RD Rolling Median
Overview
The Rolling Median chart shows how duplicate precision changes through time, independent of grade. Each light point is one duplicate pair and each dark point the median of one despatch’s pairs; the line is the rolling median, which smooths the despatch-to-despatch scatter so a trend stands out. A rising line points to deteriorating precision and should be investigated; a falling line to improving precision.
The chart covers the despatches in the filters: the date range, or the despatches chosen with “Select by”. Only pairs whose mean is at least the threshold (the “Threshold factor” × the lower detection limit) are used.
How it is calculated
For each pair, the absolute Mean Pair Relative Difference, as the Overview uses:
|MPRD%| = |a − b| / ((a + b) / 2) × 100
where a is Sample 1 (the original) and b is Sample 2 (the duplicate). Despatches are ordered by their earliest laboratory return date (or send date, with “Despatch send”).
- Despatch median (dark points): the median |MPRD%| of all the pairs in the despatch.
- Rolling median (the line): at each despatch, the median |MPRD%| of all the pairs in that despatch and the ones before it, up to the “Rolling window” (5 to 30 despatches, 20 to start). The first few despatches have fewer before them, so their window is shorter.
The rolling median is taken over the pairs themselves, not over the despatch medians, so a despatch with many pairs counts for more than one with a single pair. A median keeps a single poor pair from moving the line, so it follows the typical precision rather than the occasional bad pair.
Limits
- Stage limit: the check stage’s RD limit from the Overview rules: your own setting, or the standard limit for the stage (±30% for field, ±20% for coarse and ±10% for pulp duplicates), or a limit guessed from its name, or the general default.
- Action limit: twice the stage limit.
A rolling median above the stage limit means the typical pair over that stretch of despatches is worse than the stage allows; above the action limit, markedly so. The stats under the chart count the despatches where the rolling median is over each.
Worked example
Three despatches of pulp duplicates (stage limit 10%, action limit 20%), with a rolling window of 3 to keep it short:
| Despatch | Pairs’ |MPRD%| | Despatch median | Rolling median |
|---|---|---|---|
| D101 | 2.1, 3.4, 2.8, 15.0 | 3.1 | 3.1 (D101’s 4 pairs) |
| D102 | 4.2, 5.0 | 4.6 | 3.8 (D101–D102’s 6 pairs) |
| D103 | 9.8, 11.4, 12.6 | 11.4 | 5.0 (D101–D103’s 9 pairs) |
D101’s 15.0% pair is well above the limits, but its despatch median stays at 3.1%. D103’s median of 11.4% is above the 10% stage limit: its pairs are consistently poorer. The rolling median rises from 3.1% to 5.0%, and would cross the stage limit if the next despatches were as poor as D103.
Interpretation
- Each light point is the |MPRD%| of an individual duplicate pair; each dark point is a despatch’s median.
- The blue line is the rolling median, and its latest value is labelled.
- The stage limit and action limit are shown for reference.
- A rising trend may indicate deteriorating precision and should be investigated, for example against changes in laboratory, preparation or sampling.
- A shorter window reacts faster to a change but is noisier; a longer one is steadier but slower to show it.