You are on page 1of 15

JOURNAL OF CLIMATOLOGY, VOL. 6, 661-675 (1986) 551.577.34(485):551 S01.

45

A HOMOGENEITY TEST APPLIED TO PRECIPITATION


DATA
HANS ALEXANDERSSON
Department of Meteorology, Box 516, S-751 20 Uppsala, Sweden

Received 2 December 1985


Revised 12 March 1986

ABSTRACT
In climate research it is important to have access to reliable data which are free from artificial trends or changes.
One way of checking the reliability of a climate series is to compare it with surrounding stations. This is the idea
behind all tests of the relative homogeneity. Here we will present a simple homogeneity test and apply it to a
precipitation data set from south-western Sweden. More precisely we will apply it to ratios between station values
and some reference values. The reference value is a form of a mean value from surrounding stations. It is found
valuable to include short and incomplete series in the reference value. The test can be used as an instrument for
quality control as far as the mean level of, for instance, precipitation is concerned. In practice it should be used
along with the available station history. Several non-homogeneities are present in these series and probably reflect
a serious source of uncertainty in studies of climatic trends and climatic change all over the world. The significant
breaks varied from 5 to 25 per cent for this data set. An example illustrates the importance of using relevant
climatic normals that refer to the present measurement conditions in constructing maps of anomalies.

KEY WORDS Homogeneity test Quality control Precipitation data control Climatic change Climatic trends
Climatic normals

INTRODUCTION
It has long been recognized that non-homogeneous climate data are a poor source of information for
climate research. Hydrologists have often preferred to use double mass curve analysis to obtain
information on the relative homogeneity of precipitation series. This technique is purely qualitative.
In the statistical literature several tests have been described which can be applied to climate data sets
to detect non-homogeneities. As an example a run test can be applied on a series of differences or
ratios between a test site and a nearby station with a homogeneous record.
Also in the older literature of statistical climatology, homogeneity tests have been described using
different ‘criteria’ (Conrad and Pollak, 1950; Brooks and Carruthers, 1953).
Almost all tests of relative homogeneity are of a type that gives no information about the most
probable date for a break and no information about the magnitude by which a series has changed.
These specifications are, however, necessary when climatologists try to homogenize data records. They
should also be based on mathematically stringent and objective methods.
A test that gives this essential information has been developed (Maronna and Yohai, 1978) and
applied on annual precipitation series for a number of stations in the north-eastern U.S.A. (Potter,
1981). In their test a difierence describes the magnitude of a possible and single break at the tested site.
Several other articles dealing with homogeneity tests can be found in the statistical literature (e.g.
Gardner, 1969; Segen and Sanderson, 1980; Willsky and Jones, 1976).
In the test used here the starting point is to form a sequence of ratios { q ; }between the test site values
and some type of mean value, here called a reference value, from the surrounding stations with
homogeneous precipitation records. The sequence of ratios is also standardized (i.e. the ratios are
reduced by the sample mean value and divided by the sample standard deviation). This new series then

0196- 1748/86/060661- 15$07.50


@ 1986 by the Royal Meteorological Society
662 H. ALEXANDERSSON

necessarily has a zero mean value and a unit standard deviation which makes a theoretical analysis very
simple. The likelihood ratio technique is used to find out critical values which give us an objective way
of classifying series as homogeneous or non-homogeneous.
There are many ways in which the simple test technique used here can be applied to data in practice.
One important problem is how to form reference values. Ideally these should be determined from good
quality stations surrounding the test site. Usually it is not possible to determine beforehand which of
the stations that are homogeneous and of good quality. Non-homogeneities may occur due to changes
of instrumentation, growing trees, new observers, leaking cans, etc. Thus this type of test has to be
run more than once. It is also intuitively clear that a reference mean value should rely most upon the
stations nearest to (or rather with largest correlation to) the actual test site. These and other problems
are discussed along with examples-some showing the danger of using non-homogeneous records in
analyses.

THE THEORY OF THE RATIO TEST


There are some advantages of using ratios of precipitation series instead of differences. First, ratios
have long been used for interpolating missing precipitation data. This is based upon the reasonable
assumption that integrand precipitation amounts from nearby stations are proportional to each other.
The starting point is to form a series of ratios { q; } between station values and some regional
reference values. As we use the ratios we need not specify the distribution of the precipitation itself.
The coefficient of variation of a series of ratios is clearly less than for the original series of annual
precipitation (see the results section). This gives a very valuable reduction of the noise in the sequences
and illustrates the main idea of testing the relative homogeneity. In practice it is hardly possible to
reduce the scatter of the ratios to really low levels. There are at least two reasons for this:
1. Small scale, convective precipitation gives irregularities within very small areas.
2. No series is perfect, and minor non-homogeneities can often be present in the series of a reference
station.
We now proceed as follows. We define a new standardized series of ratios {zi}according to

where ij is the arithmetic mean value of the ratios {qi} and sq is the sample standard deviation of this
series. This new series {zi}thus has exactly zero mean value and unit standard deviation. Now we may
define our hypotheses, Ho, the null hypothesis, and Hl, the alternative hypothesis, as
Ho: 2 E N ( 0 , l), Vi
For some 1S Y < n and pl# p2 we have
Z € N ( p 1 ,l), for i s v
Z € N ( p 2 , l), for i > Y
Here Z E N(0, 1) means that the 2 has a normal distribution with zero mean value and unit standard
deviation. Thus we have assumed that the sequence of ratios can be described by a normal distribution
and that the possible break is one single break and that it consists only of a shift of the mean level. In
equation (2), 0, p1 and p2 are mean values of normal distributions which all have unit standard
deviation. The problem is now to find a test quantity, based on equation (2), that can give us an answer
to the question: ‘When and by how many per cent and how certain is it that a series contains a
non-homogeneity?’. To do this we can use the standard technique of likelihood ratios as explained by
Lindgren (1968), for instance. Then we obtain
A HOMOGENEITY TEST 663

By differentiation the maximum of the numerator is found for p1 = Z1 and p2 = Z, where

(4)

Inserting in equation (3) gives, after some calculations,

This was also done by Hawkins (1977) although he wrote equation (5) a bit digerently, using

Here we prefer to use equation (5) which gives us our test statistic To as

Hence, after having standardized the sequence of ratios, we only have to compute the accumulated
sums of this sequence from which the arithmetic mean values Z1 and Z,, and thus Tv and T,, can be
calculated. Thus we obtain:
1. To. If To is larger than a certain critical level the series should be classified as non-homogeneous at
a certain level, e.g. with 95 per cent confidence.
2. The year (or the month and year) which is the most probable for a break. More precisely it is the
last year (or the last month and year) with the former mean precipitation level.
3. ijl and &-the mean values of the sequences of the ratios before and after the possible break. The
ratio &/ql will then give a measure of the relative change of the precipitation measurements. If
this quantity is 1.10 the catchment of the gauge has increased by 10 per cent.
In Hawkins’ paper critical levels were simulated using standard normal numbers. This is not exactly
applicable here as we made the standardization, equation (l), before we used equation (7) to find
Tv-curves.To obtain new critical values the standard normal numbers should be standardized before we
use them in simulations, i.e. if we want to find the critical level for a series of 50 values (e.g. years)
every set of 50 values should be reduced by the sample mean of the set (almost zero) and divided by the
sample standard deviation (almost one) before we simulate a series of T,. The differences of the levels
becomes negligible as the number n grows very large. For ordinary values of n (-50) the critical levels
used here, compared with those given by Hawkins, are about two per cent smaller. The critical levels
will naturally depend on the number of values, n, in the series. In Figure 1 data of critical levels are
given for up to 1000 values (if monthly values are used, such large values may be interesting) for two
significance levels, 90 and 95 per cent. Thus one out of ten homogeneous series will, just from mere
chance, exceed the Tw level. Nevertheless the 90 per cent level is preferable since real non-
homogeneities might too easily remain undetected. It is not very meaningful to use this test for series
that are shorter than about five years. If we apply it to monthly data it may be possible to make a
meaningful test on somewhat shorter series.
Hawkins also showed that the year of the most probable break did not have a uniform distribution.
More precisely: for homogeneous series (under the null hypothesis) the year of the most probable
break tends to occur near the beginning or the end of a series. This is a disadvantage of the test, and it
may be worth while to work out a ‘better’ test statistic that suppresses the ends of the series
(Alexandersson, 1984), or to simply omit the ends (Buishand, 1984).
Finally we should comment upon some rather arbitrary choices here. For instance we did not specify
664 H. ALEXANDERSSON

I
I
I 1 1 I I I I I I

TO CRITICAL L E V E L S FOR
9 . THE RATIO TEST
g

8 .- -

7 .- -

6.
- 9.80 650 8.95 10 4 0 -
50 7. 25 8.55 350 1.60 9.90 700 9.00 10.45
75 7.65 8.95 400 1 7 0 10.00 750 9.05 10.45 -
5 '-1 100 7.eo 9.15 c50 1.75 10.10 800 9.05 10.50
/ 150 8.05 9.35 500 8.85 10.~0 850 9.10 10.~0

4 -/ 200 8.20 9.55 550 8.90 10.25 900 9.10 10.50 -


250 8.35 9.70 600 8.95 10.30 950 9.10 10.50
1000 9.10 10.50
~

0 10 20 30 40 50 60 70 80 90 100 n

Figure 1. Critical levels for the ratio test as a function of the number of values and for two significance levels T,, the 90 per cent
level, and q5,the 95 per cent level

the distribution of the precipitation itself. We simply assumed that the ratios were (approximate)
normal variables. This assumption is supported by some frequency analyses made on this data set. We
also did not involve the standard deviation, by simply assuming that it is always one-regardless of
possible breaks (equation (2)). These and other assumptions could be criticized but one should then
remember that a very complex test will be hampered in its use by its mere complexity and that it is
quite meaningless to overelaborate the mathematics when we intend to use it on real climate data.

THE SET OF STATIONS AND GAUGE TYPES


The stations used here were chosen from a region in south-western Sweden which comprises a
pronounced spatial maximum of precipitation. Originally 56 stations were used, but here some of the
shortest series were excluded. In Figure 2 the stations that remained after sorting out series shorter
than 10 years are given as numbers (1-44) which can be identified in Table 11. The period analysed was
1912-1981.
In this region of south-western Sweden about 3/4 of the precipitation occurs with winds from the
south, south-west or west and the effects of the gentle hills facing south-west are very pronounced. The
annual mean precipitation rises from about 650-750 mm at coastal sites to above 1100 mm 25-30 km
from the coast. Also, during the warm season convective precipitation frequently forms some distance
from the coast. The precipitation is fairly evenly distributed throughout the year although generally
with a maximum in summer and autumn and a minimum in late winter and early spring.
The gauge used in Sweden nowadays is a can made of an aluminium alloy with a 200 cm2 orifice. It is
equipped with a wind shield. This new gauge replaced the old one which was made of galvanized iron.
In Table I some of the essential events in the history of the precipitation measurements in Sweden
are collected (Anderson, 1970). To this table we should add that all these gauges were equipped with
evaporation shields and that the height of the orifice usually has been 1.5 m above ground, although it
was quite common to place it on roofs and balconies, especially during 1859-1880. The effect of these
changes has been discussed in detail by Anderson (1964) and Eriksson (1979, 1980).
A HOMOGENEITY TEST 665

Figure 2. A map of the region and stations in south-western Sweden which have been used here

Table I. Some notes on precipitation gauges in Sweden


Year Note
~~~ ~

1859 Start of a measuring network of stations.


Collecting area 1206.5 cm'.
1873 A new can with 1000 cmz orifice was introduced.
1894 The wind shield of Swedish type was introduced
at some stations.
1908 The wind shield was gradually introduced at all
stations from this year.
late 1930s A new can with 200 cmz orifice gradually replaced
the old 1000 cmz can.
1958 A new can made of an aluminium alloy gradually replaced
the older ones made of galvanized iron.
666 H. ALEXANDERSSON

THE REFERENCE VALUE


Generally we can write Qi = Ri,df(Ri,j) (8)
where Qi is the ratio in a specific year i, Ri,ois the precipitation at the actual test site (0) and Ri,jis the
precipitation at a reference site j . Then l k
f ( R i , j )= -
k j=1
C
Ri,j (9)

is the simplest form of reference value, an unweighted arithmetic mean value from k reference stations.
In an earlier study (Alexandersson, 1984) an ‘optimal’ technique was outlined which, by using
correlation coefficients between all sites, minimized the coefficient of variation of the sequence of
ratios. This technique has a clear drawback-it becomes very laborious when more than four reference
stations are available. A simpler weighting scheme is used here and is given by

where pi is the correlation coefficient between the test site and the reference site j . It is approximated
by an exponential formula using the distance Zj, i.e.
pi = p*, (11)
For this region d was estimated as 0405 km-l. In Figure 3 an example of a correlation field is given.

~ ~

Figure 3 . The correlation coefficient field versus station no. 18, Havraryd, calculated from the series of annual precipitation
A HOMOGENEITY TEST 667

Here the correlation coefficient is calculated with a station in the middle of the region (no. 18,
Havraryd) as the centre. Note that we have some elongation of the correlation isolines parallel with the
coast, which may be somewhat surprising since the main direction of travelling precipitation systems is
from the south-west (cf. Patrinos et al., 1979).
With the formulations above (equation (8) combined with equation (9) or equation (10)) the
reference sites that can be used must be homogeneous and complete. The latter condition severely
restricts the number of reference sites when long data series are tested. But we can, however, make all
homogeneous series available as reference sites by normalizing with the site mean values, i.e.
Qi = (Ri,dRo)/f(Ri,j/Rj) (12)
Now the function f may include all homogeneous records, since the inclusion of another station will not
be affected by the absolute value at this site. Several tests have shown that this generalization is very
valuable and clearly reduces the scatter in the sequences of ratios. In the test results presented here
three combinations will be used:
(a) Equations (8) and (9)-the original technique which uses arithmetic mean values and the
homogeneous and complete records as reference sites, only 5 in this example.
(b) Equations (8), (9) and (12). Thus all homogeneous records are included by the normalizing
method.
(c) Equations (8), (lo), (11) and (12). As in (b) but weighting factors depending on the distances to
the test site are used.
In the Appendix some complementary formulae are given.

THE RESULTS
The outcome of this test can be illustrated by plots of both the sequences of ratios { q i } and the
sequences of the statistic {z}. As explained in the theory section, the final test statistic is the
maximum value of { T'}. We can use some of the longest series as examples. In these Figures (Figures
4-7) the reference value was obtained in the most satisfactory way (c).
In Figure 4 we note a very distinct break in the series of ratios for station no. 6, Knared. Such a

1'
I I I I 1 I
qi
- KNAREO, 1912 - 8 1 -
+ + +++
1.2 - + +
++ +
+-q2
- -
+ + ++ + + +
+ + + ++
1.0
- - + + +++ + +
+ + +++ +
+ ++ + ++ + +
9, + ; + ++ + ++ ++ +
- +++ ++ +
+ + + ++++ +

0.8 - + To ~ 4 6 . 2* -
Rc =853 m m
- Rep= 819 mm -
Rca= 1016 m m
I I I I 1 I I I
qi - LANGHULT, 1912-81 I -
1.2 1 +
+
+

++
+++
+ ++
I
++++ +
+ +++ +++ +
+ +
+ + +++
+ +
+ + # +
++ +
t

o'81
q,' + +
I ++ + ++ + + + I

+
+ ++

++ TO 19.7 *
i
0.6 R c = 697 mm
+
Rcp: 826 mm
Rca= 947 mm

1910
8 I
1920
I
1930 1940
I I
1950 1960
I t
1970
I
1980

Figure 5. The sequence of ratios for station no. 9, Linghult, from test version (c). See also caption to Figure 4

behaviour of a series indicates that the break is exceedingly significant. In this case it is connected with
a relocation of the site five kilometres northwards; see Figure 2 where the new location is given within
parentheses. Sometimes this may not lead to any serious shift of the mean level, but this station is
situated in a region of large gradients of the mean precipitation. In the next section we will give an
example of how such a dramatic change of about 25 per cent can affect a map of precipitation
anomalies.
For station no. 9, LHnghult, a visual inspection (Figure 5) reveals a very large dip around 1930. There
is also something like a break at the end of the 1960s. According to the tests (a) and (b) this amounts to
+18 per cent. A better estimate of this break is obtained when the dip is excluded and a test on the
latter part of the period (1946-1981) gives the more accurate value +13 per cent at about 1967/68. The
extreme dip cannot be explained by documented events, but the change of the mean level in the late
1960s may be connected with a change of gauge in 1967.
Another set of ratios, and also { T,,}, is given in Figure 6 (no. 41, Varberg). Here the situation is far
from clear. A visual inspection of the sequence of ratios gives an impression of a rather gradual
increase in the later decades and perhaps also near or somewhat before 1930. The rather diffuse
character of the increase in the later decades is supported by the flat character of the peak in the { T y }
series. The gauge at this station has been relocated several times within the city but generally it has
moved inwards from the coast to sites with better shelter from the windy region at the coast. This has
clearly resulted in some increase of the precipitation which may also be partly due to the general
increase towards the north-east and also to changes of the equipment.
Finally in Figure 7 station no. 29, GBngarebo, gives an example of a good series which seems to be
homogeneous. This station has been located at the same cottage for the whole period. There are,
however, also examples of sites which have been relocated several times, e.g. no. 8, Ljungby, without
being non-homogeneous at all according to this table. Here we could mention that a technique used by
Thompson (1984) might have rejected this record as it not only uses the mean precipitation level but
is based on a model of the frequency distribution of the precipitation. It is rather an absolute
homogeneity test (not based on comparisons between nearby stations and therefore not intended to
differ between artificial and real changes of the precipitation climate) and seems to be a bit
oversensitive according to an application on English data (Jones, 1985); see also Thompson and
Revfeim (1985).
A HOMOGENEITY TEST 669

20
T"
I I I I
X
x 1
10

0
qi
1.2
x,x , =$, 1
1 .o
41

0.8
RC = 630mm
Rcp = 606 mm
Rca = 676 mm
I I I I I I I
1910 1920 1930 1940 1950 1960 1970 1980

Figure 6. The sequences of ratios and of T, for station no. 41, Varberg, from test version (c). See also caption to Figure 4

In Table I1 all the stations in Figure 2 are listed and the results of the three versions (a), (b) and (c)
of the test are summarized using Y , the year minus 1900 of the most probable break, and T,, the test
statistic, which is followed by a dot if it is between Tw and G5 and followed by a star if it is above
&-the 95 per cent significance level. Also given is the change, in per cent, of the precipitation
according to the tests. This value is given only if the change is significant according to at least one of the
versions of the test.
One simple way of comparing the effectiveness of the different choices of reference value is to
calculate the coefficient of variation of the sequences of ratios (see the Appendix). In Table I1 the mean
value for all stations is given at the bottom of each column. For the original annual precipitation series
the mean value of the coefficient of variation is 14-72 per cent. The first test (column (a), arithmetic
mean values of the five complete and homogeneous records) gives the mean value 7.28 per cent, the

1.0

OB
+t
I

I++
1910
+
+
+'
+

++

1920
+++

I
++
++
++ ++

1930
+ +

I
+++
++
+

1940
++ +

I
++
+ +
+ +
+
+

1950
I
++

++
+++
+
++

I
1960
+

+
+

+++
+ ++ +
+ +
++

1970
I
To = 1.5
++
+ +++ +

R c = 11OOmm

I
1980
Figure 7. The sequence of ratios for station no. 29, GBngarebo, from test version (c). See also caption to Figure 4
670 H. ALEXANDERSSON

Table 11. Test results with station number, name of the station, the years (-1900) used in this test, the year, Y,
(-1900) with the most probable break, the value of T,, the test statistic, and A, the relative change in per cent. &
is marked with a dot if it is between Tw and T,, and with a star if it is above G5.A is given in parentheses if is
below Tw.Then (a) refers to the 1912-1981 test with arithmetic mean values as reference values (5 stations), (b) to
the 1912-1981 test with individual mean values as reference values (all homogeneous stations for each year) and (c)
to the 1912-1981 test with weighted, individual mean values as reference values (all homogeneous stations for each
Year)

1. Lanshult 12-80 37 27*3* -11 37 24.3* -11 38 2509~ -11


2. Bolmen 23-81 25 8.7* + 15 25 8*9* +15 25 9*3* +15
3. Bolmso 12-81 54 5.1 24 4.4 54 4.3
4. Egernahult 36-65 63 5.9 63 4.6 63 4.2
5. Hokhult 12-24 22 4.8 22 5.2 22 5.3
6. Knared 12-81 69 46-3* +25 69 45.4" +25 69 46*2* +24
7. Lidhult 12-56 30 8.3' +6 30 7.3' +6 30 7.6' +6
8. Ljungby 12-81 75 1.0 75 1.6 75 1.5
9. Ldnghult 12-81 67 20.1* +18 67 19.4* +18 40 19-7* +15
10. Majenfors 48-80 69 6.4 69 5.7 76 5.5
11. Singeshult 34-81 69 8-6* +7 69 8*7* +6 69 8.3* +6
12. Skeen 12-28, 59-77 72 9*0* +9 72 5.6 (+7) 72 5-9 (+7)
13. St. Segerstad 25-29, 45-81 61 4.8 61 5.3 61 5.4
14. Stromsnasbruk 12-80 20 7.5 (-6) 20 7.8' -6 20 7.8' -6
15. Osterang 25-37 25 4.6 25 4.9 20 4.9
16. Genevad 28-81 67 5.8 67 4.9 67 5-3
17. Brunnshult 69-81 69 7.1* +19 69 5.9' +20 69 6-1' +19
18. Havraryd 12-81 70 ll*O* +7 70 11-3* +7 70 11-7* +7
19. Hogaryd 58-68 64 4-0 64 3.9 64 4.4
20. Simldngsdalen 28-81 57 6.1 57 7.3 57 7.2
21. Froslida 52-71 56 5.4 56 3.7 56 3.5
22. Halmstad 12-81 25 6-5 25 6.8 25 7.1
23. Kinnared 12-70 38 7.7' -5 38 7.1 (-5) 38 7.3 (-5)
24. Oskarstrom 21-25, 53-81 23 5.2 (-10) 23 6.9' -11 23 7.1' -11
25. Seglehult 72-81 79 2.1 73 2.6 73 2.6
26. SmIlandsstenar 34-73 53 17-9* +10 53 18-2* +10 53 18-6* +10
27. Torup 45-51, 73-81 75 7.8* -8 75 6-5' -7 75 6.5' -7
28. VLthultsstrom 28-71 66 12.0* +10 63 11.5* +8 64 12.3* +9
29. GIngarebo 12-81 75 1.5 25 1.5 25 1.5
30. Abild 72-81 75 4.7 75 3.5 75 3.5
31. Ambjornarp 43-72 71 2.4 71 2.4 71 2.6
32. Axelfors 14-76 22 8-8* +11 22 7.9' +11 22 8-6' +11
33. Fagered 17-81 48 3.5 21 3-2 48 3.6
34. Hid 68-81 77 5.0 69 4.3 69 4-4
35. Jonstorp 25-81 41 9-8* -5 45 104* -5 45 9.9* -4
36. Mdrdaklev 35-51 39 7*8* + 10 38 7*8* +12 38 9.3* +13
37. Skogsforsen 46-81 77 4.1 67 3.0 67 3.0
38. Glommen 71-81 77 5.6' -11 77 4.5 (-8) 77 4.8 (-8)
39. Morup 57-81 75 4.9 75 4.2 75 3.8
40. Morups TInge 12-62 20 6.0 20 6.2 20 6.1
41. Varberg 12-81 58 17*9* +11 57 19*2* +11 57 19-5* +12
42. Grimeton 12-81 25 21*5* + 12 25 21-5* +12 25 22*2* +12
43. Linhult 12-81 49 0.7 49 1.o 49 1.2
44. Oxaback 12-32 22 1.7 22 1.7 22 3.0
Coefficient of variation (mean value, %) 7.28 7-08 6-86
A HOMOGENEITY TEST 671

second test (column (b), all homogeneous records used, arithmetic mean values) gives 7.08 per cent
and the third test (column (c), as in (b) but a weighting scheme is adopted) gives 6.86 per cent. Thus we
have refined the test stepwise, reducing the relative scatter in the original data from about 15 to below 7
per cent. It can be noted that for a clearly non-homogeneous record it is not possible to reduce the
coefficient of variation to lower values, however good the reference values that are inferred, because of
the break.
Although these versions of the test give very similar results the test technique in general has some
drawbacks:
1. Multiple breaks may remain undetected or obscure the result. Although it may be possible to
detect multiple breaks, visibly, as several peaks reaching above a critical level in the { T,} curve,
the mathematics of the test is in principle strictly applicable only to a single break. The same
comment is valid also for the test used by Potter (1981). Multiple breaks may, although they
obscure the mathematics, be valuable indicators of multiple changes of the measurement
conditions.
2. Really poor series may be accepted as homogeneous just because of a large coefficient of
variation-a large relative spread which includes erroneous values. On the other hand high quality
series can reveal quite small changes (e.g. due to a change of equipment) which are discovered
due to a small relative spread.
3. Trends, dips (cf. Figure 5 ) , etc. are not rightly recognized by this single break test. To handle
trends, dips, multiple breaks etc. quite a complex set of test techniques is required.

CORRECTED CLIMATIC NORMALS


If analyses of precipitation are made in terms of percentages of normal values one should of course use
normals that refer to the present measurement conditions. This cannot always be perfectly fulfilled
owing to the variety of changes that may suddenly occur. But let us take one example referring to the
area considered here.
During 1980 June, November and December were extremely wet in south-west Sweden, giving a
substantial surplus for the year as a whole. In Figure 8 a map with the precipitation for 1980 in southern

I
1
ob 50

120
100
120 120
150hm

1b. 1Lo
1

Figure 8. A map of the precipitation in 1980 as percentage of the average (Monthly and Yearly Summary of Weather and Water
Supply in Sweden, annual, SMHI)
672 H. ALEXANDERSSON

Table 111. The annual precipitation 1980 in mm, in column A as


percentage of an uncorrected average (1912-1981) and in column B
as percentage of a corrected average (1912-1981) after a significant
break. If no significant break has been detected the same figure
appears in the two columns
1980
Station precipitation (mm)
6. Knared 1284 151 126
9. Ldnghult 1237 138 120
10. Majenfors 1114 127 127
14. Stromsnasbruk 947 114 120
16. Genevad 917 125 125
17. Brunnshult 1363 121 120
18. Havraryd 1375 128 121
20. Simldngsdalen 1249 121 121

Sweden as percentage of the average precipitation is given (Monthly and Yearly Summary of Weather
and Water Supply in Sweden, annual). Comparing with Figure 2 we find a very pronounced maximum
(>150 per cent) around station no. 6, Knared. In Table 3 we have used both uncorrected values
(1912-1981) and values corrected (1912-1981) to the level after a significant break to calculate the
percentage precipitation (test (c)) for the stations nearest to Knared. For Linghult the dip (discussed
earlier) is excluded to obtain better estimates. For Station no. 14, Stromsnasbruk, a trend correction is
made, as a visual inspection shows a clear and gentle decrease during the whole period (Alexandersson,
1984). A slow trend is not properly handled by this test.
Thus we have reduced the anomalies from the wide range 114-151 per cent to the more plausible
range 120-127 and within this limited region. The erroneous or at least much exaggerated maximum
was mainly caused by the well marked discontinuity of the Knared series. To use up-to-date climatic
normals thus makes it both more correct and easier to draw maps of anomalies.
The precise technique used here to estimate missing values and to obtain current climatic normals
that refer to some common period is a bit awkward in its details but is briefly summarized in the
Appendix.
OTHER APPLICATIONS
Monthly data have also been tested in the same manner. Then all months have been put together in a
sequence of 12n values, n being the number of years. Generally these tests support the results obtained
using annual values, although somewhat more significant breaks are obtained. In some cases false
breaks may occur using monthly data because of dry months giving very uncertain ratios which distort
the test. In fact a very dry month should rather be omitted or added to the next month to avoid such
difficulties.
Instead of testing all months in a long sequence one may pick out one month at a time and test it.
The result of such a test is not very encouraging (breaks are often a bit differently dated), which also
shows that individual corrections for separate months are not easy to derive with sufficient accuracy.
One reason for this could be that sometimes several say April months can be very dry, giving uncertain
ratios. It is unfortunate that tests on separate months are difficult to perform because in high latitudes it
is quite common that changes are most pronounced in winter months due to the difficulties in
measuring snow. But, for the time being, it is recommended to use the same correction factor for all
months.
Alternatively the homogeneity test can be applied on annual sums but then individual, monthly
correction factors (&/ijl, see Appendix) can be calculated separately using the date of the break from
the annual test. Using this method, i.e. fming the break at the year given by the test using annual sums
and then calculating correction factors for individual months, we obtain the figures in Table Iv for two
A HOMOGENEITY' TEST 673

Table IV. Correction factors for individual months where the break is fixed at the year given by
the test using annual sums
Station J F M A M J J A S O N D
6. Knared 1.24 1.18 1.32 1.17 1.11 1.28 1-26 1-19 1-10 1-31 1.29 1.28
41. Varberg 1.20 1-12 1-08 1.18 1-13 1-11 1-00 1-25 1-06 1.15 1-10 1-11

stations. This example indicates that individual factors are randomly scattered around the annual means
(1.24 for Knared and 1.12 for Varberg, see Table 11). Individual monthly correction factors seem to
introduce uncertainty, but as a rule without giving especially valuable information about seasonal
differences.
All data series which can be considered approximately Gaussian-r transformed to approximately
Gaussian--can be tested in the way the ratios were tested here. Let us say that we have a series of
precipitation values from a station which we know is homogeneous relative to the surrounding stations.
Then we may use this test to investigate climate changes. As absolute precipitation series, especially if
monthly values are used, are positively skewed a transformation should be performed.
One suggestion is to use square roots. In fact the natural and simple compound Poisson-exponential
model discussed and used by some writers (e.g. Le Cam, 1961; Buishand, 1977; Ozturk, 1981;
(Alexandersson, 1985) has an approximate normal transformation which is essentially a square root.
Tests made by the author (Alexandersson, 1985) revealed no significant changes of annual precipitation
values within this region. Considering monthly data July and August seem to have been drier since the
1960s whereas in the same period October-December have been unusually wet. The same tendency for
summer months is apparent in an English record (Woodley, 1985). Both regions (south-western
England and south-western Sweden) obtain most of their precipitation when the flow is rather strong
and zonal. The very latest years not included here have somewhat changed this pattern of dry summers.

SUMMARY
A simple homogeneity test is explained and used on a precipitation data set from south-western
Sweden. The test is of a very elementary form from a statistical point of view. There is, however, such
a variety of possible inhomogeneities in precipitation data that it does not seem very fruitful to use
more advanced statistical techniques. To test the relative homogeneity it is important to obtain as many
and as good reference sites around the test site as possible. This optimization of the available records
seems quite important and a technique to incorporate incomplete series also, is explained and found
very valuable. The basic principles of testing the relative homogeneity can be used also to test the
absolute homogeneity, i.e. possible climatic changes. Earlier tests of homogeneity usually give neither
the date of a significant break nor the amount of the change. Thus the stringent type of test used here
can be applied frequently, e.g. to check the relative homogeneity of an existing climatic data network
continuously. In practice it should be used along with careful notations about changes of instrumenta-
tion, station moves and so on.
ACKNOWLEDGEMENTS

The author wishes to thank Prof. Bengt RosCn and Prof. Ulf Hogstrom for valuable comments and
Marianne Bjorklund and Aina Ekstrom for manuscript preparation. The author would also like to
thank the editor and the referees for valuable suggestions and corrections.

APPENDIX: SOME FORMULAE


As explained in the Section on the reference value, an effective way of calculating ratios and reference
values (method (c)) was given by
Qi = (Ri,o/Ro)/f(Ri,j/Rj) (13)
674 H. ALEXANDERSSON

pi = e-d‘, (15)
where the index i denotes year (or other time unit) and j denotes the station. Rewriting (13) and (14)
we obtain

Then the denominator can be interpreted as an estimate of R,,-the value at the actual site -using the
values from the surrounding stations. Thus, ideally, the ratio for an individual year equals one. The
mean value of the { q i } series will always equal one. Let us write Zi,ofor this denominator. Then let us
assume that the actual series is considered homogeneous but that it is not complete. Then Zi,ocan be
used to fill out the missing values.
Next we assume that the actual series is complete but we want to correct the earlier years for a
significant non-homogeneity. Then we have two mean values of the series of ratio-, before the
break and (r2 afterwards. To homogenize the series we should then use
Ri,o = (q21gl)Ri,O (17)
Then the new series (with primed values before the break) will be reduced to the present mean value.
Using the inverse factor in equation (17) the values after the break can be reduced to the former mean
level.
Finally, if a series is both incomplete and includes a significant break, we may use

to obtain reasonable estimates at the actual site reduced to the present mean level during the period
when it was not in operation.
In more complex situations with apparent multiple breaks, trends or dips of the type shown in Figure
5 , a subjectively drawn curve adapted to the sequence of ratios may be used to make the series
homogeneous. But then we are back to a subjective interpretation.
The mean values of ratios used in these formulae, i.e. q1 and q2, can be calculated as arithmetic
mean values but also, in line with standard methods using proportionals, as

This gives larger weights to the largest numbers (the wettest years) than does the arithmetic mean value

The author prefers equation (19) but in practice the difference is negligible using annual precipitation
data.
In the results section the coefficient of variation was used to compare the effectiveness of the different
tests. This parameter is useful for positive variables-such as precipitation and the ratios defined here.
Let us define two variables X and Y with, for convenience, equal mean value (E) and equal variance
(V) and a correlation between them given by the correlation coefficient p. Then, using Gauss’s
approximate formula for the variance of a function, we obtain for the coefficient of variation of the
ratio YIX
Cv(Y1X)= c u m
Here Cv = * / E on the right side. As expected the coefficient of variation of the ratio YIX diminishes
when the correlation between X and Y increases. Thus the coefficient of variation can be used as a
convenient, alternative measure of the goodness of fit here, dealing with positive variables.
A HOMOGENEITY TEST 675

REFERENCES
Alexandersson, H. 1984. ‘A homogeneity test based on ratios and applied .- to precipitation
. . series’, Reports No. 79, Dept. of
Meteorol., Uppsala, Sweden, p. $5.
Alexandersson, H. 1985. ‘A simple stochastic model of the precipitation
- - process’, J. Clim. Appl. Meteorol., 24, 1285.
Anderson, T. ‘1964. ‘Further stidies on the accuracy of rain measurements’, Arkiu f o r Geofysik, 4, (16), 359.
Andersson, T. 1970. ‘Swedish temperature and precipitation records since the middle of the 19th century’, Document 0 4 : 1970,
National Swedish Building Research, p. 167.
Brooks, C. E. P. and Carruthers, N. 1953. Handbook of Statistical Methods in Meteorology, Met. Office, London, p. 412.
Buishand, T. A. 1977. ‘Stochastic modelling of daily rainfall sequences’, Mededelingen, Agricultural University, Wageningen,
The Netherlands, p. 211.
Buishand, T. A. 1984. ‘Tests for detecting a shift in the mean of hydrological time series’, J. Hydrol., 73, 51.
Conrad, V. and Pollak, L. W. 1950. Methods in Climatology, 2nd edn, Harvard University Press, Mass., p. 459.
Eriksson, B. 1979. ‘Statistical analysis of precipitation data. Part I. Area precipitation’, Reports, RMK 16, SMHI, Norrkoping,
Sweden, p. 73.
Eriksson, B. 1980. ‘Statistical analysis of precipitation data. Part 11. Frequency analysis of monthly precipitation data’, Reports,
RMK 17, SMHI, Norrkoping, Sweden, p. 150.
Gardner, L. A. 1969. ‘On detecting changes in the mean normal variates’, Ann. Math. Stat., 40, (l), 116.
Hawkins, P. M. 1977. ‘Testing a sequence of observations for a shift in location’, J. Amer. Statist. Assoc., 72, (357), 180.
Jones, P. D. 1985. ‘Comment on “Homogeneity analysis of rainfall series: an application of the use of a realistic rainfall model”,
J . Climatol., 5 , 337.
Le Cam, L. 1961. ‘A stochastic description of precipitation’, in Neymen, J. (ed.), Proc. 4th Berkley Symp. on Mathematical
Statistics and Probability, University of California Press.
Lindgren, B. W. 1968. Statistical Theory, 2nd edn, The Macmillan Company, London, p. 521.
Maronna, R. and Yohai, V. J. 1978. ‘A bivariate test for the detection of a systematic change in mean’, J . Amer. Statist. Assoc.,
73, 740.
Monthly and Yearly Summary of Weather and Water Supply in Sweden, Annual, SMHI, Norrkoping, Sweden.
Oztiirk, A. 1981. ‘On the study of a probability distribution for precipitation totals’, J . Appl. Meteorol., 20, 1499.
Patrinos, A. A. N., Chen, N. C. J. and Miller, R. L. 1979. ‘Spatial correlations of monthly rainfall: applications in climatology
and weather modification experiments’, J. Appl. Met., 18, 719.
Potter, K. W. 1981. ‘Illustration of a new test for detecting a shift in mean in precipitation series’, Mon. Wea. Rev., 109, 2040.
Segen, J. and Sanderson, A. 1980. ‘Detecting change in a time series’. ZEEE Trans. Znf. Theory, IT-26,(2), 249.
Thompson, C. S. 1984. ‘Homogeneity analysis of rainfall series: an application of the use of a realistic rainfall model’, J .
Climatol. 4, 609.
Thompson, C. S. and Revfeim, K. J. A. 1985. ‘Comments on “Homogeneity analysis of rainfall series: an application of the use
of realistic rainfall model”, J. Climatol. 5 , 579.
Willsky, A. S. and Jones, H. L. 1976. ‘A generalized likelihood ratio approach to the detection and estimation of jumps in linear
systems’, ZEEE Trans. Automat. Control, AC-21, (l), 108.
Woodley, M. R. 1985. ‘Recent changes in summer rainfall values measured at Long Ashton’, Weather, 40, (6), 180.

You might also like