Appendix: Kannisto-Thatcher Model
For the age groups 95 to 109 and the open age group of 110 years and over,
Statistics Canada models mortality using the Kannisto-Thatcher model, described in:
On the use of Kannisto model for mortality trajectory modelling at very old ages.
See also: Methods for Constructing Life Tables for Canada, Provinces and Territories.
According to the Kannisto-Thatcher model, the instantaneous probability of death at age \(x\)
is given by:
\[\begin{split}\mu(x, \Delta x, t) &= \dfrac{\alpha(t) e^{\beta(t) x}}{1 + \alpha(t) e^{\beta(t) x}} \\
&= \lim_{\Delta x \to 0}
\dfrac{P(\text{death between age $x$ and $x + \Delta x$} \mid \text{survived till $x$})}{\Delta x}\end{split}\]
In mathematical terms, \(\mu(x, \Delta x, t)\) is the hazard rate. Let’s break this down further.
Let \(F_X(x)\) be the cumulative distribution function for age at death, \(X\):
\[\begin{split}F_X(x) :&= P(\text{age at death} \leq \text{given age}) \\
&= P(X \leq x)\end{split}\]
We want the conditional probability of death between age \(x\) and \(x + \Delta x\),
given that the person has survived till age \(x\). This is given by:
\[P(x < X \leq x + \Delta x \mid X > x)\]
Recall that for a conditional probability:
\[P(A \mid B) = \dfrac{P(A \cap B)}{P(B)}\]
and so:
\[P(x < X \leq x + \Delta x \mid X > x) = \dfrac{P(x < X \leq x + \Delta x \bigcap X > x)}{P(X > x)}\]
Since \(F_X(x)\) is the cumulative distribution function, by definition it must sum to 1:
\[P(X > x) = 1 - F_X(x)\]
Since \(X > x\) if \(x < X \leq x + \Delta x\), we can rewrite the numerator as:
\[\begin{split}P(x < X \leq x + \Delta x \bigcap X > x) &= P(x < X \leq x + \Delta x) \\
&= F_X(x + \Delta x) - F_X(x)\end{split}\]
Putting it all together, we have:
\[P(x < X \leq x + \Delta x \mid X > x) =
\dfrac{F_X(x + \Delta x) - F_X(x)}{1 - F_X(x)}\]
Now, we want to find the instantaneous rate of death; the probability of death per unit time.
If we take the limit as \(\Delta x \to 0\), we will find the instantaneous probability of
death at age \(x\). To get the probability of death per unit time, we need to divide by
\(\Delta x\):
\[\mu(x, \Delta x, t) = \lim_{\Delta x \to 0} \dfrac{F_X(x + \Delta x) - F_X(x)}{\Delta x (1 - F_X(x))}\]
You will recognize the derivative of \(F_X(x)\):
\[\dfrac{d}{dx} F_X(x) = \lim_{\Delta x \to 0} \dfrac{F_X(x + \Delta x) - F_X(x)}{\Delta x}\]
and so:
\[\mu(x, \Delta x, t) = \dfrac{F_X'(x)}{1 - F_X(x)}\]
The data in the Statistics Canada mortality table is the probability of death between age
\(x\) and \(x + 1\), which is denoted as \(q_x\). This is the same as the probability
\(P(x < X \leq x + \Delta x \mid X > x)\), with \(\Delta x = 1\). We would like to solve
for \(q(x, \Delta x, t)\), using the Kannisto-Thatcher Equation for \(\mu(x)\). First,
we can write \(q(x, \Delta x, t)\) in terms of \(F_X(x)\):
\[\begin{split}q(x, \Delta x, t) &= P(x < X \leq x + 1 \mid X > x) \\
&= \dfrac{F_X(x + 1) - F_X(x)}{1 - F_X(x)}\end{split}\]
Let us define \(S_X(x)\), the survival function, for convenience:
\[\begin{split}S_X(x) &:= 1 - F_X(x) \\
&= P(X > x)\end{split}\]
Then we have:
\[\dfrac{dS}{dx} = -F_X'(x)\]
and so \(\mu(x)\) can be rewritten as:
\[\mu(x, \Delta x, t) = -\dfrac{dS}{dx}\dfrac{1}{S_X(x)}\]
Solving this first order separable linear differential equation, we have:
\[F_X(x) = 1 - k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}}\]
Math: \(F_X(x)\)
\[\begin{split}\int \dfrac{dS}{S_X} &= -\int \mu(x, \Delta x, t) dx \\
\ln(S_X(x)) &= -\int \mu(x, \Delta x, t) dx + C \\
&= -\int \dfrac{\alpha(t) e^{\beta(t) x}}{1 + \alpha(t) e^{\beta(t) x}} dx + C\end{split}\]
Letting \(u(x, t) := 1 + \alpha(t) e^{\beta(t) x}\), we have:
\[\begin{split}\ln(S_X(x)) &= - \dfrac{1}{\beta(t)} \int \dfrac{du}{u} + C \\
&= - \dfrac{1}{\beta(t)} \ln(u(x, t)) + C \\
S_X(x) &= e^C (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}} \\
&= k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}} \\
1 - F_X(x) &= k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}} \\
F_X(x) &= 1 - k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}}\end{split}\]
Now, we can substitute this into the equation for \(q(x, \Delta x, t)\):
\[\begin{split}q(x, \Delta x, t) &= \dfrac{F_X(x + \Delta x) - F_X(x)}{1 - F_X(x)} \\
&= 1 - \left(
\dfrac{1 + \alpha(t) e^{\beta(t) x}}{1 + \alpha(t) e^{\beta(t) (x + \Delta x)}}
\right)^{\frac{1}{\beta(t)}}\end{split}\]
Math: \(q(x, \Delta x, t)\)
\[\begin{split}q(x, \Delta x, t) &= \dfrac{F_X(x + \Delta x) - F_X(x)}{1 - F_X(x)} \\
&= \dfrac{
1 - k (1 + \alpha(t) e^{\beta(t) (x + \Delta x)})^{-\frac{1}{\beta(t)}} -
1 + k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}}
}{k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}}} \\
&= \dfrac{
- k (1 + \alpha(t) e^{\beta(t) (x + \Delta x)})^{-\frac{1}{\beta(t)}}
+ k (1 + \alpha(t) e^{\beta x})^{-\frac{1}{\beta(t)}}
}{k (1 + \alpha(t) e^{\beta(t) x})^{-\frac{1}{\beta(t)}}} \\
&= 1 - \left(\dfrac{1 + a e^{\beta(t) (x + \Delta x)}}{1 + \alpha(t) e^{\beta(t) x}}\right)^{-\frac{1}{\beta}} \\
&= 1 - \left(
\dfrac{1 + \alpha(t) e^{\beta(t) x}}{1 + \alpha(t) e^{\beta(t) (x + \Delta x)}}
\right)^{\frac{1}{\beta(t)}}\end{split}\]
Now, based on fitting the model to empirical data, typically we have
[Appendix D, Table 5, [Kannisto, 1994]]:
\(\beta \approx \mathcal{O}(10^{-1})\)
\(\alpha \approx \mathcal{O}(10^{-5})\)