# Plan an experiment

> Check an experimental design before you collect data: the unit you count, batch confounding, sample size under stated assumptions, and a plan locked before results exist.

## Check the design. Before the first sample.

Define the comparison, count the right experimental unit, and inspect batch confounding and sample-size assumptions. Approve a versioned plan that later data and analysis can be checked against.

## The comparison starts with the unit.

Illustrative design review: three wells from one mouse do not create three independent mice.

| Design decision | What needs checking | Why it matters |
| --- | --- | --- |
| Experimental unit | Which unit independently receives treatment? | Keep animals and repeat measurements distinct. |
| Group and batch | Were all controls assigned to one day and treatments to another? | A complete group/day confounding cannot separate the two effects. |
| Sample size | What effect, variability and allocation are assumed? | Power and required sample size depend on those inputs. |
| Timing | Which comparison and rules were fixed before data arrived? | Later changes remain later versions. |

Planning checks cover two independent groups and one continuous outcome. Study conduct, independence and measurement quality remain declared conditions; the check does not certify the entire experimental design.

## Describe. Review. Approve. Keep the original plan.

1. **Describe — Tell your AI the experiment.** Groups, units, measurements, batches, and what you hope to detect, in ordinary language or from existing protocol notes.
2. **Check — Get warnings and open questions.** Unit miscounts, confounding, and missing information are listed separately, each with a suggested fix.
3. **Size — Compare candidate designs.** See power and sample size under the assumptions you choose, side by side.
4. **Lock — Approve and lock the plan.** Only changes you approve enter the design, and the approved version is locked before data arrive.

## A plan the analysis can refer back to

- A list of design issues, open questions, and suggested fixes
- Power or sample-size results with every assumption stated
- A locked, versioned design that later analysis is checked against
- An optional two-stage plan with its decision rules

## Design checks, in detail.

### Count the right unit.

SciGround separates what you want to draw conclusions about, the experimental unit (whatever independently receives the treatment—usually the animal, not the well), and what you actually measure. It flags designs that count technical replicates—repeat measurements of the same sample—as independent, and recognizes when a well or tissue section can legitimately be the experimental unit.
- Three wells from one mouse are not three mice
- Biological, experimental, and observation units kept distinct

### Spot batch traps early.

If every control runs on Monday and every treated sample on Tuesday, the treatment effect and the day effect cannot be separated. SciGround detects complete confounding between group and batch before the experiment starts.

### Sample size, with the assumptions in view.

Calculate power or the required sample size from the effect you want to detect, the expected variability, the significance level, and the group allocation, using a standard power calculator. The rounded-up sample size is rechecked, and every assumption stays printed next to the number.

### Lock the plan before the data.

Register the comparison, experimental units, outcome, measurement unit, analysis plan, and assumptions before any data arrive. Later changes become a new version; the original plan is never quietly rewritten to look pre-specified.

### Plan an early look—safely.

Run a study in two stages with one interim decision fixed in advance. Stop early when the evidence is clear or continue to the second stage, while the overall false-positive rate stays within 5%. Every decision is recorded with the data it was based on, and a welfare stop is recorded as exactly that.

### Answer in plain words.

You don’t need to declare “independence” in statistical language. SciGround asks the questions that matter—which animals shared a cage, which samples came from the same donor—and records your answers as stated conditions, separating what your records support from what remains an assumption. Questions already answered for this version are not asked again.

## Limits and conditions

A design check is a strong safeguard, not a certificate that the study is sound. See exactly what it covers. [Read the limits](https://www.licklider.ai/sciground/plan/limits/) ([Markdown](https://www.licklider.ai/sciground/plan/limits.md)).

## Related

- [Supported analyses](https://www.licklider.ai/sciground/analyses/)
- [Explore, then confirm](https://www.licklider.ai/sciground/exploration/)
- [Start from a design](https://www.licklider.ai/docs/sciground-start/#design)