Transition Regression in a Transplant Example

Why a Transplant Model Needs Transition-Specific Effects

After a transplant, a recipient may continue with a functioning graft, experience graft failure, or die. A recipient whose graft fails may subsequently die. These pathways are not interchangeable: recipient age could be associated differently with graft failure while the graft is functioning than with death after graft failure. A single regression coefficient for age would impose the same relative effect on every transition, whether or not that restriction is plausible.

We will represent the three states as \(T\), \(G\), and \(D\): \(T\) is alive after transplant with a functioning graft, \(G\) is alive after graft failure, and \(D\) is death. The allowed transitions are \(T\to G\), \(T\to D\), and \(G\to D\). Death is absorbing in this model. The analysis begins at transplant, and time \(t\) is time since transplant. Recipient age at transplant is treated as a fixed baseline covariate and is measured in decades, centered at 50 years.

Definition: A transition-specific hazard \(\lambda_{rs}(t\mid\mathcal{H}(t))\) is the instantaneous rate of moving from state \(r\) to state \(s\) at time \(t\), conditional on the observed history \(\mathcal{H}(t)\) and on being in state \(r\) immediately before \(t\). Its units are inverse time. For example, the \(T\to G\) hazard concerns recipients who are alive with a functioning graft and have not yet left state \(T\).

The hazard describes an instantaneous transition rate among people currently eligible for that transition. It is not the probability of ever making the transition, nor does it summarize all subsequent pathways. The risk set for \(T\to G\) consists of recipients still in state \(T\); a transition to death removes someone from that risk set because that person can no longer experience graft failure while in state \(T\). Recipients in state \(G\) contribute to the \(G\to D\) risk set, not to either transition out of \(T\).

A Transition-Specific Regression Model

For each allowed transition, use its own baseline hazard and its own age coefficient. Write \(A_i=(\text{age at transplant for recipient }i-50)/10\), so a one-unit increase in \(A_i\) means an age difference of 10 years. A proportional-hazards specification is

\lambda_{rs,i}(t\mid\mathcal{H}_i(t)) = \lambda_{0,rs}(t)\exp\{\beta_{rs}A_i\}, \qquad rs\in\{TG,TD,GD\}.

Here \(\lambda_{0,rs}(t)\) is the transition-specific baseline hazard for a recipient aged 50 at transplant, and \(\beta_{rs}\) is the age coefficient for that particular transition. For instance, \(\beta_{TG}\) describes the association between age and the instantaneous \(T\to G\) rate, conditional on being in state \(T\). It is not assumed to equal \(\beta_{TD}\) or \(\beta_{GD}\).

Interpretation: For a 10-year increase in age at transplant, the transition-specific hazard ratio for transition \(r\to s\) is \(\exp(\beta_{rs})\), under the model and conditional on the same transition-specific risk set and time. A hazard ratio above 1 means a higher instantaneous transition rate for the older recipient; it does not by itself establish a higher probability of that transition by a specified follow-up time.

This model assumes a log-linear age effect and proportional hazards within each transition: the age hazard ratio for a given transition does not change with time since transplant. It also assumes the specified transition structure is adequate, age is measured appropriately, and censoring is independent of future transition times conditional on the modeled information. Other modeling choices, such as additional time-varying covariates, require corresponding assumptions and careful definition of their values across state changes.

The transition-specific coefficients can be estimated by fitting a stratified Cox model or, when an appropriate piecewise-constant baseline is used, a piecewise-exponential model. In either approach, a transition contributes its own event and risk-set information. The earlier material on Poisson representation of piecewise-exponential models gives the corresponding likelihood representation; here the central task is to specify and interpret the age effect separately for each pathway.

Worked Interpretations and Comparisons

Worked Example: Interpreting Age Effects on Three Transitions

Suppose a hypothetical analysis estimates \(\hat\beta_{TG}=0.25\), \(\hat\beta_{TD}=0.15\), and \(\hat\beta_{GD}=0.55\). These estimates are illustrative and are not results from a real transplant study. Calculate and interpret the age hazard ratio for each transition.

For a 10-year age increase, the estimated transition-specific hazard ratio is \(\exp(\hat\beta_{rs})\). Thus, for graft failure while in state \(T\),

\exp(0.25)=1.284.

Conditional on being alive with a functioning graft at the same time since transplant, the estimated instantaneous \(T\to G\) hazard is 1.284 times as high for a recipient 10 years older. For direct death from state \(T\), \(\exp(0.15)=1.162\), or an estimated 16.2% higher instantaneous \(T\to D\) hazard per 10 years. For death after graft failure, \(\exp(0.55)=1.733\), or an estimated 73.3% higher instantaneous \(G\to D\) hazard per 10 years.

These comparisons are transition-specific. The estimate 1.733 compares recipients in state \(G\) at the same time since transplant; it does not compare overall mortality probabilities between older and younger recipients. Also, the coefficient alone does not say whether the association is statistically distinguishable from zero; that requires its uncertainty.

Worked Example: Testing Whether Two Age Effects Differ

Continue with the hypothetical estimates \(\hat\beta_{TG}=0.25\) and \(\hat\beta_{GD}=0.55\). Suppose their estimated standard errors are 0.08 and 0.10, and their estimated covariance is 0.001. Test whether the age coefficients are equal using a two-sided Wald test at \(\alpha=0.05\).

The estimated contrast is \(\hat\Delta=\hat\beta_{GD}-\hat\beta_{TG}=0.55-0.25=0.30\). Its variance must include the covariance between coefficient estimates:

\widehat{\mathrm{Var}}(\hat\Delta) = 0.10^2+0.08^2-2(0.001) = 0.0100+0.0064-0.0020 = 0.0144.

Therefore \(SE(\hat\Delta)=\sqrt{0.0144}=0.12\), and the Wald statistic is \(z=0.30/0.12=2.50\). The two-sided standard normal p-value is approximately 0.0124. Since 0.0124 is below 0.05, the data provide evidence against equal age coefficients for these two transitions, subject to the model assumptions.

A 95% confidence interval for the coefficient contrast is \(0.30\pm1.96(0.12)=(0.0648,0.5352)\). Exponentiating its endpoints gives \((1.067,1.708)\). This is an interval for \(\exp(\beta_{GD}-\beta_{TG})\), the ratio of the two 10-year hazard ratios—not a hazard ratio for one transition against a shared population risk. Under the fitted model, the estimated age hazard ratio for \(G\to D\) is about 1.35 times the estimated age hazard ratio for \(T\to G\), with a 95% confidence interval from 1.067 to 1.708.

Worked Example: Comparing a 20-Year Age Difference

Using the same hypothetical coefficients, compare recipients aged 70 and 50 at transplant. Their age difference is 20 years, or two units of \(A\). For each transition, the model gives a hazard ratio of \(\exp(2\hat\beta_{rs})\).

For \(T\to G\), \(\exp(2\times0.25)=\exp(0.50)=1.649\). For \(T\to D\), \(\exp(2\times0.15)=\exp(0.30)=1.350\). For \(G\to D\), \(\exp(2\times0.55)=\exp(1.10)=3.004\).

Thus, under the model, the estimated instantaneous \(G\to D\) hazard for the 70-year-old recipient is 3.004 times that for the 50-year-old recipient, conditional on both being in state \(G\) at the same time since transplant. This is a model-based relative hazard, not an estimate of the probability that either recipient will die by a particular time.

Justification for the Hazard-Ratio Interpretation

The interpretation follows directly by dividing the two hazards within the same transition. For recipients with covariate values \(A\) and \(A+1\), the baseline hazard is the same at a fixed time \(t\), so

\frac{\lambda_{rs}(t\mid A+1)}{\lambda_{rs}(t\mid A)} = \frac{\lambda_{0,rs}(t)\exp\{\beta_{rs}(A+1)\}} {\lambda_{0,rs}(t)\exp\{\beta_{rs}A\}} = \exp(\beta_{rs}).

The baseline hazard cancels, and the age increment is one decade by construction. For an age difference of \(d\) years, the same calculation gives \(\exp\{\beta_{rs}(d/10)\}\). The argument applies separately to each transition; it does not justify dividing hazards from different transitions to obtain an overall risk comparison.

To compare two transition effects, define a contrast such as \(\Delta=\beta_{GD}-\beta_{TG}\). The variance formula for a difference is

\mathrm{Var}(\hat\Delta) = \mathrm{Var}(\hat\beta_{GD}) + \mathrm{Var}(\hat\beta_{TG}) - 2\,\mathrm{Cov}(\hat\beta_{GD},\hat\beta_{TG}).

The covariance matters because the same recipients may contribute information to more than one transition-specific risk set. A valid standard error for the contrast therefore comes from the joint covariance estimate, including any appropriate robust variance estimation for the analysis design. Comparing separate confidence intervals, or noting that one coefficient is significant and another is not, is not a test of whether the coefficients differ.

Model Checks, Alternatives, and Limitations

Diagnostics should be considered transition by transition. A proportional-hazards pattern that looks reasonable for \(T\to G\) need not hold for \(G\to D\). Examine transition-specific residual or time-interaction diagnostics for age, and plot estimated effects over time where appropriate. A marked time trend in an age effect suggests that a constant \(\beta_{rs}\) may be inadequate; possible responses include a time-varying age coefficient or a more flexible model, with the added complexity justified by the available events.

Also check whether each transition has enough events to support its model terms. Few \(G\to D\) events, for example, can produce unstable estimates and wide intervals even when the total number of deaths is large. Influence diagnostics can reveal whether a small number of recipients dominate a transition-specific estimate. Verify that transition dates, state histories, and the treatment of censoring are consistent with the stated risk sets. A model cannot repair incorrectly classified transitions or unrecorded events.

The main alternative to fully transition-specific age coefficients is a shared-effect model that imposes \(\beta_{TG}=\beta_{TD}=\beta_{GD}\). This can be more precise if the restriction is scientifically defensible, but it can conceal genuine differences. A model with transition-specific coefficients is more flexible but uses more parameters and may be imprecise when events are sparse. A formal contrast or joint test can assess restrictions; the choice should also consider scientific meaning, not only a p-value. The earlier material on transition-specific terms and shared transition effects develops those tests.

Finally, a transition-specific hazard ratio is conditional on occupying the relevant starting state and on the model's time and covariate specification. It does not alone provide an absolute transition probability, an overall survival comparison, or a causal effect of age. Those conclusions require additional quantities and assumptions. Age is not a randomized exposure, and differences in age-associated hazards may reflect other measured or unmeasured differences between recipients.

Common mistake: “The \(G\to D\) age coefficient is significant, while the \(T\to G\) coefficient is not, so age affects those transitions differently.”
Correction: Significance in one model and nonsignificance in another does not establish a difference. Estimate and test the coefficient contrast directly, using its covariance-aware standard error.
Common mistake: “The \(G\to D\) hazard ratio is 1.73, so older recipients are 73% more likely to die.”
Correction: The fitted interpretation is a 73% higher instantaneous \(G\to D\) hazard per 10 years of age, conditional on being in state \(G\). It is not a 73% increase in death probability over follow-up.

Key takeaway: Specify an age coefficient for each clinically meaningful transition, define the age increment and time origin, and interpret each hazard ratio within its own risk set. Use a direct covariance-aware contrast to compare transition effects.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In the three-state model, which recipients are in the risk set for \(G\to D\)?
If \(\hat\beta_{TG}=0.20\) per decade, what is the estimated \(T\to G\) hazard ratio for a 10-year age difference?
With the same coefficient, what is the estimated hazard ratio for a 30-year age difference?
Why is it not valid to conclude that two age effects differ just because one coefficient has \(p<0.05\) and the other has \(p>0.05\)?

Solutions. (1) The risk set includes recipients who are alive with graft failure immediately before the relevant time and have not yet died or been censored. (2) \(\exp(0.20)=1.221\), so the estimated instantaneous \(T\to G\) hazard is 1.221 times as high per 10-year age increase, conditional on being in state \(T\). (3) A 30-year difference is three decades, giving \(\exp(3\times0.20)=\exp(0.60)=1.822\). (4) The two separate tests each assess whether an individual coefficient equals zero, not whether the coefficients equal each other. Test their difference directly and use the covariance between estimates to calculate its standard error.