3 Describing two categorical variables

In Section 2 we introduced several measures for describing one variable. In this and the
following sections we will introduce some methods for describing two variables simultaneously.
In this section, in particular, we consider the case of two categorical variables.

Contingency tables
The simplest, but usually quite telling, tool for describing two categorical variables is a
contingency table. Contingency tables are also known as cross tables.
Definition 38. Let xi and yi be the values of two categorical variables associated to the ith
individual in the population U (i = 1, 2, · · · , N ), where x takes Kx different categories and y
takes Ky different categories. A contingency table is a matrix-like table that shows, in the cell
(kx , ky ), the frequency of elements taking the kx th category of x and the ky th category of y
simultaneously (for kx = 1, 2, · · · , Kx and ky = 1, 2, · · · , Ky ). 
In simple words, a contingency table is a table that shows in each cell the frequency of one
category of one variable x and one category of the second variable y. This is one of the many
situations in which things are simpler than they sound: aqui
Example 39. Let U be the population of N = 120 students taking a course in statistics. Let
xi be the grade in the first assignment (Pass or Fail) and yi be the grade in the exam (A, B,
C, D, E, F) for the ith student (i = 1, 2, · · · , N ). Table 18 shows the values of x and y in the
population of students.
x y x y x y x y x y x y x y x y
Fail F Pass F Pass F Pass F Pass E Pass D Pass C Pass B
Fail F Pass F Pass F Pass F Pass E Pass D Pass C Pass B
Fail F Pass F Pass F Pass F Pass E Pass D Pass C Pass B
Fail F Pass F Pass F Pass F Pass D Pass D Pass C Pass B
Fail F Pass F Pass F Pass F Pass D Pass D Pass C Pass B
Fail F Pass F Pass F Pass F Pass D Pass D Pass C Pass B
Fail F Pass F Pass F Fail E Pass D Pass D Pass C Pass B
Fail F Pass F Pass F Pass E Pass D Pass D Pass C Pass B
Fail F Pass F Pass F Pass E Pass D Pass D Pass C Pass B
Pass F Pass F Pass F Pass E Pass D Pass D Pass C Pass B
Pass F Pass F Pass F Pass E Pass D Pass D Pass C Fail A
Pass F Pass F Pass F Pass E Pass D Fail C Pass C Pass A
Pass F Pass F Pass F Pass E Pass D Pass C Pass C Pass A
Pass F Pass F Pass F Pass E Pass D Pass C Pass C Pass A
Pass F Pass F Pass F Pass E Pass D Pass C Fail B Pass A
Table 18: Results of N = 120 students in an assignment and an exam in statistics

Note that x takes Kx = 2 different categories and y takes Ky = 6 different categories.

Therefore these values can be summarized in a contingency table of size 2 × 6 as shown
below: 
y
A B C D E F Total
x Pass 4 10 17 23 11 42 107
Fail 1 1 1 0 1 9 13
Total 5 11 18 23 12 51 120

33
 Note that we added the totals to the rows and to the columns. These totals are known as
the marginals. They show the univariate distributions of each variable, i.e. the values of each
variable disregarding the other.

Joint relative distribution

Dividing each entry in the contingency table by N we obtain the proportion of elements
that fall in each cell. This is known as the joint relative distribution.

Example 40. Now, we divide each entry of the contingency table in Example 39 by N = 120
to obtain the joint relative distribution: 

y
A B C D E F Total
x Pass 0.0333 0.0833 0.1417 0.1917 0.0917 0.35 0.8917
Fail 0.0083 0.0083 0.0083 0 0.0083 0.075 0.1083
Total 0.0417 0.0917 0.15 0.1917 0.1 0.425 1.0000

One way for represent graphically the information provided by a contingency table or the
joint relative distribution is through mosaic plots. In a mosaic plot each cell of the table is
represented by a rectangle, whose area is proportional to the value of the cell. Figure 25 shows
a mosaic plot of the grade in the assignment and the exam for the population of 120 students.
Note, for instance, that as the number of students who passed the assignment and got A in
the exam is four times bigger than the number of students who failed the assignment and got
A in the exam, therefore, in the mosaic plot, the former is represented by a rectangle that is
four times bigger than the latter.

Figure 25: Mosaic plot of the grades in a home assignment and an exam of 120 students.

Conditional distributions
Dividing each cell by the row-totals gives the conditional distribution of y given x.

Example 41. In order to obtain the conditional distribution of the grade in the exam y
conditioned on the grade on the assignment x, we divide each entry by the corresponding
row-total:
y
A B C D E F Total
x Pass 0.0374 0.0935 0.1589 0.215 0.1028 0.3925 1.000
Fail 0.0769 0.0769 0.0769 0 0.0769 0.6923 1.000
Total 0.0417 0.0917 0.15 0.1917 0.1 0.425 1.000

34
This table allows us to see some facts that are not so evident from the previous tables, for
instance, among students who fail the assignment, almost 70% fail the exam too; while among
students who pass the assignment, only around 40% fail the exam. 

In this case, we are obtaining the distribution of y for each value of x. For instance, in
the first row we are considering only the students who passed the assignment. Among them
3.74% got A in the exam, 9.17% got B and so on. In the second row we are considering only
the students who failed the assignment. Among them 7.69% got B in the exam, 7.69 got B
and so on.
Conditional distributions can be represented graphically through stacked bar charts as
follows. As we are conditioning on x we create Kx bars of length one. Then, each bar is
subdivided according to the conditional frequencies of the categories of y. Figure 26 shows a
stacked bar chart for the distribution of the grades in the exam conditioned on the grade in
the assignment.

Figure 26: Stacked bar chart for the distribution of grades in the exam conditioned on the
grade in the assignment

Dividing each cell by the column-totals gives the conditional distribution of x given y.

Example 42. In order to obtain the conditional distribution of the grade in the assignment
x conditioned on the grade in the exam y, we divide each entry by the corresponding column-
total:

y
A B C D E F Total
x Pass 0.8 0.9091 0.9444 1 0.9167 0.8235 0.8917
Fail 0.2 0.0909 0.0556 0 0.0833 0.1765 0.1083
Total 1.00 1.00 1.00 1.00 1.00 1.00 1.00

This table says, for instance, that considering only the students who got A in the exam, 80%
of them passed the assignment too whereas 20% failed it; considering only the students who
got B in the exam, 91% of them passed the assignment too whereas 9% failed it; etc.
Figure 27 shows a stacked bar chart for the distribution of the grades in the assignment
conditioned on the grade in the exam. 

35
Figure 27: Stacked bar chart for the distribution of grades in the assignment conditioned on
the grade in the exam

4 Describing one numerical and one categorical variable

In Section 2 we introduced several measures (numerical and graphical) for describing a nu-
merical variable. If we want to describe simultaneously one numerical variable x and one
categorical variable y, basically what we do is to calculate the descriptive measures of x for
each category of y. In other words, we calculate descriptive measures of x conditioned on y.
We will illustrate the idea with an example. Let the population U of interest be the days
in 2012. We are interested in describing the relationship between the average wind speed
(= x, in miles per hour, mph) and the month of the year (y). Note that the wind speed x is
a numerical variable, whereas the month y is a categorical variable.
Although 2012 was a leap year, we have no measurements for December 31, thus the size
of our population is N = 365.
Just for completeness and as a kind of reminder, we will start by summarizing the numerical
variable wind speed alone. The third column of Table 19 shows the descriptive parameters of
the wind speed. We can see that the daily wind speed spanned from 0 to 6.8 although the
central part of the measurements spanned between 0.5 and 2.4. The daily wind speed has a
mean of 1.55, and half of the days it was below 1.10. On average, the observations deviate
from the mean by 1.32 mph. Also, the wind speed exhibits some positive skewness.

With Without
Parameter Notation outliers outliers
Min x(1) 0 0
First quartile x̆0.25,U 0.5 0.5
Mean x̄U 1.55 1.48
Median x̆U 1.10 1.1
Mode ẋU 0.5 0.5
Third quartile x̆0.75,U 2.4 2.35
Max x(N ) 6.8 5.1
Range rangex,U 6.8 5.1
Interquartile range IQRx,U 1.9 1.85
Standard deviation Sx,U 1.32 1.20
Skewness Skx,U 1.14 0.88

Table 19: Descriptive parameters of the daily average wind speed in 2012 (in mph)

36
 Figure 28 shows some graphical summaries of the daily wind speed. On the top panel we
find a dotplot, on the central–left panel we find the histogram, a boxplot is shown on the
central–right panel and a time–series plot is shown in the bottom panel. A time–series plot
is a convenient way for illustrating data that is measured in time (as the variable average
wind speed) it is constructed by placing the time (days, in our case) along the horizontal axis
and the measurements along the vertical axis. The data is skewed towards the right: most
observations take “small values” and a few observations take “large values”. This is, of course,
in agreement with the skewness which equal 1.14. This is evidenced by the “tail” that extends
towards the right in the dotplot and the histogram or by the top half of the boxplot being
much longer that the bottom half. The fact that most observations take “small values” and a
few observations take “large values” can also be evidenced in the time–series plot.

Figure 28: Dotplot (top panel), histogram (center–left), boxplot (center–right) and time–series
plot (bottom) of the daily average wind speed in Hopkins Memorial Forest during 2012.

The boxplot in Figure 28 highlights some points that can be regarded as outliers. These
observations correspond to the 25th of February, the 26th of March, the 29th and 30th of
October and the 16th and 21th of December. All summary measures were recalculating after
removing these observations from the dataset. The results are shown in the last column of
Table 19. We can see that, although the outliers have some impact in some parameters it is
not very large, thus we will keep working with the whole set of observations.
The time–series plot in Figure 28 shows another interesting fact: during summer (june to
september) the wind speed is slower than during the rest of the year. It also varies less. Let
us dig more into this fact by analyzing simultaneously the average wind speed and the month
of the year.

37
 Month x1 x̆0.25 x̄ x̆ ẋ x̆0.75 xN rangex IQRx Sx Skx
Jan 0.6 1.30 2.43 2.50 2.6 3.35 5.1 4.5 2.05 1.24 0.27
Feb 0.2 1.60 2.34 2.50 2.4, 2.6, 2.7 2.70 6.3 6.1 1.10 1.17 0.95
Mar 0.5 0.85 2.12 1.60 0.8 3.45 5.6 5.1 2.60 1.45 0.65
Apr 0.2 1.83 2.45 2.60 2.7 3.08 4.7 4.5 1.25 1.06 -0.06
May 0.0 0.50 0.98 0.70 0.5 1.25 2.8 2.8 0.75 0.73 1.03
Jun 0.2 0.50 0.83 0.70 0.5, 0.7 1.12 2.0 1.8 0.62 0.48 0.80
Jul 0.1 0.30 0.52 0.50 0.1, 0.6 0.70 1.2 1.1 0.40 0.31 0.41
Aug 0.1 0.20 0.41 0.30 0.1, 0.2 0.50 1.4 1.3 0.30 0.32 1.42
Sep 0.1 0.40 0.69 0.65 0.8 0.98 1.5 1.4 0.58 0.43 0.37
Oct 0.0 0.45 1.61 1.20 0.3 2.05 6.8 6.8 1.60 1.63 1.66
Nov 0.1 0.52 1.74 1.55 0.1 2.70 4.7 4.6 2.17 1.36 0.54
Dec 0.4 1.90 2.62 2.35 2.2 3.67 5.4 5.0 1.77 1.39 0.41

Table 20: Descriptive parameters of the daily average wind speed in 2012 (in mph) by month

Table 20 shows some descriptive measures of the daily average wind speed by month. Let
us take a look at the mean and the median: they take the smallest values during August
and they increase month after month until the end of the year. The year starts with similar
values as those in December and then the average wind speed decreases until July, with one
exception: March. It is uncertain if March 2012 was particularly calm or if that is pattern
that repeats itself every year. Something similar happens with the variability, the smallest
variability occurs during July/August and then it increases until the end of the year. Figure
29 shows side–by–side boxplots of the wind speed by month. They allow to see this behavior
in a graphical way.

Figure 29: Boxplots of the daily average wind speed in Hopkins Memorial Forest during 2012
by month.

One last thing to highlight before we close this section. Note that when we analyze all
observations together, we identified six outliers. When we analyze the data by month some
values that were considered as outliers before are not considered as outliers any longer. Also,
some values that were not outliers before, are now considered to be outliers. This is due to
the fact that a value is considered as an outlier with respect to a population. In a different
population, an outlier may not be an outlier any longer.

38
5 Measures of relationships between two variables
All the parameters that we have introduced up to this point are measures for describing one
variable at a time. In this section we introduce three measures that allow for describing the
association between two variables.

5.0.1 The correlation coefficient

In fact, the measure that we introduce in this subsection is more formally known as Pearson’s
correlation coefficient. A measure due to Karl Pearson.

Definition 43. Let xi and yi be the values of two variables associated to the ith element in
U (i = 1, 2, · · · , N ). The correlation coefficient (or simply, the correlation) between x and y
is defined as P
U (xi − x̄U )(yi − ȳU )
rxy,U ≡ P .  (11)
( U (xi − x̄U )2 U (yi − ȳU )2 )1/2
P

By construction, the correlation coefficient ranges from -1 to +1. The correlation indicates
both the strength and the sign of the linear association between x and y:

• it takes the value of zero when there is no linear relationship between x and y;

• the absolute value of the correlation indicates the strength of the linear association
between x and y: values further away from zero indicate a stronger association. A
correlation of +1 or -1 indicates a perfect linear association;

• the sign of the correlation indicates the type of association. A positive correlation in-
dicates that increments in one variable are associated to increments in the other one.
A negative correlation indicates that increments in one variable are associated to decre-
ments in the other one.

It should be noted that the correlation of a variable with itself is always equal to 1, i.e.
rxx,U = 1.

Example 44. Let U be the population of N = 10 students taking a Master course in statistics.
Let xi and yi be, respectively, the scores in a home assignment and the final exam of the ith
student (i = 1, 2, · · · , N ). Table 21 shows the observed values.

i 1 2 3 4 5 6 7 8 9 10
xi 0 9 2 24 25 23 4 20 28 5
yi 8 15 5 36 40 30 9 21 32 27
Table 21: Score of ten students in a home assignment and an exam in statistics

One more exception will be made at this point and we will introduce a second type of
chart that allows for displaying the information of two numerical variables simultaneously,
namely, the scatter plot. A scatter plot is constructed by simply placing the N points (x1 , y1 ),
(x2 , y2 ), · · · , (xN , yN ) on a cartesian plane. Figure 34 shows a scatter plot of the scores in the
assignment and the exam for the population of ten students. We see that, in general terms,
students who score a higher number of points in the assignment also score a higher number
of points in the exam, so we see a positive association between both variables, therefore, we
expect a positive correlation. (The function plot() allows to create scatter plots in R.)

39
 Figure 30: Scatter plot of the points of 10 students in an assignment and an exam.

Let us find the correlation coefficient (11) between x and y. The mean of the two variables
is x̄U = 14 and ȳU = 22.3. Let us calculate the numerator first:
X
(xi − x̄U )(yi − ȳU ) = (0 − 14)(8 − 22.3) + (9 − 14)(15 − 22.3) + · · · + (5 − 14)(27 − 22.3) =
U
200.2 + 36.5 + · · · + −42.3 = 1064.

Now we calculate the first term in the denominator

X
(xi − x̄U )2 = (0 − 14)2 + (9 − 14)2 + · · · + (5 − 14)2 = 196 + 25 + · · · + 81 = 1080.
U

And the second term in the denominator is

X
(yi −ȳU )2 = (8−22.3)2 +(15−22.3)2 +· · ·+(27−22.3)2 = 204.49+53.29+· · ·+22.09 = 1412.1.
U

This gives
P
U (xi − x̄U )(yi − ȳU ) 1064
rxy,U = P 1/2
= = 0.8616,
(1080 · 1412.1)1/2
P
( U (xi − x̄U )2 U (yi − ȳU )2 )
which means that there is a relatively high linear association between the number of points
that students get in the assignment and the exam. 
It is worth repeating that the correlation measures the linear association between two
variables. And both variables are very important for interpreting the correlation correctly:
• Correlation indicates association, nothing more than that. In particular, it does not mea-
sure causality. It is a common error to conclude causality from correlation. For instance,
if we find that there is a high and positive correlation between the amount invested in
advertisement and the amount of sales in a set of companies, it may be tempting to
conclude that more investment in advertisement leads to higher sales. Although that
may still be true, we cannot make that conclusion based on the correlation coefficient.
• Correlation measures linear association. It may be that both variables are associated
in a different way and yet the correlation between them is small simply because the
association is not linear.
In few words, the correlation measures how well the association between two variables can be
described by a straight line. Let us consider the variable x and the four variables y in Table
22. Figure 31 shows the corresponding scatter plots.

40
• The association between x and y1 (top–left panel) is clearly linear and very strong, in fact,
we have a perfect linear association. This fact is adequately described by the correlation
coefficient, we get rxy1 = −1. Note that the sign indicates a negative association.

• There is a perfect association between x and y2 (top–right panel), in fact, we have that
y2 = exp(x). However, the association is not linear and the correlation coefficient equals
rxy2 = 0.6914 which, in loose words, can be interpreted as “the association between x
and y2 can be adequately described by a straight line”.

• There is a perfect association between x and y3 (bottom–left panel), in fact, we have

that y3 = (x − 6)2 . However, the association is not linear and the correlation coefficient
equals rxy3 = 0, which can be interpreted as “the association between x and y3 cannot
be described by a straight line at all”.

• The points in the bottom–right panel show no clear association of any type, in particular,
they show no linear association. This fact is reflected by the correlation coefficient which
is rxy4 = −0.2545, indicating that a straight line will be very poor at describing the
association between both variables.

x y1 y2 y3 y4
1 11 3 25 9
2 10 7 16 7
3 9 20 9 6
4 8 55 4 1
5 7 148 1 10
6 6 403 0 5
7 5 1097 1 3
8 4 2981 4 11
9 3 8103 9 8
10 2 22026 16 4
11 1 59874 25 2
Table 22: Different types of association

Figure 31: Scatter plots of the variables in Table 22.

In R, the function cor() can be used for computing the correlation coefficient.

41
Correlation matrix Let us say that we have J variables, x1 , x2 , · · · , xJ . It is possible to
calculate J(J − 1)/2 correlations between pairs of variables and present them in a matrix that
is called the correlation matrix:
 
1 rx1 x2 ,U · · · rx1 xJ ,U
 rx x ,U
 2 1 1 · · · rx2 xJ ,U 

 .. .. . . .. 
 . . . . 
rxJ x1 ,U rxJ x2 ,U · · · 1

The ones in the diagonal are obtained by taking into account that rxj xj ,U = 1. Alternatively,
taking into account that the correlation is symmetric, i.e. that rxj xj0 ,U = rxj0 xj ,U , we can simply
write the correlation matrix as
 
1 — ··· —
 rx x ,U
 2 1 1 · · · — 
 .. .. . . .. 
 . . . .
rxJ x1 ,U rxJ x2 ,U · · · 1

42