See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/274635267

One-Way Analysis of Variance (ANOVA)

Technical Report · April 2014

CITATIONS READS

3 6,303

2 authors, including:

Peter Samuels
Birmingham City University
153 PUBLICATIONS   272 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

The relationship between self-regulated learning, learning development, academic achievement and life-long learning View project

Critical Review View project

All content following this page was uploaded by Peter Samuels on 07 April 2015.

The user has requested enhancement of the downloaded file.

 community project
encouraging academics to share statistics support resources
All stcp resources are released under a Creative Commons licence

stcp-gilchristsamuels-9

The following resources are associated:

One-Way ANOVA – Additional Material worksheet and Normality Testing worksheet

One-Way Analysis of Variance (ANOVA)

Research question type: Differences between several groups of measurements
What kind of variables: Continuous (scale/interval/ratio)
Common Applications: Comparing the means of different groups in scientific or
medical experiments when treatments, processes, materials or products are being
compared

Example: Grocery Bags1

A paper manufacturer makes grocery bags. They are interested in increasing the tensile strength
of their product. It is thought that strength is a function of the hardwood concentration in the pulp.
An investigation is carried out to compare four levels of hardwood concentration: 5%, 10%, 15%
and 20%. Six test specimens were made at each level and all 24 specimens were then tested in
random order. The results are shown below:

Hardwood Standard
Tensile strength (PSI) Mean Median
concentration (%) deviation
5 7 8 15 11 9 10 10.00 2.83 9.50
10 12 17 13 18 19 15 15.67 2.81 16.00
15 14 18 19 17 16 18 17.00 1.79 17.50
20 19 25 22 23 18 20 21.17 2.64 21.00
All 15.96 4.72

Research question: Are there differences in mean tensile strengths between the different
hardwood concentrations?
th
1. Source: Montgomery, D. and Runger, G. (2011) Applied Statistics and Probability for Engineers, 5 ed.,
Hoboken, NJ: Wiley.

© Mollie Gilchrist and Peter Samuels Reviewer: Ellen Marshall

www.statstutor.ac.uk Birmingham City University University of Sheffield
Based on material provided by Loughborough University Mathematics Learning Support Centre and
Coventry University Mathematics Support Centre
One-Way Analysis of Variance (ANOVA) Page 2 of 4

In the analysis of variance we compare the variability between the groups (how far the means are
apart) to the variability within the groups (how much natural variation there is in the
measurements). This is why it is called analysis of variance, often abbreviated to ANOVA.
Hypotheses
The null hypothesis is:
H0: There is no difference in mean tensile strength between the four hardwood
concentrations
The alternative hypothesis is:
H1: There is a difference in mean tensile strength between the four hardwood
concentrations

Steps in SPSS
Create two numerical variables called Concentration and Strength. In
the example on the right, Concentration has codes 1 for 5% hardwood
concentration, 2 for 10%, etc. and Strength is the tensile strengths in
PSI. These codes can be explained using the Values field in the
Variable View.
Testing ANOVA assumptions
ANOVA has three assumptions:
1. The observations are random samples from normal distributions.
2. The residuals for the whole data set are normally distributed. This theoretically follows from the
first assumption but it is worth testing separately with small samples.
3. The groups have equal variances.
To test the first assumption, select
Analyze – Descriptive Statistics –
Explore…, select Strength on the
Dependent List and Concentration on
the Factor List. Under Plots… select
Normality plots with tests. The
Shapiro-Wilk normality test for each
group is negative (p > 0.05, see right), indicating that we
may assume that the data is normally distributed (the
Kolmogorov-Smirnov test should not be used for small
sample sizes).
Note: For more thorough normality checking, use
histograms and boxplots for large samples and QQ plots
for small samples – see the normality testing worksheet.
To test the second assumption we first need to create the
residuals:

© Mollie Gilchrist and Peter Samuels Reviewer: Ellen Marshall

www.statstutor.ac.uk Birmingham City University University of Sheffield
Based on material provided by Loughborough University Mathematics Learning Support Centre and
Coventry University Mathematics Support Centre
One-Way Analysis of Variance (ANOVA) Page 3 of 4

• Select Analyze – General linear model – Univariate

• Add Strength as the Dependent Variable and Concentration as the Fixed Factor
• Select Save… and choose Unstandardised residuals
Now we can test the residuals for normality:
• Select Analyze – Descriptive Statistics – Explore…
• Select Standardized Residual for Strength on the Dependent List and leave the Factor List
blank
• Under Plots… select Normality plots
with tests
This again gives a non-significant
result for the Shapiro-Wilk test (see
right), indicating that we may accept the
second assumption.
To test the third assumption we need to
use a Levene’s test for equality of
variance which is contained within the
ANOVA analysis in SPSS:
• Select Analyze – Compare Means –
One-Way ANOVA
• Select Strength on the Dependent List
and Concentration on the Factor list
• Select Options… then select Homogeneity of
variance (Levene's) test in the Statistics list
This produces the output shown on the right. This is
not significant at 95% so we may accept assumption
3.
Note: for robust use of ANOVA see the additional advice sheet.
Running the ANOVA analysis
Now that we have shown that all the assumptions of ANOVA can be accepted we can run the
ANOVA analysis. However, the required output table is already provided with the Levene’s test we
already carried out.
The ANOVA table gives an F
statistic of 19.61 and a p-value of
< 0.001 (the value in the Sig.
column of 0.000 has been
rounded to 3 decimal places and
should not be quoted in this
format). Therefore we reject the null hypothesis with 99.9% confidence (note: subtract the p-value
threshold from 1 and multiply by 100 to obtain the confidence level) and conclude that there is very

© Mollie Gilchrist and Peter Samuels Reviewer: Ellen Marshall

www.statstutor.ac.uk Birmingham City University University of Sheffield
Based on material provided by Loughborough University Mathematics Learning Support Centre and
Coventry University Mathematics Support Centre
 One-Way Analysis of Variance (ANOVA) Page 4 of 4

strong evidence that the mean tensile strengths of the different groups are unequal. However, we
do not know which pairs of group mean differences are significantly different, if any. This can be
explored by performing post hoc tests.

Post hoc tests

Post hoc test are 'after the event' tests used to establish significant differences between pairs of
means after an ANOVA has been performed and was found to be statistically significant. In effect
they are variations on independent samples t-tests comparing the pairs of means.
The significance level needs to be adjusted in order to reduce the possibility of Type I errors. This
leads to many choices of types of post hoc test. Field (2009, p. 375) recommends the following:
• For equal group sizes and similar variances, use Tukey or, for guaranteed control over Type I
errors (more conservative), use Bonferroni
• For slightly different group sizes, use Gabriel
• For very different group sizes, use Hochberg’s GT2
For example we had equal group sizes so we shall use Tukey and Bonferroni:
• Select Analyze – Compare Means – One-Way ANOVA
• Select Strength on the Dependent List and Concentration on the Factor list
• Select the Post Hoc… button in the One-way ANOVA Dialog box
• Select Tukey and Bonferroni
The SPSS output indicates statistically significant differences in mean strength between:
• 5% and 10% (p < 0.01 for both methods)
• 5% and 15% (p < 0.001 for both methods)
• 5% and 20% (p < 0.001 for both methods)
• 10% and 20% (p < 0.01 for both methods)
• 15% and 20% (p < 0.05 for Tukey only)
However, there was no statistically significant difference between 10% and 15% for either method
or between 15% and 20% with Bonferroni.
Note: Four groups gives six pairs with each listed twice for both methods

Reference
Field, A. (2009) Discovering Statistics using SPSS (And sex and drugs and rock 'n' roll), 3rd ed.,
London: SAGE.

© Mollie Gilchrist and Peter Samuels Reviewer: Ellen Marshall

www.statstutor.ac.uk Birmingham City University University of Sheffield
Based on material provided by Loughborough University Mathematics Learning Support Centre and
Coventry University Mathematics Support Centre

View publication stats