- Association models are statistical methods used in bioinformatics and genetic research to investigate relationships between genetic variants and phenotypic characteristics. They are central to genome-wide association studies (GWAS), which examine whether genetic markers, particularly single-nucleotide polymorphisms (SNPs), are associated with measurable traits or disease outcomes. The choice of association model depends on the type of phenotype, study design, genetic inheritance assumptions, population structure, relatedness among individuals, and other relevant factors. Quantitative and binary traits require different statistical approaches because they represent different types of biological outcomes.
- A quantitative trait is a characteristic measured numerically across a continuous or approximately continuous range. Examples include height, body weight, blood pressure, cholesterol concentration, blood glucose, gene-expression levels, milk production, crop yield, and enzyme activity. These traits may be influenced by many genetic variants, environmental conditions, and interactions between genetic and environmental factors. A binary trait has two possible categories, commonly represented as affected or unaffected, diseased or healthy, or case or control. Examples include the presence or absence of a disease, a diagnostic outcome, or survival status at a specified assessment time. The distinction between these phenotype types is important because the statistical assumptions and interpretation of their association models differ.
- In a typical GWAS, genetic variants are tested individually for association with a phenotype. The genotype at each biallelic SNP is often coded according to the number of copies of the effect allele: 0 for homozygous reference, 1 for heterozygous, and 2 for homozygous alternate. This coding represents an additive genetic model, in which each additional copy of the effect allele is assumed to contribute the same amount to the phenotype on the model’s scale. The association analysis estimates the relationship between genotype and phenotype while accounting for relevant covariates and potential confounding factors.
- For quantitative traits, linear regression is one of the most widely used association models. It examines whether the mean phenotype changes according to the number of effect alleles carried by an individual. A basic linear regression model can be written as:
[
Y_i = \beta_0 + \beta_1G_i + \varepsilon_i
]
- Here, (Y_i) represents the quantitative phenotype of individual (i), (\beta_0) is the intercept, (G_i) is the genotype dosage, (\beta_1) is the estimated genetic effect, and (\varepsilon_i) is the residual error. Under additive genotype coding, (G_i) usually takes values of 0, 1, or 2 for directly observed diploid genotypes. For imputed variants, dosage may be a continuous value between 0 and 2, representing the expected number of effect alleles.
- The coefficient (\beta_1) estimates the average change in the phenotype associated with one additional copy of the effect allele, assuming the model is correctly specified. For example, if a SNP has an estimated effect of 1.5 cm on height, individuals carrying one additional copy of the effect allele are estimated to be, on average, 1.5 cm taller under the fitted additive model. This is a statistical association and does not by itself establish that the variant directly causes the change in height. The effect may reflect a causal variant, linkage disequilibrium with another variant, population structure, or other factors.
- In practical genetic analyses, linear regression usually includes additional covariates. These may include age, sex, ancestry-related principal components, study centre, genotyping batch, and other variables relevant to the phenotype and study design. An extended model can be expressed as:
[
Y_i = \beta_0 + \beta_1G_i +
\sum_{j=1}^{k}\gamma_j C_{ij} + \varepsilon_i
]
- Here, (C_{ij}) represents covariate (j) for individual (i), and (\gamma_j) is its estimated coefficient. Including appropriate covariates can reduce confounding and improve the precision of genetic effect estimates. However, covariates should be selected using biological knowledge and the study design. Automatically adjusting for every available variable may introduce bias, particularly if a variable lies on the causal pathway between genotype and phenotype.
- Linear regression assumes that the conditional mean of the phenotype is adequately represented by the specified model. Conventional statistical inference also relies on assumptions concerning residual behaviour, independence, and variance. In large samples, some departures from normal residual distributions may be less problematic for estimating mean effects, but heteroscedasticity, outliers, relatedness, and model misspecification can still affect standard errors and statistical tests. Researchers may use robust standard errors, transformations, generalized models, or mixed models when appropriate.
- Many quantitative traits have skewed distributions. Examples include triglyceride concentrations, hormone levels, metabolite measurements, and certain clinical biomarkers. Logarithmic, square-root, or other transformations may be used to improve model behaviour, depending on the data. After transformation, the genetic effect is interpreted on the transformed scale rather than directly in the original measurement units. Researchers should report the transformation and, where possible, provide an interpretation that is understandable in the biological context.
- Binary traits require models designed for outcomes with two possible values. Logistic regression is one of the most commonly used methods in case-control GWAS. Instead of modelling the phenotype directly as a continuous value, logistic regression models the probability that an individual belongs to one of the two outcome categories. The model uses the logit function, which transforms a probability between 0 and 1 into an unbounded scale.
- A basic logistic regression model is:
\beta_0+\beta_1G_i
]
- Here, (Y_i=1) represents the presence of the outcome, (P(Y_i=1)) is the probability of that outcome, (G_i) is the genotype dosage, and (\beta_1) is the estimated genetic effect on the log-odds scale. A positive coefficient indicates that each additional effect allele is associated with higher odds of the outcome, while a negative coefficient indicates lower odds, under the fitted model.
- The exponential of the genetic coefficient gives the odds ratio:
[
OR = e^{\beta_1}
]
- An odds ratio greater than 1 indicates higher odds of the outcome per additional effect allele, while an odds ratio below 1 indicates lower odds. An odds ratio of 1 indicates no estimated change in odds. For example, an odds ratio of 1.20 means that each additional effect allele is associated with approximately 20% higher odds of the outcome under an additive logistic model. This does not necessarily mean a 20% increase in disease probability or absolute risk, because odds and probabilities are different quantities.
- Logistic regression can also include covariates. An extended model can be written as:
\beta_0+\beta_1G_i+
\sum_{j=1}^{k}\gamma_j C_{ij}
]
- Covariates may include age, sex, ancestry-related principal components, study centre, environmental exposures, and other relevant factors. In case-control studies, the sampling design and the way cases and controls are selected should be considered when interpreting effect estimates. Logistic regression commonly estimates odds ratios, but the absolute disease probability in the general population cannot usually be obtained directly from an artificially sampled case-control dataset without additional information.
- The additive genetic model is commonly used in GWAS because many complex traits are influenced by the combined effects of numerous genetic variants. Under this model, the effect of carrying two copies of an allele is assumed to be twice the effect of carrying one copy relative to carrying none. However, genetic effects do not always follow an additive pattern. Dominance and recessive effects may also be relevant, and researchers may investigate alternative genotype models when justified by biological knowledge or study objectives.
- A dominant model groups individuals carrying at least one copy of the effect allele together. For a biallelic SNP, genotypes 0/1 and 1/1 are combined and compared with 0/0. A recessive model compares individuals with two copies of the effect allele (1/1) against those with 0/0 or 0/1. A genotypic model treats the three genotype categories separately and does not require an additive relationship between allele count and phenotype. These approaches can identify patterns that an additive model may not capture, but they involve additional statistical tests and may reduce power when genotype categories are small.
- For quantitative traits, the additive, dominant, recessive, and genotypic models can be implemented using linear regression. For binary traits, the same genotype coding strategies can be incorporated into logistic regression. Researchers should distinguish the genetic inheritance model from the statistical outcome model: linear or logistic regression describes the relationship between predictors and the phenotype, while additive, dominant, and recessive coding describes how genotype information enters that model.
- The association between a genetic variant and a phenotype is usually evaluated using a statistical hypothesis test. The null hypothesis commonly states that the genetic effect is zero, while the alternative hypothesis states that the effect differs from zero. A test statistic is calculated using the estimated genetic effect and its standard error. Common approaches include Wald tests, likelihood-ratio tests, and score tests. These tests may produce slightly different results, particularly in small samples or under complex model conditions.
- The Wald test evaluates whether an estimated regression coefficient differs significantly from zero relative to its standard error. A simplified statistic is:
[
Z = \frac{\hat{\beta}_1}{SE(\hat{\beta}_1)}
]
- Here, (\hat{\beta}_1) is the estimated genetic effect and (SE(\hat{\beta}_1)) is its standard error. Under suitable large-sample assumptions, the statistic approximately follows a standard normal distribution under the null hypothesis. The corresponding p-value indicates how incompatible the observed result is with the null model. It does not represent the probability that the null hypothesis is true or the probability that the association is causal.
- GWAS test very large numbers of genetic variants, often millions, so multiple-testing correction is essential. If thousands or millions of hypotheses are tested, some small p-values may occur by chance even when no true associations exist. The Bonferroni method controls the family-wise error rate by dividing the chosen significance level by the number of tests. False discovery rate (FDR) procedures, such as the Benjamini–Hochberg method, control the expected proportion of false discoveries among the results declared significant. In human GWAS, a p-value threshold around (5 \times 10^{-8}) is widely used for conventional genome-wide significance in many studies, although the appropriate threshold depends on the number of effective tests, the genetic architecture, and the analysis design.
- Population stratification is a major source of confounding in genetic association studies. If ancestry groups differ in allele frequencies and also differ in phenotype distributions for non-genetic or other genetic reasons, an apparent genetic association may arise without a direct causal relationship. Principal components from genome-wide genotype data are commonly included as covariates to account for broad ancestry-related structure. Mixed-model methods can additionally account for genetic relatedness and more complex population structure. The choice and number of principal components should be justified by the dataset and study design.
- Relatedness among individuals can also violate the independence assumptions of ordinary regression. Family members share genetic variants, so their observations are statistically correlated. If this correlation is ignored, standard errors may be underestimated and association results may become inflated. Linear mixed models and generalized linear mixed models can account for relatedness by incorporating a genetic relationship matrix or other random effects. These methods are particularly important in family-based cohorts, biobanks, animal breeding populations, and structured plant populations.
- Linear mixed models are widely used for quantitative traits when samples contain related individuals or complex population structure. A simplified model is:
[
Y = X\beta + Zu + \varepsilon
]
- Here, (Y) is the vector of observed phenotypes, (X) is the fixed-effect design matrix containing genotype and covariates, (\beta) is the vector of fixed-effect coefficients, (Z) is the random-effect design matrix, (u) represents random genetic or relatedness effects, and (\varepsilon) represents residual error. The covariance structure of (u) is often based on a genetic relationship matrix. Software such as GEMMA, EMMAX, BOLT-LMM, and other mixed-model tools is used for different quantitative-trait GWAS settings.
- Binary traits can also be analysed using mixed-model approaches. Generalized linear mixed models extend logistic regression by incorporating random effects, allowing researchers to account for relatedness, population structure, or other sources of correlation. These models may be computationally demanding in very large cohorts. Methods such as SAIGE and REGENIE have been developed to support large-scale genetic association analyses, including settings with imbalanced case-control ratios. Their methods, assumptions, and recommended workflows differ, so researchers should consult the relevant documentation before applying them.
- Case-control imbalance is a common challenge in binary-trait GWAS. In many studies, the number of controls is much larger than the number of cases, particularly for rare diseases. Severe imbalance can affect statistical calibration and reduce the effective information available for estimating genetic effects. The problem becomes more complicated when cases and controls differ in ancestry, age, recruitment site, or genotyping batch. Appropriate association methods, careful quality control, and sensitivity analyses are important for obtaining reliable results.
- Rare variants create additional challenges for both quantitative and binary trait association analyses. Individual rare variants may have too few minor allele copies to provide reliable estimates using standard single-variant regression. Researchers may therefore use gene-based or region-based tests that combine information across multiple variants. Methods such as burden tests, SKAT, and SKAT-O are used in rare-variant association studies. These approaches differ in their assumptions about the direction and distribution of variant effects. The choice of method depends on the genetic architecture and the biological question.
- Some phenotypes do not fit neatly into ordinary linear or binary logistic regression. Ordinal traits have ordered categories, such as mild, moderate, and severe disease. Ordinal logistic regression can model such outcomes while accounting for their ordered structure, subject to its assumptions. Count traits, such as the number of disease episodes or offspring, may require Poisson or negative-binomial regression. Time-to-event traits, such as age at disease onset or survival time, may be analysed using Cox proportional hazards models or other survival-analysis methods. Repeated measurements may require longitudinal models or mixed-effects regression. Selecting a model that matches the phenotype structure is essential for valid inference.
- Interaction analyses investigate whether the effect of a genetic variant changes according to another factor. Gene–environment interaction models, for example, test whether the association between genotype and phenotype differs across environmental exposures. A simplified quantitative interaction model can be written as:
[
Y_i = \beta_0+\beta_1G_i+\beta_2E_i+
\beta_3(G_i \times E_i)+\varepsilon_i
]
- Here, (G_i) represents genotype, (E_i) represents an environmental exposure, and (\beta_3) estimates the interaction effect on the model’s scale. For binary traits, an interaction term can be included in logistic regression. Interaction findings require careful interpretation because the scale of interaction, statistical power, measurement quality, and multiple testing can influence results. Large sample sizes are often needed to detect interactions reliably.
- The interpretation of genetic association results should distinguish statistical significance, effect size, biological significance, and causality. A very small genetic effect may achieve a highly significant p-value in a large sample, while a biologically meaningful effect may not reach statistical significance in a small study. Association does not establish causation because a detected variant may be correlated with a causal variant through linkage disequilibrium or may reflect residual confounding. Fine-mapping, functional annotation, molecular experiments, replication, and other evidence are often needed to investigate causal mechanisms.
- The PLINK 2 software package is widely used for GWAS involving quantitative and binary phenotypes. It supports linear and logistic regression, genotype dosage analysis, covariate adjustment, and several quality-control procedures. The following examples illustrate simplified association analyses using a dataset in PGEN format and a phenotype and covariate file.
- For a quantitative phenotype, linear regression can be performed using:
# GWAS for a quantitative phenotype
plink2 –pfile cohort \
–pheno phenotypes.txt \
–pheno-name height \
–covar covariates.txt \
–covar-name age,sex,PC1,PC2,PC3,PC4 \
–glm hide-covar \
–out height_gwas
- For a binary phenotype, logistic regression can be performed using:
# GWAS for a binary case-control phenotype
plink2 –pfile cohort \
–pheno phenotypes.txt \
–pheno-name disease_status \
–covar covariates.txt \
–covar-name age,sex,PC1,PC2,PC3,PC4 \
–glm hide-covar \
–out disease_gwas
- These commands are illustrative. Phenotype files must use the format and coding conventions expected by the installed PLINK 2 version. Binary case-control status must be encoded correctly, covariate values must be valid, and missing data must be handled appropriately. Researchers should also confirm which regression model is selected by the software for the phenotype coding and version being used. For cohorts with substantial relatedness, population structure, or case-control imbalance, specialized mixed-model methods may be more appropriate than ordinary regression.
- R provides additional tools for association analysis. The lm() function can fit linear regression models, while glm() can fit logistic regression models using a binomial family. The following example demonstrates a basic quantitative-trait analysis:
# Load phenotype and genotype data
data <- read.csv(“association_data.csv”)
# Fit a linear regression model
model <- lm(height ~ genotype + age + sex + PC1 + PC2, data = data)
# Display model results
summary(model)
A binary-trait model can be fitted using logistic regression:
# Fit a logistic regression model
model <- glm(
disease_status ~ genotype + age + sex + PC1 + PC2,
data = data,
family = binomial(link = “logit”)
)
# Display model results
summary(model)
# Convert the genotype coefficient to an odds ratio
exp(coef(model)[“genotype”])
- These examples demonstrate the statistical models rather than a complete GWAS pipeline. A genome-wide analysis must also address genotype quality control, population structure, multiple testing, relatedness, imputation uncertainty, and other study-specific requirements. For large-scale analyses, specialized software is generally more efficient than fitting separate models manually in R.
- Association results are commonly summarized using estimated genetic effects, standard errors, p-values, allele frequencies, sample sizes, and other quality metrics. Quantitative-trait results often report beta coefficients in the original or transformed phenotype units. Binary-trait results commonly report odds ratios and confidence intervals. Manhattan plots display association strength across genomic positions, while quantile–quantile plots compare observed and expected p-value distributions. These visualizations help identify association signals and assess possible inflation or systematic deviations, although they do not replace formal statistical evaluation.
- Replication is an important stage in association research. A genetic association discovered in one cohort should ideally be evaluated in an independent dataset with compatible phenotype definitions and genetic measurements. Replication helps determine whether the finding is robust and reduces the likelihood that it reflects chance, technical artifacts, or cohort-specific confounding. Differences in ancestry, allele frequencies, LD patterns, environmental conditions, and phenotype definitions can affect replication. Failure to replicate does not always mean that the original association is false, but it requires careful investigation.
- Association models are also widely used in animal and plant genetics. In animal breeding, quantitative-trait models investigate genetic associations with milk yield, growth, fertility, feed efficiency, disease resistance, and carcass characteristics. In plant genetics, models are used to identify variants associated with yield, flowering time, drought tolerance, disease resistance, and other agronomic traits. Population structure and relatedness are especially important in these settings because breeding populations may contain strong family relationships, breed groups, or distinct genetic lineages. Mixed models and carefully designed experiments are often essential for controlling these effects.
- For reproducible association analysis, researchers should document the phenotype definition, genotype coding, genetic inheritance model, covariates, sample inclusion criteria, genotype quality-control procedures, genome assembly, imputation reference panel, statistical software, and software version. The treatment of missing values, relatedness, population structure, and multiple testing should also be reported. Researchers should preserve the analysis scripts and intermediate results and distinguish exploratory analyses from confirmatory tests. Transparent reporting allows other investigators to evaluate the assumptions, reproduce the analysis, and assess the reliability of reported associations.
- Association models for quantitative and binary traits provide the statistical foundation for identifying relationships between genetic variation and biological characteristics. Linear regression is commonly used for quantitative traits, while logistic regression is widely used for binary outcomes. Alternative genotype models, mixed-effects methods, rare-variant tests, and specialized models extend these approaches to more complex genetic and phenotypic settings. Reliable association results depend on appropriate model selection, high-quality genotype and phenotype data, correction for population structure and relatedness, careful interpretation of effect estimates, and rigorous control of multiple testing.