Guide8 min readJuly 15, 2026

ICH Q1E Shelf Life Calculation from Accelerated Data

ICH Q1E shelf life calculation from accelerated stability data: the method, a worked example with real Arrhenius output, and when extrapolation is not OK.

Oliver Kraft

CovaSyn

ICH Q1E Shelf Life Calculation from Accelerated Data

You have six months of 40 C accelerated data, three months of 25 C long-term data, and a project team that wants a shelf life number now. The regulatory answer is that shelf life comes from long-term data. The practical answer is that you still have to make a decision today, and your accelerated arm already contains most of the information you need. This article covers how the two methods work, what ICH Q1E actually permits you to extrapolate, and where the whole approach breaks down.

The two numbers you are trying to reconcile

There are two independent routes to a 25 C shelf life, and a good stability package uses both.

Route 1 - kinetics (Arrhenius).

Extract a degradation rate at each storage temperature, regress ln(k) against 1/T, and use the fitted activation energy to project the rate at 25 C. Then solve for the time at which the projected assay hits your specification limit. This route is fast: it works as soon as the accelerated arm has enough timepoints to give a reliable rate.

Route 2 - ICH Q1E regression.

Regress the long-term 25 C data against time, take the 95% one-sided confidence limit of the mean curve, and read off where that limit crosses the acceptance criterion. This is the route that goes in the dossier. It is slower, because it is limited by how long the long-term arm has been running.

Route 1 is a prediction. Route 2 is the regulatory calculation. If they agree, you have convergent evidence and a defensible position. If they disagree badly, something in your model or your data is wrong, and it is better to find that out at month 6 than at month 24.

What ICH Q1E actually lets you extrapolate

This is the part that gets misremembered most often. Q1E does not let you file an Arrhenius number as your shelf life. It lets you extrapolate the long-term regression by a bounded amount, and it uses the accelerated and intermediate data to decide how much.

Situation at accelerated (40 C / 75% RH)	What Q1E allows
No significant change over 6 months, long-term data show little change and little variability	Up to 2X the long-term period covered, capped at +12 months, often without formal statistical analysis
No significant change at accelerated, but long-term data show change or variability	Up to 2X, capped at +12 months, with statistical analysis backing it
Significant change at accelerated, no significant change at intermediate (30 C)	Reduced extrapolation - roughly 1.5X the long-term period, capped at +6 months
Significant change at intermediate as well	No extrapolation. Shelf life is the period covered by long-term data, or shorter

Two further Q1E specifics worth having in your head: batch poolability is judged with an ANCOVA at the 0.25 significance level (a deliberately permissive threshold, because failing to pool is the conservative outcome), and the shelf life is taken from the 95% one-sided confidence limit rather than the fitted mean line. Check the Appendix A decision trees against your own case before committing to an extrapolation - they branch on drug substance versus drug product and on whether the attribute is expected to increase, decrease or stay flat.

Worked example: a small-molecule drug product

The dataset below is a representative small-molecule drug product stability set: assay as % label claim, arms at 25 C, 30 C and 40 C, specification at 95% LC. The inputs are representative. Every output quoted here is a verbatim return from the named CovaStab tool, run over the MCP gateway on that data - not a hand-fitted spreadsheet number.

Step 1 - fit the temperature dependence.

covastab_fit_arrhenius returns:

Activation energy Ea = 107.45 kJ/mol
Q10 = 4.08
Arrhenius fit R2 = 0.995

A Q10 of 4.08 means roughly a fourfold increase in degradation rate per 10 C. That single number explains why the accelerated arm is useful: at 40 C the molecule is degrading about 16 times faster than at 25 C, so six weeks of 40 C data carries about as much degradation signal as two years at 25 C. An Ea of 107 kJ/mol sits in the normal range for small-molecule hydrolytic and oxidative degradation, which is a sanity check worth doing - values far outside roughly 50 to 150 kJ/mol usually mean the rate extraction or the mechanism assumption is off.

Step 2 - project to the storage condition.

covastab_predict_at_temp returns 30.4 months to the 95% LC spec at 25 C, projected from the 40 C and 30 C rates through the Arrhenius fit.

Step 3 - run the Q1E calculation independently.

covastab_estimate_shelf performs the long-term regression and returns 28.2 months, with regression R2 = 0.9989 and a slope p-value of 1.3e-8. Its embedded poolability sub-result returns pool_all, meaning the batches were not statistically distinguishable at the Q1E threshold and were analysed as a single pooled dataset.

Step 4 - check that the kinetic model was not chosen by eye.

covastab_kinetics_model scores the candidate orders by AIC: zero order -19.68, first order -21.04, second order -21.94. Second order wins. The R2 values are almost indistinguishable (0.99892 / 0.99910 / 0.99921), which is exactly why model selection should be done on an information criterion rather than on R2.

The three candidate orders are indistinguishable by R2 to four decimals, and AIC still separates them cleanly. Pick the kinetic model on an information criterion, never by eye on R2. Source: Computed with covastab_kinetics_model over the CovaSyn MCP gateway, on the stability dataset described in this article.

The reconciliation.

30.4 months from kinetics, 28.2 months from Q1E regression - about two months apart, with the regulatory method being the more conservative of the two. That is the outcome you want. The Arrhenius number told you the answer months earlier; the Q1E number is what you write down.

Two independent routes to a 25 C shelf life agree to about two months, with the regulatory Q1E number the more conservative one. The Arrhenius value is decision support, not a filing number. Source: Computed with covastab_predict_at_temp and covastab_estimate_shelf over the CovaSyn MCP gateway, on the representative small-molecule drug product stability set described in this article (assay as % label claim, arms at 25 C, 30 C and 40 C, spec 95% LC).

What these numbers do not tell you

Be blunt about this internally, because reviewers will be.

Arrhenius assumes one dominant mechanism that does not change with temperature. If 40 C activates a pathway that is negligible at 25 C, or if a phase change, melting or moisture-uptake event occurs between the conditions, the extrapolation is invalid no matter how good the R2 is. An R2 of 0.995 on three temperature points is not proof of mechanistic constancy - it is three points.
30.4 months is a prediction, not a shelf life. It supports a decision (formulation ranking, packaging choice, go/no-go on a stability protocol). It is not a filing number and CovaStab does not make it one.
Assay is not the only stability-indicating attribute. Shelf life is set by whichever attribute fails first: total degradants, a specified impurity, dissolution, water content, appearance, microbial limits. An assay-based calculation is one input.
Solid-state and biologic products often violate the assumption outright. Amorphous dispersions, lyophilised products and proteins frequently show non-Arrhenius behaviour. Treat kinetic extrapolation there as hypothesis generation only.
The statistics do not fix a thin design. Three batches with sparse timepoints will give a wide confidence limit, and the Q1E shelf life comes from the limit, not the mean. Design the study properly; no model recovers information that was never collected.
Tooling caveat from our own testing: in the run above, the standalone covastab_poolability and covastab_extract_rates endpoints had backend issues and none of their output is quoted here. The pooling verdict comes from the poolability sub-result embedded in covastab_estimate_shelf. We would rather say that than quietly present a number we did not verify.

Frequently asked questions

How is shelf life calculated under ICH Q1E?

ICH Q1E calculates shelf life by regressing the long-term stability data against time and finding the point where the 95% one-sided confidence limit of the mean curve intersects the acceptance criterion. Batches are tested for poolability by ANCOVA at the 0.25 significance level; if poolable, all batches are regressed together. The result may then be extrapolated by a bounded amount that depends on the accelerated and intermediate data.

Can I use accelerated data to set shelf life?

Not directly. Accelerated data support extrapolation of the long-term result rather than replacing it. Under ICH Q1E, if there is no significant change at accelerated conditions over six months, you may extrapolate the long-term regression to up to twice the period covered, capped at 12 months beyond it. Arrhenius projections from accelerated data are decision-support during development, not a filing number.

What is Q10 and why does it matter for stability?

Q10 is the factor by which the degradation rate increases per 10 C rise in temperature. In the CovaStab example above, covastab_fit_arrhenius returned Q10 = 4.08 with an activation energy of 107.45 kJ/mol and a fit R2 of 0.995. A Q10 near 4 means 40 C degrades roughly 16 times faster than 25 C, which is why a short accelerated study can resolve a shelf life question early.

When is Arrhenius extrapolation not acceptable?

When the degradation mechanism is not constant across the temperature range. Typical failure cases are amorphous solids near Tg, lyophilised and protein products, systems with moisture-mediated or autocatalytic degradation, and formulations that undergo a phase or physical change between the accelerated and long-term conditions. In those cases a high Arrhenius R2 is not evidence of validity, and extrapolation should not be used to support a shelf life claim.

How far beyond the long-term data can I extrapolate?

The usual limit is twice the period covered by long-term data, and never more than 12 months beyond it, when the accelerated arm shows no significant change. If significant change occurs at accelerated but not at intermediate conditions, the allowance drops to roughly 1.5X and no more than 6 months beyond. If significant change also occurs at intermediate conditions, no extrapolation is permitted.

What does a poolability result of pool_all mean?

It means the ANCOVA found no statistically significant difference between batch slopes or intercepts at the Q1E 0.25 significance level, so all batches can be regressed as a single dataset. In the example above covastab_estimate_shelf returned pool_all alongside a shelf life of 28.2 months, a regression R2 of 0.9989 and a slope p-value of 1.3e-8. Pooling generally tightens the confidence limit and lengthens the calculated shelf life.