Skip to content

Study design and phenotype definition

Careful study design and clean phenotype definition reduce bias, improve power, and make results interpretable across cohorts.

Ascertainment bias

Ascertainment bias happens when sample selection depends on the trait or related factors.

Common sources - Clinic-based recruitment for disease studies - Enrichment of extreme phenotypes - Volunteer or biobank participation bias

Mitigation - Document inclusion criteria and recruitment pathways - Include sampling weights or covariates that capture selection - Perform sensitivity analyses with alternative inclusion rules

Case/control selection

Principles - Use consistent diagnostic criteria across sites - Exclude ambiguous cases or controls with related conditions - Avoid "super controls" that are not representative of the source population

Matching and balance - Match or adjust for age, sex, ancestry, and study site - Track case/control ratio for power and calibration

Covariate strategy

Typical covariates - Age, sex, study site or batch - Genetic ancestry PCs - Technical variables (array, center, processing date)

Guidelines - Use a prespecified covariate list - Avoid overfitting with excessive covariates - Report covariates in methods and summary statistics

Trait transformations

When to transform - Skewed quantitative traits - Heavy-tailed distributions or outliers

Common options - Log or Box-Cox transform - Rank-based inverse normal transformation (INT)

See 26_normalization for recommended workflows.

Phenotype QC

Core checks - Range and unit validation - Outlier detection - Duplicate records and inconsistent values

Example rules - Remove biologically implausible values - Flag values outside 4-6 SD for review

Missingness handling

Patterns to evaluate - Missing not at random (MNAR) by case/control status - Missingness by site or batch

Approaches - Exclude individuals with high missing phenotype rate - Multiple imputation for covariates if justified - Sensitivity analysis for MNAR risks

Multi-trait and longitudinal outcomes

Multi-trait - Use multivariate models when traits are correlated - Adjust for multiple testing or use joint tests

Longitudinal - Prefer mixed models with random effects - Model time-varying covariates and visits

Checklist

  • Define inclusion/exclusion and outcome labels before analysis
  • Predefine covariates and transformation rules
  • Record all phenotype QC filters
  • Confirm that missingness is not trait-dependent

References

  • Pirastu, N., et al. (2021). Genetic analyses identify widespread sex-differential participation bias. Nature Genetics, 53, 663-671. https://doi.org/10.1038/s41588-021-00846-7
  • Hernan, M. A., & Robins, J. M. (2020). Causal Inference: What If. Chapman & Hall/CRC.