Lesson 2: Study Design and Phenotypes

  • ID: GWAS-L02
  • Type: Conceptual + Implementation
  • Audience: Public
  • Theme: Trait definition, covariates, and study structure

What This Lesson Does

Before quality control and association testing, a GWAS must be clear about three things:

  1. What is the phenotype (trait) and how is it measured?
  2. What is the study design and sampling logic?
  3. Which covariates must be accounted for to avoid misleading associations?

A GWAS can be technically correct and still be scientifically weak if the phenotype is ambiguous or the design is unclear.

Core Principle

Genetic associations are interpreted relative to the phenotype definition.

If the phenotype is noisy, misclassified, or inconsistently measured, the analysis may produce:

  • Reduced power (true signals are harder to detect)
  • False positives (associations driven by measurement artifacts)
  • Confusing effect sizes (not comparable across samples)

The goal is to define the trait in a way that is measurable, consistent, and interpretable.

Study Design Patterns in GWAS

Common GWAS study designs include:

Quantitative trait GWAS

  • Continuous outcomes (for example BMI, blood pressure, cholesterol)
  • Usually modeled with a linear model
  • Interpretation is in units of trait change per allele

Case-control GWAS

  • Binary outcomes (case vs control)
  • Usually modeled with logistic regression
  • Interpretation is odds ratio per allele

This free track uses a quantitative trait so the modeling remains transparent.

Load the Demo Data

pheno <- read.csv("data/demo-phenotype.csv", stringsAsFactors = FALSE)
covar <- read.csv("data/demo-covariates.csv", stringsAsFactors = FALSE)

dim(pheno)
[1] 200   2
dim(covar)
[1] 200   5
head(pheno)
   sample_id    trait
1 sample-001 1.484454
2 sample-002 2.485771
3 sample-003 1.676582
4 sample-004 1.963092
5 sample-005 2.226643
6 sample-006 2.206656
head(covar)
   sample_id age sex        PC1         PC2
1 sample-001  38   F -0.7152422 -0.91065959
2 sample-002  42   F -0.7526890 -1.24657225
3 sample-003  64   F -0.9385387  0.25830482
4 sample-004  46   F -1.0525133 -0.03065893
5 sample-005  47   F -0.4371595 -1.46962895
6 sample-006  66   F  0.3311792  0.12258954

Plot Setup

source("scripts/R/cdi-plot-theme.R")
library(ggplot2)

Ensure Samples Match

stopifnot(all(pheno$sample_id %in% covar$sample_id))
stopifnot(all(covar$sample_id %in% pheno$sample_id))

df <- merge(pheno, covar, by = "sample_id")

dim(df)
[1] 200   6
head(df)
   sample_id    trait age sex        PC1         PC2
1 sample-001 1.484454  38   F -0.7152422 -0.91065959
2 sample-002 2.485771  42   F -0.7526890 -1.24657225
3 sample-003 1.676582  64   F -0.9385387  0.25830482
4 sample-004 1.963092  46   F -1.0525133 -0.03065893
5 sample-005 2.226643  47   F -0.4371595 -1.46962895
6 sample-006 2.206656  66   F  0.3311792  0.12258954

If sample IDs do not match, association testing can silently become incorrect.

Phenotype Sanity Checks

summary(df$trait)
   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
-0.3067  1.3695  2.0493  2.1273  2.8598  5.3643 
sum(is.na(df$trait))
[1] 0
ggplot(df, ggplot2::aes(x = trait)) +
  cdi_geom_histogram(bins = 30, colored = TRUE) +
  cdi_scale_histogram_fill() +
  ggplot2::labs(
    title = "Trait distribution (demo data)",
    subtitle = "Sanity check before modeling",
    x = "Trait value",
    y = "Count"
  ) +
  cdi_theme()

Interpretation

The trait appears approximately unimodal with moderate right skew.
There are no extreme outliers that would dominate a linear model.

At this stage, the question is not whether the distribution is perfectly normal.
The question is whether the trait scale is stable enough to support:

  • linear modeling assumptions
  • interpretable effect sizes
  • comparability across individuals

Minor skew does not invalidate GWAS.
Severe skew or extreme outliers would require transformation or sensitivity checks.

Principal Components and Structure

pal <- cdi_palette()

ggplot(df, ggplot2::aes(x = PC1, y = trait)) +
  ggplot2::geom_point(alpha = 0.85, color = pal$teal_light) +
  ggplot2::geom_smooth(method = "lm", se = FALSE, color = pal$highlight, linewidth = 0.9) +
  ggplot2::labs(
    title = "Trait vs PC1",
    subtitle = "If this trend is strong, unadjusted GWAS will inflate",
    x = "PC1",
    y = "Trait"
  ) +
  cdi_theme()

The visible positive association between PC1 and the trait suggests that population structure contributes to trait variation.

If SNP allele frequencies also vary along PC1, then unadjusted association tests will partially capture ancestry rather than biology.

This produces:

  • inflated test statistics
  • excess small p-values
  • false positive associations

Adjusting for principal components is part of study design, not a cosmetic correction.

pal <- cdi_palette()

ggplot(df, ggplot2::aes(x = PC2, y = trait)) +
  ggplot2::geom_point(alpha = 0.85, color = pal$teal_light) +
  ggplot2::geom_smooth(method = "lm", se = FALSE, color = pal$highlight, linewidth = 0.9) +
  ggplot2::labs(
    title = "Trait vs PC2",
    subtitle = "Structure-related covariation should be accounted for",
    x = "PC2",
    y = "Trait"
  ) +
  cdi_theme()

The association with PC2 appears weaker but still non-zero.

Even modest correlations with principal components can meaningfully affect genome-wide tests, because millions of variants are evaluated. Small structural biases can accumulate into large-scale inflation.

Quantifying Structure-Trait Correlation

cor_pc1 <- cor(df$trait, df$PC1)
cor_pc2 <- cor(df$trait, df$PC2)

cor_pc1
[1] 0.1651406
cor_pc2
[1] 0.144856

Correlation values quantify the visible trend. Even modest correlations can generate measurable inflation in GWAS if not adjusted.

What This Means for the Rest of the Guide

From this point forward:

  • All association models will adjust for age, sex, PC1, and PC2.
  • Any signal that disappears after adjustment will be interpreted cautiously.
  • We will distinguish between ancestry-driven signal and locus-specific signal.

This discipline prevents us from confusing structure with biology.