Journal of Neuroscience Methods 121 (2002) 129
/137

www.elsevier.com/locate/jneumeth

The sliding window correlation procedure for detecting hidden

correlations: existence of behavioral subgroups illustrated
with aged rats
Daniela Schulz, Joseph P. Huston 
Institute of Physiological Psychology, Center for Biological and Medical Research, University of Düsseldorf, Universitätstr. 1, D-40225 Düsseldorf,
Germany

Received 29 May 2002; received in revised form 22 July 2002; accepted 23 July 2002

Abstract

We developed the sliding window correlation procedure in order to examine populations for possible heterogeneity in the ways
two variables are related with each other. This procedure involves computing correlation coefficients (R ) for overlapping successive
segments of the covariate scores. The distribution of resulting R s reveals fluctuations in the degree and direction of R over the
sample of ranked scores. This procedure is applied to behavioral data of aged rats, which were rank-ordered according to water
maze performance, and correlated with open field exploration and conflict behavior in a light/dark chamber. Results revealed
correlation coefficients of varying magnitudes and opposing directions for different segments of the population, which were
obscured by overall correlation analysis. E.g. for the superior learners, the R s were highest between maze learning ability, increased
open field exploration and reduced anxiety in the conflict test, whereas for the intermediate learners the R s were highest for maze
learning ability related with reduced exploration and increased anxiety. Thus, the sliding window correlation distribution can be
applied in conjunction with overall correlation analysis to provide information about the potential presence and locations of
subgroups within a population, especially if overall correlation analysis does not yield significant results.
# 2002 Elsevier Science B.V. All rights reserved.

Keywords: Individual differences; Aging; Correlation analysis; Learning and memory; Water maze; Open field exploration; Conflict behavior; Light/
dark chamber

1. Introduction represent distinct subgroups of animals (Baxter and

In performing correlation analysis, one generally tests
for the presence of a relationship between whole sets of
scores. Sometimes, however, a significant relationship
exists only for a portion of the sample points under
consideration, which may be masked by the overall
correlation coefficient. For example, aged animals
exhibit large individual differences in measures of
learning and memory (Rapp and Amaral, 1992; Hase-
nöhrl et al., 1997; Sykova et al., 2002). This has raised
the question of whether the ends of the distribution of
behavioral scores, e.g. the superior and inferior learners,
 Corresponding author. Tel.: /49-211-81142; fax: /49-211-
8112024
E-mail address: huston@uni-duesseldorf.de (J.P. Huston).
0165-0270/02/$ - see front matter # 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 2 7 0 ( 0 2 ) 0 0 2 2 4 - 8
Gallagher, 1996). Such groups might be expected to
differ from each other on the basis of other behavioral
characteristics and biological substrates than learning
and memory performance itself or the neural equivalent,
respectively. Importantly, such subgroups can also differ
in terms of their relationships with other behavioral and
biological variables, such that the correlation coeffi-
cients (R ) would differ in terms of magnitude (weak
/
strong) and/or direction (positive
/negative). For exam-
ple, aged superior and inferior learners have been shown
to differ from each other in terms of reactivity to novelty
(Rowe et al., 1998), pattern of water consumption
during the light/dark cycle (Gallagher and Burwell,
1989), degree of thigmotaxis displayed in a water maze
(Schulz et al., 2002), and negative feedback regulation of
the hypothalamo-pituitary-adrenocortical axis (Issa et
al., 1990). Moreover, we have recently shown that
130 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137
hippocampal long-term potentiation, open field explora-
tory behavior, and levels of water maze learning
correlated significantly with each other in aged animals
which exhibited superior water maze performance, but
not in those with inferior performance (Schulz et al.,
2002). Subgroups of superior and inferior learners also
showed differences in correlation with levels of corti-
costerone and adrenocorticotropin hormone measured
after restraint stress (Topic et al., unpublished).
Although performing separate correlation analyses
for restricted segments of a distribution of scores (e.g.
superior and inferior learners) may pinpoint possible
fluctuations in the magnitude and/or direction of R
across the score dimension, this approach is not
sufficiently systematic. If there is variation in R , we
would like to know the exact locations of variation
along the ranked score distribution. In this article, we
will propose a simple procedure, namely the ‘sliding
window correlation’, to systematically uncover such
variation if it exists. This procedure involves carrying
out correlation analyses for partly overlapping succes-
sive segments of rank-ordered scores and to plot the
resulting R s. Such a procedure was shown below to
reveal not only the presence and the degree of hetero-
geneity in co-variation but also the course of develop-
ment of such interactions across samples of ranked
scores. To illustrate such results, we rank-ordered the
performance of aged rats in a spatial water maze
learning task and correlated this performance with
exploratory behavior in an open field. We also corre-
lated water maze learning with performance in a light/
dark choice chamber, a test for anxiety and exploration.
2. Materials and methods
2.1. Subjects
Our experiments were carried out in accordance with
the German Law on the Protection of Animals and were
approved by the state authority (Bezirksregierung
Düsseldorf). In the first experiment, 31 male 28
/30
months old Wistar rats (531.77 g9/12.01 S.E.M.)
(animal facilities, University of Düsseldorf) were ranked
in terms of their performance in the water maze as
defined by the mean time required to learn the task. In a
second experiment, 27 male 26 months old Wistar rats
(553.70 g9/14.03 S.E.M.) were ranked the same way.
Only those animals were included that did not show
obvious signs of physical weakness or illness. The 31
animals of the first experiment were maintained on a 12
h light:12 h dark cycle (lights on at 07:00 h), whereas the
27 animals of the second experiment were exposed to a
reversed cycle (lights on at 19:00 h) such that all
observations were made during the animals’ active night
cycle. Experiment 1 and 2 were conducted for different
purposes at the outset, so that any differences in their
protocols were intended. The animals were group-
housed with two to three rats per cage unless cage-
mates had died during upbringing. They had free access
to food and water.
2.2. Morris water maze
2.2.1. Apparatus
The maze was a circular 185 cm diam swimming pool
made of black polyethylene and filled 30 cm deep with
209/1 8C water. A 18 cm diam black polyethylene
platform was submerged 1.5 cm under the water surface
level. It was randomly located in the center of one of
four equally large imaginary quadrants (target quad-
rant) for all rats, but was maintained in a constant
position throughout training for each rat. A sequence of
four equidistant starting points was randomly varied for
each test day, but was the same for all animals during a
given test day. Cues located near the maze, such as
posters, a blackboard and cupboard, were available for
spatial orientation and highlighted by four 75 W bulb
lights. Diffuse ceiling lighting provided additional
illumination (ca. 6.5 lx at the surface of the pool).
Computer-based systems for behavioral analysis were
placed behind a barrier remote from the maze. A camera
(SONY, CCD-IRIS) was mounted 2 m above the maze
and all trials were videotaped.
2.2.2. Procedure
The rats were brought to a waiting room at least 30
min before the experiments and were kept in holding
cages shortly before testing. At the beginning of each
trial, they were individually placed into the pool facing
the wall. One adaptation trial of 2 min was run without
a platform present and was followed by 9 days of
training with a platform hidden. A training trial was
terminated when the animal escaped onto the platform
or after a maximum of 2 min, after which the rat was
guided to the platform by an experimenter. Once on the
platform, it was left there for 30 s. Between the trials the
animals were kept in holding cages for 60 s. After four
trials in experiment 1 and after two trials in experiment
2, the rats were dried with paper towels and heated by
two 75 W light bulbs. In experiment 2, two training
trials were run in the morning and two in the afternoon
every third day. All swimming trials were recorded on
video and analyzed via Ethovision (Noldus, Wagenin-
gen, Netherlands). Testing took place between 9:00 and
17:00 h.
2.2.3. Behavioral analysis
Time to escape onto the hidden platform and distance
traversed before reaching the platform in the water maze
were measured for each acquisition trial. Daily mean
times and distances were found to correlate highly (P -
 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137
values ranged between 0.000006 and 0.000001 in both
experiments). Therefore, only the times required to
locate the platform over the course of 9 training days
was taken as an index of learning. Higher as compared
with lower mean escape times averaged over all days of
training (total mean latencies) were used to rank the
overall level of performance.
2.3. Open field
2.3.1. Apparatus
The open field was a dimly lit (ca. 0.30 lx at the center
of the field) 60 /60 cm dark gray PVC arena with 39 cm
high walls. A video camera (SONY, DCR-TRV20E)
was mounted 1 m above the center square; all sessions
were videotaped. A broad spectrum noise generator
emitted 60 dB.
2.3.2. Procedure
Open field testing began approximately 4 days after
water maze exposure in experiment 1 and preceded
water maze exposure in experiment 2. The animals were
individually placed into the center of the open field
facing one of the walls. They were allowed to explore the
arena for 10 min on the first testing day in experiment 1
and for 20 min in experiment 2. On a 2nd testing day,
the animals were allowed 10 min. In experiment 2, the
animals were exposed to an object exploration paradigm
on a 3rd day of testing. The number of rearings
displayed during the first 10 min on day 1 of testing
will be used for correlation analyses here. In experiment
1 testing took place between 20:00 and 01:00 h and in
experiment 2 between 08:00 and 18:00 h.
2.4. Light/dark (L/D) choice chamber
2.4.1. Apparatus
A two-compartment chamber (49 /70/58 cm) was
constructed of either black or white PVC with an open
doorway between the compartments (9.5 /9.5 cm). The
Table 1
Sliding correlation applied to ranked water maze performance correlated with variable X
131
white compartment (24.5/35/29 cm) was illuminated
by a 25 W light bulb positioned circa 50 cm above the
floor. A video camera (SONY, DCR-TRV20E) was
mounted 1 m above the apparatus; all sessions were
videotaped. A broad spectrum noise generator emitting
60 dB provided masking noise.
2.4.2. Procedure
The animals in experiment 1 were run between 20:00
and 01:00 h approximately 2 days after open field
testing. Exposure time was 10 min. An animal was
placed from a holding cage into the white compartment
facing the wall opposite to the doorway. Total time
spent in the black compartment was measured among
other variables and will be considered for correlation
analyses here.
2.5. Sliding window correlation
All 31 animals of experiment 1 and 27 animals of
experiment 2 were ranked according to learning ability
(from shortest to longest total mean latencies, reflecting
superior to inferior learning ability, respectively). Spear-
man rank-order correlation analyses were then carried
out between learning ability and the corresponding
values of a given covariate, for always ten animals at a
time, starting from the first ten, and then proceeding
down the learning level rank order, always including the
next worst rank and excluding the first rank of the
previous interval (see Table 1).
The sliding window correlation procedure was carried
out between learning ability and open field exploration
(number of rearings) separately for experiment 1 and 2.
For experiment 1, the procedure was also applied to
water maze learning ability and L/D choice chamber
behavior (time spent in the black compartment) as the
covariate.
132 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137

2.6. Overall correlation analysis

In order to determine the degree of relationship

between the above mentioned variables for all animals
of the sample, overall correlation analyses using the
Spearman rank-order correlation procedure were also
carried out. The Shapiro
/Wilk test was used to check
for normality of the score distributions of the given
variables and yielded P -values larger than 0.05 (the
significance level) for water maze performance and L/D
chamber behavior in experiment 1, and water maze and
open field behavior in experiment 2. For these variables,
respectively, regression analyses for linear and curvi-
linear functions, including logarithmic, inverse, quad-
ratic, cubic, and exponential functions, were also
performed.

3. Results

The distribution of scores reflecting levels of learning

in the water maze (low total mean latencies indicate
better learning) and amount of exploration in the open
field for the animals of both experiments are shown in
Fig. 1. The means (9/S.E.M.) of the respective distribu-
tions are also shown.
Overall correlation analyses were then carried out: in
experiment 1 (Fig. 2A), the correlation between water
maze learning (total mean latencies) and open field
exploration (number of rearings) was not significant
(r//0.09; P /0.6). In experiment 2 (Fig. 2B), the r
was equal to 0.3 but not significant (P /0.1).
Since the scores of the variables in experiment 2 were
normally distributed, regression analyses were also
performed. Among all the functions tested (linear and
curvilinear), the linear, logarithmic, quadratic, and cubic
functions provided the best approximations to our data:
however only about 12% of the variance in open field
Fig. 1. Score distributions and means (9/S.E.M.) of water maze and
open field performance in aged animals. Total mean latencies (s) were
computed from mean times to platform during 9 days of training in the
water maze. Number of rearings were measured in a novel field. The
data were collected in two separate experiments (Experiment 1, n/31
and Experiment 2, n/27).
Fig. 2. Overall correlations between water maze and open field
performance. Total mean latencies (s) to platform in a water maze
and number of rearings in an open field were correlated for the animals
of Experiment 1 (n/31) (A) and Experiment 2 (n/27) (B). For
purposes of presentation, raw values were plotted and linear regression
lines fitted to the data. Spearman rank-order correlation procedures
yielded P -values /0.05 for both experiments.
exploration was explained by variability in water maze
performance (see Fig. 2B).
The sliding correlation procedure yielded 22 correla-
tion coefficients for experiment 1 (Fig. 3A) and 18 for
experiment 2 (Fig. 3B) for water maze learning ability
and open field exploration as the covariate. These were
plotted on the Y -axis for each interval of ten animals,
the intervals shifting towards the lower (inferior) end of
the rank order on the X -axis.
The results show complex distributions of correlation
coefficients for the rank-order populations of aged rats
in both experiments. In both experiments, the R s ranged
from strong positive values (representing better learning
related with more exploration) to strong negative values
(representing better learning related with less explora-
tion). In experiment 1, for the superior range of learners,
the highest R s were positive such that learning ability
was related with more exploratory activity. For the
intermediate range, the highest R s were negative such
that better learning was related with less exploratory
activity. For the most inferior learners of the popula-
tion, the coefficients were mainly in the positive range.
In experiment 2, for the intermediate learners, the
highest R s were also negative such that better learning
was related to less exploration. For the inferior range of
learners, the coefficients were both positive and nega-
tive. Thus, both experiments point to the potential
 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137 133

distribution of open field scores and carried out the

sliding window correlation procedure as before. Results
(Fig. 3C) show the distribution of sliding R s between
rank-ordered water maze performance and the rando-
mized covariate which may be expected to occur due to
chance. The overall correlation coefficient for this
condition was /0.04 (P /0.82).
For experiment 1, we also correlated water maze
learning with behavior displayed in an L/D choice
chamber, i.e. time spent in the black compartment. An
overall analysis yielded a coefficient of /0.3 (P /0.1).

Fig. 3. Sliding correlations between water maze and open field

performance. The sliding correlation procedure yielded 22 rank order
correlation coefficients between water maze learning and open field
exploration in Experiment 1 (A) and 18 in Experiment 2 (B), one for
each interval of ten animals that were ranked from superior to inferior.
Fig. 4. Overall and sliding correlations between water maze and L/D
The correlation coefficients (R s) were plotted on the Y -axis. In both
choice chamber behavior. (A) Using the Spearman correlation
experiments, positive R s indicate that water maze learning proficiency
procedure, an overall correlation analysis between total mean latencies
(derived from total mean latencies to platform) was related with more
(s) to platform in a water maze and time spent in the black
exploratory activity (number of rearings in the open field) and,
compartment (s) of an L/D choice chamber for the animals of
conversely, negative R s indicate that water maze learning proficiency
Experiment 1 (n/30) yielded a P -value /0.05. For purposes of
was related with less exploration of a novel field. A sliding correlation
presentation, raw values were plotted and a cubic regression line drawn
‘control’ was also performed (C), such that open field performance was
through the data. (B) The sliding correlation procedure yielded 21
randomized and correlated with water maze behavior for each window
rank-order correlation coefficients between water maze learning
as before. The resulting correlation coefficients were randomly
(derived from total mean latencies) and time spent in the black
scattered within the XY -plane to a degree which may be expected to
compartment of an L/D chamber, one for each interval of ten animals.
occur due to chance.
A positive coefficient indicates that water maze learning proficiency
was related with more time spent in the black compartment whereas a
presence of distinct subgroups of aged animals, for negative coefficient indicates that water maze learning proficiency was
which degree of learning was either related with more or related with less time spent in the black compartment. The circles (k)
less exploratory behavior. In order to determine whether indicate the topography the correlation distribution would have, if the
frequency that each animal contributed to the analyses was reduced by
the heterogeneity within the distributions of R s obtained sliding down the rank order by four animals at a time, instead of by
could be due to chance, we established a random one.
134 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137

Regression analysis for these variables showed that the

cubic prediction line provided the best fit to our data:
18% of the variability in L/D chamber behavior was
accounted for by variability in water maze performance
(as compared with 6% of explained variance using the
straight line) (Fig. 4A).
Then, the sliding correlation procedure was applied
and the results are shown in Fig. 4B. As before, the
correlation coefficients ranged between strong negative
and strong positive values. For the superior range of
learners, the highest R s were between better learning
related to less time spent in the black compartment of
the L/D chamber, whereas for the intermediate learners,
the highest R s were between learning ability related with
increased time spent in the black compartment. Low R s
between the variables were obtained for extremely
inferior learners. A sliding correlation control is shown
by the large circles drawn around every fourth dot of the
correlation distribution (Fig. 4B); the inclusion of only
the marked R s in the distribution would minimize the
frequency that each animal is used for correlation
analysis, holding the size of the interval constant (n /
10). In this way, the general topography of the distribu-
tion was retained, but information regarding the width
of peaks was lost.

3.1. Relationship between exploration and anxiety as a

function of water maze performance

Since behavior in the L/D choice chamber is con-

sidered to reflect conflict behavior between the natural
tendency of rats to explore the novel environment and
the preference to stay in the black compartment, i.e. a
reflection of anxiety, we hypothesized that exploratory
behavior in the open field would be related with
behavior displayed in the L/D chamber. An overall
correlation analysis between these variables yielded a
coefficient of /0.2 (P /0.3). The data are plotted in
Fig. 5A.
A variation of the sliding correlation principle is to
plot the relationship between two variables in terms of
the R s calculated for windows of a third variable. Here
we assessed the extent to which exploratory behavior
and anxiety correlate across the ranks of superior to
inferior water maze learners (Fig. 5B). For the superior
and intermediate learners, the R s were in the negative
range, suggesting more open field exploration to be
predictive of less time spent in the black compartment of
the L/D chamber. For extremely inferior performers, the
R s were in the positive range.
For the same variables as in Fig. 5B, the R s were also
computed using seven instead of ten animals as an
interval size. A smaller interval was used to obtain a
more detailed correlation distribution. The results in
Fig. 5C show a similar distribution of R s to that
obtained with a larger interval size. Many of the
Fig. 5. (A) Overall correlation between number of rearings in an open
field and time spent in the black compartment of an L/D choice
chamber for the animals of Experiment 1 (n/30) (P /0.05). Raw
values are shown and a linear regression line was drawn through the
data for illustrative purposes. (B) For the sliding correlation proce-
dure, the animals were rank-ordered according to learning ability in
the water maze, from superior to inferior. Correlation analyses were
carried out between number of rearings and time spent in the black
compartment (s) and resulted in 21 correlation coefficients, one for
each rank order interval of ten animals. (C) The sliding correlation
distribution for the same variables as in B with an interval size of seven
animals each. In the sliding correlation distributions B and C, a
positive coefficient indicates that more exploratory activity was related
with more time spent in the black compartment whereas a negative
coefficient indicates that more exploratory activity was related with
less time spent in the black compartment.
coefficients, especially at intermediate and inferior
learning levels, became larger. Some of the weaker
coefficients at superior learning levels, became smaller.
 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137
4. Discussion
The present experiments examined the relationships
between water maze learning, exploration of a novel
field and conflict behavior displayed in an L/D choice
chamber in order to illustrate the sliding window
correlation procedure. This procedure which, in contrast
to overall correlation analysis, examines relationships
between variables for successive segments of a distribu-
tion of scores, was able to uncover hidden correlations
of varying magnitudes and direction which were ob-
scured by overall correlation analysis. Whereas correla-
tion analysis for all sample points of the learning
distribution yielded low R s between the variables tested,
the sliding correlation procedure yielded high and low
and positive and negative R s for different segments of
the learning distribution, indicative of the fact that the
aged animal population as examined here, was hetero-
geneous with respect to the ways the above variables
were related to each other, and thus, may consist of
several subgroups with distinct behavioral characteris-
tics. Furthermore, the sliding correlation procedure was
able to demonstrate the exact locations and the devel-
opment of the changes in magnitude and direction of the
coefficients within the correlation distribution.
4.1. Behavioral outcomes of the sliding correlations
4.1.1. Water maze performance and open field
exploration
Upon examining the relationship between water maze
learning and open field exploration using the sliding
window correlation procedure, we found that, for
different segments of the learning distribution, water
maze performance was related with both more and less
exploratory behavior in a novel field in two separate
experiments. This variability in the relationship con-
firms previous results reported, where, for a superior but
not an inferior group of learners, water maze learning
proficiency was significantly correlated with more
exploratory activity (Schulz et al., 2002). Using the
sliding correlation procedure, we obtained high positive
R s between the same variables for the superior range of
animals in experiment 1, but not in experiment 2. Any
differences in the sliding R distributions between the
experiments, which were conducted for different pur-
poses at the outset, could be due to variations in age as
well as experimental protocols.
To our surprise, however, the sliding correlation
distributions in both experiments also revealed segments
with high R s indicating a negative relationship between
water maze performance and open field exploration,
such that better performance in the water maze was
related to lower levels of exploratory activity. The
change from a positive (at superior learning levels) to
a negative (at intermediate learning levels) relationship
135
was expressed by a regular stepwise change in the
magnitude as well as in the direction (best seen in
experiment 1) of the correlation coefficients. Explana-
tions for why water maze performance was related with
both, more and less, exploratory activity, have to be
sought in interrelationships with other behaviors, e.g.
conflict behavior (see below).
Finally, the sliding correlation procedure indicated
low and high and positive and negative R s for the
inferior segments of learners in both experiments. From
this perspective, it is of no surprise that in a previous
report, we did not find a significant correlation between
water maze performance and open field exploration for
the inferior group of water maze learners (Schulz et al.,
2002).
In summary, whereas overall correlation analyses
between water maze performance and open field ex-
ploration yielded low R s between the variables, the
sliding correlation procedure provides a differentiated
picture: for different segments of the learning level rank
order of animals, the R s were positive and negative, thus
canceling each other out in an overall analysis. The
gradual development of positive and negative peaks
along the sliding correlation distribution indicates that
the aged animals were heterogeneous with respect to the
relationships between sets of scores and that different
segments of the rank order distribution likely reflect the
existence of distinct subgroups, expressing different
behavioral phenotypes, within the aged animal popula-
tion.
4.1.2. Water maze performance and conflict behavior in
an L/D choice chamber
For experiment 1, we had the chance to examine the
animals further in terms of the relationship between
water maze performance and conflict behavior as
displayed in an L/D choice chamber. The conflict in
this behavioral paradigm consists of the motivation to
explore a novel environment which is anxiety-inducing
by nature of its being brightly lit, and to avoid the bright
chamber, thereby reducing anxiety and increasing the
time spent in the alternative dark chamber (Crawley and
Goodwin, 1980). Anxiety can be reduced by the admin-
istration of anxiolytic drugs, and in this paradigm,
reduced anxiety is expressed by proportionally de-
creased time spent in the black compartment (Costall
et al., 1989). In the present experiment, the sliding
window correlation procedure applied to learning levels
in the water maze and time spent in the dark compart-
ment of the L/D chamber yielded rather interesting
results: in the superior range of learners, high R s were
obtained in the direction that learning proficiency was
related with lower levels of anxiety. Noticeably, in the L/
D chamber paradigm, anxiety is inversely related with
exploratory activity (Costall et al., 1989). Thus, our
results may be interpreted such that better learning in
136 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137
superior learners is related to reduced anxiety and to
higher levels of exploratory activity. For this segment of
learners, we previously established a relationship in the
same direction between water maze performance and
open field exploratory activity (see Fig. 3A).
Interestingly, between the superior and intermediate
segments of learning levels, a gradual shift in the
opposite direction of correlation coefficients occurred,
such that for intermediate learning levels, high R s were
observed between water maze learning ability and
increased anxiety. Consistently, for this segment of
learning levels, we previously obtained high R s, indicat-
ing a relationship between maze learning ability and
reduced open field exploratory activity (Fig. 3A).
Furthermore, in segments of the sliding correlation
distribution of the inferior learners, water maze perfor-
mance was unrelated with anxious behavior, as it was
inconsistently related with open field exploratory activ-
ity (Fig. 3A).
Thus, once again it becomes clear as to why the
overall correlation analysis did not yield a significant
relationship between the variables tested: for different
segments of the correlation distribution the variables
were differentially related with each other, i.e. positively
and negatively; an overall correlation analysis which
included all sample points of the learning distribution
thus masked such differentiated relationships.
4.1.3. Open field exploration and conflict behavior in an
L/D choice chamber
We also applied a variation of the sliding correlation
procedure to open field exploratory activity and L/D
choice chamber behavior, since the latter is based on a
conflict between the motivation to explore and the
anxiety-induced motivation to remain in the black
compartment of the L/D chamber. Thus, we performed
the sliding correlation procedure for these variables
basing the rank order on a third variable, i.e. water maze
performance levels. Interestingly, for most intervals of
the rank order distribution, the coefficients were in the
negative range, such that less open field exploratory
activity was associated with higher levels of anxiety. By
contrast, for the inferior segment of learning levels, the
R s were mainly in the positive range. Thus, the sliding
correlation procedure was again able to detect hetero-
geneity in the relationship tested within the aged animal
population.
4.2. Sliding correlations using variable interval sizes
Initially we set our interval sizes to an arbitrary
number of ten animals to perform the sliding correlation
analyses, although, of course, other interval sizes could
also be used. In general, the smaller the interval size, the
more detailed the sliding correlation distribution will be
(as an example see Fig. 5B as compared with Fig. 5C).
To employ smaller intervals may be useful to check for
the influence of one or the other variant animal within
the learning distribution, which one suspects to con-
tribute substantially to the magnitude of a given
coefficient. Also, the optimum interval size for the
detection of peaks may vary and can be assessed, e.g.
by using smaller windows for correlation analyses. On
the other hand, with smaller interval sizes, of say five
animals or less, a given R may also become less
meaningful.
By contrast, larger interval sizes may generally
smooth out the distribution of sliding R s. Although
the general shape of this distribution is likely to be
retained with the use of larger intervals, smaller changes
within this distribution may remain undetected. How-
ever, with huge sample sizes, larger intervals may be
recommended for ease of computation. In such cases,
certain segments of the rank-order population may be
re-checked using smaller interval sizes.
4.3. Frequency of an animal contributing to the sliding
correlation
One of the potential limits appertaining to the sliding
correlation procedure concerns the frequency that a
particular animal contributes to the value of a given
correlation coefficient. Assuming an interval size of ten
animals, for instance, the first animal of the rank order
will be used only once for correlation analysis, the
second animal will be used twice, the third animal three
times and so on, up to the tenth animal, from which
point on all of the animals will be used ten times, again
up to the last nine animals of the rank order which will
be used, respectively, nine, eight, seven, etc. times. Thus,
potentially, the shape of the sliding R distribution may
be affected by the frequency with which each animal’s
weight contributes to it. Assuming that, e.g. the first
animal of the rank order, a relatively superior learner in
the present study, adds strongly to a particular direction
and strength of a correlation coefficient, the exclusion of
such an animal may shift the next R of the rank order to
a large degree and induce a strong slope in the
correlation distribution. Had the animal been included
for correlation analysis more often, a given peak
(positive or negative) may have consisted of more R s
being similar to each other, thus building a wider peak.
Although the extent of such a peak cannot be corrected
for, one may nevertheless ensure that the general shape
of the correlation distribution is not affected by this
shortcoming. Heretofore we propose to compute a
sliding correlation control by minimizing the frequency
that each animal is included for correlation analysis.
Thus, instead of sliding down the rank order by one
animal at a time, one may choose to slide by five animals
at a time (holding the size of the interval constant), or by
the number of animals which constitute one interval. In
 D. Schulz, J.P. Huston / Journal of Neuroscience Methods 121 (2002) 129
/137
the present study, this procedure was demonstrated (Fig.
4B) and yielded a similar topography of the correlation
distribution.
4.4. The sliding correlation procedure as a descriptive
tool
The sliding correlation procedure represents a de-
scriptive tool, which may be used in conjunction with
overall correlation analysis. Given the many coefficients
which are being computed by the new procedure, it
would be inappropriate without proper adjustment for
Type I error to talk about significance levels. In the
present article the magnitude of the R s can be con-
sidered to provide ‘measures of effect’. In conjunction
with overall correlation analysis, the sliding correlation
procedure can provide additional information about the
heterogeneity within correlation distributions. Such
information may be especially useful when overall
analyses do not yield significant results contrary to
hypothesis as well as for the generation of new
hypotheses. Even in case of significant overall correla-
tion coefficients, the sliding window procedure may
reveal segments of the ranked score distribution respon-
sible for the correlation and may uncover potential
locations of heterogeneity. The procedure can be applied
to any field of study employing a correlation design and
may also be used to examine the effects of treatments or
subject variables on the shape of the distribution of R s.
As one control procedure, we suggest to establish a
random distribution of scores with which to perform the
sliding correlation which then can be compared with the
one obtained with the actual data. In doing so, we
illustrated the degree of scatter that may result from
using randomized control scores. It was interesting to
observe that even with randomized scores relatively
small peaks could be observed; however, the more
gradual change in magnitude and direction of the
coefficients over a wide range of ranks, as observed
with the actual data, was absent in the control condi-
tion.
5. Conclusion
The sliding correlation procedure can uncover hetero-
geneous or non-monotonic correlation distributions
indicative of the presence of distinct subgroups within
a set of rank-ordered scores. As a complementary
method to overall correlation analysis, the sliding
window correlation procedure provides a means of
visualizing the degree of heterogeneity or homogeneity
137
of the relationship between variables across a distribu-
tion of ranked scores. We have shown that the degree as
well as direction of relationship between behavioral
measures can vary greatly over the range of ranked
scores providing a basis for hypotheses regarding the
existence of subgroups in terms of the relationship
between these variables. While we have demonstrated
the utility of this procedure on behavioral data of aged
rats in pointing to such subgroups, the method has
potential application to any field of study involving the
analysis of relationship between variables.
Acknowledgements
This work was supported by a grant from the German
National Science Foundation (Grant number: Hu306/
11-3). We would like to thank Professor J. Krauth and
Jay-Shake Li for their advice.
References
Baxter MG, Gallagher M. Neurobiological substrates of behavioral
decline: models and data analytic strategies for individual differ-
ences in aging. Neurobiol Aging 1996;17:491
/5.
Costall B, Jones BJ, Kelly ME, Naylor RJ, Tomkins DM. Exploration
of mice in a black and white test box: validation as a model of
anxiety. Pharmacol Biochem Behav 1989;32:777
/85.
Crawley J, Goodwin FK. Preliminary report of a simple animal
behavior model for the anxiolytic effects of benzodiazepines.
Pharmacol Biochem Behav 1980;13:167
/70.
Gallagher M, Burwell R. Relationship of age-related decline across
several behavioral domains. Neurobiol Aging 1989;10:691
/708.
Hasenöhrl RU, Söderström S, Mohammed AH, Ebendal T, Huston
JP. Reciprocal changes in expression of mRNA for nerve growth
factor and its receptors TrkA and LNGFR in brain of aged rats in
relation to maze learning deficits. Exp Brain Res 1997;114:205
/13.
Issa AM, Rowe W, Gauthier S, Meaney MJ. Hypothalamic-pituitary-
adrenal activity in aged, cognitively impaired and cognitively
unimpaired rats. J Neurosci 1990;10:3247
/54.
Rapp PR, Amaral DG. Individual differences in the cognitive and
neurobiological consequences of normal aging. Trends Neurosci
1992;15:340
/5.
Rowe WB, Spreekmeester E, Meaney MJ, Quirion R, Rochford J.
Reactivity to novelty in cognitively-impaired and cognitively-
unimpaired aged rats and young rats. Neuroscience 1998;83:669
/
80.
Schulz D, Huston JP, Jezek K, Haas HL, Roth-Härer A, Selbach O,
Luhmann HJ. Water maze performance, exploratory activity,
inhibitory avoidance and hippocampal plasticity in aged superior
and inferior learners. Eur J Neurosci 2002, in press.
Sykova E, Mazel T, Hasenöhrl RU, Harvey AR, Simonova Z,
Mulders WHAM, Huston JP. Learning deficits in aged rats
correlate with decrease in extracellular volume and loss of diffusion
anisotropy in hippocampus. Hippocampus 2002;12:269
/79.