Data Classification

OVERVIEW

Data classification involves combining raw data into the method can assist in selecting an appropriate number
classes or groups, with each class represented by a of classes.
unique symbol. For instance, data for a choropleth map Many cartographers have promoted the optimal
might be grouped into five classes, with each class method of classification because it does the best job of
depicted by a different shade of gray. The result of data considering how data are distributed along the number
classification is a classed map; in contrast, if each raw line (by placing similar data values in the same class and
data value is depicted by a unique symbol, an unclassed dissimilar data values in different classes). The optimal
map results. method, however, does not score well on some of the other
Although data classification is relevant to a wide vari- criteria; for example, the legend is difficult to understand
ety of mapping techniques, it is commonly discussed in and the method is unacceptable for ordinal data. As a
association with choropleth maps. Traditionally, cartog- result, it is important that you learn about the advantages
raphers argued for classed choropleth maps on two and disadvantages of other methods of classification.
grounds: readers’ inability to discriminate among many One limitation of the classification methods described
differing areal symbols, and the difficulty of creating in section 1 is that they do not consider the spatial context
unclassed maps using traditional photomechanical pro- of the data. If the purpose of classification is to simplify
cedures. Today the latter constraint has been eliminated, the appearance of a map, it can be argued that spatial
as computer hardware is capable of producing unclassed context also should be considered. In light of this, section
maps; as a result, some cartographers now question 2 covers two approaches for simplifying choropleth
whether data classification is a necessity. maps that do consider spatial context. The first uses the
In this chapter, we consider two issues relevant to the optimal classification method, but incorporates a spatial
classification of a single attribute: methods of classifica- constraint by requiring that values falling in the same
tion and how spatial context can be used to simplify the class be contiguous, whereas the second employs no clas-
appearance of choropleth maps. Section 1 considers six sification approach, but rather smooths the raw data by
common methods of data classification: equal intervals, changing values for enumeration units as a function of
quantiles, mean–standard deviation, maximum breaks, values of neighboring units.
natural breaks, and optimal. A common problem you Taking a broader view of the classification issue, Marc
will face is determining which of these methods you Armstrong and his colleagues (2003) have argued that we
should use. Criteria that can assist in selecting a classifi- should consider multiple criteria when classifying data,
cation method include: (1) whether the method considers including minimizing the difference between data values
how data are distributed along the number line, (2) ease in classes, minimizing the difference between classes for
of understanding the method, (3) ease of computation, contiguous enumeration units, maximizing the overall
(4) ease of understanding the legend, (5) whether the spatial autocorrelation in the classed map, and equalizing
method is acceptable for ordinal data, and (6) whether the area in each class. The first of these is a nonspatial

From Chapter 4 of Thematic Cartography and GeoVisualization, Third Edition. Terry A. Slocum, Robert B. McMaster, Fritz C. Kessler,
Hugh H. Howard. Copyright © 2009 by Pearson Education, Inc. All rights reserved.
65
16:05:32.
 Data Classification

constraint, while the latter three are spatial constraints. We in 1990 (U.S. Bureau of the Census 1994). Raw data
discuss Armstrong et al.’s approach in section 3. (sorted from low to high data values) and a dispersion
graph for this data set are given in Table 1 and
Figure 1, respectively. Note that the dispersion graph
1 COMMON METHODS OF DATA has a positive skew, with a distinctive outlier (i.e.,
CLASSIFICATION Dade County, which includes the city of Miami).
Before attempting to classify data, it is useful to
In this section, we consider six common methods of data consider whether the data have either a natural or a
classification: equal intervals, quantiles, mean–standard meaningful dividing point that can be used to partition
deviation, maximum breaks, natural breaks, and opti- the data (the data are bipolar. For example, a data set
mal. These methods range from those that do not con- of “percent population change” has a natural dividing
sider how data are distributed along the number line point of zero, which can be used to create two classes:
(e.g., equal intervals) to those that do (e.g., optimal). We values at or above zero and values below zero. Once
point out advantages and disadvantages of each data have been split in such a fashion, it might be
method, and develop a general set of criteria by which appropriate to apply one of the methods discussed in
the methods can be evaluated. For simplicity, we assume this section to each subset of data. Even if there is no
that the intent is to visualize an attribute for a single natural dividing point in the data, it might be desirable
point in time, as opposed to visualizing how an attribute to create a meaningful dividing point prior to classify-
changes over time. ing. For example, we might compute a mean percent-
For illustrative purposes, we will map the percent- age value for the foreign-born data and then split the
age of foreign-born residents for counties in Florida data into values above and below the mean.

TABLE 1 Percentage of foreign-born population in Florida counties in 1990

Observation County % Foreign-Born Observation County % Foreign-Born
1 Madison 0.6 35 Leon 3.7
2 Calhoun 0.8 36 Lafayette 4.1
3 Dixie 0.8 37 Okaloosa 4.3
4 Baker 0.8 38 Glades 4.5
5 Taylor 0.9 39 Highlands 4.6
6 Jefferson 1.0 40 Citrus 4.9
7 Liberty 1.0 41 Lee 5.2
8 Bradford 1.0 42 Brevard 5.3
9 Wakulla 1.2 43 Manatee 5.4
10 Gadsden 1.2 44 DeSoto 5.5
11 Gilchrist 1.3 45 Hernando 5.5
12 Gulf 1.3 46 Volusia 5.8
13 Holmes 1.4 47 Pasco 5.9
14 Suwannee 1.5 48 Alachua 5.9
15 Hamilton 1.6 49 Sarasota 6.0
16 Walton 1.6 50 Indian River 6.1
17 Nassau 1.6 51 Okeechobee 6.3
18 Columbia 1.7 52 Seminole 6.3
19 Franklin 1.8 53 St. Lucie 6.3
20 Sumter 1.9 54 Charlotte 6.3
21 Jackson 2.2 55 Hardee 6.3
22 Putnam 2.2 56 Martin 6.8
23 Sana Rosa 2.2 57 Osceola 7.1
24 Levy 2.3 58 Pinellas 7.1
25 Union 2.4 59 Orange 7.5
26 Washington 2.4 60 Hillsborough 7.6
27 Escambia 2.7 61 Flagler 8.3
28 Clay 3.1 62 Monroe 10.1
29 Bay 3.4 63 Collier 10.5
30 Duval 3.5 64 Palm Beach 12.2
31 Lake 3.5 65 Hendry 14.6
32 St. Johns 3.6 66 Broward 15.8
33 Polk 3.6 67 Dade 45.1
34 Marion 3.6

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

Equal Intervals (A)

5 10 15 20 25 30 35 40 45
Class 1 Class 2 Class 3 Class 4 Class 5

Quantiles (B)

5 10 15 20 25 30 35 40 45
Class 1 Class 3 Class 5
Class 2 Class 4

Mean–Standard Deviation (C)

5 10 15 20 25 30 35 40 45
Class 3 Class 4 Class 5

Maximum Breaks (D)

5 10 15 20 25 30 35 40 45
Class 1 Class 2 Class 4 Class 5
Class 3

Natural Breaks (E)

5 10 15 20 25 30 35 40 45
Class 1 Class 3 Class 4 Class 5
Class 2

Optimal (median) (F)

5 10 15 20 25 30 35 40 45
Class 1 Class 3 Class 4 Class 5
Class 2

FIGURE 1 Dispersion graphs of the foreign-born data shown in Table 1 along with class breaks for various methods of data
classification.

67
16:05:32.
 Data Classification

An important consideration in any method of TABLE 2 Class limit computations for equal-intervals
classification is selecting an appropriate number of classification
classes. To compare methods of classification in this sec- Class Calculated Limits Legend Limits
tion, we assume five classes, which are easy to discrimi-
Florida Foreign-Born Data (Five Classes)
nate on the gray-tone maps shown here.
Before selecting a classification method, it is essential 1 0.6 to 9.5 0.6 to 9.5
to determine precisely how you wish to portray the data 2 9.6 to 18.4 9.6 to 18.4
on the map. The precision you select will be a function of 3 18.5 to 27.3 18.5 to 27.3
4 27.4 to 36.2 27.4 to 36.2
the initial data you have available, your impression of
5 36.3 to 45.1 36.3 to 45.1
the quality of the data, and how easily you think readers
can interpret the numeric values you provide. In our Florida Foreign-Born Data (Four Classes)
case, we felt that the data were of sufficient quality to 1 0.6 to 11.725 0.6 to 11.7
report to the nearest tenth of a percent and that readers 2 11.726 to 22.850 11.8 to 22.9
of this text would be comfortable with these values. The 3 22.851 to 33.975 23.0 to 34.0
argument could certainly be made, however, that we 4 33.976 to 45.100 34.1 to 45.1
should have rounded to the nearest whole percent.
reflect the rounded raw data on which the classification
1.1 EQUAL INTERVALS is based. Because we began with values rounded to the
nearest tenth of a percent, we also should report class
In the equal intervals (or equal steps) method of classifi- limits to the nearest tenth of a percent. This step does not
cation, each class occupies an equal interval along the affect a five-class map for the foreign-born data because
number line. As a result, this method is identical to creat- the class limits are already reported to the nearest tenth.
ing a grouped-frequency table, except that cartographers (Note that the Calculated Limits and Legend Limits in
commonly distinguish between the calculated class limits Table 2 match.) To illustrate how this step might have an
and the limits actually used for mapping. The steps for impact, consider computing class limits for a four-class
computation are as follows: map for this same data set. The class interval would be
computed as follows:
Step 1. Determine the class interval, or width that each
class occupies along the number line. This is computed Range High - Low
by dividing the range of the data by the number of =
Number of classes Number of classes
classes. The result is as follows for the foreign-born data:
45.1 - 0.6
= = 11.125
Range High - Low 4
=
Number of classes Number of classes Using this class interval, we derive the calculated class
45.1 - 0.6 limits shown at the bottom of Table 2. Assuming that
= = 8.9 we wished to report class limits to the nearest 0.1, we
5
would round the calculated limits to create the legend
Step 2. Determine the upper limit of each class. The limits shown at the bottom of Table 2.
upper limit for each class is computed by repeatedly Step 5. Determine which observations fall in each
adding the class interval to the lowest value in the data. class. This involves simply comparing the raw values
(For the first class, adding the class interval 8.9 to 0.6 with the legend limits from step 4. Figure 1A presents
yields a value of 9.5.) The result is the right-hand set of the resulting classes in graphic form, and a map of the
numbers in the Calculated Limits column at the top of classified data appears in Figure 2A.
Table 2. An advantage of equal intervals is that these steps
Step 3. Determine the lower limit of each class. Lower can be completed using a calculator, or even pencil and
limits for each class are specified so that they are just paper. As a result, this method was often favored before
above the highest value in a lower-valued class (the mapping software became available. A second advan-
lower limit of class 2 is 9.6, which is 0.1 more than the tage is that the resulting equal intervals will, in some
upper limit of class 1). An alternative approach is to let cases, be easy for map users to interpret. For example, if
the lowest value of each class equal the highest value in you were making a five-class map of “percent urban
a lower-valued class, but then it may be unclear into population,” and the data ranged from 0 to 100, the
which class a data value falls if it matches the lowest resulting classes would be convenient rounded multiples
value in a class. (0–20, 21–40, 41–60, 61–80, and 81–100). For the foreign-
Step 4. Specify the class limits actually shown in the leg- born data, however, note that the mapped limits do not
end. The class limits actually shown in the legend should readily reveal what the class interval might be.

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Methods of Data Classification
Foreign Born in Florida, 1990

A B
Equal Intervals Quantiles

Percentage Percentage
Foreign Born Foreign Born
0.6 to 9.5 0.6 to 1.4
9.6 to 18.4 1.5 to 2.7
18.5 to 27.3 3.1 to 4.9
27.4 to 36.2 5.2 to 6.3
36.3 to 45.1 6.8 to 45.1

C D
Mean–Standard Deviation Maximum Breaks

Percentage Percentage
Foreign Born Foreign Born
< -6.9 0.6 to 8.3
-6.9 to -1.0 10.1 to 10.5
-0.9 to 10.9 12.2
11.0 to 16.8 14.6 to 15.8
16.9 to 45.1 45.1

E F
Natural Breaks Optimal (median)

Percentage Percentage
Foreign Born Foreign Born
0.6 to 2.7 0.6 to 2.4
3.1 to 4.9 2.7 to 4.9
5.2 to 8.3 5.2 to 8.3
10.0 to 15.8 10.1 to 15.8
45.1 45.1

FIGURE 2 Choropleth maps illustrating various methods of data classification for the foreign-born data shown in Table 1. (A)
Equal intervals, (B) quantiles, (C) mean–standard deviation, (D) maximum breaks, (E) natural breaks, and (F) optimal, based on
medians.

69
16:05:32.
 Data Classification

data until the desired number of members in a class is

obtained. Identifying class membership for the foreign-
born data is somewhat complicated because the number
of observations desired in each class, 13.4, is not an integer
0.6 9.5 18.4 27.3 36.2 45.1 value. In such a situation, one should attempt to place
approximately the same number of observations in each
class. Thus, we placed 13 observations in the first class and
FIGURE 3 A legend for the equal-intervals method that 14 observations in the second class so that the total for the
takes advantage of the continuous nature of class limits. first two classes was 27 (approximately 13.4 * 2).
Because identical data values should not be placed in
different classes, ties can complicate the quantiles
A third advantage of equal intervals is that the method. This can be seen for the foreign-born data if a
legend limits contain no missing values (or gaps): For class break is attempted between the 54th (13.4 * 4) and
the foreign-born data, the difference between the upper 55th numeric positions. Defining a break here doesn’t
value in a class and the lower value in the next class is make sense because the data value for both positions is
0.1, which is the precision of the data. Gaps, which we 6.3. Rather, it is more logical to choose a break between
will see occur in other classification methods (e.g., the 55th and 56th positions.
quantiles), might cause a reader to wonder why some of Two approaches are possible for defining legend
the data are missing from the legend. A related advan- limits for the quantiles method. One is to specify the
tage for equal intervals is that the legend limits can be lowest and highest values of members in a class; for the
simplified so that only the lowest and highest values in foreign-born data, the 14 members of the second class
the data and the upper limit of each class are shown range from 1.5 to 2.7, and so these limits would be
(Figure 3). This approach should permit faster map shown in the legend. The other approach is to compute
interpretation, but it might also create confusion con- a class boundary as an average of the highest value in a
cerning the bounds of each class (e.g., the reader might class and the lowest value of the next class; using this
wonder whether 9.5 falls in the first or second class). approach for the foreign-born data, the upper limit for
The major disadvantage of equal intervals is that the the second class would be (2.7 + 3.1)/2, or 2.9. We pre-
class limits fail to consider how data are distributed fer the former approach because it more accurately
along the number line. For example, if you inspect the reflects the range of data values falling in a class. An
dispersion graph for the foreign-born data (Figure 1), advantage of the latter approach, however, is that only
you will note that classes 3 and 4 seem to be wasted, as the lowest and highest values in the data and the upper
these classes have no observations within them. On the limit of each class would need to be shown, as was done
plus side, however, note that unusual data values (such for the equal-intervals method in Figure 3.
as the Dade County outlier) will appear in their own As with equal intervals, an advantage of quantiles is
unique class. Thus, although two classes are not used for that class limits can be computed manually. A second
the foreign-born data, the mapped distribution does advantage is that because an equal number of obser-
correctly show Dade County as being distinctly differ- vations fall in each class, the percentage of observa-
ent from the rest of the data. tions in each class will also be the same. Depending on
the number of classes, this might simplify our discus-
1.2 Quantiles sion of the mapped data. For instance, on a five-class
map we can refer to the upper or lower 20 percent of
In the quantiles method of classification, data are rank- the data, whereas on a six-class map we would have
ordered and equal numbers of observations are placed to refer to the upper or lower 16.7 percent of the
in each class. Different names for this method are used, data, which is not a convenient round number. A
depending on the number of classes; for example, four- related advantage is that the 50th percentile or
and five-class quantiles maps are referred to as quartiles median (a measure of central tendency in the data)
and quintiles, respectively. To compute the number of will be logically associated with the classes. For an odd
observations in a class, the total number of observations number of classes, the median will fall in the center of
is divided by the number of classes. For the foreign-born the middle class, whereas for an even number of
data, we have: classes, the median will fall between the two middle
Total observations classes.
Number in class = = 67/5 = 13.4 A third advantage of quantiles is that because class
Number of classes
assignment is based on rank order, quantiles are useful
To determine which observations should be placed in each for ordinal-level data. For example, if the 50 states of
class, one simply progresses through the rank-ordered the United States were ranked on “quality of life,” the

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

resulting ranks could be split into five equal groups: two different classes, and the limits of the lowest and
no numeric information would be necessary to create highest classes reflect the lowest and highest values in
the classification. A fourth advantage is that if enu- the data.
meration units are approximately the same size, each A major disadvantage of the mean–standard devia-
class will have approximately the same map area. tion method is that it works well only with data that are
The quantiles method shares the major disadvantage normally distributed. This is particularly evident with
of equal intervals: It fails to consider how the data are the foreign-born data, in which the two lowest classes
distributed along the number line. For example, note contain solely negative values and therefore have no
that for the foreign-born data, the outlier for Dade members (Figure 2C). One solution to this problem is
County is included in the same class with values of con- to transform the data, but this is inappropriate if the
siderably lower magnitude (Figure 1B). Thus, we have intention is to examine the raw data. Another disadvan-
eliminated the problem of empty classes, but distinctly tage is that the mean–standard deviation method
unlike data values have been placed in the same class. requires an understanding of some basic statistical con-
Another disadvantage of quantiles is that gaps result cepts; a message on the map or in the text indicating
(for the foreign-born data, the highest value in class 2 is that “classes were developed based on the mean and
2.7 and the lowest value in class 3 is 3.1, resulting in a standard deviation” would not be meaningful if one had
gap of 0.4; Figure 2B). Gaps are problematic because no statistical training.
the reader might wonder why they occur. However, they A distinct advantage of the mean–standard devia-
do permit the legend to reflect the range of data actu- tion method, however, is that if the data are normally
ally occurring in each class. distributed (or near normal), the mean serves as a
useful dividing point, enabling a contrast of values
1.3 Mean–Standard Deviation above and below it. This is most effectively accom-
plished if an even number of classes is used; for exam-
The mean–standard deviation method is one of several ple, a six-class map could consist of the positive
classification techniques that do consider how data are q to x
classes ( x q + s, xq + s to x
q + 2s, and 7 x q + 2s)
distributed along the number line. In this method, classes and the negative classes ( x q to x
q - s, qx - s to x
q - 2s,
are formed by repeatedly adding or subtracting the stan- and 6 x q - 2s). (For the five-class map shown in
dard deviation from the mean of the data, as shown in Figure 2C, the two middle classes were combined.)
Table 3.* As with the equal-intervals method, both cal- Another advantage is that the legend contains no gaps
culated and legend limits can be computed. Calculated that might confuse the reader.
limits are computed using the mean and standard devia-
tion values listed in column 2 of Table 3 (Normal 1.4 Maximum Breaks
Distribution Limits). To create legend limits, calculated
limits are adjusted so that identical values cannot fall in The mean–standard deviation method considers how
data are distributed along the number line in a holistic
sense—by trying to fit a normal distribution to the data.
An alternative approach is to consider individual data
values and group those that are similar (or, alterna-
TABLE 3 Class limit computations for mean–standard
deviation classification tively, avoid grouping those that are dissimilar). The
maximum breaks method is a simplistic means for
Normal accomplishing this.† In this method, raw data are
Distribution Calculated Legend
Class Limits Limits Limits
ordered from low to high, the differences between adja-
cent values are computed, and the largest of these
Florida Foreign-Born Data: xq = 4.96, s = 5.94 differences serve as class breaks.
1 6 xq - 2s 6 - 6.9 6 - 6.9 The foreign-born data provide a good illustration of
2 xq - 2s to qx - 1s -6.9 to - 1.0 - 6.9 to - 1.0 how the maximum breaks method groups similar values
3 xq - 1s to xq + 1s - 1.0 to 10.9 - 0.9 to 10.9 (and avoids grouping dissimilar values): In Figure 1, note
4 xq + 1s to xq + 2s 10.9 to 16.8 11.0 to 16.8 how the six highest data values are broken into four
5 7 xq + 2s 7 16.8 16.9 to 45.1 classes. In addition to considering how data are distrib-
uted along the number line, the maximum breaks method
is easy to compute, simply involving subtracting adjacent
* We use mean and standard deviation formulas appropriate for
a sample; because of the relatively large number of observations
in each case, similar results would be obtained with population † We have borrowed the term “maximum breaks” from the teaching
formulas. of George Jenks.

71
16:05:32.
 Data Classification

values. A disadvantage of maximum breaks is that by TABLE 4 Computing the sum of absolute deviations
paying attention only to the largest breaks, the method about class medians (ADCM)
seems to miss natural clusters of data along the number Raw Data: 11, 12, 13, 14, 31, 32, 33, 99, 100
line. For example, for the foreign-born data, the maximum
breaks method ignores the peaks and dips in class 1 that Quantiles Classification Optimal Classification
seem to reflect natural clustering there. Class Values Error Class Values Error

1 11, 12, 13 2 1 11, 12, 13, 14 4

1.5 Natural Breaks
2 14, 31, 32 18 2 31, 32, 33 2
3 33, 99, 100 67 3 99, 100 1
The natural breaks method is one solution to the failure
of maximum breaks to consider natural groupings of ADCM = 87 ADCM = 7
data. In natural breaks, graphs (e.g., the dispersion graph
or histogram) are examined visually to determine logi-
cal breaks (or, alternatively, clusters) in the data. Stated method places similar values in the same class (the first
another way, the purpose of natural breaks is to mini- class consists of 11, 12, 13, and 14, whereas the second
mize differences between data values in the same class class consists of 31, 32, and 33).
and to maximize differences between classes. Later One measure of classification error commonly used
we will see that this is also the objective of the optimal in the optimal method is the sum of absolute devia-
method, but with the optimal method, the classification tions about class medians (ADCM). Computing this
is done using a numerical measure of classification measure involves calculating the median of each class,
error, whereas with natural breaks, the classification is which is the sum of absolute deviations of class
subjective. members about each class median, and then adding
To illustrate the computation of natural breaks, the resulting sums of absolute deviations. For exam-
consider how we might divide the foreign-born data ple, for quantiles, the median in the first class is 12
into five classes (you should examine the dispersion (remember, the median is simply the middle value in
graph associated with the natural breaks method in an ordered set), and the sum of absolute deviations
Figure 1E). The highest value in the data (45.1) for the class is |11 - 12| + |12 - 12| + |13 - 12| = 2.
appears to be quite different from the rest of the data, The sum of absolute deviations for all classes is
so we will place it in a class by itself. Our next deci- 2 + 18 + 67 = 87. In contrast, ADCM for the optimal
sion is how to handle the five values ranging from 10.1 method is 4 + 2 + 1 = 7, which is obviously a smaller
to 15.8. We could divide these into two separate value and thus is indicative of a better classification.
classes (10.1–12.2 and 14.6–15.8), but then we would The data for this hypothetical example were selected
have to divide the remaining data into two classes. It so that the results would be clear-cut. In the real world,
seems easier to divide the remaining data into three the desired minimum-error classification is normally not
classes, roughly corresponding to the three peaks in obvious, so researchers have developed computer-based
the dispersion graph, and so we group the values rang- algorithms for determining possible solutions. Here we
ing from 10.1 to 15.8 in one class. We can see in this consider two algorithms: the Jenks–Caspall and the
example that an obvious problem with natural breaks Fisher–Jenks.
is that decisions on class limits are subjective, and
therefore can vary among mapmakers. The Jenks–Caspall Algorithm
The Jenks–Caspall algorithm, developed by George Jenks
1.6 Optimal and Fred Caspall (1971), is an empirical solution to the
problem of determining optimal classes. It minimizes the
The optimal classification method is a solution to the sum of absolute deviations about class means (as opposed
limitations noted for maximum and natural breaks. The to medians). The algorithm begins with an arbitrary set of
optimal method places similar data values in the same classes (say, the quantiles classes shown in Table 4), calcu-
class by minimizing an objective measure of classifica- lates a total map error analogous to ADCM (but involv-
tion error. To illustrate, consider how a small hypothet- ing the mean), and attempts to reduce this error by mov-
ical data set of nine values would be classified by the ing observations between adjacent classes. Observations
quantiles and optimal methods (Table 4). The quantiles are moved using what Jenks and Caspall termed reitera-
method assigns the same number of observations to tive and forced cycling. In reiterative cycling, movements
each class (three, in this case), and thus places similar are accomplished by determining how close an observa-
values in different classes (e.g., 14 appears with 31 and tion is to the mean of another class; for example, for the
32 in class 2, even though 14 is more similar to 11, 12, quantiles data in Table 4, the value 14 is closer to the
and 13, the members of class 1). In contrast, the optimal mean of class 1 (12) than 13 is to the mean of class 2

16:05:32.
 Data Classification

(25.7), so 14 would be moved to the first class. Movements map. They computed a measure of boundary accuracy
based on the relation of observations to class means are by dividing the sum of the n actual cliffs used on the
repeated until no further reductions in total map error classed map by the sum of the n largest cliffs occurring
can be made. in the raw data.
In forced cycling, individual observations are moved Of these three kinds of error, cartographers (including
into adjacent classes, regardless of the relation between Jenks himself) have focused on tabular error, primarily
the mean value of the class and the moved observation. because of its simplicity. We have presented the three
After a movement, a test is made to determine whether different kinds of error, however, to provide a broader
any reduction in total map error has occurred. If error perspective on the classification problem. In section 3, we
has been reduced, the new classification is considered will return to the notion of the boundary error when we
an improvement, and the movement process continues consider the possibility of using multiple criteria to deter-
in the same direction. Forcing is done in both direc- mine class intervals.
tions (from low to high classes and from high to low
classes). At the conclusion of forcing, the reiterative The Fisher–Jenks Algorithm
procedure described earlier is repeated to see whether In contrast to the empirical approach used by Jenks and
any further reductions in error are possible. Although Caspall, the Fisher–Jenks algorithm has a mathematical
this approach does not guarantee an optimal solution, foundation that guarantees an optimal solution. Walter
Jenks and Caspall (1971, 236) indicated that they were Fisher (1958) was responsible for developing the mathe-
“unable to generate, either purposefully or by accident, matical foundation, and George Jenks (1977) introduced
a better . . . representation in any set of data.” the idea to cartographers. Cartographers generally have
In addition to developing an automated algorithm chosen to recognize only Jenks for this contribution, so
for determining optimal classes, Jenks and Caspall the reader might find the algorithm referred to as
(1971, 225) introduced three criteria for selecting a “Jenks’s optimal method.”
“best” classification. They introduced these criteria by To understand the Fisher–Jenks algorithm, it is
posing three questions: worthwhile to consider how an optimal solution might
be computed using brute force. Imagine that you
1. Which map provides the reader with the most wanted to develop an optimal two-class map of the
accurate intensity values for specific places? data 1, 3, 7, 11, and 22. With such a small data set, it is
2. Which map creates the most accurate overview? easy to list all possible two-class solutions and compute
3. Which map contains boundaries that occur along associated error measures (Table 5). If the process is so
major breaks in the statistical surface? simple for a small data set, it would seem that for large
data sets, a computer could be used to determine an
Corresponding to these questions, Jenks and Caspall optimal solution by simply considering all possibilities.
discussed three kinds of error: tabular, overview, and Unfortunately, for large data sets, the number of possi-
boundary. ble solutions becomes prohibitively large; for example,
To understand these forms of error, it is helpful to Jenks and Caspall (1971, 232) calculated that for the
consider a 3-D prism map, which has no error due to 102 counties of Illinois, there would be over 1 billion
classification because each county is raised to a height possible seven-class maps.
proportional to the data. If the data were classed, error Rather than consider all solutions, the Fisher–Jenks
would arise as a result of counties in the same class algorithm takes advantage of the mathematical foun-
being raised to the same height. The difference in dation provided by Fisher, which states that any opti-
height between corresponding prisms would constitute mal partition is simply the sum of optimal partitions of
tabular error, and overview error would be the differ- subsets of the data. We illustrate this concept by con-
ence in volume between corresponding prisms. Tabular sidering some initial steps for handling the data 1, 3, 7,
error is equivalent to the error measure described pre- 11, and 22 (Table 6). For computational simplicity, we
viously (minimizing the sum of absolute deviations use the median (and associated sum of absolute devia-
about the median, or mean, of each class); thus, mini- tions); another version of the algorithm uses the mean
mizing tabular error produces an optimal classification. (and associated sum of squared deviations about the
In contrast, overview error is weighted to account for mean).
the size of enumeration units. Step 1 involves computing the sum of absolute devia-
Jenks and Caspall used the term boundary error to tions about the class median for any ordered subset of
describe the error occurring along the boundary the raw data, ignoring how these subsets might fit into a
between two enumeration units on a classed map. They particular classification. For example, the sum of absolute
argued that the highest cliffs appearing on the unclassed deviations for the first through third observations (the
prism map should ideally appear on the classed prism subset 1, 3, and 7) is |1 - 3| + |3 - 3| + |7 - 3| = 6.

73
16:05:32.
 Data Classification

TABLE 5 Computing ADCM for all potential two-class TABLE 6 Initial steps in the Fisher–Jenks algorithm for
maps optimal data classification
Raw Data: 1, 3, 7, 11, 22 Raw Data: 1, 3, 7, 11, 22

Solution 1 Solution 2 Step 1. Compute the sum of absolute deviations about the class
median for all ordered subsets of the data.
Class Values Error Class Values Error
The following matrix shows the sum of absolute deviations
1 1 0 1 1, 3 2 about the median for the i th through the j th observation;
2 3, 7, 11, 22 23 2 7, 11, 22 15 for example, if i = 1 and j = 3, then the sum is
ADCM = 23 ADCM = 17 |1 - 3| + |3 - 3| + |7 - 3| = 6. (Note that this result appears
in the first row and the third column of the matrix.) The resulting
Solution 3 Solution 4 sums of absolute deviations are commonly termed the diameter (D)
and are represented by D(i, j) ; for this example D (1,3) = 6.
Class Values Error Class Values Error
j th observation
1 1, 3, 7 6 1 1, 3, 7, 11 14
1 2 3 4 5
2 11, 22 11 2 22 0
1 0 2 6 14 29
ADCM = 17 ADCM = 14
i th 2 0 4 8 23
Solution 4 is optimal because it has the smallest total error
observation 3 0 4 15
(or ADCM).
4 0 11
5 0
This result appears in row 1, column 3, of the matrix Step 2. Compute all optimal two partitions.
shown in step 1 of Table 6. The resulting sum of absolute
(a) The results for the optimal two-class map of the complete
deviations is termed the diameter (D), and is represented
data set are as follows:
by D(i, j), where i and j identify the observations in step
1 | 3 7 11 22
1 (Hartigan 1975, Chapter 6); thus, D(i, j) for this exam-
D(1,1) + D(2,5) = 0 + 23 = 23
ple would be D(1,3) = 6. 1 3 | 7 11 22
In step 2, the optimal solution for a two-class map D(1,2) + D(3,5) = 2 + 15 = 17
of the complete data set is computed, along with opti- 1 3 7 | 11 22
mal two-class solutions for subsets of the data. D(1,3) + D(4,5) = 6 + 11 = 17
1 3 7 11 | 22
Together, these are termed the optimal two partitions.
D(1,4) + D(5,5) = 14 + 0 = 14
Calculations for the optimal two-class map for the
complete data set are shown in part (a) of step 2 (b) The following are some results for optimal two-class partitions
(Table 6), and some of the results for subsets of the of subsets of the data:
data are shown in part (b) of step 2. Although the sub- 1 | 3 7 11
D(1,1) + D(2,4) = 0 + 8 = 8
set calculations are not used in determining the opti-
1 3 | 7 11
mal two-class map, they are used to determine optimal D(1,2) + D(3,4) = 2 + 4 = 6
classifications for maps with a greater number of 1 3 7 | 11
classes, such as a three-class map, for which some cal- D(1,3) + D(4,4) = 6 + 0 = 6
culations are shown in step 3. 1|3 7
D(1,1) + D(2,3) = 0 + 4 = 4
1 3|7
Advantages and Disadvantages of Optimal D(1,2) + D(3,3) = 2 + 0 = 2
Classification Ideally, all optimal two partitions would be computed!
The obvious advantage of the optimal method is that it Step 3. Compute all optimal three-partitions (let Opt-2 represent
considers, in detail, how data are distributed along the an optimal two partition.)
number line. It is the “best” choice for classification (a) The results for the optimal three-class map of the complete
when the intention is to place like values in the same data set would be calculated. Some of the calculations follow.
class (and unlike values in different classes) based on The question marks represent optimal two-class partitions not
computed above.
the position of values along the number line.
Another advantage is that the optimal method can 1 | 3 7 11 22
D(1,1) + Opt-2 = 0 + ?
assist in determining the appropriate number of classes. 1 3 | 7 11 22
When the median is used as the measure of central ten- D(1,2) + Opt-2 = 2 + ?
dency, this is accomplished by computing the goodness 1 3 7 | 11 22
of absolute deviation fit (GADF), which is defined as: Opt-2 + D(4,5) = 2 + 11 = 13
1 3 7 11 | 22
ADCM Opt-2 + D(5,5) = 6 + 0 = 6
GADF = 1 -
ADAM (b) Calculate all other optimal three-class partitions.

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

where ADCM is the sum of absolute deviations about the foreign-born data, the curve appears to flatten out at
class medians for a particular number of classes, and about five classes. A flattening at this point indicates that
ADAM is the sum of absolute deviations about the a larger number of classes would not contribute sub-
median for the entire data set. An analogous measure can stantially to a reduction in the classification error. (One
be computed when the mean is used as the measure of should also bear in mind that a map with more classes
central tendency (and the error in a class is the sum of would be more difficult to use, as a greater number of
squared deviations about the mean), and is known as the areal symbols would have to be differentiated.)
goodness of variance fit (GVF; Robinson et al. 1984, 363). Another approach is to determine the number of
GADF ranges from 0 to 1, with 0 representing the low- classes for which the GADF first exceeds a certain
est accuracy (a one-class map) and 1 representing the value, say, 0.80 (i.e., the accuracy is 80 percent). For the
highest accuracy. If there are no ties in the data, then a foreign-born data, this approach yields a six-class map
GADF value of 1 will result only when each observation (Figure 4). Note, however, that if a more stringent value
is a separate class (an n-class map will be required, where were used, say, 0.90, a nine-class map would be required.
n is the number of classes). Ties can, however, consider- Admittedly, both of these approaches are subjective,
ably reduce the number of classes needed to achieve a but they are an improvement over choosing an arbi-
GADF of 1. For example, in the case of the foreign-born trary number of classes.
data, only 46 classes are needed (there are 67 data values). In addition to helping you determine an appropri-
It is important to note that an n-class map is equivalent ate number of classes for the optimal method, the
to an “unclassed map.” This might be confusing because GADF technique could also be used to compute the
the term unclassed suggests no classes, and there are actu- accuracy of other classification methods, and thus
ally n classes on an n-class map. “Unclassed” is commonly determine whether those methods might be appropri-
used to indicate that no classing or grouping has been ate. For example, if you were to compute similar
applied to the data; for instance, it is not necessary to run GADF values for optimal and quantiles, you might
the optimal program to create an unclassed map. choose the quantiles method because it would be eas-
GADF calculations can assist in selecting an appro- ier for the user to understand how the class limits were
priate number of classes in two ways. One approach is to created. Alternatively, GADF values might be com-
construct a graph of the number of classes against puted for various numbers of classes for quantiles to
GADF values (Figure 4) and look for a point at which a assist in determining an appropriate number of classes
curve fit to the data begins to flatten out. In the case of for that method.
Disadvantages of the optimal method include the dif-
ficulty of understanding the concept and the appearance
of gaps in the legend. Another traditional disadvantage
was that software packages generally did not include the
1.0 optimal method as an option, but this has changed (e.g.,
the popular ArcGIS package includes an optimal option,
which it terms “Natural Breaks (Jenks)”).
0.8
1.7 Criteria for Selecting a Classification Method

0.6 In discussing common methods of classification, we

GADF

have pointed out numerous criteria that might be

used to judge these methods’ usefulness. Figure 5
0.4 summarizes these criteria and rates each classification
method as very good, good, or poor on each measure
(and acceptable or unacceptable in the case of
0.2 “acceptable for ordinal data”). One problem with any
rating system is that it is a function of the computer
environment the mapmaker has available and the
0 knowledge of the map user. For Figure 5, we assume
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
that computer software is available for creating all of
Number of Classes the classification methods we have considered and
that the map user is a college-level student with a
FIGURE 4 Graph of the number of classes plotted against basic foundation in introductory statistics.
GADF values. The curve appears to flatten out at about five Note that “ease of understanding legend” is a func-
classes, indicating that a greater number of classes would not tion of whether or not there are gaps in the legend:
substantially reduce the error on the map. Remember that gaps between class limits can make the

75
16:05:32.
 Data Classification

FIGURE 5 Criteria for selecting a method of classification.

legend difficult to understand. The equal-intervals An analysis of Figure 5 reveals that there is no single
method receives a very good rating on this criterion best method of classification. Although the optimal
because not only are there no gaps, but the rounded approach is often touted as the best method, it is best
intervals can be very easy to understand (e.g., 0–25, only in terms of grouping like values together (as a
26–50, etc.).You might wish to avoid the problem of gaps function of their position along a number line) and in
by creating continuous legends for all classification selecting an appropriate number of classes. Clearly,
methods (as in Figure 3). Remember, though, that this there are several other criteria for which it is not the
approach will not indicate the actual range of values best. Ultimately, you must consider the purpose of the
falling in a class (the latter is dealt with in the criterion map and the knowledge of the intended audience
“legend values match range of data in a class”). before selecting a classification method.

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

A good illustration of the role of map purpose and between 6 and 9 and between 13 and 16. The GADF
intended audience is the effort of Cynthia Brewer and for this map is 0.76, and the complexity using the CF
Trudy Suchan to create an atlas of the 2000 U.S. Census face measure is 10/16 or 0.63.
data (Brewer 2001). Rather than use one of the stan- A spatial constraint can be applied by starting with
dard classification methods presented in this chapter, the optimal solution and then allowing data values to
Brewer and Suchan used meaningful breaks (e.g., a per- shift between classes so that the map pattern is simpli-
centage figure for the entire United States), rounded fied. One potential solution is shown in the right-hand
breaks, and breaks that were identical across a set of map, where the values 9 and 16 have been shifted from
maps. Brewer (2001, 225) argued that the resulting map the second and third classes to the first and second
set is much more useful than a “map-by-map optimiza- classes, respectively. The result is a slightly lower GADF
tion approach.” (0.70 as opposed to 0.76), but a considerably simpler
complexity (0.38 as opposed to 0.63).
2 USING SPATIAL CONTEXT TO SIMPLIFY Franky Declercq (1995) developed an algorithm
CHOROPLETH MAPS that will find the least complex map for a particular
number of classes, assuming that one is willing to shift
a certain percentage of observations between classes.
One limitation of the classification methods described
Obviously, one problem with this approach is that there
in the preceding section is that they do not consider the
is a trade-off between complexity and accuracy: As
spatial context of the data. If the purpose of classifica-
complexity decreases, accuracy will also decrease.
tion is to simplify the appearance of the map, it can be
Consequently, there is some subjectivity in determin-
argued that spatial context also should be considered.
ing what magnitude of GADF is acceptable for a
This section considers two approaches that do consider
particular application.
spatial context: one that begins with the optimal classi-
fication approach but incorporates a spatial constraint,
2.2 Spatial Context Only
and one that works solely with spatial context.
Another approach for simplifying map pattern is to
2.1 Optimal Classification with a Spatial Constraint adjust the values of enumeration units as a function of
values of surrounding units, an idea developed by Waldo
To illustrate how a spatial constraint can be combined Tobler (1973) and promoted by Adrian Herzog (1989).
with optimal classification, consider Figure 6. The top This notion is based on the concept of random error: If
portion of the figure lists raw data for 16 hypothetical we repeatedly measure a value for an enumeration unit,
enumeration units, and the bottom portion portrays we will likely get a different value each time (think of
two maps of these data. The map on the left is an opti- trying to calculate the population for a census tract for
mal classification; note that breaks for this map occur the 2000 Census—numerous variables could affect your
result). This suggests that some change in the data we
have collected is permissible (Clark and Hosking 1986,
Raw Data 14). Moreover, the spatially autocorrelated nature of
geographic data provides a mechanism for making
1, 2, 3, 4, 5, 6, 9, 10, 11, 12, 13, 16, 17, 18, 19, 20 adjustments: Nearby units can assist in determining
appropriate values for a particular enumeration unit
because similar values are likely to be located near one
Optimal Map another.
Optimal Map (with a spatial constraint)
Although this approach could be used with a wide
1 9 5 6 1 9 5 6 variety of data, it is easiest to implement for propor-
tion data (e.g., the proportion of adults who smoke cig-
20 3 16 12 20 3 16 12 arettes) because the statistical theory for such data is
well known. To illustrate, consider the hypothetical
19 4 13 11 19 4 13 11
portion of a map shown in Figure 7, where it is
18 10 17 2 18 10 17 2 assumed we wish to change the value for the central
enumeration unit as a function of the surrounding
GADF = 0.76 GADF = 0.70 units. The simplest formula for determining the value
CF = 10/16 = 0.63 CF = 6/16 = 0.38 of the central unit would be to average all four values,
but such a formula would not consider two factors:
(1) that we have presumably taken some care in col-
FIGURE 6 Applying a spatial constraint to the optimal clas- lecting the data for the central unit (and thus would
sification approach. like to place greater weight on it), and (2) that those

77
16:05:32.
 Data Classification

outside of a 95 percent confidence interval around its

original value (see Burt and Barber 1996, 272–274, for
a discussion of confidence intervals for proportions).
Second, surrounding values differing dramatically from
the central value should not be used in the calculations.
0.22
For proportion data, Herzog suggested a difference of
0.26 proportions test (see Burt and Barber 1996, 322–324),
with proportions differing significantly from the cen-
tral unit not being used. Third, the process should be
0.15 implemented in an iterative fashion, meaning that the
entire map should be simplified several times; for
example, this process could continue until all enumera-
tion units reached the bounds of their confidence inter-
vals or until the largest change made in a unit was less
0.19 than a small tolerance. Finally, the map resulting from
the iterative process should be further smoothed by
combining areas with nearly equal values.
Figure 8 illustrates the net effect of using Herzog’s
approach for the percentage of students in each com-
mune of Switzerland who study at the University of
FIGURE 7 A hypothetical set of enumeration units for which a
Zurich: The top map is an original (unsimplified)
simplified value is desired for the central enumeration unit.
unclassed map, and the bottom map is the simplified
result. Although the bottom map is still unclassed, note
that it is much simpler to interpret than the top map.
units having a longer common boundary with the cen- It should be noted that the methods we have exam-
tral unit should have greater impact. Thus, a more ined for simplifying map appearance are relatively new
appropriate formula is and have not yet been incorporated into commercial
n software. Assuming that such methods do become read-
Ve = Wc Vc + Ws a a Vi b
Li
ily available, it will be interesting to see to what extent
i=1 LT
they are used, and what impact their use has on the use
where Ve = the estimate value for the central unit of more traditional classification methods that do not
Vc = the original value for the central unit consider spatial context.
Vi = the original value for the ith
surrounding unit
Wc = the weight for the central unit 3 USING MULTIPLE CRITERIA TO DETERMINE CLASS
Ws = the weight for the surrounding units INTERVALS
Li = the length of the boundary between
the ith unit and the central unit In section 2.1, we showed how we might combine the
LT = the total length of the central optimal classification approach with a spatial con-
unit boundary straint. We utilized two criteria to accomplish this: (1)
n = the number of surrounding units minimizing the difference between data values in
classes (i.e., minimizing the tabular error, a nonspatial
If we assume Wc and Ws have values of 0.67 and 0.33, constraint), and (2) simplifying the map pattern by
respectively, then for Figure 7, the formula would be defining class limits such that contiguous enumeration
calculated as follows: units fall in the same class (a spatial constraint). Marc
Armstrong and his colleagues (2003) have argued that
Ve = .67(.15) + .33 a (.26) +
1
6 we should consider a broader set of criteria when clas-
sifying data, including minimizing the tabular error,
(.22) + (.19) b = .17
2 3
6 6 minimizing the difference between classes for contigu-
ous enumeration units (i.e., minimizing the boundary
Herzog indicated that a more complete algorithm error), maximizing the overall spatial autocorrelation
should involve several other considerations. First, some in the resulting classed map, and equalizing the area in
limit should be placed on how much change is permit- each class (which is often useful when we wish to com-
ted to the central unit. For proportion data, Herzog pare choropleth maps). Note that the latter three of
suggested that the central value should not be moved these are all spatial constraints.

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

Percentage of Students Studying

at the University of Zurich

University of Zurich

Percentage of Students Studying

at the University of Zurich

University of Zurich

FIGURE 8 A comparison of an original unclassed map (top) and a simplified map (bottom) resulting from applying
Herzog’s method for simplifying choropleth maps. (An adaptation of Figure 1 from p. 214 of Accuracy of Spatial Databases,
M. Goodchild and S. Gopal (eds.), 1989; courtesy of Taylor & Francis.)

79
16:05:32.
 Data Classification

Armstrong et al. considered these multiple criteria by criterion we choose to focus on. If you compare Figures
using the technique of genetic algorithms (Michalewicz 2 and 9, you will note that the “Minimize Tabular Error”
1998), which “do not guarantee a global optimum solu- map matches exactly the “Optimal (median)” map. This
tion, [but] normally yield very good solutions to large should not be surprising because, as we indicated in
problems” (Armstrong et al., 602). Since a full under- section 1.6, minimizing the tabular error is equivalent to
standing of genetic algorithms is beyond the scope of the producing an optimal classification. Also note that the
present text, we will focus on the results Armstrong et al. “Equalize Area in Each Class” map is similar in appear-
obtained when utilizing genetic algorithms. Figure 9 illus- ance to the “Quantiles” map. This also should not be sur-
trates the four maps for the Florida foreign-born data prising, either, because we indicated that if enumeration
that result if we consider only one of the four criteria at units are approximately the same size, each class on a
a time. As with Figure 2, we see that the appearance of quantiles map will have approximately the same map
the distribution obviously differs as a function of which area. In contrast to these similarities, note that the

Armstrong et al.’s Multiple Criteria for Data Classification

Foreign Born in Florida, 1990

A B
Minimize Tabular Error Minimize Boundary Error

Percentage Percentage
Foreign Born Foreign Born
0.6 to 2.4 0.6 to 3.6
2.7 to 4.9 3.7 to 7.6
5.2 to 8.3 8.3 to 12.2
10.1 to 15.8 14.6 to 15.8
45.1 45.1

C D
Maximize Spatial Autocorrelation Equalize Area in Each Class

Percentage Percentage
Foreign Born Foreign Born
0.6 to 0.9 0.6 to 1.6
1.0 to 3.6 1.7 to 3.5
3.7 to 10.1 3.6 to 5.4
10.5 to 12.2 5.5 to 7.5
14.6 to 45.1 7.6 to 45.1

FIGURE 9 Maps that are ideal from the standpoint of the data classification criteria considered by Armstrong et al. (2003): (A)
minimizes the tabular error, (B) minimizes the boundary error, (C) maximizes the spatial autocorrelation, and (D) equalizes the
area in each class.

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

“Maximize Spatial Autocorrelation” and “Minimize are labeled EVF and GEA. EVF stands for “equality of
Boundary Error” maps provide a different view of the variance fit” and is a measure of tabular error (in this
data than the standard methods of classification shown case, minimizing the sum of squared deviations about
in Figure 2. class means); low EVF values are desirable. The G in
Rather than focusing on only one criterion, the pro- GEA stands for the “Gini coefficient” and EA stands for
gram ChoroWare (Xiao and Armstrong 2006) can be uti- “equal areas.” The Gini coefficient is an index of dissim-
lized to trade off one criterion for another. For instance, ilarity (Hammond and McCullagh 1978; Smith 1977) that
Figure 10 shows two maps that represent trade-offs measures the difference between the areas in each class
between minimizing tabular error and equalizing the for a given map with a theoretical best map in which the
area in each class. Also shown in this figure is a scatter- area in each class is identical. Low values on GEA are
plot of potential solutions that arise when considering also desirable because a low value on the Gini coefficient
these two criteria. Note that the axes on the scatterplot tends to equalize the area in each class. Also note that

Trade-Offs Using Armstrong et al.’s Multiple Criteria

Foreign Born in Florida, 1990

A B
Trade-Off #1 (TO1) Trade-Off #2 (TO2)

Percentage Percentage
Foreign Born Foreign Born
0.6 to 2.3 0.6 to 1.7
2.4 to 4.3 1.8 to 3.7
4.5 to 6.8 4.1 to 6.0
7.1 to 15.8 6.1 to 10.5
45.1 12.2 to 45.1

Best
EVF

TO1
Best
GEA
TO2

FIGURE 10 Trade-offs using Armstrong et al.’s (2003) multiple criteria approach: (A) the EVF (the tabular error) is close to the pos-
sible minimum and the GEA is also reasonably low (classes are close to being equal in area); (B) the GEA is very low and EVF is
moderate; (C) a scatterplot of potential values of EVF and GEA determined using the genetic algorithm approach.

81
16:05:32.
 Data Classification

the Best EVF and Best GEA values (corresponding to One limitation of traditional classification methods
maps A and D, respectively, in Figure 9) fall to the (e.g., equal intervals and optimal) is that they fail to con-
extreme left and bottom of the scatterplot, reflecting the sider the spatial context of the data. If the purpose of
fact that these values are minimized. In contrast, the two classification is to simplify the appearance of a map, it
trade-off maps (TO1 and TO2) are not minimized on can be argued that spatial context should also be consid-
either criterion, but represent a balance between the cri- ered. In this chapter, we initially looked at two
teria. For instance, TO1 has a value that is very close to approaches that consider spatial context when simplify-
the minimum EVF value, but is also a reasonably low ing a choropleth map. The first used the optimal classifi-
GEA value. In contrast, TO2 has a very low GEA value cation method but incorporated a spatial constraint by
and a moderate EVF value. Selection of one of these two requiring that values falling in the same class be contigu-
solutions would be a function of the relative weight the ous, whereas the second employed no classification
user wished to specify for the two criteria. approach, but rather smoothed the raw data by changing
values for enumeration units as a function of values of
SUMMARY neighboring units. We then examined an approach based
on genetic algorithms that allowed us to consider a broad
range of criteria when selecting class intervals, including
In this chapter, we initially examined several common
the tabular error, the boundary error, the spatial auto-
methods of data classification that combine raw data
correlation of the data, and a desire to equalize the area
values into groups (or classes). Methods that do con-
in each class. We saw that optimizing one criterion (say,
sider how data are distributed along the number line
tabular error) would not necessarily optimize another
(e.g., natural breaks and optimal) are desirable
(say, boundary error), but that the program ChoroWare
because they place similar data values in the same
could be used to trade off one criterion for another.
class (and dissimilar data values in different classes).
In the simplest view, data classification involves
Methods that do not consider the distribution of data
grouping observations that have similar scores on a sin-
along the number line (e.g., equal intervals and
gle numeric attribute such as median income in
quantiles) might, however, also be desirable because
California counties. It is also possible, however, to con-
they satisfy other criteria. For example, equal intervals
sider classifying observations on the basis of multiple
is desirable because it is easy to understand, easy to
attributes; for example, we might ask which counties
compute, and has an easily understood legend (at least
have similar incomes and similar voting behavior. Such
when rounded percentage data are used, such as 0–100
a process is termed cluster analysis.
percent urban population).

FURTHER READING

Armstrong, M. P., Xiao, N., and Bennett, D. A. (2003) “Using Cromley, R. G. (1996) “A comparison of optimal classification
genetic algorithms to create multicriteria class intervals for strategies for choroplethic displays of spatially aggregated
choropleth maps.” Annals of the Association of American data.” International Journal of Geographical Information
Geographers 93, no. 3:595–623. Systems 10, no. 4:405–424.
Uses a variety of data sets to illustrate the notion of considering Illustrates a variety of criteria for optimally classifying data by
multiple criteria in data classification. treating classification as an integer programming problem; also
Brewer, C. (2001) “Reflections on mapping Census 2000.” see Cromley and Mrozinski (1997; 1999).
Cartography and Geographic Information Science 28, no. Evans, I. S. (1977) “The selection of class intervals.”
4:213–235. Transactions, Institute of British Geographers (New Series) 2,
Describes how meaningful breaks, rounded breaks, and breaks
no. 1:98–124.
that were identical across a set of maps were used to design an A classic article on methods of data classification.
atlas of 2000 U.S. Census data.
Herzog, A. (1989) “Modeling reliability on statistical surfaces
Coulson, M. R. C. (1987) “In the matter of class intervals for by polygon filtering.” In Accuracy of Spatial Databases, ed. by
choropleth maps: With particular reference to the work of M. Goodchild and S. Gopal, pp. 209–218. London: Taylor &
George F. Jenks.” Cartographica 24, no. 2:16–39. Francis.
Summarizes data classification methods, with an emphasis on A detailed discussion of Herzog’s method for simplifying patterns
Jenks’s work. on choropleth maps.

Slocum, Terry A., et al. Thematic Cartography and Geovisualization:

16:05:32.
 Data Classification

Jenks, G. F., and Caspall, F. C. (1971) “Error on choroplethic Monmonier, M. S. (1982) “Flat laxity, optimization, and
maps: Definition, measurement, reduction.” Annals, Associ- rounding in the selection of class intervals,” Cartographica
ation of American Geographers 61, no. 2:217–244. 19, no. 1:16–27.
The classic article on optimal data classification. Argues that round-number breaks should be used with optimal
data classification to create a more readable map.
Lindberg, M. B. (1990) “Fisher: A Turbo Pascal unit for optimal
partitions.” Computers & Geosciences 16, no. 5:717–732.
Describes a computer implementation of the Fisher–Jenks optimal
method of data classification.

GLOSSARY

boundary error: if classed data are conceived as a prism maximum breaks: a method of data classification in which
map, boundary error describes how close the resulting cliffs the largest differences between ordered observations are used
come to matching the cliffs on an unclassed prism map of the to define classes.
data. mean–standard deviation: a method of data classification in
class interval: in creating an equal-interval map, the width which the mean and standard deviation of the data are used
that each class occupies along the number line. to define classes.
classed map: a map in which data are grouped into classes of natural breaks: a method of data classification in which a
similar value, and the same symbol is assigned to all members graphical plot of the data (e.g., a histogram) is examined to
of each class (e.g., a data set with 100 different data values determine natural groupings of data.
might be depicted using only five shades of gray). optimal: a method of classification in which like values are
cluster analysis: a mathematical method for grouping obser- placed in the same class by minimizing an objective measure
vations (say, counties) based on their scores on a set of attrib- of classification error (e.g., by minimizing the sum of absolute
utes. deviations about class medians).
equal-intervals: a method of data classification in which overview error: if classed data are conceived of as a prism
each class occupies an equal portion of the number line. map, overview error is the difference in volume between this
Fisher–Jenks algorithm: a method for classifying data in map and an unclassed prism map of the data.
which an optimal classification is guaranteed by essentially quantiles: a method of data classification in which an equal
considering all possible classifications of the data. number of observations is placed in each class.
goodness of absolute deviation fit (GADF): a measure of random error: the notion that if we repeatedly measure a
the accuracy of a classed choropleth map when the median is value for an enumeration unit, we will likely get a different
used as a measure of central tendency. value each time.
goodness of variance fit (GVF): a measure of the accuracy tabular error: if classed data are conceived as a prism map,
of a classed choropleth map when the mean is used as a mea- tabular error is the difference in height between prisms on
sure of central tendency. this map and those on an unclassed prism map of the data.
Jenks–Caspall algorithm: an optimal data classification that unclassed map: a map in which data are not grouped into
is achieved by moving observations between classes through classes of similar value, and thus each data value can theoret-
trial-and-error processes known as reiterative and forced ically be represented by a different symbol (e.g., a data set
cycling. with 100 different data values might be depicted using 100 dif-
ferent shades of gray on a choropleth map).

REFERENCES

Armstrong, M. P., Xiao, N., and Bennett, D. A. (2003) “Using Cromley, R. G., and Mrozinski, R. D. (1997) “An evaluation
genetic algorithms to create multicriteria class intervals for of classification schemes based on the statistical versus the
choropleth maps.” Annals of the Association of American spatial structure properties of geographic distributions in
Geographers 93, no. 3:595–623. choropleth mapping.” 1997 ACSM/ASPRS Annual
Brewer, C. A. (2001) “Reflections on mapping Census 2000.” Convention & Exposition, Technical Papers, Volume 5
Cartography and Geographic Information Science 28, no. (Auto-Carto 13), Seattle, WA, pp. 76–85.
4:213–235. Cromley, R. G., and Mrozinski, R. D. (1999) “The classifica-
Burt, J. E., and Barber, G. M. (1996) Elementary Statistics for tion of ordinal data for choropleth mapping.” The
Geographers (2nd ed.). New York: Guilford. Cartographic Journal 36, no. 2:101–109.
Clark, W. A. V., and Hosking, P. L. (1986) Statistical Methods Declercq, F. A. N. (1995) “Choropleth map accuracy and the
for Geographers. New York: Wiley. number of class intervals.” Proceedings of the 17th

83
16:05:32.