Conditioning the Future on the Present State

Why the Present State Can Summarize the Past

A multi-state process records how an individual moves among states over time. Two individuals may be in the same state at a given time but have arrived there by different routes. One may have entered the illness state recently; another may have experienced an earlier episode before recovering and relapsing. The Markov property makes a precise claim about what those different histories imply: when the comparison is made at the same time and in the same state, the conditional law of the future is the same.

This is a statement about a model, not a general fact about clinical data. It says that, for predicting future states within the model, the information in the recorded present state is sufficient to summarize the relevant information in the past. It does not say the past histories are identical, or that the future is known with certainty.

Definition: Let $X(t)$ denote the state occupied at time $t$, taking values in a state space $\mathcal{S}$. Let $\mathcal{F}_t$ represent the information in the process history up to and including time $t$. The process is Markov if, for each $u>0$ and each set of states $B$, the conditional probability of being in $B$ at time $t+u$, given the history through time $t$, depends on that history only through $X(t)$: $$ P\{X(t+u)\in B\mid \mathcal{F}_t\} = P\{X(t+u)\in B\mid X(t)\}. $$ The property is required at the times under consideration, and the conditioning on the present state is understood wherever that state occurs with positive probability.

For a future event involving several times, the corresponding Markov statement is that its conditional probability given the history through $t$ depends on that history only through $X(t)$. Equivalently, the conditional distribution of the future process, starting from $t$, is determined by the state at $t$. The single-time equation above is a useful way to state the property; the multi-time implication follows by applying the Markov property successively.

From the Markov Property to Equal Conditional Futures

To make the comparison precise, consider two possible histories, represented by events $H_1$ and $H_2$, that are known by time $t$. Suppose both histories end in state $i$: $H_k\subseteq\{X(t)=i\}$ for $k=1,2$. Suppose also that each history has positive probability. For any future event $F$, the Markov property says that its conditional probability given the complete history through $t$ is a function of the current state. On histories ending in $i$, that function has the same value.

\begin{aligned} P(F\mid H_k) &= E\!\left[P(F\mid\mathcal{F}_t)\mid H_k\right]\\ &= E\!\left[P(F\mid X(t))\mid H_k\right]\\ &= P(F\mid X(t)=i), \qquad k=1,2. \end{aligned}

The first equality is the law of iterated expectations: conditioning first on all information through $t$, then on the less detailed event $H_k$, gives the conditional probability given $H_k$. The second uses the Markov property. The last holds because $X(t)=i$ throughout $H_k$, so the conditional probability given the present state is the same value on that history. Thus,

P(F\mid H_1)=P(F\mid X(t)=i)=P(F\mid H_2).

The future event $F$ can concern a state at one future time, a sequence of future states, or a trajectory feature such as visiting a specified state before another. For example, taking $F=\{X(t+u)=j\}$ gives equal probabilities of being in state $j$ after $u$ time units. The equality holds for each future event of interest, so it expresses equality of conditional future distributions—not just equality of one particular transition probability.

Key implication: At a fixed time $t$, any two histories that end in the same state have the same conditional future law under a Markov model. The histories may differ in their earlier states, event counts, or times of entry into the current state. Those differences do not affect the conditional future law unless they are represented in the state or otherwise included in the model.

Worked Examples

Worked Example: Different Routes to the Illness State

Consider a hypothetical illness-death process with states $H$ (healthy), $I$ (ill), and $D$ (dead). At a fixed time $t$, one individual’s history is $H\to I$, while another’s is $H\to I\to H\to I$. Both are currently in $I$. Assume the state definition is adequate and the process satisfies the Markov property. For the next interval, suppose the conditional transition probabilities from $I$ are 0.70 to remain in $I$, 0.30 to enter $D$, and 0 to return directly to $H$.

For the future event $F=\{X(t+1)=D\}$, the probability given either history is the probability of moving from the current state $I$ to $D$:

P(F\mid H\to I)=P(F\mid H\to I\to H\to I)=0.30.

The histories are different, but they end in the same state at the same time, so the Markov property gives the same next-interval probability. This conclusion depends on the stated model assumption: if recurrence history changes the death risk and that information is not represented in $X(t)$, the Markov property for this state definition would not hold.

Worked Example: A Two-Interval Future Distribution

Use the same hypothetical states $H,I,D$, with $D$ absorbing. For individuals in state $I$ at time $t$, suppose the transition probabilities over the first interval are 0.70 to $I$ and 0.30 to $D$. For those still in $I$ at $t+1$, suppose the probabilities over the next interval are 0.60 to $I$ and 0.40 to $D$. Assume the process is Markov at both times, and compare two positive-probability histories that are both in $I$ at $t$.

To be in $I$ at $t+2$, an individual must remain in $I$ over both intervals. The probability is

P\{X(t+2)=I\mid X(t)=I\} =0.70(0.60)=0.42.

Death by $t+2$ can occur in the first interval, with probability 0.30, or in the second interval after remaining in $I$, with probability $0.70(0.40)=0.28$. Therefore,

P\{X(t+2)=D\mid X(t)=I\} =0.30+0.70(0.40)=0.58.

These probabilities sum to 1, since $0.42+0.58=1$. Both histories ending in $I$ at time $t$ have the same two-interval distribution: 0.42 in $I$ and 0.58 in $D$. The calculations use transition probabilities for each specific interval; they do not require the probabilities to be constant over time.

Worked Example: A Future Event Along Several Routes

Suppose a hypothetical process has state $R$ (at risk) and two absorbing event states $A$ and $B$. At time $t$, two different histories end in $R$. For the first interval, assume that an individual in $R$ has probability 0.20 of moving to $A$, 0.10 of moving to $B$, and 0.70 of remaining in $R$. For the second interval, an individual still in $R$ has probabilities 0.25 to $A$, 0.15 to $B$, and 0.60 to remain in $R$. Assume the Markov property at the interval boundaries.

The probability of being in $A$ by the end of two intervals is the probability of moving to $A$ in the first interval plus the probability of remaining in $R$ and then moving to $A$:

P\{X(t+2)=A\mid X(t)=R\} =0.20+0.70(0.25)=0.375.

Similarly, the probability of being in $B$ is

P\{X(t+2)=B\mid X(t)=R\} =0.10+0.70(0.15)=0.205.

The probability of remaining in $R$ is $0.70(0.60)=0.42$. These probabilities sum to $0.375+0.205+0.42=1$. Since both histories start this calculation from the same state $R$ at the same time $t$, both have the same conditional probabilities for these future outcomes under the model.

Assumptions, Scope, and Comparisons

The implication requires a correctly specified state process and the Markov property for that process. In particular, the state must carry the information needed to make the conditional future law independent of earlier history. The histories being compared must end in the same state at the same time, and the conditioning events must be meaningful (in the examples above, they have positive probability).

The same-time condition matters. In a time-inhomogeneous model, transition probabilities may depend on the calendar time. Even for the same state $i$, the conditional future law starting at time $t_1$ need not equal the law starting at a different time $t_2$. The Markov implication compares histories at a common time; it does not assert that future laws are identical across different starting times.

The claim is also about future distributions, not about past histories. It does not imply that the two histories are equally likely, that they contain the same information about earlier outcomes, or that the individuals have the same future realized path. Nor does it imply that a transition probability is constant across times or across individuals. It says only that, within the specified model, conditioning on additional history does not change the future law once the present state and comparison time are fixed.

A model that retains more history can be considered when the current state does not adequately summarize the past. For example, one might refine a state to distinguish clinically relevant subgroups or record duration in the current state. Another approach is to use a model that explicitly allows future risk to depend on history. These alternatives change the conditioning information; they are not consequences of the basic Markov implication.

Common Pitfalls and Limitations

Comparing different times. Equal current states alone do not imply equal future laws when the comparison times differ. Make the time origin of each conditional distribution explicit.
Assuming the implication proves the model. Equal predictions in a small sample do not establish that the process is Markov. The equality follows if the Markov assumption holds; it is not evidence by itself that the assumption is correct.
Ignoring omitted history. If duration in the current state or prior episodes affect the future and are not represented in $X(t)$, histories ending in the same recorded state may have different conditional futures. The chosen state may then be too coarse for a Markov model.
Conditioning on an impossible history. If a history has probability zero, its conditional probability is not defined by the elementary ratio $P(F\cap H)/P(H)$. Conditional distributions given continuously valued histories are defined through regular conditional distributions and are identified only almost surely; arbitrary zero-probability histories cannot automatically be compared point by point.
Confusing shared distributions with shared outcomes. Two individuals can have identical conditional future distributions and still experience different future state paths.

The practical importance of the result is that it gives the present state a clear predictive role. If the state definition and Markov assumption are appropriate, future transition calculations can start from the current state rather than separately tracking every possible route by which it was reached. If they are not appropriate, that simplification can conceal real history dependence.

Check Your Understanding

Answer each question using the same-time comparison and the Markov assumption carefully.

Two histories end in state $i$ at time $t$. What does the Markov property imply about their conditional probabilities of a specified future event $F$?
In the illness-death example, calculate the probability of remaining ill at $t+2$, given illness at $t$, using the two stated interval probabilities.
In the $R,A,B$ example, calculate the probability of being in $B$ by the end of two intervals.
Does the implication guarantee equal future laws for two histories ending in the same state at different times? Why or why not?
If a history has probability zero, can its conditional future probability be computed by dividing $P(F\cap H)$ by $P(H)$?

Solutions. (1) For histories $H_1,H_2$ of positive probability ending in $i$ at $t$, $P(F\mid H_1)=P(F\mid X(t)=i)=P(F\mid H_2)$. (2) Remaining ill requires staying ill in both intervals, so $0.70(0.60)=0.42$. (3) Moving to $B$ in the first interval has probability 0.10; remaining in $R$ and then moving to $B$ has probability $0.70(0.15)=0.105$. Thus the total is $0.10+0.105=0.205$. (4) No. The result compares histories at the same time; a time-inhomogeneous model can have different future laws at different starting times. (5) No. The ratio is undefined when $P(H)=0$; conditional distributions for such cases require more care and are generally specified only almost surely.