GWAS Summary Statistics

The core quantitative output for each single nucleotide polymorphism (SNP), representing the magnitude and direction of its association with the phenotype. For quantitative traits, this is typically the beta coefficient from a linear regression, indicating the change in trait value per copy of the effect allele. For binary traits, it is often the log odds ratio from logistic regression. Accurate effect sizes are critical for calculating causal estimates in two-sample Mendelian randomization.

The standard error (SE) quantifies the uncertainty around the effect size estimate. It is essential for downstream analyses because it determines the weight each genetic variant receives in meta-analytic methods like inverse-variance weighting (IVW). A smaller standard error indicates a more precise estimate, giving that variant greater influence in the combined causal effect calculation. Without SEs, it is impossible to distinguish true signals from statistical noise.

The p-value tests the null hypothesis that a genetic variant has no association with the trait. GWAS apply a stringent genome-wide significance threshold, conventionally p < 5 × 10⁻⁸, to correct for testing millions of independent variants. Summary statistics must report p-values to enable the selection of valid instrumental variables for Mendelian randomization. Variants that do not meet this threshold are typically excluded to avoid weak instrument bias.

Each variant record must unambiguously identify the effect allele (the allele whose effect is being measured) and the non-effect allele. Mismatched alleles or ambiguous strand orientation (A/T or C/G SNPs) between exposure and outcome datasets are a primary source of error in two-sample MR. Harmonization algorithms align effect alleles and flip effect directions to ensure consistent referencing, a non-negotiable quality control step before causal analysis.

The effective sample size (N) for each variant is crucial for calculating statistical power. For imputed genotypes, the INFO score or Rsq metric measures the confidence in the imputed genotype, ranging from 0 to 1. Filtering on INFO > 0.8 or 0.9 is standard practice to remove poorly imputed variants that introduce measurement error. Larger sample sizes yield more precise effect estimates, strengthening the instruments used in Mendelian randomization.

Precise variant identification requires chromosome, base-pair position (build GRCh37/hg19 or GRCh38/hg38), and reference SNP ID (rsID). This information is used to calculate the linkage disequilibrium (LD) structure between variants. In Mendelian randomization, instruments must be pruned for LD (e.g., r² < 0.001) to ensure independence. Clumping algorithms use these coordinates and LD reference panels to select a single representative variant per genomic locus.

The core components of a summary statistic file. The effect allele (or alt allele) is the specific nucleotide variant being tested for association. The beta (β) represents the estimated effect size—the change in the phenotype per copy of the effect allele. For quantitative traits, this is a linear regression coefficient; for binary traits, it is often a log odds ratio. Always verify the coding strand and whether the beta is reported relative to the effect allele or the reference allele.

The standard error (SE) quantifies the uncertainty in the estimated effect size (beta). A smaller SE indicates higher precision, typically driven by larger sample sizes or higher variant frequency. The SE is critical for downstream analyses like inverse-variance weighting (IVW) in Mendelian randomization, where variants with larger SEs are down-weighted. Always check for inflated SEs that may indicate genotyping errors or poor imputation quality.

The p-value tests the null hypothesis that the variant has no association with the trait. In GWAS, the standard genome-wide significance threshold is 5 × 10⁻⁸, correcting for approximately one million independent tests. Summary statistics often include p-values from Wald tests or score tests. Be cautious of p-values near zero in logistic regression, which may indicate complete separation or model failure rather than true biological signal.

The minor allele frequency (MAF) is the frequency of the less common allele in the study population. It is essential for quality control: variants with very low MAF (e.g., < 1%) have unstable effect estimates and inflated standard errors. Many summary statistic formats include the effect allele frequency (EAF). Mismatches in EAF between exposure and outcome datasets in two-sample Mendelian randomization can indicate strand alignment errors or population differences.

Most GWAS summary statistics include imputed genotypes. The INFO score (or Rsq) measures the quality of imputation, ranging from 0 to 1. A low INFO score (e.g., < 0.3) indicates high uncertainty in the imputed genotype, which biases effect estimates toward the null and inflates standard errors. Standard practice is to filter variants with INFO < 0.6 or 0.8 before downstream analysis. Always check the imputation reference panel (e.g., 1000 Genomes, HRC, TOPMed) to assess coverage.

The effective sample size (Neff) accounts for imbalanced case-control ratios in binary trait GWAS. It is calculated as Neff = 4 / (1/N_cases + 1/N_controls). Using raw sample size instead of Neff can overstate statistical power. Summary statistics should clearly report the number of cases and controls. For meta-analyses, the per-variant sample size may vary due to differential missingness across genotyping arrays.