Basic principles of graphing data

483

Point of View

Basic principles of graphing data

Marcin Kozak
Warsaw University of Life Sciences Dept. of Experimental Design and Bioinformatics Nowoursynowska
159 02-776 Warsaw, Poland. E-mail <[email protected]>
ABSTRACT: Data visualization is a very important aspect of data analysis and of presentation. Focusing on
the latter, this paper discusses various elements of constructing graphs for publications. Bad and good graphs
are compared, and a checklist with graphical elements to be used while creating graphs is proposed.
Key words: visualization, graphs, statistics

Princpios bsicos na confeco de grficos

RESUMO: A visualizao de dados um aspecto importante da anlise de dados e de sua apresentao. Nesse
ponto de vista enfocam-se vrios elementos da construo de grficos para publicaes. Grficos ruins e bons
so comparados, e uma lista de checagem de elementos de composio grfica proposta para ser utilizada
durante a confeco de grficos.
Palavras-chave: visualizao, grficos, estatstica

Introduction
Agricultural scientists produce abundant results.
Some of them are important, others may be less important, still others may be negligible. The way one presents the results, then, counts for several reasons. First,
one needs to emphasize those important ones. Second,
for most data, interpretation is easiest with graphs, and
it is far easier based on a good than a bad graph. Just as
statistical analysis can provide false conclusions (Huff,
1954; Kozak, 2009a), poor graphs can misinform, sometimes leading to serious misinterpretations.
Huff (1954) was probably the first to discuss that
graphing statistical data can be a tool of misinformation.
Later some of his notions found counter-arguments, yet
the main ideas still hold: graphing must be correct and
convey true information and interpretation. Scientific
graphs are not to be beautiful; they are to be informative. Of course, ugly graphs will not convey any interesting message, so elegance should not be disregarded
(Tufte 1991; 1997; 2001 and 2006).
Tufte (1991; 1997; 2001 and 2006), Cleveland (1993; 1994),
Jacoby (1997; 1998) and Wilkinson (2005), among others,
discuss graphing data in great detail. Harris (1999) is a useful reference for information graphics. Yet scientific literature in the 21st century is full of poor graphs, some of which
are incomprehensible while others can even unintentionally falsify information. Data visualization is a developing
research area, and what was considered a good graph 30
years ago does not have to be good these days. In this paper basic information is offered about graphing data with
the help of which authors should be able to construct sufficiently good scientific graphs. Focus will mainly be
placed on graphical rather than statistical aspects, so remember that when constructing a graph, you must take all

care to ensure it delivers the message you intend it to, from

both scientific and statistical points of view. Note also that
visualization data with the purpose of data analysis and interpretation at the computer screen does not follow exactly
the same rules as those described in this paper deals
mainly with graphing for others, that is, for publications.
You will see good and bad graphs, what can go
wrong when visualizing data, and why it is wrong. At
the same time you will see how the problem can be
overcome, and what to do to ensure the graph will be
fine. In the next section, General principles of graphing data, a general vision of graphing scientific data
is presented. Next, in the section Specific principles
of graphing data some detailed rules to be followed
are presented to ensure if a graph is efficient. Do not
forget, however, that these are really basic principles,
as the title of the paper says; if one wants to learn more
about this topic, one should refer to sources cited in
this paper. The last section offers a short conclusion,
while the Appendix lists data sets used in this paper
and the information on the software employed to analyze and graph these data. Note that in figure captions
more information than is normally needed is given, for
example the type of graph. Each figure caption informs
whether the figure (or its panels) represents a good or
rather poor style.

General principles of graphing data

The general and most important principle of graphing data is to construct a graph so that it conveys the
message in the most efficient way, and the message one
wants it to convey. Thus, all elements of the graph
should be helpful but not distracting, and important aspects should be emphasized but not hidden. This is the

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

484

Kozak

most general rule, and all the following rules account

for this general one. Remember that all details count:
viewers of a graph will follow your ideas if you help
them. So if you dont feel an expert in visualization,
think twice - or rather thrice - before deciding not to follow any of the rules.
Before making a graph, think over what message it
is to convey and whether it is needed whatsoever. In general, graphing three numbers is not advisable; no matter
how amazing this can seem, one can find quite a few
such graphs, especially piecharts and barplots. In fact,
two numbers are sometimes graphed. Two numbers will
be most efficiently presented within a sentence, while
few numbers within a text-table (which is a simple table
inserted within the text - Tufte, 2001; Kozak, 2009b) or a
regular table. Tufte (2001) argues that even a dozen or
so numbers are better represented by a table than a
graph, but this is arguable - in general for 10 and even
fewer numbers to be compared a graph can be preferred,
although it depends on various things, including what
the numbers represent.
Make graphs as simple and as complex as it is needed
to deliver the message. But neither be afraid of a complex graph nor make it more complex than it should be.
Remember that the human eye can work efficiently with
very composite images, detecting large amount of information in small spaces (Tufte, 2001). However, always
be careful to make the graph readable and understandable: avoid clutter (see, e.g., Reynolds et al., 2009; Silva,
2009c).
An efficient way to proceed while constructing a
graph is to:
1) figure out the contexts and the message
2) figure out the way of presenting it, so the type, layout and style of graph to be used

3) construct the graph

4) check Tables 1 and 2 and revise the graph accordingly
5) check whether the revised version conveys the message you want it to convey
6) check Tables 1 and 2 and revise the graph accordingly
7) and so on
Table 1, mentioned before, offers basic elements to
check while constructing a graph - they will be discussed
in the following section. They do not refer to every type
of graph, but it will be best to check everything that is
listed there, and decide what does and what does not
apply to your graph. Table 2 cites Clevelands (1994)
very useful principles of graphing data.
Note how similar the process of graph construction
is to the process of writing: figure out the message and
style, construct, revise, revise, revise, revise Revising
cannot be ignored, and exactly like in writing, it can be
a source of inspiration, leading to results and conclusions one could not even imagine before. The above algorithm of graph construction together with Tables 1
and 2 will be especially useful for non-experienced graph
constructors, because those experienced ones follow it
intuitively.

Specific principles of graphing data

Data points are very important elements of many
graphs, especially scatterplot and line plot with lines
superimposed on data points. Symbols representing
them together with their size and color should be carefully chosen so that the points could be easily seen,
and the patterns present in the data (or their lack)
could be easily noticed. Figure 1, picturing logarithm
of the volume of cherry trees against the logarithm of

Table 1 Elements to check while constructing a graph. Refer to the text for details.
Data points

symbol, size, color, overlap

Lines

type, width, color, overlap

Color

check whether needed at all; check whether all elements are easily distinguished
type of box, aspect, minimum and maximum value, data rectangle and scale-line rectangle, tick marks
(number of, location, direction, length and width), tick marks' labels (font type, length, rotation,
numbering style, abbreviations of text labels)
check whether needed, position, box around, size, elements within (see above for "data points" and
"lines", and "labels" within "box and axes")
of the graph; of bars, cross-hatching, shading, color

Box and axes

Legend
Background
Text inside the graph

Dimensions

check whether needed; font style, size and color

check whether informative, explaining everything that is needed to understand the graph and the
message it delivers
check for over-dimensionality

Captions

Reference line

check whether needed, line type, width and color

Grid lines

check whether needed, line type, width and color

Error bars

check if the information is given on what they represent; line type, width and color, type of ending
all the above elements; layout (rows, columns, pages), choice of panel variables and conditioning
variable(s), panel order

Trellis display

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

Basic principles of graphing data

485

Table 2 Clear vision principles given by Cleveland (1994).

1. Make the data stand out. Avoid superfluity.
2. Use visually prominent graphical elements to show the data.
3. Use a pair of scale lines for each variable. Make the data rectangle slightly smaller than the scale-line rectangle. Tick marks
should point outward.
4. Do not clutter the interior of the scale-line rectangle.
5. Do not overdo the number of tick marks.
6. Use a reference line when there is an important value that must be seen across the entire graph, but do not let the line
interfere with the data.
7. Do not allow data labels in the interior of the scale-line rectangle to interfere with the quantitative data or to clutter the
graph.
8. Avoid putting notes and keys inside the scale-line rectangle. Put a key outside, and put notes in the caption or in the text.
9. Overlapping plotting symbols must be visually distinguishable.
10. Superposed data sets must be readily visually assembled.
11. Visual clarity must be preserved under reduction and reproduction.
12. Put major conclusions into graphical form. Make captions comprehensive and informative.

Figure 1 A scatterplot with a regression line superimposed,

representing trees volume against girth, both after
a logarithmic transformation (data set trees). The
graph is full of drawbacks: filled squares as plotting
symbols are difficult to distinguish; the box consists
of two scale lines only; the data rectangle is too large
so the points are clustered within a small area; the
tick mark labels on the y-axis are vertical instead of
horizontal even though they are just one-digit
numbers; the tick marks point inward. Note that
the regression line extrapolates the values outside
the ranges observed in the experiment, which can
be dangerous.

girth (data trees), is poor. The choice of filled squares

as symbols is bad because it is difficult to distinguish
the points. (By the way, a filled rhombus, which is
the default symbol for a scatterplot in Microsoft Excel, is an equally poor choice.) Compare it with Figure 2, whose readability is greatly enhanced thanks
to open circles used as symbolsin fact, open circles
should be considered the best symbol in case of some
overlap in scatterplots for non-grouped data. (Note
that for dotplots, in which no overlap occurs, closed
circles are usually used as plotting symbolsFigures

Figure 2 A scatterplot with a regression line superimposed,

based on the same data as in Figure 1, but all the
drawbacks mentioned there are overcome. It is
considered a GOOD GRAPH.

3 and 4). Sometimes, however, overlap can be a serious problemsee Figure 5.

When data are grouped, the choice of plotting symbols matters too: c, z, }, should work well for little
overlap, and c, +, <, s, w for serious overlap (Cleveland, 1994); but this will strongly depend on what is offered by software one uses. Varying color for groups can
be even more efficient, not to mention combining plotting symbol and color. Sometimes decreasing the size
of symbols will suffice, but a special technique to deal
with overlapping points is jittering (Cleveland, 1994),
which was applied in the right panel of Figure 5. This
technique consists of adding a small amount of random
noise to the data, thanks to which the overlapping points
are usually easier to distinguish. There are also other
techniques to deal with serious overlap when there are
many points to draw, examples being sunflower plot
(Cleveland and McGill, 1984), graphing binned data

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

486

Figure 3 A vertical dotplot representing mean area under the

disease progress curve (AUDPC) of 16 clones in one
environment (Hastings, FL; data set haynes). The
clones are ordered by decreasing yield. A drawback
of this graph is the legend: because the clones are
presented in the x-axis, their names are too long to
be presented without rotation or abbreviation. The
abbreviated names are used here, which calls for the
legend to explain the abbreviations.

Kozak

Figure 4 A horizontal dotplot representing the same data as

in Figure 3. The drawback of Figure 3 is overcome
here: vertical dotplot enables one to use sufficiently
long labels so that no legend is needed. Note that
this Figure is smaller than the previous one, yet it is
better readable. It is considered a GOOD GRAPH.

Figure 5 Scatterplots representing replicated data of number of insects caught after applying various insecticides (data set InsectsSprays).
A drawback of the left panel is the overlap of points. This is overcome in the right panel, where jittering is employed
(Observe the GOOD GRAPH IN RIGHT PANEL).

(Carr et al., 1987), a combination of these two (Dupont

and Plummer, 2003), or graphing regions of bivariate
kernel density (Hydman, 1996). Yet another idea is to
use the trellis display (Cleveland, 1993) for grouped
data, in which case the groups will be plotted in different panels, alleviating the overlap of points from different groups; this type of display will be discussed
later.
Similar issues ought to be raised for lines: they
should be easily distinguished, for which varying line
type (e.g., solid, dashed, dotted), width and color can
be used. Lines can be plotted together with data points,
in which case their efficient combinations should be

chosen. In Figure 6 three options are presented for this:

in the top panel, each of the five lines representing the
growth of a chicken is plotted with the same style of
line and plotting symbol. Note that up to about 10 days
the lines are difficult to be distinguished, and it is practically impossible to notice here what can be seen in
the middle and bottom panels: that the lightest chicken
at the end of the experiment was one of the heaviest
during the first days. Color in the bottom panel seems
to work better than varying line type and plotting symbol in the middle panel, but both succeed to convey
the most important information. Note that not all the
information can be extracted from these plots. In this

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

Basic principles of graphing data

particular situation it does not seem a problem, but a

simple solution might be to plot a logarithm of body
weight on the y-axis - see Figure 7, where the lines are
slightly easier to be distinguished for the first days of

487

the experiment (although in many situations the gain

in readability will be much larger).
Color can be powerful, but do not overuse it. Save
color differentiation for situations when it helps, and
dont use it when it adds nothing to the graph. For example, would it make sense to use different colors for
the bars in Figure 8, picturing mean count of insects
caught after applying six insecticides? Of course not
not helping to differentiate the insecticides, it would be
redundant and distracting. But sometimes color can be
very helpful, supporting a viewer by clear group differentiation (e.g., Figures 7, 10 and 15, and Lammel et al.,
2007; Jaradat, 2007; 2009). Remember that color will help

Figure 7 A line plot representing the same data as in the

bottom panel of Figure 6, but body weight is presented
in natural logarithms. This helps one to compare the
lines in the early days of growth. The aspect ratio has
been chosen according to the baking to the 45 rule
note that the aspect ratio is different than in the plots
in Figure 6, because of differentscales for variables in
these two plots. It is a GOOD GRAPH.

Figure 6 Line plots representing changes in body weight of

five chickens through time (data set ChickWeight).
The problem in the top panel is that all five lines are
plotted with the same line style and plotting symbol.
This is overcome in THE MIDDLE AND BOTTOM PANELS
(GOOD GRAPHS). In all graphs, the aspect ratio has
been chosen according to the baking to the 45
rule.

Figure 8 A barplot representing mean count of insects caught

after applications of six insecticides (data set
InsectSprays). Error bars represent standard error
(SE) of the mean estimated with generalized least
squares (Pinheiro and Bates, 2000). Drawbacks
The insecticides are ordered alphabetically even
though their names have no meaning; tick mark
labels at the vertical axis can be presented vertically.

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

488

only if other graphical elements are well arranged. For

example, Vzquez et al. (2009) had eight groups to distinguish in a scatterplot, and they did that by varying
plotting symbols and their color. Unfortunately, the
large size of symbols made it practically impossible to
compare the groups.
Box (frame) around a graph plays an important role:
it consists of axes (scales) that work as rulers along
which data are graphed (Cleveland, 1994). Box and axes
help see the data and locate coordinates of data points.
Whether the box should consist of two axes, which is
a tradition, or four axes, which is more efficient (Cleveland, 1994), is a matter of discussion. A four-axis box
is better, unless one is forced to do otherwise (for example, by a journal editor, which sometimes happens).
Compare Figures 1 and 2 and note how well the full
box in the latter defines the plotting region compared
to the two-axis box in the former. A very important aspect of the box is that the data rectangle (so the region
used for plotting) should be slightly smaller than the
box itself (Cleveland, 1994) - otherwise the plotting symbols will interfere with the axes (see Figure 9 and Virdis
et al., 2009), although Tuftes (1997, p. 45) graph is a
very elegant with exception to this rule. Remember also
that the data rectangle should be effectively usedfor
a variable ranging from 50 to 105 it would seldom be
wise to start the axis at 10 and end at 170 (see also Figure 1). Of course, there are exceptions to this rule, when
there are reasons to start and/or end the axes at particular values. In addition, the appearance of tick marks
matters: most software has some algorithm of choosing their number and location, but remember not to
use too many of them. They also should neither be too
long nor too short, and they should point outwards.
Refer to Table 1 for further details.

Figure 9 A scatterplot with a regression line superimposed,

based on the same data as in Figure 2, but the
drawback of this graph is that the data rectangle is
the same as the box, so four points are graphed at
the axes.

Kozak

Aspect ratio of a data rectangle is its height divided

by width (Cleveland, 1994). It is the aspect ratio that decides about the shape of the graph, so it should be chosen with care. Try not to choose the aspect ratio so that
the graph fills the space available in the publication, but
choose it so that a graph conveys a message. Tufte (2001)
suggested following the nature of data if there is any hint
about the shape of the graph, and otherwise making the
graph wider than tall about 50%. Cleveland (1994) proposed a rule of banking to 45, which he showed to be
very useful for line plots and scatterplots with a superimposed locally weighted regression line; this method
optimizes judging the rate of change. Sometimes banking to 45 does not work because it makes the graph not
readable, but the point is that the aspect ratio is important. In general, unless there are particular reasons for
that, the box should be neither too narrow nor too wide
compared to its height. It also strongly depends on the
type of plot: for example, for scatterplots square boxes
might be good - too many scatterplots in the scientific
literature are far too wide. Sometimes a so-called isometric aspect ratio can be employed, in which the relation between physical distance on the device and distance in the data scale are forced to be the same for both
axes (Sarkar, 2008) - see Figure 10 (which will be discussed in detail in the next section), in which petal
length and width are in cm, so the isometric scale helps
compare these two traits.
Order of elements matters. This is very important
for tables (Ehrenberg, 1977), but equally important for
groups compared by means of barplots, dotplots,
boxplots and the like. It is generally acknowledged that

Figure 10 A scatterplot representing petal length versus petal

width of three iris species (data set iris). The shape
of the graph results from the isometric scales. Look
that it is a GOOD GRAPH.

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

Basic principles of graphing data

comparing several numbers ordered alphabetically by

labels, like here:
A
B
C
D

101.2
112.1
110.2
102.1

is more difficult than comparing these same numbers

ordered by size, like here:
B
C
D
A

112.1
110.2
102.1
101.2

Compare the two boxplots presented in Figure 11.

Clearly the ordering by median insect count makes the
comparison of insecticides easier than the alphabetical
ordering. Of course, you must decide if alphabetical ordering by labels does not matter indeed. Note also that
for several graphs you may want to apply the same ordering to facilitate the general interpretation of the groups.
Text in graphs appears in various places: axis and tick
mark labels, legend (sometimes), title (sometimes, usually for multi-panel graphs, like in Figure 11), and data
point labels (see the comment below and Tables 1 and
2). It also appears in figure captions, but we will discuss
them later. Use a font that will be readable for everyone; in most situations sans-serif fonts (e.g., Arial) work
best for graphs. Horizontal labels are easier to read than
vertical or rotated ones (cf. Figures 1 and 2)so horizontal barplots and dotplots (cf. Figures 3 and 4) will
usually be better than vertical ones. If numbers are too
long to present them horizontally on the y-axis, consider
dividing them by some reasonable factor. This does not
mean that one should never present vertical tick mark
labels at the y-axis, but that in most situations this will
not be the best solution. Try to avoid putting text within
the data rectangle when it can distract a readers atten-

489

tion (e.g., Mohapatra and Kariali 2008); for example, regression equations are far too often put inside graphs
(Fiorio and Dematte, 2009; Pahlavani et al., 2009; Simoes
et al., 2009). However, Tuftes (2006, p. 120) graph is an
excellent exception to this rule. Sometimes, labels are
needed inside the data rectangle, examples being biplots
(e.g., Lammel et al., 2007), but one must remember about
Clevelands (1994) rule of not allowing such data labels
to interfere with the quantitative data or to clutter the
graph (Table 2).
Do not overuse legends. If all the information necessary can be included in the graph itself, then there is no
need of legend whatsoever. For example, Silva et al.
(2009b) decided not to use a legend in a situation when
most others would. Jonsson and Aoyama (2009) would
have needed a legend had they used a horizontal version
of the barplot instead of the vertical one, while
Ambrosano et al. (2009) chose the vertical barplot and so
had to use the legend. Compare Figures 3 and 4, representing the same data: mean area under the disease
progress curve of 16 potato clones in one environment.
In the former, the clones are presented at the horizontal
axis (note that this is usually done for barplots). The
clones names are too long to be presented in full, so they
are abbreviated and the legend is needed to explain the
abbreviations; even despite that, the font size needed to
be reduced. In Figure 4, a vertical version of the dotplot
is drawn with full names, so no legend is needed. Of
course, it does not mean that legends should never be
usedsometimes they are indeed required, for example
when the grouped data are plotted within one graph (e.g.,
Figure 10). Quite often the legend is included in a figure
caption (which saves space), but placing it near the graph
(like in Figure 10) facilitates reading. Nonetheless, remember Clevelands advice of avoiding putting a legend within
the box (Cleveland, 1994, Table 2; cf. Macedo et al., 2009).
When the legend is presented, it should rather not be surrounded by a box, as it is sometimes done.

Figure 11 Boxplots representing number of insects caught after applications of six insecticides (data set InsectSprays). In left panel,
alphabetical ordering is employed, while in right ordering by count. Observe the GOOD GRAPH IN RIGHT PANEL.
Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

490

Use figure captions not only to give the short title

of a graph, but also to comment on the graph. It can be
very efficient to draw viewers attention to some interesting aspects of the graph, while the same information
in the text has less power (the reader has to move between the text and figure). In fact, the more self-standing the caption, the better, but of course the balance
needs to be struck, and repeating everything from the
text does not have to be the best idea. (You also need to
remember that not all journals will accept long captions.)
In general, use white background for graphs in publications: in most situations it will be the best color if
the publication is printed on white paper. Thus for example grey, the default color of the background in MS
Excel 2003 charts, is generally a poor choice.
With barplots one needs to be very careful. All-toooften do authors choose cross-hatching to help the
viewer distinguish bars (e.g., Ambang et al., 2009;
Czubaszek, 2009; El-Shafey et al., 2009; Ofosu-Anim
and Leitch, 2009; Takahashi et al., 2009; Wiewira,
2009), the effect of which is rather poor: instead of easy
differentiation of groups the viewer is attacked with
what Tufte (2001) calls moir effects. Moir effects
are the optical noise that makes the bars difficult to
compare, and the effects the bars represent difficult to
grasp. Tufte (2001) calls such junk elements the
chartjunk. Cross-hatching, then, should be replaced
with shades of grey or color (Tufte, 2001); good examples are those by Karlidag et al. (2009) and Silva et
al. (2009a).
Bars in barplots should normally start at zero. This
is because in barplots we compare heights of bars, so
they must be meaningful: unless they start at zero, they
usually have no sensible meaning. Note that this does
not apply for dotplots (Figure 4), in which the viewer
compares the position of the points along the appropriate axis.
Error bars are very often used with statistical data.
Some rules should be followed when presenting them.
First of all, the information (usually provided in a figure caption) on what the error bars represent is criticalwithout it the reader will not be able to make any
use of the bars. Is it a standard deviation, standard error, confidence interval, Tukeys HSD, or still something else? Figures 8 and 12 show standard errors (SEs)
while Figure 13 confidence intervals (CIs)note how
lengths of SE and CI error bars differ. The choice is
extremely important, but this topic is beyond the scope
of this paper; refer to Cumming et al. (2004) and
Cumming and Finch (2005) for a discussion on the use
and interpretation of error bars. But one important aspect deserves to be mentioned: this is not true that two
overlapping confidence intervals indicate a statistically
significant difference at the corresponding significance
level (see Cumming and Finch, 2005 for the details).
What also matters when including error bars is the
way they are presented. Compare Figures 8 and 12 to
see that error bars are easier to read in dotplots than

Kozak

in barplots. If there are too many error bars to show

(e.g. Maci-Vicente et al., 2009, Scarpari and Beauclair,
2009), a good idea might be to present them in grey
(quite likely making the data points in black better visible), and in the case of overlap, to employ jittering.
For symmetric error bars only a half of them can be
presented, as Giacomini et al. (2009) did for standard
errors of the mean in line plots. They had two lines
that were superimposed on data points: for the upper
line the error bars pointed up, while for the bottom
line pointed down (in addition the lines and error bars
for the two groups differed in color), which avoided
overlap. However, if too many error bars make the
graph be in such a clutter that nothing can be seen in
it (including data points), the data should be presented
in some other way. Note that error bars in Figures 12
and 13 have different endings. Both are good, although

Figure 12 A dotplot representing the same data as in Figure 8,

but instead of the barplot, the dotplot is used (data
set InsectSprays). The insecticides are ordered by
decreasing mean count of insects. It is considered a
GOOD GRAPH.

Figure 13 A dotplot representing the same data as in Figure

12, but instead of standard errors of the means,
95% confidence intervals (CI) are presented. It is a
GOOD GRAPH.

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

491

Basic principles of graphing data

those in Figure 12 are preferred rather than those in Figure 13 because of easiness of comparison. The latter,
however, have this advantage that they look like a confidence interval closed within parentheses, which is the
way numeric confidence intervals are usually presented.
For multivariate data sets, a so-called trellis display
(Cleveland, 1993, 1994) can be of help. It consists of panels arranged into rows, columns and pages, each panel
containing a plot of the same type for subsets of the data
set defined in a particular way. (Note that in publications
trellis displays are seldom divided into pages.) This display can be very helpful in data analysis, interpretation
and presentation. The trellis display works well with
grouped data, in which case panels represent values of a
categorical conditioning variable(s), but also with quantitative predictors, in which case a conditioning variable(s)
is cut into intervals, and each panel represents those data
points for which the value of the conditioning variable
fits into the interval. Panels can represent various types
of plots, including scatterplots, dot plots, boxplots, line
plots and others. The trellis display can also be very useful in analyzing data from designed experiments (Cleveland and Fuentes, 1997). Most of the comments concerning graphing given before are equally important for this
type of display, but there are many other rules that can
or should be followedthe reader can refer for example
to Cleveland (1993) and Cleveland and Fuentes (1997). Figure 14, which will be discussed later, is an example of the
trellis display for a scatterplot of two quantitative variables and a categorical conditioning variable.
Some additional elements can be used to enhance a
graph. These could be for example a reference line (see
Table 2 for Clevelands [1994] comment) or, for multipanel graphs (especially scatterplots), visual grid. Re-

member that a visual grid seldom has to do with table

look-up of values (this could be the case for example for
barplots); it usually aims to enhance between-panel visual comparisons (Cleveland, 1994) - Figure 14.
There is no space for over-dimensionality. If you
are constructing a barplot, forget about 3D-barplots (e.g.,
Moussa and Abdel-Aziz, 2008; Abou Khalifa, 2009;
Brunnings et al., 2009; Marcinkowska et al., 2009). Do
note that regular barplots also add an unimportant dimension: the bars width. Hence Clevelands dotplots
(Cleveland, 1994) will usually do better (cf. Figures 8,
12 and 13).
Last but not least, make it a rule: there is no space
for piecharts in agricultural sciences. The piechart has
been criticized by many (Cleveland, 1994; Tufte, 2001;
Sarkar, 2008) - which does not mean it has not been defended (e.g. Spence and Lewandowski, 1990; Friendly,
1994). Nevertheless, for regular agricultural data
piecharts should not be used, and this should be considered a rule with no exception. In every single case there
will be a chart that works at least equally well as the
piechart, and in the abundance of situations it will work
much better, let it be a barplot, a dotplot and the like.
In summary, pay attention to every single element
of the graph, checking its size, color, position and cooperation with other elements. After one has prepared
quite a number of graphs, this will come unconsciously.
But before this time comes, one might wish to follow
the rules given in this paper.

Additional examples
Several examples of good graphs have been provided
(Figures 2, 4, 7, 10, 12, 13, 14, 15; right panels of Figures

Figure 14 A trellis display portraying relationships of fruitfly longevity and thorax length in different experimental groups
(representing various sexual activities), with a regression lines superimposed (data set fruitfly). The regression lines were
obtained through generalized least squares with a power variance function to deal with heterogeneity of residuals, and
the final model included a common slope for thorax length intercept and varying intercept across activities. This is a
GOOD GRAPH.
Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

492

5 and 11; and middle and bottom panels of Figure 6).

Here, three figures are discussed (two of which have already been referred to) in detail.
Example 1
Figure 10 shows an association between length and
width of petals of three Iris species. The three species
are plotted in the same panel, so three colors are used
to distinguish them; a legend is needed, then. It was
placed next to the graph, not in the figure caption: this
facilitates quick recognition which species is represented
by which color. Note that the species in the legend are
presented in the same order as in the graph: I. virginica is
in the top position, while I. setosa in the bottom one. This
very subtle thing facilitates reading the graph. Jitter helps
notice overlapping points. The box consists of four scale
lines with tick marks pointing outwards, the top and right
axes having tick marks without labels. The untypical
shape of the box (narrow and high) is due to the isometric scales: petal length and width are presented in the same
and comparable units (cm), and one cm has the same
length at both scales. It works well by facilitating comparison of the two traits: we can immediately see that
petal length is much more variable than petal width, information that would not be so obvious without the isometric scales. So do not be afraid of untypical aspect ratios of graphs if they are helpful indeed.
Example 2
Figure 15 presents a comparison of within-species
distributions of petal length of the three Iris species. It
has all properties of a good graph we have discussed
above: color is used to differentiate the groups (varying
line and point types might work too); the full box is used,
with tick marks on each scale and tick mark labels on

Figure 15 Kernel density of petal length of three Iris species

(data set iris). It is considered a GOOD GRAPH.

Kozak

left and bottom ones; legend is added to introduce the

colors corresponding to the species. But note also two
additional interesting elements of this graph. This time
the legend is placed above the graph, and the order of
the species there is the same as in the graph itself: I.
setosa, with the shortest petals, is to the left, while I.
virginica, with the longest petals, is to the right of the
legend; this small aspect helps read the graph. Second,
the so-called rug is added (Chambers and Hastie, 1992),
which shows individual points - this is useful for example for outlier detection. The rug is jittered.
Example 3
The last example will be more biological than agricultural, describing association between fruitfly male
longevity and thorax length, depending upon sexual activity. First we will model the relationship, and then we
will visualize the results.
A linear model includes longevity as the dependent
variable and thorax length and sexual activity as predictor variables, the former being continuous and the latter categorical. So, it is an analysis of covariance model.
Because thorax length ranged between 0.64 and 0.94,
in the modeling 0.80 was subtracted from the variable
to allow sensible inference for intercepts for activities.
After fitting a model with varying intercepts and slopes
it appeared that the residuals were heterogeneous, so this
needed to be taken into account. The plot of standardized residuals versus fitted values suggested a power variance function structure, which indeed worked well after estimation with generalized least squares (Pinheiro
and Bates, 2000). However, the slopes seemed quite similar, and the Akaike Information Criterion showed that
the model with the common slope and varying intercepts showed the best fit; the power variance function
still worked best to deal with heterogeneity of residuals. In this model both sexual activity and thorax length
affected longevity of the fruitfly male.
Figure 14 portrays the data and the model fitted, with
the help of the trellis display framework. We can see
how powerful such a display can be when there are too
many groups to be presented within one graph. (Here,
124 points and 5 lines within one panel would cause clutter, while the trellis display effectively compares the
sexual activity groups.) All elements in the plot follow
the rules given above and in Tables 1 and 2. Notice some
additional aspects: the panels are ordered by the increasing intercept: the bottom left represents the high group
(a fly kept with eight virgin fruitflies), which has the
lowest intercept, and the upper middle panel represents
the many group (a fly kept with eight pregnant
fruitflies), which has the highest intercept. Mild jittering was employed for thorax length to deal with overlap of points, and grid lines facilitate panel-to-panel
comparisons. Note that although the model was fitted
for the thorax length subtracted by 0.80, this is not seen
in the graph - the tick mark labels were simply changed
from (0.1, 0.0, 0.1) to (0.7, 0.8, 0.9).

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

493

Basic principles of graphing data

Conclusion
Graphing data is not a simple matter. In fact, this
point of view is just a small portion of the knowledge of visualizing scientific data, but it is believed to
be sufficient for authors constructing simple graphs for
their publications. Whole books have been devoted to
the topic, best-known ones being the classics by Tufte
(1991; 1997; 2001 and 2006), Cleveland (1993; 1994) and
Wilkinson (2005). Jacoby (1997; 1998) also presents
some interesting ideas, while Harris (1999) is a useful
reference of information graphics.
Most principles presented in this paper are so basic that they apply to most graph types. However, be
aware that there are other types of graphs that offer
various possibilities of visualizing data. One should always choose that graph which is best for each data:
best in terms of data presentation and interpretation,
but also easiness of reading. When a type of graph is
chosen, all its details should be carefully selected so
that the graph conveys the message its author planned
it to.

Appendix
Data sets used in the examples. All these data sets come
from R (R Development Core Team, 2009), from the data
sets of the same names.
ChickWeight
The data from a repeated-measure experiment on the effect of diet on early growth of chickens (Crowder and
Hand, 1990).
fruitfly
The data on longevity and sexual activity of male
fruitflies, originating from Partridge and Farquhar
(1981), downloaded from the faraway package (Faraway,
2009) of R. See also the paper by Hanley and Shapiro
(1994), which discusses interesting aspects of this data
set and its usefulness in teaching statistics.
haynes
The data on phenotypic stability of resistance to late
blight of potato clones; the variable considered is area
under the disease progress curve (AUDPC) (Haynes et
al., 1998). Data were taken from the agricolae package
(Mendiburu, 2010) of R.
InsectsSprays
The data originating from Beall (1942). They include a
number of insects counted after application of six insecticides, in 12 replications.
iris
A famous iris data set (Anderson 1935; Fisher, 1936), providing the measurements (in cm) of sepal length and
width and petal length and width for 50 flowers from
each of three species of iris: Iris setosa, I. versicolor and I.
virginica.

trees
The data on measurements of the girth, height and volume of timber in 31 felled black cherry trees, provided
by Ryan et al. (1976).
Software used for graphing and data analysis. The statistical analysis for Example 3 was performed with linear
mixed-effects modeling framework offered by the nlme
package (Pinheiro and Bates, 2000) of R (R Development
Core Team, 2009). The graphs were constructed in the
graphics (Murrell, 2005) and lattice packages of R
(Sarkar, 2008).

References
Abou Khalifa, A.A.B. 2009. Physiological evaluation of some hybrid
rice varieties under different sowing dates. Australian Journal
of Crop Science 3: 178-183.
Ambang, Z.; Ndongo, B.; Amayana, D.; Djil, B.; Ngoh, J.B.;
Chewachong, G.M. 2009. Combined effect of host plant resistance
and insecticide application on the development of cowpea viral
diseases. Australian Journal of Crop Science 3: 167-172.
Ambrosano, E.J.; Trivelin, P.C.O.; Cantarella, H.; Ambrosano,
G.M.B.; Schammass, E.A.; Muraoka, T.; Guirado, N.; rossi, F.
2009. Nitrogen supply to corn from sunn hemp and velvet bean
green manures. Scientia Agricola 66: 386-394.
Anderson, E. 1935. The irises of the Gaspe Peninsula. Bulletin of
the American Iris Society 59: 2-5.
Beall, G. 1942. The transformation of data from entomological
field experiments. Biometrika 29: 243262.
Brunnings, A.M.; Datnoff, L.E.; Ma, J.F.; Mitani, N.; Nagamura,
Y.; Rathinasabapathi, B.; Kirst, M. 2009. Differential gene
expression of rice in response to silicon and rice blast fungus
Magnaporthe oryzae. Annals of Applied Biology 155: 161-170.
Carr, D.; Littlefield, R.J.; Nicholson, W.L.; J.S., Littlefield. 1987.
Scatterplot matrix techniques for large N. Journal of the
American Statistical Association 82: 424-436.
Cleveland, W.S. 1993. Visualizing Data. Hobart Press, Summit,
NJ, USA. 360 p.
Cleveland, W.S. 1994. The Elements of Graphing Data. 2ed. Hobart
Press, Summit, NJ, USA. 323 p.
Cleveland, W.S.; Fuentes, M. 1997. Trellis Display: Modeling Data
from Designed Experiments; Technical Report. BellLabs, Paris,
France.
Cleveland, W.S.; McGill, R. 1884. The Many Faces of a Scatterplot.
Journal of the American Statistical Association. 79: 807-822.
Chambers, J.M.; Hastie, T.J. 1992. Statistical Models in S. Brooks/
Cole Wadsworth, Florence, KY. 608 p.
Crowder, M.; Hand, D. 1990. Analysis of Repeated Measures. Chapman
and Hall, NY, USA. 257 p.
Cumming, G.; Finch, S. 2005. Inference by eye: confidence intervals
and how to read pictures of data. American Psychologist 60:
170-180.
Cumming, G.; Williams, J.; Fidler, F. 2004. Replication, and
researchers understanding of condence intervals and standard
error bars. Understanding Statistics 3: 299-311.
Czubaszek, A. 2009. The effects of genotype and environment on
selected traits of oat grain and flour. Plant Breeding and Seed
Science 60: 45-60.
Dupont, W.D.; Plummer, W.D.J. 2003. Density distribution
sunower plots. Journal of Statistical Software 8: 1-5.
Ehrenberg, A.S.C. 1977. Rudiments of numeracy. Journal of the
Royal Statistical Society. Serie. A. 140: 277-297.
EL-Shafey, N.M.; Hassaneen, R.A.; Gabr, M.M.A.; EL-Sheihy, O.
2009. Pre-exposure to gamma rays alleviates the harmful effect
of drought on the embryo-derived rice calli. Australian Journal
of Crop Science 3: 268-277.

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010

494

Faraway, J. 2009. Faraway: functions and datasets for books;. R

package
version
1.0.4.
Available
at:
http://
www.maths.bath.ac.uk/~jjf23/ [Accessed Dec. 18, 2009]
Fiorio, P.R.; Dematte, J.A.M. 2009. Orbital and laboratory spectral
data to optimize soil analysis. Scientia Agricola 66: 250-257.
Fisher, R.A. 1936. The use of multiple measurements in taxonomic
problems. Annals of Eugenics 7: 179-188.
Friendly, M. 1994. A Fourfold Display for 2 by 2 by K Tables York
University, Psychology Department, Toronto, Canada.
(Technical Repport 217).
Giacomini, A.A.; Da Silva, S.C.; Sarmento, D.O.L; Zeferino, C.V.;
Da Trindade, J.K.; Souza Jnior, S.J.; Guarda, V.A.; Sbrissia,
A.F.; Nascimento Jnior, D. 2009. Components of the leaf area
index of marandu palisadegrass swards subjected to strategies
of intermittent stocking. Scientia Agricola 66: 721-732.
Hanley, J.A.; Shapiro, S.H. 1994. Sexual activity and life span of
male fruit flies: a data set that gets attention. Journal of Statistical
Education 2.
Harris, R.L. 1999. Information Graphics: A Comprehensive Illustrated
Reference. Oxford University Press, Oxford, UK. , 448 p.
Haynes, K.G.; Lambert, D.H.; Christ, B.J.; Weingartner, D.P.;
Douches, D.S.; Backlund, J.E.; Fry, W.; Stevenson, W. 1998.
Phenotypic stability of resistance to late blight in potato clones
evaluated at eight sites in the United States. American Journal
of Potato Research 75: 211-217.
Huff, D. 1954. How to lie with statistics. WW Norton, New York,
MY, USA. 142p.
Hydman, R.J. 1996. Computing and graphing highest density
regions. The American Statistician 50: 120-126.
Jacoby, W.G. 1997. Statistical Graphics for Univariate and Bivariate
Data: Statistical Graphics. Sage, Thousand Oaks, CA, USA. 97 p.
Jacoby, W.G. 1998. Statistical Graphics for Visualizing Multivariate
Data. Sage, Thousand Oaks, CA, USA. 103 p.
Jaradat, A.A. 2007. Predictive grain yield models based on canopy
structure and structural plasticity. Communications in
Biometry and Crop Science 2: 7489.
Jaradat, A.A. 2009. Modeling biomass allocation and grain yield in
bread and durum wheat under abiotic stress. Australian Journal
of Crop Science 3: 237-248.
Jonsson, C.N.; Aoyama, H. 2009. Extraction, partial
characterization and Susceptibility to Hg2+ of acid phosphatase
from the microalgae Pseudokirchneriella subcapitata. Scientia
Agricola 66: 634-642.
Karlidag, H.; Yildirim, E.; Turan, M. 2009. Salicylic acid ameliorates
the adverse effect of salt stress on strawberry. Scientia Agricola
66: 180-187.
Kozak, M. 2009a. Analyzing one-way experiments: a piece of cake
or a pain in the neck? Scientia Agricola 66: 556-562.
Kozak, M. 2009b. Text-table: an undervalued and underused tool
for communicating information. European Science Editing 35:
103-105b.
Lammel, D.R.; Brancalion, P.H.S.; Dias, C.T.S.; Cardoso, E.J.B.N.
2007. Rhizobia and other legume nodule bacteria richness in
brazilian Araucaria angustifolia forest. Scientia Agricola 64: 400408.
Macedo, O.J.; Barbin, D.; Mourao, G.B. 2009. Genetic parameters
for post weaning growth of Nellore cattle using polinomyals
and trigonometric functions in random regression models.
Scientia Agricola 66: 522-528.
Maci-Vicente, J.G.; Rosso, L.C.; Ciancio, A.; Jansson, H.B; LopezLorca, L.V. 2009. Colonisation of barley roots by endophytic
Fusarium equiseti and Pochonia chlamydosporia: Effects on plant
growth and disease. Annals of Applied Biology 155: 391-401.
Marcinkowska, J.; Boros, L.; Wawel, A. 2009. Response of pea (Pisum
sativum L.) cultivars and lines to seed infection by Ascochyta
blight fungi. Plant Breeding and Seed Science 59: 75-86.
Mendiburu, F. Agricolae: Statistical Procedures for Agricultural
Research. R: 2010. Package Version 1.0-9. Available at: http://
CRAN.R-project.org/package=agricolae [Accessed Dec. 18,
2009]

Kozak
Moussa, H.R.; Abdel-Aziz, S.M. 2008. Comparative response of
drought tolerant and drought sensitive maize genotypes to water
stress. Australian Journal of Crop Science 1: 31-36. Murrell, P.
2005. R Graphics. Chapman Hall, New York, NY, USA. 328 p.
Ofosu-Anim, J.; Leitch, M. 2009. Relative efficacy of organic
manures in spring barley (Hordeum vulgare L.) production.
Australian Journal of Crop Science 3: 13-19.
Pahlavani, M.H.; Miri, A.A.; Kaziem, G. 2009. Response of oil and
protein content to seed size in cotton(Gossypium hirsutum L.,
cv. Sahel). Plant Breeding and Seed Science 59: 53-64.
Pinheiro, J.P.; Bates, D.M. 2000. Mixed-effects models in S and SPlus. Springer, New York, NY,USA. 528 p.
Partridge, L.; Farquhar, M. 1981. Sexual activity and the lifespan of
male fruitflies. Nature 294: 580-581.
R Development Core Team. 2009. R: A language and environment
for statistical computing. R Foundation for Statistical
Computing, Vienna, Austria
Reynolds, M.; Manes, Y.; Izanloo, A.; Langridge, P. 2009.
Phenotyping approaches for physiological breeding and gene
discovery in wheat. Annals of Applied Biology 155: 309-320.
Ryan, T.A.; Joiner, B.L.; Ryan, B.F. 1976. The Minitab Student
Handbook. Duxbury Press, Pacific Grove, CA, USA.
Sarkar, D. Lattice Multivariate Data Visualization with R. 2008.
Springer, New York, NY,USA. 265 p.
Scarpari, M.S.; Beauclair, E.G.F. 2009. Physiological model to
estimate the maturity of sugarcane. Scientia Agricola 66: 622-628.
Silva, S.C.; Oliveira Bueno, A.A.; Carnevalli, R.A.; Uebele, M.C.;
Bueno, F.O.; Hodgson, J.; MAtthew, C.; Arnold, G.C.; M,
J.P.G. 2009a. Sward structural characteristics and herbage
accumulation of Panicum maximum cv. Mombaa subjected to
rotational stocking managements. Scientia Agricola 66: 8-19
Silva, M.M.; Libardi, P.L.; Fernandes, F.C.S.. 2009c. Nitrogen doses
and water balance components at phenological stages of corn.
Scientia Agricola 66: 512-521.
Silva, R.B.T.R.; Ns, I.A.; Moura, D.J. 2009b. Broiler and swine
production: animal welfare legislation scenario. Scientia Agricola
66: 713-720
Simoes, M.S.; Rocha, J.V.; Lamparell, R.A.C. 2009. Orbital spectral
variables, growth analysis and sugarcane yield. Scientia Agricola
66: 451-461.
Spence, I.; Lewandowsky, S. 1990. Graphical perception. p. 13
57, In: Fox ,J.; Long, J.S., eds. Modern methods of data analysis,
Sage Thousand Oaks, CA, USA.
Takahashi, D.; ditt, R.F.; Lambais, M.R. 2009. Cloning of putative
ureG genes from Glomus intraradices and urease activities in tobacco
arbuscular mycorrhizal roots. Scientia Agricola 66: 258-266.
Tufte, E.R. 2001. The Visual Display of Quantitative Information.
2ed. Graphics Press, Cheshire, CT, USA. 199 p.
Tufte, E.R. 1991. Envisioning Information. Graphics Press, Cheshire,
CT, USA. 126 p.
Tufte, E.R. 1997. Visual Explanations. Graphics Press, Cheshire,
CT, USA. 157 p.
Tufte, E.R. 2006. Beautiful Evidence. Graphics Press, Cheshire,
CT, USA., 213 p.
Vzquez, E.V.; Vieira, S.R.; De Maria, I.C.; Gonzlez, A.P. 2009.
Geostatistical analysis of microrelief of an Oxisol as a function
of tillage and cumulative rainfall. Scientia Agricola. 66: 225-232.
Virdis, A.; Motzo, R.; Giunta, F. 2009. Key phonological events in
globe artichoke (Cynara cardunculus var. scolymus) development.
Annals of Applied Biology 155: 419-429.
Wiewira, B. 2009. Long-time storage effect on the seed health of
spring barley grains. Plant Breeding and Seed Science 59: 3-12.
Wilkinson, L. 2005. The Grammar of Graphics. 2ed. SpringerVerlag, New York, NY, USA. 690 p.

Received December 19, 2009

Accepted January 05, 2010

Sci. Agric. (Piracicaba, Braz.), v.67, n.4, p.483-494, July/August 2010