Recalibrating Baseline Transition Risks

Why Recalibrate a Transition Baseline?

An external validation population can have a different underlying rate of moving from one state to another, even when the original model’s relative covariate effects remain useful. For example, a transition may occur more often in the validation setting because its baseline risk is higher, while a particular covariate still multiplies that transition’s hazard by the same factor as before. Baseline recalibration updates the transition’s underlying hazard to reflect the new population while keeping the original covariate coefficients fixed.

This is a targeted update, not a refit of the entire model. It is appropriate when the validation data support estimating a new baseline but are not being used to replace the original relative-effect estimates. The transition’s baseline is updated using observed transition and risk-time information; predictions for individuals then combine that baseline with the retained covariate effects.

Definition: For an allowed transition from state $r$ to state $s$, a transition-specific proportional-hazards model has intensity $$ \lambda_{rs}(t\mid x)=\lambda_{0,rs}(t)\exp(\beta_{rs}^{\mathsf T}x). $$ Here, $\lambda_{rs}(t\mid x)$ is the instantaneous rate of transition $r\to s$ at time $t$, conditional on being in state $r$ and having covariate vector $x$; $\lambda_{0,rs}(t)$ is the baseline transition hazard; and $\beta_{rs}$ is the transition-specific coefficient vector. Baseline recalibration estimates a new $\lambda_{0,rs}(t)$ from validation data while treating $\beta_{rs}$ as fixed at its original-model value.

The baseline is the hazard for a subject whose covariates are zero under the chosen coding. It is not generally the average hazard or the population’s average transition probability. A transition probability also depends on the covariates, time, and any other ways to leave the current state.

Estimating a Validation Baseline With Fixed Effects

Suppose the original model provides $\beta_{rs}$, and validation follow-up records which subjects are at risk for transition $r\to s$ and when that transition occurs. Write $Y_r(u)$ for the indicator that a subject is in state $r$ and under observation just before time $u$, and $N_{rs}(t)$ for the number of observed $r\to s$ transitions by time $t$. For a tied event time $u$, the Breslow increment of the updated baseline cumulative hazard is

\Delta\widehat{\Lambda}_{0,rs}(u) = \frac{d_{rs}(u)} {\displaystyle\sum_i Y_{ir}(u)\exp(\beta_{rs}^{\mathsf T}x_i)}, \qquad \widehat{\Lambda}_{0,rs}(t)= \sum_{u\le t}\Delta\widehat{\Lambda}_{0,rs}(u),

where $d_{rs}(u)$ is the number of $r\to s$ transitions at time $u$, and the coefficients in the denominator are the fixed original-model coefficients. The denominator is a covariate-weighted risk set: subjects with higher modeled relative transition rates contribute more to the baseline-hazard estimation. The resulting validation prediction for a subject with covariates $x$ uses $\widehat{\Lambda}_{0,rs}(t)\exp(\beta_{rs}^{\mathsf T}x)$.

The standard assumptions matter. The transition model must use the same covariate definitions and transition structure in the validation population; the relative effects must be adequately transported; and censoring or loss to follow-up must not selectively hide transitions after conditioning on the information used to model observation. The risk set must correctly represent who is in state $r$ and observable at each time. If the model is semi-Markov, the time origin and sojourn-time information must also match the model’s specification.

Why the Baseline Estimate Has This Form

A short likelihood argument explains the weighted risk-set denominator. Consider one event time $u$, and let the unknown baseline cumulative-hazard increment be $a$. Under the fixed-effects model, subject $i$’s event intensity at that time is proportional to $a\exp(\beta_{rs}^{\mathsf T}x_i)$. In the usual multiplicative-intensity likelihood, the terms involving $a$ have log-likelihood

\ell(a)=d_{rs}(u)\log(a) -a\sum_iY_{ir}(u)\exp(\beta_{rs}^{\mathsf T}x_i) +\text{terms not involving }a.

Differentiating and setting the score to zero gives

\frac{\partial \ell(a)}{\partial a} =\frac{d_{rs}(u)}{a} -\sum_iY_{ir}(u)\exp(\beta_{rs}^{\mathsf T}x_i)=0, \qquad \widehat a= \frac{d_{rs}(u)} {\sum_iY_{ir}(u)\exp(\beta_{rs}^{\mathsf T}x_i)}.

This is the Breslow increment. With no transition at $u$, the numerator is zero, so that time contributes no increment. With tied transitions, the same expression is the Breslow treatment of the tied event set. The argument shows why updating the baseline can account for a changed overall transition level without changing the model’s relative covariate effects.

Worked Example: A Two-Time Breslow Update

Consider a hypothetical transition from state A to state B. The original model has one binary covariate $x\in\{0,1\}$ and fixed coefficient $\beta=\log(2)$, so the relative hazards are $1$ for $x=0$ and $2$ for $x=1$. At time 1, subjects A and B with $x=0$, and subject C with $x=1$, are all at risk; C makes the transition. At time 2, A and B remain at risk and A makes the transition. Assume follow-up and censoring are correctly recorded, and there are no other transitions out of A during these event times.

At time 1 the weighted risk-set total is $1+1+2=4$, and one transition occurs. Thus

\Delta\widehat{\Lambda}_{0,AB}(1)=\frac{1}{4}=0.25.

At time 2, only A and B are at risk, both with $x=0$. The weighted total is $1+1=2$, and one transition occurs:

\Delta\widehat{\Lambda}_{0,AB}(2)=\frac{1}{2}=0.50.

The updated cumulative baseline hazard through time 2 is $0.25+0.50=0.75$. For $x=1$, the cumulative transition hazard is $2(0.75)=1.50$. If A-to-B is the subject’s only possible exit from A, the probability of remaining in A through time 2 is $\exp(-1.50)=0.2231$, and the transition probability is

P\{X(2)=B\mid X(0)=A,x=1\} =1-\exp(-1.50)=0.7769.

This last conversion relies on the stated single-exit assumption. If other transitions out of A are possible, this expression is not the A-to-B probability: the other exit hazards also affect the probability of still being in A and must be included in the transition-probability calculation.

Baseline Shape and a Constrained Scale Update

Re-estimating the baseline hazard over time is flexible, but it requires enough validation events across follow-up to estimate its shape. A more restrictive alternative is to keep the original baseline shape and multiply it by a constant $c$:

\lambda^{\mathrm{new}}_{0,rs}(t)=c\,\lambda^{\mathrm{old}}_{0,rs}(t).

This preserves the timing pattern of the original baseline and changes only its overall level. It can be useful when validation data are limited, but the constraint is strong: it assumes the old and new baseline hazards differ by the same factor throughout follow-up. A time-varying difference cannot be repaired by one constant. The scale $c$ must be estimated from validation data, rather than chosen merely because it makes one summary risk look closer.

Worked Example: Updating a Constant Baseline Rate

In a hypothetical one-year validation interval, assume the A-to-B transition has a constant baseline intensity and is the only exit from A. The original coefficient is $\beta=\log(2)$. Validation data contain 6 person-years at $x=0$ and 2 person-years at $x=1$, with 3 observed A-to-B transitions. Assume the piecewise-exponential model is appropriate, follow-up is observed correctly, and censoring is independent of the transition process conditional on the modeled covariates.

The covariate-weighted person-time is $6(1)+2(2)=10$ weighted person-years. With the coefficient fixed, the log-likelihood terms for the validation baseline rate $\lambda_0$ are $3\log(\lambda_0)-10\lambda_0$, apart from constants. Its derivative is $3/\lambda_0-10$; setting this to zero gives

\widehat{\lambda}_{0,AB}^{\mathrm{new}} =\frac{3}{10}=0.3000\text{ per year}.

Suppose the original baseline cumulative hazard over the same one-year interval was $0.20$. For a subject with $x=1$, the original model’s transition probability under the single-exit assumption is $1-\exp\{-0.20(2)\}=0.3297$. The recalibrated probability is

1-\exp\{-0.3000(2)(1\text{ year})\} =1-\exp(-0.6000)=0.4512.

The increase reflects the higher validation baseline rate. The relative hazard for $x=1$ versus $x=0$ remains 2; no coefficient was re-estimated. The calculation is for this one-year interval and assumes no competing exit from A.

Using Updated Hazards in a Multi-State Prediction

Transition-specific hazards are components of a multi-state prediction, not probabilities in isolation. After recalibrating a transition’s baseline, use the updated hazard together with the model’s other relevant transition hazards to obtain state-occupation or transition probabilities. A baseline-hazard increment is not itself a probability. Even when all effects are fixed, changing a baseline can change the probabilities of reaching later states because it changes how quickly individuals enter or leave states.

The following example has a single route, A to B to C, with constant transition intensities. This deliberately avoids competing exits so that each probability calculation can be checked directly. It illustrates that updating baselines for two transitions can alter a multi-step prediction while leaving both relative effects unchanged.

Worked Example: Recalibrating Two Successive Transitions

In a hypothetical validation cohort, the weighted exposure for A-to-B is 20 weighted person-years, with 6 observed transitions. For B-to-C it is 8 weighted person-years, with 4 observed transitions. Suppose both original coefficients equal $\log(2)$, so $x=1$ doubles each transition intensity. Assume constant intensities within the prediction interval, no other allowed transitions, and adequately observed state-specific exposure.

The validation baseline intensity estimates are $6/20=0.3000$ per year for A-to-B and $4/8=0.5000$ per year for B-to-C. For an individual with $x=1$, the corresponding intensities are $a=0.6000$ and $b=1.000$ per year. Starting in A, the probability of being in B at time $t$ is the probability of entering B at time $u$ and then remaining there until $t$:

P\{X(t)=B\mid X(0)=A,x=1\} =\int_0^t a e^{-au}e^{-b(t-u)}\,du =\frac{a(e^{-at}-e^{-bt})}{b-a},\quad a\ne b.

At $t=1$ year, substitution gives

\frac{0.6000(e^{-0.6000}-e^{-1.000})}{1.000-0.6000} = 1.5(0.5488-0.3679) =0.2714.

For comparison, if the original baseline intensities were 0.2000 and 0.4000 per year, the same individual’s intensities would be 0.4000 and 0.8000 per year, giving $e^{-0.4}-e^{-0.8}=0.2210$ for occupancy of B at one year. In both calculations, the covariate doubles each transition intensity. The difference comes from the updated baselines, not from a change in relative effects.

Diagnostics, Alternatives, and Limitations

A baseline update is only credible if the fixed relative effects remain plausible in the validation population. Examine transition-specific event counts and risk sets over time, and check whether estimates are driven by a few late events or sparse risk sets. Compare observed transition experience with predictions across follow-up and across covariate values. Systematic discrepancies that vary with covariate level may indicate that keeping the original effects fixed is inadequate; discrepancies that vary strongly over time may indicate that a constant-scale update is inadequate. Such patterns are reasons to investigate the model, not proof of a particular alternative.

Use the update only when the data support it. A very small number of transitions can produce unstable baseline estimates, especially at late times. Informative censoring, incorrect state histories, inconsistent time origins, or changes in covariate measurement can also invalidate the risk sets or the transport assumption. Uncertainty in the recalibrated predictions should account for estimation from the validation sample; treating the estimated baseline as known understates uncertainty.

Approach	What changes	Key limitation
Full transition-baseline re-estimation	The baseline hazard over time is estimated from validation data; the original coefficients are fixed.	Needs adequate transition information across follow-up.
Constant baseline-scale update	One multiplier changes the original baseline hazard’s overall level.	Cannot capture a change in baseline shape over time.
Refit covariate effects	Coefficients as well as the baseline can change.	Requires enough validation information and no longer retains the original relative effects.
Calibration-slope update	Relative effects are rescaled rather than held fixed.	It addresses a different updating question from baseline-only recalibration.

Baseline-only updating is distinct from estimating a calibration slope: here the original coefficients are held fixed by design. It is also distinct from separately forcing predicted probabilities for competing destinations to match observed proportions. Cause-specific hazards and transition probabilities are different quantities; probabilities must be computed from the joint transition model, not obtained by treating each transition hazard as an independent probability.

Common Mistakes

Using the wrong risk set. A subject contributes to the $r\to s$ risk set only while in state $r$ and under observation. The correction is to construct risk sets from the observed state history and the model’s time origin.
Changing the coefficient while claiming to retain it. Baseline recalibration uses the original $\beta_{rs}$ in the weighted risk set and prediction. Re-estimating it is a different update.
Calling a cumulative hazard a probability. A cumulative hazard can exceed 1. Convert it to a probability only under the relevant exit structure; with competing exits, calculate probabilities using all relevant transition hazards.
Assuming one multiplier fixes every discrepancy. A constant baseline scale cannot repair a changed time pattern or an incorrect covariate effect.

Key takeaway: Recalibrating a transition baseline updates the validation population’s underlying transition intensity while preserving the original relative covariate effects. The update is meaningful only when risk sets, time origins, observation, and transport assumptions are defensible; probabilities must then be derived from the full transition model.

Check Your Understanding

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

At a transition time, three subjects are at risk: two have $x=0$ and one has $x=1$. The fixed coefficient is $\log(2)$, and one transition occurs. What is the Breslow baseline-hazard increment?
Why is an estimated baseline cumulative hazard of 0.75 not automatically a transition probability of 75%?
In a one-year interval, a fixed-effects model has 5 observed transitions and 20 weighted person-years. Under a constant baseline intensity, what is the estimated baseline intensity, including units?
State one assumption needed to interpret a baseline-only update as transporting the original model’s relative effects.
If discrepancies between predictions and validation outcomes grow steadily over time, why might a single baseline multiplier be inadequate?

Solutions. (1) The weighted risk-set total is $2(1)+1(2)=4$, so the increment is $1/4=0.2500$. (2) A cumulative hazard is not a probability. With one possible exit, the transition probability is $1-\exp(-0.75)=0.5276$; with competing exits, all relevant hazards are needed. (3) The estimate is $5/20=0.2500$ per year. (4) For example, the original transition-specific covariate effect must remain adequately valid in the validation population, conditional on the model’s covariates and state history. (5) A single multiplier changes the baseline level uniformly; it cannot represent a difference in baseline shape over time.