The HistData Challenge: Re-Visioning Historical Graphics

Challenges

This idea first took shape the Minard Challenge that stemmed from my paper Friendly (2002), “Visions and Re-Visions of Charles Joseph Minard”. This reviewed decades of attempts – by many different people, using many different tools – to reproduce and extend Minard’s famous graphic of Napoleon’s Russian campaign. That paper, and the gallery of re-visions that grew out of it, is the model for this vignette: a catalog of kinds of re-vision, applied not just to Minard but to any dataset in this package, plus a running inventory of what’s already been attempted (in sandbox/, by outside contributors, in companion projects) and what’s still open.

I should mention that the idea of a data visualization “Challenge” is an important one in this age of online communication. Typically, they present a dataset, or a content theme, or just an image, and invite participants to participate; sometimes prizes are awarded.

Some examples illustrate the range of possibilities from such endeavors:

• ASA Statistical Graphics Expositions: starting in the early 1980s – the well-known 1983 Data Exposition (Donoho & Ramos’s cars/mpg dataset) is often cited as the first, though a 1982 precursor is also mentioned in some accounts – the Section on Statistical Graphics of the American Statistical Association began a series of poster/paper sessions at the Joint Statistical Meetings organized around analysis of a single curated dataset, intended as an opportunity to share perspectives, approaches to the problem, and highlight developments in software. Some notable examples are the HistData::Pollen data contained here and the baseball salary/performance data from the 1988 Exposition (based on the 1986 season), now shipped in the vcd package as vcd::Baseball.

• Hotelling Society: I was teaching a graduate course, Psychology 6140: Multivariate Data Analysis at the time, and I had the idea to use the baseball data again to challenge my students to “strut their stuff” and show off what they had learned in the course in a mock-conference setting I called the “Society for Exploration of Multivariate Data”, or, in honor of Harold Hotelling, the “Hotelling Society”. I used the Baseball dataset for this, with data provided in SAS and SPSS format.

  But students were encouraged to extend the scope with other baseball-related datasets or questions. Students dressed the part, a “7th inning stretch” included beer and pizza, and a fictitious guest speaker, Izzy A. Ringer, presented conclusive proof that Shoeless Joe Jackson did NOT throw the 1919 World Series – the Black Sox Scandal, in which Jackson and seven Chicago White Sox teammates were accused of intentionally losing to the Cincinnati Reds. The upshot was that we all had fun with this, and it contributed to understanding of use of methods of the course and conveying results for a conference presentation.

  The baseball data itself outlived the course: working with it for the Hotelling Society is what led, eventually, to developing the Lahman R package, which packages Sean Lahman’s baseball database directly for R users – itself a small case of one challenge (teach a course, make the data usable) turning into a durable, reusable piece of data for baseball research. This is the same trajectory HistData itself follows on a larger scale.

• Kaggle Competitions: machine-learning contests run on the Kaggle data science platform, from beginner “Getting Started” competitions like Titanic: Machine Learning from Disaster (predict who survived, from passenger data) to prize-money contests sponsored by companies and research groups. Not about historical re-visioning specifically, but the same competitive/participatory spirit – a public dataset, a shared task, and a leaderboard of attempts to compare against.

• MakeoverMonday: probably the closest in spirit to this vignette’s own idea of re-visioning. Started in 2016 by Andy Kriebel and Eva Murray, a new dataset – almost always one already published somewhere as a chart, often a mediocre one – is posted every Sunday, and participants are asked “to work with a given data set and create better, more effective visualizations and help us make information more accessible.” Where 30DayChartChallenge and TidyTuesday hand you a bare dataset, MakeoverMonday hands you an existing graphic and asks you to improve on it – the same path this vignette takes with Minard, Perozzo, and Snow, just aimed at recent charts instead of historical ones.

• 30DayChartChallenge: one chart a day for the month of April, each day keyed to a themed prompt (e.g. “part-to-whole,” “historical,” “3D”). Started in 2021 by Cédric Scherer and Dominic Royé as a follow-on to the #30DayMapChallenge. John Russell’s recreations of several HistData datasets, used throughout this vignette, came from the 2025 edition – see the sandbox/ inventory below.

• TidyTuesday: a weekly data project run by the R4DS (“R for Data Science”) online learning community since 2018 – a new real-world dataset posted every Monday, participants share their own wrangling/visualization of it. Less about re-visioning one specific historical graphic and more about practicing data skills on whatever dataset comes up that week.

• #DuBoisChallenge: launched in 2021 by Allen Hillery, Anthony Starks, and Sekou Tyler specifically to recreate W. E. B. Du Bois’s hand-drawn data visualizations from the 1900 Paris Exposition using modern tools. Closest in spirit to this vignette’s own framing of all the challenges above – explicitly about revisiting one historical body of graphic work, not a general prompt-of-the-day format.

A tiny motivating example: Arbuthnot

Before anything as ambitious as tackling Minard or John Snow, a simple example makes a small point of statistical historiography concretely. John Arbuthnot’s 1710 table of annual christenings in London (1629-1710) looks, on a first read, like a straightforward count. But Sandy Zabell (1976) found that the total number of christenings (Males + Females) for 1674 and 1704 are identical – not approximately close, identical, down to the last birth – strongly suggesting one year’s entry was copied from the other, whether by Arbuthnot himself or a later transcriber.

data(Arbuthnot)
Arbuthnot$Christenings <- Arbuthnot$Males + Arbuthnot$Females
Arbuthnot[Arbuthnot$Year %in% c(1674, 1704),
          c("Year", "Males", "Females", "Christenings", "Ratio")]

Males, Females, Christenings, and the derived Ratio all match exactly; only Plague and Mortality (not shown) differ. A table alone makes this easy to miss among 82 rows. Plotting Christenings by Year makes it much harder to miss: the series climbs steadily over these 80 years, nearly doubling, so a value perfectly ordinary for the 1670s becomes a jarring plunge when it recurs, unchanged, thirty years later. The horizontal line marks the shared level. And, following Zabell & Wainer’s Figure 2 (2002), shows it passing through only these two points:

copy_level <- Arbuthnot$Christenings[Arbuthnot$Year == 1704]

ggplot(Arbuthnot, aes(x = Year, y = Christenings)) +
  geom_line(color = "grey50") +
  geom_point(color = "grey50") +
  geom_hline(yintercept = copy_level, linetype = "dotted", color = "black") +
  annotate(
    "segment", x = 1677, xend = 1701, y = copy_level, yend = copy_level,
    color = "darkred", linewidth = 1.3,
    arrow = arrow(length = unit(0.3, "cm"), type = "closed")
  ) +
  geom_point(
    data = subset(Arbuthnot, Year %in% c(1674, 1704)),
    color = "firebrick", size = 4
  ) +
  labs(
    title = "Arbuthnot's annual christenings in London, 1629-1710",
    subtitle = "1674 and 1704 (red) are identical -- likely a copying error",
    y = "Christenings (Males + Females)"
  ) +
  theme_minimal()

Christenings, not the sex Ratio (Males/Females), is the quantity that makes the anomaly visible, even though Ratio also matches exactly between the two years. Since Males and Females were copied together, their ratio just reproduces 1674’s perfectly ordinary value – nothing in the Ratio series flags 1704 as unusual. (The one genuinely striking Ratio value in this stretch, 1703’s record low of 1.011, is unrelated noise, no part of this story.) It’s Christenings, because it climbs steadily across the whole span, that turns a copying error into a visible outlier – exactly the quantity Zabell & Wainer’s Figure 1 plots (2002).

That’s re-visioning in miniature: looking again at old data, graphically, shows something tabulation alone hides. Zabell & Wainer (2002) state the scale of the anomaly plainly: 1704 has about 4,000 fewer christenings than its neighbors predict, and the correct figure – 15,895 (8,153 boys, 7,742 girls) – sits comfortably between 1703 and 1705, exactly where a corrected value should. Arbuthnot, they argue, would surely have caught and fixed so glaring an error had he ever plotted his own data: graphed, it stands out “literally, like a sore thumb.”

That he apparently never did is itself a historiographic finding, not a personal failing: it’s evidence that graphing data simply wasn’t yet part of a statistician’s toolbox in 1710, before Playfair invented the line graph in his Commercial and Political Atlas (1786). Re-visioning the graph doesn’t just fix a 300-year-old data error; it exposes a working assumption of Arbuthnot’s era that no longer holds today.

See Zabell (1976) for the original discovery and an earlier version of this graph, Zabell & Wainer (2002) for the fuller published account excerpted above, and Wainer & Spence (2005) for a further retelling.

The graph above could be improved to tell a richer historical story – 1704 is far from the only irregularity in it, just the only one without an obvious explanation. Zabell & Wainer’s own Figure 1 (2002) labels several other anomalies directly on the plot:

• 1643-1660: Christenings falls from about 10,400 in 1642 to a low of 5,612 in 1650 – nearly halved – coinciding with the English Civil War (1642-1651) and the execution of Charles I in 1649. Timing alone doesn’t fully explain it, though: the war ended in 1651, yet the series doesn’t climb back past 10,000 until 1662, over a decade into the Interregnum and Restoration. Zabell & Wainer note that Graunt himself offered a more complex account: Anglican ministers, in this politically fraught stretch, likely declined to register children born to Catholics or Protestant dissenters. Civil war explains the drop’s onset better than its decade-long persistence.
• 1665-1666: a sharper, shorter dip – from 11,722 in 1664 down to 9,972 in 1665 and 8,997 in 1666, before rebounding to 10,938 in 1667 – lines up with the Great Plague of London (1665, which killed roughly a fifth of the city’s population) immediately followed by the Great Fire of London (September 1666). Unlike the Civil War’s decade of depressed numbers, this one snaps back within a year or two once the immediate crises passed.

Neither needed a Zabell-style detective plot to explain – they’re visible, dateable historical events, which is exactly why Zabell & Wainer’s Figure 1 labels them directly rather than treating them as mysteries. 1704 stood out precisely because it didn’t have an obvious explanation like these – until the plot revealed it wasn’t a historical event at all, just a clerical error.

What is re-visioning?

Historians of statistics have practiced “re-visioning” long before this vignette put a name to it – going back to an original historical graphic or dataset itself, not just its published conclusions, to understand it on its own terms, in its own time. Some examples here follow the same pattern as Arbuthnot above, at a larger scale.

Stephen Stigler does this throughout The History of Statistics (1986) and in papers devoted specifically to it: “Francis Galton’s Account of the Invention of Correlation” (1989) re-examines Galton’s own scatter diagram of parent/child heights to work out how Galton actually arrived at correlation, and “Regression Towards the Mean, Historically Considered” (1997) does the same for Galton’s geometric argument for regression – reconstructing the reasoning behind a graphic, not just restating its result.

Stigler has also made the case for why this kind of return-to-the-source work matters, not just demonstrated it. “Data Have a Limited Shelf Life” (2019) traces three textbook-famous datasets – Quetelet’s Scottish soldiers’ chest measurements, von Bortkiewicz’s Prussian horse-kick deaths, and the Cushny-Peebles sleep data behind Gosset’s first published t-test – back to their original sources, and finds each one, after decades of uncritical reuse, had drifted away from the messy, selective, error-ridden reality of its collection into something merely “ornamental”: a fixed textbook prop, its provenance and caveats long forgotten (Gosset’s own data, for instance, reached Fisher’s textbook with two columns mislabeled as the wrong drugs entirely, an error nobody caught for 27 years). Old, dead data aren’t worthless, he’s careful to add – like museum pieces, their value is just a different kind from a live dataset still doing scientific work – but he warns that “big data” is no less susceptible to going “ornamental,” and may be harder to catch: “dead data do not represent the worst case; zombie data may be even worse for scientific use.” Re-visioning, in this vignette’s sense, is exactly the antidote he’s calling for: asking where a dataset actually came from is what keeps it from going ornamental in the first place.

David Bellhouse (with Laura Murray) does something similar for a much earlier and more painstaking case: Murray & Bellhouse (2017) work out how Edmond Halley actually constructed his 1701 map of magnetic declination – a map, not a table – using only the methods and instruments available to Halley at the time.

James Hanley (McGill) has made a research program out of exactly this kind of return-to-the-source work, publishing detailed re-analyses and re-created datasets that go well beyond what the original publications contained. Two examples tie directly into HistData:

• Hanley, Julien & Moodie (2008), “Student’s z, t, and s: What if Gosset had R?,” traces the height and finger-length data originally gathered by Macdonell (1901) and used by “Student” (Gosset, 1908) to introduce the t-distribution – the same data shipped here as HistData::Macdonell.
• Hanley’s re-enactment of Gosset’s yeast-cell counting, following the method described in Gosset’s 1907 Biometrika article, connects directly to HistData::Yeast – also the subject of John Russell’s sandbox/Yeast.R recreation, picked up again in the case studies below.

Digital history

Martin Grandjean’s “The Digital Historian’s Craft” (2026) makes the case at the level of the whole discipline: digital methods are new tools in the historian’s kit, to be appreciated for what they bring to telling the story. Using them well remains a modern, personal craft at the crossroads of technical mastery and the older skills of source criticism, interpretation, and synthesis. Re-visioning, as practiced in this vignette, is one small corner of that same craft.

1. Software/language ports

The same graphic, redrawn with different tools, each an opportunity and challenge to express the drawing in software. The Minard Gallery included examples using:

• Mathematica,
• a formal Grammar-of-Graphics specification (see below), ggplot2,
• SAS/IML (Rick Wicklin’s version – see his own answer to What was he thinking? in “What They Were Thinking”),
• Protovis, by Mike Bostock, superseded by D3.js, whose modern home is Observable).

The content doesn’t change; what changes is what the choice of tool reveals about the graphic’s underlying structure and how you think about doing it in that language environment. A grammar-of-graphics implementation, for instance, forces you to name the mapping from data to mapping into visual channels explicitly, in a way the original hand-drawn version never had to.

2. Design variations

Same content, same tool family, different design decisions: legibility fixes, a graphic designer’s redraw, an alternate visual metaphor (Minard’s gallery includes a pictograph version emphasizing human cost over cartographic accuracy). This is the category where taste and communication goals do the most work.

3. Interactive/web-based

Tooltips, drill-down, linked views, map overlays: these are re-visions that trade the fixed, printed page for something a reader can query, interact with and explore. The Minard gallery includes an image map with linked text and a drill-down SAS GMAP; this is also illustrated for the plotly version of the Perozzo data described below.

4. Temporal/animated

A step-by-step reveal, a third axis given over to time, a narrated video. Minard’s original already compresses time into distance along the march route; an animated re-vision instead lets time play out, trading the compression for a different kind of legibility. (See “Open invitations” below for the deferred gganimate idea for Perozzo.)

A concrete example: Ron Kennett’s “moving bubble” animated re-vision, which replaces Minard’s static flow band with a single traveling bubble sized to the army’s current strength, sweeping along the march route as time advances. (Image on file, Kennet-moving-bubble.gif; embedding pending the general image-hosting/licensing decision noted above.)

5. Alternative representations

A genuinely different chart type for the same underlying data: not a redraw, a rethink– “How could I do this better?”. The Snow-density.R case study below (density contours instead of point symbols for cholera deaths) is this category in a small example.

6. Data packaged for reuse

The original figures plus the underlying data, released in a form someone else can build on. This is, quite literally, what the HistData package itself does for all 34 of its datasets – arguably the least glamorous but most consequential category, since every other kind of re-vision in this list depends on having data to play with first.

A software/language port, built from Minard.cities/Minard.troops/Minard.temp

As noted in “Software/language ports” above, the march itself decomposes into Minard.cities (places), Minard.troops (the flow band – each row a vertex, survivors giving the band’s width, direction its color, group which sub-path it belongs to), and Minard.temp (the temperature series, not used in this panel). Porting to ggplot2 is then mostly a matter of mapping survivors to linewidth and direction to colour:

data(Minard.cities); data(Minard.troops)

ggplot(Minard.troops, aes(x = long, y = lat)) +
  geom_path(aes(linewidth = survivors, colour = direction, group = group),
            lineend = "round") +
  scale_linewidth(range = c(0.5, 15), guide = "none") +
  scale_colour_manual(values = c(A = "grey50", R = "firebrick"), guide = "none") +
  geom_point(data = Minard.cities, size = 1.2) +
  geom_text(data = Minard.cities, aes(label = city), vjust = 1.5, hjust = 0.5, size = 3) +
  labs(title = "Napoleon's Russian campaign of 1812", x = NULL, y = NULL) +
  theme_minimal() +
  theme(axis.text = element_blank(), panel.grid = element_blank())

Grey is the advance on Moscow, red the retreat; band width still carries troop strength, exactly as in the 1869 original. Not shown here is the separate graph of the decline in temperature Napoleon’s troops suffered on the retreat, contained in HistData::Minard.temp. The help file ?Minard shows ggplot2 steps for both together, and for combining these into a single graphic.

Wilkinson’s GoG specification

Minard’s graphic has a more important claim to fame beyond the many re-visions described here: Leland Wilkinson used it as a capstone illustration of The Grammar of Graphics (1999): Proof that a formal grammar could specify a graphic this intricate (a tapering, bidirectional flow band keyed to two variables at once, sharing its horizontal axis with an aligned temperature series) compactly enough to state once, and precisely enough for a computer to draw it unassisted.

He used this version as cover image for the book. He was saying, “HELLO WORLD: I can do this!!”. See (Friendly 2022) for an account of my conversation (Colorless Green Graphs Sleep Furiously) with Lee about the thinking that went into his work on the grammar of graphics.

Wilkinson’s rendering of Minard’s map (Wilkinson, 1999, Fig. 15.1)

His code for this comprised dplyr-like TRANSformation steps, followed by two FRAMEs for the flow portion and the subscripted line graph of temperature on the retreat.

TRANS: date = miss(date, lont, "linear")
TRANS: ldate = lag(date, -1)
TRANS: days = diff(date, ldate)

FRAME: lonc*latc
GRAPH: point(label(city), size(0))
GRAPH: path(position(lonp*latp), size(survivors), color(direction),
            split(group))
GUIDE: legend(color(direction))

FRAME: lonc*temp
COORD: stretch(dim1, dim2, 1, .2)
COORD: translate(dim1, dim2, 0, -.5)
GRAPH: path(position(lont*temp), label(date), texture.granularity(days),
            color.brightness(.5))
GUIDE: axis1(label("Longitude"))
GUIDE: axis2(label("Temperature"))

Set next to the ggplot2 call above, the correspondence is nearly line-for-line: GRAPH: path(position(lonp*latp), size(survivors), color(direction), split(group)) is geom_path(aes(x = long, y = lat, linewidth = survivors, colour = direction, group = group)) – the same data-to-channel mapping (position, size, color, grouping), just spelled in a different grammar’s syntax. ggplot2, as a direct implementation of Wilkinson’s grammar, doesn’t just resemble this specification; it is a runnable version of it, decades on.

A few more, from the gallery

The rest of Friendly (2002)’s catalog isn’t reproduced here (some entries are third-party work with licensing not yet sorted out for redistribution – see issues/vignettes/Minard-images.md), but is worth naming by taxonomy category, all viewable at https://www.datavis.ca/gallery/re-minard.php:

• Design variation: Gene Zelazny’s pictograph redesign – soldier silhouettes standing in for the flow band, a Napoleon portrait and Moscow’s onion domes flanking the numbers, thermometer icons instead of a temperature line. Same data, a completely different set of design decisions about what to foreground.
• Interactive/web-based: a Protovis port overlaying the flow band directly on a pannable, zoomable Google Map of the actual terrain – the natural next step past a static basemap, and in the same category as the plotly version of Perozzo above.
• Temporal/animated: Ron Kennett’s moving-bubble re-vision, already named in the taxonomy section above.

Re-vision 1: ggplot2 raster + contour

A straightforward “alternative representation” re-vision: give up the 3D illusion entirely in favor of a 2D heatmap with isolines, which trades the stereogram’s dramatic silhouette for exact, directly-comparable values.

ggplot(Perozzo, aes(x = Year, y = Age, z = Survivors)) +
  geom_raster(aes(fill = Survivors)) +
  geom_contour(color = "white", alpha = 0.6) +
  scale_fill_distiller(palette = "Blues", direction = 1) +
  scale_y_reverse() +
  labs(title = "Perozzo's survivorship grid, as a heatmap") +
  theme_minimal()

Re-vision 2: a fake-3D isometric shear

A “design variation”: keep the 3D look Perozzo was after, but build it as an ordinary 2D ggplot via a manual isometric shear (offsetting each Age slice horizontally and vertically as a function of Age), rather than persp()’s true 3D projection. Unlike persp(), this gets full access to ggplot’s color scales, legends, and layering.

shear <- 0.6   # horizontal points per year of age; tune to taste
lift  <- 400   # vertical Survivors-units per year of age

Perozzo_iso <- transform(Perozzo,
  x_iso = Year + shear * Age,
  y_iso = Survivors + lift * Age
)

ggplot(Perozzo_iso, aes(x = x_iso, y = y_iso, group = Age, color = Age)) +
  geom_line() +
  scale_color_distiller(palette = "Blues", direction = 1) +
  labs(
    title = "Perozzo's stereogram as a 2D isometric shear",
    x = "Year (sheared by Age)", y = "Survivors (lifted by Age)"
  ) +
  theme_minimal() +
  theme(axis.text.x = element_blank(), axis.text.y = element_blank())

The shear/lift constants are chosen by eye for a legible “mountain” silhouette, not fit to Perozzo’s own oblique axis convention – a stylized fake-3D, not a literal reproduction of his projection geometry.

Re-vision 3: an interactive plotly surface

An “interactive/web-based” re-vision, and arguably the most direct modern analogue of a physical stereogram: something a reader can actually rotate, rather than a single fixed viewing angle.

library(plotly)
plot_ly(x = ~years, y = ~ages, z = ~Pmat, type = "surface") |>
  layout(scene = list(
    xaxis = list(title = "Year"),
    yaxis = list(title = "Age"),
    zaxis = list(title = "Survivors")
  ))

Still open

The one Perozzo idea from R/Perozzo.R’s @details not attempted anywhere above is animating across Year, letting each birth cohort’s diagonal “arrive” in sequence rather than showing all of them at once (the gganimate idea in “Open invitations” below).