Pseudoreplication: Readings Are Not Experiments

Lazic examined neuroscience examples in which observations shared a subject, hierarchical organization or temporal and spatial correlation. The problem arose when inference ignored those relationships. Dependence can exist even if every row contains a different measurement and no value has literally been duplicated. Methodological study.

In a hypothetical example, several images from the same tissue describe different fields, but do not themselves create new organisms. The analysis must preserve those fields' shared origin.

Aarts and colleagues studied nested designs and showed that ignoring their structure can increase the frequency of incorrect statistically significant conclusions. The size of this problem depends on data organization and correlation; no universal percentage applies to every experiment. Methodological study.

The intuition is that observations resembling one another because of shared origin do not provide the same information as independent observations. Treating them equally can make uncertainty appear smaller than the design justifies.

Zimmerman and colleagues documented dependence among cells from the same individual in single-cell studies. They also examined methods that incorporate that structure. Numerous cells help describe cellular heterogeneity, but do not automatically make each cell an independent individual when comparing groups of people. Methodological study.

Scope matters: the comparison between statistical procedures belongs to the conditions evaluated in the article. It does not establish that one method is always superior for every cellular dataset.

Options can include summarizing observations at the relevant level or using a model representing their groupings. The choice depends on what is being estimated and how the data were generated. Averaging everything without considering that question can also lose important information.

As a hypothetical example, one effect applied to organisms and another applied to derived preparations may require different levels within the same analysis. A hierarchical model recognizes dependence but cannot manufacture independent groups that were never present in the experiment.

Look for origin identifiers, the number of units per condition and the number of observations within each unit. The analysis description should explain how it incorporated those relationships. Reporting only the total number of cells or images leaves the basis of inference incomplete.

Repeated measurements remain valuable: they describe within-unit variation and can improve estimation for each unit. The error arises when that information is used to support an independence or generalization that the design does not justify.