Quantifying Uncertainty In Factor Interpretation

Why Factor Interpretations Need Uncertainty

A factor model summarizes structured variation across subjects and omics features, but an estimated score or loading is not known without uncertainty. A subject may have a high estimated score while the posterior distribution still includes values near zero. A feature may have a mostly positive loading, yet retain a meaningful probability of a negative loading. These distinctions matter: an interpretation should reflect what the fitted model supports, not only the direction of a point estimate.

Let \(z_{ik}\) denote the score of subject \(i\) on factor \(k\), and let \(\lambda_{jk}\) denote the loading of feature \(j\) on that factor. In a multi-omics setting, \(j\) identifies a feature in a specified view. A generic factor representation includes terms such as \(z_{ik}\lambda_{jk}\) in the modeled signal for a measurement. This tutorial focuses on uncertainty in the scores and loadings themselves, not on decomposing total variance or checking model fit.

A posterior interval is conditional on the model, prior, data, and parameterization used to obtain it. It is not a statement that there is a 95% chance the fixed parameter lies in this particular interval under repeated sampling. Nor is a sign probability the probability that a scientific hypothesis is true independently of the model. These summaries express model-based uncertainty and should be reported with that scope.

Alignment and Conditions for Meaningful Summaries

Posterior summaries require draws that refer to the same factor across iterations. Factor models can have sign and permutation ambiguity: reversing a factor's sign while reversing its scores leaves the fitted products unchanged, and exchanging factor labels can leave the model unchanged. If draws are pooled before matching and orientation are resolved, a loading that is consistently positive in one orientation can appear to have a posterior centered near zero simply because equivalent draws use both orientations.

Therefore, summarize draws only after applying a consistent matching and orientation convention, as in the earlier material on matching factors under sign and permutation ambiguity. The scale must also be comparable across draws. If the model permits arbitrary rescaling of scores and loadings, a score interval is not interpretable until a scale constraint or common normalization has been imposed. Sign probabilities also depend on the chosen orientation: they have meaning relative to that convention, not as an orientation-free biological property.

The basic summaries assume that posterior draws adequately represent the target posterior distribution. Convergence and mixing should be assessed using the earlier material on diagnosing convergence in factor estimation. Strong autocorrelation does not necessarily bias the empirical quantiles, but it reduces the effective information in the draws and makes Monte Carlo error larger than a calculation based on the raw draw count would suggest. These summaries also condition on the specified model and prior; sensitivity to those choices remains relevant.

Computing Intervals and Sign Probabilities

For \(M\) aligned draws \(\theta^{(1)},\ldots,\theta^{(M)}\), sort the draws and take the empirical 2.5th and 97.5th percentiles to form a central 95% interval. Estimate the positive sign probability by the fraction of draws greater than zero:

Here \(I(\cdot)\) equals 1 when its condition holds and 0 otherwise. For independent posterior draws, the Monte Carlo standard error of this estimated probability is approximately \(\sqrt{\hat{p}(1-\hat{p})/M}\), where \(\hat{p}\) is the estimated sign probability. With correlated MCMC draws, replace \(M\) with an appropriate effective sample size for the indicator sequence, or use a suitable chain-based Monte Carlo error estimate.

An interval and a sign probability answer related but different questions. An interval describes the range and spread of posterior values. A sign probability summarizes how much posterior mass lies on one side of zero. A narrow interval entirely above zero indicates both limited posterior spread and a positive direction in the chosen orientation. A wider interval that crosses zero can coexist with a high positive sign probability: most draws may be positive while a smaller number of more extreme negative draws still place a lower quantile below zero.

From Individual Parameters to Biological Patterns

Biological interpretations often involve a pattern across several features rather than one loading. Report uncertainty for the specific loadings that support the claim, and distinguish uncertainty about each loading from uncertainty about the whole pattern. Marginal sign probabilities cannot generally be multiplied to obtain a joint probability, because loadings in a factor model may be dependent in the posterior.

This kind of qualification is especially important when a pattern is selected after inspecting many features. The uncertainty summaries for selected features do not automatically account for the selection process or provide a calibrated error rate for all possible patterns. Use them to describe posterior support under the fitted model, not as a substitute for a separate multiplicity or validation analysis.

Diagnostics, Alternatives, and Limitations

Before interpreting an interval or sign probability, check whether the summary is behaving plausibly. Useful warning signs include intervals that change substantially across chains, sign probabilities that vary with factor alignment, high autocorrelation in the sign indicator, and intervals that are implausibly narrow given weak identification. For a claimed multi-feature pattern, calculate its joint draw-based probability directly rather than inferring it from marginal summaries.

• Unresolved sign or label switching: intervals can be misleading if equivalent factor orientations or labels are mixed together. Align draws first and state the orientation convention.
• Poor effective sample size: a large raw draw count can overstate Monte Carlo precision when draws are highly correlated. Check convergence and Monte Carlo error for both parameters and sign indicators.
• Intervals crossing zero: this does not mean that positive and negative values are equally probable. Inspect the sign probability and the interval endpoints together.
• Overconfident biological language: a high posterior probability for a loading direction is conditional on the model and does not establish a causal mechanism or independent replication.

A posterior interval differs from a frequentist confidence interval in its probability interpretation: it summarizes posterior probability conditional on the model and observed data. Bootstrap stability instead asks how estimates vary under a resampling scheme and is sensitive to the resampling design; it is not the same as posterior uncertainty conditional on one fitted model. These approaches can complement one another, but neither should be relabeled as the other.

Posterior summaries have further limits. They can appear precise when the model is misspecified, when the prior strongly constrains parameters, or when factor identity is weakly determined. Sign probabilities depend on orientation, and magnitudes depend on scale. A posterior interval for a loading alone also does not measure the biological importance of that feature. Interpretations should combine uncertainty summaries with the scientific context and the earlier material on interpreting loading patterns across omics and pathway enrichment.