Welcome to the World of Transition Intensities!
Hello there! Today, we are going to dive into one of the most fundamental tools in an actuary's toolkit: Maximum Likelihood Estimators (MLEs) for transition intensities. If you have ever wondered how insurance companies actually calculate the "risk" of someone moving from a 'Healthy' state to a 'Sick' state, or from 'Alive' to 'Dead', this is where the magic happens.
Don't worry if the math looks a bit intimidating at first. We are going to break it down step-by-step. By the end of this, you’ll see that finding the "best" estimate for a risk rate is actually quite logical!
What exactly is a Transition Intensity?
Before we derive anything, let’s quickly recap what we are looking at. In survival models, a transition intensity (often denoted by the Greek letter \(\mu\) or \(\sigma\)) represents the instantaneous rate of moving from one state to another.
Think of it like a speedometer. If you are driving at 60 mph, it doesn't mean you will drive for an hour; it just tells you how fast you are going at that exact moment. Similarly, \(\mu\) tells us the "force" or "speed" of mortality or morbidity at a specific time.
The Observational Plan and Statistics
To estimate transition intensities, we set up an observational plan. We observe a finite sample of \(N\) individuals over a finite investigation period. For each person, we record:
- When they enter the observation period (and in which state).
- The exact times of any transitions between states.
- When they leave observation (either through reaching the end of the study period, withdrawing, or entering an absorbing state).
From this plan, two key summary statistics emerge for each state \(i\) and transition \(i \to j\):
1. Waiting time (total central exposure) \(v_i\): the total time spent by all individuals while occupying state \(i\) and under observation.
2. Transition count \(d_{ij}\): the total number of transitions observed directly from state \(i\) to state \(j\).
The Goal: Maximum Likelihood Estimation (MLE)
In the real world, we observe these movements across our sample. We have the data, but we don't know the true value of \(\mu\). The Maximum Likelihood method asks: "Which value of \(\mu\) makes the data we actually observed the most likely to have happened?"
Step 1: Building the Likelihood Function
To find the MLE, we first write down the Likelihood Function, \(L(\mu)\). For a constant intensity \(\mu\) in a single-decrement model (e.g., Alive to Dead):
1. Staying in the state: The probability that an individual stays in the state for an observed duration \(t\) is \(e^{-\mu t}\). Multiplying over all individuals gives \(e^{-\mu v}\), where \(v = \sum t\) is the total waiting time.
2. Transitioning: If an individual transitions at a given moment, that event contributes a factor proportional to \(\mu\). For \(d\) total transitions, this contributes \(\mu^d\).
Combining these independent contributions yields the likelihood function:
\(L(\mu) = C \mu^d e^{-\mu v}\)
where \(C\) is an arbitrary constant independent of \(\mu\).
Step 2: The Log-Likelihood Function
To simplify differentiation, we take the natural logarithm of the likelihood function (ignoring additive constants that do not depend on \(\mu\)):
\(l(\mu) = \ln(L(\mu)) = d \ln(\mu) - \mu v\)
Step 3: Finding the Maximum
To maximize the log-likelihood, we take the first derivative with respect to \(\mu\) and set it to zero:
\(\frac{dl}{d\mu} = \frac{d}{\mu} - v = 0\)
Solving for \(\mu\) gives the MLE:
\(\hat{\mu} = \frac{d}{v}\)
Key Takeaway: The Maximum Likelihood Estimator for a transition intensity is simply the Total Number of Transitions divided by the Total Time at Risk.
Step 4: The Sanity Check (Second Derivative)
To confirm that \(\hat{\mu}\) maximizes the likelihood rather than minimizes it, we check the second derivative:
\(\frac{d^2l}{d\mu^2} = -\frac{d}{\mu^2}\)
Since \(d > 0\) and \(\mu^2 > 0\), the second derivative is strictly negative, confirming a unique local and global maximum.
Asymptotic Distribution of the Estimator
From maximum likelihood theory, as the sample size grows large, the estimator \(\hat{\mu}\) is asymptotically normally distributed:
\(\hat{\mu} \sim N\left(\mu, \frac{1}{I(\mu)}\right)\)
where \(I(\mu)\) is the expected Fisher information:
\(I(\mu) = -E\left[\frac{d^2l}{d\mu^2}\right] = E\left[\frac{D}{\mu^2}\right] = \frac{E[D]}{\mu^2}\)
Since \(E[D] = \mu E[V]\), we have \(I(\mu) = \frac{E[V]}{\mu}\). Thus, the asymptotic variance is:
\(\text{Var}(\hat{\mu}) \approx \frac{\mu}{E[V]} = \frac{\mu^2}{E[D]}\)
In practice, replacing parameters with our observed statistics gives the estimated variance:
\(\widehat{\text{Var}}(\hat{\mu}) = \frac{d}{v^2} = \frac{\hat{\mu}}{v}\)
Generalizing to Multiple States and Joint Distribution
What if individuals move between multiple states (e.g., Healthy, Sick, Dead)? Suppose we have a Markov model with states \(i, j, \dots\) and constant transition intensities \(\mu_{ij}\). The full likelihood function is:
\(L = \prod_{i} \prod_{j \neq i} \mu_{ij}^{d_{ij}} \exp(-v_i \mu_{ij})\)
Taking logs:
\(l = \sum_{i} \sum_{j \neq i} \left( d_{ij} \ln(\mu_{ij}) - \mu_{ij} v_i \right)\)
Differentiating with respect to each intensity \(\mu_{ij}\):
\(\frac{\partial l}{\partial \mu_{ij}} = \frac{d_{ij}}{\mu_{ij}} - v_i = 0 \implies \hat{\mu}_{ij} = \frac{d_{ij}}{v_i}\)
Notice what happens to the second-order partial derivatives:
- For the same transition: \(\frac{\partial^2 l}{\partial \mu_{ij}^2} = -\frac{d_{ij}}{\mu_{ij}^2}\)
- For distinct transitions \((i, j) \neq (r, s)\): \(\frac{\partial^2 l}{\partial \mu_{ij} \partial \mu_{rs}} = 0\)
Because all cross-partial derivatives are zero, the information matrix is diagonal. This establishes a vital result: the maximum likelihood estimators \(\hat{\mu}_{ij}\) are asymptotically independent, with asymptotic joint distribution:
\(\hat{\mu}_{ij} \sim N\left(\mu_{ij}, \frac{\mu_{ij}}{E[V_i]}\right)\)
The Poisson Approximation in a Single Decrement
There is another powerful way to view this estimator. In a single-decrement model over an observed waiting time \(v\), we can treat the number of transitions \(D\) as an approximate Poisson random variable with parameter \(\lambda = \mu v\):
\(P(D = d) = \frac{(\mu v)^d e^{-\mu v}}{d!}\)
Let’s write out the log-likelihood under this Poisson model:
\(l_{\text{Poisson}}(\mu) = d \ln(\mu) + d \ln(v) - \mu v - \ln(d!)\)
Differentiating with respect to \(\mu\):
\(\frac{dl_{\text{Poisson}}}{d\mu} = \frac{d}{\mu} - v = 0 \implies \hat{\mu} = \frac{d}{v}\)
The Poisson approximation leads to the exact same estimator and the same variance structure! This equivalence is widely used in actuarial graduation and generalised linear modelling.
Common Mistakes to Avoid
- Mixing up \(d\) and \(v\): Remember that the rate is "Events per Time," so \(d\) (events) goes on top, and \(v\) (time) goes on the bottom.
- Assuming exposures are identical: In multi-state models, \(v_i\) depends only on the time spent in state \(i\), not time spent in other states.
- Forgetting the asymptotic independence: Candidates often overlook that estimators for different transition intensities have zero asymptotic covariance because the cross-derivatives of the log-likelihood vanish.
Quick Review Box
Single intensity MLE: \(\hat{\mu} = \frac{d}{v}\)
Multiple intensities MLE: \(\hat{\mu}_{ij} = \frac{d_{ij}}{v_i}\)
Asymptotic variance: \(\text{Var}(\hat{\mu}) \approx \frac{\mu}{E[V]} \approx \frac{d}{v^2}\)
Asymptotic covariance: \(\text{Cov}(\hat{\mu}_{ij}, \hat{\mu}_{rs}) = 0\) for \((i, j) \neq (r, s)\)
Poisson approximation: Assuming \(D \sim \text{Poisson}(\mu v)\) gives the identical estimator \(\hat{\mu} = d/v\).
Summary and Key Takeaways
1. The observational plan records individual transition histories to produce the total waiting time \(v\) and transition counts \(d\).
2. Maximizing the log-likelihood function yields \(\hat{\mu} = d/v\) (or \(\hat{\mu}_{ij} = d_{ij}/v_i\) in multi-state systems).
3. The estimators are asymptotically normally distributed and mutually independent across distinct transitions.
4. Modelling transitions as a Poisson process produces the exact same maximum likelihood estimator, bridging survival models with generalized linear models.