Why Uncertainty Must Travel Along the Path
A multi-omic pathway effect is often assembled from several estimated links. For example, one molecular layer may be associated with a second layer, which is in turn associated with a third. If the link effects are \(a\) and \(b\), the product \(ab\) summarizes the effect transmitted along that two-link path under the specified path model. The product is not known exactly: both link estimates have sampling or posterior uncertainty, and they may be dependent.
The main task here is to propagate that uncertainty through the product. A calculation that treats uncertain estimates as fixed numbers understates what is known about the pathway effect. A calculation that ignores covariance can also be wrong, because estimates may use the same participants, correlated molecular measurements, or a shared model.
This is an uncertainty calculation for a defined product. It does not establish that the path is causally identified, that the component effects are unbiased, or that the chosen pathway is biologically correct. Those questions depend on the design and assumptions of the molecular mediation model. Here, the component estimates and their joint uncertainty are taken as inputs.
Delta-Method Uncertainty
The delta method approximates the variability of a smooth function of estimated quantities by linearizing that function near the estimates. Let \(\hat{\boldsymbol{\beta}}=(\hat{\beta}_1,\ldots,\hat{\beta}_K)^\mathsf{T}\) estimate \(\boldsymbol{\beta}\), with estimated covariance matrix \(\widehat{\Sigma}\). For the product function \(g(\boldsymbol{\beta})=\prod_j\beta_j\), the derivative with respect to component \(j\) is the product of all the other components.
The approximation is useful when the joint estimator is reasonably close to multivariate normal and the product is not so nonlinear over the relevant range that a local linear approximation is misleading. A Wald interval is then \(\hat{\theta}\pm z_{1-\alpha/2}SE(\hat{\theta})\), where \(SE(\hat{\theta})\) is the square root of the estimated variance. The interval can extend below zero even when all estimated component effects are positive; that is possible for a symmetric normal approximation.
For independent estimates, the covariance term is zero. Independence should follow from the estimation design or be supported by the joint covariance estimate; it should not be assumed simply because effects refer to different omic layers. For more than two links, the matrix formula includes every variance and covariance contribution.
Worked Example: A Three-Link Product by the Delta Method
Suppose three hypothetical standardized molecular links have estimates \(\hat{\beta}_1=0.20\), \(\hat{\beta}_2=0.50\), and \(\hat{\beta}_3=0.40\), with standard errors 0.05, 0.10, and 0.08. Assume, for this example, that the estimates come from independent samples, so their estimated covariances are zero. The path product is $$ \hat{\theta}=0.20(0.50)(0.40)=0.0400. $$
The gradient at the estimates is \((0.50\cdot0.40,\;0.20\cdot0.40,\;0.20\cdot0.50)=(0.20,\;0.0800,\;0.100)\). With zero covariances, the estimated variance is $$ \widehat{\mathrm{Var}}(\hat{\theta}) =(0.20)^2(0.05)^2+(0.0800)^2(0.10)^2+(0.100)^2(0.08)^2 =0.000100+0.0000640+0.0000640=0.000228. $$ Thus \(SE(\hat{\theta})=\sqrt{0.000228}=0.0151\). A 95% Wald interval is $$ 0.0400\pm1.96(0.0151)=0.0400\pm0.0296=(0.0104,\;0.0696). $$
Under the stated independence and approximation assumptions, the estimated standardized path product is 0.0400, with a 95% Wald interval from 0.0104 to 0.0696. The interval describes uncertainty in the product, not uncertainty in any one link alone.
Why Covariance Can Change the Answer
When component estimates are dependent, covariance contributes to the uncertainty of their product. Positive covariance between two positive estimated links increases the delta-method variance through the \(2\hat a\hat b\,\widehat{\mathrm{Cov}}(\hat a,\hat b)\) term. Negative covariance decreases it. The sign and size of that contribution depend on the estimates as well as their covariance.
Worked Example: Accounting for Dependence in Two Links
Suppose two hypothetical positive link estimates are \(\hat a=0.30\) and \(\hat b=0.40\), with standard errors 0.08 and 0.10. They are estimated in the same participants, and their estimated correlation is 0.60. Therefore their estimated covariance is $$ \widehat{\mathrm{Cov}}(\hat a,\hat b) =0.60(0.08)(0.10)=0.00480. $$ The product estimate is \(\hat a\hat b=0.30(0.40)=0.120\). Its estimated variance is $$ (0.40)^2(0.08)^2+(0.30)^2(0.10)^2 +2(0.30)(0.40)(0.00480) =0.001024+0.000900+0.001152=0.003076. $$ Thus \(SE=\sqrt{0.003076}=0.0555\), and the 95% Wald interval is $$ 0.120\pm1.96(0.0555)=(0.0113,\;0.229). $$
If covariance were incorrectly set to zero, the variance would be only \(0.001924\), giving \(SE=0.0439\). In this example, ignoring positive dependence understates the estimated standard error. The same-subject design is precisely why the covariance needs to be retained.
Paired Bootstrap Propagation
A bootstrap propagates sampling uncertainty by repeatedly resampling observational units, refitting the component effects, and recalculating the product. If links are estimated from the same participants, resample participants jointly across the relevant omic layers. This preserves the dependence induced by shared units. Resampling each layer independently would break that dependence and generally target the wrong joint sampling distribution.
The bootstrap distribution of the recalculated products can be summarized by its standard deviation or by empirical quantiles. A percentile interval uses the \(\alpha/2\) and \(1-\alpha/2\) quantiles of the bootstrap products for a nominal \(100(1-\alpha)\%\) interval. This interval is not automatically superior to a delta-method interval: its quality still depends on a suitable sampling design, a valid resampling scheme, and enough bootstrap replicates. For independent cohorts, resampling is carried out separately within each cohort; for shared participants, the participant-level records must remain paired.
Resample the independent biological units appropriate to the design, such as participants, keeping all linked omic measurements together when they come from the same participant.
In each resample, re-estimate all component effects using the same modeling procedure used for the observed data.
For bootstrap replicate \(r\), compute \(\hat{\theta}^{*(r)}=\prod_j\hat{\beta}_j^{*(r)}\).
Use their spread to estimate standard error, or use specified empirical quantiles to form a bootstrap interval.
Worked Example: A Paired Bootstrap for a Two-Link Path
Consider the two-link pathway in which both links are estimated from the same hypothetical cohort of participants. The observed estimates are \(\hat a=0.30\) and \(\hat b=0.40\), so the observed product is \(0.120\). Suppose 2,000 participant-level bootstrap resamples are drawn, with each participant’s measurements for both layers kept together. In one illustrative resample, the refitted estimates are \(a^*=0.24\) and \(b^*=0.46\), giving $$ \theta^*=a^*b^*=0.24(0.46)=0.1104. $$
Suppose the 2,000 hypothetical replicate products have a 2.5th percentile of 0.018 and a 97.5th percentile of 0.238. The percentile bootstrap interval is therefore \((0.018,\;0.238)\). Its endpoints are empirical quantiles of the resampled products; they are not calculated by adding and subtracting 1.96 standard errors. The width and asymmetry of this interval reflect the observed joint resampling distribution, including dependence between the fitted links.
The example assumes that participants are independent sampling units and that the bootstrap repeats the entire fitting procedure. If participants are clustered, for example within families, resampling individuals independently would not respect the design; the resampling unit must instead reflect the independent clusters.
Joint Posterior Simulation
When a fitted model supplies joint posterior draws for the component effects, uncertainty in the path product can be propagated by calculating the product within every draw. Joint draws matter: they preserve posterior dependence among effects. Multiplying posterior means gives one summary value, but it does not provide the posterior distribution of the product. In general, the product of posterior means is not the posterior mean or median of the product.
A posterior interval summarizes uncertainty conditional on the specified model, prior, and observed data. A posterior probability such as \(P(\theta>0\mid\text{data})\) is also conditional on those inputs. It is not a frequentist confidence level. This simulation approach does not require a normal approximation for the product, although the quality of its conclusions still depends on the fitted model and the adequacy of the simulation.
Worked Example: Propagating Three Joint Posterior Effects
Suppose a hypothetical model produces 10,000 joint posterior draws for three standardized links. The posterior means of the individual links are 0.15, 0.35, and 0.45. Their product is $$ 0.15(0.35)(0.45)=0.0236, $$ but this calculation alone is not a summary of the product’s uncertainty. For one illustrative joint draw, let the three effects be 0.10, 0.42, and 0.51. The path product for that draw is $$ 0.10(0.42)(0.51)=0.0214. $$
Suppose multiplying the three effects within each of the 10,000 joint draws gives a posterior median of 0.021, a 95% credible interval from \(-0.012\) to 0.068, and \(P(\theta>0\mid\text{data})=0.91\). Conditional on this model, prior, and data, 95% of the posterior probability for the path product lies between \(-0.012\) and 0.068; the posterior probability that the product is positive is 0.91. The credible interval includes zero, so the probability statement is more informative than reporting only the signs of the posterior means.
These results are hypothetical summaries of joint posterior draws. They cannot be reconstructed from the three marginal means alone: the joint draws, including their dependence, are needed to obtain the product distribution.
Choosing and Comparing the Methods
| Method | What is propagated | Useful when | Main caution |
|---|---|---|---|
| Delta method | Estimated covariance of the component effects, through a local linear approximation | Estimates are approximately jointly normal and the product is not too nonlinear over their uncertainty range | Can misrepresent skewness or uncertainty when effects are imprecise or near zero |
| Paired bootstrap | Sampling variation from repeated resampling and refitting | Observational units and dependence structure can be resampled appropriately | Invalid resampling units or too few replicates distort the product distribution |
| Joint posterior simulation | The joint posterior distribution of component effects | A model supplies well-behaved joint draws and a posterior summary is desired | Results depend on the model, prior, data, and adequate simulation |
The delta method is transparent and computationally efficient, especially when a reliable covariance matrix is available. Bootstrap and posterior simulation can retain features of a non-normal product distribution that a symmetric Wald interval cannot. Neither simulation method repairs an invalid component model, and neither should discard dependence among links. A practical analysis may compare methods when nonlinear behavior is plausible, while stating clearly which uncertainty interpretation each method provides.
Derivation and a Common Pitfall
For the two-link case, write the estimation errors as \(u=\hat a-a\) and \(v=\hat b-b\). Expanding the product gives $$ \hat a\hat b=(a+u)(b+v)=ab+bu+av+uv. $$ The first-order delta approximation retains the linear terms \(bu+av\) and omits the second-order term \(uv\). The variance of those retained terms is $$ b^2\mathrm{Var}(u)+a^2\mathrm{Var}(v)+2ab\,\mathrm{Cov}(u,v). $$ Replacing unknown parameters and covariance quantities by their estimates gives the two-link formula above. The omitted term explains why the approximation can be less accurate when estimation errors are large or the product’s distribution is notably skewed.
Common mistake: “The two links were estimated separately, so their covariance is zero.” Separate coefficient estimates are not necessarily statistically independent; they may use the same participants or share estimated nuisance quantities. The correction is to use the joint covariance from the fitted model, or use a resampling or simulation procedure that preserves the dependence. Setting covariance to zero is justified only when independence is warranted by the estimation design or model.
Check Your Understanding
Answer each question using the assumptions and scale of the stated pathway.
- For three component effects, what is the derivative of their product with respect to the second effect?
- Two positive link estimates have positive covariance. What term in the delta-method variance reflects this dependence?
- Why should measurements from the same participant remain paired in a bootstrap of a multi-omic path?
- What does a 95% posterior credible interval describe, and how does its interpretation differ from a 95% confidence interval?
- Give one condition under which the delta-method interval may be unreliable for a product.