10: One-Way ANOVA

10: One-Way ANOVA

Objectives

Upon successful completion of this lesson, you should be able to:

  • Explain why it is not appropriate to conduct multiple independent t tests to compare the means of more than two independent groups
  • Use Minitab to construct a probability plot for an F distribution
  • Use Minitab to perform a one-way between groups ANOVA with Tukey's pairwise comparisons
  • Interpret the results of a one-way between groups ANOVA
  • Interpret the results of Tukey's pairwise comparisons

In previous lessons you learned how to compare the means of two independent groups. In this lesson, we will learn how to compare the means of more than two independent groups. This procedure is known as a one-way between groups analysis of variance, or more often as a "one-way ANOVA."

Why not multiple independent t-tests?

A frequently asked question is, "why not just perform multiple two independent samples \(t\) tests?" If you were to perform multiple independent \(t\) tests instead of a one-way between groups ANOVA you would need to perform more tests. For \(k\) independent groups there are \(\frac{k(k-1)}{2}\) possible pairs. If you had 5 independent groups, that would equal \(\frac{5(5-1)}{2}=10\) independent t tests! And, those 10 independent t tests would not give you information about the independent variable overall. Most importantly, multiple \(t\) tests would lead to a greater chance of making a Type I error. By using an ANOVA, you avoid inflating \(\alpha\) and you avoid increasing the likelihood of a Type I error.


10.1 - Introduction to the F Distribution

10.1 - Introduction to the F Distribution

One-way ANOVAs, along with a number of other statistical tests, use the F distribution. Earlier in this course you learned about the \(z\) and \(t\) distributions. You computed \(z\) and \(t\) test statistics and used those values to look up p-values using statistical software. Similarly, in this lesson you are going to compute F test statistics. The F test statistic can be used to determine the p-value for a one-way ANOVA.

The video below gives a brief introduction to the F distribution and walks you through two examples of using Minitab to find the p-values for given F test statistics. The steps for creating a distribution plot to find the area under the F distribution are the same as the steps for finding the area under the \(z\) or \(t\) distribution. For the F distribution we will always be looking for a right-tailed probability. Later in this lesson we will see that this area is the p-value.

The F distribution has two different degrees of freedom: between groups and within groups. Minitab will call these the numerator and denominator degrees of freedom, respectively. Within groups is also referred to as error. 

Between Groups (Numerator) Degrees of Freedom

\(df_{between}=k-1\)

\(k\) = number of groups

Within Groups (Denominator, Error) Degrees of Freedom

\(df_{within}=n-k\)

\(n\) = total sample size with all groups combined

\(k\) = number of groups

Minitab®  – Creating an F Distribution

Scenario: An F test statistic of 2.57 is computed with 3 and 246 degrees of freedom. What is the p-value for this test?

We can create a distribution plot. Our distribution is the F distribution. The numerator df (\(df_1\)) is 3 and the denominator df (\(df_2\)) is 246. We want to shade the area in the right tail. Our “X Value” is 2.57.

  1. Open Minitab
  2. Select Graph > Probability Distribution Plot > View Probability
  3. Change the Distribution to F
  4. Fill in the Numerator degrees of freedom with 3 and the Denominator degrees of freedom with 246
  5. Select the Options button
  6. Select A specified x value
  7. Use the default Right tail 
  8. For the X value enter 2.57
  9. OK and OK

Distribution Plot of Density vs X - Fm df1=3, df2=246

The area beyond an F-value of 2.57 with 3 and 246 degrees of freedom is 0.05487. The p-value for this F test is 0.05487.

Note: When you conduct an ANOVA in Minitab, the software will compute this p-value for you.


10.2 - Hypothesis Testing

10.2 - Hypothesis Testing

A one-way between groups ANOVA is used to compare the means of more than two independent groups. A one-way between groups ANOVA comparing just two groups will give you the same results at the independent \(t\) test that you conducted in Lesson 8. We will use the five step hypothesis testing procedure again in this lesson.

1. Check assumptions and write hypotheses

The assumptions for a one-way between groups ANOVA are:

  1. Samples are independent
  2. The response variable is approximately normally distributed for each group or all group sample sizes are at least 30
  3. The population variances are equal across responses for the group levels (if the largest sample standard deviation divided by the smallest sample standard deviation is not greater than two, then assume that the population variances are equal)

Given that you are comparing \(k\) independent groups, the null and alternative hypotheses are:

\(H_{0}: \mu_1 = \mu_2 = \cdots = \mu_k\)
\(H_{a}:\) Not all \(\mu_\cdot\) are equal

In other words, the null hypothesis is that at all of the groups' population means are equal. The alternative is that they are not all equal; there are at least two population means that are not equal to one another.

2. Calculate the test statistic

ANOVA uses an F test statistic. Hand calculations for ANOVAs require many steps. In this class, you will be working primarily with Minitab outputs.

Conceptually, the F statistic is a ratio: \(F=\frac{Between\;groups\;variability}{Within\;groups\;variability}\). Numerically this translates to \(F=\frac{MS_{Between}}{MS_{Within}}\). In other words how much do individuals in different groups vary from one another over how much to individuals within groups vary from one another.

Statistical software will compute the F ratio for you and produce what is known as an ANOVA source table. The ANOVA source table will give you information about the variability between groups and within groups. The table below gives you all of the formulas, but you will not be responsible for performing these calculations by hand. Minitab will do all of these calculations for you and provide you with the full ANOVA source table.

Source df SS MS F p
Between Groups (Factor) \(k-1\) \(\sum_{k}n_k(\overline{x}_k-\overline{x}_\cdot)^2\) \(\dfrac{SS_{Between}}{df_{Between}}\) \(\dfrac{MS_{Between}}{MS_{Within}}\) Area to the right of Fk-1, n-k
Within Groups (Error) \(n-k\) \(\sum_k \sum_i(x_{ik}-\overline{x}_k)^2\) \(\dfrac{SS_{Within}}{df_{Within}}\)    
Total \(n-1\) \(\sum_k \sum_i(x_{ik}-\overline{x}_\cdot)^2\)      
Legend
\(k\) Number of groups
\(n\) Total sample size (all groups combined)
\(n_k\) Sample size of group \(k\)
\(\overline{x}_k\) Sample mean of group \(k\)
\(\overline{x}_\cdot\) Grand mean (i.e., mean for all groups combined)
SS Sum of squares
MS Mean square
df Degrees of freedom
F F-ratio (the test statistic)

Some of the terms in the table above should look familiar, while others will be new to you. The sum of squares that appears in the ANOVA source table is similar to the sum of squares that you computed in Lesson 2 when computing variance and standard deviation. Recall, the sum of squares is the squared difference between each score and the mean. Here, there are three different sum of squares each measuring a different type of variability.

The ANOVA source table also has three different degrees of freedom: \(df_{between}\), \(df_{within}\), and \(df_{total}\). If you were to look up an F value using statistical software you would need to know two of these degrees of freedom: \(df_1 = df_{between}\) and \(df_2=df_{within}\).

3. Determine the p-value

When performing an  ANOVA using statistical software, you will be given the p-value in the ANOVA source table. If performing an ANOVA by hand, you would use the F distribution. Similar to the t distribution, the F distribution varies depending on degrees of freedom.

4. Make a decision

If \(p \leq \alpha\) reject the null hypothesis. If \(p>\alpha\) fail to reject the null hypothesis.

5. State a "real world" conclusion.

Based on your decision in Step 4, write a conclusion in terms of the original research question.


10.3 - Pairwise Comparisons

10.3 - Pairwise Comparisons

While the results of a one-way between groups ANOVA will tell you if there is what is known as a main effect of the explanatory variable, the initial results will not tell you which groups are different from one another. In order to determine which groups are different from one another, a post-hoc test is needed. Post-hoc tests are conducted after an ANOVA to determine which groups differ from one another. There are many different post-hoc analyses that could be performed following an ANOVA. Here, we will learn about one of the most common tests known as Tukey's Honestly Significant Differences (HSD) Test.

Most statistical software, including Minitab, will compute Tukey's pairwise comparisons for you. This specific post-hoc test makes all possible pairwise comparisons. In this class we will be relying on statistical software to perform these analyses, if you are interested in seeing how the calculations are performed, this information is contained in the notes for STAT 502: Analysis of Variance and Design of Experiments. This analysis takes into account the fact that multiple tests are being performed and makes the necessary adjustments to ensure that Type I error is not inflated.

In the examples later in this lesson you will see a number of Tukey post-hoc tests. Next, you will also learn how to obtain these results using Minitab.

For each pairwise comparison, \(H_0: \mu_i - \mu_j=0\) and \(H_a: \mu_i - \mu_j \ne 0\).


10.4 - Minitab: One-Way ANOVA

10.4 - Minitab: One-Way ANOVA

In one research study, 20 young pigs are assigned at random among 4 experimental groups. Each group is fed a different diet. (This design is a completely randomized design.) The data are the pigs' weights in kg after being raised on these diets for 10 months. We wish to determine if there are any differences in mean pig weights for the 4 diets.

  • \(H_0: \mu_1 = \mu_2 = \mu_3 = \mu_4\)
  • \(H_a:\) Not all \(\mu\) are equal
Pig Weight Data
Feed_1 Feed_2 Feed_3 Feed_4
60.8 68.3 102.6 87.9
57.1 67.7 102.2 84.7
65.0 74.0 100.5 83.2
58.7 66.3 97.5 85.8
61.8 69.9 98.9 90.3

Contained in the Minitab file: ANOVA_ex.mpx

Note that in this file the data were entered so that each group is in its own column. In other words, the responses are in a separate column for each factor level. In later examples, you will see that Minitab will also conduct a one-way ANOVA if the responses are all in one column with the factor codes in another column. 

Minitab®  – One-Way ANOVA

To perform an Analysis of Variance (ANOVA) test in Minitab:

  1. Open the Minitab file: ANOVA_ex.mpx
  2. From the menu bar, select Stat > ANOVA > One-Way.
  3. Click the drop-down menu and select 'Responses are in a separate column for each factor level'.
  4. Enter the variables Feed_1Feed_2Feed_3, and Feed_4 to insert them into the Responses box.
  5. Choose the Comparisons button and check the box next to Tukey. Under Results also select Tests.
  6. OK and OK

The result should be the following output:

Method
Null hypothesis All means are equal
Alternative hypothesis At least one mean is different
Significance level \(\alpha=0.05\)

Equal variances were assumed for the analysis

Factor Information
Factor Levels Values
Factor 4 Feed_1, Feed_2, Feed_3, Feed_4
Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Factor 3 4703.2 1567.73 206.72 0.000
Error 16 121.3 7.58    
Total 19 4824.5      
Model Summary
S R-sq R-sq(adj) R-sq(pred)
2.75386 97.48% 97.01% 96.07%
Means
Factor N Mean StDev 95% CI
Feed_1 5 60.68 3.03 (58.07, 63.29)
Feed_2 5 69.24 2.96 (66.63, 71.85)
Feed_3 5 100.340 2.164 (97.729, 102.951)
Feed_4 5 86.38 2.78 (83.77, 88.99)

Pooled StDev = 2.75386

Grouping Information Using the Tukey Method and 95% Confidence
Factor N Mean Grouping
Feed_3 5 100.34 A      
Feed_4 5 86.38   B    
Feed_2 5 69.24     C  
Feed_1 5 60.68       D

Means that do not share a letter are significantly different.

Difference of
Levels
Difference
of Means
SE of
Difference
95% CI T-Value Adjusted
P-Value
Feed_2 - Feed_1 8.56 1.74 (3.57, 13.55) 4.91 0.001
Feed_3 - Feed_1 39.66 1.74 (34.67, 44.65) 22.77 0.000
Feed_4 - Feed_1 25.70 1.74 (20.71, 30.69) 14.76 0.000
Feed_3 - Feed_2 31.10 1.74 (26.11, 36.09) 17.86 0.000
Feed_4 - Feed_2 17.14 1.74 (12.15, 22.13) 9.84 0.000
Feed_4 - Feed_3 -13.96 1.74 (-18.95, -8.97) -8.02 0.000
Tukey Simultaneous Tests for Differences of Means

Individual confidence level = 98.87%

ANOVA Output - Example 3

ANOVA Output - Example 4


10.5 - Example: SAT-Math Scores by Award Preference

10.5 - Example: SAT-Math Scores by Award Preference

In this example, we are comparing the SAT scores of students who said that they would prefer to win an Academy Award, a Nobel Prize, or an Olympic gold medal.

The example uses the StudentSurvey dataset provided by the Lock5 textbook.

Let's apply the five-step hypothesis testing process to this example.

1. Check assumptions and write hypotheses

The assumptions for a one-way between-groups ANOVA are:

Assumption: Samples are independent: Each student selected either Olympic, Academy, or Nobel. Each student is in only one group and those groups are in no way matched or paired. This assumption is met.

 Assumption: The response variable is approximately normally distributed for each group or all group sample sizes are at least 30: To check this we can construct a histogram with groups in Minitab. These plots show that the distributions are all approximately normal. 

Histograms of SAT grouped by award

 Assumption: The population variances are equal across responses for the group levels (if the largest sample standard deviation divided by the smallest sample standard deviation is not greater than two, then assume that the population variances are equal). Again we can use Minitab to look at the standard deviations across the groups. The largest standard deviation is 151.3 and the smallest is 114.1 for a ratio of 1.33 which is less than 2. So this assumption is met.

Statistics

            

Award

N

N*

Mean

SE Mean

StDev

Minimum

Q1

Median

Q3

Maximum

Academy

31

0

1191.0

27.2

151.3

820.0

1070.0

1200.0

1300.0

1530.0

Nobel

149

0

1239.1

9.38

114.5

920.0

1160.0

1230.0

1310.0

1550.0

Olympic

182

0

1176.7

8.46

114.1

800.0

1100.0

1190.0

1250.0

1470.0

Use Minitab to run a one-way ANOVA with the Minitab file: StudentSurvey.mpx

The following will describe the output within the context of the five-step hypothesis testing process.

Given that you are comparing \(k\) independent groups, the null and alternative hypotheses are:

Method

Null hypothesis

All means are equal

Alternative hypothesis

At least one mean is different

Significance level

\(\alpha=0.05\)

Equal variances were assumed for the analysis

2. Calculate the test statistic

You should get the following output:

Factor Information

Factor

Levels

Values

Award

3

Academy, Nobel, Olympic

The factor information tells us that our factor is the Award.

Analysis of Variance

Source

DF

Adj SS

Adj MS

F-Value

P-Value

Award

2

323269

162134

11.67

0.000

Error

359

4986078

13889

  

Total

361

5310347

   

The analysis of variance table is also known as our ANOVA source table. The source of the Award is our between-groups variation so the DF K -1 or 3-1 = 2. Error is our within-groups variation so the degrees of freedom are n - K or 362-3 = 359. The 'Adj' stands for adjusted. For a one-way ANOVA, nothing is being adjusted. The F-value is our F-test statistic, which in this case is 11.67 with a p-value of 0.000. The F value could be written as F (2, 359) = 11.67.

Means

Factor

N

Mean

StDev

95% CI

Academy

31

1191.0

151.3

(1149.3, 1232.6)

Nobel

149

1239.11

114.54

(1220.13, 1258.10)

Olympic

182

1176.73

114.13

(1159.55, 1193.91)

Pooled StDev = 117.851

The Means table is just a table of descriptive statistics. The sample size, mean, and the standard deviation is presented for each group. The last column is an unadjusted 95% confidence interval. You should not refer to these confidence intervals since these do not take into account that there are three different groups and three different confidence intervals being computed at the same time. The confidence intervals that we're interested in come with our Tukey pairwise comparison.

 need to know two of these degrees of freedom: \(df_1 = df_{between}\) and \(df_2=df_{within}\).

3. Determine the p-value

The p-value is 0.000 from the Minitab output.

4. Make a decision

\(p \leq \alpha\) so reject the null hypothesis.

5. State a "real world" conclusion.

We can conclude that the group means are not all equal.

Remember that ANOVA is just an omnibus test. It only tells us there is a difference somewhere. To determine where the differences are you would have to look at the Tukey pairwise comparison output.

Grouping Information Using the Tukey Method and 95% Confidence

Award

N

Mean

Grouping

Nobel

149

1239.11

A

 

Academy

31

1190.97

A

B

Olympic

182

1176.73

 

B

Means that do not share a letter are significantly different.

In the grouping table, Nobel and Olympic do not share a grouping letter. The means of these two groups are significantly different. Both the Nobel and Academy belong to Group A and both Academy and Olympic belong to Group B. These pairs are not significantly different.

Tukey Simultaneous Tests for Differences of Means
      

Difference of Levels

Difference
of Means

SE of
Difference

95% CI

T-Value

Adjusted
P-Value

Nobel - Academy

48.1

23.3

(-6.3, 102.6)

2.07

0.096

Olympic - Academy

-14.2

22.9

(-67.8, 39.4)

-0.62

0.808

Olympic - Nobel

-62.4

13.0

(-92.9, -31.9)

-4.79

0.000

Individual confidence level = 98.87%

The Tukey Simultaneous Tests table provides the adjusted 95% confidence intervals. These are the confidence intervals and p-values you should be looking at when you're making pairwise comparisons. These are adjusted to take into account that there are three different pairwise tests being performed simultaneously. To determine which groups are statistically significant we can look at the adjusted p-value. In this case, the only p-value less than 0.05 is the Olympic - Nobel pairwise comparison. The means of the Olympic and Nobel group are significantly different.

ANOVA Output - Example 3

In the Tukey Simultaneous 95% CI graph any interval not containing zero indicates a statistically significant difference. In this case, we see that the Olympic - Nobel group does not contain zero so this pairing has a statistically significant difference.


10.6 - Example: Exam Grade by Professor

10.6 - Example: Exam Grade by Professor

Scenario

Three professors were each teaching one section of a course. They all gave the same final exam and they want to know if there are any differences between their sections’ mean scores.

Step 1: Write Hypotheses

\(H_0:\mu_1=\mu_2=\mu_3\)

\(H_a: Not\;all\;\mu\;are\;equal\)

Means
Instructor N Mean StDev 95% CI
Dr. Al 60 68.367 17.719 (63.977, 72.756)
Dr. Oh 87 71.448 16.702 (67.803, 75.094)
Dr. Pa 98 67.939 17.465 (64.504, 71.373)

Pooled StDev = 17.2609

The standard deviations for all three classes are all similar.

Step 2: Compute Test Statistic
  1. Open the Minitab file: ANOVA_Exam_Profs.mpx
  2. Using Minitab: Stat > ANOVA > One-Way

The result is the following ANOVA source table:

Analysis of Variance
Source DF Adj SS Adj MS F-Value P-Value
Instructor 2 635.3 317.7 1.07 0.346
Error 242 72101.1 297.9    
Total 244 72736.4      

F (2, 242) = 1.07

Step 3: Find the p-Value

From our ANOVA source table, p = .346

Step 4: Make a Decision

Because \(p > \alpha\), we fail to reject the null hypothesis.

Step 5: State Conclusion

There is not enough evidence to conclude that the mean scores from the three different professors’ sections are different.

Tukey Pairwise Comparisons

There is some debate as to whether pairwise comparisons are appropriate when the overall one-way ANOVA is not statistically significant. Some argue that if the overall ANOVA is not significant then pairwise comparisons are not necessary. Others argue that if the pairwise comparisons were planned before the ANOVA was conducted (i.e., "a priori") then they are appropriate.

The results of our Tukey pairwise comparisons were as follows:

Grouping Information Using the Tukey Method and 95% Confidence
Instructor N Mean Grouping
Dr. Oh 87 71.448 A
Dr. Al 60 68.367 A
Dr. Pa 98 67.939 A

Means that do not share a letter are significantly different.

Tukey Simultaneous Tests for Differences of Means
Difference of Levels Difference of Means SE of Difference 95% CI T-Value Adjusted P-Value
Dr. Oh-Dr. Al 3.08 2.90 (-3.70, 9.86) 1.06 0.537
Dr. Pa-Dr. Al -0.43 2.83 (-7.05, 6.19) -0.15 0.987
Dr. Pa-Dr. Oh -3.51 2.54 (-9.46, 2.44) -1.38 0.351

Individual confidence level = 97.99%

Looking at the first table, all three instructors are in group A. Means that share a letter are not significantly different from one another (i.e., they are in the same group). Because all three instructors share the letter A, there are no significantly different pairs of instructors.

We could also look at the second table which gives us the t-test statistic and adjusted p-value for each possible pairwise comparison. This p-value is adjusted to take into account that multiple tests are being conducted. You can compare these p-values to the standard alpha level of .05.  All p-values are greater than .05, therefore no pairs are significantly different from one another. 


10.7 - Lesson 10 Summary

10.7 - Lesson 10 Summary

Objectives

Upon successful completion of this lesson, you should be able to:

  • Explain why it is not appropriate to conduct multiple independent t tests to compare the means of more than two independent groups
  • Use Minitab to construct a probability plot for an F distribution
  • Use Minitab to perform a one-way between groups ANOVA with Tukey's pairwise comparisons
  • Interpret the results of a one-way between groups ANOVA
  • Interpret the results of Tukey's pairwise comparisons

In this lesson you learned how to compare the means of three or more groups using a one-way between groups ANOVA. A one-way between groups ANOVA is used instead of multiple independent \(t\) tests in order to avoid increasing the likelihood of committing a Type I error.

An ANOVA provides information about the explanatory variable overall, but not about differences between the different levels of that variable. In order to compare the different pairs we need to conduct a post-hoc analysis such as Tukey's HSD test. 


Legend
[1]Link
Has Tooltip/Popover
 Toggleable Visibility