Experimental units and visual independence

Summary

Biological data often have complex structure (e.g. samples nested within donors or repeated measures). A plot should not treat all observations as independent if they are not. The experimental unit (the entity randomized or independently sampled) and observation unit (measured item) must be identified. For example, multiple cells from the same patient are correlated: plotting them as independent points can exaggerate power. Always encode grouping or pairing: use separate panels for subjects, or connect paired points. The plot must preserve correct unit context to avoid pseudoreplication.

Core rules

Identify true replicates: Determine what was independently sampled (e.g. individual donors, litters, cultures).
Map group structure: If observations are nested (e.g. cells within a donor), use jittered points colored by donor or faceted by donor to indicate grouping.
Paired data: For matched samples (pre/post, mother/child), connect paired points or plot differences, not separate group means.
Random effects: If batch or plate is a factor, consider facet or random jitter to reveal batch effects.
Sample size notes: Always show n at appropriate level (e.g. n=5 donors, n=100 cells). Do not confuse “points” count with true independent replicates.

Required context

Knowledge of the study design: which level was randomized, what constitutes a replicate vs subsample.
Any blocking or batch variables (date, plate ID) that induce non-independence.
Subject identifiers if data are paired or longitudinal.

AI behaviour

Check independence: Before aggregating or testing, confirm if observations share an experimental unit. If unsure, ask or infer from metadata (e.g. same patient ID).
Annotate plots: Indicate paired lines or color-code grouping by replicate source.
Flag pseudoreplication: If an analysis or plot assumes independence incorrectly, raise an alert.
Preserve design: Don’t average replicates unless asked; instead, show variability at the replicate level.

Common failure modes

Unmarked repeats: Plotting all cell-level points without marking which donor they came from, and treating them as independent replicates.
Disconnected pairs: Showing pre/post means in separate boxplots without linking subjects hides within-subject trends.
Ignoring batch: Overlooking plate effects that cause clustering in a scatterplot.
Overaggregation: Aggregating data at donor level when the question targets cell-level distribution, losing granularity.

Authoritative standards

Experimental design principles: Follow ARRIVE/SIMRO guidelines: replicate = independent unit.
Biology Skills experimental-design: Validate independence as in the experimental-design skill.

Examples

Example: Cell vs donor plot

Data: Gene expression of single cells from multiple donors.
Check: Color points by donor ID or facet by donor; do not calculate one mean per donor without indicating variety.
Good outcome: UMAP scatter where each donor has a distinct color legend (legend labels = Donor IDs), or separate panels per donor showing distribution.

Example: Paired measurements

Data: Blood pressure before and after treatment in same patients.
Check: Use paired scatter/line plot or connected dot plot.
Good outcome: Each patient is a line connecting “before” to “after” point, rather than two unconnected boxplots labeled “Before/After”.

Sources

Biology Skills: experimental unit and replication rules.
Foundational stats/design (ARRIVE, etc.) – e.g. “Experimental unit is the biological entity independently assigned” (ARRIVE guidelines).