Interpreting Kolmogorov Terms as Insurance-State Flows

From Transition Intensities to Insurance-State Flows

In a multiple-state model, an insured may move among states such as active, disabled, and dead. A transition intensity describes the instantaneous rate of movement out of one state and into another. The forward Kolmogorov equations translate those rates into changes in the probabilities of occupying each state. Reading the equations as inflows minus outflows gives them a direct insurance interpretation.

The distinction between a transition intensity and a probability flow matters. An intensity is a rate conditional on currently being in its starting state. To obtain the population-level flow per unit time, multiply the intensity by the probability of occupying that starting state.

Definition: Let \(p_j(t)\) be the probability that an insured is in state \(j\) at time \(t\). For distinct states \(i\) and \(j\), let \(\mu_{ij}(t)\) be the transition intensity from \(i\) to \(j\), measured per unit time. The instantaneous probability flow from \(i\) to \(j\) at time \(t\) is \(p_i(t)\mu_{ij}(t)\). It is a probability change per unit time, not a probability by itself.

For a finite-state model, the forward equation for occupancy of state \(j\) is the sum of flows into \(j\) from other states minus the sum of flows out of \(j\) to other states:

p_j'(t) = \sum_{i\ne j}p_i(t)\mu_{ij}(t) - p_j(t)\sum_{k\ne j}\mu_{jk}(t).

A transition from \(j\) to itself is not included: it does not change occupancy. If a transition is impossible in the model, its intensity is zero. The equations apply under the continuous-time Markov model represented by the specified intensities; they do not account for unmodelled transitions or for occupancy changes that are not represented by the states.

Why Each Term Is a Probability Flow

The inflow term \(p_i(t)\mu_{ij}(t)\) follows from conditioning on the state at time \(t\). For a small positive interval \(h\), an insured in state \(i\) has probability approximately \(\mu_{ij}(t)h\) of moving to \(j\), to first order in \(h\). The probability of being in \(i\) at time \(t\) is \(p_i(t)\). Thus the probability of a transition from \(i\) to \(j\) over that interval is approximately \(p_i(t)\mu_{ij}(t)h\). Dividing by \(h\) gives the instantaneous flow rate.

To justify the balance equation, count the ways occupancy of \(j\) can change over \([t,t+h]\). The probability in \(j\) at time \(t+h\) includes those already in \(j\) who make no exit, plus those who enter \(j\) from another state. To first order, the exit probability from \(j\) is \(p_j(t)h\sum_{k\ne j}\mu_{jk}(t)\), while inflows are \(h\sum_{i\ne j}p_i(t)\mu_{ij}(t)\). Events with two or more transitions over an interval of length \(h\) contribute only higher-order terms under the usual finite-intensity model. Therefore,

p_j(t+h)-p_j(t) = h\left( \sum_{i\ne j}p_i(t)\mu_{ij}(t) - p_j(t)\sum_{k\ne j}\mu_{jk}(t) \right)+o(h).

Dividing by \(h\) and taking the limit as \(h\) decreases to zero gives the forward equation. This argument also explains why the starting-state probability multiplies each intensity.

Key interpretation: Every term \(p_i(t)\mu_{ij}(t)\) is a flow from state \(i\) to state \(j\). It appears with a positive sign in the equation for \(p_j'(t)\) and with a negative sign in the equation for \(p_i'(t)\).

Worked Insurance-State Flow Calculations

Worked Example: Active, Disabled, and Dead Occupancy

A hypothetical insurance model has three states: active \(A\), disabled \(D\), and dead \(L\). At time \(t\), the occupancy probabilities are \(p_A(t)=0.70\), \(p_D(t)=0.25\), and \(p_L(t)=0.05\). The transition intensities per year are \(\mu_{AD}=0.12\), \(\mu_{AL}=0.01\), \(\mu_{DA}=0.24\), and \(\mu_{DL}=0.04\). Death is absorbing. Find all three occupancy derivatives at this instant.

Method. For each state, list its incoming flows and outgoing flows. The flow from active to disabled is \(0.70(0.12)=0.084\) per year. The flow from active to dead is \(0.70(0.01)=0.007\) per year. The flow from disabled to active is \(0.25(0.24)=0.060\) per year, and the flow from disabled to dead is \(0.25(0.04)=0.010\) per year.

For the active state, the total exit intensity is \(0.12+0.01=0.13\) per year. Its only inflow is from disabled, so \(p_A'(t)=0.060-0.70(0.13)=0.060-0.091=-0.031\) per year.

For the disabled state, the inflow is from active and the exits are to active and dead: \(p_D'(t)=0.70(0.12)-0.25(0.24+0.04)=0.084-0.070=0.014\) per year.

For the dead state, there are no exits, and the inflows are from active and disabled: \(p_L'(t)=0.70(0.01)+0.25(0.04)=0.007+0.010=0.017\) per year.

Check. The derivatives sum to \(-0.031+0.014+0.017=0\), as required for probabilities whose total remains one. The method avoids treating an intensity as a flow: for example, \(0.12\) is not the active-to-disabled flow; the flow is \(0.70(0.12)\).

Worked Example: Separating Competing Transitions from Active

In a hypothetical long-term care model, the states are active \(A\), receiving care \(C\), and dead \(L\). At a particular instant, the probabilities are \(p_A=0.60\), \(p_C=0.30\), and \(p_L=0.10\). The intensities per year are \(\mu_{AC}=0.08\), \(\mu_{AL}=0.01\), \(\mu_{CA}=0.20\), and \(\mu_{CL}=0.04\); dead is absorbing. Calculate the derivatives and identify the net effect for each state.

Method. The active state has two competing exits, to care and to death, so its total exit intensity is \(0.08+0.01=0.09\). Its inflow is the recovery flow from care, \(0.30(0.20)=0.060\). Therefore, \(p_A'=-0.60(0.09)+0.060=-0.054+0.060=0.006\) per year.

The care-state inflow is \(0.60(0.08)=0.048\). Its total exit intensity is \(0.20+0.04=0.24\), giving an outflow of \(0.30(0.24)=0.072\). Hence, \(p_C'=0.048-0.072=-0.024\) per year.

The death-state inflows are \(0.60(0.01)=0.006\) from active and \(0.30(0.04)=0.012\) from care. Thus \(p_L'=0.006+0.012=0.018\) per year.

Check and interpretation. The sum is \(0.006-0.024+0.018=0\). At this instant, active occupancy is increasing slightly because recovery inflow exceeds active exits, while care occupancy is decreasing. The positive derivative for active does not mean that no active individuals are leaving; it means total inflow exceeds total outflow.

Worked Example: Interpreting Flows with Time-Varying Intensities

A hypothetical disability model has active, disabled, and dead states. At time \(t=5\) years, the probabilities are \(p_A=0.65\), \(p_D=0.30\), and \(p_L=0.05\). The time-dependent intensities at that instant are \(\mu_{AD}(5)=0.10\), \(\mu_{AL}(5)=0.02\), \(\mu_{DA}(5)=0.25\), and \(\mu_{DL}(5)=0.05\), all per year. Find the instantaneous derivatives and use them to give first-order occupancy estimates \(0.01\) year later.

Method. Use the intensities evaluated at \(t=5\), rather than treating them as constant beyond that instant. The active exits total \(0.10+0.02=0.12\), and the disabled-to-active inflow is \(0.30(0.25)=0.075\). Thus \(p_A'(5)=-0.65(0.12)+0.075=-0.078+0.075=-0.003\) per year.

The disabled inflow is \(0.65(0.10)=0.065\), while disabled exits total \(0.30(0.25+0.05)=0.090\). Therefore \(p_D'(5)=0.065-0.090=-0.025\) per year. The death inflows are \(0.65(0.02)=0.013\) and \(0.30(0.05)=0.015\), so \(p_L'(5)=0.028\) per year.

For \(h=0.01\) year, the first-order approximation is \(p_j(5+h)\approx p_j(5)+h p_j'(5)\). It gives \(p_A(5.01)\approx0.65+0.01(-0.003)=0.64997\), \(p_D(5.01)\approx0.30+0.01(-0.025)=0.29975\), and \(p_L(5.01)\approx0.05+0.01(0.028)=0.05028\).

Check and limitation. These estimates sum to \(0.64997+0.29975+0.05028=1.00000\). They are first-order estimates, not exact transition probabilities over the interval; changing intensities and multiple transitions can affect the higher-order terms. The calculation avoids using an instantaneous derivative as though it were an exact finite-interval change.

Reading and Checking the Equations

A practical way to interpret any state equation is to make a flow ledger. Write each transition as “current occupancy probability multiplied by its transition intensity.” Then place that flow in the ledger as an outflow from its origin and an inflow to its destination. This keeps the signs tied to the direction of movement, not to the name of the state.

Identify the state whose derivative is required.
Only transitions entering that state add occupancy; only transitions leaving it reduce occupancy.

Calculate each flow.
Multiply the probability of occupying the origin state by the applicable transition intensity at the time in question.

Combine inflows and outflows.
Add incoming flows and subtract outgoing flows to obtain the occupancy derivative per unit time.

Check the system.
When all states are included and probabilities sum to one, the sum of their derivatives should be zero.

The conservation check follows because each transition flow is counted once positively at its destination and once negatively at its origin. Thus all transition terms cancel when the state equations are summed. In a model with an absorbing dead state, death flows still cancel across the full system: they leave a living state and enter the dead state. Absorption means there are no flows out of dead, not that the death-state probability cannot change.

Common exam trap: Do not subtract a transition intensity directly from an occupancy probability or treat the intensity itself as a population flow. For example, with \(p_A=0.80\), \(p_D=0.20\), \(\mu_{AD}=0.10\), and \(\mu_{DA}=0.30\), the incorrect calculation \(p_D'=0.10-0.30=-0.20\) omits the origin-state probabilities. The correct calculation is \(p_D'=0.80(0.10)-0.20(0.30)=0.08-0.06=0.02\) per year, and \(p_A'=-0.02\) per year. The signs are also easy to reverse: the active-to-disabled flow increases disabled occupancy and decreases active occupancy.

Interpreting the forward equations this way is useful whenever a model has insurance states with different transition patterns. It gives a direct check on equations before solving them and helps explain whether an occupancy probability is rising or falling at a given time. The rates and probabilities must use the same time scale: if intensities are per year, the resulting derivatives are probability changes per year.

Check Your Understanding

For each hypothetical model, calculate the requested instantaneous derivatives from the stated occupancies and intensities.

A two-state model has active and disabled states, with \(p_A=0.80\), \(p_D=0.20\), \(\mu_{AD}=0.10\), and \(\mu_{DA}=0.30\) per year. Find both occupancy derivatives. (Answer: \(p_A'=-0.020\), \(p_D'=0.020\) per year.)
A model has active, disabled, and dead states, with occupancies \(0.70, 0.30, 0\). The only transitions are active to disabled at \(0.12\) per year and disabled to active at \(0.24\) per year. Find the disabled-state derivative and explain whether it is increasing or decreasing. (Answer: \(p_D'=0.70(0.12)-0.30(0.24)=0.084-0.072=0.012\) per year; disabled occupancy is increasing.)
A hypothetical care model has active, care, and dead states with occupancies \(0.60,0.25,0.15\). Its intensities per year are \(\mu_{AC}=0.05\), \(\mu_{AL}=0.02\), \(\mu_{CA}=0.40\), and \(\mu_{CL}=0.08\), with dead absorbing. Find all three derivatives and check their sum. (Answer: \(p_A'=-(0.05+0.02)(0.60)+0.40(0.25)=0.058\), \(p_C'=0.05(0.60)-(0.40+0.08)(0.25)=-0.090\), \(p_L'=0.02(0.60)+0.08(0.25)=0.032\) per year; the sum is \(0\).)
At time \(t=3\), a time-inhomogeneous model has active, disabled, and dead probabilities \(0.50,0.40,0.10\). At that instant, \(\mu_{AD}=0.20\), \(\mu_{AL}=0.03\), \(\mu_{DA}=0.10\), and \(\mu_{DL}=0.07\), per year. Find the three instantaneous derivatives. (Answer: \(p_A'=-0.23(0.50)+0.10(0.40)=-0.075\), \(p_D'=0.20(0.50)-0.17(0.40)=0.032\), \(p_L'=0.03(0.50)+0.07(0.40)=0.043\) per year.)
A hypothetical model has active, disabled, and dead states, with probabilities \(0.55,0.35,0.10\). Intensities per year are \(\mu_{AD}=0.06\), \(\mu_{AL}=0.02\), \(\mu_{DA}=0.18\), and \(\mu_{DL}=0.04\); dead is absorbing. Calculate the derivatives, then give first-order occupancy estimates \(0.02\) year later. (Answer: \(p_A'=-0.08(0.55)+0.18(0.35)=0.019\), \(p_D'=0.06(0.55)-0.22(0.35)=-0.044\), and \(p_L'=0.02(0.55)+0.04(0.35)=0.025\) per year; estimates are \(0.55038, 0.34912, 0.10050\), respectively, and sum to \(1\).)

What You'll Learn