Why Multi-Omic Mediation Is Easy to Overinterpret
A multi-omic analysis can connect a genetic variant to gene expression, protein abundance, a metabolite, and a clinical outcome. A fitted model may yield a sequence of coefficients and an estimated indirect effect, yet those calculations alone do not establish that the molecular measurements form a causal chain. Correlated assays, noisy measurements, weak associations, and unmeasured common causes can all make a pathway claim look more definitive than the evidence warrants.
The practical task is not to dismiss mediation estimates, but to separate what the model estimates from what the evidence supports. An estimated product of path coefficients is a statistical summary under a specified model. Calling it a causal indirect effect requires additional assumptions about measurement, confounding, model form, and the direction of the relationships.
For example, in linear models without exposure–mediator interaction, one might write \(M=\alpha_M+aX+\varepsilon_M\) and \(Y=\alpha_Y+c'X+bM+\varepsilon_Y\). This notation does not itself ensure that \(a\), \(b\), or \(ab\) have causal interpretations. A nonzero product can be compatible with mediation, but it can also arise from bias, measurement error, or chance.
Correlated Assays Are Not Automatically Separate Pathways
Expression, protein, and metabolite assays can be strongly correlated because they measure related biology, share technical variation, or respond to common causes. When correlated mediators enter the same outcome model, each coefficient describes the association with the outcome conditional on the other included mediators. It is not the same quantity as the association from a model containing that mediator alone.
This distinction matters for pathway interpretation. A coefficient from a single-mediator model may partly reflect the signal carried by another correlated assay. Conversely, coefficients in a joint model may be imprecise or unstable when assays are highly correlated. Neither pattern, by itself, tells us that one assay is the uniquely causal mediator.
Worked Example: Two Correlated Assays
Suppose a hypothetical study reports standardized measurements \(M_1\), \(M_2\), and \(Y\). The correlation between \(M_1\) and \(M_2\) is \(0.90\), and the correlation of each mediator with \(Y\) is \(0.50\). Compare the single-mediator associations with the coefficients from a joint linear regression of \(Y\) on both mediators.
With standardized variables, the single-mediator regression coefficient equals the correlation. Thus, the coefficient for \(M_1\) in a model containing only \(M_1\) is \(0.50\); likewise, the coefficient for \(M_2\) in a model containing only \(M_2\) is \(0.50\).
In the joint regression, the coefficient vector solves the two-predictor normal equations. The correlation matrix of the predictors and their correlations with the outcome give:
The symmetric solution has \(\hat{\beta}_1=\hat{\beta}_2\). Substituting that value into either equation gives \(1.90\hat{\beta}_1=0.50\), so each joint coefficient is \(0.50/1.90=0.2632\). The joint model therefore assigns a smaller conditional coefficient to each assay than the separate models do.
Now suppose the estimated genetic associations with the two assays are each \(a_1=a_2=0.20\). Multiplying these by the joint coefficients gives \(0.20(0.2632)=0.0526\) for each model-based path product. The arithmetic does not establish that there are two distinct causal routes: the assays are highly correlated, and the data summaries alone do not identify why. A defensible report would describe the conditional associations and assay correlation, then qualify any pathway interpretation.
A common error is to treat each assay’s single-mediator association as an independent pathway and add the resulting products. That can count shared signal more than once. If a joint mediator model is used, its conditional coefficients also need careful interpretation: their meaning depends on the model and the causal assumptions, not just on including both assays in a regression.
Weak Associations Require Cautious Pathway Language
An indirect-effect estimate is a product. If either component is weakly estimated, the product can be uncertain even when its point estimate is nonzero. In particular, a small \(p\)-value for one component does not establish that the product is estimated precisely, and a nonsignificant component test does not prove that the corresponding association is exactly zero.
The following calculation uses the delta-method approximation for the standard error of a product. If \(\hat{a}\) and \(\hat{b}\) have standard errors \(SE(\hat{a})\) and \(SE(\hat{b})\), and their estimation covariance is negligible, then:
This is an approximation, not a universal description of the product’s sampling distribution. If the estimates are dependent, the covariance contribution must also be considered; for a weak or highly uncertain path, a method suited to the product’s distribution may be preferable.
Worked Example: A Small Product With Substantial Uncertainty
In a hypothetical analysis, suppose \(\hat{a}=0.040\) with \(SE(\hat{a})=0.030\), and \(\hat{b}=0.30\) with \(SE(\hat{b})=0.12\). Assume for this illustrative calculation that the two coefficient estimates have negligible covariance. Calculate the product and its approximate standard error.
The product is \(\hat{a}\hat{b}=0.040(0.30)=0.0120\). Using the delta-method formula:
The approximate Wald statistic is \(0.0120/0.0102=1.18\). Under a standard normal approximation, the two-sided \(p\)-value is about \(0.24\), and an approximate 95% Wald interval is \(0.0120\pm1.96(0.0102)=[-0.0080,\,0.0320]\). This interval is an approximation under the stated assumptions; it does not show that the true indirect effect has a 95% probability of lying inside it.
The evidence is compatible with a range of product values, including zero. It would be too strong to say that the molecular mediator carries a demonstrated genetic effect. It would also be too strong to conclude that mediation is impossible. A more accurate statement is that this analysis provides an imprecise estimate and does not establish a nonzero indirect effect.
Measurement Error Can Distort a Path
A measured assay is not necessarily the underlying molecular quantity of interest. Under classical measurement error, the observed value equals the true value plus an error term that has mean zero and is independent of the true value and other model variables. In a regression of an outcome on a noisy mediator, this error commonly attenuates the mediator coefficient toward zero. The effect on a mediation estimate depends on which variables are measured with error and on the model.
In the simple model below, \(X\) is measured without error and the mediator is measured with independent additive error. The \(X\)-to-measured-mediator coefficient remains \(a\), because independent error adds variability but no expected association with \(X\). The conditional mediator-to-outcome coefficient is attenuated according to the reliability of the mediator’s residual variation after conditioning on \(X\).
Worked Example: Error in the Mediator Measurement
Consider a hypothetical model \(M=0.40X+\varepsilon_M\) and \(Y=0.10X+0.50M+\varepsilon_Y\), with \(\operatorname{Var}(\varepsilon_M)=1\). Suppose the assay records \(M^*=M+u\), where \(u\) is independent of \(X\), \(M\), and \(\varepsilon_Y\), and \(\operatorname{Var}(u)=1\). Find the expected mediator coefficient when \(Y\) is regressed on \(X\) and \(M^*\).
After conditioning on \(X\), the remaining variation in \(M\) is \(\varepsilon_M\), and the remaining variation in \(M^*\) is \(\varepsilon_M+u\). Their covariance is \(1\), while the variance of \(\varepsilon_M+u\) is \(1+1=2\). Consequently, the conditional regression coefficient for \(M^*\) is:
The expected coefficient of \(X\) in the regression of \(M^*\) on \(X\) remains \(0.40\), because the independent measurement error has no association with \(X\). The estimated path product is therefore centered at \(0.40(0.25)=0.10\), rather than the true-model product \(0.40(0.50)=0.20\). This calculation relies on the specified linear model and independent additive error; more complex error or correlated assay noise need not produce this exact pattern.
A small estimated mediation product may thus reflect a weak biological relationship, noisy measurement, or both. Without information about assay reliability or a measurement-error model, it is generally not possible to determine which explanation applies from the product alone.
Untested Mediator–Outcome Confounding
Even a well-measured mediator can be confounded with the outcome. A mediator–outcome confounder is a variable that affects, or is associated with, both \(M\) and \(Y\). If it is not adequately controlled, the fitted \(b\) coefficient may combine the mediator’s relationship with the outcome and the confounder’s relationship with the outcome. An estimated \(ab\) can then misrepresent the mediated effect.
This issue is not resolved merely because \(X\) is genetic. The interpretation of a genetic exposure and the interpretation of the mediator–outcome path involve distinct assumptions. In particular, a genetic association with a molecular mediator does not by itself establish that the mediator causes the outcome.
Worked Example: An Unmeasured Common Cause Inflates the Mediator Slope
Suppose, hypothetically, that \(U\) is independent of \(X\), with variance 1, and that:
where \(\varepsilon_M\) has variance 1, and the errors are mutually independent and independent of \(X\) and \(U\). The coefficient of \(M\) in the outcome-generating equation is \(0.50\), but \(U\) also affects \(Y\). Calculate the mediator slope from a regression of \(Y\) on \(X\) and \(M\) that omits \(U\).
After conditioning on \(X\), the mediator’s residual is \(U+\varepsilon_M\), with variance \(1+1=2\). The outcome residual contains \(0.50(U+\varepsilon_M)+U+\varepsilon_Y=1.50U+0.50\varepsilon_M+\varepsilon_Y\). Its covariance with the mediator residual is \(1.50(1)+0.50(1)=2.00\). Thus, the fitted conditional slope is:
The omitted-variable regression slope is \(1.00\), although the coefficient assigned to \(M\) in the outcome-generating equation is \(0.50\). If the \(X\)-to-\(M\) coefficient is \(0.30\), the fitted product is \(0.30(1.00)=0.30\), whereas the product of the specified causal path coefficients is \(0.30(0.50)=0.15\). In this example, the difference arises because \(U\) is associated with both the mediator and outcome after conditioning on \(X\).
The calculation shows why a pathway claim requires a defensible argument about mediator–outcome confounding, not just a reported product. Adjusting for measured covariates can help only when the relevant confounders are measured and the adjustment is appropriate; the absence of an obvious confounder in a model does not establish that none exists.
How to Match the Conclusion to the Evidence
A useful interpretation states which associations were estimated, describes important limitations, and keeps causal wording conditional on the assumptions that support it. The same numerical result can warrant different wording depending on assay quality, mediator correlation, study design, and confounder control.
Say whether the reported quantity is a marginal association, a conditional regression coefficient, or a product of path estimates. Do not call all three “the pathway effect.”
Report uncertainty for the relevant coefficients and product. A nonzero point estimate alone does not establish a well-supported indirect effect.
Ask whether the molecular assays are reliable and whether correlated measurements make separate mediator interpretations difficult.
State that a causal interpretation depends on the identification assumptions, including adequate control of mediator–outcome confounding. If these are not established, describe the result as pathway-compatible evidence or an association.
For instance, if an analysis finds a genetic association with expression and a conditional expression–outcome association, a careful conclusion is that the pattern is consistent with the proposed pathway under the model’s assumptions. It is not automatically justified to say that changing expression would change the outcome. Stronger causal language requires evidence that supports the required causal ordering and identification assumptions.
Common Mistake
- Wrong reasoning: “The genetic variant is associated with the mediator, the mediator is associated with the outcome, and their coefficient product is nonzero; therefore, the assay mediates the genetic effect on the outcome.”
- Correction: Those findings describe model-based associations. Correlated assays can share signal, weak paths can yield uncertain products, measurement error can distort coefficients, and unmeasured mediator–outcome confounding can bias the conditional association. State the product and its uncertainty, then make a causal claim only to the extent that the relevant assumptions are supported.
Check Your Understanding
For each question, focus on what the evidence supports rather than on the pathway story one might hope to tell.
- Why can coefficients for two correlated molecular assays differ between separate regressions and a joint regression?
- An estimated indirect-effect product is small and has an interval that includes zero. What can be concluded, and what cannot be concluded?
- Under classical independent measurement error in a mediator, why can the mediator-to-outcome coefficient be attenuated?
- How can an unmeasured variable that affects both a mediator and an outcome distort the estimated mediator coefficient?
- Rewrite this claim more cautiously: “The protein mediates the variant’s effect on disease because both associations were significant.”