Omics Predictors In Multi-State Survival Models

Why Model More Than One Transition?

A clinical course can involve several possible transitions: a patient may move from an initial state to a complication, recover, or die. A single “event hazard” can hide important differences among these paths. An omics measurement may be associated with entering a complication state but have a different association with death or recovery. A multi-state model represents these pathways separately, then combines their fitted transition rates to predict the probability of occupying each state at a future time.

The central distinction is between a transition hazard, which describes an instantaneous rate of leaving one state for another, and a state-occupation probability, which describes the probability of being in a state at a specified time. The latter depends on all possible routes into and out of that state, not only on one transition coefficient.

Definition: Let $S(t)$ denote the state occupied at time $t$. For two distinct states $r$ and $s$, the transition intensity $q_{rs}(t\mid\mathcal H_t)$ is the conditional instantaneous rate of moving from $r$ to $s$, given the observed history $\mathcal H_t$ and that the individual is in state $r$ just before $t$. Formally, $$ q_{rs}(t\mid\mathcal H_t) = \lim_{\Delta t\downarrow 0} \frac{P\{S(t+\Delta t)=s\mid S(t)=r,\mathcal H_t\}}{\Delta t}. $$ A state-occupation probability is $P\{S(t)=s\}$, or its conditional version given a starting state and covariate history.

This tutorial uses states indexed by $r,s$, with a finite set of permitted transitions. A state can be absorbing, meaning that no transitions leave it; death is a common example. Omics variables may be measured at baseline or updated over follow-up. A baseline vector $Z$ is fixed over time. A time-updated vector $Z(t)$ must be available just before the transition time being modeled.

Transition-Specific Omics Effects

A common model for each permitted transition is a transition-stratified proportional-hazards model:

q_{rs}(t\mid Z(t),W) = q_{0,rs}(t) \exp\!\left\{\beta_{rs}^{\mathsf T}Z(t)+\gamma_{rs}^{\mathsf T}W\right\}.

Here $q_{0,rs}(t)$ is the baseline intensity for transition $r\to s$, $Z(t)$ is the omics predictor vector, and $W$ contains other covariates. The coefficients $\beta_{rs}$ are allowed to differ by transition. For a one-unit increase in omics feature $k$, holding the other modeled predictors fixed, the transition-specific hazard ratio is $\exp(\beta_{rs,k})$. If the feature is standardized, the unit is one standard deviation; otherwise, it is one unit on the stated assay scale.

The model assumes proportional hazards for each transition: the covariate multiplies that transition’s baseline intensity by a constant factor over time. It does not require the baseline intensities for different transitions to have the same shape. The model also requires correctly specified allowed transitions and a defined risk set: an individual contributes to the $r\to s$ risk set while in state $r$, and leaves that risk set upon leaving $r$, whether by transition to $s$ or to another state.

For a time-updated predictor, using $Z(t)$ requires that its value be known before the event at $t$. A value measured after a transition cannot be used to predict that transition. If assay times are irregular, the model must specify how the value is defined between measurements, such as carrying the latest available value forward. This rule does not eliminate measurement error or informative visit timing; those issues can require methods such as joint models.

Worked Example: Interpreting Two Transition Effects

Suppose two standardized omics features, $Z_1$ and $Z_2$, enter a model for transition $0\to1$, with $\beta_{01,1}=0.40$. For transition $0\to2$, suppose $\beta_{02,1}=-0.20$. Interpret the effect of a one-standard-deviation increase in $Z_1$ for each transition.

For transition $0\to1$, the conditional hazard ratio is $$ \exp(\beta_{01,1})=\exp(0.40)=1.492. $$ Thus, among individuals currently in state 0 with the same values of the other modeled predictors, a one-standard-deviation higher $Z_1$ corresponds to a 1.492-fold instantaneous rate of transition to state 1, under the proportional-hazards model.

For transition $0\to2$, the corresponding calculation is $$ \exp(\beta_{02,1})=\exp(-0.20)=0.8187. $$ A one-standard-deviation increase corresponds to a transition intensity about 18.1% lower, since $1-0.8187=0.1813$, for transition $0\to2$. These are conditional associations, not evidence that changing the feature would cause either transition rate to change.

The effects differ by destination state. Reporting one coefficient for “the event” would discard this distinction. The interpretation also relies on the proportional-hazards assumption and on the stated one-standard-deviation scale.

From Intensities to State Occupation

For a Markov multi-state model, the transition rates at time $t$, conditional on the covariates being used, depend on the current state and time, not on earlier state history beyond the current state. Form a matrix $Q(t)$ whose off-diagonal entry in row $r$, column $s$ is $q_{rs}(t)$, and whose diagonal entry is the negative sum of rates out of that state:

q_{rr}(t)=-\sum_{s\ne r}q_{rs}(t).

The state-occupation transition matrix $P(u,t)$ has entry $P_{rs}(u,t)=P\{S(t)=s\mid S(u)=r\}$. It satisfies the forward equation $dP(u,t)/dt=P(u,t)Q(t)$, with $P(u,u)=I$, the identity matrix. Each row of $P$ sums to one. Thus, if the initial state is known, the corresponding row gives predicted probabilities for all states at the prediction time.

Justification: Over a short interval $\Delta t$, the probability of one permitted transition $r\to s$ is $q_{rs}(t)\Delta t+o(\Delta t)$; the probability of remaining in $r$ is $1-\sum_{s\ne r}q_{rs}(t)\Delta t+o(\Delta t)$. These probabilities give the short-interval transition matrix $I+Q(t)\Delta t+o(\Delta t)$. Composing successive short-interval matrices yields the forward equation. This construction requires a well-defined intensity process and the Markov property for the state representation and covariates used.

When rates are constant over an interval, the transition matrix over that interval is $\exp\{Q\Delta t\}$. With time-varying rates, the solution is a product integral, or numerically a product of short-interval transition matrices. In practice, estimated transition-specific baseline hazards and coefficients are combined at the specified covariate values to construct the prediction. Uncertainty in predicted probabilities should also reflect estimation uncertainty, for example through a bootstrap or model-based simulation.

A Markov assumption can be inappropriate if risk depends on time since entering the current state or on earlier states in a way not summarized by current covariates. One remedy is a semi-Markov model with time since state entry in the intensity; another is to expand the state representation to include relevant history. The chosen model must match the prediction question.

Worked Example: Predicting Occupation in a Three-State Process

Consider a hypothetical illness-death process with state 0 (alive without complication), state 1 (alive with complication), and state 2 (dead). State 2 is absorbing. For one specified omics profile, suppose fitted constant transition intensities are $q_{01}=0.20$ per year, $q_{02}=0.10$ per year, and $q_{12}=0.30$ per year. There are no other permitted transitions. Predict state occupation after 2 years for an individual starting in state 0.

These values are assumed to apply throughout the two-year interval for this profile. The total rate out of state 0 is $0.20+0.10=0.30$ per year, and the rate out of state 1 is $0.30$ per year. Therefore, $$ P\{S(t)=0\}=e^{-0.30t}. $$ To occupy state 1 at time $t$, the individual must first enter state 1 from state 0 and then remain there. The inflow-outflow equation is $P_1'(t)=0.20P_0(t)-0.30P_1(t)$, with $P_1(0)=0$. Since $P_0(t)=e^{-0.30t}$ and the two exponential rates are equal, integration gives $P_1(t)=0.20t e^{-0.30t}$.

At $t=2$ years, $$ P_0(2)=e^{-0.60}=0.5488,\qquad P_1(2)=0.20(2)e^{-0.60}=0.2195. $$ Because the three states are exhaustive, the remaining probability is death: $$ P_2(2)=1-P_0(2)-P_1(2) =1-0.5488-0.2195=0.2317. $$ The predicted probabilities are therefore 0.5488 in state 0, 0.2195 in state 1, and 0.2317 in state 2 after 2 years. They sum to 1.0000 up to rounding. These predictions are conditional on the specified profile, constant rates, model structure, and starting state; they are not three separate transition hazard ratios.

Estimation, Diagnostics, and Alternatives

With transition-specific Cox models, coefficients can be estimated using the risk sets for each transition, with a separate baseline hazard for each transition. The partial likelihood contribution for a transition compares the covariates of the individual making that transition with those of individuals still in the origin state at that time. A baseline cumulative intensity can then be estimated for each transition and combined with the fitted coefficients to obtain transition probabilities.

Diagnostics should be transition-specific. Check whether there are enough observed transitions for stable estimation, whether the proportional-hazards assumption is plausible for each transition, and whether the fitted intensities yield sensible predictions. A transition with few events can produce imprecise coefficients even when the overall cohort is large. Compare predicted and observed state-occupation proportions at relevant times, accounting for censoring, and inspect whether predictions remain between zero and one and sum to one. Systematic disagreement suggests model misspecification, poor calibration, or an unsuitable covariate-history assumption.

Independent censoring conditional on the modeled information is needed for standard likelihood-based estimation and prediction. If follow-up ends for reasons related to unobserved transition risk, ordinary censoring treatment may be inadequate. Omics features measured with substantial error can also distort estimated effects; measurement-error methods may be needed. When many features are considered, fitting and selecting predictors must avoid using outcome information in a way that exaggerates apparent performance.

Approach	Useful when	Main limitation
Transition-specific Cox models	Different transitions plausibly have different covariate effects and baseline rates.	Requires adequate events per transition; proportional hazards must be checked.
Shared-effect multi-state model	A common omics effect across transitions is scientifically defensible and parsimonious.	Can obscure genuine differences among pathways if effects are not shared.
Parametric multi-state model	A defensible functional form for transition rates is available and smooth extrapolation is needed.	Predictions can be sensitive to baseline distributional assumptions.
Competing-risks model	The question concerns alternative first events from a single initial state.	Does not by itself represent later transitions among nonabsorbing states.

A multi-state model matters because clinical probabilities combine multiple routes through the process. A predictor can increase entry into a complication state yet also be associated with subsequent mortality; the final probability of being alive with that complication reflects both transitions. Reporting only a transition coefficient therefore does not answer a state-occupation question.

Common pitfall: “The omics feature has a hazard ratio of 1.5, so the probability of being in the complication state is 50% higher.” This is incorrect: the hazard ratio is a relative instantaneous rate for one specified transition among individuals still in its origin state. The state-occupation probability depends on all transitions, their timing, and the prediction horizon. Calculate it from the full fitted transition model.

Worked Example: A Time-Updated Omics Measurement

Suppose the hypothetical model for transition $0\to1$ is $q_{01}(t)=0.10\exp\{0.50Z(t)\}$ per year, where $Z(t)$ is an assay score in standardized units. For a patient still in state 0, the most recent pre-event score changes from 0.2 to 1.0. Compare the modeled instantaneous intensities at those values.

At $Z=0.2$, $$ q_{01}=0.10\exp(0.50\times0.2) =0.10\exp(0.10)=0.1105\text{ per year}. $$ At $Z=1.0$, $$ q_{01}=0.10\exp(0.50\times1.0) =0.10\exp(0.50)=0.1649\text{ per year}. $$ The ratio is $0.1649/0.1105\approx1.492$, which also equals $\exp\{0.50(1.0-0.2)\}=\exp(0.40)$. Thus the higher score is associated with a 1.492-fold instantaneous rate for transition $0\to1$, conditional on remaining in state 0 and on other modeled covariates. It does not imply a 1.492-fold probability of occupying state 1 at a future time.

This calculation is valid only if both scores are measured before the relevant transition, the model’s proportional effect applies over the interval, and the score’s time-updating rule is specified. If the intensity also depends on time since entering state 0, an ordinary time-inhomogeneous Markov model using only current state and calendar time may not be adequate.

Check Your Understanding

Answer each question, then compare with the solutions below.

For a standardized omics feature with $\beta_{12}=0.30$, calculate and interpret the hazard ratio for transition $1\to2$.
Why can a transition-specific hazard ratio not be read directly as a state-occupation probability ratio?
In the three-state example, what is the total intensity of leaving state 0, and what does it represent?
Name one reason the Markov assumption could fail and one modeling response.
A score is assayed only after a patient enters state 1. Can that value be used as a time-updated predictor of the earlier $0\to1$ transition? Explain.

Solutions. (1) $\exp(0.30)=1.350$. A one-standard-deviation increase is associated with a 1.350-fold instantaneous rate for transition $1\to2$, conditional on the model covariates and being in state 1. (2) The occupation probability depends on all transitions into and out of the state and on the time horizon, whereas a hazard ratio compares one transition’s instantaneous rates. (3) The total is $0.20+0.10=0.30$ per year, the instantaneous rate of leaving state 0 by either permitted route. (4) Dependence on time since state entry can violate the Markov representation; a semi-Markov model can include that duration. (5) No. The value occurs after the transition, so it is unavailable at the prediction time and could introduce information from the future.

Omics Predictors In Multi-State Survival Models

What You'll Learn