Shrinkage as a Response to Noisy GWAS Effects

Why Shrink Noisy GWAS Effects?

A polygenic risk score combines many estimated genetic effects. Even when most individual effects are small, their estimation noise can accumulate across variants. Using each GWAS estimate without adjustment is simple, but it can give excessive influence to estimates that happened to be large because of sampling variation. Shrinkage reduces the magnitude of estimates, usually drawing them toward zero. It can lower the variability of scores in new samples, at the cost of introducing bias when true effects are nonzero.

This tutorial focuses on that trade-off. It does not imply that every shrunk score predicts better: performance depends on the quality of the estimates, the shrinkage rule, linkage disequilibrium (LD), and the target population. The earlier material on defining a polygenic risk score and on how LD changes marginal GWAS effects supplies the context for interpreting the weights.

A Sampling Model and the Bias-Variance Trade-Off

For a first analysis, consider one effect estimate with sampling model \(\hat{\beta}=\beta+\varepsilon\), where \(E(\varepsilon)=0\) and \(\operatorname{Var}(\varepsilon)=v\). A common approximation in a sufficiently large GWAS is \(\hat{\beta}\sim N(\beta,v)\), with \(v\) estimated from the squared standard error. The model treats the variant coding and effect scale as fixed and assumes the reported standard error adequately describes sampling uncertainty. It is a one-effect approximation: it does not account by itself for LD between variants, population structure, or systematic bias.

For the shrunk estimate \(\tilde{\beta}=a\hat{\beta}\), its expected value is \(a\beta\). Thus its bias is \((a-1)\beta\), and its variance is \(a^2v\). Squared-error risk is the sum of squared bias and variance:

The first term is the price of pulling a genuine effect toward zero; the second is the remaining sampling variance. Unshrunk estimates have zero bias under this model, but variance \(v\). Shrinkage is useful when the variance reduction outweighs the added squared bias for the prediction task.

To verify the result, expand the risk as a quadratic in \(a\). Its derivative is \(-2(1-a)\beta^2+2av\). Setting it to zero gives \(\beta^2=a(\beta^2+v)\), hence \(a^*=\beta^2/(\beta^2+v)\). The second derivative is \(2(\beta^2+v)>0\) unless both signal and sampling variance are zero, so this stationary point minimizes the risk. At this value the minimum risk is \(\beta^2v/(\beta^2+v)\), which is less than the unshrunk risk \(v\) when \(\beta^2\) is finite and \(v>0\). This is an oracle result, not a directly usable rule: it depends on the unknown true effect.

Common Shrinkage and Ridge-Like Weights

A single factor \(a\) makes the basic idea transparent, but practical genomic methods need a rule for choosing weights across many variants. In an idealized setting with uncorrelated, standardized predictors, ridge regression shrinks coefficients continuously. With \(n\) observations and objective equal to residual sum of squares plus \(\lambda\sum_j\beta_j^2\), each orthogonal coefficient is multiplied by \(n/(n+\lambda)\). The exact factor changes if the loss function is scaled differently; therefore, a ridge penalty value is meaningful only together with the objective convention.

With correlated genotypes, ridge does not generally shrink each marginal GWAS estimate by the same scalar. In matrix form its coefficients depend jointly on the predictor correlation matrix and the penalty. This distinction matters because marginal GWAS effects can reflect LD with other causal or associated variants. A uniform multiplier can illustrate noise reduction, but it cannot in general resolve LD-induced differences between marginal and joint effects.

Bayesian Posterior Estimates: A Related Form of Shrinkage

A Bayesian estimate combines a likelihood, which describes sampling information, with a prior distribution for effects. The posterior mean is often used as a weight. For a prior centered at zero, estimates with substantial sampling uncertainty can be pulled more strongly toward zero than precise estimates. This resembles ridge shrinkage, but the amount of shrinkage follows from the prior and likelihood assumptions rather than from a penalty chosen in a regression objective.

A simple normal example makes the relationship concrete. Suppose \(\hat{\beta}\mid\beta\sim N(\beta,v)\) and assign \(\beta\sim N(0,\tau^2)\), independently of the sampling error. The posterior mean is \(\tau^2/(\tau^2+v)\hat{\beta}\). This elementary scalar result illustrates the form of shrinkage; specifying an effect-size prior for genome-wide variants and accounting for LD require additional modeling.

Assumptions, Diagnostics, and Alternatives

Shrinkage is most defensible when the effect estimates are noisy and the shrinkage rule is estimated without using the final test data. The simple calculations above assume an appropriate sampling model, a zero-centered target for shrinkage, and squared-error loss. They also assume that uncertainty estimates are meaningful. If effect estimates have systematic bias, for example from population structure or allele-coding errors, multiplying them toward zero does not repair the source of bias.

In practice, inspect the distribution of standard errors and effect estimates, confirm consistent allele coding and effect scales, and check whether the chosen weights behave sensibly across independent validation data. A marked loss of predictive accuracy after shrinkage can indicate over-shrinkage, a poorly chosen penalty or prior, or a mismatch between discovery and target settings. Conversely, apparent gains confined to the data used to tune the rule are not reliable evidence of generalization. LD-related diagnostics and careful validation remain necessary; shrinkage alone does not guarantee portability.

These approaches answer related but distinct questions. Shrinkage addresses estimation noise by modifying weights; it does not determine whether the score is well calibrated or clinically useful. The best method depends on the intended prediction target, the available reference data, and validation that respects the separation between training, tuning, and testing.

Common Mistakes and Limitations

• Claiming shrinkage must improve every estimate. It reduces expected squared error only under specified assumptions and for an appropriate amount of shrinkage; a particular realized estimate can become worse.
• Calling a smaller estimate a more accurate estimate by definition. Shrinking toward zero adds bias when the true effect is nonzero. The justification is a possible reduction in total risk, not the belief that effects are always zero.
• Applying the scalar formula as if variants were independent. Correlated variants require attention to LD; joint ridge weights are not generally obtained by independently scaling marginal effects.
• Tuning on the test cohort. Choosing a shrinkage factor using test outcomes makes the reported test performance optimistic. Tune using training or validation data and reserve test data for final evaluation.

The central limitation is that no shrinkage rule can create information absent from the discovery data. It can stabilize noisy weights under a useful model, but its success depends on assumptions about effect distributions, sampling uncertainty, LD, and the population in which the score will be used.