TrueGate Bio · Analytics Case Work

Assay precision: where does the variability come from?

A variance components analysis of an intermediate precision study, and why retraining analysts would not have fixed it.

Synthetic data, built to demonstrate the method

The question

A potency assay fails its intermediate precision criterion of 8% CV. The proposed fix is retraining the analysts. Will that work?

The analysis

The study ran every combination of 3 analysts, 6 days, and 2 instruments in triplicate (108 results). A random-effects model splits the total variance into its sources, so each one can be sized separately.

2026-09-25T02:57:46.466092 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
SourceCV contributionShare of variance
Day to day2.5%53%
Instrument0.8%6%
Analyst0.9%7%
Repeatability2.0%35%
Repeatability (within run)2.0%
Intermediate precision (all sources)3.4%
2026-09-25T02:57:46.508337 image/svg+xml Matplotlib v3.10.9, https://matplotlib.org/
Why it matters. Analysts account for a small share of the variability. Day-to-day effects dominate: something that changes each day, such as reagent or standard preparation, plate or cell condition, or instrument warm-up, moves the whole run. Retraining analysts would cost time and leave the failure in place.

What to do

  • Investigate what changes between days: fresh reference standard and reagent preparation, cell passage or density, and system suitability results by run.
  • Consider controls that absorb day effects, such as reporting relative to a same-plate reference or tightening run acceptance criteria.
  • Repeat the variance components study after the fix, and set the criterion from the method's intended use, not from the data.

How it was done

Synthetic relative potency results from a fully crossed precision design. Variance components estimated by REML in a mixed model with analyst, day, and instrument as random effects; CVs expressed against the overall mean, consistent with ICH Q2 intermediate precision. Download the data (CSV).

Back to all work   Discuss a similar problem