davidgasquez.com
📚 Personal bits of knowledge
1# Data Visualization
2
3Charts can be more memorable, shareable, and quickly understood than written explanations. They help explore data, explain concepts, and [share information effectively](https://www.scientificdiscovery.dev/p/salonis-guide-to-data-visualization). Clear visuals strengthen [[Communication]] and [[Dashboards]].
4
5## Why Visualize Data
6
7- Visualizing data helps spot patterns, trends, and unusual data points that are hard to see in averages or summaries alone. A chart can reveal what an aggregate hides (e.g: [Anscombe's quartet](https://en.wikipedia.org/wiki/Anscombe%27s_quartet)).
8- Diagrams can explain concepts faster than text. A few-second visual can replace a long, confused explanation.
9- A good chart communicates faster than 1,000 words, but that power comes with responsibility. Misleading charts spread just as easily as accurate ones.
10- Plotting data helps spot potential errors and artefacts before publishing.
11
12## Chart Type Selection
13
14- Pick the chart type and [[Metrics|metric]] that answers the exact question; rates, counts, and shares reveal different truths, so show multiple small views when one cut feels incomplete.
15- Use familiar or practical units (minutes, not standard deviations) when possible. They're easier to interpret and sense-check.
16
17## Clarity
18
19- Keep labels horizontal and close to the data. Direct labels beat legends.
20- Don't include a legend, instead, [label data series directly](https://www.eugenewei.com/blog/2017/11/13/remove-the-legend) in the plot area (usually to the right of the most recent data point). Exception: many categories referring to many elements (e.g., maps).
21- Use small multiples when too many lines overlap. Splitting into panels makes individual trends easier to follow, though it trades off direct comparison between entities.
22- Sort categories logically (inherent order) or alphabetically (easier to skim).
23- [Data looks better naked](https://www.darkhorseanalytics.com/blog/data-looks-better-naked).
24 - Reduce non-data-ink as much as possible without losing communicative power.
25- Don't include more precision than needed.
26- Format axis labels to match the figures being measured (e.g., currency for dollars).
27- Look at axis label spacing and increase intervals if crowded.
28
29## Color
30
31- Match colors to concepts (plants → green, bad → red) so readers aren't forced into a [Stroop test](https://en.wikipedia.org/wiki/Stroop_effect).
32- Use [color-blind friendly palettes](https://davidmathlogic.com/colorblind/). About 4-5% of the population has some form of color blindness.
33- Direct labeling also helps color-blind readers distinguish categories.
34
35## Axes
36
37- Start your y-axis at zero (assuming no negative values).
38- Avoid deceptive scale tricks.
39 - Leave breathing room on axes instead of extreme zoom.
40 - The lowest point shouldn't appear to be the lowest possible value.
41- Pair relative effects with absolute numbers (or prediction intervals instead of confidence intervals) to show real-world risk.
42
43## Context
44
45- Include explanations for anomalous events directly on the graph.
46- For unfamiliar chart types, guide readers with annotations. Add a mini-tutorial if needed.
47- Include targets as asymptotes to help audiences see if you're on track.
48- Make the chart standalone. Add purpose, units, timeframe, and source so it can travel without losing meaning and slot into [[Dashboards]] or memos without extra explanation.
49- Titles for graphs should be the conclusion or key takeaway.
50- Always note the data source below the graph.
51
52## Reproducibility
53
54- Publish provenance with the chart (data source, assumptions, and ideally a link to code) so others can verify or reuse it and keep [[Data Practices]] consistent.
55- A chart with no source isn't much better than claiming a trend was revealed in a dream.
56
57## Common Pitfalls
58
59- Skip arrows or other glyphs that imply trends you can't support.
60- Don't use 3D charts. They distract and make values harder to read.
61- Avoid confidence intervals when showing variability. They're often [misinterpreted as ranges](https://www.dangoldstein.com/papers/Hofman_Goldstein_Hullman_Visualizing_Uncertainty_Mislead_Scientific.pdf). Consider prediction intervals or underlying percentages instead.
62- Try not to have too many data series; 5-8 is the usual limit depending on clustering.
63
64## Guiding Questions
65
66Ask yourself when creating a visualization:
67
681. Is my chart type meaningful for the question?
692. Can I make it clearer?
703. If complicated, can I guide the viewer through it?
714. Does the chart work as a standalone?
725. Is my chart's presentation justifiable?
736. Is my chart reproducible?
74
75## Tools
76
77- [Datawrapper](https://www.datawrapper.de/). Quick interactive charts with great defaults.
78- [Raw Graphs](https://app.rawgraphs.io/). Open-source, unusual chart types.
79- [Observable Plot](https://observablehq.com/plot/). JavaScript-based exploratory charts.
80- [Kepler](https://kepler.gl/). Geospatial visualization.
81
82## Resources
83
84- [Saloni's Guide to Data Visualization](https://www.scientificdiscovery.dev/p/salonis-guide-to-data-visualization)
85- [The Data Visualization Catalogue](https://datavizcatalogue.com/) and [Project](https://datavizproject.com/)
86- [Visualization Curriculum](https://jjallaire.github.io/visualization-curriculum/)
87- [Guides for Visualizing Reality](https://flowingdata.com/2020/06/01/guides-for-visualizing-reality/)
88- [Datawrapper's Do's and Don'ts](https://www.datawrapper.de/blog/category/datavis-dos-and-donts)
89- [The Science of Visual Data Communication: What Works](https://journals.sagepub.com/doi/10.1177/15291006211051956)