Lecture

Class: M.Sc. – SEM 2

Subject : Statistical and Risk Modelling 1
Chapter: Unit 3 Chapter 1
Chapter Name: Graduation 1

1
Today’s Agenda
1. Graduation

2. Need for graduation

3. Origins of graduation and smoothing

4. Assumptions

5. Comparison with other experiences

6. Procedure to test comparison

7. Desirable features and issues of graduation

8. Testing for smoothness

1. The test

2
Continued…
9. Statistical test for testing graduation
1. Chi-Squared test
2. Limitations of Chi-Squared test
3. Deviations and standardized deviations
4. Standardized deviations test
5. Signs test
6. Cumulative Deviations test
7. Grouping of Signs test
8. Serials Correlations test

10. Methods for Graduation

3
Introduction
We’ve previously learnt about a few models and the way crude mortality rates are modelled. Now we’ll try to
refine them for our use.

Crude Rates against Age

4
1 Graduation
Graduation is a technique used to get a smooth and justifiable mortality rates. (which is also called as
Graduated Mortality Rate).

The crude mortality rates derived from a mortality investigation will not be the final rates that are published for
use in actuarial calculations. The rates will have to pass through a further process called graduation. Graduation
results in a ‘smoothing’ of the crude rates.

The aims of graduation are: –

• To produce a smooth set of rates that are suitable for a particular purpose.
• To remove random sampling error.
• To use information available from adjacent ages

5
2 Need for graduation
Let us understand why do we need to perform the process of graduation.

Theoretical argument:

i. We believe that mortality varies smoothly with age.

ii. Therefore the crude estimate at any age carries information about mortality at adjacent ages. i.e. a crude
estimate of 𝜇𝑥 for any age x also carries information about the values of 𝜇𝑥−1 and 𝜇𝑥+1 .
iii. By smoothing, we can make use of the data at adjacent ages to improve the estimate at each age. Thus, it
reduces sampling errors

6
2 Need for graduation
While theoretical arguments provide a good reason to graduate rates, the real reason why all companies
perform graduation is not theoretical but practical reasons.

Practical argument:
• Mortality data is that which we will use in the life tables to compute financial quantities, such as premiums
for life insurance contracts.
• It is very desirable that such quantities progress smoothly with age, since irregularities (jumps or other
anomalies) are hard to justify in practice.
• Before applying the mortality rates to any financial problem, we need to consider their suitability. Thus we
need to make comparisons with standard tables and other experiences. It is often the case that a mortality
experience must be adjusted in some way before use, in which case there is little point in maintaining the
roughness of the crude estimates.

7
3 Where it all began!!
Actuaries have had a very long association with smoothing. The Gompertz model (Gompertz, 1825) could be
considered an early smoothing method, albeit a very simple one.
Makeham’s extension (Makeham, 1860) of the Gompertz model improved the fit to mortality tables, but neither
model was sufficiently flexible to be applicable outside a limited age range, say 40 to 90.
There were many other efforts, all grouped under the general heading of mathematical formulae, of which Perks
(1932) is perhaps the best known. The basic idea is that adding a parameter will improve the fit, so we have the
Gompertz model with two parameters, Makeham’s with three and the two Perks formulae with four and five
parameters.
English Life Tables No. 11 and No. 12 were graduated using a mathematical formula with seven parameters.

8
4 Assumptions
Consider that we have data for all ages in our investigation.

Using the Poisson or multiple-state model, for x = 𝑥1 , 𝑥2 , … . . 𝑥𝑚 we have:

Number of deaths at age x nearest birthday = 𝐷𝑥

Central exposed to risk at age x nearest birthday = 𝐸𝑥𝑐

Crude estimate of the force of mortality at exact age = 𝜇ෞ𝑥

Approximate distribution that we use is

𝐷𝑥 ~𝑁𝑜𝑟𝑚𝑎𝑙(𝐸𝑥𝑐 𝜇𝑥 , 𝐸𝑥𝑐 𝜇𝑥 )
Graduated rates are denoted as 𝜇ሶ 𝑥

9
5 Comparison with other experiences
• In mortality investigations and insurance studies, we often need to keep a check on the consistency of our
data with other known experiences. It is important to see whether our recent experience is consistent with
past experiences or with published life tables.

• By consistency, we cover two concepts: the shape of the mortality curve over the range of ages and the level
of mortality rates.

Comparison with Standard tables.

• Published life tables based on large amounts of data are called standard tables. These mortality tables use a
large number of factors to predict the likelihood of death in an individual. Mortality tables are used heavily
by insurance companies.

10
6 Procedure to test comparison
• Notation: The superscript ‘s’ will denote a quantity from a published standard table, eg 𝜇𝑥𝑠

• Our hypothesis is that the mortality rates being tested are consistent with those from the standard table. We
will reject this null hypothesis if we find evidence that the rates being tested are significantly different from
those in the standard table.

• We can derive tests of this hypothesis using the distributional assumptions made earlier. Under the
hypothesis we have,

𝑫𝒙 ~𝑵 𝑬𝒄𝒙 𝝁𝒔𝒙 , 𝑬𝒄𝒙 𝝁𝒔𝒙 approximately

• Thus we can find test statistics making comparisons of actual deaths 𝑑𝑥 in our data and the expected deaths
given by this distributional assumption There are suitable tests that can help us to work further

11
7 Desirable features and issues of Graduation
Three desirable features of a graduation are: –
i. Smoothness
ii. Adherence to data
iii. Suitability for the purpose to hand

• We need to make a balance between smoothing and adherence to data. At one extreme, we could easily
smooth the crude estimates by ignoring the data altogether; we want to avoid such extremes since we want
the graduation to be representative of the experience.
• If the graduation process results in rates that are smooth but show little adherence to the data, then we say
that the rates are over-graduated.
• If insufficient smoothing has been carried out then it is called under-graduation.

12
7 Desirable features and issues of Graduation

• Smoothness and adherence to data are usually conflicting requirements. Perfect smoothness (extreme
example: a straight line) pays little or no attention to the data, while perfect adherence to the data means no
smoothing at all.
• Over graduation - If the graduation process results in rates that are smooth but show little adherence to
the data, then we say that the rates are overgraduated. The graph in Graduation A (see above) is very
smooth, but it tends to overestimate the crude rates at the younger ages and underestimates them at the
older ages.
• Under graduation – Graduation where insufficient smoothing has been carried out. This will tend to
produce a curve of inadequate smoothness, but better adherence to data. The graph in Graduation B (which
uses the same data as Graduation A) adheres very closely to the crude rates, but it twists and turns erratically
13
8 Testing for Smoothness
• The achievement of smoothness is one of the main purposes of graduation. The ‘art’ of graduation lies in
finding a satisfactory compromise between smoothness and adherence to data.

Smooth Graduation
• Mathematical smoothness is usually defined in terms of differentiability, but this is of little use in graduation
work because many functions that misbehave wildly between integer ages are nevertheless differentiable
many times.
• The test for smoothness will be used as a check for undergraduation.

14
8.1 The test
To test smoothness we will find the third-order differences of the graduated rates.
The first difference Δ𝜇ሶ 𝑥 = 𝜇ሶ 𝑥+1 − 𝜇ሶ 𝑥
The second difference Δ2 𝜇ሶ 𝑥 = Δ𝜇ሶ 𝑥+1 − Δ𝜇ሶ 𝑥
The third difference Δ3 𝜇ሶ 𝑥 = Δ2 𝜇ሶ 𝑥+1 − Δ2 𝜇ሶ 𝑥

The criterion for the graduated rates to qualify the smoothness test is that the third differences should:
• be small in magnitude compared with the quantities themselves; and
• progress regularly.

15
Question 1
Compare the smoothness of the rates in Graduations A and B over the age range 30 to 35:
Looking at the graph in the next slide, also comment about undergraduation and overgraduation of rates

16
Solution
For A, the third differences are very small, which indicates that the graduated rates are very smooth.
The third differences are larger for Graduation B than for Graduation A (especially when x = 30 ) and they
progress in a less regular manner. This indicates that Graduation B is not as smooth as Graduation A.

17
Graduation A
We are going to use Graduation A &
Graduation B for a few of the tests to follow.
These charts represent their data.
𝜇
Graduation A assumed that ln(𝑒 𝑥 − 1) could
be modelled as 𝛼 + 𝛽𝑥 (2 parameters), which
was fitted using the method of least squares.

18
Graduation B
We are going to use Graduation A &
Graduation B for a few of the tests to follow.
These charts represent their data.
𝜇
Graduation B assumed that ln(𝑒 𝑥 − 1) could
be modelled as a polynomial of degree 10 (11
parameters), which was fitted using the
method of least squares.

19
9 Statistical tests for testing graduation
Statistical test process

Hypotheses - The null hypothesis (denoted by 𝐻0 ) corresponds to a neutral conclusion. In graduation tests, the
null hypothesis will correspond to a statement that some aspect of a proposed graduation is ‘OK’. In graduation
tests, the alternative hypothesis (𝐻1 ) will correspond to a statement that some aspect of a proposed graduation
is ‘no good’.

We then perform the required statistical test and find the probability value – (p value).

P – Value - A significance level must be selected at the beginning of the test. The significance level usually used
is 5%, which means that if 𝐻0 were true, the value of the test statistic would only be this extreme by chance 1
time in 20. A statistical test may be one-tailed or two-tailed, depending on the nature of the test and the feature
we are interested in.

20
9.1 Chi-Squared Test
Chi-squared test is a test for the overall fitness of fit.
A chi-squared test can be used to assess whether the observed numbers of individuals in a specified categories
are consistent with a model that predicts the expected numbers in each category.

𝑂−𝐸 2
Test Statistic - The chi-squared test statistic is: ∑ 𝐸
where,
• O is the observed number in a particular category
• E is the corresponding expected number predicted by the assumed probabilities
• the sum is over all possible categories .

This statistic has a chi-squared distribution (approximately), which is tabulated in the statistics section of the
Actuarial Tables.

Calculate Degrees of freedom as = No. of categories – number of parameters

21
Question 2
Perform the Chi Square Goodness of Fit test on the below data

Age Crude Rates Central Exposed to Graduated Rates

Risk
41 0.02 100 0.022
42 0.03 200 0.0305
43 0.033 150 0.034
44 0.067 120 0.068
45 0.07 100 0.075
46 0.075 100 0.08

One – parameter estimation has been performed to calculate the graduated rates.

Do you see an issue with your conclusion in the above test? If yes, what?
22
9.2 Limitations of Chi-Squared Test
The 𝜒 2 test will fail to detect several defects that could be of considerable financial importance. (These
comments apply particularly when we are testing a graduation, and for ease of exposition we will write as if that
were the case.)
i. There could be a few large deviations offset by a lot of very small deviations. In other words, the 𝜒 2 test
could be satisfied although the data do not satisfy the distributional assumptions that underlie it. This is, in
essence, because the 𝜒 2 statistic summarises a lot of information in a single figure.
ii. The graduation might be biased above or below the data by a small amount. The 𝜒 2 statistic can often fail
to detect consistent bias if it is small, but we should still wish to avoid it.
iii. Even if the graduation is not biased as a whole, there could be significant groups of consecutive ages
(called runs or clumps) over which it is biased up or down. This is still to be avoided.
iv. It should be noted that because the 𝜒 2 test is based on squared deviations, it tells us nothing about the
direction of any bias or the nature of any lack of adherence to data of a graduation, even if the bias is large
or the lack of adherence manifest. To ascertain this there is no substitute for an inspection of the
experience.

23
Question
Apply the chi-squared test to Graduation A.

24
Solution
From the table of values for Graduation A, we see that:
∑ 𝑍𝑋2 = 43.17
In this example, it is not difficult to work out how many degrees of freedom to use. There are 20 ages. We have
not constrained the totals. The graduated rates have been calculated by estimating 2 parameters. So, the
number of degrees of freedom is 20 – 2 = 18 .
From the Tables, the upper 5% point for the 𝜒18 2
distribution is 28.87. The observed value of the test statistic
exceeds this, so we reject the null hypothesis. (In fact, the test statistic also exceeds 42.31, the upper 0.1%
point.)
So, we conclude that the mortality experience does not conform to a formula of the type assumed in the
graduation.

25
9.3 Deviations and standardized deviations
When we are testing the adherence to data of a graduation, we assume the normal distribution is used as an
approximation to the Poisson distributions.
𝑫𝒙 ~𝑵(𝑬𝒄𝒙 𝝁ሶ 𝒙 , 𝑬𝒄𝒙 𝝁ሶ 𝒙 )

• Deviations and standardized deviations:

The deviation at age x is defined to be: Actual deaths – Expected deaths
i.e. 𝑫𝒙 − 𝑬𝒄𝒙 𝝁ሶ 𝒙
and the standardized deviation, denoted 𝑧𝑥 is:
𝑫𝒙 − 𝑬𝒄𝒙 𝝁ሶ 𝒙
𝒁𝒙 =
𝑬𝒄𝒙 𝝁ሶ 𝒙

In case of comparison with standard table replace 𝜇ሶ 𝑥 by 𝜇𝑥𝑠 , above.

26
9.4 Standardized deviations test
The test looks at the distribution of the values of the standardized deviations.

• Step 1 – Calculate the standardized deviations 𝑧𝑥 for each age .

• Step 2 - Divide the real (number) line into any convenient intervals. Plot or count the number of
standardized deviations falling into each of the ranges.
• Step 3 - We can then compare: the observed number of the 𝑧𝑥 that fall in each interval; and the expected
number of the 𝑧𝑥 that should fall in each interval, under the hypothesis.

The hypothesis states that the 𝑧𝑥 values are realizations of a standard normal random variable. If the
standardized deviations do not appear to conform to a standard normal distribution, this indicates that
mortality rates do not conform to the model with the rates assumed in the graduation.

27
9.4 Standardized deviations test
In this example, the expected numbers are:

Interval (−∞, −3) (-3,-2) (-2,-1) (-1,0) (0,1) (1,2) (2,3) (3, ∞)
Expected
0 0.02m 0.14m 0.34m 0.34m 0.14m 0.02m 0
number

28
9.4 Standardized deviations test
• Step 4 -To formalise the comparison, we can form a 𝜒 2 statistic (nothing to do with the use of the 𝜒 2 test
mentioned previously):
𝑨𝒄𝒕𝒖𝒂𝒍 − 𝑬𝒙𝒑𝒆𝒄𝒕𝒆𝒅 𝟐
𝝌𝟐𝒄𝒂𝒍 = ෍
𝑬𝒙𝒑𝒆𝒄𝒕𝒆𝒅
𝒂𝒍𝒍 𝒊𝒏𝒕𝒆𝒓𝒗𝒂𝒍𝒔

which here should have a 𝜒 2 distribution with 7 degrees of freedom (since we have used 8 intervals).

Note how this differs from the way we previously applied the 𝜒 2 test, which was to test whether the observed
numbers of deaths were consistent with a given set of graduated rates. Here, we are testing whether the
observed pattern of the individual standardised deviations (ie the numbers falling in each interval) is consistent
with a standard normal distribution.

• If the number of age groups is small, we should use a smaller number of intervals, ensuring that the
expected number of standardised deviations in each interval is not less than five (recommended) but should
not be less than one (as a rule of thumb), and we then reduce the number of degrees of freedom in the 𝜒 2
test appropriately.

29
Question
Analyse the distribution of the standardised
deviations for Graduation A.

30
Solution
The observed and expected numbers in each range are shown in the table below.

There are only 7 values in the range ( -2/3,2/3) . So, there appear to be too few values in the centre of the
distribution and too many in the tails. This might indicate overgraduation (an inappropriate graduation formula)
or the presence of duplicates.
The distribution of the standardised deviations is fairly symmetrical, with 10 positive and 10 negative values. So
there is no evidence of bias in the graduated rates.

31
9.5 Signs test
The signs test is a simple test for overall bias. In other words, this test checks whether the graduated rates are
too high or too low.

• Step 1 - Count how many of the graduated rates lie above/below the crude rates. We will do this by looking
at the signs of the individual standardized deviations.
• Step 2 - Calculate the probability value for the test by finding the probability of obtaining a split of
positive/negative values as extreme as observed.

Define the test statistic: P = Number of 𝑧𝑥 that are positive

Under the hypothesis, P ~ Binomial(m,½) . So the probability function of P is:
𝒎
𝒎 𝟏
𝑷 𝑷=𝒙 = , 𝒙 = 𝟎, 𝟏, … , 𝒎
𝒙 𝟐

An excess of either negative or positive deviations is a defect, so we apply a two-tailed test.

32
Question
State the conclusion that can be drawn from an
examination of the signs of the deviations for
Graduation B.

33
Solution
There are 12 positive and 8 negative values for Graduation B. Here we have more positive deviations than
expected. So the p-value is:

2𝑃 𝑝 ≥ 12 = 2 1 − 𝑃 𝑃 ≤ 11 = 2 1 − 0.7483 = 0.5034

The value of 0.7483 comes from page 188 of the Tables.

Since this is (much) greater than 5%, there is very little evidence of bias in the graduated rates.

34
9.6 Cumulative Deviations test
The cumulative deviations test detects overall bias or long runs of deviations of the same sign.
Applying normal approximation to the deviation and standardizing gives us :
∑∀𝒙 𝑫𝒙 − 𝑬𝒄𝒙 𝝁ሶ 𝒙
~𝑵(𝟎, 𝟏)
∑∀𝒙 𝑬𝒄𝒙 𝝁ሶ 𝒙
• Step 1 - Calculate ∑dx (the total observed deaths) and 𝐸 ∑𝑐𝑥 𝜇ሶ 𝑋 (the total expected deaths), where the sum is
over the selected age range.
∑𝒅𝒙 −∑ 𝑬𝒄𝑿 𝝁ሶ 𝒙
• Step 2 - Calculate the test statistic and use this to determine the p-value using the tables for
∑ 𝑬𝒄𝑿 𝝁ሶ 𝒙
the standard normal distribution.

35
Question
State the conclusion that can be drawn from
applying the cumulative deviations test to the whole
age range of Graduation A.

36
Solution
From the table given on Graduation A, the value of the test statistic is:

{∑𝑑𝑋 − ∑(𝐸𝑋𝑐 𝜇𝑋ሶ )] 1561 − 1516.50

= = 1.143
∑ 𝐸𝑋𝑐 𝜇ሶ 𝑋 1516.50

This is a two-tailed test, so we compare the value of the test statistic with the upper and lower 2.5% points of
N(0,1) , ie ±1.96 . As −1.96 < 1.143 < 1.96 , there is insufficient evidence to reject the null hypothesis.

So, the cumulative deviations test does not provide evidence that the graduated rates are biased.

37
9.7 Grouping of Signs test
This test is used mainly to test for over graduation. Lets look at how the test is done…

• Step 1 - Determine the sign of the deviation at each age.

• Step 2 - Count the number of groups of positive signs ( = G ).
• Step 3 - The hypothesis is that the given 𝑛1 positive deviations and 𝑛2 negative deviations are in random
order. Since every pair of positive groups must be separated by a negative group, the numbers of positive
and negative groups will be small or large alike, so a one-tailed test is appropriate. We should find the
smallest k such that:
𝑘 𝑛1 − 1 𝑛2 + 1
𝑡
෍ 𝑡−1 𝑚 ≥ 0.05
𝑡=1 𝑛1

and say that the test has been failed (at the 5% level) if G < k .
If there are too few runs, this indicates that the rates are over graduated.

Alternatively, we could look up the critical value of the test on page 189 of the Tables. If the number of groups
of positive deviations is less than or equal to the critical value given in the Tables, we reject the null hypothesis.

38
Question
Test Graduation A for overgraduation using the
grouping of signs test.

39
Solution
Here there are 20 age groups, with 10 positive and 10 negative deviations. From page 189 of the Tables, we see
that the critical value is 3 when 𝑛1 =10 and 𝑛2 =10 . Looking at the column of 𝑧𝑋 values on page 12, we see that
there are 3 runs of positive deviations. So we reject the null hypothesis at the 5% significance level and conclude
that there is evidence of grouping of deviations of the same sign.

Alternatively, we could carry out the test using a normal approximation. The expected number of positive runs
10 10+1 10𝑋10 2
is: = 5.5 and the variance is: = 1.25
10+10 10+10 3

The p-value is P(G≤ 3) . Using N(5.5,1.25) as an approximation to the distribution of G and incorporating a
continuity correction:
3.5 − 5.5
𝑃 𝐺 ≤ 3 ≅ 𝑃 𝑁 0,1 ≤ = Φ −1.78885 = 1 − Φ 1.78885 = 3.7%
1.25
Since the p-value is less than 5%, we reject the null hypothesis at the 5% level and conclude that there is
evidence of grouping of deviations of the same sign.

40
9.8 Serial Correlations test
We have this test to test for clumping of deviations of the same sign. If clumping is present, then the graduation
has the wrong shape.

• Step 1 - Calculate the standardized deviations 𝑧𝑋 for each age or age group.
• Step 2 - Calculate the serial correlation coefficients using the formula:
𝟏
∑𝒎−𝒋 (𝒛 − 𝒛ത )(𝒁𝒊+𝒋 − 𝒛ത )
𝒎 − 𝒋 𝒊=𝟏 𝒊
𝒓𝒋 =
𝟏 𝒎
∑ 𝒛 − 𝒛ത 𝟐
𝒎 𝒊=𝟏 𝒊
𝟏
where 𝒛ത = ∑𝒎 𝒛 is the overall average of 𝑧𝑋 for the m ages
𝒎 𝒊=𝟏 𝒊
We can use the fact that 𝑟𝑗 must take values in the range -1 ≤ 𝑟𝑗 ≤ 1 to check the calculations for
reasonableness, j is the lag.

• Step 3 - Multiply by √𝑚 to obtain the value of the test statistic and compare this with the percentage points
of the standard normal distribution.
If the test statistic is ‘too positive’, this indicates that the rates are over graduated.

41
Question
Carry out the serial correlation test at lag 1 for
Graduation A.

42
Solution
The mean of the individual standardised deviations is:
49
1 1
𝑧ҧ = ෍ 𝑍𝑋 = 2.27 + 2.69 + ⋯ + 1.99 = 0.18
20 20
𝑋=30
The denominator of 𝑟1 is:
49
1
ҧ
෍ ((𝑍𝑋 − 𝑍))^2 = 2.13
20
𝑋=30
The numerator of 𝑟1 𝑖𝑠:
48
1
ҧ 𝑋+1 − 𝑍)ҧ = 0.94
෍ (𝑍𝑋 − 𝑍)(𝑍
19
𝑋=30
So:
0.94
𝑟1 = = 0.44
2.13
And the value of the test statistic is
20 𝑋 0.44 = 1.97
This is more than 1.6449, the upper 5% point of the standard normal distribution. So there is evidence of
grouping of deviations of the same sign.

43
Question 3
A graduation of the mortality experience of the male population of a region of the United Kingdom has been
carried out using a formula with 3 parameters. The following is an extract from the results.

44
Question 3
Continued…

i. Use the chi-squared test to test the adherence of the graduated rates to the data. State clearly the null
hypothesis you are testing and comment on the result.
ii. Perform two other tests that detect different aspects of the adherence of the graduation to the data. For
each test state clearly the features of the graduation that the test is able to detect, and comment on your
results.

45
Solution
i. Chi-squared test
The null hypothesis is:
𝐻0 : the graduated rates are the true underlying mortality rates for the population
We calculate the individual standardized deviations at each age using the formula:
𝜃𝑥 − 𝐸𝑥𝑐 𝜇ሶ 𝑥
𝑍𝑥 =
𝐸𝑥𝑐 𝜇ሶ 𝑥
The ISDs are:
-0.845, 0.554, -0.408, 2.146, 2.328, 0.904, 1.796,-0.407, -0.321
The test statistic for the chi-squared test (based on unrounded 𝑧𝑥 values) is:
∑𝑧𝑥2 = 15.53
We now compare this with a 𝜒 2 distribution. We were given data from 9 ages. Since the graduation was
carried out using a formula with 3 parameters, we lose 3 degrees of freedom. So we are left with 6 degrees
of freedom.
From the Tables, we see that the upper 5% point of 𝜒62 is 12.59. As the value of the test statistic is greater
than this, we reject the null hypothesis and conclude that the graduated rates do not provide a good fit to
the data. In particular, it looks like the graduated rates are too low for ages 17 to 20.

46
Solution
ii. Two other tests
You can take your pick here from the individual standardized deviations test, the signs test, the cumulative
deviations test, the grouping of signs test and the serial correlation test.

The null hypothesis for all the tests is:

𝐻0 : the graduated rates are the true underlying mortality rates for the population

ISD Test
This is a good all round test that detects most of the problems that might be present in a graduation
including any outliers.
For this test we compare the 𝑧𝑥 values with a standard normal distribution:

47
Solution
Continued…
There are 4 things to consider here:
• Outliers-there are no ISDs greater than 3 in absolute value, which is good; however, with fewer than 20 age
groups, we should be suspicious about any ISD greater than 2 in magnitude, and here we have 2 ISDs
greater than 2.
• The balance of positive and negative deviations- this is OK
• Symmetry- the distribution is a bit positively skewed, which is not so good.
2 2 1 4
• Proportion of ISDs lying in the range (− , ) should be - it is here, which is OK.
3 3 2 9

• The graduated rates fail this test as the ISDs do not appear to be normally distributed. In particular, the
graduated rates appear to be too low at ages 17 and 18.

Signs test
• This is a simple two-tailed test for overall bias.
• There should be roughly equal numbers of positive and negative ISDs. Under the null hypothesis, the
number of positive deviations has a Binomial(9,0.5) distribution.
• We have 5 positives and 4 negatives, which is fine.
• So we do not reject the null hypothesis and we conclude that there is no evidence of overall bias in the
graduated rates.
48
Solution
Continued…
Cumulative deviations test
• This is a two-tailed test for overall bias.
• The observed value of the test statistic is:
∑𝜃𝑥 − ∑ 𝐸𝑥𝑐 𝜇ሶ 𝑥 91 − 73.704
= = 2.015
∑ 𝐸𝑥𝑐 𝜇ሶ 𝑥 73.704
• For a test at the 5% significance level, we compare the value of the test statistic with the lower and upper
2.5% points of N(0,1) , ie with ±1.96 . Since 2.015 is greater than 1.96, we reject the null hypothesis and
conclude that the graduated rates are too low overall.

• Make sure that you don’t do both of the signs test and the cumulative deviations test as they both test for
the same thing.

49
Solution
Continued…
Grouping of signs test
• This is a one-tailed test that detects clumping of deviations of the same sign.
• The observed number of positive deviations is 5, and the observed number of negative deviations is 4.
• From the Tables, we find that the critical value is 1 and we reject the null hypothesis if the observed number
of positive runs is less than or equal to this.
• The observed number of positive runs is 2, so we do not reject the null hypothesis in this case, and we
conclude that there is no evidence of grouping of signs.

50
Solution
Continued…
Serial correlation test
• This is an alternative test for grouping of signs, but it takes much longer to carry out this test so it’s
recommended that you do don’t do it unless you absolutely have to. Make sure that you don’t carry out
both the grouping of signs test and the lag-1 serial correlation test since they test for the same thing.
• This is a one-tailed test that detects clumping of deviations of the same sign.
• The serial correlation coefficient at lag 1 is:
1 21
∑𝑥=14(𝑧𝑥 − 𝑧)(𝑧
ҧ 𝑥+1 − 𝑧)ҧ 0.2165
𝑟1 = 8 = = 0.1643
1 22 1.3172
∑ 𝑧 − 𝑧ҧ 2
9 𝑥=14 𝑥
• and the value of the test statistic is:
𝑟1 𝑚 = 0.1643𝑋3 = 0.493
• As we are only testing for positive correlation, we compare the value of the test statistic with 1.6449, the
upper 5% point of N(0,1) . We find that there is insufficient evidence to reject the null hypothesis or, in other
words, there is no evidence of grouping of signs.

Comment
• The graduation has not fully taken into account the accident hump, ie the increase in mortality around the
late teens and early twenties.
51
10 Methods for Graduation
• The process of graduation can be done using different methods.

• We will look at three methods of carrying out a graduation:

▪ graduation by parametric formula
▪ graduation by reference to a standard table
▪ graduation using spline functions.

• The most appropriate method of graduation to use will depend on the quality of the data available and the
purpose for which the graduated rates will be used.
• We will look at these methods in the next chapter

52
Recap
• Graduation is a technique used to get a smooth and justifiable mortality rates. (which is also called as
Graduated Mortality Rate).
• The aims of graduation are: –
• To produce a smooth set of rates that are suitable for a particular purpose.
• To remove random sampling error.
• To use information available from adjacent ages
• In mortality investigations and insurance studies, we often need to keep a check on the consistency of our
data with other known experiences. It is important to see whether our recent experience is consistent with
past experiences or with published life tables.
• By consistency, we cover two concepts: the shape of the mortality curve over the range of ages and the level
of mortality rates.
• Three desirable features of a graduation are: – i. Smoothness; ii. Adherence to data & iii. Suitability for the
purpose to hand.
• We need to make a balance between smoothing and adherence to data. At one extreme, we could easily
smooth the crude estimates by ignoring the data altogether; we want to avoid such extremes since we want
the graduation to be representative of the experience.
• If the graduation process results in rates that are smooth but show little adherence to the data, then we say
that the rates are over-graduated.
• If insufficient smoothing has been carried out then it is called under-graduation.
53
Continued…
• Chi-squared test is a test for the overall fitness of fit. It can be used to assess whether the observed numbers
of individuals in a specified categories are consistent with a model that predicts the expected numbers in
each category.
• Standardized deviations test- The test looks at the distribution of the values of the standardized deviations.
• The signs test is a simple test for overall bias. In other words, this test checks whether the graduated rates
are too high or too low.
• The cumulative deviations test detects overall bias or long runs of deviations of the same sign.
• Grouping of signs test is used mainly to test for over graduation.
• Serial correlations test tests for clumping of deviations of the same sign. If clumping is present, then the
graduation has the wrong shape.

54