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

Biological statistics

Use this skill when a statistical result depends on how biological observations entered the dataset, which observations contributed to the denominator, or to whom the result is intended to generalise.

The central question is:

What process generated these observations, and to which population or biological units does the inference actually apply?

Core rules

  • Distinguish source population, eligible population, recruited population, measured population, analysed population, and target population when they differ.
  • An estimate inherits the process that selected the observations used to calculate it.
  • Preserve eligibility, referral, recruitment, consent, survival, follow-up, measurement, exclusion, attrition, and QC mechanisms when they can affect inclusion.
  • A control is defined by its ascertainment criteria, not by the label control.
  • Control does not automatically mean healthy, phenotype-negative, disease-free for life, population-representative, or exchangeable with cases except for the exposure of interest.
  • Do not use a denominator without establishing what observations were eligible to contribute to it.
  • Selection related to exposure, outcome, or common causes can change observed associations.
  • Treat missingness and technical QC as potential selection mechanisms when exclusion can correlate with biology, ancestry, disease severity, sample quality, site, or other relevant variables.
  • Distinguish internal validity from transportability to another target population.
  • Do not assume a predictive model, frequency estimate, or treatment effect transports unchanged to a different population.
  • Preserve relatedness, repeated measurements, clustering, and hierarchical dependence when they affect effective independence.
  • Population structure and ancestry can alter genetic associations and frequency estimates; adjustment strategy must match the inferential target.
  • Statistical precision does not repair an undefined target population or selection mechanism.

AI behaviour

Before interpreting or generalising an estimate, establish:

  1. what population generated the eligible observations,
  2. how subjects or samples became included,
  3. who or what was excluded before and after measurement,
  4. what the denominator counts,
  5. what case, control, exposed, or other group labels actually mean,
  6. whether missingness or QC changed the analysed population,
  7. whether observations are independent,
  8. which population or biological units the requested conclusion targets.

When a cohort is recruited through a specialist clinic, registry, biobank, hospital, referral network, or voluntary study, do not silently reinterpret it as a random sample of the general population.

When controls were not systematically assessed for the phenotype, do not call them phenotype-negative.

When age, treatment, survival, or follow-up affects manifestation, do not convert non-observation into lifelong absence.

When a result is transported to another population, state the assumptions or limitations rather than presenting transportability as automatic.

References

Read the relevant reference when the task depends on it:

Table of contents