MEDICINE

REVIEW ARTICLE

Descriptive Statistics
The Specification of Statistical Measures and Their Presentation in Tables and Graphs

Part 7 of a Series on Evaluation of Scientific Publications

Albert Spriestersbach, Bernd Rhrig, Jean-Baptist du Prel, Aslihan Gerhold-Ay, Maria Blettner

SUMMARY
Background: Descriptive statistics are an essential part of
A set of medical data is based on a collection of the
data of individual cases or objects, also called
observation units or statistical units. Every case, for
biometric analysis and a prerequisite for the understanding
of further statistical evaluations, including the drawing of
example every study participant, patient, every experi-
inferences. When data are well presented, it is usually mental animal, every tooth or every cell shows compa-
obvious whether the author has collected and evaluated rable parameters (such as body weight, gender, erosion,
them correctly and in keeping with accepted practice in pH). Each of these parameters, also called variables,
the field. has a specific parameter value (gender = male, age =
30 years, weight = 70 kg) for each observation unit (for
Methods: Statistical variables in medicine may be of either
example the patient). The aim of descriptive statistics
the metric (continuous, quantitative) or categorical
is to summarize the data, so that they can be clearly
(nominal, ordinal) type. Easily understandable examples
illustrated (13).
are given. Basic techniques for the statistical description
The property of a parameter is specified by its so-
of collected data are presented and illustrated with
called scale of measure. Generally two types of para-
examples.
meters are distinguished. A variable has a metric level
Results: The goal of a scientific study must always be (= quantitative data) if it can be counted, measured or
clearly defined. The definition of the target value or clinical weighed in a physical unit (as in cm or kg) or at least
endpoint determines the level of measurement of the can be recorded in whole numbers. Data with a metric
variables in question. Nearly all variables, whatever their
scale of measure can be further classified into conti-
level of measurement, can be usefully presented
nuous and discrete variables. In contrast to discrete
graphically and numerically. The level of measurement
variables, continuous variables can take any value.
determines what types of diagrams and statistical values
Examples for metric continuous parameters are body
are appropriate. There are also different ways of presenting
height in cm, blood pressure in mmHg or the creatinine
combinations of two independent variables graphically and
concentration in mg/L. One example for a metric
numerically.
discrete parameter is the number of erythrocytes per
Conclusions: The description of collected data is microliter of blood.
indispensable. If the data are of good quality, valid and The gender of man cannot be measured, but is clas-
important conclusions can already be drawn when they are sified into two categories. Parameters which can be
properly described. Furthermore, data description provides classified into two or more categories are described as
a basis for inferential statistics. categorical parameters (= qualitative data). A further
Key words: statistics, data analysis, biostatistics, classification of a categorical parameter is into nominal
publication characteristics (unordered) and ordinal characteristics
(ordered according to rank). Good basic portrayals of
Cite this as: Dtsch Arztebl Int 2009; 106(36): 57883 the descriptive statistics of medical data can be found
DOI: 10.3238/arztebl.2009.0578 in text books (49). Figure 1 gives a review of types of
parameters, as well as graphs to be used and statistical
Institut fr Medizinische Biometrie, Epidemiologie und Informatik, Johannes measures.
Gutenberg-Universitt Mainz: Prof. Dr. rer. nat. Blettner, Spriestersbach, Different procedures are necessary for the statistical
Gerhold-Ay evaluation of metric and categorical parameters in graphic
Zentrum Prventive Pdiatrie, Zentrum fr Kinder- und Jugendmedizin, Univer- and tabular forms. The graphs used here and the evalua-
sittsmedizin der Johannes Gutenberg-Universitt Mainz: Dr. med. du Prel,
M.P.H. tion tables were created with the statistics package SPSS
MDK Rheinland-Pfalz, Referat Rehabilitation/Biometrie, Alzey: Dr. rer. nat. for WINDOWS (Version 15). As example, we are using
Rhrig data of 176 sportsmen and women.

578 Dtsch Arztebl Int 2009; 106(36): 57883

Deutsches rzteblatt International
 MEDICINE

FIGURE 1 FIGURE 2

Example for a box plot

distribution of body weight in kg in 176 sportsmen and

women.
Sketch of parameter types and suitable statistical measures for descriptive presentation
1.1.2. The histogram
A histogram shows the distribution form of the measured
values of a continuous variable. In a normal distribution,
this takes the form of a "Gauss bell curve" (Figure 3a).
The present (measured) values are classified into an
Results appropriate number of classes (3). If the number of classes
1. Description of a continuous parameter is not "naturally" given, it is recommended to choose the
A continuous parameter is a quantitative measure. The number of classes as the square root of the case number
value is measured on a continuous scale in arbitrarily N. If you had, for example, 49 cases, seven classes
small intermediate steps (3). In this article, quantitative would be chosen for the histogram on which the measured
parameters are treated as continuous parameters. The values are distributed. Within each class, the measured
size of the continuous parameters is presented by a mea- values are counted and illustrated as column in the figure.
surement unit (= size unit, physical unit), for example Figure 3 shows five schematic examples for distribution
body height in cm or body weight in kg. forms in each histogram.
Initially, graphic presentations of the distribution of a In a histogram it is recognizable whether the data are
continuous variable are created in the form of box plots symmetrically distributed around the mean value (Figure
and histograms. Such diagrams make it possible for the 3a). However, when the histogram has a left peak (Figure
researcher to get an initial visual impression of the dis- 3b) or a right peak (Figure 3c), the values have a "skew"
tribution of the collected parameters. distribution. In some situations, it may happen that
several peaks are recognizable in a histogram (Figure 3d
1.1. Graphic presentation of a continuous parameter and e).

1.1.1. The box plot diagram 1.2. Numerical description of a continuous parameter
Box plots offer a visual impression of the position of the The distribution of a recorded continuous parameter can
first and third quartiles (25th and 75th percentile) and of be numerically described with the following statistical
the median (central value). Also minimum, maximum, measures: Minimum, maximum, quartiles (with median),
and the breadth of scatter of all case values of a contin- range (difference between maximum and minimum),
uous parameter are recognizable. 50% of the values of skewness (indicates whether the distribution is symmetric
distribution are within the box (= interquartile range). A or not), arithmetic mean value, and standard deviation
box with a greater interquartile range indicates greater (= square root of the variance) (6). When the parameter
scatter of the values. Figure 2 shows an example of the "skewness" is between 1 and +1, the distribution is

Dtsch Arztebl Int 2009; 106(36): 57883

Deutsches rzteblatt International 579
MEDICINE

FIGURE 3 2.1.2. The bar diagram as graphic presentation of a categorical

parameter
The bar diagram offers an alternative form of presenta-
tion. In contrast to the pie diagram, the frequency values
are read on the y-axis. This figure can present absolute
or relative frequencies. It is possible to compare the
heights of the bars directly, which is not possible with
the segments in the pie diagram. In contrast to the histo-
gram for continuous parameters, there are no class inter-
vals for the value ranges of the measurements on the
Examples of the types of distribution in histograms x-axis of the bar diagram. On the contrary, each bar
a) Normal distribution (symmetric) creates one unit completed to the right and to the left,
b) left peak (= skewed to the right)
according to its value. Thus, one single bar refers either
c) right peak (= skewed to the left)
to the female or to the male sex in Figure 4. In contrast
d) two peaks (symmetric)
e) several peaks to the histogram, the single bars or columns of a bar dia-
gram should contain gaps.

2.2. Numerical description of a categorical parameter

symmetric, but when it is under 1 or above +1 the values Both absolute and relative frequencies of a categorical
are distributed with a right or a left peak (Box). parameter can be given in a frequency table. In Table 1,
we show our population of sportswomen and sportsmen
2. Description of a categorical parameter once again: In this SPSS Version of a frequency table,
you can find the absolute frequencies in the "frequency"
2.1. Graphic presentation of a categorical parameter column. If given, the number of missing values is listed
here. In the columns "percentage" and "valid percent-
2.1.1. The pie diagram as graphic presentation of a categorical ages," you can choose whether the missing values
parameter should be listed or not as a separate category. In the
The pie diagram or circular diagram is a popular form of column "cumulated percentages," the successive cumu-
presentation for the distribution of characteristics of a lated relative frequencies are presented. They are only
parameter classified into groups. The number of seg- important in parameters of ordinal scale level with more
ments in one pie diagram corresponds to the number of than two values. The two last columns ("valid and
possible characteristics (= steps) of these variables. So, cumulated percentages") are generally not suitable for
a pie diagram would have two segments for "sex." Their presentation in a publication.
proportion in the total pie corresponds to their relative
percentage. 3. Description of correlations
Until now, we have only considered single variables, i.e.,
we have described our data as "univariate." No parameter
was related to any other. It is also possible to describe po-
BOX tential correlations between two variables, for example
between body weight (continuous) and body height (con-
tinuous). It is also possible to relate two categorical para-
meters or a metric and a categorical parameter.
3.1. Description of the correlation of two continuous
parameters
One of the variables is allocated to each of the two axes
in the scatter plot (point cloud diagram). If there is a cor-
relation, it is expressed in the trend of the point cloud to
form an ellipse. If there were a 100 percent linear corre-
lation between two parameters, all points would lie on a
straight line. In our example (10), the measured values
of bone density on the proximal site of measurement
(variable: 'spa_prox') and the values of bone density on
the distal site of measurement (variable: 'spa_dist') are
correlated with each other (Figure 5). If there were no
correlation, the area of the diagram would be unstructured
and covered with points.
The degree of correlation can be numerically expressed
as linear correlation with a coefficient of correlation.
This index can have values between 1 and +1. If both
distributions are symmetric, the coefficient of correlation

580 Dtsch Arztebl Int 2009; 106(36): 57883

Deutsches rzteblatt International
 MEDICINE

can be calculated according to Pearson. Otherwise, the FIGURE 4 Example of a bar

coefficient of correlation according to Spearman is diagram of two
appropriate. This is not directly calculated from the groups
values but from their ranks. In linear distribution, the
coefficient of correlation should be calculated according
to Pearson. The Pearson coefficient of correlation is cal-
culated as 0.886 for the measurement of bone density.
This is a very strong correlation.

3.2. Description of the correlation of two categorical

parameters
A grouped bar diagram is suitable for the graphical pre-
sentation of the correlation between two categorical
parameters. The standard form uses the x-axis for entry
of the single categories of the first variable and portrays
within these categories the absolute (number) or relative
frequencies (percentage) of the second variable through
different colored bars. In Figure 6, a possible correlation
between smokers and (smokers') cough is to be studied.
The bars represent relative frequencies in percentages.
They are more suitable for comparison than the absolute
frequencies, as the groups are often not of the same size.
In this figure, a bar diagram is presented and it is recog-
nizable that smokers (with 71%) more often have cough
than non-smokers (29%). TABLE 1
The numerical description of information about ab-
solute and relative frequencies of the two correlated Sex
categorical variables is summarized in a cross table. We Frequency Percentage Valid Cumulated
differentiate between line variables (smokers and non- percentages percentages
smokers) and column variables (cough). It is possible to Valid Male 102 58.0 58.0 58.0
give the absolute and relative frequencies, i.e. line and
Female 74 42.0 42.0 100.0
column percentages, within the cells (Table 2).
You can see from this presentation that 21/29 = 72% Total 176 100.0 100.0
of non-smokers do not cough but 8/29 = 28% cough.
The cumulative percentage for all lines is 100% and the
cumulative percentage for all columns is also 100%.
The four field table helps the reader to understand the
reference values and the results.
FIGURE 5
To simplify the interpretation, it should be made clear
which variable is the target parameter. In our example, it
was important to know whether smoking influences the
occurrence of cough. Thus, cough is the (dependent) tar-
get parameter and smoking the (independent) possible
factor. In the interest of clarity of the cross table, it is
recommended to write the factor in the cells and the
target parameter in the columns. Then you can dispense
with the percentages in the columns. In this way, the
cross table is clearer and easier to understand. In prac-
tice, medical expertise is required to be able to formulate
a sensible question. This results in the arrangement of
the variables. Further parameters which can be calculated
from the four field table (relative risk, odds ratio) will be
discussed in a later publication.

3.3. Correlation between a continuous and a categorical

parameter
Several box plots can be presented in one diagram with
the help of a statistics program like SPSS, STATISTICA,
or SAS. Thus, it is possible to compare groups for which
for example the continuous parameter "weight in kg" Example of a scatter diagram

Dtsch Arztebl Int 2009; 106(36): 57883

Deutsches rzteblatt International 581
MEDICINE

Example of a FIGURE 6 was recorded. In Figure 7, the distribution of the weight

grouped bar of 74 sportswomen was compared with the distribution
diagram of the weight of 102 sportsmen. When comparing the
two groups, a trend is recognizable: Men tend to have
higher weights than women (consideration of medians).
The values are more scattered with men than with
women, as the box plot for men is more stretched than
for women.
As described in chapter 1.2., suitable measures are
numerically calculated for the comparison of the groups
of the two sexes.

Discussion
The exact description of data collected in a study is sen-
sible and important. The correct descriptive presentation
of the results is the first step in evaluating and graphi-
cally presenting the results (79, 11). The description is
the basis of the biometric evaluation and is the indis-
pensable starting point for further methodological pro-
cedures such as statistical significance tests. The
descriptive presentation of study results usually occupies
most of the space in publications. The description
covers the graphical and tabular presentation of the
TABLE 2 results. The exact assessment of the scale level of the
parameter is important as the scale level determines the
Distribution of the parameters smoking and cough type and procedure, both in the descriptive, and in the
explorative (= generating of hypothesis) and confirma-
Cough
No Yes Total tory (= biometric testing of hypothesis) evaluation. The
Smokers No Number 21 8 29
selection of a suitable statistical test procedure for con-
% of smokers 72% 28% 100% trolling the significance is determined by the scale level
% with cough 75% 29% 52% of the investigated parameters.
Yes Number 7 20 27 In normally distributed data, the arithmetic mean
% of smokers 26% 74% 100% value is the same as the median. The skewness has the
% with cough 25% 71% 48% value zero. Unfortunately there rarely is a normal distri-
Total Number 28 28 56 bution in natural systems as in parameters collected in
% of smokers 50% 50% 100% patients. It is sensible to give the arithmetic mean value,
% with cough 100% 100% 100%
as well as the median for continuous data. A normal dis-
tribution cannot be assumed when the two values are
very different. The arithmetic mean value cannot be cal-
culated in data with a purely ordinal scale. It is often
asked whether graphical or numerical presentations are
Example of grouped FIGURE 7 preferable in data description. Graphics serve to give a
box plots first impression and to visually illustrate the situation of
distribution parameters. It can be difficult to read the
exact values of the median or the percentiles on the y-axis
in a box plot diagram. For this reason, calculation and
presentation of the exact statistical characteristic values
is indispensable.
In the individual case, the information of further bio-
metric measures is of course valuable, including measures
not mentioned in the article. Examples would be effect
sizes, confidence intervals, Cohen's kappa, relative risk,
and cumulated values.
The use of suitable validated statistics software like
SPSS or SAS is recommended for statistical evaluation
of data.

Conflict of interest statement

The authors declare that there is not conflict of interest according to the
guidelines of the International Committee of Medical Journal Editors.

582 Dtsch Arztebl Int 2009; 106(36): 57883

Deutsches rzteblatt International
 MEDICINE

Manuscript received on 4 February 2009, revised version accepted on 7. Trampisch HJ, Windeler J: Medizinische Statistik. Berlin, Heidelberg,
16 March 2009. New York: Springer 2000; 2. Auflage: 5282.
Translated from the original German by Rodney A. Yeates, M.A., Ph.D. 8. Hilgers RD, Bauer P, Schreiber V: Einfhrung in die Medizinische Sta-
tistik. Berlin, Heidelberg, New York: Springer 2003; 343.
REFERENCES 9. Altman DG: Practical statistics for medical research. Boca Raton,
London, New York, Washington D.C.: Chapman & Hall/CRC 1999;
1. Greenfield MLVH, Kuhn JE, Wojtys EM: A statistics primer: descriptive 1045.
measures for continuous data. Am J Sports Med 1997; 25: 7203.
10. Zawalski R: Messung der Hautfaltendicke am Handrcken mit Hilfe
2. McHugh ML: Descriptive statistics, part I: level of measurement. einer Mikrometerschraube [Dissertation]. Mainz: Fachbereich Medi-
JSPN 2003; 8: 357. zin der Johannes Gutenberg-Universitt; 1997.
3. Overholser BR, Sowinski KM: Biostatistics primer: part I. Nutr Clin 11. Du Prel JB, Rhrig BBM: Kritisches Lesen wissenschaftlicher Artikel.
Pract 2007; 22: 62935. Dtsch Arztebl Int 2009; 106: 1005.
4. SPSS Incorporated: SPSS 16.0 Schneller Einstieg. Dublin: SPSS Inc.
Coresponding author
2007; 5562.
Prof. Dr. rer. nat. Maria Blettner
5. Bortz J: Statistik fr Sozialwissenschaftler. Berlin Heidelberg New Institut fr Medizinische Biometrie,
York: Springer 1999; 5. Auflage: 1747. Epidemiologie und Informatik
Johannes Gutenberg-Universitt
6. Sachs L: Angewandte Statistik: Anwendung statistischer Methoden. 55101 Mainz, Germany
Berlin, Heidelberg, New York: Springer 2004; 11. Auflage: 1177. [email protected]

Dtsch Arztebl Int 2009; 106(36): 57883

Deutsches rzteblatt International 583