Summary Statistics

Each row must uniquely identify a genetic variant using standardized fields:

• rsID (dbSNP identifier): The reference SNP cluster ID, providing a universal label for the variant.
• Chromosome & Base-Pair Position: Genomic coordinates mapped to a specific reference genome build, such as GRCh37 (hg19) or GRCh38.
• Effect Allele (A1): The allele to which the reported effect size applies, also called the coded allele or reference allele in some software packages.
• Non-Effect Allele (A2): The alternative allele at the same locus, sometimes labeled as the other allele.

Consistent allele coding across cohorts is critical; strand flips or allele mismatches during meta-analysis or PRS calculation can reverse the direction of effect and destroy predictive accuracy.

The magnitude and direction of a variant's association with the trait are captured by the beta coefficient (β) and the odds ratio (OR).

• Beta (β): Represents the change in the quantitative trait per copy of the effect allele. For a binary trait, it is the log-odds ratio from logistic regression.
• Odds Ratio (OR): An exponentiated beta (OR = e^β) used in case-control studies. An OR > 1 indicates the effect allele increases disease risk.
• Standard Error (SE): Quantifies the uncertainty around the beta estimate, reflecting sample size and allele frequency. A smaller SE indicates a more precise estimate.

These fields are the direct inputs for the weighted allele scoring that defines a polygenic risk score.

The probability that the observed association is due to chance is reported as a p-value, often on a -log10 scale.

• P-value: Derived from the beta and its standard error, typically using a Wald test or score test. The genome-wide significance threshold is conventionally set at p < 5 × 10⁻⁸ to correct for multiple testing of ~1 million independent variants.
• -log10(p): A transformed metric where a value of 7.3 corresponds to the genome-wide threshold. This scale is used for visualization in Manhattan plots.
• Z-score: Calculated as β / SE, indicating the direction and strength of association in units of standard deviation.

P-values are the primary metric for the clumping and thresholding (C+T) method, where only variants below a user-defined significance threshold are retained for PRS calculation.

Quality control fields ensure that only well-measured variants are included in downstream analyses:

• Effect Allele Frequency (EAF): The frequency of the effect allele in the study population. This is essential for identifying rare variants and for harmonizing allele coding across datasets.
• Imputation INFO Score: A metric ranging from 0 to 1 that quantifies the confidence in statistically inferred genotypes. An INFO > 0.8 is a common filter to exclude poorly imputed variants.
• Minor Allele Frequency (MAF): The frequency of the less common allele. PRS analyses often filter out variants with MAF < 1% due to unstable effect size estimates.

Filtering on these metrics removes noisy variants that degrade PRS predictive performance.

Metadata fields describe the study population and statistical power:

• N (Sample Size): The total number of individuals included in the analysis for that variant. Larger sample sizes yield smaller standard errors and more robust p-values.
• N Cases / N Controls: For binary traits, the breakdown of affected and unaffected individuals. This ratio impacts the power to detect associations.
• Cohort / Study Identifier: In a meta-analysis file, a column indicating the originating study for each variant, enabling leave-one-out analyses.

Sample size directly determines the statistical power of the GWAS and, consequently, the maximum predictive ceiling of any PRS derived from it.

The effect allele is the specific nucleotide base to which the estimated effect size is assigned. The beta coefficient represents the log-odds change in disease risk (for binary traits) or the unit change in a quantitative trait per additional copy of the effect allele. Together, they form the core directional signal extracted from GWAS. A positive beta indicates risk-increasing, while a negative beta indicates a protective effect. Accurate allele alignment between the summary statistics and the target genotype data is critical; strand flips or allele mismatches will invert the direction of effect and catastrophically degrade PRS performance.

The standard error quantifies the precision of the estimated beta coefficient, reflecting sampling variability. It is inversely proportional to the square root of the sample size. The p-value is derived from the ratio of beta to its standard error (the Z-score) and tests the null hypothesis that the variant has no effect. Key considerations:

• Larger sample sizes shrink standard errors, yielding more significant p-values
• P-values are often reported on a -log10 scale in Manhattan plots
• Winner's Curse: selecting variants by p-value threshold systematically inflates effect estimates, requiring correction

The effect allele frequency (EAF) reports how common the effect allele is in the GWAS cohort. This is essential for downstream quality control because:

• Rare variants (MAF < 1%) have unstable effect estimates with large standard errors
• Mismatches between EAF in the base data and the target population can indicate allele coding errors
• The INFO score (imputation quality metric) ranges from 0 to 1 and measures confidence in statistically inferred genotypes. Filtering on INFO > 0.8 is standard practice to remove poorly imputed variants that introduce noise into PRS calculations.

The genomic inflation factor (λ) is a global quality control metric computed as the ratio of the median observed chi-squared test statistic to the expected median under the null distribution. A λ value of 1.0 indicates no systemic bias. Values substantially above 1.0 suggest:

• Population stratification: systematic allele frequency differences between cases and controls due to ancestry
• Cryptic relatedness: unrecognized familial relationships inflating test statistics
• Polygenic signal: for highly polygenic traits with large sample sizes, moderate inflation (λ < 1.1) is expected and does not indicate confounding. LD score regression can distinguish true polygenicity from bias.

Statistical power in GWAS is primarily driven by sample size. The effective sample size (Neff) accounts for imbalanced case-control ratios and is calculated as Neff = 4 / (1/Ncases + 1/Ncontrols). For meta-analyses combining multiple cohorts, Neff approximates the equivalent sample size of a balanced study. Key implications:

• Larger Neff yields smaller standard errors and more precise beta estimates
• PRS predictive accuracy scales with the discovery GWAS sample size
• Meta-analysis summary statistics should report per-variant sample sizes, as not all variants are measured in every contributing cohort

Strand ambiguity occurs when A/T or C/G SNPs cannot be resolved by allele matching alone, as the forward-strand allele is identical to the reverse-strand complement. This creates a critical data integration challenge when aligning summary statistics to a target genotype panel. Allele harmonization is the systematic process of:

• Matching alleles by identity, then by complement
• Removing ambiguous SNPs where allele frequency cannot disambiguate orientation
• Flipping beta signs when the effect allele in the summary statistics matches the reference panel's non-effect allele Failure to harmonize properly introduces sign errors that destroy PRS predictive power.