Reporting Uncertainty in Integrated Molecular Evidence

Why Reporting Is Part of the Analysis

An integrated molecular result is not fully interpretable from a point estimate alone. A reader also needs to know what quantity was estimated, which donors and molecular layers contributed, how missing measurements were handled, what tissues the data represent, and how uncertain the estimate is. Without these details, even a numerically precise result can be difficult to reproduce or easy to overinterpret.

This tutorial treats reporting as a statistical task: define the target quantity, make the data structure visible, describe assumptions that affect the estimate, and show how conclusions change under reasonable alternatives. The goal is not to make every analysis appear conclusive. It is to make clear which claims the evidence supports and which remain uncertain.

Definition: An integrated molecular estimand is the precisely specified population quantity that an analysis combining molecular layers aims to estimate. A report should state its target population, molecular pathway or combination of evidence, outcome and scale, contrast, and units. An estimate is a sample-based result for that estimand; it is not itself the estimand.

Specify the Estimand Before Giving the Estimate

A useful report begins with a sentence that says what the target quantity means in context. For example, an analysis might target a genotype-associated change in an outcome carried along a specified expression-to-protein-to-metabolite path. If its path coefficients are \(a\), \(b\), \(c\), and \(d\), the path-specific product is \(\theta=abcd\). The report must say what each coefficient represents, the scale of each molecular variable, and the units of the outcome and genetic contrast.

Standardizing molecular intermediates does not make every quantity unitless. If genotype is measured in allele copies and the outcome is measured in mmol/L, the product has units of mmol/L per allele. Nor does naming an indirect path establish that it is causal. A causal interpretation depends on identification assumptions, including assumptions about confounding and the ordering of the variables; the numerical definition of the estimand should not conceal those requirements.

Worked Example: State a Path-Specific Estimand and Its Uncertainty

Suppose a hypothetical analysis uses standardized expression, protein, and metabolite measures. The estimated path coefficients are \(a=0.20\) SD per allele, \(b=0.30\) SD protein per SD expression, \(c=0.25\) SD metabolite per SD protein, and \(d=0.40\) mmol/L per SD metabolite for the outcome. The target is the model-defined path-specific effect per additional allele.

Point estimate. The path product is

\hat{\theta}=\hat{a}\hat{b}\hat{c}\hat{d} =(0.20)(0.30)(0.25)(0.40) =0.0060\ \text{mmol/L per allele}.

The units follow by cancellation: the standardized expression and protein units cancel across their coefficients, as does the standardized metabolite unit, leaving mmol/L per allele.

Suppose a donor-level bootstrap with 2,000 resamples gives a 95% percentile confidence interval from \(-0.002\) to \(0.015\) mmol/L per allele. A careful report would say that the estimated path-specific effect is 0.0060 mmol/L per allele, with a 95% bootstrap confidence interval of \(-0.002\) to \(0.015\). Because the interval includes zero, these data do not rule out a zero path effect at the corresponding two-sided 5% level under the analysis and interval procedure. The interval is not a 95% probability that the fixed effect lies within these particular endpoints.

The bootstrap resamples donors, not individual assay values, so measurements from the same donor remain together. Its interpretation also depends on donors being an appropriate independent sampling unit and the fitted analysis being repeated consistently in each resample. The interval reflects sampling variability under that procedure; it does not automatically account for every modeling choice or bias from violated assumptions.

Report Who Contributed to Each Layer

For integrated evidence, the number of measurements is not necessarily the number of independent people. Report the number of unique donors and, for each molecular layer and outcome, the number with usable measurements. Also describe overlap: give the number of donors shared across relevant layers, especially the number with all measurements required for the integrated analysis. If different steps use different subsets, say which donors contribute to each estimate.

Overlap matters for uncertainty as well as transparency. Estimates from two tissues or molecular layers measured in some of the same donors are generally dependent. Combining them as though they came from independent samples can misstate the uncertainty of a sum or contrast. A report should identify the covariance method, or another method that preserves the observed dependence, rather than reporting only marginal standard errors.

Worked Example: Explain Sample Overlap and Its Effect on an Integrated Interval

In a hypothetical study of 240 donors, expression is available for 240, protein for 196, metabolite for 178, and the outcome for 224. A total of 142 donors have all four measurements. These counts describe assay availability, but do not by themselves show how estimates from different tissues relate.

Suppose two tissue-specific estimates, both in mmol/L per allele, are \(\hat{\theta}_1=0.012\) with \(SE=0.005\) and \(\hat{\theta}_2=0.008\) with \(SE=0.004\). The target integrated quantity is their sum, \(\hat{\theta}_{\mathrm{sum}}=0.020\). If the tissues share donors, suppose a donor-level joint bootstrap estimates their covariance as \(0.000012\) in squared units. Then

SE(\hat{\theta}_{\mathrm{sum}}) =\sqrt{SE(\hat{\theta}_1)^2+SE(\hat{\theta}_2)^2+ 2\operatorname{Cov}(\hat{\theta}_1,\hat{\theta}_2)}

=\sqrt{(0.005)^2+(0.004)^2+2(0.000012)} =\sqrt{0.000065} =0.00806\ \text{mmol/L per allele}.

Using a normal approximation, the 95% interval is \(0.020\pm1.96(0.00806)\), or approximately \(0.0042\) to \(0.0358\) mmol/L per allele. This calculation assumes the joint estimate is adequately approximated by a normal distribution and the covariance estimate is suitable. Ignoring the positive covariance would instead give \(SE=\sqrt{0.000041}=0.00640\), an artificially narrower interval in this example. The report should therefore state the overlap and how dependence was handled, not just the final interval.

Make Missingness Assumptions Visible

A missingness report should distinguish observed assay availability from assumptions used to estimate the target. State how many values are missing in each layer, how the analysis handles incomplete records, and which variables are included in that handling. Under a missing-at-random assumption, missingness may depend on observed variables included in the analysis, but not on the unobserved value itself after conditioning on those variables. This is an assumption, not a conclusion established merely by fitting a likelihood.

If missingness may depend on an unobserved molecular value even after conditioning on observed information, the data are not missing at random under that model. The report should then describe sensitivity analyses that encode plausible departures from the primary assumption. Explain the direction and scale of each departure, the analysis refitted, and whether the substantive conclusion changes. A sensitivity result does not identify which assumption is true; it shows how results vary across specified assumptions.

Worked Example: Report Sensitivity to Missing Metabolite Values

Return to the hypothetical cohort of 240 donors, in which 178 have a metabolite measurement and 62 do not. Suppose the primary likelihood analysis assumes that missingness is at random conditional on genotype, age, sex, batch, and the observed molecular layers. It estimates the path-specific effect as \(0.006\) mmol/L per allele, with a 95% interval from \(-0.002\) to \(0.015\).

To assess sensitivity to nonrandom missingness, suppose the analysis is refitted under two explicit pattern-mixture scenarios. In one, the unobserved metabolite values are shifted upward by 0.5 SD relative to the primary model's conditional predictions; in the other, they are shifted downward by 0.5 SD. Suppose the resulting estimates and 95% intervals are \(0.009\), with interval \(0.000\) to \(0.019\), and \(0.003\), with interval \(-0.007\) to \(0.013\), respectively.

The report should state that these are estimates under specified alternative assumptions, not observations of the missing values. Here the point estimate changes across scenarios, while the intervals include or touch zero. The evidence therefore does not support a claim that the conclusion is robustly separated from zero across these missingness assumptions. The shift size, direction, model, and interval method should be reported so another analyst can reproduce the comparison.

Tissue Scope and the Meaning of an Interval

State the tissue or tissues directly represented by the measurements. A result from whole blood is evidence about the analyzed whole-blood measurements; it does not, by itself, establish the same molecular relationship in liver, brain, or every cell type. If evidence is integrated across tissues, report tissue-specific estimates as well as any combined quantity, and explain the rule used to combine them. Do not let a pooled estimate erase tissue-specific differences or imply that all tissues were measured.

Name the interval type and construction. A frequentist confidence interval describes a procedure that, under its assumptions and repeated sampling, covers the fixed estimand at the stated rate. A Bayesian credible interval instead summarizes posterior probability for the parameter conditional on the model, prior, and data. These are different interpretations and should not be used interchangeably. For either kind, give the point estimate, endpoints, units, and method; state whether the interval accounts for dependence, missingness, and uncertainty propagated through the integrated analysis.

Limits on Causal Interpretation

Integrated molecular evidence can be consistent with a proposed pathway without proving that the pathway causes the outcome. A causal claim requires the identification assumptions for the chosen estimand to be credible in the study context. In particular, report the assumptions relevant to confounding, measurement, selection, and the ordering of molecular layers, and distinguish them from results directly checked in the data. An interval quantifies uncertainty conditional on the analysis; it does not repair bias from an invalid identification assumption.

Prefer language such as “the estimated integrated path was…” or “the result was consistent with…” when the design and assumptions do not warrant a causal conclusion. Avoid saying that a molecular layer “explains” an outcome merely because an estimated path is nonzero or an interval excludes zero. Statistical evidence, tissue scope, and causal interpretation are separate parts of the report.

A Reproducible Reporting Sequence

A concise report can still be complete if it follows a consistent sequence. Record the analysis choices that make the estimate reproducible, including the units and transformations used for each layer, the donors included at each step, and the method used to propagate uncertainty. A useful reporting sequence is:

Name the estimand.
Define the target population, genetic contrast, integrated path or combination, outcome scale, and units.

Describe the data used.
Give unique-donor counts, assay-specific sample sizes, cross-layer overlap, tissue scope, and the handling of incomplete records.

State assumptions and uncertainty methods.
Identify missingness and causal assumptions, the interval type, and how uncertainty and dependence were propagated.

Show sensitivity and limits.
Report the alternatives examined, how estimates changed, and what the data do not establish.

Common Reporting Pitfalls

Reporting a number without defining its target. A path product is not self-explanatory. Name the path, contrast, population, outcome scale, and units.
Treating assay counts as independent sample sizes. Several layers can come from the same donors. Report overlap and account for dependence when calculating integrated uncertainty.
Calling missing-at-random an observed fact. It is a modeling assumption conditional on specified observed variables. Describe it and examine plausible departures where warranted.
Using an interval as a causal guarantee. An interval addresses uncertainty under a model and its assumptions; it does not establish those assumptions or rule out bias.
Generalizing beyond measured tissues. A finding in one tissue has that tissue's scope unless further evidence supports broader claims.
Describing a confidence interval as a probability statement about the fixed parameter. State the procedure's repeated-sampling interpretation, or label a Bayesian credible interval and its conditional interpretation correctly.

Key takeaway: A reproducible report of integrated molecular evidence connects a well-defined estimand to its contributing donors, tissues, assumptions, uncertainty interval, and sensitivity results. Its conclusions should be no broader or more causal than those elements support.

Check Your Understanding

Use these questions to check whether a report makes its uncertainty and scope interpretable.

What details are needed to define a path-specific estimand whose molecular intermediates are standardized but whose outcome is measured in mmol/L?
Why can the standard error of a sum of tissue-specific estimates differ from the value obtained by treating them as independent?
What does a missing-at-random assumption mean after conditioning on observed variables, and why is it not directly verified by fitting a likelihood?
How should a sensitivity analysis describe a shift applied to missing molecular values?
What does a 95% frequentist confidence interval mean, and what does it not mean about the fixed estimand?
Why should a whole-blood result not automatically be described as a result for every tissue?