964 VOLUME 14J O U R N A L O F C L I M A T E

q 2001 American Meteorological Society

Quality Control and Homogeneity of Precipitation Data in the Southwest of Europe

J. FIDEL GONZÁLEZ-ROUCO

Departamento de Astrofı́sica y Ciencias de la Atmósfera, UCM, Madrid, Spain, and
GKSS Forschungzentrum, Geesthacht, Germany

J. LUIS JIMÉNEZ

Departamento de Astrofı́sica y Ciencias de la Atmósfera, UCM, Madrid, Spain

VICENTE QUESADA

Departamento de Estadı́stica e Investigación Operativa, UCM, Madrid, Spain

FRANCISCO VALERO

Departmento de Astrofı́sica y Ciencias de la Atmósfera, UCM, Madrid, Spain

(Manuscript received 19 July 1999, in final form 19 April 2000)

ABSTRACT

A quality control process involving outliers processing, homogenization, and interpolation has been applied
to 95 monthly precipitation series in the Iberian Peninsula, southern France, and northern Africa during the
period 1899–1989. A detailed description of the procedure results is provided and the impact of adjustments
on trend estimation is discussed.

Outliers have been censored by trimming extreme values. Homogeneity adjustments have been developed by
applying the Standard Normal Homogeneity Test in combination with an objective methodology to select ref-
erence series.

The spatial distribution of outliers indicates that they are due to climate variability rather than measurement
errors. After carrying out the homogeneity procedure, 40% of the series were found to be homogeneous, 49.5%
became homogeneous after one adjustment, and 9.5% after two adjustments. About 30% of the inhomogeneities
could be traced to information in the scarce history files.

It is shown that these data present severe homogeneity problems and that applying outliers and homogeneity
adjustments greatly changes the patterns of trends for this area.

1. Introduction

In the last few decades there has been an increasing
concern regarding the changes in climate caused by an-
thropogenic emissions of greenhouse gases and aerosols
(Houghton et al. 1992, 1995). This interest has been
mainly based on the results of climate models. In a
similar fashion, there has been a rise in the activity
directed toward climate change detection in observed
data (Santer et al. 1995; Hegerl et al. 1996). It is es-
sential, for this purpose, to use high quality observa-
tions. Thus outliers and homogenization arise as im-
portant issues.

Corresponding author address: J. Fidel González-Rouco, GKSS
Forschungzentrum, GUA, GEB. 38, Max-Planck-Str. 1, 21502 Gees-
thacht, Germany.
E-mail: fidel@gkss.de

Historically, the identification of outliers has been the
primary emphasis of quality control work (Filippov
1968; Grant and Leavenworth 1972). Outliers are ob-
servations very distant from the mean value that can be
due to measurement errors or to extreme meteorological
events. Several approaches that focus on temporal and/
or spatial variability can be applied in order to identify
outliers and diagnose whether they are erroneous or not
(Barnett and Lewis 1994; Eischeid et al. 1995; Peterson
et al. 1998a). When outliers are undoubtedly erroneous
measurements those extreme data can be rejected and
the problem is converted into one of missing data treat-
ment. When outliers have a physical background the
question arises whether they should be corrected or not
(Barnett and Lewis 1994). On the one hand, extreme
data carry very valuable climatological information that
should not be dismissed. On the other hand, many sta-
tistical techniques of common use are not resistant, that
is, they are sensitive to the presence of outliers that can



1 MARCH 2001 965G O N Z Á L E Z - R O U C O E T A L .

affect the estimation of sample statistics (Lanzante
1996). Such is the case of all moments-based statistics
(Peterson et al. 1998a).

The last issue can be approached in several ways. The
most objective approach would be always to use tech-
niques that are resistant (Lanzante 1996). Unfortunately
this is not the case for many of the available methods.
An alternative and rather radical solution would be re-
moving all outliers. This would allow a safe use of
nonresistant statistics at the expense of losing interesting
climatological information. Another alternative that
would combine keeping the information of extreme data
and showing consideration for nonresistant methods
would be to censor outliers by means of replacing them
by some threshold value that keeps the information of
an extreme event and yet does not have such an im-
portant influence on nonresistant statistics (Barnett and
Lewis 1994). The original value of the outliers can be
later restored for specific studies concerning extreme
values. This last subjective approach has been the one
used herein as a preprocessing for the nonresistant ho-
mogenization techniques employed in this work.

A homogeneous climate series is defined as one where
variations are caused only by changes in weather and
climate (Conrad and Pollak 1950). This would mean
that the measurements have been consistently registered
by the same method, with the same instrumentation, at
the same time of day and place and in the same envi-
ronment. Unfortunately, this is seldom the case and at
most old observatories, instruments, location, observer,
or other environmental factors have been altered (Jones
et al. 1985; Karl and Williams 1987; Peterson et al.
1998a). In the case of precipitation, progressive im-
provements of instrumentation can also introduce arti-
ficial systematic increases (Hanssen-Bauer and Førland
1994), and thus long-term variations and trends should
be interpreted cautiously (Houghton et al. 1992). Ac-
curate long-term climate analysis requires homogeneous
data, thus, inhomogeneities must be detected and in-
homogeneous series be excluded or adjusted.

Several statistical tests have been developed that al-
low detection of inhomogeneities in time series. Some
depend on the metadata (Karl and Williams 1987;
Rhoades and Salinger 1993) and others use them as
additional information when the station history is poorly
documented or missing at all (Buisband 1982; Alex-
anderson 1986). Some tests are univariant, while others
use information from reference time series that are com-
pared with the test series to decide upon its quality. A
comprehensive review on this topic has been presented
by Peterson et al. (1998b).

Considerable research has been accomplished on the
characteristics of low-frequency climate variability of
precipitation in the western Mediterranean (Maheras
1988; Valero et al. 1993; Zorita et al. 1992; Fernández-
Mills 1995; Valero et al. 1996b; Rodriguez-Puebla et
al. 1998; Serrano et al. 1999; González-Rouco et al.
2000). However, homogeneity assessments of historical

precipitation data during the twentieth century have sel-
dom been performed (Almarza et al. 1996; Garrido et
al. 1996; Esteban-Parra et al. 1998). Thus, a detailed
description of this problem from the perspective of an
objective methodology using all the information from
a high number of stations is still to come.

In this work, we focus on the detection and treatment
of outliers and inhomogeneities in a dataset of precip-
itation in the southwest of Europe. The adjustment of
outliers has been performed by taking into account dis-
persion in the time series by means of windsorization,
that is, trimming the values that exceed a certain thresh-
old (Barnett and Lewis 1994). For detection and pro-
cessing of inhomogeneities we have used a procedure
based upon the Standard Normal Homogeneity Test de-
veloped by Alexanderson (1986). This method assumes
invariance of the ratio of the values of the test series
and those of a reference series built using information
from several neighboring stations. An inhomogeneity in
the test series will be revealed by a systematic change
in this ratio. However, if there exists no a priori knowl-
edge of which ones are the homogeneous time series,
the test has to be applied several times to decide which
stations have good quality records and can be used as
a reference. We have followed a procedure divided in
six steps that allows identifying reference series and
adjusting inhomogeneous ones. The first five steps fol-
low those of Hanssen-Bauer and Førland (1994) and the
sixth one has been added to account for series with two
inhomogeneities.

It is worthwhile making a comment on the limits of
the corrections applied in this work. The procedures
herein are applied with the intention of improving the
quality of the data, but not of providing an error-free
dataset. The use of the term ‘‘correct’’ should not be
taken to imply a level of perfection in the adjustments
that does not exist in any dataset. Future inclusion of
new data and/or the use of other methods will hopefully
lead to further improvements.

Data are described in the next section. A description
of the methodology is presented in section 3. Results
concerning outliers and homogenization are presented
in the following section as well as a description of their
impact on trends. Finally conclusions are summed up
and discussed.

2. Data

We have used long time series of monthly precipi-
tation totals (mm) covering most of the Iberian Penin-
sula, south of France and north of Africa and spanning
the period 1899–1989. Iberian rain gauge data were
provided by the Instituto Nacional de Meteorologia
(Spain). The Carbon Dioxide Information Analysis Cen-
ter (United States) and Metéofrance (France) supplied
African and French precipitation data, respectively. Fig-
ure 1 shows the spatial distribution of the 95 observa-
tories; the corresponding names and lengths of the as-



966 VOLUME 14J O U R N A L O F C L I M A T E

FIG. 1. Spatial distribution of precipitation stations. Numbers refer
to those of Table 1. Circles point time series in which information
from secondary sites was used.

sociated time series are listed in Table 1. We have se-
lected the period 1899–1989 for this first version of the
data.

Noncircled numbers in Fig. 1 stand for sites in which
time series were used unchanged to apply the procedures
described in the next section. In some cases (circled
identifiers in Fig. 1) long data gaps were present and/
or temporal extension of the time series was desirable.
In order to improve the quality of these records and also
the spatial coverage of the dataset additional information
from 43 secondary sites (nonnumbered rows in Table
1) was used. Thus, initial corrections following the ratio
method (Mitchell et al. 1966) were calculated and ap-
plied using data from the overlapping periods between
the base stations and the additional observatories. The
lower threshold value for correlation between the base
station time series and the time series of a candidate
secondary site to be used for data interpolation was 0.8.

3. Methodology

In this section a brief description is provided on the
development of outliers and homogeneity corrections.

a. Outliers

Historically, the identification of outliers has been the
primary emphasis of quality control work (Grant and
Leavenworth 1972). Outliers can be erroneous mea-
surements or correct extreme values. In this version of
the dataset the approach has not been searching for er-
roneous measurements, but reducing the size of the dis-
tribution tails in order to make a safer use of the non-
resistant homogenization techniques used later. This has
been performed by trimming extreme data by a thresh-
old value. This extreme value keeps the information of
an extreme event, but its reduced magnitude in com-
parison with the original value has less impact on non-
resistant statistics (Lanzante 1996). In the next section

the spatial distribution of values identified as outliers
will be described showing that in most cases the
trimmed values have a physical meaning.

Outliers were identified as those values trespassing a
maximum threshold for each time series (Trenberth and
Paolino 1980; Peterson et al. 1998a) defined by

Pout 5 q0.75 1 3IQR, (1)

where q0.75 is the third quartile and IQR the interquartilic
range. The IQR has been used in quality control of
climate data (Eischeid et al. 1995) because it is resistant
to outliers. Values over Pout were substituted by this
limit. This way of proceeding reduces the bias caused
by outliers and yet keeps the information of extreme
events (Barnett and Lewis 1994).

b. Homogenization

1) THE STANDARD NORMAL HOMOGENEITY TEST

The homogenization procedure is based on the ap-
plication of the Standard Normal Homogeneity Test (Al-
exanderson 1986). This test is based on the assumption
that precipitation amounts at the station being tested
(test station) and some regional average values are pro-
portional to each other. This relationship is expressed
in terms of the ratio q between the test station normal-
ized precipitation values and those of a regional time
series defined as a weighted average of several neigh-
boring reference stations. Thus, the ratio q in a specific
year may be denoted as

Fiq 5 , i 5 1, . . . , n, (2)i Gi

n being the number of time steps, and Fi and Gi functions
of the precipitation at the test and reference stations that
are defined as

ki Qij
yO j QP j51 jiF 5 G 5 , (3)i i kiP

yO j
j51

Pi being the precipitation at the test station, Qij the pre-
cipitation at the jth reference station, and y j a weight
factor for the jth reference station. In this work y j was
defined as the square of the correlation coefficient be-
tween the test series and the jth reference series. Here
ki is the number of reference sites used in time step i
that varies as a result of different length of time series
and the existence of missing data. Overbars denote av-
erages throughout the observational period.

In order to apply the test the qi values are standardized
by their sample mean, q , and standard deviation, sq:

(q 2 q )iz 5 .i sq

This series has zero mean and unit standard deviation



1 MARCH 2001 967G O N Z Á L E Z - R O U C O E T A L .

and is assumed to be normally distributed, N(0, 1). An
inhomogeneity in the test series would affect the series
of ratios and thus zi. Therefore, the homogeneity test is
based on the definition of the following hypotheses.

Null hypothesis, (H0): The test series is homogeneous.
Any subset of zi is distributed as N(0, 1).

Alternative hypothesis, (H1): The test series is inho-
mogeneous. There is an inhomogeneity in year m such
that the m first years of the zi have mean value m1 and
the (n 2 m) last years have mean value m2, with the
standard deviation 1 in both temporal segments.

These definitions enable the test to detect and adjust
only one inhomogeneity in the original test series. A
test statistic, Tm, is computed for each of the n 2 1
possible change points:

Tm 5 m 1 (n 2 m) , m 5 1, . . . , n 2 1,2 2z z1 2 (4)

where z 1 and z 2 are the mean values of zi during the m
first and the (n 2 m) last years, respectively. A high T
value in year m implies that m1 and m2 depart signifi-
cantly from zero, making H0 unlikely.

The maximum Tm value in the time series will be
denoted by Tx. The probability that Tx exceeds a certain
value if H0 is correct depends only on the length of the
time series. The critical T values for the 5% and 10%
significance levels (T95 and T90, respectively) are given
by Alexanderson (1986). In practice, the test is applied
making use also of the existing information about the
history of the station, the metadata. A precipitation se-
ries is then classified as inhomogeneous when (i) it con-
tains an inhomogeneity significant at the 5% level 5 yr
or more from either end of the series, and (ii) the series
contains an inhomogeneity significant at the 10% level,
which is explainable by metadata.

The interaction with the metadata information at the
10% level reduces the probability of rejecting homo-
geneous series, thus increasing the power of the test.
The reason for not accepting unexplained inhomoge-
neities close to the ends of the series is an increased
probability for high T values near the ends (Hawkins
1977).

Once a series is identified as inhomogeneous, the data
before the inhomogeneity date are corrected through
multiplying them by

qaf 5 , (5)
qb

where qa and qb are the mean values of qi after and
before the inhomogeneity. When the date of a change
in the historical records was close to the year corre-
sponding to Tx, the historical date was used for the
correction. Otherwise, the year corresponding to Tz was
assumed.

Equations (2) and (3) can also be used for interpo-
lating missing data (Alexanderson 1986; Valero et al.
1997a). Assuming q(t) . q 5 1 for missing data, the
estimations of the missing Fi values, F̂i, can be obtained
from

F̂i 5 qiGi . q Gi 5 Gi, (6)

and thus precipitation for the test station can be esti-
mated from

P̂i 5 PGi. (7)

2) SELECTION OF REFERENCE STATIONS

The above description states how factors can be cal-
culated to produce an homogeneous test series by using
information for several reference stations. However,
usually no a priori knowledge exists on which ones
should form the reference series and the test has to be
run in an iterative way several times to decide upon
which are the best quality series. We have followed the
same procedure as Hanssen-Bauer and Forland (1994)
for this purpose. We have added one step to their pro-
cedure in order to account also for series with two in-
homogeneities. Figure 2 highlights the steps of the meth-
od. The three first stages are essentially devoted to im-
prove the selection of reference stations:

Step 1: A first homogeneity assessment. In this step
the test is applied to each time series. The reference
series will be those that correlate best with the test sta-
tion. Since reference stations were not necessarily ho-
mogeneous, the result is a first approach to make a clas-
sification of homogeneous (H1; see Fig. 2) and inho-
mogeneous (I1) series.

Step 2: Using adjusted reference series. Corrections
are made to the I1 set following Eq. (5). Subsequently
the test is applied to each one of the original time series
again though using an improved dataset as reference
series: the H1 group and the corrected I1 series. Since
this correction reduces the amount of inhomogeneities
in the reference set this produces a new, somewhat op-
timized, classification into homogeneous (H2) and in-
homogeneous (I2) series. The only difference with step
1 is using adjusted reference series.

Step 3: Applying the test to corrected series. The I2
group is corrected and then the test is applied only to
these adjusted time series; the pool of possible reference
stations is built with the H2 set and the adjusted I2
stations. The result is a group of homogeneous time
series that have been corrected once (HC3) and a subset
of inhomogeneous series with more than one inhomo-
geneity (IC3).

At this stage of the procedure all time series can be
grouped into an homogeneous group (H2 and HC3) and
an inhomogeneous one (IC3). The next two steps revise
this classification by using as a reference only time se-
ries that have proved to be homogeneous after applying
the test.

Step 4: Testing again using only homogeneous series.
All series are checked again: the homogeneous H2 group
and the corrected I2 group. The reference series are only
those that have proved their homogeneity so far, namely,
the H2 and HC3 sets. The H2 set provides two new
subsets after applying the test, a group of homogeneous



968 VOLUME 14J O U R N A L O F C L I M A T E

TABLE 1. Names and periods of the time series used in the com-
pilation of the precipitation dataset. Numbers match those in Fig. 1.
Correlation values between main sites and additional secondary ob-
servatories are also shown.

No. Name Period
Correlation

values

1 Lyon 1841–1994
2 Grenoble 1833–1976
3 Bourdeaux 1899–1994
4 Gap 1899–1992
5 Bagnols Les Bains 1899–1993
6 Joyeuse 1900–93

Service des crues 1900–30 0.98
7 Cahors 1899–1993
8 Avignon 1899–1993
9 Saint Server 1899–1993

10 Nice 1899–1993
Rue Gioffredo 1891–1920 0.90

11 Toulouse 1899–1994
12 Peirehorade 1851–1992
13 Montpellier 1899–1993
14 Gijon 1913–94

Cabo Peñas 1914–22 0.76
15 Castropol 1907–94

Tapia de Casariego 1907–25 0.80
Castropol 1923–72 0.95

16 Santander 1912–91
17 Marselle 1899–1994
18 Revel 1899–1973
19 Coruña 1899–1993

Bilbao sondica 1899–1993
Guernica 1925–30 0.86
Larrasquitu 1925–46 0.86
Machichaco 1913–30 0.77
Basauri 1940–50 0.93

21 San sebastian 1899–1993
22 Toulon 1899–1994
23 Lugo 1913–94

Mundı́n 1913–21 0.92
Brigós 1922–39 0.94
Sarria 1952–66 0.91

24 Reinosa 1911–91
25 Camporredondo 1917–92
26 Santiago 1906–93

Universidad 1899–1950 0.89
27 Cerv. pisuerga 1912–86
28 Pamplona 1899–1993
29 Canfranc 1910–94
30 Perpignan 1899–1993

Pont Rouge 1899–1923 0.98
31 Oña 1900–94
32 Leon 1904–93
33 Capdella 1915–91
34 Logroño 1911–94

Instituto 1911–47 0.85
35 Pontevedra 1899–1994

Salcedo 1949–85 0.96
36 Burgos 1899–1991
37 Orense (Granja) 1899–1994

Santuario Molas 1924–25 0.90
Beariz 1926–35 0.87
Villamayor 1937–50 0.90
Ginzo 1969–72 0.91

38 Huesca 1899–1994
39 Palencia 1913–88
40 Gerona 1911–91

Instituto 1911–69 0.89
41 Soria 1899–1994
42 Zaragoza 1899–1994

TABLE 1. (Continued)

No. Name Period
Correlation

values

Facultad de ciencias 1899–1941 0.89
43 Valladolid 1899–1994
44 Lerida 1913–91
45 Vid de aranda 1899–1994

Col. Franciscanos 1914–39 0.85
Aranda del Duero 1943–94 0.84

46 Zamora 1909–94
47 Barcelona 1899–1994

Universidad 1899–1937 0.91
48 Daroca 1899–1993
49 Segovia 1899–1994
50 Salamanca 1899–1993

Observatorio 1899–1944 0.94
51 Tortosa 1899–1991
52 Avila 1900–94
53 Guadalajara 1912–91

Atienza 1937–45 0.81
54 Madrid 1899–1994
55 Teruel 1899–1994

Ariño 1936–43 0.90
56 Hervas 1913–94

Bejar 1913–79 0.90
Aldea del Camino 1951–54 0.90

57 Coimbra 1899–1989
58 Cuenca 1908–90
59 Castellón 1911–93
60 Talavera 1913–90

Candeleda 1936–40 0.90
Robledillo de la Vera 1940–42 0.91
Marrupe 1944–47 0.87
Alcañizo 1959–62 0.96
Cabañuelas 1989–90 0.85

61 Mahón 1899–1994
62 Toledo 1908–91
63 Palma mallorca 1899–1993
64 Caceres 1904–91
65 Valencia 1899–1994
66 Ciudad real 1904–91
67 Albacete 1899–1991
68 Badajoz 1899–1994
69 Alange 1913–94

Valencia del Ventoso 1913–34 0.89
Miajadas 1917–50 0.91

70 Almaden 1913–94
Pozoblanco 1936–94 0.84

71 Lisboa 1899–1989
72 Alicante 1899–1993
73 Cehegim 1913–94
74 Murcia 1899–1993
75 Cazorla 1911–94
76 Aracena 1913–94

Galaroza 1960–94 0.95
Valdeazufre 1964–94 0.96

77 Cordoba 1911–94
78 Jaen 1899–1993
79 Huelva 1903–94
80 Granada 1899–1993

Alhendim 1977–90 0.93
81 Sevilla 1899–1994
82 Almeria 1910–94

Estación sismológica 1909–32 0.89
Ciudad Jardı́n 1933–67 0.89

83 Grazalema 1912–94
84 Malaga 1900–95
85 Dar-el-beida 1838–1994
86 Algiers 1889–1994



1 MARCH 2001 969G O N Z Á L E Z - R O U C O E T A L .

TABLE 1. (Continued)

No. Name Period
Correlation

values

87 San Fernando 1899–1989
88 Constantinopla 1899–1994
89 Gibraltar 1899–1994
90 Tanger 1899–1989
91 Oran 1899–1994
92 Tlemcem 1899–1989
93 Biskra 1902–94
94 Djelfa 1899–1994
95 Casablanca 1902–89

FIG. 2. Diagram showing the six steps of the homogenization procedure [after Hanssen-Bauer and Førland (1994)]. Test
series are tested by comparing with reference series and sorted into output groups at each step of the procedure for
subsequent use. The output groups are identified by the result of the test (H for homogeneous and I for inhomogeneous),
by the number of corrections applied to the input test series (C for one correction and CC for two corrections) and by the
number of the step in which the test is performed. Independently of this convention symbols indicate the number of
corrections applied to any input series.

series (H4) and a group of series that contain at least
one inhomogeneity (I4). The adjusted I2 is divided into
a group of homogeneous series (HC4) and a set of series
that have at least two inhomogeneities (IC4).

Step 5: Last corrections for one inhomogeneity. The
inhomogeneous group I4 is corrected and checked for
new inhomogeneities. The reference series are taken
from the just updated homogeneous (H4) and homo-
geneous after one correction (HC4) sets.

In steps 3–5 corrections for series with one inho-

mogeneity are applied. Step 6 has been incorporated to
correct two inhomogeneities.

Step 6: Correcting two inhomogeneities. The test is
applied to parts of the series containing only one in-
homogeneity. First, the data before the first inhomo-
geneity in series belonging to the groups IC4 and IC5
are omitted and the remaining data are corrected using
as reference series those classified as homogeneous in
the five preceding steps (H4, HC4, HC5); second, the
test is run with all the data in order to correct the first
inhomogeneity.

4. Results

a. Distribution of outliers

Figure 3a shows the distribution of Pout values [Eq.
(1)]. This plot describes the variability of the data that
reaches maximum values in the northwest and south of
the Iberian Peninsula and most of the Pyrenees. Data
dispersion achieves its lowest values in the central Pla-
teau and along the Mediterranean coast where the fre-
quency of extreme events is higher (Linés 1970; Valero



970 VOLUME 14J O U R N A L O F C L I M A T E

FIG. 3. (a) Spatial distribution of Pout (mm) values. (b) Spatial
distribution of the absolute number of outliers in each station. Contour
interval is 100 mm. (c) Time evolution of the monthly absolute num-
ber of outliers in the area of interest. Contour interval is 5 units.

TABLE 2. Number and type of series involved in each step of the
homogenization procedure.

Step

Input series

Test
Refer-
ence

Output series

H HC HCC I IC ICC

1
2
3
4
5
6

95
95
53
95

4
10

95
95
95
87
82
85

43
42
—
38
—
—

—
—
45
44

3
—

—
—
—
—
—
9

52
53
—
4

—
—

—
—
8
9
1

—

—
—
—
—
—
1

et al. 1997b). This agrees with the distribution of outliers
in Fig. 3b that shows a clear maximum in the Mediter-
ranean area of influence.

The total number of corrected values was 519: 12%
of them were produced in spring, 3% in summer, 52%
in autumn, and 33% in winter. Figure 3c shows the
corresponding monthly distribution, characterized by a
maximum in October and a minimum in the summer.

The spatial distributions of the number of outliers for
each season (not shown here) reveals that maxima are
located in the east coast in spring and autumn in a similar

fashion as Fig. 3b. In winter they spread over the west
and south of the Iberian Peninsula and for the summer
months there is no apparent structure but isolated cases
in the central plateau.

The overall agreement of the areas and timing of out-
liers with those of extreme events supports the idea of
physical mechanisms as causes for the detected anom-
alous data rather than human-induced errors.

b. Homogenization

For this version of the dataset, homogeneity correc-
tions were developed from the annually averaged time
series and subsequently applied to monthly data. Month-
ly corrections can also be incorporated through aligning
the whole series as a sequence of 12n values, n being
the number of years or through considering individual
monthly series (Alexanderson 1986). In the former case,
results support those obtained using annual values, al-
though somewhat more significant breaks are obtained.
Also, there are cases of dry months giving very uncer-
tain ratios that distort the test. This is also the situation
when treating individual monthly series.

This is relevant when applying the homogenization
process described in the last section because it is a non-
robust method (distribution dependent). The qi ratios in
Eq. (2) are assumed to follow a Gaussian distribution
and deviations from it could damage the performance
of the test.

The problem of very dry months, a frequent one in
the southern stations, is reduced when using annual data.
After applying a x2 test for the goodness of the fit (Sne-
decor and Cochran 1972) of the qi to a normal distri-
bution, it was found that for a significance level of 0.1,
88% of the qi series in step 1 [see section 3b(2)] met
the Gaussian hypothesis. Since the non-Gaussian 12%
is in the range of the significance level, the approxi-
mation was made that the set of qi were normally dis-
tributed.

Table 2 shows the number and type of series involved
in each step (column 1) of the homogenization proce-
dure. The number of time series tested in each step is
shown in column 2 and the number of series making
part of the reference group appears in column 3. The
quantity of resulting homogeneous (H) or inhomoge-



1 MARCH 2001 971G O N Z Á L E Z - R O U C O E T A L .

FIG. 4. (a) Type of time series after applying the homogenization
procedure: homogeneous (H), homogeneous after one (HC) or two
(HCC) corrections, and inhomogeneous after two corrections (ICC).
Numbers in the legend stand for the step of the procedure. Solid
symbols indicate agreement with metadata. Isolines indicate the num-
ber of reference stations used. Contour interval is 2 units. (b) Cor-
relation values between F and G [see Eq. (3)]. Contour interval is
0.1. (c) Distribution of correction factors within the studied period.
Crosses stand for inhomogeneities of unknown origin, circles cor-
respond to relocations or interpolations of data from distant sites, and
triangles to changes in height, the histogram of inhomogeneity dates
both for the detected changes (black bars) and those registered in the
metadata (white bars).

neous (I) time series are indicated in columns 4–9, either
after one (HC, IC) or after two (HCC, ICC) corrections.

In the two first steps almost the same results are ob-
tained. The most relevant changes are shifts in the clas-
sification of some series that switch from homogeneous
in step 1 (H1) to inhomogeneous in step 2 (I2; stations
42, 51, 72, and 90) or vice versa (stations 13, 23, 34,
79, and 94). In the third stage the 53 inhomogeneous
series obtained in step 2 are corrected and tested pro-
ducing a subset of 45 homogeneous after one correction
(HC3) and 8 inhomogeneous (IC3) series.

Following the methodology previously described, in
the fourth step the subset of 42 homogeneous series in
step 2 (H2) was tested and provided 38 purely homo-
geneous series without any corrections (H4) and 4 series
that contain at least one inhomogeneity (I4). The 53 I2
series were corrected and tested again and provided 44
series homogeneous after one correction (HC4) and 9
series with at least two inhomogeneities (IC4).

During the fifth stage of the method the 4 I4-type
series obtained in step 4 were corrected and tested. The
output was a set of 3 homogeneous series (HC5) and 1
with at least 2 inhomogeneities (IC5).

In the last step, the procedure was applied to the 10
series with more than one inhomogeneity (IC4 and IC5)
being 9 of them successfully corrected (HCC6) and only
one (station 69, Alange), which showed more than two
inhomogeneities (ICC6). This last station was rejected
from the dataset. Here 40% of the series were homo-
geneous, an additional 49.5% became homogeneous af-
ter one correction, and 9.5% more after two corrections.

Figure 4a describes the type (symbols) of inhomo-
geneity of each time series. It can be appreciated that
there is no evident spatial structure in the distribution
of inhomogeneities. The same figure shows the number
of reference stations used for the analysis of each series
(solid lines). This number was imposed to be always
between 5 and 10 and was chosen as a function of the
correlation (always higher than 0.5) between the test
station and any candidate reference station. There is a
clear structure, with the highest values in the west of
the Iberian Peninsula and the south of France where
correlations between pairs of time series are also higher
than in any other area. This also corresponds with the
areas of highest correlation values between the F and
the G series [Eq. (3)] that are shown in Fig. 4b. The
high correlations between F and G also support our
criterion for selecting reference stations.

Figure 4c shows the temporal distribution of correc-
tion factors. Inhomogeneities that are not supported by
metadata are identified with a cross sign and those that
agree with the history of the station with a circle or a
triangle corresponding to a relocation or a change in
height, respectively. The histogram shows the distri-
bution of the changes during the century, both for the
detected inhomogeneities (black bars) and for the chang-
es registered in the metadata (white bars).

More complete history files of stations would be de-



972 VOLUME 14J O U R N A L O F C L I M A T E

sirable to discuss the potential influence of changes in
instrumentation, observer, and other factors upon the
quality of the dataset. Unfortunately, only relocation and
changes in the height of a few stations could be related
to identified inhomogeneities (30% of the total). The
cases in which data from external sites were used to fill
gaps in the time series, both at the early stages of this
work and by the original sources (see section 2) have
also been considered as a type of relocation in the com-
parison of inhomogeneities and metadata. Relocation
can lead to inhomogeneities as a result of the difference
in annual precipitation in both the old and the new po-
sition as well as differences in sheltering conditions. In
70% of the inhomogeneities identified as relocations an
increase of catchment (factor . 1) has been produced.
However, the mean of all correction factors in Fig. 4c
does not depart significantly from one, therefore, there
is no bias in the occurrence of inhomogeneities either
to increase or decrease precipitation means on the whole
dataset. This is due to the fact that unknown changes
(crosses in Fig. 4c) have led to a decrease in catchment
(factor , 1) in 66% of the cases. Nevertheless, changes
in specific sites have played a decisive role as will be
shown later.

The three changes in height (triangles) seem to agree
somehow in the type of inhomogeneity. The two first
ones occur in La Coruña (1916) and San Sebastián,
Spain (1918) and reveal an increase in catchment related
to a new position of the pluviometer in a lower and
probably more sheltered location. The third one (Gi-
braltar in 1935) corresponds to a change of the pluvi-
ometer to a higher elevation, in principle less sheltered,
and thus favoring a decrease in catchment.

Some of the inhomogeneities lie on the 1.0 value line.
All of them correspond to cases in which two inho-
mogeneities were detected, but only the data between
them had to be corrected, thus in such cases the first
inhomogeneity is identified in Fig. 4c with value 1.0.

Finally it is also worth pointing out that 80% of the
inhomogeneities occur before 1950. Such a bias is also
present in the histogram of the medatata (white bars).
Though the metadata archives are incomplete they can
be taken as an indication that there have been fewer
changes in the more recent decades. This supports the
smaller amount of inhomogeneities after 1950.

Figure 5 gives some typical examples of the appli-
cation of the Standard Normal Homogeneity Test to
several types of time series showing values for the T
statistic [Eq. (4)] and the series of ratios qi [Eq. (2)].
Figure 5a displays results for Murcia, Spain (station
number 74), classified as homogeneous in the fourth
step of the procedure (H4) and shows no significant
changes in the T statistic; there is a secondary maximum
in 1955 that seems to agree with a change in the position
of the rain gauge registered in the metadata. The series
of ratios, q, does not show apparent displacements from
the mean value. Palma de Mallorca, Spain (Fig. 5b), is
an example of homogeneous after one correction series

(HC4). In step 2 of the procedure (dots), the test reveals
a wide T maximum centered in 1940, possibly related
to a change in the position of the rain gauge in 1937.
This maximum disappears after the correction factor is
applied to the values before 1937. The q values also
reveal this inhomogeneity: noncorrected values in dots
cluster above (below) 1 for most years before (after) the
change while for the corrected ones (continuous line)
this steplike behavior is reduced.

Another example (Fig. 5c) of homogeneous series
after applying one correction is La Coruña (number 19).
In this case the test statistic shows several significant
maxima. The most important one takes place in 1915
and agrees with a change in the height of the rain gauge;
the second peak occurs in 1942 and does not match with
any entry in the historical records. The first peak dis-
appears and the second one is not significant after ap-
plying the correction factor to the data before 1915. The
q series shows some displacement from the mean value
before 1915 (dots in the oval), which also disappears
after the correction is applied. Hervás (number 56), one
of the series reconstructed for the dataset, is yet another
example of HC4-type series. The T statistic shows (Fig.
5d) a low-amplitude maximum in 1983 not supported
by the available metadata that drops out in step 4 of the
procedure.

Bilbao, Spain (number 20), described by Fig. 5e is
an example of a series that became homogeneous after
two corrections. The values of the T statistic in the
second step (hollow dots) show a maximum centered
in 1946 that agrees with a reference in the metadata in
1945. After the correction of this inhomogeneity, a sec-
ond maximum rises over the significance level in the
fourth step (solid dots). This peak agrees with a change
in the position of the rain gauge in 1920. In the sixth
step both inhomogeneities are corrected and the T sta-
tistic shows no important peaks. The q values also show
these changes: it can be noticed how noncorrected data
(dots) depart more strongly from unity than the cor-
rected ones (solid line), particularly during the period
between both inhomogeneities (highlighted with the
oval).

Finally, results for Alange, Spain (number 69), are
shown also in Fig. 5f. This is the only station which
remained inhomogeneous after two corrections (ICC).
The T values exhibit significant maxima in years 1934,
1917, and 1955 corresponding to the steps 2, 4, and 6
of the procedure, respectively. The series of q ratios
show an appreciable displacement from the mean value
only in step 2 (hollow dots in the oval). The inhomo-
geneity in 1917 is supported by only a single peak value
and the maximum in step 6 hardly reaches the T95 sig-
nificance level. This accounts for the low magnitude of
these inhomogeneities and their lack of evidence in the
series of ratios for steps 4 and 6. The three inhomo-
geneities are probably related to data interpolations from
the secondary sites.

The correction factors deduced from the analysis of



1 MARCH 2001 973G O N Z Á L E Z - R O U C O E T A L .

FIG. 5. Values of T and q for (a) Murcia, (b) Palma de Mallorca, (c) La Coruña, (d) Hervás, (e) Bilbao, and (f ) Alange. Dates of metada-
ta are indicated with arrows in the horizontal axis. Here q data before the inhomogeneities are highlighted in the oval.



974 VOLUME 14J O U R N A L O F C L I M A T E

FIG. 6. Temporal distribution of the annual percentage of missing
data.

FIG. 7. Trends and outputs of a centered 5-yr moving average for
monthly precipitation values in La Coruña before (dashed line) and
after (solid line) the homogeneity correction.

the annual data were applied to the monthly data and
interpolation for missing values was also accomplished
following Eqs. (6) and (7). Figure 6 shows the temporal
distribution of missing values in the dataset. The highest
percentages are reached at the beginning of the century
due to the lower amount of stations at this time. There
is a secondary maximum in the 1970s and 1980s due
to the shut down of measurements in some stations. A
small maximum close to 1940 related to the Spanish
Civil War (1936–39) can also be appreciated. After the
interpolation was accomplished Camporredondo, Spain
(number 25), and Algiers, North Africa (number 86),
were eliminated from the dataset because of the long
data gaps present and neighboring stations that can de-
scribe the local climate features were available in both
cases.

c. Trends

This section provides a description of some effects
of the adjustments. The effects of outliers adjustments
on trends and on the homogenization procedure are de-
scribed as well as the effects of homogenization cor-
rections on trends. A more complete description of
trends, including a discussion of connections to general
circulation changes, is beyond the scope of the present
work. Some results with the same dataset are presented
by González-Rouco et al. (2000).

Trends are obtained through the calculation of least
square linear regressions of precipitation data against
time. A significance t test (Snedecor and Cochran 1972)
for constant precipitation is applied. The effective sam-
ple size and autocorrelation in the residuals are also
taken into account. As precipitation time series often
show high autocorrelation, values are not independent
of each other and the effective number of data has also
been used in the testing of significance (Trenberth 1984;
Zwiers and von Storch 1995). Also, residuals resulting
from the least square regression are usually assumed to
be white noise, but this is not often the case. When

residuals show autocorrelation both the slope and its
error are badly estimated. We have used a Durbin–Wat-
son method (Durbin and Watson 1971; Valero et al.
1996a) to account for the existence of autocorrelation
in the residuals. An AR(1) model has been assumed for
the errors of the least square regression whenever sig-
nificant autocorrelations were detected.

Figure 7 shows the effect of the homogeneity cor-
rection for the time series from La Coruña. The solid
lines correspond to a low-pass filter output and a linear
fit to the corrected data; the dashed line is similarly
obtained from the original series. The linear fit to non-
corrected data shows a significant (a 5 0.05) slope of
3.6 mm decade21 that is reduced to a nonsignificant
value of 0.9 mm decade21 after the correction.

Figure 8 shows trends for the dataset in different stag-
es: for the original data, without corrections (Fig. 8a),
after applying outliers corrections (Fig. 8b), after ap-
plying outliers and homogeneity corrections (Fig. 8c),
and after interpolating data in the homogenized dataset
for the period 1899–1989. In Figs. 8a–c results for each
time series correspond only to the period of existence
of data within the time interval 1899–1989 (i.e., Gre-
noble, France: 1899–1976, see Table 1), therefore, in-
terpreting spatial structures from these plots should be
done with caution since the time span of the time series
can change from site to site.

In the noncorrected data (Fig. 8a) 30% of the trends
are significant. In general, positive trends seem to dom-
inate, particularly in the north of the Iberian Peninsula.
Figure 8b shows trends after the corrections of outliers
were applied. Nearly the same pattern is obtained, ex-
cept in some specific cases where changes in the mag-
nitude and significance of slopes arise: Perpignan,
France (number 30), Palencia (39) and Leon, Spain (32),
as the only change in sign, and Castellón, Spain (50).

In order to check the influence of outliers corrections
in the homogenization procedure, step 1 in Fig. 2 was
rerun with the original data without outliers corrections.
The test results were different in eight cases: stations



1 MARCH 2001 975G O N Z Á L E Z - R O U C O E T A L .

FIG. 8. Trends in the (a) original dataset, (b) after correcting outliers, (c) after correcting outliers and inhomogeneities.
Annual means for the periods of existence of data in each station have been used for the calculations. (d) Trends in the
corrected dataset after interpolating missing data for the period 1899–1989 (inputs were also annual means). A solid line
for value zero is plotted for easy visualization of different regions.

38, 43, 47, 71, and 84 switched from inhomogeneous
to homogeneous and 72, 87, and 94 from homogeneous
to inhomogeneous. In all these cases T values were very
close to the significance level. If outliers were not
trimmed the series were pushed to the significance or
nonsignificance side depending on the specific charac-
teristics of the time series. Outliers cause an increase
in the mean of the period in which they occur. If this
rise in mean sharpens (shortens) the differences between
the period before and after a given change, it will in-
crease (decrease) the T values and the significance of
the change. The authors consider that it is safer to take
decisions upon the quality of the time series that depend
mostly on its overall characteristics rather than on the
extreme behavior of a few data. This is also valid for
the calculation of trends, since we are interested in re-

sults that describe the long-term behavior of the time
series and are not so sensitive to individual data. There-
fore we think that the use of trimmed data is justified.

Figure 8c shows the trends after the data have been
corrected for outliers and inhomogeneities. If we com-
pare it to Fig. 8b we can describe the changes as dra-
matic. Significance is reached in only eight sites and
the picture is more uniform. In most sites the value of
trends is negligible and there are almost no negative
trends left; positive trends in the north of the Iberian
Peninsula still remain.

Figure 8d shows trends of the corrected and inter-
polated data for the period 1899–1989. A solid zero line
is plotted for easier visualization of regions of different
sign. Though significance is reached only at a few sta-
tions, a highly coherent spatial behavior arises in this



976 VOLUME 14J O U R N A L O F C L I M A T E

FIG. 9. Trends in the final dataset for (a) winter, (b) spring, (c) summer, and (d) fall. (e) Winter trends for the period
1920–89. Monthly data for each season were used as inputs. Full symbols highlight significant trends (95% significance
level). A solid line for value zero is plotted for easy visualization of different regions.

plot. All stations close to the Atlantic coast show pos-
itive trends as do most French stations. On the east coast
of the Iberian Peninsula, positive values also arise. The
Pyrenees, the French Mediterranean coast, north of Af-
rica, and quite a few stations in the middle and south
of Spain show negative values.

These annual trends result from the cumulative effects
of seasonal trends. Some insight into the latter is ob-
tained from Fig. 9. Slopes have been obtained using
monthly anomalies for each season (winter: December–
February; spring: March–May; summer: June–August;
autumn: September–November). Anomalies were cal-
culated by subtracting the long-term monthly means to
filter out the mean annual cycle.

Trends in winter (Fig. 9a) show positive values over

nearly all the area except for the north of Africa. Sig-
nificance is reached in 40% of the trends. In spring (Fig.
9b) the positive trends lose significance in the northern
half of the Iberian Peninsula and give way to negative
trends in the southern half. In the summer (Fig. 9c) the
pattern alternates areas of positive and negative values.
In autumn (Fig. 9d), negative, often significant, precip-
itation trends are obtained nearly everywhere except for
some sites on the east coast and in the northwest. The
decrease of precipitation in autumn and spring and in-
crease in winter indicates strengthening of the annual
cycle. These results agree with those of other authors
that report a significant downward trend during this cen-
tury, both in the intensity of the zonal circulation in the
North Atlantic and in the amplitude of its annual cycle,



1 MARCH 2001 977G O N Z Á L E Z - R O U C O E T A L .

these being ascribed by changes in the North Atlantic
oscillation (NAO) index (Lamb and Peppler 1987; Hur-
rell 1995). When the NAO index annual amplitude is
less than normal, that is, when smaller autumn–winter
and/or greater NAO for summer is observed (Makro-
giannis et al. 1991) the amplitude of the annual wave
in precipitation in the Iberian Peninsula increases (Val-
ero et al. 1996b).

It is worth remarking on the possible influence of
missing data on trends. Figure 6b shows that changes
in the early part of the century can have a considerable
influence on trends. According to Fig. 6a, 40% of the
data are missing during the first decade of the century
and are interpolated using information from the existing
60%. There is a possibility that this interpolation biases
the magnitude and sign of the trends toward spatial con-
tinuity. This has been explored through calculating
trends for the period 1920–89, thus avoiding the prob-
lem of long missing data gaps in some stations. Figure
9e shows the result for the wintertime. Though the mag-
nitude of trends is not so high as in Fig. 9a, regional
agreement is still the general feature. The magnitude of
trends is smaller than in Fig. 9a, but this could also be
due to natural variability. In addition to the interpolation
method we suggest that the homogenization procedure
could influence some regional agreement. The funda-
mental hypothesis in the Standard Normal Homogeneity
Test states that the ratios between a test station and some
regional precipitation average should be constant. It is
possible that this assumption leads to removing distinct
local trends in stations imbedded in microclimates and
where changes in long-term trends are not necessarily
related to homogeneity problems.

5. Discussion and conclusions

Outliers and inhomogeneities have been taken into
account as potential quality problems in a dataset of 95
precipitation time series in southwestern Europe.

Outliers adjustments have been carried out by trim-
ming very extreme data with the aim of reducing large
distribution tails as a means of preprocessing previous
to homogeneity corrections. There are, however, caveats
to the use of this subjective approach since it involves
dismissing interesting climatological information about
extreme events and their temporal and spatial distri-
bution. Nevertheless, it is worth noting that this does
not mean a ‘‘point of no return’’ since the original value
of the extreme data can be returned to the dataset at any
time, also incorporating the homogeneity corrections.

The outliers adjustments applied herein lead to chang-
es of minor importance in trend assessment as shown
in the previous section. Concerning the impact of out-
liers adjustments in the homogeneity test, this becomes
evident in cases were the T values are very close to the
significance level. In such situations the effect of a few
extreme values can lead to significance or not signifi-
cance of an inhomogeneity, depending on the case.

Though it might seem contradictory to support correc-
tions that have such a limited effect, it is the authors’
opinion that the use of trimmed data is legitimate be-
cause decisions concerning trends and especially ho-
mogeneity are then taken with regard to the long-term
behavior of the series rather than the marginal extreme
properties of a few values.

Analysis of the homogeneity of the time series found
40% were homogeneous, an additional 49.5% were ho-
mogeneous after applying one correction factor and
9.5% more after applying two correction factors. In 30%
of the cases the inhomogeneities could be linked to some
historical change in the rain gauge, in spite of the me-
tadata being scarce. Examples have been presented that
provide evidence that precipitation records from the area
of study show severe homogeneity problems. There is
considerable influence of homogeneity corrections in
the estimation of trends, emphasizing that the results of
any long-term variability study from this area are very
dependent upon the use of homogenized or nonhomo-
genized data.

Trends for the corrected data present great regional
agreement. Winter values show positive and usually sig-
nificant trends over most of the study area. Significant
negative trends show up mainly in autumn. These trends
seem to agree with those described in the literature for
the zonal circulation in the North Atlantic (Makrogian-
nis et al. 1991; Valero et al. 1996b). A brief description
of trends in relationship to the outliers and homogeneity
adjustments made on the dataset has been provided. A
deeper insight into long-term trends and their relation-
ship to general circulation is beyond the scope of this
work.

The Standard Normal Homogeneity Test is based on
the hypothesis that the ratios between the test-station
values and some regional precipitation average should
be constant. It is possible that this underlying assump-
tion leads to removing distinct local trends in stations
embedded in microclimates and not necessarily related
to inhomogeneity problems. It should also be pointed
that the output of such a quality control procedure is
not expected to be an error-free dataset, but is one where
the quality of the time series has been significantly im-
proved. On the other hand, different methodologies can
lead to partially different results, since not only the error
associated with any homogeneity test, but also the var-
ious methods and criteria that can be chosen for the
selection of reference stations add uncertainty. Different
countries and institutions have chosen methodologies
that specifically suit their purposes (Peterson et al.
1998b). In this sense, work intercomparing different
methods is desired and still to come.

Acknowledgments. This work was supported by Pro-
ject Number CLI97-0558, Comunidad de Madrid, and
Fundación Ramón Areces. The authors also thank P.
Brandt, P. Cuesta, M. Widmann, and E. Zorita for their
comments on the original version of the manuscript. We



978 VOLUME 14J O U R N A L O F C L I M A T E

thank also F. Zwiers and the two anonymous reviewers
for their valuable comments and suggestions.

REFERENCES

Alexanderson, H. A., 1986: A homogeneity test applied to precipi-
tation data. J. Climatol., 6, 661–675.

Almarza, C., J. A. López, and C. Flores, 1996: Homogeneidad y
variabilidad de los registros históricos de precipitación de Es-
paña. Tech. Rep. A-143, Instituto Nacional de Meteorologı́a,
Madrid, Spain, 318 pp. [Available from Instituto Nacional de
Meteorologı́a, Madrid, Spain.]

Barnett, V., and T. Lewis, 1994: Outliers in Statistical Data. 3d ed.
J. Wiley and Sons, 604 pp.

Buishand, T. A., 1982: Some methods for testing the homogeneity
of rainfall records. J. Hydrol., 58, 11–27.

Conrad, V., and C. Pollak, 1950: Methods in Climatology. 2d ed.
Harvard University Press, 459 pp.

Durbin, J., and G. S. Watson, 1971: Testing for serial correlation in
least squares regression, III. Biometrika, 37, 1–19.

Eischeid, J., C. B. Baker, T. Karl, and H. F. Dı́az, 1995: The quality
control of long-term climatological data using objective data
analysis. J. Appl. Meteor., 34, 2787–2795.

Esteban-Parra, M. J., F. S. Rodrigo, and J. Castro-Diez, 1998: Spatial
and temporal patterns of precipitation in Spain for the period
1880–1992. Int. J. Climatol., 18, 1557–1574.

Fernández-Mills, G., 1995: Principal component analysis of precip-
itation and rainfall regionalization in Spain. Theor. Appl. Cli-
matol., 50, 117–131.

Filippov, V. V., 1968: Quality control procedures for meteorológical
data. Tech. Rep. 26, WMO, Geneva, Switzerland, 38 pp.

Garrido, J., J. A. Garcia, and V. L. Mateos, 1996: Homogeneidad y
variabilidad natural de series largas de precipitación. An. Fı́s.
92, 19–31.

González-Rouco, J. F., H. Heyen, E. Zorita, and F. Valero, 2000:
Agreement between observed rainfall trends and climate change
simulations in the southwest of Europe. J. Climate, 13, 3057–
3065.

Grant, E. L., and R. S. Leavenworth, 1972: Statistical Quality Con-
trol. McGraw-Hill, 764 pp.

Hanssen-Bauer, I., and J. Førland, 1994: Homogenizing long Nor-
wegian precipitation series. J. Climate, 7, 1001–1013.

Hawkins, P. M., 1977: Testing a sequence of observations for a shift
in location. J. Amer. Stat. Assoc., 72, 180–186.

Hegerl, G. C., H. von Storch, K. Hasselmann, B. D. Santer, U. Cu-
basch, and P. D. Jones, 1996: Detecting anthropogenic climate
change with an optimal fingerprint method. J. Climate, 9, 2281–
2306.

Houghton, J., B. A. Callendar, and S. K. Varney, Eds., 1992: Climate
Change: The Supplemntary Report to the IPCC Scientific Ass-
esment. Cambridge University Press, 205 pp.
, L. G. Meira Filho, B. A. Callander, N. Harris, A. Kattenberg,
and K. Maskell, Eds., 1995: Climate Change 1995: The Science
of Climate Change. Cambridge University Press, 572 pp.

Hurrell, J. W., 1995: Decadal trends in the North Atlantic Oscillation:
Regional temperatures and precipitation. Science, 269, 676–679.

Jones, P., and Coauthors, 1985: A grid point surface air temperature
data set for the northern hemisphere. Tech. Rep. TRO22, U.S.
Department of Energy, Carbon Dioxide Research Division,
Washington, DC, 251 pp.

Karl, T. R., and C. N. Williams, 1987: An approach to adjusting
climatological time series for discontinuous inhomogeneities. J.
Climate Appl. Meteor., 26, 1744–1763.

Lamb, P. J., and R. Peppler, 1987: North Atlantic Oscillation: Concept
and application. Bull. Amer. Meteor. Soc., 68, 1218–1225.

Lanzante, J. R., 1996: Resistant, robust and non-parametric tech-
niques for the analysis of climate data: Theory and examples,
including applications to historical radiosonde station data. Int.
J. Climatol., 16, 1197–1226.

Linés, A., 1970: The Climate of the Iberian Peninsula. Climates of
Northern and Western Europe, C. Wallen, Ed., World Survey of
Climatology, Vol. 5, Elsevier, 195–239.

Maheras, P., 1988: Changes in precipitation conditions in the western
Mediterranean over the last century. J. Climatol., 8, 179–189.

Makrogiannis, T. J., H. S. Sahsamanoglou, A. A. Flocas, and A.
Bloutsos, 1991: Analysis of the monthly zonal index values and
longterm changes in the North Atlantic and Europe. Int. J. Cli-
matol., 11, 493–503.

Mitchell, J. M., B. Dzerdzeevski, H. Flohn, W. L. Hofmeyr, H. H.
Lamb, K. N. Rao, and C. C. Wallen, 1966: Climate change. Tech.
Rep. 79, WMO, Geneva, Switzerland, 79 pp.

Peterson, T. C., R. Vose, R. Schmoyer, and V. Razuvaev, 1998a:
Global Historical Climatology Network (GHCN) quality control
of monthly temperature data. Int. J. Climatol, 18, 1169–1179.
, and Coauthors, 1998b: Homogeneity adjustments of in situ
atmospheric climate data: A review. Int. J. Climatol., 18, 1493–
1517.

Rhoades, D. A., and M. J. Salinger, 1993: Adjustment of temperature
and rainfall records for site changes. Int. J. Climatol., 13, 899–
913.

Rodriguez-Puebla, C., A. H. Encinas, S. Nieto, and J. Garmendia,
1998: Spatial and temporal patterns of annual precipitation var-
ibility over the Iberian Peninsula. Int. J. Climatol., 18, 299–316.

Santer, B. D., K. E. Taylor, T. M. Wigley, J. E. Penner, P. D. Jones,
and U. Cubash, 1995: Towards the detection and attribution of
an anthropogenic effect on climate. Climate Dyn., 12, 77–100.

Serrano, A., J. A. Garcı́a, V. L. Mateos, M. L. Cancillo, and J. Garrido,
1999: Monthly modes of variation of precipitation over the Ibe-
rian peninsula. J. Climate, 12, 2894–2919.

Snedecor, G. W., and W. G. Cochran, 1972: Statistical Methods. 6th
ed. Iowa University Press, 593 pp.

Trenberth, K. E., 1984: Some effects of finite sample size and per-
sistence on meteorological statistics. Part I: Autocorrelations.
Mon. Wea. Rev., 112, 2359–2368.
, and D. A. Paolino, 1980: The Northern Hemisphere sea-level
pressure data set: Trends, errors and discontinuities. Mon. Wea.
Rev., 108, 856–872.

Valero, F., J. A. Garcı́a-Miguel, M. L. Martin, and Y. Luna, 1993:
Detecting long-term variations in precipitation in the souteastern
Iberian Peninsula. Ann. Geophys., 11, 837–845.
, J. F. González, F. J. Doblas, and J. A. Garcı́a-Miguel, 1996a:
A method for the reconstruction and temporal extension of cli-
matological time series. Int. J. Climatol., 16, 213–227.
, F. J. Doblas, J. F. González, and J. A. Garcı́a-Miguel, 1996b:
On long-term evolution of seasonal precipitation in southwestern
Europe. Ann. Geophys., 14, 976–985.
, R. M. Rodrı́guez, and J. A. Garcı́a-Miguel, 1997a: Filling data
algorithms in urban air pollution monitoring networks. J. En-
viron. Eng. Div. ASCE, 103, 395–403.
, M. Y. Luna, and M. Martin, 1997b: An overwiew of a heavy
rain event in Southeastern Iberia: The role of large-scale me-
teorological conditions. Ann. Geophys., 15, 494–502.

Zorita, E., V. Kharin, and H. von Storch, 1992: The atmospheric
circulation and sea surface temperature in the North Atlantic
area in winter: Their interaction and relevance for Iberian pre-
cipitation. J. Climate, 5, 1097–1108.

Zwiers, F. W., and H. von Storch, 1995: Taking serieal correlations
into account in tests of the mean. J. Climate, 8, 336–351.