TrueGate Bio · Analytics Case Work

Process capability on a drifting process

Why a comfortable Ppk across 40 batches can hide a process that is losing margin, and what the control chart shows instead.

Synthetic data, built to demonstrate the method

The question

A continued process verification report shows a capability index (Ppk) for monomer purity comfortably above target across 40 batches, and concludes the process is capable. Is it?

The analysis

2026-09-25T02:50:25.439553 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
MeasureValue
Ppk over all 40 batches, as reported1.62
Trend in monomer per batch (regression slope)-0.039 % per batch, p < 0.001
First run rule signal (8 consecutive below center line)batch 22
Ppk, most recent 12 batches (looks even better: a short window hides the trend)2.13
Batch where the trend brings the 3-sigma band to the lower specabout 63

A capability index assumes a stable process: one mean and one spread. This process is drifting down. Ppk over all 40 batches averages early and late batches into one number that no single batch resembles, and a shorter recent window looks even better because it is too short to show the trend. Either way, the index describes the past, not where the process is heading.

Why it matters. Both capability numbers say capable, and neither can see that the mean is moving. The run rules fired at batch 22, and at the current rate the process reaches the edge of specification in roughly 23 more batches. Capability should only be reported once the process is shown to be in control.

What to do

  • Open an investigation on the trend signal, not on the capability number.
  • Report capability only for a period that passes control-chart run rules, and state that period.
  • Use the projected crossing point to set how urgently the root cause needs to be found, and trend the mean itself, not just capability.

How it was done

Synthetic monomer results for 40 batches with a slow downward drift. Individuals chart with limits from batches 1 to 20 (moving-range sigma), Western Electric run rule of eight consecutive points on one side, linear regression for the trend, and Ppk computed as (mean minus LSL) / (3 x overall SD). Download the data (CSV).

Back to all work   Discuss a similar problem