A meaningful effect.
Choose the smallest difference that matters biologically—not simply the effect you hope to see or the biggest effect in a published paper.
How many animals does your study really need? Explore sample size and statistical power for common in vivo designs—with the assumptions in plain sight.
Seven common designs. Transparent methods.
A clearer conversation about your next study.
Set your assumptions to estimate sample size.
An estimate is only as useful as the question and assumptions behind it.
Choose the smallest difference that matters biologically—not simply the effect you hope to see or the biggest effect in a published paper.
Use comparable prior data when available. Standard deviation is not standard error, and small pilots can give unstable estimates.
Sample size counts independent experimental units. More measurements from the same animal do not necessarily mean more independent information.
Each scenario maps to a specific planned analysis. The calculator is not a substitute for choosing that analysis.
Read our sample size guide ↗Continuous normally distributed outcomes (paired: normally distributed differences). Student’s and paired t-test power use the noncentral t distribution. Welch power uses a Satterthwaite degrees-of-freedom approximation. Equal group sizes; two-sided tests.
Balanced groups, normally distributed outcomes, and a common SD. Power is for one prespecified treatment-versus-control comparison with a two-sided Dunnett critical value controlling familywise α across all treatment-versus-control comparisons. The critical value is calculated by deterministic numerical quadrature; sample-size decisions add a conservative 0.001 percentage-point power margin for numerical uncertainty. It is not the probability of detecting any or all treatments.
Two randomized groups, one continuous baseline covariate, common linear baseline–endpoint relationship, and common residual variance. Uses residual SD = endpoint SD × √(1 − r²), with 2n − 3 residual degrees of freedom and an approximate noncentral t calculation. Not a general longitudinal model.
Two independent groups and a two-sided, uncorrected pooled-score test for equal proportions. Power uses a large-sample normal approximation, not Fisher’s exact test. Sparse expected cells are flagged; exact small-sample planning needs specialist review.
Schoenfeld’s event-based approximation for a two-sided log-rank comparison, equal allocation, proportional hazards, and independent censoring. Total enrollment is estimated from the average probability of observing an event by study end. Required events are not the same as required animals.
Four independent, equally sized cells with normally distributed outcomes and a common SD. Tests one prespecified difference-in-differences interaction using a noncentral t distribution with 4n − 4 degrees of freedom. This is not a calculation for the main effects or each separate cell comparison.
Clustered or litter-randomized designs, general repeated-measures mixed models, count or ordinal models, unequal allocation, interim looks, non-inferiority, and multiple primary endpoints. A single prespecified per-animal summary such as AUC can use the continuous-outcome scenario only when its own SD and planned analysis are appropriate.
Up to 10,000 analyzable units per group or cell; 2–5 treatment groups for Dunnett; effects up to five SDs; baseline correlation from −0.95 to 0.95. Noncentral-t calculations are bounded to noncentrality ≤37, avoiding the library’s extreme-tail approximation. Values outside the supported range are rejected rather than silently substituted. Numerical functions use lib-r-math.js 3.0.2 (MIT). The pinned library loads from jsDelivr; calculations run in a page-owned browser worker.
Look for prior experiments using a comparable model, endpoint, and protocol, or get help planning a pilot. There is no universal SD or “medium effect” that safely replaces study-specific assumptions. Explore a plausible range rather than trusting a single guess.
No. It means an 80% chance of detecting the specified effect if that effect and the other assumptions are true. A real study may have a smaller effect, more variability, or a different missing-data pattern.
The copyable summary records the method and assumptions to help draft your justification. It does not constitute ethics approval or statistical validation of your design. Include the evidence behind your inputs and have the plan reviewed.
The calculator itself runs locally in your browser and does not upload or store your study inputs. The website’s existing analytics and privacy policy still apply. Copying a summary places it on your device’s clipboard.
Find comparable in vivo studies and the experimental conditions behind them. Build your next protocol from evidence, not guesswork.