7 A.H.

Pripp, Statistics in Food Science and Nutrition, SpringerBriefs in Food,

Health, and Nutrition, DOI 10.1007/978-1-4614-5010-8_2,
Springer Science+Business Media New York 2013
Abstract The best way to learn statistics is by taking courses and working with
data. Some books may also be helpful. A rst step in applied statistics is usually to
describe and summarize data using estimates and descriptive plots. The principle
behind p -values and statistical inference in general is covered with a schematic
overview of statistical tests and models.
Keywords Recommended textbooks Descriptive statistics p -values Statistical
models
2.1 Recommended Textbooks on Statistics
How does one learn statistics, epidemiology, and experimental design? The recom-
mended approach is, of course, to take (university) courses and combine it with applied
use. In the same way it takes considerable effort and time to become trained in food
technology or chemistry or as a physician, learning statistics both the mathematical
theory and applied use takes time and effort. Some courses or books that promise to
teach statistics without requiring much time and that neglect all the fundamental
aspects of the subject could be deceiving. Learning the technical use of statistical
software without some fundamental knowledge of what these methods express and
the basics of calculations may leave the statistical analysis part in a black box.
Appropriate statistical analysis and a robust experimental design should be the oppo-
site of a black box it should shed light upon data and give clear insights. It should
ideally not be Harry Potter magic!
A comprehensive introduction to statistics and experimental design goes some-
what beyond the scope of this brief text. Therefore, this section will refer the reader to
several excellent textbooks on the subject available from Springer. Readers with
access to a university library service should be able to obtain these texts online through
Chapter 2
Methods and Principles of Statistical Analysis
8 2 Methods and Principles of Statistical Analysis
www.springerlink.com . Readers unfamiliar with the general aspects of statistics and
experimental design or who have not taken introductory courses are encouraged to
study some of these textbooks. An overview of the principles of descriptive statistics,
statistical inference (e.g., estimations and p -values), classic tests, and statistical mod-
els is given later, but it is assumed that the reader has a basic knowledge of these
principles.
2.1.1 Applied Statistics, Epidemiology,
and Experimental Design
Statistics for Non-Statisticians by Madsen ( 2011 ) is an excellent introductory text-
book for those new to the eld. It covers the collection and presentation of data,
basic statistical concepts, descriptive statistics, probability distributions (with an
emphasis on the normal distribution), and statistical tests. The free spreadsheet soft-
ware OpenOf ce is used throughout the text. Additional material on statistical soft-
ware, more comprehensive explanations on probability theory, and statistical
methods and examples are provided in appendices and at the textbooks Web site.
At 160 page, the textbook is not overwhelming. Readers with different interests,
either in applied statistics or in mathematical-statistical concepts, are told which
parts to read. Readers unfamiliar with statistics are highly encouraged to read this
text or a similar introductory textbook on statistics.
Applied Statistics Using SPSS, STATISTICA, MATLAB and R by Marques de S
( 2007 ) is another recommended textbook, although it goes into somewhat more depth
on mathematical-statistical principles. However, it provides a very useful introduction
to using these four key statistical softwares for applied statistics. Combined with soft-
ware manuals, it will give the reader an improved understanding of how to conduct
descriptive statistics and tests. Both SPSS and STATISTICA have menu-based sys-
tems in addition to allowing users to write command lines (syntaxes). MATLAB and
R might have a steeper learning curve and assume a more in-depth understanding of
mathematical-statistical concepts, but they have many advanced functions and are
used widely in statistical research. R is available for free and can be downloaded on
the Internet. This is sometimes a great advantage and makes the user independent of
updated licenses. Those who wish to make the effort to learn and use R will be part of
a large statistical community (R Development Core Team 2012 ) . It may, however,
take some effort if one is unfamiliar with computer programming.
Biostatistics with R: An Introduction to Statistics Through Biological Data
by Shahbaba ( 2012 ) gives a very useful step-by-step introduction to the R software
platform using biological data. The many statistical methods available through
so-called R packages and the (free) availability of the software makes it very
attractive, but its somewhat more complicated structure compared to commercial
software like SPSS, STATISTICA, or STATA might make it less relevant for those
who use mostly so-called standard methods and have access to commercial
software.
9 2.1 Recommended Textbooks on Statistics
Various regression methods play a very important part in the analysis of biological
data including food science and technology and nutrition research. Regression
Methods in Biostatistics by Vittinghoff et al. ( 2012 ) gives an introduction to
explorative and descriptive statistics and basic statistical methods. Linear, logis-
tic, survival, and repeated measures models are covered without a too-overwhelming
focus on mathematics and with applications to biological data. The software
STATA that is widely used in biostatistics and epidemiology is used throughout
the book.
Those who work much with nutrition, clinical trials, and epidemiology with
respect to food will nd very useful topics in textbooks such as Statistics Applied to
Clinical Trials by Cleophas and Zwinderman ( 2012 ) and A Pocket Guide to
Epidemiology by Kleinbaum et al. ( 2007 ) . These books cover concepts that are sta-
tistical in nature but more related to clinical research and epidemiology. Clinical
research is in many ways a scienti c knowledge triangle comprised of medicine,
biostatistics, and epidemiology.
Lehmann ( 2011 ) in his book Fisher, Neyman, and the Creation of Classical
Statistics gives historical background of the scientists that laid the foundation for
statistical analysis Ronald A. Fisher, Karl Pearson, William Sealy Gosset, and
Egon S. Pearson. Those with some insight into classical statistical methods and with
a historic interest in the subject should derive much pleasure from reading about the
discoveries that we sometimes take for granted in quantitative research. The text is
not targeted at food applications but, without going into all the mathematics, pro-
vides a historical introduction to the development of statistics and experimental
design. Some of the methods presented from their historical perspective might be
dif cult to follow if one is unfamiliar with statistics. However, a more comprehen-
sive description of the lady tasting tea experiment is provided together with the
many important concepts later discussed in relation to food science and technology.
2.1.2 Advanced Text on the Theoretical
Foundation in Statistics
Numerous textbooks have a more theoretical approach to statistics, and many are col-
lected in the series Springer Texts in Statistics . Modern Mathematical Statistics with
Applications by Devore and Berk ( 2012 ) provides comprehensive coverage of the
theoretical foundations of statistics. Another recommended text that gives an over-
view of the mathematics in a shorter format is the SpringerBrief A Concise Guide to
Statistics by Kaltenbach ( 2011 ) . These two and other textbooks with an emphasis on
mathematical statistics are useful for exploring the fundamentals of statistical science
with a more mathematical than applied approach to data analysis. However, most
readers with a life science or biology-oriented background may nd the formulas,
notations, and equations challenging. Applied knowledge and mathematical knowl-
edge often go hand in hand. It is usually more inspiring to learn the basic foundation
if there is an applied motivation for a speci c method. Many readers might therefore
10 2 Methods and Principles of Statistical Analysis
wish to consult textbooks with a more mathematical approach on a-need-to-know
basis and begin with the previously recommended texts on applied use.
2.2 Describing Data
Food scientists encounter many types of data. Consumers report their preferences,
sensory panels give scores on avor and taste characteristics, laboratories provide
chemical and microbial data, and management sets speci c targets on production
costs and expected sales. Analysis of all these data begins with a basic understand-
ing of their statistical nature.
The rst step in choosing an appropriate statistical method is to recognize the
type of data. From a basic statistical point of view there are two main types of data
categorical and numerical. We will discuss them thoroughly before continuing
with more speci c types of data like ranks, percentages, and ratios. Many data sets
contain missing data and extreme observations often called outliers. They also pro-
vide information and require attention.
To illustrate the different types of data and how to describe them, we will use
yogurt as an example. Yogurt is a dairy product made from pasteurized and homoge-
nized milk, forti ed to increase dry matter, and fermented with the lactic acid bacteria
Streptococcus thermophilus and Lactobacillus delbrueckii subspecies bulgaricus . The
lactic acid bacteria ferment lactose into lactic acid, which lowers pH and makes the
milk protein form a structural network, giving the typical texture of fresh fermented
milk products. It is fermented at about 45C for 57 h. A large proportion of yogurts
also add fruit, jam, and avor. There are two major yogurt processing technologies
stirring and setting. Stirred yogurt is fermented to a low pH and thicker texture in a vat
and then pumped into packages, while set yogurt is pumped into packages right after
lactic acid bacteria have been added to the milk; the development of a low pH and the
formation of a gel-like texture take part in the package. The fat content can change
from 0% fat to as high as 10% for some traditional types. It is a common nutritious
food item throughout the world with a balanced content of milk proteins, dairy fats,
and vitamins. In some yogurt products, especially the nonfat types, food thickeners
are added to improve texture and mouthfeel (for a comprehensive coverage of yogurt
technology see, e.g., Tamine and Robinson 2007 ) .
2.2.1 Categorical Data
In Table 2.1 categorical and numerical data from ten yogurt samples are presented
to illustrate types of data. The rst three variables ( avor, added thickener, and fat
content) are all derived from categorical data. Observations that can be grouped into
categories are thus called categorical data. Statistically they contain less informa-
tion than numerical data (to be covered later) but are often easier to interpret and
11 2.2 Describing Data
understand. Low-fat yogurt conveys more clearly the fat content to most consumers
than the exact fat content. Consumers like to know the fat content relative to other
varieties and not the exact amount. Categorical data are statistically divided into
three groups nominal, binary, and ordinal data. Knowing the type of data one is
dealing with is essential because that dictates the type of statistical analysis and tests
one will perform.
Data that fall under nominal variables (e.g., avor in Table 2.1 ) are comprised
of categories, but there is no clear order or rank. Perhaps one person prefers straw-
berry over vanilla, but from a statistical point of view there is no obvious order to
yogurt avors. Other typical examples of numerical data in food science are food
group (e.g., dairy, meat, vegetables), method of conservation (e.g., canned, dried,
vacuum packed), and retail (e.g., supermarket, restaurant, fast-food chain).
Statistically, nominal variables contain less information than ordinal or numerical
variables. Thus, statistical methods developed for nominal variables can be used
on other types of data, but with lower ef ciency than other more appropriate or
ef cient methods.
If measurements can only be grouped into two mutually exclusive groups, then
the data are called binary (also called dichotomous). In Table 2.1 the variable added
thickener contains binary data. As long as the data can be grouped into only two
categories, they should be treated statistically as binary data. Binary data can always
be reported in the form of yes or no . Sometimes for binary data, yes and no are coded
as 1 and 0, respectively. It is not necessary, but it is convenient in certain statistical
analysis, especially when using statistical software. Binary variables are statistically
often associated with giving the risk of something. One example is the risk of food-
borne disease bacteria (also called pathogenic bacteria) in a yogurt sample. Pathogenic
bacteria are either detected or not. However, from a statistical point of view the risk
is estimated on a scale of 0 to 1, but for individual observations the risk is either pres-
ent (pathogenic bacteria detected) or not (pathogenic bacteria not detected). Thus, it
could then be presented as binary data for individual observations.
Table 2.1 Types of data and variables given by some yogurts samples
Type of data Categorical Numerical
Type of variable Nominal Binary Ordinal Discrete Continuous
Sample Flavor
Added
thickener Fat content
Preference
(1: low, 5: high) pH
1 Plain Yes Fat free 1 4.41
2 Strawberry Yes Low fat 4 4.21
3 Blackberry No Medium 3 4.35
4 Vanilla No Full fat 4
5 Vanilla Yes Full fat 4 4.15
6 No Low fat 3 4.38
7 Strawberry Yes Fat free 2 4.22
8 Vanilla Yes Fat free 2 4.31
9 Plain No Medium 2 4.22
10 Strawberry No Full fat 6.41
12 2 Methods and Principles of Statistical Analysis
Data presented by their relative order of magnitude, such as the variable fat
content in Table 2.1 , are ordinal. Fat content expressed as fat free, low fat, medium
fat, or full fat has a natural order. Since it has a natural order of magnitude with more
than two categories, it contains more statistical information than nominal and binary
data. Ordinal data can be simpli ed into binary data e.g., reduced fat (combining
the categories fat free, low fat, and medium fat) or nonreduced (full fat), but with a
concomitant loss of information. Statistical methods used on nominal data can also
be used on ordinal data, but again with a loss of statistical information and ef ciency.
If ordinal data can take only two categories, e.g., thick or thin, they should be
considered binary.
2.2.2 Numerical Data
Observations that are measurable on a scale are numerical data. In Table 2.1 , two
types of numerical data are illustrated. These are discrete or continuous. In applied
statistics both discrete and ordinal data are sometimes analyzed using methods
developed for continuous data, even though it is not always appropriate according
to statistical theory. Numerical data contain more statistical information than cate-
gorical data. Statistical methods suitable for categorical data analysis can therefore
be applied to numerical data, but again with a loss of information. Therefore, it is
common to apply other methods that take advantage of their additional statistical
information compared with categorical data.
Participants in a sensory test may score samples on their preference using only
integers like 1, 2, 3, 4, or 5. Observations that can take only integers (no decimals)
are denoted discrete data. The distance between discrete variables is assumed to be
the same. For instance the difference in preference between a score of 2 and 3 is
assumed to be the same as the difference between scores 4 and 5. It is therefore pos-
sible to estimate, for example, the average and sum of discrete variables. If the dis-
tance cannot be assumed equal, discrete data should instead be treated as ordinal.
The pH of yogurt samples is an example of continuous data. Continuous data are
measured on a scale and can be expressed with decimals. They contain more statisti-
cal information than the other types of data in Table 2.1 . Thus, statistical methods
applied to categorical or discrete data can be used on variables with continuous data,
but not vice versa. For example, the continuous data on pH can be divided into those
falling below and those falling above pH 4.6 and thereby be regarded as binary data
and analyzed using methods for such data. However, if we have only information in
our database about whether the yogurt sample is below or above pH 4.6, it is not
possible to make such binary data continuous data. Thus, it is always useful to save
the original continuous data even though they may be divided into categories for
certain analysis. One may perhaps need the original datas additional statistical
information at a later stage. Many advanced statistical methods like regression were
rst developed for continuous data as an outcome and then later expanded for use
with categorical data.
13 2.2 Describing Data
2.2.3 Other Types of Data
Understanding the properties of categorical and numerical data serves as the
foundation of quantitative and statistical analysis. However, in applied work with
statistics one often encounters other speci c types of data that require our attention.
Some examples are missing data, ranks, ratios, and outliers. They have certain
properties that one should be aware of.
Missing data are unobserved observations. Technical problems during laboratory
analysis or participants not answering all questions in a survey are typical reasons
for missing data. In Table 2.1 yogurt samples 4, 6, and 10 have missing data for
some of the variables. A main issue with missing data is whether there is an underly-
ing reason why data are missing for some observations.
Statistical research on the effect of missing data is driven by medical statistics.
It is a very important issue in both epidemiology and clinical studies and especially
with longitudinal data (Song 2007 ; Ibrahim and Molenberghs 2009 ) . What if a
large proportion of those patients that do not experience any health improvement
of a new drug drop out of a clinical trial? Statistical analysis could then be in uenced
greatly by the proportion of missing data and the biased medical conclusions that
were reached. Missing data should therefore never be simply neglected or just
replaced by a given value (e.g., the mean of nonmissing data) without further inves-
tigation. The issue of missing data is likewise important in food science and nutri-
tion. We will use the terminology developed in medical statistics to understand
how missing data could be approached.
Let us assume that we are conducting a survey on a new yogurt brand. We
want to examine how fat content in uences sensory preferences. A randomly
selected group of 500 consumers is asked to complete a questionnaire about
food consumption habits including their consumption of different yogurts.
However, only 300 questionnaires are returned. Thus, we have 200 missing
observations in our data set. According to statistical theory on missing data,
these 200 missing observations can be classi ed as missing completely at random
(MCAR), missing at random (MAR), or missing not at random (MNAR). This
terminology is, unfortunately, not self-explanatory and somewhat confusing.
However, one may say generally that it concerns the probability that an observa-
tion is missing.
Missing completely at random (MCAR) : It is assumed here that the probability of
missing data is unrelated to the possible value of a given missing observation (given
that the observation was not missing and was actually made) or any other observa-
tions in ones data set. For instance, if the 200 missing observations were randomly
lost, then it is unlikely that the probability to be missing is related to the preference
scores of yogurt or any selected demographic data. Perhaps the box with the last 200
questionnaires was accidentally thrown away! For MCAR any piece of data is just
as likely to be missing as any other piece of data. The nice feature is that the statisti-
cal estimates and resulting conclusions are not biased by the missing data. Fewer
observations give increased uncertainty (i.e., reduced statistical power or conse-
14 2 Methods and Principles of Statistical Analysis
quently broader con dence intervals), but what we nd is unbiased. They may just
remain missing in your data set in further statistical analyses. All statistical analyses
with MCAR give unbiased information on what in uences yogurt preferences.
Missing at random (MAR) : It is also assumed that the probability of missing data is
unrelated to the possible value of a given missing observation (given that the observa-
tion was not missing and was actually made) but related to some other observed data in
the data set. For example, if younger participants are less likely to complete the ques-
tionnaire than older ones, the overall analysis will be biased with more answers from
older participants. However, separate analysis of young and old participants will be
unbiased. A simple analysis to detect possible MAR in the data set entails examining
the proportion of missing data between key baseline characteristics. Such characteris-
tics in a survey could be the age, gender, and occupation of the participants.
Missing not at random (MNAR) : Missing data known as MNAR present a more
serious problem! It is assumed here that the probability of a missing observation is
related to the possible value of a given missing observation (given that the observation
was not missing and was actually made) or other unobserved or missing data. Thus, it
is very dif cult to say how missing data could in uence ones statistical analysis. If
participants who prefer low-fat yogurt are less likely to complete the questionnaire,
then the results will be biased, but the information that it is due to their low preference
for low-fat yogurt is lacking! The overall results will be biased and incorrect conclu-
sions could be reached.
Whenever there are missing data, one needs to determine if there is a pattern in
the missingness and try to explain why the data are missing. In Table 2.1 data on
preference are missing for yogurt sample 10. However, the pH is exceptionally high.
Perhaps something went wrong during the manufacture of the yogurt and the lactic
acid bacteria did not ferment the lactose into lactic acid and so did not lower the pH.
That could explain why preference was not examined for this sample. Therefore,
always try to gather information to explain why data are missing. The best strategy
is always to design a study in a way that minimizes the risk for missing data.
Especially in consumer surveys and sensory analysis, it is common to rank food
samples. Ranking represents a relationship between a set of items such that, for any
two items, the rst is either ranked higher than, lower than, or equal to the second.
For example, a consumer might be asked to rank ve yogurt samples based on pref-
erence. This is an alternative to just giving a preference score for each sample. If
there is no de ned universal scale for the measurements, it is also feasible to use
ranking for comparison of samples. Statistically speaking, data based on ranking
have a lot in common with ordinal data, but they may be better analyzed using meth-
ods that take into account the ranks given by each consumer. It is therefore important
to recognize rankings from other types of data.
A ratio is a relationship between two numbers of the same kind. We might estimate
the ratio of calorie intake from dairy products compared with that from vegetables.
Percentage is closely related as it is expressed as a fraction of 100. In Latin per cent
means per hundred . Both ratios and percentages are sometimes treated as continuous
data in statistical analysis, but this should be done with great caution. The statistical
properties might be different around the extremes of 0 or 100%. Therefore, it is important
15 2.3 Summarizing Data
to examine ratio and percentage data to assess how they should be treated statistically.
Sometimes ratios and percentages are divided into ordinal categories if they cannot be
properly analyzed with methods for continuous data.
Take a closer look at the data for sample 10 in Table 2.1 . All the other samples
have pH measurements around 4.5, but the pH of sample 10 is 6.41. It is numeri-
cally very distant from the other pH data. Thus, it might be statistically de ned as
an outlier, but it is not without scienti c information. Since it deviates consider-
ably from the other samples, the sample is likely not comparable with the other
ones. This could be a sample without proper growth of the lactic acid bacteria that
produce the acid to lower the pH during fermentation. Outliers need to be exam-
ined closely (just like missing data) and be treated with caution. With the unnatu-
ral high pH value of sample 10 compared with the other samples, the average pH
of all ten samples would not be a good description of the typical pH value among
the samples. Therefore, it might be excluded or assessed separately in further
statistical analysis.
2.3 Summarizing Data
2.3.1 Contingency Tables (Cross Tabs) for Categorical Data
A contingency table is very useful for describing and comparing categorical
variables. Table 2.2 is a contingency table with exempli ed data to illustrate a
comparison of preferences for low- or full-fat yogurt between men and women.
The number of men and women in these data is different, so it is very useful to
provide the percentage distribution in addition to the actual numbers. It makes
the results much easier to read and interpret. Statistically, it does not matter
which categorical variable is presented in rows or columns. However, it is rather
common to have the variable de ning the outcome of interest (preferred type of
yogurt in our example) in columns and the explanation (gender of survey par-
ticipants) in rows (Agresti 2002 ) . In these illustrative data, women seem on
average to prefer low-fat yogurt, and men seem to prefer full-fat yogurt. Perhaps
this is a coincidence just for these 100 people, or is it a sign of a general differ-
ence in preference for yogurt types among men and women? Formal statistical
tests and models are needed to evaluate this.
2.3.2 The Most Representative Value of Continuous Data
Let us examine again the pH measurements of our ten yogurt samples. Remember,
we have missing data for sample 4; therefore, we have only nine data observations.
Reordering the pH data in ascending yields 4.15, 4.21, 4.22, 4.22, 4.31, 4.35, 4.38,
16 2 Methods and Principles of Statistical Analysis
4.41, and 6.41. What single number represents the most typical value in this data
set? For continuous data, the most typical value, or what is referred to in statistics
as the central location, is usually given as either the mean or median. The mean is
the sum of values divided by the number of values (the mean is also known as the
standard average). It is de ned for a given variable X with n observations as

=
=

1
1
n
i
i
x x
n

and is estimated in our example as

+ + + + + + + +
= =
4.15 4.21 4.22 4.22 . 4.35 4.38 4.41 6.41
mean 4.52.
9
4 31


The single outlier measurement of pH 6.41 has a relatively large in uence on the
estimated mean. An alternative to the mean could be to use the median. The median is
the numeric values separating the upper half of the sample or, in other words, the value
in the middle of our data set. The median is found by ranking all the observations from
lowest to highest value and then picking the middle one. If there is an even number of
observations and thus no single middle value, then the median is de ned as the mean
of the two middle values. In our example the middle value is 4.31 (indicated by bold
typeface in the equation estimating the mean). A rather informal approach to deciding
whether to use the mean or median for continuous data is to estimate them both. If the
median is close to the mean, then one can usually use the mean, but if they are sub-
stantially different, then the median is usually the better choice.
2.3.3 Spread and Variation of Continuous Data
Describing the central location or the most typical value is telling only half the
story. One also needs to describe the spread or variation in the data. For continuous
data it is common to use the standard deviation or simply the maximum and mini-
mum values. These might not be so intuitive as the mean and median. If the data set
has no extreme outliers or a so-called skewed distribution (many very high or low
Table 2.2 Comparison of two categorical variables
Low-fat yogurt Full-fat yogurt Total
Men 12 (30%) 28 (70%) 40 (100%)
Women 45 (75%) 15 (25%) 60 (100%)
Total 60 (60%) 40 (40%) 100 (100%)
17 2.4 Descriptive Plots
values compared with the rest of the data), it is common to use the standard devia-
tion. It can be estimated for a given variable X with n observations as

( )
=
-
=
-

2
1
Standard deviation (SD)
1
n
i
i
x x
n


If we exclude the extreme pH value in sample 10 (regarded as an outlier), then
the new mean of our remaining eight data points on pH is estimated to be 4.28 and
the standard deviation is estimated as

( ) ( ) ( )
- + - + + -
= =
-
2 2 2
4.41 4.28 4.21 4.28 ... 4.22 4.28
SD 0.09
8 1


Fortunately, most spreadsheets such as Excel or OpenOf ce or statistical soft-
ware can estimate standard deviations and other statistics ef ciently and lessen the
need to know the exact estimation formulas and computing techniques. If we assume
that our data are more or less normally distributed, then a distance of one standard
deviation from the mean will contain approximately 65% of our data. Two standard
deviations from the mean will contain approximately 95% of our data. This is the
main reason why continuous data are often described using the mean and standard
deviation.
If a data set has a skewed distribution or contains many outliers or extreme val-
ues, it is more common to describe the data as the median, with the spread repre-
sented by the minimum and maximum values. To reduce the effect of extreme
values, the so-called interquartile range is an alternative measure of spread in data.
It is equal to the difference between the third and rst quartiles. It can be found by
ranking all the observations in ascending order. For the sake of simplicity, let us
assume one has 100 observations. The lower boundary of the interquartile range is
at the border of the rst 25% of observations in this example observation 25 if they
are ranked in ascending order. The higher boundary of the interquartile range is at
the border of the rst 75% of observations in this example observation 75 if they
are ranked in ascending order.
2.4 Descriptive Plots
The adage a picture is worth a thousand words refers to the idea that a complex
idea can be conveyed with just a single still image. Actually, some attribute this
quote to the Emperor Napoleon Bonaparte, who allegedly said, Un bon croquis
vaut mieux quun long discours (a good sketch is better than a long speech). We
might venture to rephrase Napoleon to describe data A good plot is worth more
18 2 Methods and Principles of Statistical Analysis
M
e
a
s
u
r
e
d

v
a
l
u
e
5
123456789
1
0
1
1
1
2
1
3
1
4
1
5
1
6
1
7
1
8
1
9
2
0
0
a d
e
b
c
5
10
15
Fig. 2.1 Typically used descriptive plots. The plots are ( a ) bar chart, ( b ) box plot, ( c ) line plot,
( d ) histogram, and ( e ) scatterplot. All plots were made using data in Box 5.1 in Chap. 5
than a thousand data. Plots are very useful for describing the properties of data.
It is recommended that these be explored before further formal statistical analysis is
conducted. Some examples of descriptive plots are given in Fig. 2.1

19 2.4 Descriptive Plots
2.4.1 Bar Chart
A bar chart or bar graph is a chart with rectangular bars with lengths proportional to
the values they represent. They can be plotted vertically or horizontally. For categori-
cal data the length of the bar is usually the number of observations or the percentage
distribution, and for discrete or continuous data the length of the bar is usually the
mean or median with (error) lines sometimes representing the variation expressed as,
for example, the standard deviation or minimum and maximum values. Bar charts are
very useful for presenting data in a comprehensible way to a nonstatistical audience.
Bar charts are therefore often used in the mass media to describe data.
2.4.2 Histograms
Sometimes it is useful to know more about the exact spread and distribution of a
data set. Are there many outliers, or is the data distribution equally spread out? To
know more about this, one could make a histogram, which is a simple graphical way
of presenting a complete set of observation in which the number (or percentage
frequency) of observations is plotted for intervals of values.
2.4.3 Box Plots
A box plot (also known as a box-and-whisker diagram) is a very ef cient way of
describing numerical data. It is often used in applied statistical analysis but is not as
intuitive for nonstatistical readers. The plot is based on a ve-number summary of a
data set: the smallest observation (minimum), the lower quartile (cutoff value of the
lowest 25% of observations if ranked in ascending order), the median, the upper
quartile (cutoff value of the rst 75% of observations if ranked in ascending order),
and the highest observation (maximum). Often the whiskers may indicate the
2.5% and 97.5% values with outliers and extreme values indicated by individual
dots. Box plots provide more information about the distribution than bar charts.
If the line indicating the median is not in the middle of the box, then this is usually
a sign of a skewed distribution.
2.4.4 Scatterplots
Scatterplots are very useful for displaying the relationship between two numerical
variables. These plots are also sometimes called XY-scatter or XY-plots in certain
software. A scatterplot is a simple graph in which the values of one variable are
20 2 Methods and Principles of Statistical Analysis
plotted against those of the other. These plots are often the rst step in the statistical
analysis of the correlation between variables and subsequent regression analysis.
2.4.5 Line Plots
A line plot or graph displays information as a series of data points connected by
lines. Depending on what is to be illustrated, the data points can be single observa-
tions or statistical estimates as, for example the mean, median, or sum. As with the
bar chart, vertical lines representing data variation, for example standard deviation,
may then be used. Line plots are often used if one is dealing with repeated measure-
ments over a given time span.
2.5 Statistical Inference (the p -Value Stuff)
Descriptive statistics are used to present and summarize ndings. This may form the
basis for decision making and conclusions in, for example, scienti c and academic
reports, recommendations to governmental agencies, or advice for industrial pro-
duction and food development. However, what if the ndings were just due to a
coincidence? If the experiment were repeated and new data collected, a different
conclusion might be reached. With statistical methods it is necessary to assess
whether ndings are due to randomness and coincidence or are representative of the
true or underlying effect. One set of tools is called statistical tests (or inference)
and form the basis of p -values and con dence intervals.
The basis is a hypothesis that could be rejected in relation to an alternative hypothesis
given certain conditions. In statistical sciences these hypotheses are known as the null
hypothesis (typically a conservative hypothesis of no real difference between samples,
no correlation, etc.) and the alternative hypothesis (i.e., that the null hypothesis is not in
reality true). The principle is to assume that the null hypothesis is true. Methods based
on mathematical statistics have been developed to estimate the probability of outcomes
that are at least as rare as the observed outcomes, given the assumption that the null
hypothesis is true. This probability is the well-known p -value. If this probability is small
(typical less than 5%), then the null hypothesis is typically rejected in favor of the
alternative hypothesis. The level of this probability before the null hypothesis is rejected
is called the signi cance level (often denoted a ).
The relationship between the (unknown) reality if the null hypothesis is true or
not and the decision to accept or reject the null hypothesis is shown in Table 2.3 .
Two types of error can be made Type I and Type II errors. The signi cance level
a is typically set low (e.g., 5%) to avoid Type I errors that from a methodological
point of view are regarded as being more serious than Type II errors. The null
hypothesis is usually very conservative and assumes, for example, no difference
between groups or no correlation. The Type II error is denoted by b . The statistical
21 2.7 Overview of Statistical Models
power is the ability of a test to detect a true effect, i.e., reject the null hypothesis if
the alternative hypothesis is true. Thus, this is the opposite of a Type II error and
consequently equal to 1- b .
2.6 Overview of Classical Statistical Tests
Classical statistical tests are pervasive in research literature. More complex and gen-
eral statistical models can often express the same information as these tests. Table 2.4
presents a list of some common statistical tests. It goes beyond the scope of this
brief text to explain the statistical and mathematical foundations of these tests, but
they are covered in several of the recommended textbooks. Modern software often
has menu-based dialogs to help one determine the correct test. However, a basic
understanding of their properties is still important.
2.7 Overview of Statistical Models
Generally speaking, so-called linear statistical models state that your outcome of
interest (or a mathematical transformation of it) can be predicted by a linear combi-
nation of explanatory variables, each of which is multiplied by a parameter (some-
times called a coef cient and often denoted b ). To avoid having the outcome be
estimated as zero if all explanatory variables are zero, a constant intercept (often
denoted b
0
) is included. The outcome variable of interest is often called the depen-
dent variable, while the explanatory variables that can predict the outcome are called
independent variables.
The terminology in statistics and experimental design may sometimes be some-
what confusing. In all practical applications, models like linear regression, analysis of
covariance (ANCOVA), analysis of variance (ANOVA), or general linear models
(GLM) are very similar. Their different terminology is due as much to the historical
tradition in statistical science as to differences in methodology. Many of these models
with their different names and terminologies can be expressed within the framework
of generalized linear models. It was common to develop mathematical methods to
estimate parameter values and p -values that could be calculated manually by hand and
Table 2.3 Two types of statistical errors: Types I and II errors and their relationship to signi cance
level a and the statistical power (1- b )
Null hypothesis
(H
0
) is true Alternative hypothesis (H
1
) is true
Accept null hypothesis Correct decision Type II error: b
Reject null hypothesis Type I error: a Correct decision
22 2 Methods and Principles of Statistical Analysis
T
a
b
l
e

2
.
4

P
r
o
p
o
s
e
d

s
t
a
t
i
s
t
i
c
a
l

t
e
s
t
s

o
r

m
o
d
e
l
s

d
e
p
e
n
d
i
n
g

o
n

p
r
o
p
e
r
t
i
e
s

o
f

t
h
e

o
u
t
c
o
m
e

a
n
d

e
x
p
l
a
n
a
t
o
r
y

v
a
r
i
a
b
l
e
.

N
o
n
p
a
r
a
m
e
t
r
i
c

a
l
t
e
r
n
a
t
i
v
e

i
s

g
i
v
e
n

i
n

b
r
a
c
k
e
t
s

i
f

a
s
s
u
m
p
t
i
o
n
s

o
n

n
o
r
m
a
l

d
i
s
t
r
i
b
u
t
i
o
n
s

a
r
e

n
o
t

v
a
l
i
d
.

T
h
e

n
u
m
b
e
r

o
f

m
e
n
t
i
o
n
e
d

t
e
s
t
s

i
s

l
i
m
i
t
e
d

a
n
d

r
e
c
o
m
m
e
n
d
a
t
i
o
n
s

m
a
y

v
a
r
y

d
e
p
e
n
d
i
n
g

o
n

t
h
e

n
a
t
u
r
e

o
f

t
h
e

d
a
t
a

a
n
d

p
u
r
p
o
s
e

o
f

a
n
a
l
y
s
i
s

P
u
r
p
o
s
e

w
i
t
h

s
t
a
t
i
s
t
i
c
a
l

a
n
a
l
y
s
i
s

T
y
p
e

o
f

o
u
t
c
o
m
e

d
a
t
a

N
o
m
i
n
a
l

B
i
n
a
r
y

O
r
d
i
n
a
l

D
i
s
c
r
e
t
e

C
o
n
t
i
n
u
o
u
s

A
g
a
i
n
s
t

s
p
e
c
i

c

n
u
l
l

h
y
p
o
t
h
e
s
i
s

a
b
o
u
t

e
x
p
e
c
t
e
d

m
e
a
n

o
r

p
r
o
p
o
r
t
i
o
n

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

B
i
n
o
m
i
a
l

t
e
s
t

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

O
n
e

s
a
m
p
l
e

t
-
t
e
s
t

O
n
e

s
a
m
p
l
e

t
-
t
e
s
t

R
e
l
a
t
i
o
n
s
h
i
p

w
i
t
h

c
o
n
t
i
n
u
o
u
s

e
x
p
l
a
n
a
t
o
r
y

v
a
r
i
a
b
l
e


U
s
e

a

s
t
a
t
i
s
t
i
c
a
l

m
o
d
e
l

U
s
e

a

s
t
a
t
i
s
t
i
c
a
l

m
o
d
e
l

S
p
e
a
r
m
a
n

c
o
r
r
e
l
a
t
i
o
n

P
e
a
r
s
o
n

(
S
p
e
a
r
m
a
n
)

c
o
r
r
e
l
a
t
i
o
n

P
e
a
r
s
o
n

(
S
p
e
a
r
m
a
n
)

c
o
r
r
e
l
a
t
i
o
n

D
i
f
f
e
r
e
n
c
e

i
n

e
x
p
e
c
t
e
d

m
e
a
n

o
r

p
r
o
p
o
r
t
i
o
n
s

b
e
t
w
e
e
n

t
w
o

g
r
o
u
p
s

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

f
o
r

c
r
o
s
s

t
a
b
s

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

f
o
r

c
r
o
s
s
t
a
b
s

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

f
o
r

c
r
o
s
s
t
a
b
s

T
w
o
-
s
a
m
p
l
e

t
-
t
e
s
t

(
M
a
n
n

W
h
i
t
n
e
y

U

t
e
s
t
)

T
w
o
-
s
a
m
p
l
e

t
-
t
e
s
t

(
M
a
n
n

W
h
i
t
n
e
y

U

t
e
s
t
)

D
i
f
f
e
r
e
n
c
e

b
e
t
w
e
e
n

m
e
a
n

o
r

p
r
o
p
o
r
t
i
o
n
s

b
e
t
w
e
e
n

m
o
r
e

t
h
a
n

t
w
o

g
r
o
u
p
s

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

f
o
r

c
r
o
s
s
t
a
b
s

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

f
o
r

c
r
o
s
s
t
a
b
s

C
h
i
-
s
q
u
a
r
e
d

t
e
s
t

f
o
r

c
r
o
s
s
t
a
b
s

A
n
a
l
y
s
i
s

o
f

v
a
r
i
a
n
c
e

(
K
r
u
s
k
a
l

W
a
l
l
i
s

H

t
e
s
t
)

A
n
a
l
y
s
i
s

o
f

v
a
r
i
a
n
c
e

(
K
r
u
s
k
a
l

W
a
l
l
i
s

H

t
e
s
t
)

A
n
a
l
y
z
e
d

a
s

l
i
n
e
a
r

s
t
a
t
i
s
t
i
c
a
l

m
o
d
e
l

M
u
l
t
i
n
o
m
i
a
l

l
o
g
i
s
t
i
c

r
e
g
r
e
s
s
i
o
n

B
i
n
a
r
y

l
o
g
i
s
t
i
c

r
e
g
r
e
s
s
i
o
n

O
r
d
i
n
a
l

l
o
g
i
s
t
i
c

r
e
g
r
e
s
s
i
o
n

L
i
n
e
a
r

r
e
g
r
e
s
s
i
o
n
/
g
e
n
e
r
a
l

l
i
n
e
a
r

m
o
d
e
l

L
i
n
e
a
r

r
e
g
r
e
s
s
i
o
n
/
g
e
n
e
r
a
l

l
i
n
e
a
r

m
o
d
e
l

T
w
o

c
l
u
s
t
e
r
e
d

o
r

r
e
p
e
a
t
e
d

m
e
a
s
u
r
e
m
e
n
t
s

M
c
N
e
m
a
r

B
o
w
k
e
r

t
e
s
t

M
c
N
e
m
a
r

t
e
s
t

M
c
N
e
m
a
r

B
o
w
k
e
r

t
e
s
t

P
a
i
r
e
d

s
a
m
p
l
e

t
-
t
e
s
t

(
W
i
l
c
o
x
o
n

s
i
g
n
e
d
-
r
a
n
k

t
e
s
t
)

P
a
i
r
e
d

s
a
m
p
l
e

t
-
t
e
s
t

(
W
i
l
c
o
x
o
n

s
i
g
n
e
d
-
r
a
n
k

t
e
s
t
)

S
t
a
t
i
s
t
i
c
a
l

m
o
d
e
l

f
o
r

c
l
u
s
t
e
r
e
d

o
r

r
e
p
e
a
t
e
d

m
e
a
s
u
r
e
m
e
n
t
s

M
i
x
e
d

m
u
l
t
i
n
o
m
i
a
l

l
o
g
i
s
t
i
c

r
e
g
r
e
s
s
i
o
n

o
r

G
E
E

M
i
x
e
d

b
i
n
a
r
y

l
o
g
i
s
t
i
c

r
e
g
r
e
s
s
i
o
n

o
r

G
E
E

M
i
x
e
d

o
r
d
i
n
a
l

l
o
g
i
s
t
i
c

r
e
g
r
e
s
s
i
o
n

o
r

G
E
E

L
i
n
e
a
r

m
i
x
e
d

m
o
d
e
l

o
r

G
E
E

L
i
n
e
a
r

m
i
x
e
d

m
o
d
e
l

o
r

G
E
E

G
E
E

g
e
n
e
r
a
l
i
z
e
d

e
s
t
i
m
a
t
i
n
g

e
q
u
a
t
i
o
n
s

23 References
with the help of statistical tables. Most graduates in statistics are familiar with such
methods for simple regression and ANOVA methods. However, recent innovations in
mathematical statistics, and not least computers and software, have in an applied sense
replaced such manual methods. These computer-assisted methods are usually based
on the theory of so-called likelihood functions and involve nding their maximum
values by using iterations. In other words, these are methods where computer software
is needed for most applied circumstances. The theory behind maximum-likelihood
estimations is covered in several of the more advanced recommended textbooks.
Linear statistical models are often described within the framework of generalized
linear models. The type of model is determined by the properties of the outcome
variable. A dependent variable with continuous data is usually expressed with an
identity link and is often referred to by more traditional terms such as linear regression
or analysis of variance. If the dependent variable is binary, then it is usually expressed
by a logit link and is often referred to by the more traditional term logistic regression .
Count data use a log link and the statistical model is traditionally referred to as
Poisson regression (e.g., Dobsen and Barnett 2008 ) .
References
Agresti A (2002) Categorical data analysis, 2nd edn. Wiley, Hoboken
Cleophas TJ, Zwinderman AH (2012) Statistics applied to clinical studies, 5th edn. Springer,
Dordrecht
Devore JL, Kenneth N (2012) Modern mathematical statistics with applications, 2nd edn. Springer,
New York
Dobsen AJ, Barnett A (2008) An introduction to generalized linear models, 3rd edn. CRC Press,
London
Ibrahim JG, Molenberghs G (2009) Missing data methods in longitudinal studies: a review. Test
18:143. doi: 10.1007/s11749-0090138-x
Kaltenbach HM (2011) A concise guide to statistics. Springer, New York
Kleinbaum DG, Sullivan K, Barker N (2007) A pocket guide to epidemiology. Springer, New
York
Lehmann EL (2011) Fisher, Neyman, and the creation of classical statistics. Springer, New York
Madsen B (2011) Statistics for non-statisticians. Springer, Heidelberg
Marques de S JP (2007) Applied statistics using SPSS, STATISTICA, MATLAB and R, 2nd edn.
Springer, Berlin
Shahbaba R (2012) Biostatistics with R: an introduction to statistics through biological data.
Springer, New York
Song PXK (2007) Missing data in longitudinal studies. In: Correlated data analysis: modeling,
analytics, and applications. Springer, New York
Tamine AY, Robinson RK (2007) Tamine and Robinsons yoghurt science and technology, 3rd edn.
CRC Press, Cambridge
R Development Core Team (2012) The R project for statistical computing. http://www.r-project.
org . Accessed 30 Apr 2012
Vittinghoff E, Glidden DV, Shiboski SC, McCulloch CE (2012) Regression methods in biostatistics:
linear, logistic, survival and repeated measures models, 2nd edn. Springer, New York
http://www.springer.com/978-1-4614-5009-2