The question
Three registration batches show slightly different degradation rates. The draft proposes a shelf life from all the data pooled together, extrapolated beyond the 24 months of real-time data. Is pooling justified, and what shelf life does the data actually support?
The analysis
ICH Q1E sets the order: test whether the batches share a slope, then whether they share an intercept, each at a significance level of 0.25 so that a real difference is not waved through. Only if both pass can the data be pooled.
| Poolability test | p-value | Result at 0.25 |
|---|---|---|
| Equal slopes (batch x time interaction) | 0.001 | Slopes differ: do not pool |
| Equal intercepts (given common slope) | 0.031 | Not reached: slopes already differ |
| Batch | Slope (% per month) | Where the 95% lower bound meets 95.0% |
|---|---|---|
| Batch 1 | -0.126 | 32.3 months |
| Batch 2 | -0.142 | 29.3 months |
| Batch 3 | -0.248 | 20.1 months |
| Pooled (not justified) | -0.172 | 26.6 months |
What to do
- Base the proposed shelf life on the worst batch, or justify a common slope with more data.
- Investigate why one batch degrades faster; it may be a process or packaging signal, not noise.
- Keep any extrapolation within Q1E limits and commit to confirming with ongoing real-time data.
How it was done
Synthetic assay data for three batches at seven time points. ANCOVA (ordinary least squares) for slope and intercept equality at alpha 0.25; one-sided 95% lower confidence bound on the mean for each regression. You can estimate a single batch in the shelf-life tool. Download the data (CSV).