In short
A control chart tracks an indicator period after period and separates ordinary variation from real change. You point at a date, a variable and a denominator; the widget picks the fitting chart, draws its limits, and flags in red whatever falls outside the ordinary.
Why not a month-on-month comparison
“Infections are up 20% this month”: on small numbers that sentence is almost always noise. Comparing two consecutive points, or a month against last year, produces constant alarms and reactions to variation that means nothing.
A control chart answers differently. It establishes what the process usually does, draws limits around that habit, and flags only what falls outside. It is the reference method for hospital quality monitoring.
Choosing the chart
You can leave the setting on Auto: the widget infers the chart from the nature of the indicator and the shape of the data. Eleven choices are offered.
| Chart | For which indicator |
|---|---|
| p, P′ | A proportion of cases. The P′ variant widens the limits when the data vary more than the model predicts. |
| np | A count of events, when denominators are equal from one period to the next. |
| u, U′ | A rate against an exposure — per 1,000 patient-days, for instance. |
| c | A count with a constant denominator. |
| I | A measured quantity: a length of stay, a delay, a score. |
| EWMA | Slow drifts, which a classic chart takes a long time to see. |
| g, t | Rare events: you track the number of days or the time between two occurrences. |
Rare events are the classic trap
In a unit where an infection occurs twice a quarter, a monthly rate is almost entirely noise. The g and t charts answer the right question — “how long without an event?” — and read the opposite way round: a rising line is an improvement.
The denominator
This is the setting that deserves the most attention, because it decides what the indicator means. Four possibilities:
- Cases — one row, one case.
- A duration column — when the exposure is already computed in your data.
- Patient-days — computed by overlapping stays across the period, from admission and discharge dates.
- Device-days — the same principle, between insertion and removal.
One row per stay, not per measurement
The One row per setting deduplicates the data before computing. Without it, a patient with ten samples would count ten times in the denominator.
Reading the chart
The centre line is what the process usually does; the limits bound its ordinary variation. Flagged points appear in red, and a counter summarises them under the chart.
Staircase limits are normal
When the denominator changes from one period to the next — 80 patients one month, 300 the next — the limits tighten on large periods and widen on small ones. A proportion computed on 300 cases is far more precise than on 80, and the chart shows it.
Flat limits over a varying denominator are the single most common error in hand-made charts.
The signal rules
Three sets of rules are offered.
Anhøj — recommended
Two rules whose threshold adapts to the length of the series: a run that stays too long on one side of the centre line, and too few crossings of it.
Fixed run length
A threshold you set, plus a monotone trend rule. Pick this to reproduce an existing chart.
Out-of-limit points only
The minimum. A point beyond the limits is flagged whatever the setting: that is the chart itself, not an added rule.
Why not the eight classic rules
Stacking all the Western Electric rules makes a perfectly stable process alarm constantly — and mostly teaches teams to ignore the chart. The rules kept by default were calibrated to avoid that.
A signal means "look", not "act"
It says the variation observed is probably not due to chance. What caused it — a change of practice, a change of coding, an outbreak, a new clinician — requires going back to the data.
Freezing the limits
The Freeze limits after setting defines a baseline period: limits are computed up to that date and then carried forward unchanged. This is the clean way to show the effect of an intervention — the limits describe the “before”, and the “after” reads against it.
A vertical line marks the split on the chart.
The warnings
The widget shows above the chart whatever could mislead the reading, rather than letting a misleading chart through:
- Too few periods — the limits are provisional; about twenty are needed.
- Events too rare — a g-chart would read better.
- Overdispersion — the data vary more than the model predicts, and a P′ or U′ chart will raise fewer false alarms.
- Baseline too short — every period was used instead.
- Option ignored — a setting does not apply to the nature of the indicator chosen.
What does not exist yet
- Annotations — marking an intervention on the chart.
- Several phases — only one baseline split is possible.
- XbarS, CUSUM, funnel plots and risk adjustment are under consideration, with no timeline.
A recent feature
This analysis is complete and covered by its tests, but it has not yet been exercised by real use in the application. If its behaviour surprises you, that is a useful thing to report.
Going further
- Analysis widgets — comparing groups rather than tracking time.
- Built-in widgets — charts and indicators without statistics.
- Datasets — preparing the table a control chart expects.