Confounder Adjustment

A graphical rule from causal directed acyclic graphs (DAGs) used to identify a sufficient set of variables to adjust for. A set of covariates satisfies the backdoor criterion if it blocks all spurious paths between exposure and outcome while leaving true causal paths open. In multi-omics, this prevents adjusting for colliders or mediators that would introduce collider stratification bias.

A balancing technique that condenses multiple confounders into a single scalar: the probability of exposure given covariates. Key implementations include:

• Propensity Score Matching (PSM): Pairing exposed and unexposed samples with similar scores
• Inverse Probability of Treatment Weighting (IPTW): Creating a pseudo-population where exposure is independent of confounders
• Stratification: Grouping samples into quintiles of propensity score

In proteogenomics, propensity scores can balance tumor stage and batch across treatment arms.

A method for addressing unmeasured confounding by leveraging a variable that:

• Is strongly associated with the exposure
• Has no direct effect on the outcome except through the exposure
• Is independent of unmeasured confounders

In Mendelian Randomization (MR), genetic variants (SNPs) serve as instruments to test causal effects of protein or metabolite levels on disease, bypassing environmental confounders entirely.

A technique using negative control outcomes (variables known not to be causally affected by the exposure) or negative control exposures to detect and correct residual confounding. If an association is observed where none should exist, the magnitude quantifies unadjusted bias. In multi-omics, negative control proteins or metabolites can calibrate false discovery rate estimates in high-dimensional association studies.

Regularized regression techniques designed for scenarios where the number of potential confounders exceeds the sample size:

• Double LASSO: Applies L1 regularization to both the exposure model and outcome model, selecting confounders that predict either
• Debiased Lasso: Provides valid p-values and confidence intervals after high-dimensional adjustment
• Post-double selection: Reduces omitted variable bias in wide omics datasets with limited n

In multi-omics integration, batch effects—systematic technical variation from processing date, plate, or sequencing lane—are a primary source of confounding. Methods like ComBat and Harmony adjust for batch while preserving biological variability. Failure to account for batch confounds can produce spurious cross-omics correlations that disappear upon technical replication, undermining biomarker reproducibility.

Batch effects represent one of the most pervasive confounders in multi-omics studies, where technical variation from different processing sites, sequencing runs, or sample preparation protocols can overwhelm biological signal.

• ComBat and Harmony algorithms adjust for known batch variables while preserving biological variability
• Applied routinely in The Cancer Genome Atlas (TCGA) and Genotype-Tissue Expression (GTEx) consortia
• Failure to adjust leads to clustering by processing site rather than disease subtype
• Modern implementations use negative control samples and replicate bridging to estimate batch effect magnitude

Propensity score methods estimate the probability of treatment or exposure assignment given observed covariates, enabling balanced comparisons in observational clinical-molecular datasets.

• Propensity score matching pairs treated and untreated patients with similar covariate profiles
• Inverse probability of treatment weighting (IPTW) reweights the entire cohort to create a pseudo-randomized population
• Used to adjust for confounding by indication when linking genomic biomarkers to treatment outcomes in real-world data
• Covariate balance diagnostics (standardized mean differences < 0.1) confirm successful adjustment

Surrogate Variable Analysis (SVA) identifies and estimates latent sources of expression heterogeneity that represent unmeasured confounders in high-dimensional omics data.

• Constructs surrogate variables from the residual expression matrix after removing primary signal
• Captures unknown technical artifacts (e.g., sample degradation, ambient temperature) and unmodeled biological variation
• sva and RUVseq (Remove Unwanted Variation) packages implement these methods in R/Bioconductor
• Critical for studies where all confounders cannot be measured or documented prospectively
• Enables recovery of true biological signal when batch annotations are incomplete or missing

Principal component analysis (PCA) on genotype data captures population structure, the most critical confounder in genome-wide association studies where allele frequency differences between ancestral groups create spurious phenotype associations.

• Top 10-20 principal components are included as fixed-effect covariates in linear or logistic regression models
• EIGENSTRAT software implements PCA correction with Tracy-Widom statistic for selecting significant PCs
• Linear mixed models (LMMs) like BOLT-LMM simultaneously adjust for population structure and cryptic relatedness
• Standard practice in all modern GWAS including UK Biobank and FinnGen analyses
• Prevents the classic confounding where both lactose tolerance and height appear associated due to shared ancestry