thebench.bio #1 Ecosystem for the Bio-Tech Industry
All Quality Control questions

Question 18 of 20 Advanced Interpretation

When would a method transfer fail despite acceptable average recovery?

Short answer

When the average hides what matters: a bias that offsets at different levels, much higher variability at the receiving site, a failure on real samples but not spiked ones, or too few results to show equivalence. The transfer criteria should test bias and precision separately, ideally with an equivalence approach.

Why

A mean recovery of 100 percent across a transfer can come from results of 94 and 106 percent, which no one would accept for a 98 to 102 percent specification. Transfer exists to show the receiving laboratory can make the same decisions as the sending one. WHO's 2022 technology transfer guideline asks for a protocol with an experimental design and acceptance criteria agreed before execution, and for the number of analysts and parameters evaluated to be defined and justified.

ICH Q2(R2) says transfer can use comparative analysis of representative samples. Several patterns sink a transfer even with a good average. Precision at the receiving site may be twice the sending site's, which will generate OOS results near limits.

Bias may depend on level, high at low concentrations and low at high ones, so the average looks fine while impurity results near the reporting threshold are wrong. Spiked placebo may recover well while real product, with its matrix and degradants, behaves differently. And a test of whether the two means are statistically different can pass simply because there were too few replicates to detect a difference.

A stronger design defines an equivalence margin for the mean difference that is small relative to the specification width, plus a precision criterion, and includes real lots that span the range. When a transfer fails this way, look first at method steps that differ between sites, such as equipment models, columns, water quality or analyst interpretation.

A strong answer covers

  • Averages can hide offsetting bias and high variability
  • Separate criteria for bias and precision
  • Real samples across the range, not only spikes
  • Equivalence testing rather than failing to find a difference
  • Predefined criteria in the protocol

Common mistakes

  • Passing a transfer because a t-test found no significant difference with few replicates.
  • Using only spiked samples and never testing real product lots.
  • Setting acceptance criteria unrelated to the specification width.