Why Use a Mixture Prior?
Genome-wide association study (GWAS) estimates are noisy, and linkage disequilibrium (LD) means that a marginal association estimate can reflect correlated variants as well as the variant’s own effect. The earlier material on shrinkage and LDpred2 established why estimates should be regularized while accounting for correlation among SNPs. SBayesR combines these ideas using summary statistics and an explicit prior that allows effects to have different plausible sizes.
The prior is not a claim that every SNP belongs to a known biological class. Rather, it represents a flexible distribution across the genome: some effects may be exactly zero, many may be small, and a smaller share may be larger. The model estimates how much posterior support each SNP has for the available effect-size components, while using the LD matrix to distinguish correlated signals.
The mixture weights and the component variances have different meanings. A weight \(\pi_c\) describes the prior fraction assigned to component \(c\); it does not describe the magnitude of an effect in that component. The multiplier \(\gamma_c\) sets the component variance relative to \(\tau^2\). For example, the \(\gamma_c=0.01\) component has standard deviation \(0.1\tau\), while the \(\gamma_c=1\) component has standard deviation \(\tau\). Component membership is uncertain: the data update the probability that a SNP’s effect comes from each component.
Effect sizes are interpreted on the scale used to construct the summary statistics. With standardized genotype and phenotype variables, \(\beta_j\) is measured in phenotype standard deviations per genotype standard deviation. With other scaling, the effect units change accordingly; the mixture structure itself does not make coefficients unit-free.
Summary Statistics and the LD-Aware Likelihood
Let \(\hat{\boldsymbol b}\) be the vector of marginal GWAS effect estimates for a set of \(m\) SNPs, and let \(R\) be their LD correlation matrix from an appropriate reference panel. Let \(\boldsymbol\beta\) denote the corresponding joint effects. In a standardized linear model with a common sample size \(n\), a useful summary-statistic approximation is
This approximation assumes that genotypes and phenotype are consistently standardized, the summary statistics refer to the same harmonized SNP alleles, and the LD matrix adequately represents the discovery sample. It also uses a common-sample-size, linear-model approximation with residual variance treated as one. In applications, the likelihood scaling must match the GWAS design and the supplied summary-statistic conventions; the displayed form is not a universal likelihood for every phenotype or meta-analysis.
The mean \(R\boldsymbol\beta\) explains why marginal estimates are not generally equal to joint effects. For SNP \(j\), the approximate marginal estimate is \(\hat b_j \approx \beta_j+\sum_{k\ne j}r_{jk}\beta_k\), where \(r_{jk}\) is the LD correlation between SNPs \(j\) and \(k\). Thus the estimate at one SNP can carry signal from correlated SNPs. SBayesR uses the full LD-aware likelihood, rather than treating all marginal estimates as unrelated observations.
How the Mixture Enters Estimation
The likelihood contributes information about the joint effects, while the mixture prior regularizes them. A SNP whose data are weakly informative may retain substantial posterior probability at the zero or small-effect components. A SNP with stronger evidence can receive more posterior probability on a wider component. Its final estimated effect is typically summarized by a posterior mean, which averages over both effect values and component uncertainty.
A short derivation makes the role of LD explicit. Under the approximate likelihood above, its log-density, up to terms that do not depend on \(\boldsymbol\beta\), is
The equality follows by expanding the quadratic and using \(R^\mathsf{T}=R\). In particular, the term \(\boldsymbol\beta^\mathsf{T}R\boldsymbol\beta\) includes cross-products between correlated SNP effects. If \(R\) were replaced by the identity matrix, those cross-products would disappear, and the model would no longer use LD to separate correlated effects.
For a coordinate update, hold all effects except \(\beta_j\) fixed and define the LD-adjusted residual summary \(z_j=\hat b_j-\sum_{k\ne j}r_{jk}\beta_k\). Since \(R_{jj}=1\), the likelihood contribution for that coordinate is proportional to
This expression shows the conditional estimate being updated against a residual that accounts for the current effects of correlated SNPs. SBayesR uses sampling and model updates to combine this likelihood with the spike-and-mixture prior across the SNP set. The resulting posterior distribution, not a single component assignment treated as certain, is the basis for effect summaries.
Worked Example: Averaging Across Effect Components
Suppose a hypothetical SBayesR update for one SNP gives posterior probabilities \(0.34264\), \(0.39516\), and \(0.26220\) for three components, with corresponding conditional posterior means \(0\), \(0.010\), and \(0.018\). These are illustrative posterior summaries, not measurements from a real study. The probabilities sum to one: \(0.34264+0.39516+0.26220=1.00000\).
The posterior mean is the probability-weighted average of the component-specific means:
Thus the posterior mean is \(0.008671\) after rounding to six decimal places, in the effect units of the summary statistics. It is neither the effect of the most probable component nor a guarantee that the SNP has a nonzero effect. It averages over uncertainty about component membership and effect size.
Worked Example: LD Changes the Joint Estimates
Consider a hypothetical two-SNP analysis with standardized summary statistics \(\hat{\boldsymbol b}=(0.12,0.10)^\mathsf{T}\), sample size \(n=100\), and LD matrix \(R=\begin{pmatrix}1&0.8\\0.8&1\end{pmatrix}\). To isolate the role of LD, use a common zero-centered normal prior with precision \(\lambda=100\) for this calculation. This is a Gaussian-prior illustration of the LD-aware likelihood, not a replacement for SBayesR’s mixture prior.
The posterior mode minimizes the negative log-likelihood plus the prior penalty. Differentiating the quadratic objective and setting the gradient to zero gives \((nR+\lambda I)\hat{\boldsymbol\beta}=n\hat{\boldsymbol b}\). Substitution yields
The determinant is \(200(200)-80(80)=33{,}600\), so
If the two marginal estimates were instead shrunk independently by the same prior precision, the results would be \(100(0.12)/(100+100)=0.0600\) and \(100(0.10)/(100+100)=0.0500\). The joint estimates differ because the LD-aware calculation recognizes that the two marginal statistics share correlated signal. This example demonstrates why the LD matrix matters; it does not establish that the reference LD or the assumed prior is suitable for a particular analysis.
Worked Example: Overall Prior Variance Across SNPs
Suppose the mixture weights for the four components are \((0.90,0.05,0.03,0.02)\), with variance multipliers \((0,0.01,0.1,1)\), and take \(\tau^2=0.0011404255\). The average prior variance is the weighted average of the component variances:
The corresponding prior standard deviation across SNPs is \(\sqrt{0.0000268}=0.005176871\ldots\), or approximately \(0.005177\) in the same effect units. This summary describes the marginal spread induced by the mixture prior; it does not mean that every component has this standard deviation. The component-specific variances remain \(0\), \(0.01\tau^2\), \(0.1\tau^2\), and \(\tau^2\).
Assumptions, Diagnostics, and Failure Modes
The interpretation of SBayesR depends on more than the mixture specification. The summary statistics and LD matrix must use compatible SNP identifiers, allele coding, and variant sets. Effect alleles must be harmonized so the signs of \(\hat b_j\) and the corresponding LD relationships are meaningful together. The reference panel should represent the LD structure in the discovery sample well enough for the likelihood approximation to be useful. A mismatch can cause the model to attribute a correlated signal to the wrong SNPs or distort the amount of shrinkage.
The displayed likelihood also presumes a suitable sample-size and variance scale. If the true study has varying sample sizes, substantial sample overlap, meta-analysis-specific uncertainty, or a non-linear phenotype model, a single common-\(n\) approximation may not capture the dependence and uncertainty in the summary statistics. Model inputs and software options should be checked against the GWAS design rather than assuming that every reported \(\hat b_j\) has the same sampling variance.
Useful checks include confirming that allele harmonization and SNP matching are complete, that the LD matrix is symmetric with plausible correlations, and that its variant order matches the summary-statistic order. Results that change sharply under small changes to the reference panel or SNP set can signal instability or LD mismatch. Very large or implausible estimated effects, unusual component allocation patterns, or poor agreement between observed and model-implied summary statistics also warrant investigation. These checks diagnose concerns; they do not by themselves prove that a model is correct.
A mixture prior is flexible relative to a single normal prior, but it does not guarantee that the true effects follow the specified mixture. The discrete variance multipliers approximate a range of effect sizes rather than representing known biological categories. The zero component is a modeling device, not proof that a SNP is biologically inactive. Nor can LD-aware estimation recover reliable distinctions between strongly correlated variants when the data contain little information to distinguish them.
How SBayesR Differs From Other Shrinkage Choices
| Approach | Prior or treatment of effects | Use of LD |
|---|---|---|
| Independent shrinkage | Each marginal estimate is shrunk without modeling correlated SNP effects jointly. | Does not use the full LD structure in the joint likelihood. |
| SBayesR | A point mass and several normal components allow posterior support for zero, small, and larger effects. | Uses the LD matrix in a summary-statistic likelihood. |
| PRS-CS | Uses continuous shrinkage, allowing a continuous range of shrinkage strengths. | Models LD in its summary-statistic framework. |
The comparison is about modeling choices, not a guarantee of which method predicts best. SBayesR’s discrete mixture makes its component interpretation direct, while continuous-shrinkage methods represent effect scales without a small fixed set of variance components. Relative performance depends on data quality, LD compatibility, model assumptions, and the prediction target.
Common Mistakes
- Reading a variance multiplier as an effect multiplier. The value \(0.1\) scales a component variance; its standard deviation is scaled by \(\sqrt{0.1}\), not by \(0.1\).
- Treating mixture weights as posterior assignments. The \(\pi_c\) values are prior probabilities. Posterior component probabilities also depend on the summary statistics, LD, and other effects.
- Interpreting the zero component as established biological absence. It is a prior component that permits exact zero effects in the model; posterior support for it remains conditional on the data and model.
- Using a convenient but mismatched LD panel. An LD-aware likelihood is only as relevant as its LD matrix is to the discovery data and its allele and variant coding.
- Equating marginal and joint effects. With LD, a marginal estimate can reflect multiple correlated joint effects, so direct shrinkage of marginal estimates is not generally the same as joint estimation.
Check Your Understanding
Try each problem, then compare your reasoning with the solutions below.
- In the prior with variance multipliers \(0,0.01,0.1,1\), what does the multiplier \(0.01\) scale?
- Why can a marginal GWAS estimate differ from the corresponding joint effect when two SNPs have nonzero LD?
- A SNP has posterior component probabilities \(0.2\) and \(0.8\), with conditional posterior means \(0\) and \(0.015\). Calculate its posterior mean.
- Name two ways that a poorly matched LD reference can affect SBayesR estimates.
Solutions. (1) It scales the component variance relative to \(\tau^2\), giving variance \(0.01\tau^2\) and standard deviation \(0.1\tau\). (2) The marginal estimate includes contributions from correlated SNP effects, approximately \(\hat b_j\approx\beta_j+\sum_{k\ne j}r_{jk}\beta_k\). (3) The mean is \(0.2(0)+0.8(0.015)=0.012\), in the effect units used for the estimates. (4) LD mismatch can misattribute correlated signal across SNPs and can distort shrinkage or posterior component support.