The Dashboard Nobody Reads
A regional health network emails a weekly PDF dashboard to 30 clinic managers. The current version contains a dual-axis combo chart plotting appointment volume as bars on the left axis and no-show rate as a line on the right axis, with the bar axis starting at 4,000 instead of zero. Below it is a 3-D pie with 11 slices for cancellation reasons, then a 47-row table sorted alphabetically by clinic name. Key measures carry a red/amber/green tile comparing no-show rate to a 4.0% target. Three managers with red-green color vision deficiency report they cannot read the tiles. One small clinic at 4.1% shows red; it books about 60 appointments a month. Executives say they 'cannot tell whether things are getting better.'
1. Critique the combo chart and the pie, and state the visualization principle each one violates.
Read the worked answer
Combo chart - two violations.
- The bar axis starts at 4,000. Bar length encodes magnitude, so a truncated baseline misrepresents the ratio between values: a 4,000 and a 4,400 month appear as wildly different heights rather than a 10% difference. Bars must start at zero; only position-encoded charts (lines, dots) may crop the axis.
- Dual axes let the author pick two independent scales. The line's apparent relationship to the bars is an artifact of the chosen limits - nudge the right axis and the 'story' changes. Viewers read correlation that the data does not contain.
- Fix: two stacked panels sharing one x-axis (volume on top, rate below), or index both series to the first period on a single axis. If a combo must stay, zero-base the bars and label both axes loudly.
3-D pie - two violations.
- A pie encodes value as angle and area, which humans judge poorly - worse than length. Perspective and depth add parallax error, so the slices nearest the viewer look larger. There is no valid reason for 3-D.
- 11 slices exceeds the ~5–7 categories anyone can rank. Small slices become indistinguishable slivers; a legend forces eye-scanning back and forth.
- Fix: sorted horizontal bar chart, descending, with the six smallest categories rolled into 'Other (n categories)' and a footnote listing them.
2. Redesign the page for the three stated audiences and fix the accessibility failure.
Read the worked answer
Accessibility (WCAG 1.4.1 - do not use color alone):
- Red/amber/green tiles are the classic failure. Encode status redundantly: an icon (▲/●/▼ or a check/warning), a text label ('Above target'), and a pattern or shape change - not hue alone.
- Use a colour-vision-safe palette (e.g., blue–orange diverging) and verify contrast ≥ 4.5:1 for text.
- Do not rely on hover tooltips in a PDF; PDFs have no interaction state.
Audience-driven layout:
- Clinic managers need 'my clinic, this week, versus my own history': one page per clinic, a small-multiples trend of the last 12 weeks, and the two or three measures they can act on.
- Executives need 'are we improving, where is the spread': a single trend line for the network, a distribution or ranked bar of clinics, and exceptions called out in text.
- The 47-row alphabetical table serves nobody. Alphabetical order conveys no information. Sort by exception severity (largest gap to target, with confidence band) or group by region with subtotals. Add sparklines for trend.
Answer the actual complaint: executives cannot tell if things are getting better because the dashboard shows state, not change. Add a change view - current period vs. prior period, plus a 12-point trend - and annotate it. One annotated sentence ('Network no-show rate fell 0.6 points over 8 weeks, driven by three clinics') will do more than any chart.
3. The small clinic at 4.1% against a 4.0% target shows red. Is that a fair signal? What should replace the fixed-threshold tile?
Read the worked answer
No - that red tile is almost certainly noise.
At ~60 appointments a month, 4.1% is about 2.5 no-shows. The 95% confidence interval on a rate estimated from 60 trials at p≈0.04 is roughly ±4.7 points - the clinic cannot be distinguished from 0% or 9% with that denominator. Flagging it red next to a 9,000-appointment clinic at exactly 4.1% treats a coin flip and a real signal as identical.
Worse, this is a ranking trap: small units will always produce extreme rates, so a naive top-10-by-rate list is dominated by tiny denominators. Managers learn to ignore the flags, which is how dashboards die.
Replace with a variance-aware view:
- Funnel plot: x = appointment volume, y = no-show rate, with 95% and 99.8% binomial control limits drawn as curves that narrow as volume grows. A clinic is flagged only if it falls outside the funnel. Small clinics get wide bands automatically.
- Or a p-chart across time with 3-sigma limits, which also answers 'is this getting better' - a special-cause shift is visible, ordinary month-to-month jitter is not.
- Show the denominator next to every rate, and sort by distance outside the control limit rather than by raw rate.
Report the caveat: control limits assume independence and stable case mix. If no-show risk varies by appointment type, risk-adjust or stratify before flagging - otherwise you will punish clinics that serve the highest-risk populations.