Calculate maximum likelihood estimates of transition intensities and probabilities for multiple state and multiple decrement models, assuming piecewise constant transition intensities.

You watch a portfolio of disability-income policies for a year. Twenty-three active lives became disabled, nine died directly from the active state, and fourteen disabled lives died. The active cohort logged 1,840 person-years of exposure; the disabled cohort logged 312. Those five numbers, plus the assumption that intensities are constant inside an age band, are everything you need to build the MLE.

A life under observation contributes two kinds of information. While occupying state $i$, it contributes a survivorship factor $\exp (-{\mu }^{i\bullet }t)$, where ${\mu }^{i\bullet }={\sum }_{j\ne i}{\mu }^{ij}$ is the total force out of $i$. At a transition $i\to j$, it contributes the intensity ${\mu }^{ij}$.

$$L=\prod _i\left[\exp \left(-{\mu }^{i\bullet }{v}^i\right)\prod _{j\ne i}{\left({\mu }^{ij}\right)}^{d^{ij}}\right]$$

Here $v^i$ is the total central exposure in state $i$ (sum of time all lives spent there) and $d^{ij}$ counts observed $i\to j$ transitions.

Logging converts the product into a sum, so each $(i,j)$ pair contributes an independent term you can maximize on its own.

Common mistakes

• Dividing by wrong exposure. ${\hat{\mu }}^{12}$ uses disabled exposure 312, not active 1,840. Mixing them gives 14/1840 = 0.00761, far too low.
• Forgetting censored time. A life observed for only half the band contributes 0.5 person-years, not 1.0. Skipping censoring inflates denominators.
• Using $d^{ij}/(d^{ij}+\text{still in }i)$. That is an empirical probability, not an intensity. The denominator must be exposure time, not headcount.

Bottom line

• Core MLE: ${\hat{\mu }}^{ij}=d^{ij}/v^i$. Transitions out of $i$ into $j$ over total waiting time in $i$.
• Independent estimation: the likelihood factors by origin-destination pair, so each ${\mu }^{ij}$ is maximized on its own.
• Piecewise constant means ${\mu }^{ij}$ is flat inside each age band; you re-estimate for each band.
• Right exposure: the denominator is time in origin state $i$ only. Time in state $j$ never enters ${\hat{\mu }}^{ik}$ for $i\ne j$.

Exam shortcut

Write down a "transitions over exposure" table before any algebra. Rows are origin states, columns are destinations, cells are $d^{ij}/v^i$. The diagonal stays blank. DECISION: One band asked → use $e^{-{\mu }^{i\bullet }h}$ and the ratio formula directly. Multiple bands asked → chain via ${}_{a+b}{p}^{ii}={}_a{p}^{ii}\cdot {}_b{p}^{ii}$ with each factor band-specific. For competing decrements, $q^{(j)}=({\mu }^{(j)}/{\mu }^{(\tau )})(1-e^{-{\mu }^{(\tau )}})$. Memorize the split-then-scale pattern; it recurs everywhere intensities are constant.

The full lesson (about 1,798 words, 12 min read) adds 2 worked examples, all 6 common mistakes, a self-check, free in the app.

Learning objectives

• 1h

Browse all free Exam ALTAM lessons or jump into free Exam ALTAM practice questions.