Module request: ICH Q1E stability analysis, shelf-life estimation and extrapolation

[Maged-Magdy]

Thank you for updating the module. I tested version 1.2.1 on my Mac and have a few observations. Could you please check them and let me know if I have misunderstood any settings?

I compared the results with ICH Q1E and calculations made separately in R. I also used test data with known results. This was a check of selected examples, not a complete formal validation.
1. Shelf-life limits
With 36 months of data, some refrigerated settings gave a limit of 48 months where I expected 42. The setting for significant change between months 3 and 6 allowed an extension where I expected none. A frozen-storage setting also added six months.
Could you check these against ICH Q1E sections 2.5.1 and 2.5.2? The expected limit depends on the selected conditions.
2. Supporting data
Some room-temperature results still said “Fully Supported” when supporting data were unchecked.
Could you check this against sections 2.4.1.2 and 2.4.2.1? Those sections require supporting data for the extensions tested.
Official ICH Q1E guideline
3. Model selection
In one test with additional factors, the results recommended Model 1, but the shelf-life table used Model 3. Selecting Model 1 manually changed the result from 45.2 to 31.4 months.
The separate calculation agreed with the manual Model 1 result. Could you check that the recommended model is the one used throughout the calculation?
Also, one batch/package combination had only one measurement, but the analysis still gave a result using separate slopes. Please check how cases with too little data are handled.
4. Confidence limits and the search range
With Packaging added, selecting two-sided 95% confidence limits gave the same result as lower-only limits. The separate calculation gave different results. Could you check the two-sided calculation? ICH Q1E section 2.6 explains the distinction.
Another test reported 120 months, although the confidence line had not reached the specification limit. It appeared to be the end of the search range. Could the output say “beyond the search range” in this situation?
5. Temperature calculations
With test data generated using Ea = 80, selecting log transformation gave 74.219. Calculating the rates from the logged values gave 80; using the raw values reproduced 74.219.
Could you check whether the selected transformation is also used in the temperature calculations?
Differences between batches also affected the results. In another test, combining batches gave 58.132 instead of the known 80. When batches had different energies, the combined calculation underestimated the fastest batch’s degradation rate. Could you check how batch differences are handled before combining their results?
6. Ea sensitivity table
The row assuming Ea = 50 changed when I changed only the main Ea setting, although the observed data stayed the same.
Could you check why that row depends on the main Ea value, or explain how it should be interpreted?
7. Input checks
Please check these cases:
- Month imported as a nominal variable was treated as category numbers instead of the actual months.
- A zero upper specification limit with log transformation was effectively ignored.
- Temperatures below absolute zero and infinite temperature values were accepted in some tests.
These were reproduced with saved examples. For the temperature conversion, I used the NIST reference.
8. Tables and plots
Adding factors removed some batch rows, equations and crossing markers. Some plots also ended before the reported crossing.
When a secondary attribute failed validation, the output correctly removed the numerical result but still said “All CQAs Met.”
Could you check these displays and clearly distinguish the statistical estimate from the permitted shelf life?
9. What the Ea confidence interval includes
The reported interval matched the calculation used by the module. My question is whether it also accounts for uncertainty in the estimated degradation rates.
Could you explain what it includes? This is a request for clarification, not a confirmed calculation error. I used NIST’s uncertainty guidance when exploring this point.
Basic examples also gave the expected results, and some of the observations occurred in the older source too.

I have saved the test data, jamovi files and separate calculations. I can share the examples so you can reproduce the results and check both the calculations and my interpretation.
https://drive.google.com/file/d/1zKumWJ ... sp=sharing