Why Cross-Omic Interactions Need Careful Interpretation
A feature from one omics layer may have a different association with the hazard of an event at different levels of a feature from another layer. A Cox model can represent this effect modification with a product term. The coefficient on that term is not, by itself, the hazard ratio for either feature: it describes how one feature’s hazard ratio changes as the other feature changes.
Interaction estimates can also be fragile. A fitted model may produce a notable coefficient even when few observed events occur among subjects with particular combinations of the two features. The number and distribution of events, the risk sets in which events occur, and the uncertainty of the estimate all matter when deciding how strongly to interpret apparent interaction evidence.
Model and Assumptions
Let \(T_i\) be the event time for subject \(i\), and let \(\delta_i=1\) if the event is observed and \(\delta_i=0\) if the observation is censored. Let \(X_i\) and \(Z_i\) be measured features from different omics layers, and let \(W_i\) denote any additional covariates included in the model. A simple interaction model is
Here \(h_i(t\mid\cdot)\) is the subject-specific hazard at time \(t\), \(h_0(t)\) is an unspecified baseline hazard, and \(\beta_X,\beta_Z,\theta,\boldsymbol{\gamma}\) are regression coefficients. The model assumes proportional hazards: for any two subjects with fixed covariates, their hazard ratio does not change over time. It also assumes the log hazard is linear in the listed terms, including the product, and that subjects’ outcomes are independent conditional on the model covariates. Censoring must be non-informative conditional on those covariates for the usual Cox partial-likelihood inference to apply.
The features and the product term must be defined before fitting. For continuous features, units and scales affect coefficient interpretation; standardizing each feature, for example, makes a one-unit contrast correspond to one standard deviation if that is how the variables were standardized. Centering can make the main effects easier to interpret. For binary features, the product term is one only when both features equal one.
What the Interaction Coefficient Means
At a fixed value \(z\) of \(Z\), a one-unit increase in \(X\) changes the log hazard by \(\beta_X+\theta z\). Thus its conditional hazard ratio is \(\exp(\beta_X+\theta z)\). At two values \(z_1\) and \(z_0\), the ratio of those conditional hazard ratios is
For a one-unit contrast in \(Z\), the ratio is \(\exp(\theta)\). So \(\theta=0\) means the hazard ratio for a one-unit change in \(X\) does not vary with \(Z\) on the multiplicative hazard scale. It does not mean that the two features are unrelated, nor does it rule out effect modification on another scale.
With uncentered continuous features, \(\beta_X\) is the log hazard ratio for a one-unit increase in \(X\) when \(Z=0\), and \(\beta_Z\) is the corresponding effect when \(X=0\). If zero is not a scientifically meaningful or observed feature value, center the features at stated reference values before fitting or clearly report the conditional reference values. Centering changes the interpretation of the main-effect coefficients, but does not change the interaction contrast per unit change in each feature.
Worked Example: Interpreting an Interaction With Standardized Features
Suppose \(X\) and \(Z\) are standardized features, so one unit represents one standard deviation. A fitted Cox model has \(\beta_X=0.30\), \(\beta_Z=-0.20\), and \(\theta=0.50\). Interpret the association of \(X\) with hazard at \(Z=0\) and \(Z=1\).
At \(Z=0\), the log hazard change for a one-standard-deviation increase in \(X\) is \(0.30+0.50(0)=0.30\). The conditional hazard ratio is \(\exp(0.30)=1.350\).
At \(Z=1\), that log hazard change is \(0.30+0.50(1)=0.80\). The conditional hazard ratio is \(\exp(0.80)=2.226\). The ratio of these two hazard ratios is \(2.226/1.350\approx1.649\), which agrees with \(\exp\{0.50(1-0)\}=\exp(0.50)=1.649\).
In this hypothetical model, the estimated hazard ratio for a one-standard-deviation increase in \(X\) is larger at \(Z=1\) than at \(Z=0\). This is an association conditional on the modeled covariates; it does not show that changing either feature would cause a change in hazard.
Testing and Quantifying Interaction Evidence
A common inferential target is \(H_0:\theta=0\) against \(H_a:\theta\ne0\). A Wald test uses the estimate \(\hat{\theta}\) and its standard error \(SE(\hat{\theta})\), typically obtained from the fitted Cox model. Under the null and regular large-sample conditions, the Wald statistic is approximately standard normal. A confidence interval for the coefficient is \(\hat{\theta}\pm1.96SE(\hat{\theta})\) for an approximate 95% interval; exponentiating its endpoints gives an interval for the one-unit interaction ratio \(\exp(\theta)\).
A likelihood-ratio test compares the partial likelihood for a model containing the interaction with that for the nested model omitting it. With one added interaction coefficient, twice the difference in maximized partial log likelihoods is approximately chi-squared with one degree of freedom under the null and regularity conditions. Both approaches rely on adequate information for the parameters and appropriate model assumptions. With sparse event support, their approximations and the stability of the estimate deserve particular scrutiny.
Worked Example: Wald Evidence and an Interaction-Ratio Interval
In a hypothetical fitted Cox model, suppose the interaction estimate is \(\hat{\theta}=0.90\), with \(SE(\hat{\theta})=0.45\). The Wald statistic is
The two-sided normal-approximation p-value is approximately \(0.0455\). The estimated one-unit interaction ratio is \(\exp(0.90)=2.460\). The coefficient-scale 95% interval is
Exponentiating gives the interval for the interaction ratio:
Under the model and approximation, the interval corresponds to interaction ratios ranging from about 1.018 to 5.942 per one-unit contrast in each feature. It is wide, so the magnitude is imprecisely estimated. The p-value is the probability, under the null and the test’s assumptions, of a statistic at least as extreme as the observed one; it is not the probability that the null is true.
Why Event Support Matters
Cox regression obtains information through comparisons between the subject experiencing an event and other subjects still at risk at that event time. Therefore, total sample size alone does not establish that an interaction is well supported. Relevant considerations include the total number of observed events relative to the number of estimated parameters, the spread of the two features, how often their combinations occur, and whether those combinations appear in informative risk sets.
A simple count of events in four groups formed by binary features can reveal a sparse combination, but those counts do not reproduce the Cox partial likelihood: event times, censoring, and risk-set membership also matter. Likewise, a global events-per-parameter ratio is a rough warning indicator, not a universal adequacy threshold. Sparse combinations can lead to unstable estimates, large standard errors, sensitivity to a few events, and poor approximation by standard tests. Strong correlation among features or between a product term and its main effects can add to the problem.
Worked Example: A Sparse Joint Feature Group
Consider a hypothetical cohort of 200 subjects with two binary features. The group counts are: neither feature present, 80 subjects and 18 events; only \(X\) present, 45 subjects and 6 events; only \(Z\) present, 50 subjects and 7 events; both present, 25 subjects and 4 events. There are \(18+6+7+4=35\) observed events. A model with \(X\), \(Z\), and \(XZ\) has three regression coefficients, giving a crude overall ratio of \(35/3\approx11.7\) events per coefficient.
That overall ratio does not remove the concern that the joint-present group has only four events. Suppose a fitted model reports the interaction estimate and standard error from the preceding example. The estimate may be driven substantially by a small number of event comparisons involving that group. The group counts alone cannot establish that this happened, because the relevant Cox risk sets depend on event times and censoring; they do identify a reason to inspect the risk-set data and the estimate’s stability.
A careful report would give the interaction estimate and interval, describe the event distribution, and state that the estimate may be sensitive to sparse support. It would not treat the nominal p-value just below 0.05 as decisive evidence. Sensitivity checks, such as examining influential observations or fitting a prespecified penalized model, can help assess instability, but do not create information absent from the data.
A Comparison of Modeling Choices
| Approach | What it represents | Key caution |
|---|---|---|
| Additive Cox model, without \(XZ\) | Each feature has a constant conditional hazard ratio over the other feature’s values. | It may miss effect modification if the interaction is real. |
| Cox model with \(XZ\) | The log hazard includes a product term, allowing the conditional effect of each feature to vary linearly with the other. | Needs sufficient event and risk-set support; interpretation depends on feature scale and reference values. |
| Separate subgroup models | Estimates associations within selected strata of one feature. | Can discard information, yield imprecise comparisons, and make formal differences between effects difficult to assess. |
| Penalized interaction model | Regularization can stabilize estimation when many candidate terms or correlated features are considered. | Penalty choice affects estimates; ordinary unpenalized p-values do not automatically account for selection or shrinkage. |
For a single prespecified interaction, a conventional Cox model is often the clearest way to state the target and its uncertainty when event support is adequate. In a high-dimensional search over many feature pairs, selecting an interaction because it looks promising and then reporting its ordinary Wald p-value as if it had been prespecified exaggerates evidence. Candidate selection, multiplicity, and any regularization must be reflected in the analysis plan and interpretation.
Diagnostics, Limitations, and Common Pitfalls
Before interpreting an interaction, check that the model’s proportional-hazards assumption is plausible for its terms, that the product and main effects are correctly constructed, and that the data contain variation in both features and their combinations. Inspect event counts and risk-set representation across relevant combinations. Examine standard errors, intervals, convergence warnings, and sensitivity to influential observations. An extremely large coefficient or standard error, unstable signs under small specification changes, or near-empty feature combinations are warning signs, not evidence of a dramatic biological interaction.
The product term captures a particular form of modification: linear variation in the log hazard with the other feature. It does not cover every possible relationship, and the interaction is scale-specific. A zero product-term coefficient rules out this form of modification on the modeled log-hazard scale, not all possible forms of dependence. The model also remains observational: a fitted interaction is an association and should not be stated as a causal mechanism without additional design and assumptions.
A Practical Assessment Sequence
Specify the features, their units or coding, and the values being compared. Center continuous features at meaningful references when useful.
Include both main effects when including their product, along with prespecified covariates. State the proportional-hazards and censoring assumptions.
Report the interaction coefficient, its standard error, a confidence interval, and a prespecified test if one is used. Translate the coefficient into the corresponding ratio of hazard ratios.
Summarize events and risk-set representation across combinations, inspect diagnostics, and assess sensitivity to sparse groups or influential observations.
Distinguish an estimated association from causation, and state when event sparsity or uncertainty limits interpretation.
Check Your Understanding
Try each problem, then compare your reasoning with the solutions below.
- If \(\hat{\theta}=0.40\), what is the estimated ratio of hazard ratios for a one-unit increase in each feature?
- Why does centering \(Z\) change the interpretation of \(\beta_X\) but not the interaction ratio for a one-unit change in each feature?
- A model has 60 events and estimates four regression coefficients. Calculate the crude events-per-coefficient ratio, and explain why it cannot by itself establish adequate interaction support.
- What does a 95% confidence interval for \(\exp(\theta)\) describe in this model?
Solutions. (1) \(\exp(0.40)=1.492\), so the estimated hazard ratio for a one-unit change in \(X\) is 1.492 times as large at a \(Z\) value one unit higher. (2) Centering changes which value of \(Z\) is represented by zero, and hence the reference at which the \(X\) main effect applies. The slope of the \(X\) effect with respect to \(Z\), \(\theta\), is unchanged by a shift of origin, so \(\exp(\theta)\) is unchanged. (3) \(60/4=15\) events per coefficient. This is only a crude summary: events may be sparse for a particular feature combination, and Cox information depends on event times and risk sets. (4) Under the fitted model and its assumptions, it gives a range of values compatible with the data for the ratio of conditional hazard ratios for the stated one-unit contrasts; it is not a 95% probability statement about the fixed parameter.