Why Cognitive Impairment Needs an Explicit Model Choice
A long-term care contract may make benefits available when a person meets an activities-of-daily-living criterion, a cognitive-impairment criterion, or either one. The model must represent those conditions in a way that matches the benefits being valued. If the model records only whether a person is eligible, it may be sufficient for benefit calculations that do not depend on the reason for eligibility. If it also needs to distinguish cognitive impairment from functional impairment, the state definition must preserve that information.
The model is not a clinical diagnosis. It is an actuarial classification of health conditions relevant to the hypothetical contract. The contract’s definition must determine what counts as qualifying cognitive impairment; a diagnosis label alone does not establish eligibility unless the assumed contract says that it does.
Use continuous-time transition intensities, measured per year, for the examples below. For a continuous-time Markov model, the intensity \(q_{jk}\) is the instantaneous rate of transition from state \(j\) to state \(k\), conditional on currently being in state \(j\). The diagonal entry of the intensity matrix is \(q_{jj}=-\sum_{k\ne j}q_{jk}\), so each row sums to zero. Unless stated otherwise, the examples assume constant intensities over the period considered.
Two Ways to Represent Eligibility
A folded representation combines the reasons for eligibility. For example, the model can have a healthy, not-eligible state \(H\), an eligible state \(E\), and death \(D\). A transition from \(H\) to \(E\) can occur because of either qualifying ADL impairment or qualifying cognitive impairment. The model then retains eligibility status, but not the trigger that caused it.
A separate-state representation distinguishes the relevant conditions. Let \(A\) mean ADL-eligible but not cognitively impaired, \(C\) mean cognitively impaired but not ADL-eligible, and \(AC\) mean that both criteria are met. With \(H\) and \(D\), this model can record both eligibility and the reason or reasons for it. It can also allow transitions such as cognitive impairment developing while a person is already ADL-eligible.
| Model choice | Information retained | Key assumption or consequence |
|---|---|---|
| Folded eligible state \(E\) | Whether at least one eligibility trigger is met | Trigger-specific differences can be discarded only if they are not needed for benefits or future transition rates |
| Separate states \(A,C,AC\) | Which trigger or combination of triggers is present | Each allowed transition between these states needs an intensity or an explicit structural exclusion |
| Collapsed eligible group | Only the combined eligible/not-eligible classification | The aggregate is Markov only if the relevant transition rates agree across states being combined |
The choice is not simply a matter of detail. Folding the states can make a model smaller, but it is inappropriate if cognitive impairment changes the benefit amount, mortality assumption, recovery pattern, or other future transition rates. Conversely, retaining separate trigger states adds parameters and requires assumptions about transitions that a folded model may not need.
Worked Example: Folding Two Eligibility Triggers
In a hypothetical LTC model, a person starts in \(H\). The constant intensities from \(H\) are \(0.06\) per year to ADL eligibility, \(0.04\) per year to cognitive eligibility, and \(0.02\) per year to death. Either eligibility trigger moves the person into a single state \(E\). Calculate the probability that the first transition out of \(H\) is an eligibility transition within one year, and split it by trigger.
Method and assumptions. The three exits from \(H\) are competing transitions. Their total intensity is \(0.06+0.04+0.02=0.12\), so the probability of remaining in \(H\) for time \(t\) is \(e^{-0.12t}\). The model assumes that the trigger causing entry is not needed after entry to \(E\).
Over a short interval of length \(dt\), the probability of an ADL transition, given that the person is still in \(H\), is approximately \(0.06\,dt\). Thus the probability of an ADL transition as the first exit by time 1 is the integral of the probability of still being in \(H\) times the ADL intensity:
For the cognitive trigger, the same calculation gives
Adding the two mutually exclusive first-transition probabilities gives \(0.056540+0.037693=0.094233\). Equivalently, the combined eligible intensity is \(0.06+0.04=0.10\), and
As a check, the probability of any first exit from \(H\) by one year is \(1-e^{-0.12}=0.113080\). The death-first probability is \(\frac{0.02}{0.12}(1-e^{-0.12})=0.018847\), and \(0.094233+0.018847=0.113080\). The method avoids treating death as if it could not occur before eligibility.
Separate States and Their Transition Assumptions
A separate-state model makes it possible to distinguish the triggers and their overlap. For example, from \(A\), cognitive impairment can develop and move the person to \(AC\); from \(C\), ADL eligibility can develop and also move the person to \(AC\). Recovery can be represented as a transition from \(AC\) to \(A\) if cognitive impairment ceases to meet the criterion, or from \(AC\) to \(C\) if the ADL criterion ceases to be met. Each such transition is a model assumption, not an automatic consequence of defining the states.
The following hypothetical intensity matrix gives one possible specification. The state order is \(H,A,C,AC,D\), and \(D\) is absorbing. All nonzero off-diagonal entries are annual intensities.
For instance, from \(A\), the model allows a transition back to \(H\) at intensity \(0.06\), a transition to \(AC\) at \(0.04\), and death at \(0.04\). The diagonal entry is \(-0.14\), the negative sum of those three intensities. The \(H\)-to-\(AC\) intensity is set to zero: this specification excludes a direct simultaneous onset of both criteria. A model could allow that transition instead, but it would require an assumed intensity.
Worked Example: Calculating First Eligibility in a Separate-State Model
Use the matrix above. A person starts in \(H\). Find the probability that the first transition out of \(H\) within one year is to either \(A\) or \(C\). Also calculate the contributions from each destination.
Method and assumptions. From \(H\), the outgoing intensities are \(0.07\) to \(A\), \(0.05\) to \(C\), and \(0.03\) to death. Their total is \(0.15\), so \(H\)-occupancy up to the first transition is \(e^{-0.15t}\). Because the person begins in \(H\), transitions from \(A,C,\) or \(AC\) do not affect this first-transition calculation.
The first transition to \(A\) by one year has probability
The first transition to \(C\) by one year has probability
Therefore, the probability of a first transition to an eligible state is \(0.065003+0.046431=0.111434\). Directly, this is \(\frac{0.12}{0.15}(1-e^{-0.15})=0.111434\). The probability of death as the first transition is \(\frac{0.03}{0.15}(1-e^{-0.15})=0.027858\). These add to \(1-e^{-0.15}=0.139292\), the probability of any first transition by one year. The answer is not the probability of occupying an eligible state at the end of the year, because subsequent recovery or death can change occupancy.
Why Competing Triggers Are Not Always Additive
When a folded model treats the trigger as a single first transition out of \(H\), the trigger-specific intensities add: an ADL exit and a cognitive exit cannot both be the first exit at the same instant. But probabilities of ever meeting each trigger over a period are different quantities. A person can meet one criterion and later meet the other. Adding those two marginal probabilities without accounting for overlap counts people who meet both twice.
Worked Example: Correcting for Overlap Between Triggers
Consider a hypothetical model with no death or recovery over one year. ADL impairment first arises according to a constant intensity of \(0.10\) per year, and cognitive impairment according to an independent constant intensity of \(0.08\) per year. The states are \(H,A,C,AC\): if the person has already met one criterion, the other can still arise. Calculate the probability that at least one trigger has occurred by one year.
Method and assumptions. The two onset times are independent exponential random variables with rates \(0.10\) and \(0.08\). The probability that the ADL trigger has occurred is \(1-e^{-0.10}=0.095163\); the probability that the cognitive trigger has occurred is \(1-e^{-0.08}=0.076884\). Independence means the probability both have occurred is the product:
The inclusion-exclusion rule gives the probability of at least one trigger:
A check comes from the probability that neither independent onset has occurred, \(e^{-0.10}e^{-0.08}=e^{-0.18}\). Thus the probability of at least one is \(1-e^{-0.18}=0.164730\). The method avoids double-counting people who meet both criteria.
When Can Eligible States Be Collapsed?
A common modeling question is whether \(A,C,\) and \(AC\) can be replaced by one eligible state \(E\). The key issue is whether the aggregated process still has the Markov property: given its current aggregated state, its future transition behavior must not depend on which underlying state it came from. In a continuous-time Markov model, a necessary aggregation condition is that, for any two states being grouped together, their total intensities into each other aggregate group agree. In particular, their total intensities from the eligible group to the ineligible group must match.
If the intensity of recovery from \(A\) differs from the intensity of recovery from \(C\), the chance of leaving eligibility over a short interval depends on which state within \(E\) the person occupies. An averaged recovery intensity may approximate a model for a specified mix of states, but that mix can change over time. A fixed average does not generally preserve the transition law.
Worked Example: Testing Whether an Eligible Aggregate Is Markov
Suppose \(A\) and \(C\) are both eligible, and the only transition from either state is recovery to \(H\). The recovery intensity is \(0.20\) per year from \(A\) and \(0.05\) per year from \(C\). At time zero, a hypothetical group is split equally between \(A\) and \(C\). Compare the probability of still being eligible after one year and two years with what a single constant recovery intensity fitted to the one-year result would predict.
Method and assumptions. Each state has an exponential time to recovery with its stated constant intensity. The probability of remaining eligible for one year, averaged over the initial 50–50 mixture, is
A single constant recovery intensity \(q\) fitted to this one-year probability must satisfy \(e^{-q}=0.884980\), so \(q=-\ln(0.884980)\), approximately \(0.1222\) per year. The resulting two-year prediction is \(e^{-2q}=(0.884980)^2=0.7832\).
The actual two-year probability under the separate states is
The fitted single rate matches the first-year mixture by construction, but its two-year prediction differs. This reflects the changing mix of \(A\) and \(C\) among people who remain eligible. The aggregate \(E\) therefore does not have one constant recovery intensity that reproduces these probabilities at both durations. The mistake to avoid is treating an initial weighted average as a transition intensity that remains valid at all future times.
Common Exam Traps and Model Checks
- Adding marginal trigger probabilities without subtracting overlap. In the independent-onset example, adding \(0.095163+0.076884=0.172047\) overstates the probability of at least one trigger. Subtracting the overlap, \(0.007317\), gives the correct \(0.164730\).
- Confusing a first-transition probability with an occupancy probability. A first transition from \(H\) to an eligible state can be followed by recovery or death before the valuation time. First-entry probability and end-time occupancy are not interchangeable.
- Combining states that have different future transition rates. If cognitive status changes mortality, recovery, or benefit eligibility, a folded state may discard information needed to specify the model correctly.
- Leaving transitions implicit. A separate-state design should state whether simultaneous onset is allowed, whether each trigger can recover, and whether a person can relapse. A zero intensity is a substantive structural assumption.
- Equating a clinical label with the contract trigger. The state must follow the hypothetical contract’s operational definition, including any assumed evidence or functional conditions, rather than an unspecified diagnosis alone.
This modeling choice matters because transition probabilities, expected benefit eligibility, and the interpretation of state-specific assumptions depend on the state definitions. A folded state is suitable when eligibility alone is sufficient and the grouped states have compatible transition behavior. Separate states are needed when the trigger itself affects benefits or future transitions. State definitions and intensities should be checked together: a detailed state graph is useful only when its transition assumptions are explicit and consistent.
Check Your Understanding
For each problem, use the stated hypothetical model assumptions and give probabilities to four decimal places.
- From \(H\), the intensities to \(A\), \(C\), and death are \(0.09\), \(0.06\), and \(0.05\) per year. Calculate the probability that the first transition is to either eligible state within two years. (Answer: \(0.2473\))
- From \(H\), the intensities to \(A\), \(C\), and death are \(0.04\), \(0.06\), and \(0.02\) per year. What is the probability that the first transition is to \(C\) within half a year? (Answer: \(0.0291\))
- Independent ADL and cognitive onset times have constant intensities \(0.12\) and \(0.03\) per year. With no recovery or death, find the probability at least one trigger has occurred by two years. (Answer: \(0.2592\))
- A folded model has \(H\)-to-eligible intensities \(0.08\) for ADL and \(0.07\) for cognitive impairment, and \(H\)-to-death intensity \(0.05\), all per year. Find the probability the first transition is an eligibility transition within one year. (Answer: \(0.1360\))
- Two eligible states have recovery intensities \(0.10\) per year from \(A\) and \(0.30\) per year from \(C\). Can they be combined into one Markov eligible state with a single recovery intensity while retaining the exact transition behavior from either starting state? State the reason. (Answer: No; the instantaneous exit intensities differ, so the future exit behavior depends on which underlying eligible state is occupied.)