Positivity and Extreme Weights in Incomplete Omics Studies

Why Positivity Matters for Modality Weighting

Inverse-probability weighting can adjust for differences between subjects whose omics data are available and those whose data are not, provided that the observed subjects represent every part of the target population after conditioning on measured covariates. Positivity is the support condition behind that requirement. If a subgroup has no chance of contributing the modality needed for an analysis, its outcomes cannot be recovered by assigning larger weights to subjects from other subgroups.

In an incomplete multi-omics study, availability may concern one modality or a set of modalities needed jointly. A subject can have one assay available but still be unavailable for an analysis requiring two assays. Positivity must therefore be checked for the actual analysis and its target population, not inferred from a general statement that “some omics data are available.”

For a joint analysis requiring modalities \(A\) and \(B\), \(R=1\) means that both are available. The relevant probability is then \(P(R_A=1,R_B=1\mid X)\), not either modality’s marginal availability probability. The same principle applies to a feature or analysis whose observation indicator differs from that of another analysis. The conditioning variables must be available for the target subjects and adequate for the missing-at-random condition used by the weighting analysis; positivity alone does not make that condition true.

Why the Weighted Estimator Needs Support

For a target-population mean \(\theta=E(Y)\), the normalized inverse-probability-weighted estimator is

Here \(\hat e(X_i)\) estimates the probability that subject \(i\)'s required modality data are observed. Only observed outcomes contribute to the numerator. The weighting argument also requires conditional exchangeability, often expressed as \(Y\perp R\mid X\): among subjects with the same \(X\), availability does not further select subjects according to \(Y\). It requires a correctly specified or sufficiently accurate availability model as well. Positivity is a separate requirement: even if those conditions hold, the observed data must include subjects at every relevant covariate value.

The justification is easiest to see using the true \(e(X)\). If \(Y\perp R\mid X\) and \(e(X)>0\), then

Taking expectations gives \(E[RY/e(X)]=E(Y)\). Similarly, \(E[R/e(X)]=1\), so the population ratio of the weighted outcome total to the weighted observation total is \(E(Y)\). This is the basis of the normalized estimator; it is not an exact finite-sample identity. If \(e(X)=0\) in a target stratum, the ratio \(R/e(X)\) is undefined there, and no observed \(Y\) values from that stratum can contribute. If \(e(X)\) is merely very small, the identity still holds under the assumptions, but the finite-sample estimator can be highly unstable.

Diagnosing Sparse Support and Extreme Weights

A small estimated availability probability produces a large inverse weight. A few such weights can dominate a weighted estimate, so inspect both availability and the consequences of weighting. Useful checks include tabulating observed and target counts across covariate strata and modality-availability patterns; examining the distribution of \(\hat e(X_i)\) and \(1/\hat e(X_i)\); and reporting the maximum and upper quantiles of the weights. Also check whether weighted covariate distributions are better aligned with the target population. Balance is informative, but good balance does not establish positivity or prove that the unobserved outcomes are recoverable.

An effective sample size summarizes weight concentration. For the observed subjects with weights \(w_i=1/\hat e(X_i)\), define

If all weights are equal, this equals the number of observed subjects. As weights become more concentrated, it decreases. It is a diagnostic summary—not a replacement sample size for every variance calculation, nor a test of the weighting assumptions.

Very small fitted probabilities can arise from genuinely rare availability, a small sample, a poorly specified availability model, or a combination. Inspect model fit, but do not treat a good fit as a solution to sparse support. A model can predict availability well overall while assigning tiny probabilities to a clinically or biologically important subgroup. Conversely, a smooth fitted probability bounded away from zero does not show that the subgroup actually supplied observed outcomes.

Responding to Positivity Problems

The appropriate response depends on whether the problem is structural or practical. When the required modality is never observed in a target stratum, weighting cannot identify an unrestricted target-population mean from the observed outcomes. A narrower target population with adequate support may be defensible if it answers a clearly stated scientific question; it changes the estimand and should be reported as such. Alternatively, improving data collection to obtain modality measurements in the unsupported group directly addresses the design gap.

When availability is rare but nonzero, options include revising an unstable availability model where scientifically justified, collecting more data in sparse strata, or presenting sensitivity analyses that show how results change under reasonable choices. Weight truncation or capping limits the influence of large weights, but it generally trades variance for bias: the modified weights no longer exactly represent the original inverse-probability correction. Report the rule used and compare the resulting estimate with the untruncated estimate. Truncation does not solve structural nonpositivity.

Complete-case analysis, weighted analysis, and model-based approaches such as likelihood or multiple imputation rely on different assumptions; none can recover information that is wholly absent without additional assumptions. In particular, a model-based analysis may extrapolate into an unsupported stratum, but its answer then depends on that model rather than empirical coverage there. The earlier material on inverse-probability weighting for modality availability introduced the weighting approach; here the central practical lesson is to establish support before interpreting its estimate.

Common Mistakes and Limits

• Confusing good model fit with positivity. Predicting observed availability accurately does not guarantee that every target stratum contains observed modality data.
• Calling every large weight a proof of bias. Extreme weights signal potential instability and weak practical support. Bias depends on the assumptions and the data-generating process, not on the weight size alone.
• Assuming marginal availability is enough. If an analysis requires two modalities jointly, check the probability of observing both, conditional on relevant covariates.
• Treating truncation as a repair for zero probability. Capping weights can reduce variance in sparse but supported data; it cannot supply outcomes for a stratum with no observed subjects.
• Ignoring the target population. Positivity is defined relative to the covariate values in that population. Restricting the population may restore support, but it changes the question being answered.

No diagnostic can establish the missing-at-random assumption from observed data alone, and weight summaries cannot certify that the availability model is correct. They can expose unsupported or weakly supported regions where weighted conclusions are especially dependent on assumptions or modeling choices.