Skip to main content Link Menu Expand (external link) Document Search Copy Copied
Biology Skills

Experimental design

Use this skill when a biological conclusion depends on the structure of an experiment rather than only on the number of measurements produced.

The central question is:

What was genuinely independent, and could the design distinguish the claimed biological effect from its technical alternatives?

Core rules

  • Identify the experimental unit before interpreting sample size or independence.
  • The unit measured is not automatically the unit independently assigned to treatment or condition.
  • Distinguish biological replication, technical replication, subsampling, repeated measurement, and duplicated computation.
  • Do not infer independent n from the number of rows, cells, reads, fields, wells, images, libraries, or other measurements.
  • Preserve nesting and clustering when multiple observations derive from the same donor, animal, specimen, culture, library, batch, or experimental unit.
  • Match the statistical model to the level at which treatment, exposure, or sampling was independently varied.
  • Distinguish the intended experimental design from the realised design after exclusion, assay failure, missingness, and QC.
  • Do not rely on post hoc batch correction to identify a biological effect when batch and biological condition are completely confounded.
  • Controls must address the alternative explanation relevant to the claim; a nominal control label does not guarantee process matching.
  • Technical replication can estimate technical variability but does not by itself establish biological reproducibility.
  • An orthogonal validation is strongest when it does not simply reproduce the same upstream failure mechanism under a different tool or assay name.
  • Preserve randomisation, blocking, processing order, plate, lane, batch, and other design variables when they can affect the result.

AI behaviour

Before treating observations as independent, determine:

  1. what unit was independently sampled or assigned,
  2. what unit received the intervention or exposure,
  3. which observations share a biological source,
  4. which observations share a technical process,
  5. which repeated measurements or subsamples belong to the same unit,
  6. what n actually counts.

If treatment and batch are aligned, state the confounding rather than assuming normalization can uniquely recover the missing comparison.

When a study reports many cells, reads, images, wells, or fields from a small number of donors or animals, preserve the higher-level biological structure.

When describing replication, name the replication level rather than using replicate without qualification.

References

Read the relevant reference when the task depends on it:

Table of contents