Data visualizations based on Tufte’s principles
Above all else show the data
A working deck MER-201 and MER-102 are invented studies. The rules are not.
Edward Tufte’s ten rules for data graphics, each demonstrated on a figure a clinical statistician actually makes, almost always as a pair: the junk version on the left, the revised version on the right. Then the kill list, drawn the same way, and the checks to run before a figure goes in the CSR.
MER-201 progression-free survival, N = 240. Two curves, two labels, five ticks. Everything else on this slide waits for rule 3.
Rule 1 · Show the data
Rule 01 VDQI, p. 92 — “Above all else show the data”
As most tools draw it
The same 120 patients
The bar shows 2 numbers from 120 patients and hides the rest: the spread, the placebo responders, the active treatment progressors. The dot plot shows all 120 and the medians in red. Same data, opposite amount of information.
Rules 2 + 3 · Maximize data-ink, erase the rest
Rules 02 + 03 VDQI, p. 96 — data-ink is ink that would change if the data changed
Default output
After erasure
The audit behind both rules: change every event time and ask what moves. The curves move; the frame, grid, legend and minor ticks do not, and they are erased in that order. What replaced them is data: direct labels, censor marks, medians where they occur, and the at-risk counts, which are the honest width of the right-hand tail.
Rules 4 + 5 · Say it once, draw it honestly
Rules 04 + 05 VDQI, p. 100 — one encoding per quantity, and a lie factor between 0.95 and 1.05
Five statements, drawn 12x too big
Position once, with its interval
The left panel breaks both rules at once. The bar states 42.3 five ways: top-edge position, fill height, printed label, the gridline it touches, the axis tick beside it. And its axis starts at 35, so an 11% relative difference is drawn as 135%, a lie factor of 12. The right panel says each number once, as position on a scale anchored at zero, with the decimals printed and the 95% interval drawn. The intervals overlap by nine points, which is the finding the truncated bars were built to hide.
Rule 6 · Small multiples
Rule 06 VDQI, p. 67 — “Compared to what?”
Twelve lines, one panel
Six panels, one line each
Mean % change from baseline, six labs, two arms. On the left the ALT signal is one strand in a twelve-line knot with a two-part legend. On the right, six panels on one shared scale, active treatment in ink, placebo in gray; the ALT climb to +24% is the safety story, so ALT alone carries the accent.
Rule 7 · Layer and separate
Rule 07 Envisioning Information — data loudest, scaffolding faintest
Grid at the weight of the data
Grid one step back
Same three dose groups. On the left the grid is drawn at the weight of the data and every crossing makes a small visual event Tufte calls 1 + 1 = 3. On the right the layers are ranked: data in ink, labels next, two faint gridlines last. Nothing crosses anything at equal weight.
Rule 8 · Micro and macro readings
Rule 08 Envisioning Information — “To clarify, add detail”
The summary alone
Every patient, and the summary
eGFR, active treatment arm, 30 patients over 26 weeks, same y scale in both panels. The mean is flat, and the left panel ends the story there. The right panel keeps the macro reading (a stable band) and adds the micro one: three patients decline more than 30%, one past 40%, in red. Means can be stable while patients are not.
Rule 9 · The smallest effective difference
Rule 09 Visual Explanations, p. 73 — distinctions as subtle as possible, still clear
Three saturated hues, one legend
Gray context, one accent
The proposed phase 3 dose is 20 mg, so 20 mg carries the accent and the other arms become gray context, labeled where their lines end. The left panel separates three groups with three saturated hues and makes the reader look up each one; the reading is identical, only the effort changes.
Rule 10 · Integrate words and data
Rule 10 The forest plot is the most Tufte figure in the CSR: a table and a graphic, fused
PFS by subgroup, MER-201. Label, n, estimate and picture share each row; the eye never leaves the line it is reading. No legend, no second panel, no page flip to Table 14.2.4. The point sits on a log scale where a ratio belongs, the interval is drawn rather than footnoted, and the prior platinum CI crossing 1 is legible at a glance.
Kill · The response pie
The kill list, drawn 1 of 3
Angles, matched to a legend
Positions, sorted and labeled
Which is larger, stable disease or partial response? The pie asks you to compare two angles 90 degrees apart, via a legend. The dot plot answers before the question forms, gives the exact counts, and marks the responders, the number the study lives on, in red.
Kill · The dual-axis overlay
The kill list, drawn 2 of 3
Two scales, one panel
Two panels, one clock
The left panel invites a correlation the data never claimed: rescale either axis and the crossing point moves anywhere you like. The right panel keeps the shared clock, gives each measure its own honest scale, and lets the reader align the timings without being steered.
Kill · The rainbow heatmap
The kill list, drawn 3 of 3
Hue has no order
One hue, light to dark, sorted
ALT as multiples of ULN, 24 patients, 26 weeks. Is yellow higher than green? The rainbow forces the reader to keep re-checking the legend, and its brightest band lands mid-scale. One hue running light to dark reads without a key, and sorting patients by peak puts the four who crossed 2 xULN, labeled at week 26, in one place instead of scattered.