Assaad et al.

SpringerPlus (2015) 4:33

DOI 10.1186/s40064-015-0795-z
a SpringerOpen Journal

SOFTWARE Open Access

Rapid publication-ready MS-Word tables for

two-way ANOVA
Houssein I Assaad1,2*, Yongqing Hou3, Lan Zhou1, Raymond J Carroll1 and Guoyao Wu2,3*

Abstract
Background: Statistical tables are an essential component of scientific papers and reports in biomedical and
agricultural sciences. Measurements in these tables are summarized as mean ± SEM for each treatment group.
Results from pairwise-comparison tests are often included using letter displays, in which treatment means that
are not significantly different, are followed by a common letter. However, the traditional manual processes for
computation and presentation of statistically significant outcomes in MS Word tables using a letter-based algorithm
are tedious and prone to errors.
Results: Using the R package ‘Shiny’, we present a web-based program freely available online, at https://houssein-
assaad.shinyapps.io/TwoWayANOVA/. No download is required. The program is capable of rapidly generating
publication-ready tables containing two-way analysis of variance (ANOVA) results. Additionally, the software can
perform multiple comparisons of means using the Duncan, Student-Newman-Keuls, Tukey Kramer, Westfall, and
Fisher’s least significant difference (LSD) tests. If the LSD test is selected, multiple methods (e.g., Bonferroni and
Holm) are available for adjusting p-values. Significance statements resulting from all pairwise comparisons are
included in the table using the popular letter display algorithm. With the application of our software, the
procedures of ANOVA can be completed within seconds using a web-browser, preferably Mozilla Firefox or Google
Chrome, and a few mouse clicks. To our awareness, none of the currently available commercial (e.g., Stata, SPSS
and SAS) or open-source software (e.g., R and Python) can perform such a rapid task without advanced knowledge
of the corresponding programming language.
Conclusions: The new and user-friendly program described in this paper should help scientists perform statistical
analysis and rapidly generate publication-ready MS-Word tables for two-way ANOVA. Our software is expected to
facilitate research in agriculture, biomedicine, and other fields of life sciences.
Keywords: Two-way ANOVA; Multiple comparisons; Online software; Statistical analysis; Biology; Agriculture; R; Shiny

Introduction 2014a, b; Meininger and Wu 1997; Wu 1997). An example

Since the publication of our paper on rapid generation of . of such tables is shown in Table 1, reporting the effects of
rtf tables from one-way analysis of variance (ANOVA) diet [low fat (LF) or high fat (HF)] and body weight classi-
(Assaad et al. 2014a), we have received multiple requests fication (lean or overweight) on concentrations of amino
to construct a similar software package that can handle acids in the plasma of Sprague-Dawley rats. The table is
data arising from two-way ANOVA designs. These statis- entirely generated by our program, including the default
tical tables are common in a variety of scientific applica- caption at the bottom. Other table formats are also offered
tions, including agricultural, biological, and biomedical by the software and will be described in details in the sub-
studies (Steel et al. 1997; Jobgen et al. 2009a; Wang et al. sequent sections (see Table 2, for instance).
The present work focuses on generating publication-
ready tables from two-way ANOVA models where mea-
* Correspondence: [email protected]; [email protected]
1
Department of Statistics, Texas A & M University, College Station, TX 77843, surements are summarized as mean ± SEM for each
USA treatment group. Researchers will have an option to in-
2
Department of Animal Science, Texas A & M University, College Station, TX clude post-hoc test results in these tables using a letter-
77843, USA
Full list of author information is available at the end of the article based algorithm (Piepho 2004) to indicate which treatment

© 2015 Assaad et al.; licensee Springer. This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction
in any medium, provided the original work is properly credited.
Assaad et al. SpringerPlus (2015) 4:33 Page 2 of 9

Table 1 Effects of diet and weight classification on amino acid concentrations (nmol/ml) in the rat plasma
HF LF P-value
Variable
Lean Overweight Lean Overweight Diet Weight D × W1
Asp 45.3 ± 2.39 47.6 ± 2.97 46.2 ± 2.96 45.3 ± 3.45 0.824 0.824 0.604
Glu 88.4 ± 3.23 88.6 ± 1.16 87.8 ± 2.58 87.6 ± 3.18 0.757 0.984 0.951
Asn 128 ± 5.71 114 ± 10.4 120 ± 5.62 133 ± 6.58 0.477 0.922 0.068
Ser 359 ± 10.3a 294 ± 4.39b 353 ± 7.43a 292 ± 3.76b 0.576 <0.001 0.807
Gln 562 ± 18.3b 645 ± 11.1a 559 ± 9.43b 655 ± 19.6a 0.801 <0.001 0.676
His 124 ± 3.18 115 ± 2.95 120 ± 4.08 130 ± 5.89 0.236 0.906 0.028
Gly 392 ± 6.91a 305 ± 7.3b 384 ± 7.32a 297 ± 5.68b 0.254 <0.001 0.955
Thr 379 ± 7.42 359 ± 11.6 381 ± 8.62 376 ± 12.6 0.365 0.223 0.481
Cit 79.9 ± 2.66a 53.8 ± 1.63c 69.6 ± 3.37b 73.9 ± 1.84ab 0.057 <0.001 <0.001
b c a c
Arg 251 ± 8.09 197 ± 6.4 279 ± 7.46 219 ± 4.46 0.001 <0.001 0.642
β-Ala 13.6 ± 0.884b 31.4 ± 1.52a 13.1 ± 0.611b 26.8 ± 1.91a 0.064 <0.001 0.123
a c a b
Taurine 648 ± 17.2 469 ± 13.1 670 ± 11.9 572 ± 12 <0.001 <0.001 0.006
Ala 471 ± 10.2b 403 ± 13.6c 492 ± 7.54b 585 ± 9.22a <0.001 0.252 <0.001
Tyr 131 ± 4.25 135 ± 4.04 125 ± 5.7 139 ± 5.32 0.865 0.069 0.268
Trp 110 ± 3.28 114 ± 4.78 115 ± 2.61 113 ± 3.4 0.658 0.703 0.42
Met 108 ± 5.39 112 ± 3.65 107 ± 4.83 110 ± 4.8 0.77 0.521 0.991
Val 253 ± 9.61c 331 ± 6.74a 249 ± 10.1c 289 ± 9.63b 0.017 <0.001 0.048
Phe 111 ± 3.7 108 ± 3.84 106 ± 3.81 113 ± 3.44 0.988 0.645 0.211
Ile 152 ± 4.19b 206 ± 4.01a 144 ± 4.77b 195 ± 3.68a 0.035 <0.001 0.674
b a b a
Leu 212 ± 4.56 308 ± 7.56 216 ± 6.34 302 ± 5.79 0.83 <0.001 0.465
Orn 68.8 ± 1.58b 81.4 ± 1.51a 67.8 ± 2.06b 83.3 ± 1.25a 0.786 <0.001 0.381
b a b a
Pro 277 ± 7.04 342 ± 8.59 286 ± 5.32 334 ± 4.52 0.96 <0.001 0.196
Cys 165 ± 4.53b 197 ± 7a 157 ± 3.2b 194 ± 6.58a 0.279 <0.001 0.648
c a bc ab
Lys 252 ± 8.82 291 ± 5.17 259 ± 8.32 282 ± 5.42 0.957 <0.001 0.273
Values are means ± SEM, n = 9 per treatment group. Male Sprague-Dawley rats (Charles River Laboratories) were fed a low-fat (LF) or high-fat (HF) diet between 4
and 13 weeks of age, as described by Jobgen et al. (2009a, b). At 13 weeks of age, five hours after the last feeding, blood samples were obtained from the tail vein
of box-restrained conscious rats using a microhematocrit (Wu 1995). The plasma was analyzed for amino acids using high-performance liquid chromatography
(Rezaei et al. 2013; Wu and Meininger 2008). Classification of rats as lean or overweight was performed using the Cluster analysis of body weights, as described by
Assaad et al. (2014b).
a-c
Means in a row without a common superscript letter differ (P < 0.05) as analyzed by two-way ANOVA and the TUKEY test. 1D × W = Diet × Weight interaction effect.

groups are significantly different. With this algorithm, of aspartic acid (Asp) in the plasma. We believe that
treatment means that are not significantly different, as our new software will save biologists, and other scien-
reported by an all-pairwise comparison procedure, are tists in general, an ample amount of time by avoiding
followed by a common superscript letter, e.g., a, b and c the manual addition of the superscript letters (see
(see Table 1). In other terms, two treatments without a Table 1) derived from the appropriate statistical tests.
common letter are statistically significant at the chosen This offers a distinct advantage over the traditional
level of significance α (e.g., α = 0.05 or 0.01). For ex- manual processes for computation and presentation of
ample, the effects of the HF-Lean and HF-Overweight results in tables that are not only tedious but are also
treatments on the concentration of lysine in the plasma prone to errors.
(see last row of Table 1) are significantly different as A Google search of the words “Online two way ANOVA
they do not share a common superscript. On the other calculator” reveals several online toolsa that are capable of
hand, LF-Lean and LF-Overweight have the same effect performing two-way ANOVA. Despite their user-friendly
on lysine since they share the superscript ‘b’. By con- interface, these programs have serious limitations. Particu-
vention, when all treatments have the same effect on a larly, most of them cannot carry out post-hoc testing of
certain response variable, no superscripting is used. any kind. Some can conduct the Tukey test but do not
This is the case in the first row of Table 1, where all translate the testing results into a compact letter display;
treatments have the same effect on the concentration hence the user will have to do the translation by hand.
Assaad et al. SpringerPlus (2015) 4:33 Page 3 of 9

Table 2 Summary of multiple comparison methods software should be used only when these assumptions
LSD Highest error rate and power of any method. In general, it are nearly satisfied. We also illustrate the functionality of
controls the FWER in the weak sense; when there are 3 the software via a step-by-step approach using different
treatment groups, the FWER is controlled in the strong sense.
toy datasets to cover most table designs encountered in
DC Error-rate and power intermediate between SNK and LSD. research papers. The toy datasets are available on the
Controls the FWER in the weak sense.
software webpage and can be downloaded from there.
SNK Error-rate and power intermediate between TK and DC.
Controls the FWER in the weak sense.
Various tips and concluding remarks are given towards
the end of this article.
TK SSP, Lowest error rate and power*, controls the FWER in the
strong-sense
Background and materials
BF SSP, Controls the FWER in the strong sense, but it is too
conservative (reduces the number of true positives)
Holm SWP, Stepwise extension of BF; hence, it is more powerful. It
1 Two-way ANOVA
should always be preferred over BF; controls the FWER in the
strong sense. It doesn’t take logical constraints or correlations The main purpose of this section is to present a brief
into account.
non-technical description of two-way ANOVA and
Westfall More powerful than any MCP controlling the FWER in the introduce the statistical terms that will be used through-
strong sense. However, it is computationally expensive.
out the rest of this paper. The reader should refer to
The table was adapted from Christensen (2011) with modifications.
BF = Bonferroni; DC: Duncan method; LSD = Least significant difference;
standard experimental design textbooks for a more in-
SNK = Student-Newman-Keuls; TK = Tukey Kramer (or Tukey HSD in balanced depth treatment of the subject (Kutner et al. 2005 and
designs); SSP = Single-step procedure; SWP = step-wise procedure. Montgomery 2012). Two-way ANOVA, also known as
*
When compared with the classical LSD, SNK, DC.
two-factor ANOVA, is concerned with the investigation
of the simultaneous effects of two nominal variables, say
While these online tools may be suitable for pedagogical A and B, called factors. These factors can take different
purposes, their major drawback remains in their inability values known as levels. Each combination of a factor
to export results to an RTF reader in a publication-ready level of A and a factor level of B is a treatment. For in-
format similar to that of Table 1, making their usage in re- stance, in Table 1, there are two factors, diet and body
search impractical and thus unlikely. Also, several soft- weight classification. Factors, diet and body weight, have
ware packages (e.g. R, SAS, Stata, SPSS, JMP, etc.) can two levels each, LF and HF for the diet and Lean and
conduct two-way ANOVA, followed by post-hoc analysis. Overweight for body weight. This leads to four treatments:
To our knowledge, none of them is capable of exporting LF-Lean, LF-Overweight, HF-Lean, and HF-Overweight.
the multiple comparisons results to an RTF reader in a In general, if factor A has a levels and factor B has b levels,
format similar to that of Table 1 without advanced know- the total number of treatments is ab. The variable under
ledge of the corresponding programming language. study is often referred to as response variable. In the same
When working on the software, we received consider- previous example, the amino acids (e.g., Asp, Ser, and
able and valuable assistance from several R (R Core Team Gln), are all response variables. The effect of a factor is
2014) packages. We would like to acknowledge the for- defined to be the change in response resulting from a
midable efforts on the part of the developers of the follow- change in the level of the factor. This is frequently called a
ing packages: grifExtra (Auguie 2012), XLConnect (Mirai main effect. In some experiments, we may observe that
Solutions 2014), agricolae (Mendiburu 2014), rtf (Schaffer the difference in response between the levels of factor A is
2013), multcomp (Hothorn et al. 2008), plyr (Wickham not the same at all levels of factor B. When this occurs,
2011), ggplot2 (Wickham 2009) and shiny (RStudio 2013). there is interaction between the two factors. There are
Without the availability of these R packages, this software three hypothesis of interest in two-way designs, namely,
would not have been developed. the significance of the main effects of factors A and B, as
In the remaining sections, we present necessary back- well as their interaction. The p-values of these tests are re-
ground materials for two-way ANOVA, followed by a ported in the last three columns of Table 1. Two-way
brief summary of multiple comparison techniques. We ANOVA assume that all observations are independent
do not attempt to provide a full description of all testing from each other. Also, measurements corresponding to a
procedures that have been presented in classical text- treatment group arise from a population having a normal
books on experimental designs. Instead, our main goal is distribution with possibly different means but the same
to highlight some of the limitations of the statistical tests variance across all treatment groups. When interaction is
included in the software and help the researcher decide statistically significant (P < 0.05), all pairwise-comparisons
on a test that is more suitable for his/her data. We are usually carried out on the treatment means. The latter
would also like to underline the necessary assumptions may still be of interest and pairwise-comparisons between
required by two-way ANOVA and to emphasize that the treatment means can be made even when the two factors
Assaad et al. SpringerPlus (2015) 4:33 Page 4 of 9

do not interact (Wei et al. 2012). For instance, if diet and Power can be improved by extending single-stepc MCP
weight classification do not interact, it may still be import- into stepwise procedures via the closure method (Westfall
ant to determine whether being lean and having a HF in- et al. 2011). For example, the stepwise Holm procedure
take has the same effect on AA concentrations in the (see step 4 in the next section) is an extension of the
plasma as the LF diet in overweight rats. Our program of- single-step BF test. By construction, stepwise procedures
fers a variety of statistical tests to perform these pairwise are more powerful and control the FWER in the strong
comparisons using the cell mean modelb (Kutner et al. sense. MCP with power higher than the Holm procedure
2005). It also gives the option to create the summary table are available when there are logical restrictionsd among
without the post-hoc analysis if it is not of interest to the the hypotheses as is the case of all pairwise comparisons.
researcher. In this case, the output will be the same as Westfall (1997) extended the Holm’s procedure by incorp-
Table 1 but without the superscript letters. orating logical restrictions and accounting for random
correlations between the hypotheses being tested. Because
2 Multiple comparisons methods the method uses extensive simulation to calculate p-
values, computation usually requires more time than other
Multiple testing problems arise frequently in biome- MCP. This discussion on multiple comparison methods is
dical and agricultural research (Hou et al. 2015; Wang summarized in Table 2.
et al. 2015a, 2015b), and it is important to address them
appropriately. In this section, we merely scratch the sur-
face of the complex topic of multiple hypotheses testing; The software
the interested reader may find the books by Westfall The software (see Figure 1) is available at https://
et al. (2011) and Bretz et al. (2010) extremely helpful. houssein-assaad.shinyapps.io/TwoWayANOVA/. It can
These books offer the most up-to-date coverage of the handle balanced and unbalanced designs. One compli-
subject and provide a plenty of SAS and R code to help cation encountered in unbalanced two-way designs is
the researcher implement these methods. Hypothesis that p-value computation depends on the order in which
testing involves two types of errors. A type I error (also factors appear in the dataset. In this case, Everitt and
called false positive) occurs when we declare an effect Hothorn (2010) suggest to present two tables corre-
when none exists. Similarly, a type II error (false nega- sponding to the different order of appearance of the two
tive) occurs if we fail to detect a truly existing effect. factors in the data set. It is worth mentioning that our
Multiple testing refers to testing more than one hypo- software program will not work when there is only one
thesis in a particular study. Multiple testing procedures observation for each treatment. This is because two-way
are often designed to control the family-wise error rate ANOVA cannot be conducted unless we assume the two
(FWER) of incorrectly rejecting at least one hypothesis factors do not interact. Because scientists are interested
in a given group of tests. In other words, the FWER is not only in the main effects of two factors, but also in
the probability of committing at least one Type I error their interaction, we decided not to include this scenario
in multiple testing. The majority of the classical multiple in our program. The software will, however, display a
comparison procedures (MCP), such as DC, LSD and message alerting the user that two-way ANOVA will not
SNK, control the FWER in the weak sense, i.e. when be conducted in this case and a table will not be gener-
the p-values calculations are carried out under the as- ated. While the program will work in the presence of
sumption that all null hypotheses are true. In practice, missing values, it may generate inaccurate SEM values.
however, it is unlikely that this assumption will hold, This is because sample sizes, which appear in the de-
therefore allowing the FWER to exceed the usual 5% nominator of the SEM, are not the same across different
value. Thus, a stronger control for the FWER under less response variables when missing values are present. This
restrictive assumptions is needed. If, for a given MCP, issue could be fixed at the expense of sacrificing the sam-
the FWER is controlled under any partial configuration ple size display in the caption below each table. As we
of true and false null hypotheses, the error is controlled plan on constantly updating the software, a better solu-
in the strong sense. For instance, TK and BF control the tion to handle missing values should be incorporated in
FWER in the strong sense but suffer from a low power. the next version when released. Here, we describe, via a
Namely, TK and BF are more likely to declare true hy- step-by-step approach, the detailed functionality of our
potheses as being true, but might also fail to identify program. The software can handle two scenarios where
false hypotheses as being false. This trade-off between data should be arranged accordingly to obtain the cor-
power and FWER control is the hardest issue to deal rect output. For illustration purposes, data sets cor-
with in multiple comparisons. Ideally, it is desired to responding to each scenario can be downloaded from
pick a testing procedure that controls the FWER in the the software webpage under the ‘Data Files’ panel (see
strong sense, while aiming for the highest possible power. Figure 2). We distinguish the following settings:
Assaad et al. SpringerPlus (2015) 4:33 Page 5 of 9

Figure 1 A screenshot of the software for scenario (S1). The Excel file Plasma.xls can be downloaded from the “Data Files” panel. Because this
file contains a single dataset, the “Single dataset” option is selected (see step 6 above).

Figure 2 A screenshot of the software for scenario (S2). The file workbook.xlsx can be downloaded from the “Data Files” Panel. In the
“Choose a Data Format” drop-down menu, the option “Workbook (multiple sheets)” is selected since the file contains multiple sheets/datasets
(see step 6 above).
Assaad et al. SpringerPlus (2015) 4:33 Page 6 of 9

(S1) A single dataset in a single-sheet workbook group and one pooled SEM for all treatment groups
(see file Plasma.xls). (see Table 2). The detailed computation of the
(S2) Multiple data sets arranged within multiple Excel pooled SEM is provided in the next section.
sheets (one data per sheet) and saved in one Excel 6. Specify a data format: For (S1), select ‘Single dataset’,
workbook (see file workbook.xlsx). whereas for (S2), select ‘Workbook (multiple
sheets)’.
The user of our software should follow the steps below 7. Click on the ‘Get. rtf table’ to download the table
(see Figures 1 and 2): with all statistical results included. Interaction plots
(only available for the data set in the first sheet in an
1. Upload an excel workbook (both .xls and .xlsx Excel workbook currently) can also be downloaded
format are supported) and select the level of in .pdf format by clicking on the ‘Get .pdf
significance α. interaction plots’ button.
2. Indicate the number of significant digitse for the
means and SEM and the number of decimal places The publication-ready table should now open in the
for the p-values to be displayed in the table. By user’s default .rtf reader (e.g. MS Word). The table can
default, 3 significant digits and 3 decimal places are now be edited as desired (adding/removing columns,
used for means/SEM and p-values respectively. rows or borders, merging, and centering, etc.).
3. By default, the software will perform post-hoc
pairwise comparisons (PHC) of all treatment means Regarding the pooled SEM
and report the results in each table row. Thus, The main purpose of this section is to describe how the
superscripts are added to each treatment group cell, pooled SEM (PSEM) is computed in balanced and un-
describing significance or the lack of it. Select balanced designs. In general, pooled SEM should only
“ANOVA without PHC” if you wish to construct a be used when the design is balanced for reasons that
table without PHC in which case the superscripts will become clear in the definition below. However, we
will not be reported. Note also that, in this case, the have decided to report a “pooled SEM” for unbalanced
“choose a statistical test” drop-down list will designs if the researcher is seeking a more compact
disappear as PHC tests will not be conducted. table design (one pooled SEM column as opposed to
4. If “ANOVA with PHC” was selected in step 3, SEM in every treatment column, see Tables 1 and 3).
specify a statistical test to perform all pairwise When the design is balanced, i.e. when there are n sub-
comparisons. Currently available tests are Tukey’s jects assigned to every treatment, the pooled SEM for
HSD, Duncan, Student-Newman-Kleus (SNK), any pairwise comparison is the same and is computed
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Westfall, and the least significant difference (LSD)f. as follows: PSEM ¼ MSE  n2 where MSE is the mean
If you select the (LSD) test, multiple methods, such as
Bonferroni (BF) and Holm are available for adjusting square error. For unbalanced designs, the software will
p-values. These are actually the Bonferroni and compute a PSEM for each pairwise comparison and then
Holm tests described in standard experimental report the highest one. For example, the PSEM for com-
design books. paring treatment i with treatment j is calculated using the
5. Choose an output format for the table: Two formats following formula:
are widely used in the literature. By selecting ‘Per
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi
group SEM’, the table will report the mean and SEM
1 1
for each group (see Table 1). The ‘Pooled SEM’ PSEM ij ¼ MSE  þ
ni nj
option will only report the means for each treatment

Table 3 An excerpt from Table 1 illustrating the “Pooled SEM” table format
HF LF Pooled P-value
Variable
Lean Overweight Lean Overweight SEM Diet Weight D × W1
Asp 45.3 47.6 46.2 45.3 4.2 0.824 0.824 0.604
a b a b
Ser 359 294 353 292 9.87 0.576 <0.001 0.807
Gln 562b 645a 559b 655a 21.6 0.801 <0.001 0.676
b c b a
Ala 471 403 492 585 14.7 <0.001 0.252 <0.001
Met 108 112 107 110 6.66 0.77 0.521 0.991
a-c
Values are means and pooled SEM, n = 9 per treatment group. Means in a row without a common superscript letter differ (P < 0.05) as analyzed by two-way
ANOVA and the TUKEY test. 1D × W = Diet × Weight interaction effect.
Assaad et al. SpringerPlus (2015) 4:33 Page 7 of 9

Where ni and ni are the sample sizes for treatments level comes first in the generated table. For example,
i and j. This process is repeated for every pairwise- HF appears before LF and Lean before Overweight. If,
comparison and the maximum is reported. for some reason, there is a need to change the default
level ordering, the following trick is useful: In the data
Common issues and useful tricks set Excel file, add the number 1 at the beginning of the
Valid variable or factor level names consist of letters, level to be shown first, the number 2 at the beginning
numbers and the dot or underline characters (typically of the level to be shown next, and so on. For example,
used to replace spaces between words). All names should assume the user wants the level Overweight to be shown
start with a letter or the dot not followed by a number. before Lean in Table 1 (thus disobeying the alphabetical
For instance, names such as “.2Ala” are not valid and the order rule). Replace Overweight with 1Overweight and
software might modify it in order to properly function. Lean with 2Lean throughout the whole Excel fileg (See file
Spaces in variable names should be avoided (replace them Change_Level_order.xlsx). The software will generate a
with a dot or an underscore). Greek letters should be modification of Table 1 shown in Table 4. The user can
avoided as they are not rendered properly in the tables. now delete the added digits.
You can add them later on after the table is generated.
Also, if the length of a variable’s name in the dataset is lar- Concluding remarks
ger than 10 characters, which might be the rule rather This paper presented a free web-based program capable
than the exception in many cases in biological studies, the of generating publication-ready RTF tables for two-
software will abbreviate the variable’s name. This can lead way analysis. These tables are often prepared for writ-
to ambiguous or unpleasant terms. We, therefore, advise ing agricultural, biological and medical science papers.
researchers to subjectively assign descriptive abbreviations Significance statements resulting from an all-pairwise
for variables with long names before loading their dataset comparison procedure are indicated by the popular
into the software. superscript letters display, which also allows for the
ranking of the treatment means. The software can
Factor levels should be described using letters, not numbers handle an Excel workbook with multiple datasets saved
This is how the program distinguishes between numer- in multiple sheets, creating one table per dataset. Two
ical variables and categorical/factor variables. For in- of the most commonly used formats for tables (see
stance, in the sample dataset Plasma.xls, the factor ‘diet’ Tables 1 and 3) in biomedical journals are also sup-
with the two levels low-fat and high-fat should be coded, ported by our software. The user has full control over
for instance as ‘LF’ and ‘HF’, not 0s (for low-fat) and 1s the number of significant digits for the treatment
(for high-fat). The latter will lead to an error stating the means, SEM, and the number of decimal points for the
program was unable to find two factors in your dataset p-values shown in the table. In addition, the program
(since one of them is treated as a numerical variable), appends an automatic informative caption at the bot-
and thus cannot perform two-way ANOVA. tom of every table it generates. Due to its user-friendly
interface, the software spare researchers a considerable
Factors and Levels ordering in the table amount of time and eliminate errors introduced by hu-
By default, the software will use the order in which fac- man input. To summarize, this software provide freely
tors appear in the user’s data set to decide which factor available statistics tools to facilitate research in many
comes on top of the output table. For instance, in scientific fields. Future work may involve adding mul-
Table 1, the factor diet with levels HF and LF occupies tiple comparisons procedures that are more powerful
the top row, while the factor weight with levels Lean than Holm but less computationally expensive than
and Overweight comes next. This is because in the data Westfall. Also, we may adjust our code to handle two-
set excel file (Plasma.xlsx), diet appears before weight. way ANOVA in the presence of missing data; however,
Changing the order of appearance in the original data this might be at the expense of not reporting sample
set will be reflected in the output table. Also, the soft- sizes in the caption as they are no longer the same
ware will use alphabetical order to choose which factor across different response variables.

Table 4 A modification of Table 1 that shows how to alter the positions of the levels of the factor “Weight”
HF LF P-value
Variable
1Overweight 2Lean 1Overweight 2Lean Diet Weight D × W1
Asp 47.6 ± 2.97 45.3 ± 2.39 45.3 ± 3.45 46.2 ± 2.96 0.824 0.824 0.604
Ser 294 ± 4.39b 359 ± 10.3a 292 ± 3.76b 353 ± 7.43a 0.576 <0.001 0.807
a-b
Values are means ± SEM, n = 9 per treatment group. Means in a row without a common superscript letter differ (P < 0.05) as analyzed by two-way ANOVA and
the TUKEY test. 1D × W = Diet × Weight interaction effect.
Assaad et al. SpringerPlus (2015) 4:33 Page 8 of 9

Availability and requirements Abbreviations

ANOVA: Analysis of variance; SAS: Statistical analysis system; SD: Standard
deviation; SEM: Standard error of the mean; PSEM: Pooled standard error of
 Project name: Rapid publication-ready MS Word the mean; AA: Amino acids; LF: Low-fat; HF: High- fat.
tables for two-way ANOVA.
 Project home page: https://houssein-assaad. Competing interests
The authors declare that they have no competing interests.
shinyapps.io/TwoWayANOVA/
 Operating system(s): Platform independent. Authors’ contributions
 Programming language: R, HTML/CSS, RTF. HA and GW conceived this project. HA wrote the software R code. YH, LZ
 Other requirements: internet connection, Mozilla and RJC contributed to software development. HA, YH, LZ, RJC, and GW
wrote the manuscript. HA and GW had the primary responsibility for the final
Firefox, or Google Chrome. content. All authors read and approved the final manuscript.
 Any restriction to use by non-academics: None.
Acknowledgments
We thank Dr. Carmen Tekwe for her helpful comments on this paper. Dr. H.
Endnotes Assaad and Dr. R.J. Carroll were supported by a postdoctoral training grant
a
See http://faculty.vassar.edu/lowry/anova2x2.html, http:// (R25T-CA090301) and a research grant (R37-CA057030) from the National
scistatcalc.blogspot.com/2013/11/two-factor-anova-test- Cancer Institute, respectively. Dr. Y. Hou was supported by the National Basic
Research Program of China (no. 2012CB126305), the National Natural Science
calculator.html. Foundation of China (no. 31372319), the Hubei Provincial Research and
b
Assume there are two factors A and B with a and b Development Program (no. 2010BB023), and the Natural Science Foundation
levels, respectively. In the cell mean model, the two fac- of Hubei Province (no. 2013CFA097, 2013CFB325, 2012FFB04805 and
2011CDA131). Dr. G. Wu was supported by Agriculture and Food Research
tors are collapsed into one factor with ab levels (each Initiative Competitive Grants (no. 2011-67015-20028 and 2014-67015-21770) from
level correspond to a treatment combination) and com- the USDA National Institute of Food and Agriculture, a Hatch project from Texas A
putation is now carried out using One-way ANOVA. & M AgriLife Research (H-8200), and the Hubei Hundred Talent program.
c
When testing multiple hypotheses, a test procedure Author details
is called a single-step method if the rejection or non- 1
Department of Statistics, Texas A & M University, College Station, TX 77843,
rejection of a null hypothesis does not take the decision USA. 2Department of Animal Science, Texas A & M University, College
Station, TX 77843, USA. 3Hubei Key Laboratory of Animal Nutrition and Feed
of any other hypothesis into account, e.g. the BF and Science, Wuhan Polytechnic University, Wuhan 430023, China.
TK tests. On the other hand, step-wise methods differ
from single-step procedures in that the results of a Received: 8 December 2014 Accepted: 5 January 2015
given test depend upon the results of other tests, e.g.,
Holm. References
d
For instance, in a three group ANOVA, H12: M1 = M2, Assaad H, Zhou L, Carroll RJ, Wu G (2014a) Rapid publication-ready MS-Word
H13: M1 = M3, and H23: M2 ≠ M3 cannot be simultaneously tables for two-way ANOVA. SpringerPlus 3:474
Assaad H, Yao K, Tekwe CD, Feng S, Bazer FW, Zhou L, Carroll RJ, Meininger CJ,
true. Thus the hypotheses are logically restricted. Choos- Wu G (2014b) Analysis of energy expenditure in diet-induced obese rats.
ing a test that does not account for these logical con- Frontiers Biosci (Landmark Edition) 19:967–985
straints might lead to loss in power and problems with the Auguie B (2012) gridExtra: functions in Grid graphics. R package version 0.9.1.
http://CRAN.R-project.org/package=gridExtra
interpretation of the test results. Bretz F, Hothorn T, Westfall P (2010) Multiple comparisons using R. CRC Press,
e
For the sake of completeness, we summarize the basic Boca Raton, FL
rules for significant digits. Christensen R (2011) Plane answers to complex questions, the theory of linear
models, 4th edition. Springer, New York
R Core Team (2014) R: A language and environment for statistical computing.
1. All nonzero digits are significant. R Foundation for Statistical Computing, Vienna, Austria. URL http://www.
2. All zeros between significant digits are significant. R-project.org/
Everitt BS, Hothorn T (2010) A handbook of statistical analyses using R,
3. All zeros which are both to the right of the decimal 2nd edition. CRC press, Boca Raton, FL
point and to the right of all non-zero significant Hothorn T, Bretz F, Westfall P (2008) Simultaneous inference in general
digits are themselves significant. parametric models. Biom J 50:346–363
Hou YQ, Wang L, Yi D, Wu G (2015) N-acetylcysteine and intestinal health: a focus
on mechanisms of its actions. Front Biosci 20:872–891
For example, 3211 has four significant digit (using rule Jobgen WJ, Meininger CJ, Jobgen SC, Li P, Lee MJ, Smith SB, Spencer TE,
1), 400.06 has five significant digits (using rules 1 and 2), Fried SK, Wu G (2009a) Dietary L-arginine supplementation reduces white-fat
gain and enhances skeletal muscle and brown fat masses in diet-induced
3.2000 has five significant digits (using rules 1 and 3), obese rats. J Nutr 139:230–237
and 0.003 has 1 significant (using rule 1). Jobgen W, Fu WJ, Gao H, Li P, Meininger CJ, Smith SB, Spencer TE, Wu G (2009b)
f High fat feeding and dietary L-arginine supplementation differentially regulate
The LSD test is carried out by choosing the LSD
gene expression in rat white adipose tissue. Amino Acids 37:187–198
option in step 4 and setting the method for adjusting Kutner KH, Nachtsheim CJ, Neter J, Li W (2005) Applied linear statistical models.
p-values to ‘none’ (which is the default selection). McGrow-Hill, Irwin
g
Use the “Find and Replace” feature in Excel to achieve Meininger CJ, Wu G (1997) L-Glutamine inhibits nitric oxide synthesis in bovine
venular endothelial cells. J Pharmacol Exp Ther 281:448–453
this task. This feature can be accessed on PC by clicking Mendiburu F (2014) agricolae: statistical procedures for agricultural research. R
CTRL + F and then go to the “Replace” tab. package version 1.1-7, http://CRAN.R-project.org/package=agricolae
Assaad et al. SpringerPlus (2015) 4:33 Page 9 of 9

Mirai Solutions GmbH (2014) XLConnect: Excel connector for R. R package

version 0.2-7, http://CRAN.R-project.org/package=XLConnect
Montgomery DC (2012) Design and analysis of experiments, 8th edition. Wiley,
New York
Piepho HP (2004) An algorithm for a letter-based representation of all-pairwise
comparisons. J Comput Graph Stat 13:456–466
Rezaei R, Knabe DA, Tekwe CD, Dahanayaka S, Ficken MD, Fielder SE, Eide SJ,
Lovering SLW (2013) Dietary supplementation with monosodium glutamate
is safe and improves growth performance in postweaning pigs. Amino Acids
44:911–923
RStudio Inc (2013) shiny: web application framework for R. R package version
0.8.0, http://CRAN.R-project.org/package=shiny
Schaffer ME (2013) rtf: rich text format (RTF) output. R package version 0.4–11,
http://CRAN.R-project.org/package=rtf
Steel RGD, Torrie JH, Dickey DA (1997) Principles and procedures of statistics; a
biometrical approach, 3rd edition. McGraw-Hill, Boston
Wang XQ, Frank JW, Little DR, Dunlap KA, Satterfiled MC, Burghardt RC,
Hansen TR, Wu G, Bazer FW (2014a) Functional role of arginine during the
peri-implantation period of pregnancy. I. Consequences of loss of function
of arginine transporter SLC7A1 mRNA in ovine conceptus trophectoderm.
FASEB J 28:2852–2863
Wang WW, Dai ZL, Wu ZL, Lin G, Jia SC, Hu SD, Dahanayaka S, Wu G (2014b)
Glycine is a nutritionally essential amino acid for maximal growth of milk-fed
young pigs. Amino Acids 46:2037–2045
Wang L, Hou YQ, Yi D, Ding BY, Zhao D, Wang ZX, Zhu HL, Liu YL, Gong J,
Assaad H, Wu G (2015a) Dietary oleum cinnamomi alleviates intestinal injury.
Front Biosci 20:814–828
Wang H, Zhang C, Wu G, Sun YL, Wang B, He BB, Dai ZL, Wu ZL (2015b)
Glutamine enhances tight-junction protein expression and modulates CRF
signaling in the jejunum of weanling piglets. J Nutr 145:25–31
Wei JW, Carroll RJ, Harden KK, Wu G (2012) Comparisons of treatment means when
factors do not interact in two-factorial studies. Amino Acids 42:2031–2035
Westfall P (1997) Multiple testing of general contrasts using logical constraints
and correlations. JASA 9(437):299–306
Westfall P, Tobias R, Wolfinger R (2011) Multiple comparisons and multiple
testing using SAS, 2nd edition. SAS Institute, Cary, NC
Wickham H (2009) ggplot2: Elegant Graphics for Data Analysis. Springer, New York
Wickham H (2011) The Split-Apply-Combine Strategy for Data Analysis. Journal of
Statistical Software, 40(1), 1-29. URL http://www.jstatsoft.org/v40/i01/
Wu G (1995) Nitric oxide synthesis and the effect of aminoguanidine and
NG-monomethyl-L-arginine on the onset of diabetes in the spontaneously
diabetic BB rat. Diabetes 44:360–364
Wu G (1997) Synthesis of citrulline and arginine from proline in enterocytes of
postnatal pigs. Am J Physiol Gastrointest Liver Physiol 272:G1382–G1390
Wu G, Meininger CJ (2008) Analysis of citrulline, arginine, and methylarginines using
high-performance liquid chromatography. Methods Enzymol 440:177–189

Submit your manuscript to a

journal and benefit from:
7 Convenient online submission
7 Rigorous peer review
7 Immediate publication on acceptance
7 Open access: articles freely available online
7 High visibility within the field
7 Retaining the copyright to your article

Submit your next manuscript at 7 springeropen.com