• Data distribution shape (bimodal, skewed, etc.)
• Individual data points
• Outliers

Why It Matters

Two datasets can have identical means and standard errors while having completely different distributions. Bar plots mask this critical information, potentially leading to incorrect conclusions.

The Solution

Use visualization methods that show the full data distribution:

Recommendation: Combine approaches—overlay individual points on box plots to show both summary statistics and raw data.

Pitfall 2: Violin Plots Without Internal Structure

The Problem

Violin plots show distribution shape but can hide important details:

• Sample size differences between groups
• Actual data points
• Bimodal vs. uniform distributions can look similar

The Solution

Always pair violin plots with:

• Internal box plots: Show median and quartiles
• Jittered points: Show actual observations
• Sample size annotations: Make n explicit

ggplot(data, aes(x = group, y = value)) +
  geom_violin() +
  geom_boxplot(width = 0.1) +
  geom_jitter(width = 0.1, alpha = 0.3)

Pitfall 3: Wrong Color Scales

The Problem

Choosing the wrong color scale type obscures patterns or creates false ones:

• Sequential scales (light→dark): For data with a natural ordering from low to high
• Diverging scales (color→white→color): For data with a meaningful midpoint

Using sequential scales for diverging data (or vice versa) misrepresents the underlying patterns.

Color Scale Selection Guide

The Solution

1. Identify whether your data has a meaningful midpoint
2. Choose the appropriate scale type
3. Center diverging scales on the meaningful midpoint (often 0)

Pitfall 4: Unordered Heatmaps

The Problem

Heatmaps with arbitrary row/column ordering hide patterns that become obvious when properly organized. The human eye struggles to identify clusters and correlations in randomly ordered matrices.

The Solution

Apply hierarchical clustering or other ordering algorithms:

# Using pheatmap for automatic clustering
pheatmap(data_matrix,
         cluster_rows = TRUE,
         cluster_cols = TRUE,
         clustering_method = "complete")

Ordering Strategies

Pitfall 5: Pie Charts for Comparison

The Problem

Pie charts require viewers to compare angles and areas—tasks humans perform poorly. Common issues:

• Difficult to compare non-adjacent slices
• Small differences are invisible
• More than 5-6 categories become unreadable
• 3D effects make interpretation even harder

When Pie Charts Work

• Showing parts of a whole (exactly 100%)
• Only 2-3 categories
• Differences are large (>10% between slices)
• Exact values are labeled

Better Alternatives

Pitfall 6: Ignoring Colorblind Accessibility

The Problem

Approximately 8% of men and 0.5% of women have some form of color vision deficiency. Red-green colorblindness (deuteranopia/protanopia) is most common, making standard red-green palettes inaccessible.

Problematic Combinations

• Red vs. Green
• Green vs. Brown
• Blue vs. Purple
• Light green vs. Yellow

Colorblind-Safe Palettes

Additional Strategies

1. Use shapes in addition to colors: Different point shapes remain distinguishable
2. Add patterns to fills: Hatching, dots, etc.
3. Direct labeling: Label lines/points directly instead of using legends
4. Test your plots: Use colorblindness simulators

Pitfall 7: Dual Y-Axes

The Problem

Dual y-axes can:

• Create false correlations through arbitrary scaling
• Mislead viewers about relationships
• Make accurate comparison impossible

The same data can show positive correlation, negative correlation, or no correlation depending on how the axes are scaled.

Better Alternatives

Pitfall 8: Arbitrary Histogram Bins

The Problem

Histogram appearance depends heavily on bin width:

• Too few bins: Hide distribution features
• Too many bins: Create noise
• Arbitrary boundaries: Can hide or create patterns

The Solution

1. Try multiple bin widths: Ensure conclusions are robust
2. Use principled methods: Freedman-Diaconis, Sturges, Scott's rule
3. Consider kernel density: Smooth alternative to histograms
4. Report sensitivity: Note if patterns depend on binning

Pitfall 9: Omitting Zero on Bar Charts

The Problem

Bar charts encode values as lengths. Truncating the y-axis:

• Exaggerates small differences
• Misleads about relative magnitudes
• Violates the principle that length should be proportional to value

When Truncation Is Acceptable

• Line charts (showing trends, not absolute values)
• When the baseline is explicitly noted
• Scientific contexts where zero is meaningless

Best Practice

For bar charts: always start at zero or use a different visualization type.

Pitfall 10: 3D Charts

The Problem

3D effects on 2D displays:

• Distort relative sizes (perspective effects)
• Occlude data points
• Add visual noise without information
• Make precise reading impossible

The Only Exception

Actual 3D data (x, y, z coordinates) may benefit from 3D visualization, but even then:

• Interactive rotation is essential
• Multiple 2D projections often work better
• Consider contour plots as alternatives

Pitfall 11: Connected Scatter Plots Without Time

The Problem

Lines between points imply:

• Continuity
• Temporal or ordinal sequence
• Meaningful interpolation

Connecting unordered points creates misleading visual patterns.

When to Connect Points

Pitfall 12: Overplotting

The Problem

When many points overlap:

• True density is hidden
• Outliers obscure or are obscured
• Patterns become invisible

Solutions by Data Size

Pitfall 13: Rainbow Color Scales

The Problem

Rainbow (jet) color scales:

• Are not perceptually uniform
• Create false boundaries
• Are colorblind-unfriendly
• Distort data interpretation

Better Alternatives

• Viridis family: Perceptually uniform, colorblind-safe
• Cubehelix: Monotonic lightness change
• Scientific color maps: Designed for specific data types

Pitfall 14: Missing Uncertainty

The Problem

Point estimates without uncertainty measures:

• Overstate precision
• Hide variability
• Prevent proper interpretation

What to Show

Always specify which measure you're displaying.

Quick Reference Checklist

Before finalizing any visualization:

• Does the chart type match the data and question?
• Is the full distribution visible (not just summary statistics)?
• Are colors colorblind-accessible?
• Is the color scale type (sequential/diverging) appropriate?
• Are categories meaningfully ordered?
• Does the y-axis start at zero (for bar charts)?
• Is uncertainty shown where appropriate?
• Are there any 3D effects that should be removed?
• Would a different chart type tell the story better?

Recommended Tools

R Packages

• ggplot2: Grammar of graphics
• viridis: Colorblind-safe palettes
• patchwork: Combining multiple plots
• ggridges: Ridge plots for distributions

Python Libraries

• matplotlib: Base plotting
• seaborn: Statistical visualization
• plotnine: ggplot2 for Python
• altair: Declarative visualization

Colorblindness Testing

• Coblis: Color blindness simulator
• Viz Palette: Interactive palette builder
• ColorBrewer: Pre-tested palettes

Conclusion

Good data visualization is both an art and a science. The principles outlined here—showing full distributions, using appropriate color scales, ensuring accessibility, and matching visualization type to data structure—will help you create graphics that communicate honestly and effectively.

Remember: the goal is not to make data look impressive, but to reveal truth. When in doubt, choose clarity over cleverness, and always consider your audience.

Reference: Li, Chenxin. "Friends Don't Let Friends Make Bad Graphs." GitHub Repository. https://github.com/cxli233/FriendsDontLetFriends

• Slide: 1103_Vis.pdf