INSTITUTE AND FACULTY OF ACTUARIES

CURRICULUM 2019

SPECIMEN SOLUTIONS

Subject CS1A – Actuarial Statistics

 Institute and Faculty of Actuaries

1
(i) Sample median is not affected by the fact that the last two observations are
censored.

It is therefore given by the 5.5th ranked observation, i.e. (355 + 379) / 2 =

367
days. [2]

(ii) We know that the last two observations have minimum values 432 and 463.

Using these two values the sample mean would be equal to 3679/10 = 367.9.

So, the sample mean is at least equal to 367.9 days. [2]

[Total 4]

2
0 ∞
1 1 (t −1) x
∫ e(t +1) x dx +
2∫
(i) =
M X (t ) E=
(etX ) e dx
2
−∞ 0

0 ∞
1  e(t +1) x  1  e(t −1) x 
=   +  
2  t + 1  2  t − 1 
−∞ 0

and for | t |< 1

1 1 1  1
M X (t )=  − = [3]
2  t +1 t −1  1− t 2

′
(ii) { }
M ′X (t ) =(1 − t 2 ) −1 =−(1 − t 2 ) −2 (−2t ) =2t (1 − t 2 ) −2

⇒ E ( X ) = M ′X (0) = 0
′
{
M ′′X (t ) = 2t (1 − t 2 ) −2 } = 2(1 − t 2 −2
) + 2t (−2)(1 − t 2 ) −3 (−2t )

=2(1 − t 2 ) −2 + 8t 2 (1 − t 2 ) −3

⇒ E ( X 2 ) = M ′′X (0) = 2
V ( X ) = E( X 2 ) − E 2 ( X ) = 2

(Alternatively, based on a series expansion:

M X (t ) =1 + t 2 + t 4 + ... ⇒ E ( X ) =0 and E ( X 2 ) =2 and the variance follows.)

[3]
[Total 6]

Page 2
3
(i) P[Y = 2] = 0.25 + 0.05 = 0.3 [1]

(ii) P[X = 0] = 0.75 and

P[{X = 0} ∩ {Y = 2}] = 0.25 ≠ 0.225 = 0.3 * 0.75 = P[X = 0] * P[Y = 2]

Therefore X and Y are not independent.

(Other joint probabilities could be used, but not those with Y = 1.) [2]

(iii) The probability function is

r 1 2 4 8 9 18
P(R = r) 0.25 0.05 0.3 0.1 0.2 0.1
[3]
[Total 6]

4
(i) … X n are independent
The random variables X1 , , 
and identically distributed with X i ~ N (µ, σ2 ) [2]

(ii) X and S 2 are independent

X ~ N (µ, σ2 / n)

( n − 1) S 2 ~ χ 2n−1 [3]
σ 2

=
(iii) tk N (0,1) / χ k2 / k where N (0,1) and χ 2k are independent

X −µ
This result can be applied here, and we get ~ tn−1 [2]
S/ n
[Total 7]

5
(i) P(µ > 180) = P(N(187, 102) > 180)

 180 − 187 
= P  N (0,1) > 
 10 

Page 3
 = P(N(0,1) > − 0.7)

= 0.75804 [2]

(ii) We know that µ|x ~ N (µ* , σ*2 )

 80 ×182 187   80 1 
Where µ* =  + 2   2 + 2= 182.14
 15 2
10   15 10 

1
And σ*2 = = 2.73556 = 1.65402
80 1
2
+ 2
15 10

so P(µ > 180) = P(N(182.14, 1.6542) > 180)

 180 − 182.14 
= P  N (0,1) > 
 1.654 

= P(N(0,1) > -1.29192)

= 0.38 × 0.9032 + 0.62 × 0.90147

= 0.90180 [4]

(iii) The probability has risen, reflecting our much greater certainty over the value
of µ as a result of taking a large sample.

This is despite the fact that our mean belief about µ has fallen, which a priori
might make a lower value of µ more likely.

The posterior distribution has thinner tails / lower volatility, since we have
increased credibility around the mean. [2]
[Total 8]

6
Let the prior distribution of µ have a Gamma distribution with parameters α and λ
as per the tables.
α α
Then = 50 and 2 = 152
λ λ
50
Then dividing the first by the second= λ = 0.22222
152
And so α= 50 × 0.22222= 11.111111

Page 4
 The posterior distribution of µ is then given by

f (µ x) ∝ f ( x µ) f (µ)

∝ e−10µ × µ630 × µ10.11111e−0.22222µ

∝ µ640.11111e−10.22222µ

Which is the pdf of a Gamma distribution with parameters α ' =641.11111 and
λ '= 10.22222

Now under all or nothing loss, the Bayesian estimate is given by the mode of the
posterior distribution. So we must find the maximum of

f ( x) = x 640.11111e−10.2222 x (we may ignore constants here)

Differentiating:

(
e−10.22222 x −10.2222 x 640.11111 + 640.1111x 639.11111
f '( x) = )
= x 639.1111e−10.22222 x (−10.2222 x + 640.11111)

And setting this equal to zero we get

640.111111
=x = 62.62 [7]
10.22222

7
(i) The link function here is g (µ)= log µ . [1]

(ii) (a) The linear predictor is αi + βx where the intercept αi for i = 1, 2

depends on gender.

(b) The linear predictor is αi + βi x where βi for i = 1, 2 also depends on

gender, so that both parameters depend on gender. [4]
[Total 5]

8
(i) From the definition of the gamma density given in the question

α
αα − y
f ( y) = y α−1
e µ
µα Γ(α)

Page 5
  y  
= exp  − − log µ  α + ( α − 1) log y + α log α − log Γ(α) ) 
 µ  

 ( yθ − b ( θ )) 
= exp  + c( y, ϕ) 
 a (ϕ) 

where:
1
θ=−
µ

ϕ=α

1
a ( ϕ) =
ϕ

b ( θ ) = − log ( −θ )

c ( y, ϕ ) = ( ϕ − 1) log y + ϕ log ϕ − log Γ(ϕ).

Hence the distribution has the right form for a member of an exponential
family.

1
The natural parameter is − 1 . The canonical link function is . [5]
µ µ

(ii) Using the information given, we can calculate the deviance differences and
compare that with the differences of the degrees of freedom for each of the
nested models. If the decrease in the deviances is greater than twice the
difference in degrees of freedom this suggests an improvement.

Model Scaled Degrees of Difference

Deviance freedom in scaled
deviance
1 900 12
Age 789 10 111
Age +location 544 7 245
Age * location 541 1 3

From the table we can see that the interaction model does not indicate any
improvement hence the recommended model would be Age +location. [6]

[Total 11]

Page 6
9
In thousands:

 2.5 2.5   1 1
(i)  21 − t0.025,24 5 ,   21 + t0.025,24 5  =
 21 − 2.064 2 ,   21 + 2.064 2 

[19.968,   22.032] [3]

(ii) α ≤ v H1: α > 20

H 0 : 20
(or, H0: α = 20v H1: α > 20)

x − α 0 21 − 20
Test statistic: t=
 2 = => 1.711 =
t0.05,24
s / n 2.5 × 0.2

21 − 20
=
2 > 1.711 =
t0.05,24
2.5 × 0.2

We reject the null hypothesis. [3]

21 − α 0
(iii) = 1.711,     2=
1 − α 0 0.8555,     
= α 0 20.1445 [2]
2.5 × 0.2

(iv) Test H0: λ = 0.6 v H1: λ≠ 0.6

Test statistic (based on normal approximation to Poisson) is:

x − λ0 0.5 − 0.6 −0.1

z= = = =−1.29 ∈ [ −1.96,1 .96]
λ0 / n 0.6 /100 0.077

0.5 − 0.6 −0.1

= =−1.29 ∈ [ −1.96,1 .96]
0.6 /100 0.077

0.5
0.5 + − 0.6
(or, with continuity correction z = 100 = −1.226 )
0.6 /100

λ=
The null hypothesis H 0 : 0.6 cannot be rejected for the year 2011. [3]

(v) Test H 0 : λ 2012 ≤ λ 2011 v H1 : λ 2012 > λ 2011

(or, H 0 : λ 2012 =

λ 2011 v H1 : λ 2012 > λ 2011 )

Page 7
 Overall sample mean λ̂ = 0.55

λ 2012 − λ 2011 0.6 − 0.5 0.1

Test statistics now =
is: z  0.9535 = = = < 1.64
λˆ λˆ 1.1/100 0.104
+
n1 n2

λ 2012 ≤ λ 2011 cannot be rejected at the 5% level.

The null hypothesis H 0 :  
Therefore, we do not have empirical evidence to suggest that the alternative
λ 2012 > λ 2011 is true. [3]
[Total 14]

10
(i) We have:

∞ ∞ ∞
ac a ac a  −( a −1)  ∞
E[X ] −a
= ∫ xf X ( x)dx = ∫ x a +1 dx = ac ∫ x dx = −
a
x
x a −1  c
c c c

and for a > 1

ac a ac
E[X ] =
− (0 − c − a +1 ) =. [2]
a −1 a −1

x x
ac a
(ii) =
FX ( x) ∫=
f X (t )dt ∫ a +1 dt
t
c c

which gives

a
x c
−c t − a  =
FX ( x) = a
−c a ( x − a − c − a ) =
1−   , x≥c
 c x

[OR differentiate FX ( x) to obtain f X ( x ) ] [2]

(iii) The likelihood function is given by:

n n n
ac a
=
L(a ) ∏= ∏ x a+1
=
f X ( xi ) a n na
c ∏ xi−(a+1)
=i 1 =i 1 i =i 1

and

Page 8
 n
(a ) n log(a ) + na log(c) − (a + 1)∑ log( xi )
l=
i =1

For the MLE:

n
+ n log(c) − ∑ log( xi ) = 0
n
l ′(a ) = 0 ⇒
a i =1

n n
=⇒ aˆ n
= n
,
x 
∑ log( xi ) − n log(c) ∑ log  ci 
=i 1 =i 1

n
and for c = 2.5, aˆ = n
[3]
 x 
∑ log  2.5i 
i =1

(iv) For the asymptotic variance we use the Cramer-Rao lower bound:

n n
l ′′(a ) = − 2
, and E l ′′ ( a )  = −
a a2

giving

a2
{ }
−1
V [ aˆ ] − E l ′′ ( a ) 
= = .
n

Hence, asymptotically, aˆ ~ N (a, a 2 n) . [4]

(v) Size of claim in the following year will be given by 1.05X

 4   4 
So we want P(1.05 X > 4) =P  X >  =1 − FX  
 1.05   1.05 

and using FX given in the question

6
 1.05 × 2.5 
=
P(1.05 X > 4)  =  0.0799 . [3]
 4 
[Total 14]

Page 9
11
(i) Scatterplot with suitable axes and clearly labelled:

There does not appear to be much of a relationship, perhaps a slight increasing

linear relationship but it is weak with quite a bit of scatter. [3]

(ii) n = 16

1362
Stt =1496 − = 340
16

14.1602
S yy = 12.531946 − = 0.000346
16

(136)(14.160)
Sty =
120.518 − =
0.158
16
Sty 0.158
=
βˆ = = 0.0004647
Stt 340

14.160 136
αˆ = y − βˆ t = − (0.0004647) = 0.88105
16 16

Fitted line is y = 0.88105 + 0.000465t [4]

2
σˆ 2 1 Sty
(iii) (a) s.e. ( β̂ ) = =
where σˆ2
( S yy − )
Stt n−2 Stt

Page 10
 1 0.1582
=σˆ 2 (0.000346=
− ) 0.0000195
14 340

0.0000195
=
∴ s.e.(βˆ ) = 0.000239
340

(b) Null hypothesis of “no linear relationship” is equivalent to H0: b = 0

βˆ
We use t = ~ t14 under H0: b = 0
s.e.(βˆ )

0.000465
=
Observed t = 1.95 and t0.025,14 = 2.145
0.000239

So we must accept H0: no linear relationship at the 5% level.

(c) 95% CI is 0.000465 ± 2.145 × 0.000239

giving 0.000465 ± 0.000513 or (−0.000048, 0.000978) [5]

0.000487
(iv) (a) =
Observed t = 2.21 – this is greater than t0.025,14 = 2.145
0.000220

So we reject H0: no linear relationship at the 5% level.

(b) 95% CI is 0.000487 ± 2.145 × 0.000220

giving 0.000487 ± 0.000472 or (0.000015, 0.000959)

The two CIs overlap substantially, so there is no evidence to suggest

that the slopes are different.

(c) Although the tests have different conclusions at the 5% level, the 100m
observed t is only just inside the critical value of 2.145 and the 200m
one is just outside. This in fact agrees with, rather than contradicts, the
conclusion that the slopes are not different. [6]
[Total 18]

END OF MARKING SCHEDULE

Page 11