Statistical & Population Genetics Foundational to Advanced 18 min read

Genome-Wide Association Studies (GWAS): The Mathematical Engine of Trait Mapping

By Kuan-Hao Chao·Published August 2026

How do 3 billion base pairs of human DNA translate into phenotypic diversity, disease predisposition, and complex human traits? Genome-Wide Association Studies (GWAS) provide the foundational statistical and algorithmic bridge connecting population-scale genotype matrices to quantitative traits and clinical risk.

Executive Summary & Core Mechanisms

GWAS tests millions of genetic variants across cohorts of tens to hundreds of thousands of individuals using additive linear/logistic regression. To discover genuine biological signals, a GWAS must rigorously solve two major statistical challenges: population stratification confounding (via Ancestry PCA or Linear Mixed Models) and massive multiple testing (via the genome-wide Bonferroni threshold $p < 5 \times 10^{-8}$ ).

1. The Genotype Dosage Matrix & Single-Variant Testing

At the core of every GWAS is a high-dimensional cohort dataset containing $N$ individuals assayed across $M$ single nucleotide polymorphisms (SNPs). For an individual $i$ and variant $j$ , genotype data is encoded as an additive dosage:

Additive Genotype Dosage Encoding

G_{ij} \in \{0, 1, 2\} \quad \longrightarrow \quad \text{Count of Minor (Effect) Alleles}

0 = \text{Ref/Ref (Homozygous reference)}

1 = \text{Ref/Alt (Heterozygous)}

2 = \text{Alt/Alt (Homozygous alternative)}

While the additive model is the standard assumption in human genetics due to parsimony and power, alternative genetic architectures can also be tested:

Additive Model: $G_{ij} \in \{0, 1, 2\}$ (each minor allele linearly shifts phenotype by $\beta_j$ ).
Dominant Model: $G_{ij} \in \{0, 1, 1\}$ (one copy is sufficient to confer full phenotypic effect).
Recessive Model: $G_{ij} \in \{0, 0, 1\}$ (requires two copies of the minor allele).

Quantitative Traits: Ordinary Least Squares (OLS) Regression

For a continuous quantitative phenotype $\mathbf{y} \in \mathbb{R}^N$ (e.g. standing height, LDL cholesterol, or fasting glucose), we formulate a single-variant Ordinary Least Squares (OLS) regression model:

Single-Variant Linear Regression Model

y_i = \alpha + \beta_j X_{ij} + \sum_{k=1}^K \gamma_k Z_{ik} + \epsilon_i, \quad \epsilon_i \sim \mathcal{N}(0, \sigma^2)

X_{ij}

= Additive genotype dosage |

\beta_j

= Effect size per minor allele |

Z_{ik}

= Covariates (Age, Sex, Ancestry PCs)

Using the Frisch-Waugh-Lovell theorem to project out the covariate matrix $\mathbf{Z}$ , the analytical Ordinary Least Squares estimate of the effect size $\hat{\beta}_j$ and its variance are given by:

OLS Estimator, Standard Error & Statistical Power

\hat{\beta}_j = \frac{\sum_{i=1}^N (X_{ij} - \bar{X}_j)(y_i - \bar{y})}{\sum_{i=1}^N (X_{ij} - \bar{X}_j)^2} = \frac{\text{Cov}(X_j, y)}{\text{Var}(X_j)}

\text{Var}(\hat{\beta}_j) = \frac{\sigma^2}{2 N p_j (1 - p_j)} \implies \text{SE}(\hat{\beta}_j) = \frac{\sigma}{\sqrt{2 N p_j (1 - p_j)}}

t_j = \frac{\hat{\beta}_j}{\text{SE}(\hat{\beta}_j)} \sim \text{Student's } t(df = N - K - 2) \quad \longrightarrow \quad p_j = 2 \cdot [1 - \Phi(|t_j|)]

Under Hardy-Weinberg equilibrium with minor allele frequency

p_j = \text{MAF}

, the genotype variance is

\text{Var}(X_j) = 2p_j(1 - p_j)

. Notice that discovery power is directly proportional to

\sqrt{N \cdot p_j (1 - p_j) \cdot \beta_j^2}

Case-Control Binary Traits: Logistic Regression & Odds Ratios

For binary clinical disease status $y_i \in \{0, 1\}$ (e.g. Type 2 Diabetes case vs. control), we fit a Logistic Regression model via Maximum Likelihood Estimation (MLE):

Logistic Regression & Odds Ratio (OR)

\ln\left(\frac{P(y_i = 1)}{1 - P(y_i = 1)}\right) = \alpha + \beta_j X_{ij} + \sum_{k=1}^K \gamma_k Z_{ik}

\text{Odds Ratio (OR)} = e^{\beta_j}, \quad 95\% \text{ CI} = \left[ e^{\hat{\beta}_j - 1.96 \cdot \text{SE}}, \; e^{\hat{\beta}_j + 1.96 \cdot \text{SE}} \right]

2. The Population Stratification Conundrum ("The Chopsticks Gene")

The most notorious pitfall in statistical genetics is population stratification. In 1994, Eric Lander and Nicholas Schork described a classic thought experiment: imagine running a GWAS for chopstick dexterity in a mixed cohort of European and East Asian individuals living in San Francisco.

The Chopsticks Problem & Spurious Association

Chopstick usage is a cultural trait strongly correlated with East Asian ancestry. Simultaneously, thousands of completely unrelated alleles across the genome have divergent frequencies between Europe and East Asia due to historical genetic drift. A naive regression will find hundreds of thousands of false-positive "chopstick genes" across every chromosome!

Mathematically, omitted ancestry $A$ confounds the relationship because $\text{Cov}(X, A) \neq 0$ and $\text{Cov}(y, A) \neq 0$ :

Mathematical Decomposition of Ancestry Confounding Bias

\mathbb{E}[\hat{\beta}_{\text{naive}}] = \beta_j + \frac{\text{Cov}(X_j, A)}{\text{Var}(X_j)} \cdot \gamma_{\text{Ancestry}}

When ancestry effect

\gamma_{\text{Ancestry}} \neq 0

, the naive estimator is heavily biased away from 0 even if the true genetic effect

\beta_j = 0

Ancestry Principal Component Analysis (EIGENSTRAT)

To eliminate stratification, Price et al. (2006) introduced EIGENSTRAT. We construct the normalized genotype matrix $\mathbf{G}^*$ where each entry is standardized by its allele frequency $p_j$ :

Genotype Standardization & SVD Decomposition

G^*_{ij} = \frac{G_{ij} - 2p_j}{\sqrt{2p_j(1 - p_j)}} \quad \implies \quad \mathbf{G}^* = \mathbf{U} \mathbf{\Sigma} \mathbf{V}^T

\mathbf{PC}_k = \mathbf{U}_k \cdot \sigma_k \quad (k = 1, \dots, 10)

Top eigenvectors

\mathbf{PC}_1, \dots, \mathbf{PC}_{10}

capture continental and sub-continental ancestry axes. Including them as covariates in regression orthogonalizes the association test.

Linear Mixed Models (LMM: EMMAX / BOLT-LMM)

When cohorts contain cryptic relatedness, family structures, or complex population structure, Linear Mixed Models (LMM) replace fixed PC covariates with a random polygenic background effect:

Linear Mixed Model Formulation

\mathbf{y} = \mathbf{X}\beta + \mathbf{g}_j \gamma_j + \mathbf{u} + \mathbf{\epsilon}, \quad \mathbf{u} \sim \mathcal{N}(0, \sigma_g^2 \mathbf{K}), \quad \mathbf{\epsilon} \sim \mathcal{N}(0, \sigma_e^2 \mathbf{I})

\mathbf{K} = \frac{1}{M} \mathbf{G}^* (\mathbf{G}^*)^T \quad \text{(Empirical Genetic Relationship Matrix)}

\sigma_g^2

= Genetic variance component |

\sigma_e^2

= Environmental variance |

h^2_{\text{SNP}} = \frac{\sigma_g^2}{\sigma_g^2 + \sigma_e^2}

(Narrow-sense SNP Heritability)

Genomic Inflation Factor (λ_GC) & Q-Q Plot Diagnostics

To audit whether a GWAS suffers from uncorrected confounding, we calculate the Genomic Inflation Factor ( $\lambda_{GC}$ ):

Genomic Inflation Factor Formula

\lambda_{GC} = \frac{\text{median}(\chi^2_{\text{obs}, 1}, \dots, \chi^2_{\text{obs}, M})}{0.454936}

0.454936

is the theoretical median of a 1-degree-of-freedom Chi-Square distribution

\chi^2(df = 1)

under the null hypothesis. A pristine study yields

\lambda_{GC} \le 1.05

3. Genome-Wide Multiple Testing & The Manhattan Plot

A typical GWAS evaluates between $1{,}000{,}000$ and $10{,}000{,}000$ imputed genetic variants simultaneously. If a standard statistical threshold of $\alpha = 0.05$ were used, testing $10^6$ null variants would produce 50,000 false-positive discoveries purely by chance.

Derivation of the Genome-Wide Significance Threshold

\alpha_{\text{GWAS}} = \frac{\alpha_{\text{FWER}}}{M_{\text{eff}}} = \frac{0.05}{1{,}000{,}000} = 5 \times 10^{-8}

-\log_{10}(5 \times 10^{-8}) \approx 7.301

M_{\text{eff}} \approx 10^6

represents the effective number of independent correlation blocks across the human genome due to Linkage Disequilibrium (LD).

To visualize all 22 autosomes simultaneously, GWAS results are plotted as a Manhattan Plot: genomic coordinates are aligned along the horizontal x-axis from Chromosome 1 to 22, while the vertical y-axis displays $-\log_{10}(P)$ . Strong association signals rise above the background noise like skyscrapers on the Manhattan skyline.

4. Linkage Disequilibrium (LD), Recombination & Fine-Mapping

A common misconception is that a GWAS skyscraper peak points to the exact single causal nucleotide. Because chromosomes are inherited in chunks during meiosis, alleles close to each other are transmitted together in Linkage Disequilibrium (LD) blocks.

Linkage Disequilibrium Metrics (D, r², D')

D = p_{AB} - p_A \cdot p_B, \quad r^2 = \frac{D^2}{p_A (1 - p_A) p_B (1 - p_B)}, \quad D' = \frac{D}{D_{\max}}

r^2 = 1.0

represents complete correlation (perfect proxy SNPs).

r^2 \ge 0.8

defines high-confidence LD candidate tags.

LD Score Regression (LDSC): Polygenicity vs. Confounding

How do we distinguish whether a slightly elevated $\lambda_{GC} = 1.15$ is caused by genuine polygenicity (thousands of real small-effect variants) or residual confounding? LD Score Regression (Bulik-Sullivan et al., 2015) exploits the fact that variants in high-LD regions tag more genetic variance:

LD Score Regression Equation

\mathbb{E}[\chi^2_j] = 1 + \frac{N h^2_{\text{SNP}}}{M} \cdot \ell_j + N a

\ell_j = \sum_k r^2_{jk}

is the LD score of SNP

j

. The slope reveals true heritability

h^2_{\text{SNP}}

, while the intercept

1 + Na

measures pure uncorrected confounding (

a = 0

under perfect control)!

Statistical Fine-Mapping & 95% Credible Sets

Methods like SuSiE (Sum of Single Effects) and PAINTOR use Bayesian posterior modeling to construct a 95% Credible Set — the minimal set of variants that contains the true causal mutation with 95% certainty: $\sum_{j \in \text{CS}_{95\%}} \text{PIP}_j \ge 0.95$ Integrating functional epigenomics (eQTLs, ChIP-seq, OpenSpliceAI deep learning splice models) pinpoints the causal molecular mechanism.

5. Polygenic Risk Scores (PRS) & Clinical Translation

For complex polygenic conditions (e.g. coronary artery disease, breast cancer, type 2 diabetes), no single variant causes the disease. Instead, thousands of subtle variants across the genome contribute small additive increments to an individual's overall liability:

Polygenic Risk Score (PRS) Formulation

\text{PRS}_i = \sum_{j=1}^M \hat{\beta}_j \cdot G_{ij}

\hat{\beta}_j

= Marginal effect weight from discovery GWAS |

G_{ij}

= Individual's dosage for variant

j

Clinical Utility of Polygenic Risk Scores

Individuals falling in the top 5% of the PRS distribution often face disease risks equivalent to rare, monogenic high-penetrance mutations (such as LDLR in familial hypercholesterolemia or BRCA1 in breast cancer). PRS enables personalized early intervention, lifestyle modifications, and proactive clinical screening decades before clinical symptom onset.

6. Summary Cheat-Sheet & Key Invariants

GWAS Component	Mathematical Formulation	Intuition & Purpose	Diagnostic Benchmark
Additive Dosage	$G_{ij} \in \{0, 1, 2\}$	Count of minor effect alleles	MAF ≥ 1%, HWE p > 10⁻⁶
Association Test	$t_j = \hat{\beta}_j / \text{SE}(\hat{\beta}_j)$	Evaluates $H_0: \beta_j = 0$	OLS for quantitative, Logistic for binary
Ancestry PCA	$\mathbf{G}^* = \mathbf{U} \mathbf{\Sigma} \mathbf{V}^T$	Decomposes continental ancestry	Top 10–20 PCs as covariates
Linear Mixed Model	$\mathbf{y} = \mathbf{X}\beta + \mathbf{u} + \mathbf{\epsilon}$	Random polygenic effect with GRM $\mathbf{K}$	Estimates SNP heritability $h^2_{\text{SNP}}$
Genomic Inflation	$\lambda_{GC} = \frac{\text{median}(\chi^2)}{0.455}$	Measures confounding bias	Controlled: $\lambda_{GC} \le 1.05$
Multiple Testing	$\alpha = \frac{0.05}{10^6} = 5 \times 10^{-8}$	Bonferroni correction for 10⁶ LD blocks	Threshold: $p < 5 \times 10^{-8}$
Linkage Disequilibrium	$r^2 = \frac{D^2}{p_A q_A p_B q_B}$	Measures allele co-inheritance	Fine-mapping tag: $r^2 \ge 0.8$
Polygenic Score	$\text{PRS} = \sum \hat{\beta}_j \cdot G_j$	Aggregates genome-wide polygenic risk	Clinical high risk: Top 5% quintile