Professional Documents
Culture Documents
Subject CS2
Revision Notes
For the 2019 exams
covering
CONTENTS
Contents Page
Copyright agreement
Legal action will be taken if these terms are infringed. In addition, we may
seek to take disciplinary action through the profession or through your
employer.
These conditions remain in force after you have finished using the course.
These chapter numbers refer to the 2019 edition of the ActEd course notes.
The numbering of the syllabus items is the same as that used by the Institute
and Faculty of Actuaries.
OVERVIEW
Exam questions on this topic can ask you to calculate transition probabilities
or derive expressions for such probabilities in terms of forces of transition.
To do this, we might write down and solve differential equations or integral
equations. Exam questions may be based around examples that are not
described in the Core Reading, eg machines breaking down or spaces
available in a car park. Try not to be put off by this – the same principles
apply in every case.
CORE READING
All of the Core Reading for the topics covered in this booklet is contained in
this section.
The text given in Arial Bold Italic font is additional Core Reading that is not
directly related to the topic being discussed..
____________
Notation
Different authors tend to use different notation for the same quantities,
and the Markov model is an example of this. Actuaries have often use
notations derived from the standardised International Actuarial
Notation, in which the ‘transition rate’ is the force of mortality m x , and
the corresponding probabilities are the life table probabilities t px and
t qx . Moreover the index x is generally understood to indicate age
(eg age now, or age at policy inception) and the index t indicates
duration since age x . Probabilists and statisticians tend to use
different notations.
The probability that an event occurs during the short time interval from
time t to time t + h is approximately equal to l h for small h ; the
parameter l is called the rate of the Poisson process.
P ÈÎNt + h - Nt = 1| Ft ˘˚ = l h + o(h )
f (h )
where the statement that f (h) = o(h) as h Æ 0 means lim = 0.
h Æ0 h
____________
It is far from obvious that the process defined above coincides with the
Poisson process characterised in Booklet 1 as having independent
stationary Poisson-distributed increments. That is one of the
properties that we shall prove.
____________
5 Proof
e - lt (l t )
j
p j (t ) = (4.2)
j!
p j (t + h) = P (Nt + h = j )
= P (Nt = j and Nt + h = Nt )
+P (Nt = j - 1 and Nt + h = Nt + 1) + o(h)
= p j (t )(1 - l h) + p j -1(t )l h + o(h )
dp j
= - l p j (t ) + l p j -1(t ) (4.3)
dt
dp0
= - l p0 (t ) (4.4)
dt
with p0 (0) = 1 .
6 In view of the fact that the increments of N are stationary and are
independent of the past, this result may be generalised to a statement
that Nt + s - Ns is a Poisson random variable with mean l t ,
independent of anything that has occurred before time s .
____________
Inter-event times
7 Since the Poisson process Nt changes only by unit upward jumps, its
sample paths are fully characterised by the times at which the jumps
take place. Denote by T0 , T1, T2 , .... the successive inter-event times (or
holding times), a sequence of random variables.
Xt
t
T0 T1 T2 T3 T4 T5 T6
Proof
9 P (T0 > t ) is the probability that no events occur between time 0 and
(
pij (t ) = P X t = j X 0 = i )
obey the Chapman-Kolmogorov equations:
P È X 0 = i , X t1 = j1, X t2 = j2 , ..., X tn = jn ˘
Î ˚
= qi pij1 (t1) p j1 j2 (t2 - t1) p jn -1 jn (t n - t n -1)
and:
P ÈÎ X t1 = j1, X t2 = j2 , ..., X tn = jn ˘˚
= Â qi pij1 (t1) p j1 j2 (t 2 - t1) p jn -1 jn (t n - t n -1)
i ŒS
____________
18 Noting that:
Ï0 if i π j
pij (0) = d ij = Ì (4.6)
Ó1 if i = j
d pij (h) - d ij
mij = pij (t )Ωt = 0 = lim
dt h Æ0 h
____________
20 The interpretation of the first line of (4.7) is simply that the probability
of a transition from i to j during any short time interval [ s , s + h ] is
proportional to h ; hence the name transition rate given to mij .
____________
mii = - Â mij
j πi
____________
d
pij (t ) = Â pik (t ) mkj for all i , j (4.8)
dt k ŒS
____________
d
P (t ) = P (t ) A
dt
Recall that A is often called the generator matrix of the Markov jump
process.
If the state space S is finite, (4.8) gives for each fixed i a finite linear
system of differential equations (in fact the index i enters only through
the initial condition (4.6)). Accordingly, for given transition rates mij ,
Equation (4.8) has a unique solution compatible with (4.6). For this
reason Markov models are normally formulated simply by specifying
their transition rates mij .
____________
d
P (t ) = A P (t )
dt
The forward equations are more useful in numerical work for actuarial
applications because we usually have an initial condition such as
knowing that a policyholder is healthy when a policy is sold, thus we
want equations that we can solve forwards in time from that starting
point.
____________
27 Consider the Markov jump process with state space S = {0, 1, 2, ...} and
transition rates:
Ï-l if j = i
Ô
mij = Ìl if j = i + 1
Ô0 otherwise
Ó
0 1 2 3 i i+1
____________
Ê -l l 0ˆ
Á -l l ˜
Á ˜
A=Á -l l ˜
Á ˜
Á ˜
ÁË 0 ˜¯
____________
ÏÔ pi¢0 (t ) = - l pi 0 (t )
Ì
ÔÓ pij¢ (t ) = l pij -1(t ) - l pij (t ), j > 0
Holding times
li = - mii = Â mij
j πi
In other words:
P ÈÎT0 > t | X 0 = i ˘˚ = e - li t
____________
P ÈÎ X t + h = j , t < T0 £ t + h | X 0 = i ˘˚
= P ÈÎ X t + h = j , T0 > t | X 0 = i ˘˚
= P ÈÎ X t + h = j | X 0 = i , T0 > t ˘˚ P ÈÎT0 > t | X 0 = i ˘˚
= P ÈÎ X t + h = j | X s = i , 0 £ s £ t ˘˚ e - li t
- li t
= Pij (h )e
mij e - li t
This proves two results at once: the probability that the jump out of i
to state j is:
m
P ÈÎ X T0 = j | X 0 = i ˘˚ =
ij
li
( j π i)
1
36 Note that the mean holding time of state j is ; this is an important
lj
thing to remember when assigning numerical values to the transition
rates.
____________
Xˆ n = X T0 +T1 ++Tn -1
39 The only way in which the jump chain differs from a standard Markov
chain is when the jump process {X t , t ≥ 0} encounters an absorbing
state. From that time on it makes no further transitions, implying that
time stops for the jump chain. In order to deal with the jump chain
entirely within the framework of Markov chains it is permissible to treat
the absorbing state in the same way as for a Markov chain, so that
transitions continue to occur but the chain remains in the same state
after the transition.
____________
0 = ACTIVE
01 02
1 = DEAD 2 = RETIRED
We have:
m01
p01 ( x , x + t ) = È1 - e - ( m01 + m02 )t ˘
m01 + m02 Î ˚
m02
p02 ( x , x + t ) = È1 - e -( m01 + m02 )t ˘
m01 + m02 Î ˚
(t)
H: Healthy S: Sick
(t)
(t) v(t)
D: Dead
43 Define:
N
We also need to define totals V = Â Vi (and so on).
i =1
____________
L( m ,n , s , r ) = e - ( m +s )v e - (n + r )w m dn us s r r
____________
D U S R
m = , n = , s = , r =
V W V W
____________
Ê m ˆ
m ~ Normal Á m , etc
Ë E[V ] ˜¯
49 The calculation of the estimates m , etc, requires the total waiting time
to be computed. This can be done exactly in some circumstances, but,
if the exposure data are in census form, the simple census formulae in
Booklet 5 provide estimates. Multiple state models are, therefore,
especially well suited to the data available in many actuarial
investigations.
____________
Pij (s , t ) = P ÈÎ X t = j | X s = i ˘˚ (s £ t )
which obey a version of the Chapman-Kolmogorov equations, written
in matrix form as:
52 We see that the only difference between this case and the
time-homogeneous case studied earlier is that the transition rates
mij (s ) are allowed to change over time.
____________
∂
P ( s , t ) = P ( s , t ) A (t )
∂t
∂
P (s , t ) = - A (s ) P (s , t )
∂s
____________
mii (s ) = - Â mij (s )
j πi
55 A two-state model
Consider the following survival model: transition from the alive state A
to the dead state D takes place at rate m AD (t ) , which has been
abbreviated to m (t ) here, since it is the only transition in the model.
(t)
A D
Ê - m (t ) m (t )ˆ
56 Since A (t ) = Á the forward equations give:
Ë 0 0 ˜¯
∂
p (s , t ) = - m (t ) pAA ( s , t )
∂ t AA
____________
Ê t ˆ
pAA ( s , t ) = exp Á - Ú m ( x ) dx ˜
Ë s ¯
____________
Ê s +w ˆ Ê w ˆ
w ps = pAA ( s , s + w ) = exp Á - Ú m ( x )dx ˜ = exp Á - Ú m ( s + y )dy ˜
Ë s ¯ Ë 0 ¯
(5.1)
{Rs > w , X s = i } = {X u = i , s £ u £ s + w }
____________
Formula (5.1) gives the probability that Rs > w given that the state at
time s is A.
____________
Ê s +w ˆ
P ÎÈRs > w | X s = i ˚˘ = exp Á - Ú li (t ) dt ˜ (5.2)
Ë s ¯
____________
X s+ = X s + Rs
mij (s + w )
P È X s+ = j | X s = i , Rs = w ˘ = (5.3)
Î ˚ li (s + w )
____________
61 The above is more than a neat picture for the behaviour of Markov
jump processes: it is also a powerful computational tool. Indeed,
conditioning on Rs and X s+ we have using the law of total probability:
pij ( s , t ) = P ÈÎ X t = j | X s = i ˘˚
t -s s +w
li (u )du
e Ús
-
= Â Ú mil (s + w ) P È X t = j | X s = i , Rs = w , X s+ = l ˘dw
Î ˚
l πi 0
and therefore:
t -s s +w
li (u )du
e Ús
-
pij ( s , t ) = Â Ú mil (s + w ) Plj (s + w , t ) dw (5.4)
l πi 0
Ê s +w ˆ
li (s + w ) exp Á - Ú li (u ) du ˜
Ë s ¯
state i l j
time s s + w t
____________
Ê t ˆ
63 When i = j there is an additional term exp Á - Ú li (u ) du ˜ because the
Ë s ¯
process can remain in state i throughout [ s , t ] .
____________
If instead of considering the first jump after s one focuses on the last
jump before t , one can obtain an intuitive derivation of the integrated
form of the forward equations.
____________
t
t -s - Ú l j (u )du
pij ( s , t ) = Â Ú Pik (s , t - w ) mk j (t - w )e t -w dw (5.5)
kπ j 0
____________
{Ct w, Xt = j} = { X u = j, t – w u t}
____________
Applications
Marriage
M
(t)
(t)
v(t) r(t)
d(t)
B W D
(t)
(t)
(t) (t)
____________
P ÎÈ X t = M , Ct > w | X s = B ˚˘
t -s
= Ú ÈÎ pBB (s , t - v ) a (t - v ) + pBW (s , t - v ) r (t - v )
w
t
+ pBD (s , t - v ) r (t - v ) ˘˚ e Út -v
- ( m (u )+n (u ) +d (u ))dudv
state B k M
time s t – v t – w t
(t)
H: Healthy S: Sick
(t)
(t) v(t)
D: Dead
____________
Ê -s (t ) - m (t ) s (t ) m (t )ˆ
A (t ) = Á r (t ) - r (t ) - n (t ) n (t ) ˜
Á ˜
ËÁ 0 0 0 ˜¯
____________
70 In particular:
lH (t ) = s (t ) + m (t ) lS (t ) = r (t ) + n (t ) and lD = 0
____________
Ê t ˆ
( )
P ÈÎRs > t - s | X s = H ˘˚ = exp Á - Ú s (u ) + m (u ) du ˜ (5.6)
Ë s ¯
and:
Ê t ˆ
( )
P ÎÈRs > t - s | X s = S ˚˘ = exp Á - Ú r (u ) + n (u ) du ˜
Ë s ¯
____________
s +w
t -s - Ú (s (u ) + m (u ))du
pHS ( s , t ) = Ú PSS ( s + w , t ) s ( s + w ) e s dw
0
state H S S
time s s + w t
____________
P ÈÎ X t = S , Ct > w | X s = H ˘˚
t
t -s - Ú ( r (u ) +n (u ))du
= Ú pH H (s , t - v ) s (t - v ) e t -v
dv
w
____________
P ÈÎ X t = H | X s = S , Cs = w ˚˘ = P ÎÈ X t = H | X s = S ˚˘
____________
(t)
H: Healthy S: Sick
(t Ct)
D: Dead
____________
77 This appears to take us outside the scope of this unit, as the value
of Ct must now be incorporated into the state, so that the state space
is not countable (ie discrete as opposed to continuous) any more.
However, the framework from the start of this section can still be used
provided that there is careful conditioning on the relevant current
holding time.
state H S
time s – w s u t
w – s + u
____________
t
- Ú ( r (u ,w - s +u )+n (u ,w - s +u ))du
78 P ÎÈ X t = S , Rs > t - s | X s = S , Cs = w ˚˘ = e s
____________
pSw H (s , t ) = P ÈÎ X t = H | X s = S , Cs = w ˘˚
v
t - Ú ( r (u ,w - s +u ) +n (u ,w - s +u ))du
= Úe s r (v , w - s + v ) pHH (v , t )dv
s
state H S H H
time s – w s u v t
w – s + u
____________
Example
81 Divide the total time T into k equal intervals. If the Poisson process
model fits, the number of claims arriving in the k intervals should
form a sequence of independent Poisson variates, each with
mean lT / k . There are two things to test here:
____________
Thus data for a whole year may be divided into months, giving 12
estimated claim arrival rates. Tests of goodness of fit should be
carried out for each month separately, but tests for serial correlation
should use the whole data set at once.
____________
First the li may be estimated: look at the data for the durations of
87 Tests for goodness of fit are more problematical, if only because there
is a vast collection of possible alternative hypotheses. It is reasonable
to test whether the visits to a given state really are exponentially
distributed: a c 2 goodness of fit test will do this.
____________
It is also reasonable to test whether the jump chain really does exhibit
the Markov property: see Booklet 2 for a discussion. But there are
other implications of the Markov structure that should be tested and
the procedure is not always clear.
____________
Other tests, such as testing whether the first visit to a given state is
significantly longer than subsequent visits, are also best treated
graphically.
____________
Simulation
90 Approximate method
Divide time into very short intervals of width h , say, where h mij is
much smaller than 1 for each i and j . The transition matrix P (h ) of
the Markov chain has entries approximately given by:
pij* (h ) = d ij + h mij
91 Exact method
____________
Time-inhomogeneous processes
This section contains all the relevant exam questions from 2008 to 2017 that
are related to the topics covered in this booklet.
Solutions are given after the questions. These give enough information for
you to check your answer, including working, and also show you what an
outline examination answer should look like. Further information may be
available in the Examiners’ Report, ASET or Course Notes. (ASET can be
ordered from ActEd.)
We first provide you with a cross-reference grid that indicates the main
subject areas of each exam question. You can use this, if you wish, to
select the questions that relate just to those aspects of the topic that you
may be particularly interested in reviewing.
Alternatively, you can choose to ignore the grid, and attempt each question
without having any clues as to its content.
Cross-reference grid
Question attempted
Generator matrix
Poisson process
Markov property
Differential eqns
Formulation as
Defn of MJP /
Waiting times
Integral eqns
probabilities
Jump chain
Occupancy
Transitionn
Question
diagram
MLEs
MJP
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
The probabilities that the customer is online and offline at time t are PON (t )
and POFF (t ) respectively.
¢ (t ) .
(ii) Write down Kolmogorov’s forward equation for POFF [2]
(iii) Solve the equation in part (ii) to obtain a formula for the probability that a
customer is offline at time t , given that they were offline at time 0. [3]
(iv) Calculate the expected proportion of time spent online over the period
[0, t ] . [HINT: Consider the expected value of an indicator function
which takes the value 1 if offline and 0 otherwise.] [4]
An investigation was carried out into the relationship between sickness and
mortality in an historical population of working class men. The investigation
used a three-state model with the states:
1 Healthy
2 Sick
3 Dead
(i) Draw a diagram showing the three states and the possible transitions
between them. [2]
∂ 23 21 13 22 23
t px = t p x m x + t + t px m x + t [5]
∂t
(iii) Write down the likelihood of the data in the investigation in terms of the
transition rates and the waiting times in the Healthy and Sick states,
under the assumption that the transition rates are constant. [3]
(iv) Derive the maximum likelihood estimator of the transition rate from Sick
to Dead. [3]
(a) the value of the constant transition rate from Sick to Dead
(b) 95 per cent confidence intervals around this transition rate. [4]
[Total 17]
In the village of Selborne in Southern England in the year 1637 the number
of babies born each month was as follows:
January 2 July 5
February 1 August 1
March 1 September 0
April 2 October 2
May 1 November 0
June 2 December 3
Data show that over the 20 years before 1637 there was an average of 1.5
births per month. You may assume that births in the village historically
follow a Poisson process.
An historian has suggested that the large number of births in July 1637 is
unusual.
(i) Carry out a test of the historian’s suggestion, stating your conclusion. [4]
(ii) Comment on the assumption that births follow a Poisson process. [1]
[Total 5]
(i) Describe the nature of the state space and time space for this process.
[2]
1i
t px the probability that a life is in state i at age x + t , given they are in
state 1 at age x
(iii) Write down equations which could be used to determine the evolution of
1i
t px (for each i ) appropriate for:
(a) x + t < 65
(b) x + t = 65
(c) x + t > 65 . [6]
[Total 10]
(i) If the number of cats currently infected is x , explain why the number of
possible pairings of cats which could result in a new flea infection is
x (10 - x ) . [1]
(ii) Show how the number of infected cats at any time, X (t ) , can be
formulated as a Markov jump process, specifying:
(iii) State the distribution of the holding times of the Markov jump process. [2]
(iv) Calculate the expected time until all the cats have fleas, starting from a
single flea-infected cat. [2]
[Total 9]
1 Alive
m 12
x +t
m14
x +t
m13
x +t
∂ 12 12 11
t px = m x + t t px [5]
∂t
The investigation was carried out separately for each year of age, and the
transition intensities were assumed to be constant within each single year of
age.
(ii) (a) Write down, defining all the terms you use, the likelihood for the
transition intensities.
The investigation produced the following data for persons aged 64 last
birthday:
(iii) (a) Calculate the maximum likelihood estimate (MLE) of the force of
mortality from heart disease at age 64 last birthday.
(iv) Discuss how you might use this model to analyse the impact of risk
factors on the death rate from heart disease and suggest, giving
reasons, a suitable alternative model. [3]
[Total 16]
The company models the arrival of complaints using a Poisson process with
rate 1.25 per working day.
(i) List the assumptions underlying the Poisson process model. [2]
(iii) Define a state space under which the number of outstanding complaints
can be modelled as a Markov jump process. [2]
(iv) Discuss the appropriateness of using the model for this purpose, with
reference to the assumptions being made. [3]
[Total 9]
(i) Draw a transition diagram for the multiple state model. [2]
(iii) Determine integral expressions, in terms of the transition rates and any
expressions previously determined, for:
(a) PHH ( x, x + t )
(b) PHI ( x, x + t )
(i) Sketch a diagram showing the possible transitions between the states.[2]
(ii) Write down the likelihood of the data, assuming transition rates between
states are constant over time, clearly defining all terms you use. [3]
Figures over the first year of the new system based on those who applied for
a provisional licence during that time in one area showed the following:
Person-months in Learner State 1,161
Person-months in Restricted State 1,940
Number of transitions from Learner to Restricted 382
Number of transitions from Restricted to Learner 230
Number of transitions from Restricted to Qualified 110
Number of transitions from Learner to Qualified 217
(iii) (a) Derive the maximum likelihood estimator of the transition rate from
Restricted to Learner.
(b) Estimate the constant transition rate from Restricted to Learner. [3]
[Total 8]
The transition rate for the hazard of the specified event is a constant 0.1.
Whilst policies are eligible for reinstatement, the transition rate for
resumption of cover through paying a reinstatement premium is 0.05.
(ii) (a) Explain why a model with state space {Cover In Force, Suspended,
Lapsed} does not possess the Markov property.
(b) Suggest, giving reasons, additional state(s) such that the expanded
system would possess the Markov property. [3]
(iv) Derive the probability that a policy remains in the Cover In Force state
continuously from time 0 to time t . [2]
(ii) Derive from first principles the Kolmogorov differential equation for first
marriages. [5]
(iii) Write down the likelihood of the data in terms of the waiting times in
each state, the numbers of transitions of each type, and the transition
intensities, assuming the transition intensities are constant. [3]
(iv) Derive the maximum likelihood estimator of the rate of first marriage. [2]
[Total 11]
At a certain airport, taxis for the city centre depart from a single terminus.
The taxis are all of the same make and model, and each can seat four
passengers (not including the driver). The terminus is arranged so that
empty taxis queue in a single line, and passengers must join the front taxi in
the line. As soon as it is full, each taxi departs. A strict environmental law
forbids any taxi from departing unless it is full. Taxis are so numerous that
there is always at least one taxi waiting in line.
(i) Explain how the number of passengers waiting in the front taxi can be
modelled as a Markov jump process. [2]
(iii) Calculate the expected time a passenger arriving at the terminus will
have to wait until his or her taxi departs. [4]
(iv) Write down the transition matrix of the Markov jump chain describing the
number of passengers in the front taxi after the vehicle replacement. [2]
(v) Calculate the expected waiting time for a passenger arriving at the
terminus after the vehicle scrappage scheme and compare this with
your answer to part (iii). [3]
[Total 15]
ij
The disease can be fatal, but most sufferers recover. Let t px be the
probability that a person in state i at age x is in state j at age x + t . Let
m xi j+ t be the transition intensity from state i to state j at age x + t .
∂ 13 11 13 12 23
t px = t p x m x + t + t p x m x + t [4]
∂t
The study revealed that sufferers who contract the disease a second or
subsequent time are more likely to die, and less likely to recover, than first-
time sufferers.
(iii) Draw a diagram showing the states and possible transitions of a model
which allows for this effect yet retains the Markov property. [3]
[Total 9]
(b) Write down the estimated generator matrix of the model. [4]
The generator matrix, with rates expressed per annum, for the process is
estimated as:
A T P D
A Ê -0.15 0.1 0.02 0.03ˆ
T Á 0.45 -0.6 0.1 0.05˜
Á ˜
P Á 0 0 -0.2 0.2 ˜
Á ˜
D Ë 0 0 0 0 ¯
(v) Calculate the expected future duration spent in state P , for a person
currently in state A . [2]
[Total 11]
An investigation was conducted into the effect marriage has on mortality and
a model was constructed with three states: 1 Single, 2 Married and 3 Dead.
It is assumed that transition rates between states are constant.
(i) Sketch a diagram showing the possible transitions between states. [2]
(ii) Write down an expression for the likelihood of the data in terms of
transition rates and waiting times, defining all the terms you use. [3]
The following data were collected from information on males and females in
their thirties.
(iii) Derive the maximum likelihood estimator of the transition rate from
Single to Dead. [4]
(iv) Estimate the constant transition rate from Single to Dead and its
variance. [2]
[Total 11]
(i) Write down the generator matrix of the Markov jump process. [1]
Let t psij be the probability that the process is in state j at time s + t given
∂ ∂
(iv) Write down Kolmogorov’s forward equations for pLL , pLL
∂t t s ∂t t s
∂
and pLH . [2]
∂t t s
(v) Derive an expression for the time after which there is a greater than
50% chance of having experienced a period of high equity price
volatility. [2]
(vi) Solve the Kolmogorov equation to obtain an expression for t p0LL . [4]
[Total 12]
On a small distant planet lives a race of aliens. The aliens can die in one of
two ways, either through illness, or by being sacrificed according to the
ancient custom of the planet. Aliens who die from either cause may, some
time later, become zombies.
(i) Draw a multiple-state diagram with four states illustrating the process by
which aliens die and become zombies, labelling the four states and the
possible transitions between them. [2]
(ii) Write down the likelihood of the process in terms of the transition
intensities, the numbers of events observed and the waiting times in the
relevant states, clearly defining all the terms you use. [4]
(iii) Derive the maximum likelihood estimator of the death rate from illness.
[3]
The aliens take censuses of their population every ten years (where the year
is an ‘alien year’, which is the length of time their planet takes to orbit their
sun). On 1 January in alien year 46,567, there were 3,189 live aliens in the
population. On 1 January in alien year 46,577 there were 2,811 live aliens in
the population. During the intervening ten alien years, a total of 3,690 aliens
died from illness and 2,310 were sacrificed, and the annual death rates from
illness and sacrifice were constant and the same for each alien.
(iv) Estimate the annual death rates from illness and from sacrifice over the
ten alien years between alien years 46,567 and 46,577. [2]
The rate at which aliens who have died from either cause become zombies
is 0.1 per alien year.
(v) Calculate the probabilities that an alien alive in alien year 46,567 will,
ten alien years later:
During a football match, the referee can caution players if they commit an
offence by showing them a yellow card. If a player commits a second
offence which the referee deems worthy of a caution, they are shown a red
card, and are sent off the pitch and take no further part in the match. If the
referee considers a particularly serious offence has been committed, he can
show a red card to a player who has not previously been cautioned, and
send the player off immediately.
The football team manager can also decide to substitute one player for
another at any point in the match so that the substituted player takes no
further part in the match. Due to the risk of a player being sent off, the
manager is more likely to substitute a player who has been shown a yellow
card. Experience shows that players who have been shown a yellow card
play more carefully to try to avoid a second offence.
The rate at which uncautioned players are shown a yellow card is 1/10 per
hour.
The rate at which those players who have already been shown a yellow card
are shown a red card is 1/15 per hour.
The rate at which uncautioned players are shown a red card is 1/40 per
hour.
The rate at which players are substituted is 1/10 per hour if they have not
been shown a yellow card, and 1/5 if they have been shown a yellow card.
(ii) Write down the compact form of the Kolmogorov forward equations,
specifying the generator matrix. [3]
(iii) Solve the Kolmogorov equation for the probability that a player who
starts the match remains in the game for the whole match without being
shown a yellow card or a red card. [2]
(iv) Calculate the probability that a player who starts the match is sent off
during the match without previously having been cautioned. [3]
Consider a match that continued indefinitely rather than ending after 1.5
hours.
(v) (a) Derive the probability that in this instance a player is sent off without
previously having been cautioned.
A certain non-fatal medical condition affects adults. Adults with the condition
suffer frequent episodes of blurred vision. A study was carried out among a
group of adults known to have the condition. The study lasted one year, and
each participant in the study was asked to record the duration of each
episode of blurred vision. All participants remained under observation for
the entire year.
The data from the study were analysed using a two-state Markov model with
states:
Let the transition rate from state i to state j at time x + t be m xij + t , and let
the probability that a person in state i at time x will be in state j at time
x + t be t pxij .
(ii) Derive from first principles the Kolmogorov forward equation for the
transition from state 1 to state 2. [5]
(iv) Estimate the probability that an adult with the condition who is presently
not suffering from blurred vision will be suffering from blurred vision in 3
days’ time. [6]
[Total 13]
Outside an apartment block there is a small car park with three parking
spaces. A prospective purchaser of an apartment in the block is concerned
about how often he would return in his car to find that there was no empty
parking space available. He decides to model the number of parking spaces
free at any time using a time homogeneous Markov jump process where:
the probability that a car will arrive seeking a parking space in a short
interval dt is A.dt o(dt )
for each car which is currently parked, the probability that its owner
drives the car away in a short interval dt is B.dt o(dt )
where A, B 0 .
(i) Specify the state space for the above process. [1]
(iii) Write down the generator matrix for the process. [2]
(iv) Derive the probability that, given all the parking spaces are full, they will
remain full for at least the next two hours. [2]
(vi) Specify the transition matrix for the jump chain associated with this
process. [2]
(vii) Determine the probability that all the spaces become full before any cars
are driven away. [1]
(viii) Derive the probability that the car park becomes full before the car park
becomes empty. [3]
(i) Draw a diagram showing the possible transitions between the four
states. [2]
∂
∂t
( p )=
24
t x
21 14
t px m x + t + t px22 m x24+ t [5]
(iii) Show that there is a statistically significant difference (at the 95%
confidence level) between the death rates from heart disease for
persons with and without the condition. [5]
[Total 12]
In a computer game a player starts with three lives. Events in the game
which cause the player to lose a life occur with a probability m dt + o(dt ) in a
small time interval dt . However, the player can also find extra lives. The
probability of finding an extra life in a small time interval dt is l dt + o(dt ) .
The game ends when a player runs out of lives.
(i) Outline the state space for the process which describes the number of
lives a player has. [1]
(ii) Draw a transition graph for the process, including the relevant transition
rates. [3]
(v) Determine the transition matrix for the jump chain associated with the
process. [2]
(vi) Determine the probability that a game ends without the player finding an
extra life. [1]
[Total 10]
(i) Draw a multiple state diagram illustrating the process, labelling the
states and possible transitions between states. [2]
(ii) Express the likelihood of the process in terms of the transition intensities
and other observable quantities, defining all the terms you use. [4]
(iii) Derive the maximum likelihood estimator of the rate of first time
sickness. [2]
Three years ago medical students visited the island where the disease was
first discovered and found that, of the population of 2,500 people, 860 had
suffered from the disease but recovered. They asked the leaders of the
island to keep records of the occurrence and the outcome of each incidence
of the disease. The students intended to return exactly three years later to
collect the information.
(iv) Derive an expression (in terms of the transition intensities) for the
probability that an islander who has never suffered from the disease will
still be alive in three years’ time. [4]
(v) Set out the information which the students would need when they
returned three years later in order to calculate the rate of sickness from
the disease. [2]
[Total 14]
If a repaired phone breaks again the company, in line with its customer
charter, will not attempt to repair it again, and so discards the phone and
replaces it with a new one.
The company considers the rate at which phones break down to vary
according to whether a phone has previously been repaired as follows:
(ii) Draw a transition diagram for the possible transitions between the
states, including the associated transition rates. [2]
(iv) Solve the equations in part (iii) to obtain PNB (t ) and PR (t ) . [4]
(v) Calculate the probability that a phone has not been discarded by time t .
[1]
[Total 10]
(i) Write down, defining all the terms you use, the likelihood for the
transition intensities. [3]
(ii) Derive the maximum likelihood estimator of the force of mortality from
heart disease for Obese people. [3]
The investigation has followed several thousand people aged 50-59 years
for five years and has the following data:
The doctors want to promote healthy living and therefore wish to claim that
Obese people have a much higher chance, statistically, of dying from heart
disease than do people who are Not obese.
(iii) Test whether this claim is true at the 90% confidence level. [5]
[Total 11]
(i) State the condition needed for a Markov Jump Process to be time
inhomogeneous. [1]
where t is measured in hours from the time the child arrived in the morning,
0 £ t £ 8.
(iii) Calculate the probability that the child is Good for all the time up until
time t . [3]
(iv) Calculate the time by which there is at least a 50% chance of the child
having been Naughty at some point. [2]
(i) Draw a transition diagram for the chosen model, including the transition
rates. [2]
Infected trees display clear symptoms of the disease. This has enabled the
researcher to record the following data on trees in the area of his study:
(iv) Derive the maximum likelihood estimator of the mortality rate caused by
Citrus Greening, t . [3]
(i) Derive the total annual premium charged by the insurance company on
the portfolio. [1]
(ii) Show that the probability that the insurance company has insufficient
assets to pay the next claim made is given by:
È 1 Ê Sˆ˘
1 - exp Í - Á1 - Z ˜¯ ˙
Î 1.5 Ë ˚
[3]
[Total 4]
(i) Describe how transition rates can be estimated under multiple state
models with constant transition rates, including a statement of the data
required. [3]
A specialist insurance policy provides cover only for the theft of valuable
items (such as jewellery) stored in safety deposit boxes in a bank vault. The
premium for cover for an item worth £C is paid in advance. If a claim is
made, the cover ceases.
µ
Active Theft
policy claim
If the item is no longer stored in the safety deposit box (for example, if the
item is sold) then the insurance cover lapses. The transition rate for lapses
of such policies is a constant l.
(iv) Draw a transition diagram for a revised process allowing for lapses. [2]
(v) Derive the revised expected cost of claims incurred by time T . [3]
[Total 11]
Only parts (i)-(iii) of this question are given here. Part (iv) involves exposed
to risk and is included in Booklet 5.
(i) Draw a transition diagram with three states which could be used to
analyse data from this scheme. [2]
(ii) Give the likelihood of the data, defining all the terms you use. [4]
(iii) Derive the maximum likelihood estimator of the rate of falling sick. [3]
[Total 9]
(i) Show from first principles, that if pij ( x, t ) is the probability of being in
state j at time t conditional on being in state i at time x, that:
∂
pHH ( x, t ) = pHH ( x, t ) ( -s (t ) - m (t )) + PHS ( x, t ) r (t , Ct ) [5]
∂t
(ii) Determine the probability that a life is in the Healthy state throughout the
period 0 to t if the life is in the Healthy state at time 0. [2]
P ÈÎ X t = H X s = S,Cs = w ˘˚
y
t - Ú ( r (u,w - s +u ) +u (u,w - s +u ))du
= Úe 0 u ( y , w - s + y ) pHH ( y , t ) dy
0
The solutions presented here are just outline solutions for you to use to
check your answers. See ASET for full solutions.
There are 2 states, offline and online, so the process has a discrete state
space.
Transitions between states are possible at any time. So the time set is
continuous.
¢ (t ) = 0.8PON (t ) - 0.2POFF (t )
POFF
PON (t ) = 1 - POFF (t )
We then move all the POFF (t ) terms over to the LHS so that we can use the
integrating factor method:
¢ (t ) + POFF (t ) = 0.8
POFF
et POFF
¢ (t ) + et POFF (t ) = 0.8 et
We are told in the question that the customer was offline at time 0. So:
POFF (0) = 1
1 = 0.8 + C
ie:
C = 0.2
So:
and:
Then:
and the expected amount of time spent offline over the period [0, t ] is:
t
= È0.8s - 0.2e - s ˘
Î ˚0
= 0.8t - 0.2e - t + 0.2
The expected amount of time spent online over the period [0, t ] is:
( )
t - 0.8t - 0.2e -t + 0.2 = 0.2t + 0.2e - t - 0.2
and the expected proportion of time spent online over the period [0, t ] is:
-t
0.2t + 0.2e - t - 0.2 0.2 t + e - 1
=
( )
t t
(v)(a) Graph
To help you draw the graph, think about the series expansion of e - t :
t2 t3
e -t = 1 - t + + -
2! 3 !
Ê e -t 1ˆ
You should then see that f (0) = 0 and f (t ) = 0.2 Á1 + - ˜ Æ 0.2 as
Ë t t¯
t Æ•.
f(t)
0.25
0.2
0.15
0.1
0.05
2 4 6 8 10 12 14 16 18 20 t
The graph starts at 0 as expected, since we’re assuming that the customer is
offline at time 0. As time goes on the expected proportion of time spent
online increases and tends to 0.2 as t Æ • . Again, this is as expected,
since:
m01 0.2
= = 0.2
m01 + m10 0.2 + 0.8
m 12
x
1: Healthy 2: Sick
m x21
m 13
x m x23
3: Dead
(ii) Derivation
23
t + h px = t px21 h p13 22 23 23 33
x + t + t p x h p x + t + t p x h px + t
33
Since state 3 is the dead state h px + t = 1 . We now assume that, for i π j :
ij
h px + t = h m xij + t + o(h )
So:
23
t + h px (
= t px21 h m13 ) 22 23
( ) 23
x + t + o ( h ) + t px h m x + t + o ( h ) + t px
Rearranging gives:
23
t + h px - t px23 o( h )
= t px21 m13 22 23
x + t + t px m x + t +
h h
Letting h Æ 0 , we get:
∂ 23 21 13 22 23
t px = t px m x + t + t p x m x + t
∂t
(iii) Likelihood
( )e -t ( m )
(m ) (m ) (m ) (m )
- t1 m 12 + m 13 21
+ m 23 n12 n21 n13 n23
12 21 13 23
L = Ce
2
where:
C is a constant
( ) (
ln L = ln C - t1 m 12 + m 13 - t2 m 21 + m 23 )
+ n12 ln m 12 + n21 ln m 21 + n13 ln m 13 + n23 ln m 23
∂ ln L n23
= -t 2 +
∂m 23 m 23
n23
mˆ 23 =
t2
∂ 2 ln L n23
=- < 0 fi max
( ) ( )
2 2
23
∂ m m 23
n23
So the maximum likelihood estimate of m 23 is mˆ 23 = , and the
t2
N23
corresponding maximum likelihood estimator is m 23 = , where N23 is
T2
the random variable number of transitions from sick to dead and T2 is the
random variable waiting time in the sick state.
n23 40 2
mˆ 23 = = = = 0.28571
t2 140 7
mˆ 23 ± 1.96 var m 23 ( )
The variance is estimated by:
2
Ê 2ˆ
(m )
2
ˆ 23 ÁË 7 ˜¯ 1
= = = 0.0020408
n23 40 490
2 1
± 1.96 = (0.19717, 0.37426)
7 490
(i) Test to see if July 1637 had an unusually high number of births
First we can look at the single observation of 5 births in July 1637. This
represents an increment of the process, and so we need to test the following
hypotheses:
H0 : births per month follow a Poisson distribution with parameter 1.5
H1 : births per month do not follow a Poisson distribution with parameter 1.5
Then, according to H0 :
Pr ( X ≥ 5) = 1 - Pr ( X £ 4) = 1 - 0.98142 = 0.01858
This value is sufficiently small (considerably less than 5%) for us to reject
H0 . Therefore, on the basis of this one observation in isolation, we would
conclude that the July 1637 value is atypically large.
However, this ignores the other information that we have, ie that there are 11
other increments all of smaller value than this. So we can see that 8% (one
in twelve) of our observed increments have values of 5 or more, compared
with an expected frequency of 2%, which makes the observation much more
reasonable.
We can also examine whether the data as a whole conform to sampling from
a Poisson distribution with mean 1.5 per month, using a chi-squared
goodness-of-fit test. To do this, we would need to group the data, so that
the expected number in each cell was at least 5, so as to make the test
valid. We obtain the following, according to H0 :
( )
P c 32 ≥ 1.869 = 0.60
( )
and P c 32 ≥ 1.424 = 0.70
As our value is between 1.424 and 1.869, then there is more than a 60%
probability of obtaining data like these according to H0 , so there is very little
evidence in support of rejecting this hypothesis. Therefore, in the context of
the year’s experience as a whole, the July 1637 observation does seem
perfectly reasonable.
One requirement for a Poisson process is that no more than one event can
occur at a single moment in time: however, multiple births do occur, which
invalidates this assumption.
There may be seasonal reasons for this (ie differences according to month
of year, which may be largely related to climate), and also there may be
different birth rates by calendar year (eg as a result of significant
demographic changes caused by epidemics, food shortages, war, etc, or by
progressive changes over time, such as from medical improvements to
survival rates feeding back into lower fertility rates, etc).
A larger sample than the one here would be necessary in order to test these
possibilities.
Transitions occur at any point in continuous time both before and after
age 65.
However, at exact age 65, all lives in any state other than dead transfer into
the normal retired state. This occurs at a fixed point in time, and therefore is
a discrete time transition.
3 2 4
It is not clear from the question whether persons in State 2 will be allowed to
start an ill-health pension before age 65 should they become sick, as we are
not informed of the scheme rules in this regard. We have assumed that this
is possible, and so transition from State 2 to State 3 is included in the above
diagram and is allowed for as appropriate in all the formulae that follow.
(a) x + t < 65
∂ 11
∂t
11 12
( 13 15
t px = - t px m x + t + m x + t + m x + t )
∂ 12
∂t
11 12 12
(
23 25
t px = t px m x + t - t px m x + t + m x + t )
∂ 13 11 13 12 23 13 35
t px = t p x m x + t + t px m x + t - t p x m x + t
∂t
14
t px =0
∂ 15 11 15 12 25 13 35
t px = t p x m x + t + t px m x + t + t p x m x + t
∂t
(b) x t 65
11
t px = t p12
x =0
∂ 13 13 35
t p x = - t px m x + t
∂t
14
t px = t- p11 12
x + t - px (x + t - indicates the instant
immediately preceding age 65)
∂ 15 13 35 14 45
t p x = t p x m x + t + t px m x + t
∂t
∂ 13 13 35
t px = - t px m x + t
∂t
∂ 14 14 45
t px = - t px m x + t
∂t
∂ 15 13 35 14 45
t px = t px m x + t + t p x m x + t
∂t
For the possibility of a new infection to arise, an infected cat would have to
Ê xˆ
meet with an uninfected cat. There are Á ˜ = x ways of choosing one cat
Ë 1¯
Ê10 - x ˆ
out of the x infected cats and Á = 10 - x ways of choosing one
Ë 1 ˜¯
uninfected cat out of the 10 - x uninfected cats. So the number of possible
pairings that could result in a new infection is x (10 - x ) .
(ii) How the number of infected cats can be formulated as a Markov jump
process
The state space is {1, 2,...,10} , assuming that we have to start with at least
one infected cat. So the state space is discrete.
d
P ( t ) = P (t ) A
dt
Ï i (10 - i )
Ô m
90 for j = i + 1, i = 1, 2,..., 9
Ô
Ô i (10 - i )
mij = Ì - m
90 for j = i , i = 1, 2,..., 9
Ô
Ô 0 otherwise
Ô
Ó
d
P (t ) = A P (t )
dt
Let Ti denote the holding time in State i , i = 1, 2,..., 9 . Then the expected
time until all the cats have fleas, starting from a single infected cat is:
90 Ê 1 1 1 1 1 1 1 1 1ˆ
= + + + + + + + +
m ÁË 9 16 21 24 25 24 21 16 9 ˜¯
50.92
=
m
12
t + h px = t p11 12 12 22
x h px + t + t px h p x + t
22
h px + t =1
So we have:
12
t + h px = t p11 12 12
x h px + t + t px
12
h px + t = h m 12
x + t + o( h )
So:
12
t + h px = t p11 12 12
x h m x + t + t px + o ( h )
12
t + h px - t p12 o( h )
x
= t p11 12
x mx +t +
h h
∂ 12 11 12
t px = t px m x + t
∂t
o( h )
since lim =0.
h Æ0 h
( )
(m ) (m ) (m )
- t1 m 12 + m 13 + m 14 n12 n13 n14
12 13 14
L = Ce
where:
C is a constant
t1 is the total observed waiting time in State 1 for all the lives in the
investigation
( )
ln L = ln C - t1 m 12 + m13 + m14 + n12 ln m12 + n13 ln m13 + n14 ln m14
∂ n12
ln L = -t1 +
∂m 12 m12
n12
m12 =
t1
n12
So the maximum likelihood estimate of m12 is mˆ 12 = and the
t1
N12
corresponding maximum likelihood estimator is , where N12 is m 12 =
T1
the random variable representing the number of transitions from State 1 to
State 2 and T1 is the random variable representing the waiting time in
State 1.
The maximum likelihood estimate of the force of mortality from heart disease
at age 64 last birthday is:
34
= 0.031925
1, 065
mˆ 12 34
= = 2.9976 ¥ 10 -5
t1 1, 0652
( mˆ )
2
12
=
(34 / 1,065)2 =
34
n12 34 1,0652
mˆ 12 ± 1.96 var m 12 =( ) 34
1, 065
± 1.96
34
1, 0652
= (0.0212, 0.0427)
To analyse the impact of different risk factors on the death rate from heart
disease, we would need to subdivide our observed lives into homogeneous
groups, ie groups that have the same characteristics. We would then
estimate the transition rates for each group separately. We could compare
the estimated death rate from heart disease for the different groups to
assess the impact of risk factors on this mortality rate.
1. N (0) = 0
- l (t - s ) n
e È l (t - s )˘
P ÎÈN (t ) - N (s ) = n ˚˘ = Î ˚
n!
5 ¥ 1.25 = 6.25
Type of
Straightforward Medium Complicated
complaint
Probability 0.6 0.3 0.1
Expected
1
response time 1 4
(in person-days) 2
1
0.6 ¥ + 0.3 ¥ 1 + 0.1 ¥ 4 = 1 person-day
2
If the company wants to make sure that it responds to each complaint within
a certain number of days, the state space of the model needs to include a
duration component, which measures the time since each complaint arrived.
The model in part (iii) does not include this feature and so it is not
appropriate for this task.
(x)
Healthy Infected
∂
P (s, t ) = P (s, t ) A (t )
∂t
where P (s, t ) is the matrix of transition probabilities over the interval (s, t )
and:
Ê - (s (t ) + m (t )) s (t ) 0 m (t ) ˆ
Á 0 -u ( t ) r ( t ) u ( t ) - r (t )˜˜
A (t ) = Á
Á 0 0 0 0 ˜
Á ˜
Ë 0 0 0 0 ¯
A(t ) is the generator matrix at time t assuming that the states are in the
order given in the question, ie Healthy, Infected, Dead (from disease), Dead
(not from disease).
Î 0
t
( )
PHH ( x, x + t ) = exp ÈÍ - Ú s ( x + s ) + m ( x + s ) ds ˘˙
˚
where:
PHD(from disease) ( x, x + t )
where:
Working from the description given, we see that the model looks like this:
L R
v
Here ‘L’ denotes a learner, ‘R’ denotes a restricted driver and ‘Q’ denotes a
qualified driver.
(ii) Likelihood
By analogy with the HSD model, (with the H, S and D states replaced with
L, R and Q), we can label the transition rates using the usual Greek letters.
The likelihood function is then:
where:
Ti is the total number of person-months spent in state i ( i = L or R )
nij is the total number of transitions from state i to state j for the
permitted transitions
K is a constant of proportionality.
∂ n
log L = -TR + RL = 0
∂r r
nRL
rˆ =
TR
We can confirm that this does give a maximum value (as a function of the
parameter r ) by showing that the second derivative is negative:
∂2 nRL
2
log L = - < 0 (provided that nRL > 0 )
∂r r2
The estimate of the transition rate from Restricted to Learner is the MLE
of r , which equals:
230
rˆ = = 0.1186
1, 940
(i) Time-homogeneous?
The transition rates for this process take constant values. So, in general, the
probability that the process will be in state j at time t , given that it is in
state i at time s , depends only on the length of the time interval t - s .
So, for calculations that do not involve the suspended state, for example
calculating the expected time until the first catastrophic event, it would be
appropriate to model the situation using a time-homogeneous approach.
However, for calculations where the suspended state may be involved, we
need to allow for the sudden change in the rate.
Additional clarification:
To illustrate the point, suppose we let pSS (s, t ) denote the probability that
the policy remains in the suspended state until time t , given that it was in
the suspended state at time s .
Now consider the probabilities pSS (0.25,0.75) and pSS (0.75,1.25) , which
both have t - s = 0.5 , so they are both probabilities of remaining suspended
for another 6 months.
Because 0.75 < 1 , for the first probability, the one-year limit cannot have
taken effect yet, so pSS (0.25,0.75) would equal e -0.05 ¥ 0.5 , ie e -0.05(t - s ) .
e -0.05(t - s ) , provided the policy became suspended after time 0.25 (and so
we haven’t hit the one-year limit yet), but it would equal zero if it became
suspended before time 0.25 (because time 1.25 is more than one year later).
So, even when we know the values of s and t , the second probability is not
fully defined without knowing when the policy became suspended (or
equivalently, the duration since suspension occurred).
The Examiners’ Report gives this numerical illustration and concludes that
the model is time-inhomogeneous because some probabilities do not just
depend on the value of t - s .
Note, however, that the question just asks which approach would be more
appropriate for modelling the situation, which makes the question more
open-ended. So, provided you argued your case satisfactorily, you should
have picked up some marks whether you said it was time-homogeneous or
time-inhomogeneous.
This model would not be Markov because risks that are in the Cover In
Force state include both risks that have not experienced a catastrophic
event and those that have, but were subsequently reinstated. The future
probabilities for these two types are not the same, since the first type still has
the option for future reinstatement, whereas the second type does not.
To preserve the Markov property, we need to split the Cover In Force state
into the two different types:
If we label the suspended state as S and the lapsed state as L, the model
looks like this:
(up to duration 1)
C R
0.05
0.1
0.1
S L
This situation will arise if a catastrophic event occurred at some point during
the time period (t - 1, t ) and the risk has not been reinstated by time t .
Event
CCCCCCCCCCCCCCCCCCCCCCCCCC SSSSSSSSSSS
0 t-1 s t
Suppose the catastrophic event occurred during the short time interval
(s, s + h ) , where t - 1 £ s < t .
The sequence of events required to create this situation must have been:
the risk remained in state C continuously until time s
t
Út -1 pCC (s )
¥ 0.1
ds ¥ pSS (t - s )
Jump to state S
Remain in state C during the short period ( s,s + h ) Remain in state S
until time s until time t
t t
0.1Ú e -0.1s e -0.05( t - s )ds = 0.1e -0.05t Ú e -0.05s ds
t -1 t -1
t
-0.05t
È e -0.05s ˘
= 0.1e Í- ˙
ÎÍ 0.05 ˚˙ t -1
(i) Diagram
Widowed
Divorced
Widowed
Divorced
N = never married
M = married
W = widowed
D = divorced
ij
t px = probability that a woman is in state j at age x + t given that she
is in state i at age x .
ij
for i π j and small h , h px + t = h m xij + t + o(h )
NM
t + h px = t pxNN h pxNM NM MM NW
+ t + t px h px + t + t px
WM ND DM
h px + t + t px h p x + t
= t pxNN h pxNM NM MM
+ t + t p x h px + t
= t pxNN h pxNM NM
+ t + t px 1 - h pxMW ( MD
+ t - h px + t )
= t pxNN h m xNM NM
+ t + t px (1 - h m MW
x +t - h m xMD
+t ) + o( h )
Rearranging gives:
( )
NM
t + h px - t pxNM o( h )
= t pxNN m xNM NM
+ t - t px m xMW MD
+t + mx +t +
h h
o( h )
Since Æ 0 as h Æ 0 , we have:
h
∂
∂t
NM
t px = t pxNN m xNM NM
+ t - t px m xMW MD
+t + mx +t ( )
(iii) Likelihood function
NM MW
+ m MD ) - tW mWM - tD m DM
L = Ce - tN m e - tM ( m e e
( ) (m ) (m ) (m ) (m )
nNM nMW nMD nWM nDM
¥ m NM MW MD WM DM
where:
C is a constant
ln L = ln C - tN m NM - tM ( m MW + m MD ) - tW m WM - tD m DM
+ nNM ln m NM + nMW ln m MW + nMD ln m MD + nWM ln mWM
+ nDM ln m DM
nNM
m NM =
tN
∂ 2 ln L nNM
=- < 0 fi max
( ) ( )
2 2
NM
∂ m m NM
nNM
So, the MLE is mˆ NM = and the corresponding maximum likelihood
tN
NNM
estimator is m NM = , where NNM and TN are random variables.
TN
Alternatively, we could say that the Markov property holds because arrivals
occur according to a Poisson process and Poisson processes are Markov
because they have independent increments.
0 1
3 2
Ê -b b 0 0 ˆ
Á 0 -b b 0 ˜
A=Á ˜
Á 0 0 -b b ˜
Á ˜
Ë b 0 0 -b ¯
d
P (t ) = P (t ) A
dt
where:
pij (t ) is the probability of going from State i to State j in a time period t, and
A is the generator matrix from part (ii)(a).
d
pij (t ) = pi , j -1(t ) b - pij (t )b for i = 0,1, 2, 3 and j = 1,2,3
dt
d
and pi 0 (t ) = pi 3 (t ) b - pi 0 (t )b for i = 0,1, 2, 3
dt
Let:
( )
E Tj =
1
b
.
Then the total expected waiting time for a new passenger is:
p 0 m1 + p 1 m2 + p 2 m3
where:
1
pj = = probability there are j passengers waiting in the front taxi
4
Now:
3
m1 = E (T1) + E (T2 ) + E (T3 ) =
b
2
m2 = E (T2 ) + E (T3 ) =
b
1
m3 = E (T3 ) =
b
So, the expected waiting time for a passenger arriving at the terminus is:
1Ê 3 2 1ˆ 6 3
+ + = = minutes
4 ÁË b b b ˜¯ 4 b 2b
The transition diagram for the number of passengers waiting in the front taxi
now looks like this:
0 1
0.5
4 2
0.5
3
The jump chain is the sequence of states that the process occupies. Its
transition matrix is:
Ê0 1 0 0 0ˆ
Á0 0 1 0 0˜
Á ˜
Á½ 0 0 ½ 0˜
Á ˜
Á0 0 0 0 1˜
ÁË 1 0 0 0 0˜¯
If the front taxi is a 3-seater, the total expected waiting time is:
p 0 m1 + p 1 m2
where:
1 2 1
p 0 = p1 = p 2 = , m1 = , m2 =
3 b b
1Ê 2 1ˆ 3 1
So the expected waiting time = Á + ˜= = minutes.
3 Ë b b ¯ 3b b
If the front taxi is a 5-seater, the total expected waiting time is:
p 0 m1 + p 1 m2 + p 2 m3 + p 3 m4
where:
1
p 0 = p1 = p 2 = p 3 = p 4 =
5
4 3 2 1
m1 = , m2 = , m3 = , m4 =
b b b b
1 Ê 4 3 2 1 ˆ 10 2
Á + + + ˜= = minutes
5 Ë b b b b ¯ 5b b
We know that:
The arrival rate is the same for 3-passenger and 5-passenger taxis.
3
The expected waiting time at the front of queue is for a 3-passenger
b
taxi.
5
The expected waiting time at the front of queue is for a 5-passenger
b
taxi.
So, the probability that the front taxi takes 3 passengers when a new
3
person arrives at the terminus is and the probability it takes 5
8
5
passengers is .
8
So the total expected waiting time for a passenger arriving at the terminus is:
1 3 2 5 13
E= ¥ + ¥ = minutes
b 8 b 8 8b
1Ê2 1 4 3 2 1 ˆ 13
Á + + 0 + + + + + 0˜ =
8Ëb b b b b b ¯ 8b
as before.
This is longer than the expected waiting time before the scrappage scheme
3
( minutes). So, based on this evidence alone, the service has
2b
deteriorated following the scrappage scheme.
1=H 2=S
3=D
In this diagram:
‘H’ denotes a healthy person not suffering from the disease
‘S’ denotes a sick person suffering from the disease
‘D’ denotes a dead person.
13
∂ 13 t + h px - t p13
x
t px = lim
∂t h Æ0 h
13
Because this is a Markov process, we can ‘split’ the probability t + h px at
time t using the Chapman-Kolmogorov equation:
13
t + h px = t p11 13 12 23 13 33
x ¥ h p x + t + t px ¥ h px + t + t px ¥ h p x + t
ij
h px + t = h m xij + t + o(h ) , i π j
So the three probabilities on the RHS of this equation that relate to the short
time period of length h are:
13
h px + t = h m 13
x + t + o( h )
23
h px + t = h m x23+ t + o(h )
33
h px + t =1
∂ 13
t px
∂t
13
t + h px - t p13
x
= lim
h Æ0 h
= lim
(p 11 13
t x ¥ h px + t + t p12 23 13 33
) 13
x ¥ h p x + t + t px ¥ h px + t - t p x
h Æ0 h
= lim
11
( 13
) 12
(
23
) 13 13
t px ¥ h m x + t + o( h ) + t px ¥ h m x + t + o( h ) + t p x ¥ 1 - t p x
h Æ0 h
11 13 12 23
∂ 13 t px ¥ h m x + t + t px ¥ h m x + t + o( h )
t px = lim
∂t h Æ0 h
11 13 12 23 o( h )
= t px m x + t + t px m x + t + lim
h Æ0 h
= t p11 13 12 23
x m x + t + t px m x + t
(iii) Diagram
We now know that individuals who are suffering from the disease have
different transition rates, depending on whether they have had the disease
before or not. So to preserve the Markov property, we need to have
separate states for those who are suffering from the disease for the first time
and those who have had it before. This also means that we need to split out
the ‘healthy’ state too, so that we can distinguish between those who have
never had the disease and those who have had it before.
3=D
In this diagram the numbers in the names of the states indicate how many
times the person has had the disease up to that point. So:
‘H0’ denotes a healthy person who has never suffered from the disease
‘S1’ denotes a sick person suffering from the disease for the first time
‘H1+’ denotes a person who is currently healthy but has suffered from
the disease at least once
‘S2+’ denotes a sick person suffering from the disease for the second or
subsequent time
‘D’ denotes a dead person.
The parameters of the model are the transition rates mij between the states,
pij ( h )
where mij = lim , iπ j.
hÆ0 h
So here there are six parameters: m AB , m AC , mBA , mBC , mCA and mCB .
The model assumes that the transition rates have constant values.
The model is Markov, so that the transition rates depend only on the current
state.
nij
The MLE of the transition rate from state i to state j is mˆij = , where nij
ti
is the number of transitions from state i to state j , and ti is the total
waiting time in state i .
So here the parameter estimates (which are rates ‘per hour’) are:
60 12
mˆ AB = = = 1.714
35 7
45 9
mˆ AC = = = 1.286
35 7
50 1
mˆBA = = = 0.333
150 3
25 1
mˆBC = = = 0.167
150 6
55 11
mˆCA = = = 0.262
210 42
15 1
mˆCB = = = 0.071
210 14
A B C
A B C
È 12 9˘
A Í -3 7 7˙ A È -3 1.714 1.286 ˘
or Í ˙
B Í 31 - 21 1˙ B Í 0.333 -0.5 0.167 ˙
Í 6˙
C Í 11 ÍÎ0.262 0.071 -0.333 ˙˚
1
- 31 ˙ C
Î 42 14 ˚
mij
If the current state is i , the jump probability to state j is pij = ,
li
where li is the sum of the transition rates leaving state i . So, if there
are ni transitions out of state i , the number of transitions to state j will
mˆ
(
have a Binomial ni , pˆ ij ) distribution, where pˆij = lˆij .
i
A
0.45
0.1 0.02
0.03
T P
0.1
0.2
0.05
D
F ( A) = 2 F (T ) + 1
3 5
0.45 0.1
F (T ) = 0.6
F ( A) + F (P ) + 0.05
0.6 0.6
F
(D )
=0 =1
3 1
F (T ) = 4
F ( A) + 12
F ( A) = 2 F (T ) + 1
3 5
= 2
3 { 3
4
1 +
F ( A) + 12 } 1
5
F ( A) = 1 1 +
F ( A) + 18 1
2 5
1 23
fi 2
F ( A) = 90
23
fi F ( A) = 45
= 0.5111
3 1 3 23 1 7
F (T ) = 4
F ( A) + 12 = 4
¥ 45
+ 12 = 15
= 0.4667
Looking at the diagram, we can see that a person in state A can visit
state P at most once (since, once they left state P they would have to go to
the absorbing state D ).
If they do visit state P the time spent there will have an exponential
distribution with parameter lP = 0.2 , so that the mean waiting time in
1 1
state P will be = =5.
lP 0.2
22 ¥5 = 22 = 2.444 years
45 9
(i) Diagram
12
1 (Single) 2 (Married)
21
13 23
3 (Dead)
(ii) Likelihood
L μ e -( m12 + m13 )t1 e -( m21 + m23 )t2 m12n12 m13n13 m21n21 m23n23
where:
mij is the transition rate from state i to state j
∂ n
log L = -t1 + 13
∂m13 m13
Setting this equal to zero to find the maximum likelihood estimator gives:
n13 n13
-t1 + = 0 fi m13 =
m13 t1
We can check that this will be a maximum by considering the sign of the
second derivative:
∂2 n13
log L = - <0
2
∂m13 m132
Substituting the observed values of n13 and t1 given in the question, we get
the following estimate of m13 :
12
mˆ13 = = 0.001165
10,298
∂2 Ê n ˆ m 2
CRLB = -1 log L = -1 Á - 13 ˜ = 13
2 2 n13
∂m13 Ë m13 ¯
0.0011652
= 1.132 ¥ 10-7 or (0.000336)2
12
Ê -m mˆ
A=Á
Ë r - r ˜¯
TL ~ Exp
and TH ~ Exp
We would record the total time spent in each state by the n equities over
the investigation, tL and tH .
We would also record the number of transitions from one state to the other
over the investigation, nLH and nHL .
nLH nHL
m= and r=
tL tH
∂ LL LH LL
t ps = t ps r - t ps m
∂t
∂ LH LL LH
t ps = t ps m - t ps r
∂t
∂ LL LL
t ps = - t ps m
∂t
(v) Expression for the time after which there is a greater than 50% chance
of having experienced a period of high equity price volatility
We know from part (ii) that TL ~ Exp . So, we require the minimum
time t for which the following is true:
e - m t < 0.5
¤ ln 0.5 > - m t
- ln 0.5 ln 2
¤ t> =
m m
ln 2
So, the required time is .
m
∂ LL LH LL
t p0 = t p0 r - t p0 m
∂t
LH
t p0 = 1 - t p0LL
∂
∂t
LL
( LL LL
t p0 = 1 - t p0 r - t p0 m)
= r - ( r + m ) t p0LL
or equivalently:
∂
t p0 + ( r + m ) t p0 = r
LL LL
∂t
È∂ LL ˘ ( r + m )t
+ ( r + m ) t p0LLe( = r e(
r + m )t r + m )t
Í ∂t t p0 ˙ e
Î ˚
∂ Ê LL ( r + m )t
∂t Ë t
p0 e = r e( )
r + m )t
)
n n
∂ Ê
e( dt = Ú r e(
LL r + m )t r + m )t
¤ Ú ∂t Ë t p0 dt
0 0
n n
r È ( r + m )t ˘
¤ ÈÍ t p0LLe(
r + m )t ˘
= e
Î ˚˙0 r + m ÎÍ ˚˙0
¤ LL ( r + m )n
n p0 e -1=
r
r+mË
Êe( r + m )n - 1
)
Substituting t for n and rearranging:
LL
t p0 -e (
- r + m )t
=
r
r+mË
Ê1 - e - ( r + m )t
)
¤ LL
t p0 =
r Ê
r+mË
1- e (
- r + m )t
+e (
- r + m )t
)
me (
- r + m )t
r
= +
r+m r+m
A D
I Z
- t A ( m AI + m AD )
L μe m AI nAI m AD nAD e -tI mIZ mIZ nIZ e -tD mDZ mDZ nDZ
ln L = -t A ( m AI + m AD ) + nAI ln m AI + nAD ln m AD
-tI mIZ + nIZ ln mIZ - tD mDZ + nDZ ln mDZ + c
∂ ln L n
= -t A + AI
∂m AI m AI
nAI
m AI =
tA
Differentiating again:
∂ 2 ln L nAI
=- <0 fi max
∂ ( m AI ) ( mAI )2
2
nAI
So the maximum likelihood estimate of m AI is mˆ AI = and the
tA
N AI
corresponding maximum likelihood estimator is m AI = (where the upper
TA
case N and T are random variables).
Assuming that the population varies linearly between census dates, we can
estimate the total time spent in state A over one year:
1
tA =
2
(3,189 + 2,811) = 3,000
nAI 369
mˆ AI = = = 0.123
tA 3,000
nAD 231
mˆ AD = = = 0.077
tA 3,000
(v)(a) Probability an alien alive in year 46,567 will still be alive in 46,577
We have:
-10( mˆ AI + mˆ AD )
pAA (10) = p AA(10) = e
-10(0.123 + 0.077)
=e
= e -2 = 0.135335
(v)(b) Probability an alien alive in year 46,567 will be dead but not a
zombie in 46,577
We can construct an integral equation for the probability that a life goes from
state A to state Z over 10 years:
10
pAZ (10) = Ú { }
pAA (t ) m AI pIZ (10 - t ) + m AD pDZ (10 - t ) dt
0
The rate at which aliens who have died from either cause become zombies
is 0.1 per alien year. This means that pIZ (t ) = pDZ (t ) and we have:
10
pAZ (10) = Ú 0.2e
-0.2t
pIZ (10 - t ) dt
0
Now, the probability of going from state I to state Z and the probability of
going from state I to state I together sum to 1. So:
-0.1(10 - t )
pIZ (10 - t ) = 1 - pII (10 - t ) = 1 - pII (10 - t ) = 1 - e
)
10
pAZ (10) = -0.2t Ê1 - e -0.1(10 - t ) dt
Ú 0.2e Ë
0
10 10
= 0.2 Ú e -0.2t dt - 0.2e -1 Ú e -0.1t dt
0 0
10 10
= È -e -0.2t ˘ + 2e -1 Èe -0.1t ˘
Î ˚0 Î ˚0
(
= 1 - e -2 - 2e -1 1 - e -1 )
= 0.39958
1
40
1 1
10 15
All OK Yellow Sent off
1
1 5
10
Substituted
P ¢(t ) = P (t ) A
The generator matrix for this model (with the states in the order ‘All OK’,
‘Yellow’, ‘Sent off’, ‘Substituted’) is:
OK È- 9 1 1 1 ˘
Í 40 10 40 10 ˙
Yellow Í 0 4
- 15 1 1 ˙
Í 15 5 ˙
Off Í 0 0 0 0˙
Sub Í ˙
ÍÎ 0 0 0 0 ˙˚
9
¢ (t ) = -
P11 P (t )
40 11
¢ (t )
P11 9
=-
P11(t ) 40
d 9
ie log P11(t ) = -
dt 40
9
log P11(t ) = - 40
t +c
When t = 0 , we have:
9
log P11(0) = - ¥0+c fic =0
40
=1
So:
9
log P11(t ) = - 40
t
and:
P11(t ) = exp - ( 9
40
t )
So the required probability is:
P11(1.5) = exp - ( 9
40 )
¥ 1.5 = 0.7136
1.5
Ú0 P11(s ) m13 ds
Using the formulae from part (iii), we can evaluate this as:
Ú0
1.5
P11(s ) m13 ds = Ú0
1.5
exp - ( 9
40 )
s ¥ 1
40
ds
( )
1.5
= 1 È- 40
exp - 9
s ˘˙
40 ÎÍ 9 40 ˚0
= 1
9
È1 - exp -
ÎÍ ( )˘˚˙ = 0.0318
27
80
Ú0
•
P11(s ) m13 ds = Ú0
•
exp - ( 9
40
s ¥ ) 1
40
ds
( )
•
= 1 È- 40
exp - 9
s ˘˙
40 ÍÎ 9 40 ˚0
1 1
= 9
ÈÎ1 - 0 ˘˚ = 9
This answer makes sense because there are three modes of exit from
state 1 (playing without having received a caution), namely: being sent off,
getting a yellow card or being substituted. Based on the rates of transition,
the probabilities of each of these occurring are in the ratio 40 1 : 1 : 1
10 10
or 1: 4 : 4 or 1 : 4 : 4 .
9 9 9
The Kolmogorov forward differential equation for the transition from state 1
∂
to state 2 is an equation for p12 . This can be defined as:
∂t t x
12
∂ 12 t + h px - t p12
x
t px = lim (1)
∂t hÆ0 h
12
Since this is a Markov process, we can ‘split up’ the probability t + h px using
the Chapman-Kolmogorov equation to get:
12
t + h px = t p11 12 12 22
x h px + t + t px h p x + t
12
h px + t = h m 12
x + t + o( h )
We also have:
22
h px + t ( )
= 1 - h px21+ t = 1 - h m x21+ t + o(h ) = 1 - h m x21+ t + o(h )
11 12 12 22 12
∂ 12 t p x h p x + t + t px h px + t - t p x
t px = lim
∂t h Æ0 h
= lim
11
t px (h m 12
x +t )
+ o(h ) + t p12 ( 21
) 12
x 1 - h m x + t + o( h ) - t px
h Æ0 h
∂
p12
= lim
11 12
( )
12 21
(
t px h m x + t + o( h ) + t px - h m x + t + o( h ) )
t x
∂t h Æ0 h
= lim
(p11 12
t x mx +t - t p12 21
)
x m x + t h + o( h )
h Æ0 h
o( h )
Since lim = 0 and the remaining h ’s cancel, we get:
h Æ0 h
∂ 12 11 12 12 21
t px = t px m x + t - t px m x + t
∂t
The probability that a person who is not currently suffering from the condition
12
will have blurred vision in exactly 3 days’ time is 3 px .
If we assume that the transition rates are constant and equal to the values
estimated in part (iii), we have:
∂ 12 11 12
t px = 0.2 t px - 0.8 t px
∂t
Since t p11 12
x + t px = 1 , this can be written as:
∂
∂t
12
( 12
)12 12
t px = 0.2 1 - t px - 0.8 t px = 0.2 - t px
∂ 12 12
t px + t px = 0.2
∂t
∂
et 12 t 12
t px + e t px = 0.2e
t
∂t
(
∂ t 12
∂t
)
e t px = 0.2et
et t p12 t
x = 0.2e + c
12
We can use the initial condition 0 px = 0 to find the constant c :
e0 0 p12 0
x = 0.2e + c fi c = -0.2
=0
So:
et t p12 t
x = 0.2e - 0.2
and:
12
t px (
= 0.2 1 - e - t )
So the required probability is:
12
3 px ( )
= 0.2 1 - e -3 = 0.1900
3B 2B B
0 1 2 3
A A A
(full) (empty)
0 1 2 3
0 3B 3B 0 0
1 A ( A 2B ) 2B 0
2 0 A ( A B) B
3 0 0 A A
(iv) Probability the car park will remain full for two hours
The probability that the car park will remain full for the next two hours is:
If a Markov jump process is examined only at the times of the transitions, the
resulting process is the jump chain associated with the jump process.
In other words, the jump chain is the sequence of states that the process is
observed to take. The time set comprises the times at which the transitions
occur.
0 1 2 3
0 0 1 0 0
A 2B
1 0 0
A 2B A 2B
A B
2 0 0
AB AB
0 0 1 0
3
(vii) Probability that all spaces become full before any cars drive away
The probability that all the spaces become full before any cars are driven
away is:
A A A2
A B A 2B ( A B )( A 2B )
(viii) Probability that the car park becomes full before it becomes empty
Let pi be the probability that the car park becomes full before it becomes
empty, given that there are currently i empty spaces (ie the process is in
state i ).
We have:
Ê A ˆ Ê A ˆ Ê A ˆ Ê 2B ˆ
p2 = Á + p
Ë A + B ˜¯ ÁË A + 2B ˜¯ ÁË A + B ˜¯ ÁË A + 2B ˜¯ 2
go from go from1 go from go from
2 to 1 to 0 (full) 2 to 1 1 to 2
Rearranging:
È 2 AB ˘ A2
p2 Í1 - ˙=
Î ( A + B )( A + 2B ) ˚ ( A + B )( A + 2B )
p2 ÎÈ( A + B )( A + 2B ) - 2 AB ˚˘ = A2
So:
A2
p2 =
( A + B )( A + 2B ) - 2 AB
A2
=
A + 3 AB + 2B 2 - 2 AB
2
A2
=
A2 + AB + 2B 2
(ix) Comment
The rates of arrival and departure will not be constant during the day. For
example, there is likely to be very little activity at night.
The events will not be random because the car owners will probably have
regular times at which they go to work and come home.
The process may not be Markov either if the car owners time their
movements based on past experience of when they think there is likely to be
a space or if they park nearby when it is full and watch out of their windows
until one of the cars leaves.
The model also assumes that vehicles are always parked in a single space,
and never over two spaces.
(i) Diagram
1 2
3 4
24
∂ 24 t + h px - t px24
t px = lim (1)
∂t h Æ0 h
24
t + h px = t px21 h p14 22 24 23 34 24 44
x +t + t p x h px +t + t px h p x +t + t px h p x +t
14
h p x +t = h m14
x +t + o(h ) and
24
h p x +t = h m x24+t + o(h)
We can see from the diagram in part (i) that transitions from state 3 to
34
state 4 are not possible. So h p x +t = 0.
We can also see from the diagram that state 4 is an absorbing state.
44
So h p x +t = 1.
21 14 22 24 23 24 24
∂ 24 t px h p x + t + t px h p x + t + t px ¥ 0 + t px ¥ 1 - t px
t px = lim
∂t h Æ0 h
= lim
21
t px (h m 14
x +t ) ( )
+ o(h ) + t px22 h m x24+ t + o(h )
h Æ0 h
o(h )
Since lim = 0 and the remaining h ’s cancel, we get:
hÆ0 h
∂
p 24 = t px21m 14 22 24
x + t + t px m x + t
∂t t x
nij
mˆij = , where nij is the number of transitions from state i to state j
ti
and ti is the waiting time in state i .
Ê m13 ˆ Ê m ˆ
m13 ~ N Á m13 , and m 23 ~ N Á m23 , 23 ˜
Ë t1 ˜¯ Ë t2 ¯
Ê m23 m13 ˆ
m 23 - m13 ~ N Á m23 - m13 , +
Ë t2 t1 ¯˜
Ê m23 m13 ˆ
m 23 - m13 ~ N Á 0, +
Ë t2 t1 ¯˜
m 23 - m13
Z= ~ N (0,1)
m23 m13
+
t2 t1
n13 10
mˆ13 = = = 0.004888
t1 2,046
and:
n23 25
mˆ 23 = = = 0.021949
t2 1,139
0.021949 - 0.004888
z= = 3.67
0.021949 0.004888
+
1,139 2,046
As this far exceeds 1.96, the 97.5th percentile of the N (0,1) distribution, we
can confidently reject H0 and conclude that there is a difference between
the death rates from heart disease for those with and without the condition.
(ii) Diagram
0 1 2 3 4
With the states in the order 0, 1, 2, …, the generator matrix for this process
is:
È0 0 0 0 ˘
Í ˙
Ím -( l + m ) l 0 ˙
Í0 m -( l + m ) l ˙
Í ˙
Í0 0 m -( l + m ) ˙
Í ˙˚
Î
A jump chain is a Markov chain derived from a Markov jump process where
we model the successive states the process can visit and the associated
probabilities.
È 1 0 0 0 ˘
Í ˙
Í m 0
l
0 ˙
Íl + m l+m ˙
Í ˙
Í 0 m l
0 ˙
Í l+m l+m ˙
Í ˙
Í 0 m
Í 0 0 ˙˙
l+m
Í ˙
ÎÍ ˚˙
(vi) Probability
The probability that a game ends without the player finding an extra life is
the probability of the jump chain following the path 3 Æ 2 Æ 1 Æ 0 . This is:
3
m m m Ê m ˆ
¥ ¥ =Á
l+m l+m l + m Ë l + m ˜¯
(i) Diagram
0.4
H S R
0.6
D
We have shown State S using a dotted line as this is a transient state with a
one-hour window. Sufferers of the disease pass through this state during
the first hour of sickness before moving on to State R or State D.
n , the mortality rate for people who have recovered from the disease.
(ii) Likelihood
where:
tH denotes the total waiting time for lives in the healthy state H
tR denotes the total waiting time for lives in the recovered state R
nHSR and nHSD denote the number of transitions observed from the
healthy state H to the recovered state R and to the dead state D (via the
sick state S)
nHD denotes the number of transitions observed from the healthy state
H to the dead state D (not passing through state S)
∂ n + nHSD
ln L = -tH + HSR
∂s s
nHSR + nHSD
sˆ =
tH
∂2 nHSR + nHSD
ln L = -
∂s 2 s2
An inhabitant of the island who has never suffered from the disease will still
be alive in 3 years’ time if they are in either state H or state R at time 3. So
the required probability is pHH (3) + pHR (3) .
pHH (3)
We can derive a formula for pHH (3) by first deriving the Kolmogorov
differential equation for pHH (t ) .
pHH (t + h ) - pHH (t )
¢ (t ) = lim
pHH
h Æ0 h
o( h )
where lim =0.
h Æ0 h
So:
d p ¢ (t )
ln pHH (t ) = HH = -(s + m )
dt pHH (t )
fi ln pHH (t ) = -(s + m )t + c
So:
pHR (3)
3
pHR (3) = Ú0 pHH (w ){ 0.4s pRR (3 - w ) + 0.6s pDR (3 - w ) + m pDR (3 - w ) }dw
3
= Ú0 pHH (w ) 0.4s pRR (3 - w ) dw
3
pHR (3) = 0.4s Ú e -(s + m )w e -n (3 -w )dw
0
3
= 0.4s e -3n Ú e -(s + m -n )w dw
0
Ê 1 - e -3(s + m -n ) ˆ
= 0.4s e -3n Á ˜
Ë s + m -n ¯
Ê e -3n - e -3(s + m ) ˆ
= 0.4s Á ˜
Ë s + m -n ¯
So the probability that an islander who has never suffered from the disease
will still be alive in 3 years’ time is:
Ê e -3n - e -3(s + m ) ˆ
pHH (3) + pHR (3) = e -3(s + m ) + 0.4s Á ˜
Ë s + m -n ¯
The students would need to be able to calculate the following statistics for
the 3-year period:
the total waiting time for lives in the healthy state H ( tH )
the number of transitions observed from the healthy state H to the dead
state D via the sick state S ( nHSD ).
(ii) Diagram
The total transition rate out of state NB is 0.1. We need to split this in
proportion to the probabilities of 0.75 going to state R (so the transition rate
is 0.75 ¥ 0.1 = 0.075 ) and 0.25 going to state D (so the transition rate is
0.25 ¥ 0.1 = 0.025 ).
0.075
NB R
0.2
0.025
D
NB R D
NB È -0.1 0.075 0.025 ˘
Í ˙
R Í 0 -0.2 0.2 ˙
D ÍÎ 0 0 0 ˙˚
Also:
and:
¢ (t ) = -0.1PNB (t )
The boundary condition for the first equation PNB is
PNB (0) = 1 , since we know with certainty that the phone is not broken
initially.
¢ (t ) d
PNB
= log PNB (t ) = -0.1
PNB (t ) dt
Integrating gives:
eÚ
0.2dt
= e0.2t
d È 0.2t
ie e PR (t )˘ = 0.075e0.1t
dt Î ˚
Integrating gives:
e0.1t
e0.2t PR (t ) = 0.075 + c = 0.75e0.1t + c
0.1
(
and PR (t ) = 0.75e -0.1t - 0.75e -0.2t = 0.75 e -0.1t - e -0.2t )
(v) Calculate the probability that a phone has not been discarded
PNB (t ) + PR (t ) + PD (t ) = 1
So the probability that the phone has not been discarded by time t is:
1 - PD (t ) = PNB (t ) + PR (t )
( )
= e -0.1t + 0.75 e -0.1t - e -0.2t = 1.75e -0.1t - 0.75e -0.2t
(i) Likelihood
L μ e -( m12 + m13 + m14 )t1e -( m21 + m23 + m24 )t2 m1212 m1313 m1414 m2121 m2323 m2424
n n n n n n
where:
mij denotes the transition rate from state i to state j
nij denotes the observed number of transitions from state i to state j
ti denotes the observed waiting time in state i .
log L = -( m12 + m13 + m14 )t1 + n13logm13 + terms not involving m13
To obtain the MLE we set this equal to zero and rearrange to get:
n13
mˆ13 =
t1
To verify that this does indeed give a maximum, we can check that the
second derivative is negative:
∂2 n13
2
log L = - 2
<0
∂m13 m13
(iii) Testing the significance of the difference between the death rates
Ê m13 ˆ Ê m ˆ
m13 ~ N Á m13 , and m 23 ~ N Á m23 , 23 ˜
Ë t1 ˜¯ Ë t2 ¯
Ê m13 m23 ˆ
m13 - m 23 ~ N Á m13 - m23 , +
Ë t1 t2 ¯˜
Ê m13 m23 ˆ
m13 - m 23 ~ N Á 0, +
Ë t1 t2 ¯˜
m13 - m 23
Z= ~ N (0,1)
m13 m23
+
t1 t2
0.012368 - 0.010492
z= = 1.564
0.012368 0.010492
+
14,392 18,109
Difficulties include:
Estimating the transition rates is more difficult, as we have to estimate a
function for each type of transition, rather than just a single value.
We will also need more data to estimate these functions accurately.
Solving the equations for the process, eg the Kolmogorov differential
equations, will be more difficult algebraically.
If the equations cannot be solved directly, it may be necessary to
employ Monte-Carlo techniques to apply the model.
Ë 0
t
¯ (
= exp Ê - Ú (0.2 + 0.04s )ds ˆ = exp -0.2t - 0.02t 2 )
(iv) Time for a 50% chance of being naughty
This is the median time spent in the Good state, which will satisfy the
equation:
pGG (t ) = 0.5
fi ( )
exp -0.2t - 0.02t 2 = 0.5
Taking logs:
Although this part of the question says ‘Derive’ (which usually means ‘Derive
from first principles’), it is only worth 2 marks, which is too few to suggest
that we have to give a full derivation from first principles.
∂
P (s, t ) = P (s, t ) A(t )
∂t
where P (s, t ) denotes the matrix of transition probabilities over the interval
(s, t ) and A(t ) is the generator matrix at time t .
and hence:
∂ ∂
pG (t ) = pGG (0, t )
∂t ∂t
= pGG (0, t ) ¥ ÈÎ -(0.2 + 0.04t )˘˚ + pGN (0, t ) ¥ (0.4 - 0.04t )
Also:
So:
∂
p (t ) = pGG (0, t ) ¥ ÈÎ -(0.2 + 0.04t )˘˚ + (1 - pGG (0, t )) ¥ (0.4 - 0.04t )
∂t G
= pGG (0, t ) ¥ ÈÎ -(0.2 + 0.04t ) - (0.4 - 0.04t )˘˚ + 0.4 - 0.04t
If we abbreviate the states to H , I , DC and DO (in the same order they are
given in the question), the transition diagram for this process is:
H I
–
DO DC
P ¢(t ) = P (t )A
H I DO DC
H È -s - m s m 0˘
Í
I Í 0 -r r - t t ˙˙
DO Í 0 0 0 0˙
Í ˙
DC ÍÎ 0 0 0 0 ˙˚
(iii) Likelihood
If ti denotes the observed waiting time in state i and nij denotes the
observed number of transitions from state i to state j , the likelihood is:
where C is a constant.
Note that nIDO = 0 because all the 40 deaths are accounted for by the 10
deaths of healthy trees and the 30 deaths from citrus greening.
We are not told the number of trees that became affected during the study,
so the value of nHI is not known. But, since we are not interested in the
infection rate, this doesn’t affect our calculations here.
∂ nIDO nIDC
ln L = - + … (2)
∂t r -t t
nIDO nIDC
=
r -t t
fi nIDO t = nIDC ( r - t )
fi (nIDO + nIDC )t = nIDC r
nIDC
fi t = r … (3)
nIDO + nIDC
Since this expression for t involves the unknown value for the parameter
r , we also need to maximise with respect to r by setting the derivative
with respect to r equal to zero:
∂ nIDO
ln L = -tI + =0
∂r r -t
fi ( r - t )tI = nIDO
nIDO
or r= +t … (4)
tI
nIDO nIDC
nIDO t = nIDC fi t =
tI tI
nIDC
So the maximum likelihood estimate for t is tˆ = and the
tI
NIDC
corresponding estimator is , where NIDC and TI are random variables
TI
representing the number of transitions from state I to state DC , and the
waiting time in state I .
∂2 nIDO nIDC
ln L = - - <0
∂t 2 ( r - t )2 t2
(v) Estimate t
30
tˆ = = 0.05 (in units of tree-months)
600
All three types of process share the Markov property, ie the probabilities for
future states depend only on the current state (not on the previous history).
Markov chains and Markov jump chains both have a discrete time set
whereas Markov jump processes have a continuous time set.
Markov chains and Markov jump chains are both specified in terms of
probabilities whereas Markov jump processes are specified in terms of
transition rates.
The diagram for a Markov chain may contain ‘loops’ (movements to the
same state) whereas Markov jump processes cannot. Markov jump chains
only contain loops where there is an absorbing state.
(ii) Probability the insurance company can’t pay the next claim
S + 1.5l Zt
The insurer will have insufficient funds to pay this claim (which will be for an
amount Z ) if
S + 1.5l Zt < Z
Z -S 1 Ê Sˆ
ie t< = ÁË 1 - Z ˜¯
1.5l Z 1.5l
So the insurer will have insufficient funds to pay the claim if the first claim
1 Ê Sˆ
occurs before time t = 1 - ˜ . The probability is:
1.5l ÁË Z¯
È 1 Ê Sˆ˘ È 1 Ê Sˆ˘
P ÍT < 1- ˙=F Í Á1 - Z ˜¯ ˙
Î 1.5 l ÁË Z ˜¯ ˚ T Î 1.5l Ë ˚
È 1 Ê Sˆ˘
= 1 - exp Í - l Á1 - Z ˜¯ ˙
Î 1.5l Ë ˚
È 1 Ê Sˆ˘
= 1 - exp Í - Á1 - Z ˜¯ ˙
Î 1.5 Ë ˚
The transition rate from state i to state j can be estimated using the
maximum likelihood method, which gives:
nij
mˆij =
ti
Here:
nij denotes the total number of transitions from state i to state j
observed during the observation period
ti denotes the total waiting time in state i , ie the total time spent by all
individuals in state i during the observation period.
A T
A Ê -m mˆ
T ÁË 0 0 ˜¯
where A and T denote the Active Policy and Theft Claim states
respectively.
or ¢ (t ) + m PAT (t ) = m
PAT
d È mt
ie e PAT (t )˘ = m e m t
dt Î ˚
T T
È e m t P (t )˘ = m e mt dt
Î AT ˚0 Ú0
T
fi e mT PAT (T ) - PAT (0) = Èe mt ˘
Î ˚0
=0
mT mT
fi e PAT (T ) = e -1
PAT (T ) = 1 - e - mT
(
C 1 - e - mT )
If we denote the Lapsed state by L and the lapse rate by l , the revised
transition diagram is:
Active Theft
policy claim
Lapse
A T L
A Ê -m - l m lˆ
T Á 0 0 0˜
Á ˜
L ÁË 0 0 0 ˜¯
¢ (t ) = -( m + l )PAA (t ) , PAT
PAA ¢ (t ) = m PAA (t ) and PAL
¢ (t ) = l PAA (t )
¢ (t )
PAA
= -( m + l )
PAA (t )
d
ie ln PAA (t ) = -( m + l )
dt
T T
ÈÎln PAA (t )˘˚ 0 = - Ú ( m + l )dt
0
¢ (t ) = m PAA (t ) = m exp ÎÈ -( m + l )t ˚˘
PAT
T
T
ÎÈ PAT (t )˚˘ 0 = Ú m exp ÈÎ -( m + l )t ˘˚ dt
0
T
È m ˘
fi PAT (T ) - PAT (0) = Í - exp ÎÈ -( m + l )t ˚˘ ˙
Î m+l ˚0
=0
m
fi PAT (T ) = È1 - exp ÈÎ -( m + l )T ˘˚ ˘˚
m+l Î
m
C È1 - exp ÈÎ -( m + l )T ˘˚ ˘˚
m+l Î
H S
(ii) Likelihood
If ti denotes the observed waiting time in state i and nij denotes the
observed number of transitions from state i to state j , the likelihood is:
where C is a constant.
∂ n
ln L = -tH + HS
∂s s
nHS
s =
tH
∂2 nHS
2
ln L = - <0
∂s s2
nHS
Hence the maximum likelihood estimate of s is sˆ = . The maximum
tH
NHS
likelihood estimator of s is s = , where NHS is the random variable
TH
representing the number of transitions from State H to State S and TH is the
random variable representing the total waiting time in State H.
∂ p ( x, t + h ) - pHH ( x, t )
pHH ( x, t ) = lim HH
∂t h Æ0 h
From the definitions of the transition rates and the law of total probability:
and:
pHH (t , t + h ) = 1 - hs (t ) - h m (t ) + o(h )
o( h )
where o(h ) represents a function for which lim =0.
h Æ0 h
= lim
hÆ0
1
h
{
pHH ( x, t ) ÈÎ 1 - hs (t ) - h m (t ) + o(h )˘˚
}
+ pHS ( x, t ) ÈÎ h r (t ,Ct ) + o(h )˘˚ - pHH ( x, t )
Ï o( h ) ¸
= lim Ì pHH ( x, t ) ÈÎ -s (t ) - m (t )˘˚ + pHS ( x, t ) r (t ,Ct ) + ˝
hÆ0 Ó h ˛
The total of the transition rates out of the Healthy state at time t equals
s (t ) + m (t ) . The probability of remaining in the Healthy state from time 0 to
time t can be found by integrating the negative of this quantity over the time
interval (0, t ) and applying an exponential function. This gives:
In order to start in state H at time 0 and finish in state S at time t , there must
be at least one jump between time 0 and time t .
Consider the first such jump, and assume that it occurs during the short time
interval (w , w + dw ) . Here, this jump can only be from state H to state S.
Start and finish in state S over the remaining period between time w
and time t (possibly not moving at all).
H H x S S
To finish, we note that the first jump could occur at any time between 0
and t . So we need to integrate over all possible values of w to obtain the
final equation:
t
pHS (0, t ) = Ú0 pHH (0, w ) ¥ s (w )dw ¥ pSS (w , t )
Ê ˆ
Ú0 exp Ë - Ú0 (s (u ) + m (u )) du ¯ s (w )pSS (w , t )dw
t w
=
In the backward form we condition on the first jump during the time period,
whereas in the forward form we condition on the last jump.
P ÈÎ X t = H X s = S,Cs = w ˘˚
y
t - Ú ( r (u,w - s + u ) +u (u,w - s +u ))du
= Úe s r ( y , w - s + y ) pHH ( y , t ) dy
s
Comparing this with the version in the question reveals that it contains some
errors:
The lower limits for both the outer and inner integrals should be s
rather than 0.
The transition rate for the jump should be r rather than u .
FACTSHEET
A Markov process with a continuous time set and discrete state space is
called a Markov jump process.
Poisson processes
P Nt h Nt 1| Ft h o(h )
P Nt h Nt 0| Ft 1 h o(h )
if j i
ij if j i 1
0 otherwise
When events in a Poisson process are of different types and each type
occurs at random with a fixed probability, then events of each particular type
form a thinned Poisson process. These thinned processes are independent
of one another and for each thinned process the rate is equal to the original
rate multiplied by the probability of getting that type of event.
Transition probabilities
For a Markov jump process X t , the transition probabilities pij (s, t ) are
defined by:
pij (s, t ) = P ( X t = j | X s = i )
Chapman-Kolmogorov equations
Transition rates
Generator matrix
Differential equations
d
pij (t ) = Â pik (t ) mkj
dt k
d
pij (t ) = Â mik pkj (t )
dt k
∂
pij (s, t ) = Â pik (s, t ) mkj (t )
∂t k
∂
pij (s, t ) = -Â mik (s ) pkj (s, t )
∂s k
Note: this is the only one of the Kolmogorov differential equations that
contains a minus sign.
pii (t ) = e - li t
where li = Â mik = - mii is the total force of transition out of State i . So the
k πi
holding time in State i follows an Exp ( li ) distribution. The expected
holding time in State i is 1/ li .
Given that a Markov jump process is in State i at time s and stays there
until time s + w , the probability that it goes into State j when it leaves
State i is:
To find the expected time to reach a given state, State k , starting from
State i , apply the following formula recursively:
1 mij
mi =
li
+ Â li j
m
j π i ,k
This formula is on page 38 of the Tables. It only works for processes that
are time-homogeneous.
Jump chain
The jump chain of a Markov process is the sequence of states that the
process enters. The time spent in each state is irrelevant.
(a) Estimate li by setting 1/ lˆi equal to the sample mean of the duration
of the visits to State i .
nij
(b) Set pˆ ij = where nij is the number of direct transitions from
 nik
k
Intregral equations
t -s
pij (s, t ) = Â Ú0 pik (s, t - w ) mkj (t - w ) p jj (t - w , t ) dw iπ j
kπ j
t -s
pij (s, t ) = Â Ú0 pii (s, s + w ) mik (s + w ) pkj (s + w , t ) dw iπ j
k πi
Testing a model
NOTES
NOTES
NOTES
NOTES
NOTES