Contamination Mixture and Robust Profile-Likelihood MR

Why combine variant-specific evidence?

In summary-data Mendelian randomization (MR), each genetic variant provides an estimate of the exposure–outcome effect. When instruments are valid, those estimates should be compatible with a common causal effect, allowing for sampling error. Horizontal pleiotropy can move some estimates away from that common value. A robust analysis therefore needs to do more than average all variants: it must also describe how it identifies the evidence it relies on and how uncertainty is calculated.

Contamination-mixture and robust profile-likelihood methods approach this problem differently. A contamination mixture looks for a group, or cluster, of variant-specific estimates that agree with one another and treats other estimates as possible contaminants. Profile likelihood instead models the observed exposure and outcome associations together, accounting for uncertainty in the exposure associations while allowing some variants to have additional outcome effects. Neither method establishes that the selected variants are valid. Their conclusions remain conditional on modeling assumptions.

Variant-specific estimates and their uncertainty

Suppose \(\hat\beta_{Xj}\) and \(\hat\beta_{Yj}\) are the estimated associations of variant \(j\) with the exposure and outcome, with standard errors \(s_{Xj}\) and \(s_{Yj}\). The variant-specific ratio estimate is \(r_j=\hat\beta_{Yj}/\hat\beta_{Xj}\). It estimates the causal effect under the instrumental-variable assumptions for that variant. In a two-sample analysis with independent exposure and outcome association estimates, a first-order delta-method standard error is:

This approximation treats the ratio as a smooth function of the two estimated associations and propagates their sampling variances. It can be unreliable when \(\hat\beta_{Xj}\) is very close to zero; profile likelihood offers a way to account more directly for uncertainty in that association. The independence condition also matters: if the exposure and outcome estimates are correlated, a covariance term is needed. The examples below assume independent estimates and variants that are sufficiently independent for the stated likelihood calculations.

Contamination mixtures: use a compatible cluster

The central idea of a contamination-mixture approach is to allow that only some instruments support the target causal effect. For a candidate cluster \(S\), the estimates \(r_j\) for \(j\in S\) are modeled as noisy observations of a shared effect \(\theta\). Estimates not in \(S\) are treated as coming from a contaminating distribution rather than being forced to determine \(\theta\). The analysis compares possible clusters, or averages over them with specified weights, and gives greatest support to a group whose estimates are mutually compatible.

One way to express the idea is to assign each candidate subset \(S\) a prior weight \(w_S\), with the weights summing to one, and use a broad density \(g_j\) for estimates treated as contaminants. Under independent ratio estimates with standard errors \(s_j\), a schematic mixture likelihood is:

where \(\phi(r;\theta,s^2)\) is the normal density at \(r\), with mean \(\theta\) and variance \(s^2\). The subset weights and contaminating distribution are modeling choices; implementations can differ in how they construct and search over subsets. The expression makes the key logic explicit: estimates in a candidate valid cluster inform \(\theta\), while estimates classified as contaminants need not agree with it. The method is most persuasive when a clear cluster has substantial support, rather than when the answer depends on a single finely balanced subset.

This logic requires more than a mathematical mixture. The candidate cluster must plausibly represent variants whose ratio estimates share the causal effect. If invalid variants form a tight cluster of their own, or if most variants share directional pleiotropy, the largest or best-fitting cluster need not be valid. In particular, identifying a cluster is not a test that the exclusion restriction holds.

Profile likelihood: retain exposure-association uncertainty

A profile-likelihood analysis starts from the two association estimates rather than treating the ratio estimates as error-free inputs. Let \(\beta_j\) denote the unknown true exposure association and \(\alpha_j\) a variant's direct effect on the outcome not mediated by the exposure. A basic summary-data model is:

This formulation assumes the reported standard errors are appropriate, the association estimates are approximately normal, and the exposure and outcome estimates are independent (or that their covariance is incorporated). The nuisance parameter \(\beta_j\) is the true variant–exposure association. For a valid instrument, \(\alpha_j=0\); for an invalid one, it need not be zero. A robust profile-likelihood method allows some variants to contribute poorly fitting evidence without letting them dominate the estimate. A common explicit model uses a narrow valid component and a broader contaminating component for the outcome residual, while profiling over the nuisance \(\beta_j\) values. The exact robust component and its tuning choices should be reported, not hidden under the word “robust.”

The role of profiling can be seen by temporarily considering a valid variant with \(\alpha_j=0\). Define the observed discrepancy \(d_j(\theta)=\hat\beta_{Yj}-\theta\hat\beta_{Xj}\). Under the stated independent normal model, its variance is \(s_{Yj}^2+\theta^2s_{Xj}^2\). This follows because the true \(\beta_j\) cancels from the discrepancy, while the two sampling errors contribute variances \(s_{Yj}^2\) and \(\theta^2s_{Xj}^2\). Thus the discrepancy likelihood, apart from constants, is:

With a direct effect modeled as an independent mean-zero random deviation of variance \(\tau^2\), that variance becomes \(v_j(\theta)=s_{Yj}^2+\theta^2s_{Xj}^2+\tau^2\). A robust profile likelihood can instead use a mixture or other explicitly specified robust component for outlying residuals. In either case, maximizing the likelihood over nuisance parameters is called profiling. The resulting likelihood for \(\theta\) reflects uncertainty in both associations; simply substituting \(\hat\beta_{Xj}\) as if it were known does not.

Comparing the evidence and its limits

The methods share a concern—some variants may be invalid—but emphasize different features of the data. A contamination mixture makes the search for a coherent group especially visible. Profile likelihood makes sampling uncertainty in the exposure associations explicit and can incorporate a model for residual heterogeneity or contamination. When exposure associations are weak, that uncertainty can materially change the shape of the likelihood and the resulting interval.

For example, suppose an observed discrepancy is \(d=0.08\), and compare a narrow component with variance \(0.000029\) to a broad component with variance \(0.002529\). The ratio of the narrow-component normal density to the broad-component density at this discrepancy is:

The narrow component gives essentially no relative support at this discrepancy compared with the broad component. This is a relative density comparison for the specified models and variances, not the probability that the variant is invalid. Component membership depends on the full mixture, its weights, and all variants' evidence.

Both approaches are sensitive to assumptions. Correlated variants violate the independent likelihood calculations unless their covariance is modeled; variants in linkage disequilibrium are not independent evidence. A diffuse contaminating component can reduce the influence of outliers, but cannot guarantee protection against systematic directional pleiotropy or a misleading cluster. And uncertainty intervals conditional on a selected cluster may be too narrow if they ignore the uncertainty of selecting that cluster.

Common Mistakes

• Dropping exposure-association uncertainty. Using only \(s_Y^2/\hat\beta_X^2\) in a ratio standard error ignores the second delta-method term. Include it when the independent-estimate approximation is appropriate, or use a likelihood that accounts for uncertainty in \(\hat\beta_X\).
• Calling the best-fitting cluster “valid” as a fact. A cluster is evidence under a model, not proof of the exclusion restriction. A tightly grouped invalid set can mislead the analysis.
• Interpreting a mixture weight as a validity probability without qualification. It is conditional on the chosen mixture model and prior or component weights; it is not established by the data alone.
• Ignoring dependence among variants. If variants are correlated, independent contributions overstate the amount of information unless their covariance is handled.
• Reporting a conditional interval as if it included selection uncertainty. An interval calculated after fixing a cluster generally does not automatically account for uncertainty about which variants belong to it.