Why Model a Contrast?
When two longitudinal markers carry related information, a survival model can focus on their difference rather than assigning an unrestricted hazard coefficient to each marker. For example, a clinical question might concern whether a patient's first marker is high relative to the second, rather than whether both markers are high together. A contrast represents that question directly, but it also imposes a substantive restriction: the hazard association follows the contrast direction, not every possible combination of marker values.
The markers in the hazard link are latent trajectories, not the noisy measurements recorded at individual visits. Write \(m_{ik}(t)\) for the latent value of marker \(k\) for subject \(i\) at time \(t\), as represented by the longitudinal model. Let \(a_k\) and \(s_k>0\) be fixed reference center and scale values for that marker. They define the standardized latent value \(z_{ik}(t)=[m_{ik}(t)-a_k]/s_k\). The reference values should be stated and used consistently; they might be prespecified or estimated as part of a specified modeling procedure.
Because the markers are standardized separately, one contrast unit means one reference standard deviation higher on marker 1 relative to marker 2. It does not mean one raw unit of either marker. The definition also fixes the direction: exchanging marker 1 and marker 2 changes the sign of the contrast and therefore reverses the coefficient's interpretation.
Hazard Model and Assumptions
A joint model can connect the longitudinal processes to the event hazard through this contrast. Let \(T_i\) be the event time, \(w_i\) a vector of covariates, and \(b_i\) the subject-specific random effects used in the longitudinal model. A proportional hazards specification is
Here \(h_0(t)\) is the baseline hazard, \(\gamma\) is the covariate coefficient vector, and \(\alpha\) is the contrast coefficient. The hazard is conditional on the latent trajectory represented by \(b_i\) and on the covariates. In the joint model, the latent marker values are estimated from the longitudinal data and their model; they are not replaced by the observed visit values in this expression.
Interpretation requires that the longitudinal model represent the marker trajectories adequately and that the event model's conditional hazard form be reasonable. In particular, the specified contrast must be the intended association, its effect is proportional on the hazard scale as written, and the other modeled covariate terms must be appropriate. The joint model also relies on its specified random-effects and dependence structure. Censoring must be noninformative conditional on the modeled information, or be handled by an appropriate censoring model. These assumptions do not make the association causal: unmeasured confounding or model misspecification can still affect it.
What the Contrast Coefficient Estimates
The contrast restriction becomes clearer by comparing it with a model containing separate standardized marker terms. Suppose the linear predictor included \(\alpha_1 z_{i1}(t)+\alpha_2 z_{i2}(t)\). The contrast model sets \(\alpha_1=\alpha\) and \(\alpha_2=-\alpha\), so that the marker contribution is \(\alpha[z_{i1}(t)-z_{i2}(t)]\). Thus a one-unit increase in marker 1's standardized value, with marker 2 held fixed, increases the contrast by one; a one-unit increase in marker 2 alone decreases it by one.
This is a restriction, not just a shorter way to report two unrestricted effects. In the contrast model, a one-unit increase in both standardized markers leaves the difference unchanged and therefore makes no change to this part of the hazard predictor. The contrast coefficient does not estimate a general effect of “marker 1 adjusted for marker 2” unless that phrase is understood specifically as this equal-and-opposite restriction on the standardized scale.
More generally, a prespecified weighted contrast can be written \(C_i(t)=\sum_k c_k z_{ik}(t)\), often with \(\sum_k c_k=0\) when the goal is to compare markers while removing their common level. Its coefficient is interpreted per one-unit increase in that particular weighted contrast. The chosen weights determine the scientific question; they should not be treated as interchangeable or selected after seeing which gives the most appealing result.
Worked Example: Calculating a Difference-Contrast Hazard Ratio
Suppose, hypothetically, that a fitted model uses \(D_i(t)=z_{i1}(t)-z_{i2}(t)\) with \(\hat{\alpha}=0.35\). At a particular time, subject A has \(z_{A1}(t)=0.8\) and \(z_{A2}(t)=-0.4\), while subject B has contrast \(D_B(t)=0.2\). Assume both subjects have the same modeled covariates and compare their conditional hazards at that time.
For subject A, \(D_A(t)=0.8-(-0.4)=1.2\), so the difference in contrast from subject B is \(1.2-0.2=1.0\). The model-implied hazard ratio is
Under this model, subject A's conditional instantaneous hazard is estimated to be about 1.42 times subject B's at that time, or about 41.9% higher. This comparison is conditional on the modeled covariates and other conditioning information; it is not a claim that subject A has a 41.9% greater probability of experiencing the event.
Derivation and Further Interpretations
The hazard-ratio interpretation follows directly by taking the ratio of two conditional hazards at the same time. For subjects \(i\) and \(j\), the shared baseline hazard cancels. If their covariates are equal, the covariate term cancels as well, leaving the contrast contribution:
Consequently, a difference of \(q\) contrast units corresponds to the conditional hazard ratio \(\exp(\alpha q)\). If covariates differ, their contribution must also be included: the log hazard ratio is \(\gamma^\mathsf{T}(w_i-w_j)+\alpha[D_i(t)-D_j(t)]\). The baseline cancellation does not imply that the hazard ratio is a survival-probability ratio, and with time-varying contrasts it need not be constant across time.
Worked Example: A Change in Contrast Over Time
In a hypothetical fitted model, let \(\hat{\alpha}=0.30\). For one subject, suppose the standardized marker values change from \(z_1=0.50,\ z_2=0.25\) at one time to \(z_1=0.90,\ z_2=0.05\) at a later time. To isolate the marker contribution, compare the modeled hazards at those times with the baseline and covariate contributions held fixed.
The first contrast is \(D_{\mathrm{early}}=0.50-0.25=0.25\). The later contrast is \(D_{\mathrm{late}}=0.90-0.05=0.85\), so the change is \(0.85-0.25=0.60\). The hazard ratio attributable to that change is
The model assigns about a 19.7% higher conditional hazard contribution to the later contrast than to the earlier one, when other terms are held fixed. It does not say that the event probability rose by 19.7%; that probability depends on the hazard over time and the rest of the survival model.
Worked Example: Interpreting an Estimate and Its Uncertainty
Suppose a hypothetical analysis gives \(\hat{\alpha}=0.28\) with \(SE(\hat{\alpha})=0.12\). Assume a large-sample normal approximation is suitable for this coefficient and that the standard error reflects the fitted joint model. The estimated hazard ratio per one-unit increase in the contrast is
A 95% Wald confidence interval on the log hazard-ratio scale is \(0.28\pm1.96(0.12)\). The margin is \(1.96(0.12)=0.2352\), giving endpoints \(0.0448\) and \(0.5152\). Exponentiating both endpoints gives
Thus the estimated conditional hazard ratio is 1.323 per one-unit increase in the standardized difference, with an approximate 95% confidence interval from 1.046 to 1.674. Under the model and approximation, the interval excludes a hazard ratio of 1. It is not a 95% probability statement about the fixed parameter. A Wald test of \(H_0:\alpha=0\) gives \(z=0.28/0.12=2.333\), with an approximate two-sided \(p\)-value of 0.0196. This is evidence against no contrast association under the fitted model, not evidence that the contrast causes the event.
Diagnostics, Alternatives, and Limitations
A contrast is useful when the scientific target is a relative marker profile and a lower-dimensional hazard association is appropriate. It uses one coefficient rather than separate coefficients for two markers, which can make the target easier to explain. But fewer coefficients do not guarantee a better model: the restriction may be wrong, and a contrast-only model cannot detect an association with the markers' common level.
| Hazard representation | What it allows | Main interpretive caution |
|---|---|---|
| Difference contrast, \(\alpha(z_1-z_2)\) | Equal-and-opposite standardized marker contributions | Common increases in both markers cancel |
| Separate terms, \(\alpha_1z_1+\alpha_2z_2\) | Different associations for each marker | Coefficients may be imprecise or difficult to distinguish when markers track together |
| Prespecified weighted contrast, \(\alpha\sum_k c_kz_k\) | A specified weighted comparison among markers | The weights define the target and require substantive justification |
Useful checks include confirming the marker ordering, units, reference centers, scales, and contrast sign; checking that the fitted longitudinal trajectories are plausible over the event-risk period; and assessing whether the hazard association appears compatible with the assumed proportional form. Inspect convergence and identifiability evidence before interpreting estimates. Compare the contrast restriction with a scientifically defensible alternative, such as separate marker terms, while recognizing that a more flexible model may be less precise. Large changes in the contrast estimate across reasonable specifications are a warning that the conclusion depends on modeling choices.
A further limitation is that a contrast can hide distinct patterns. For example, two subjects can have the same difference while both markers are low for one and both are high for the other. The contrast model gives those subjects the same marker contribution to the hazard. It also depends on the selected standardization: changing the scales changes what one contrast unit means and therefore changes the numerical coefficient. These are features of the estimand, not merely technical details.
Common Mistake
A common error is to read \(\alpha\) as the effect of marker 1 while “adjusting for” marker 2, then describe marker 2 as having no effect. In fact, the contrast model assigns marker 1 coefficient \(+\alpha\) and marker 2 coefficient \(-\alpha\) on the standardized scale. It imposes equal-and-opposite contributions; it does not estimate an unrestricted marker 1 association and does not show that marker 2 is irrelevant. State the contrast, its scale, and the restriction when reporting the estimate.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(\alpha=0.20\), what is the modeled conditional hazard ratio for a two-unit increase in the contrast, with other terms fixed?
- In a difference-contrast model, what happens to the marker contribution if both standardized marker values increase by 0.5?
- Why is a contrast coefficient not generally the unrestricted effect of marker 1 adjusted for marker 2?
- A 95% confidence interval for \(\alpha\) includes zero. What does that imply for the corresponding hazard-ratio interval?
Solutions. (1) The contrast change is \(q=2\), so the hazard ratio is \(\exp(0.20\times2)=\exp(0.40)\approx1.492\). (2) The difference does not change: \((z_1+0.5)-(z_2+0.5)=z_1-z_2\), so the contrast contribution is unchanged. (3) The model imposes equal-and-opposite standardized coefficients, \(\alpha_1=\alpha\) and \(\alpha_2=-\alpha\), instead of estimating two unrestricted coefficients. (4) Exponentiation is increasing and maps zero to one, so the corresponding hazard-ratio interval includes 1; the data do not exclude a null contrast association at the matching confidence level under the model and approximation.