120 Two Way Analysis of Variance (2-way ANOVA)

Two Way ANOVA is simply an extension of One Way ANOVA from Chapter 118. Instead of one, there are two categorical variables which determine the groups (there are two “explanatory factors”). In addition, it is possible to take into account the interaction effects that might exist between the factors that are considered.

120.1 Analysis based on p-values and confidence intervals

120.1.1 Software

The Two Way ANOVA R Module can be found on the publicly available website:

https://compute.wessa.net/rwasp_Two%20Factor%20ANOVA.wasp

The same R Module is also available in RFC under the “Hypotheses / Empirical Tests” menu item.

120.1.2 Data & Parameters

This R module contains the following fields:

Data X: a multivariate dataset containing quantitative data
Names of X columns: a space delimited list of names (one name for each column)
Response Variable: a positive integer value of the column in the multivariate dataset which corresponds to the response/endogenous variable (i.e. the variable we wish to explain or predict)
Factor Variable 1: a positive integer value of the column in the multivariate dataset which corresponds to the first explanatory variable (i.e. a qualitative variable containing the single-quoted group labels)
Factor Variable 2: a positive integer value of the column in the multivariate dataset which corresponds to the second explanatory variable (i.e. a qualitative variable containing the single-quoted group labels)
Include Intercept Term. This parameter can be set to the following values:
- FALSE
- TRUE

120.1.3 Output

Consider the problem of measuring the effect of a treatment with two different categories (“A” and “B”) on a response variable within an experimental setting, while simultaneously taking into account possible gender-related differences. This naming convention implies that the word “Treatment” is to be interpreted as a synonym for “explanatory variable” or “exogenous variable” (it does not necessarily have be related to drugs or medical treatments).

We wish to examine the following effects on the response variable:

the pure effect of Treatment A
the pure effect of Treatment B (i.e. gender)
the combined effect of Treatment A and B (e.g. do females benefit more from Treatment A than males?)

The results from the Two Way ANOVA analysis are shown below:

Interactive Shiny app (click to load).

Open in new tab

The ANOVA Table is used to assess the Null and Alternative Hypothesis and is based on the F-Test just like in the One Way ANOVA case. The row which corresponds to “treatment” shows an F-statistic of 95.113 with \(p \simeq 1.668e-10\). This is a marginal (averaged-over-gender) treatment effect. Because the interaction is significant (see below), this result should not be interpreted as a standalone one-way ANOVA conclusion.

The next row corresponds to the gender effect. This is also a marginal (averaged-over-treatment) main effect. Since \(p \simeq\) 5.189e-15 we reject the Null Hypothesis for the gender main effect, but its interpretation must be qualified by the significant interaction.

The row which corresponds to “treatment:gender” is related to the “combined” effect of the treatment and gender variables. Since \(p \simeq\) 9.582e-07 we reject the Null Hypothesis and conclude that the effect of the treatment for males is significantly different from the effect of the treatment for females. In other words, the effect of the treatment depends on gender, so simple effects (e.g. treatment differences within each gender) should be interpreted first.

The results shown in the Table called “Tukey multiple comparisons of means” (Tukey 1949) provide much more detailed information. We briefly discuss the most interesting rows of the table:

B-A: the difference between both groups is 0.976 which implies that subjects in group B have a higher response score than subjects in group A. Since \(0 \notin [0.771, 1.181]\) (and \(p \simeq 0 < \alpha\)) we reject the Null Hypothesis and conclude that the difference is significant. The mean response for treatment B is significantly higher than for treatment A.
M-F: the difference between both groups is -1.514 which implies that females have a higher response score than males. Since \(0 \notin [-1.719, -1.309]\) (and \(p \simeq 0 < \alpha\)) we reject the Null Hypothesis and conclude that the difference is significant. The mean response for females is significantly higher than for males.
B:F-A:F: the difference between both groups is 1.6 which implies that females in group B have a higher response score than females in group A. Since \(0 \notin [1.214, 1.986]\) (and \(p \simeq 0 < \alpha\)) we reject the Null Hypothesis and conclude that the difference is significant. The mean response for females who receive treatment B is significantly higher than for females receiving treatment A.
A:M-A:F: the difference between both groups is -0.89 which implies that females in group A have a higher response score than males in group A. Since \(0 \notin [-1.276, -0.504]\) (and \(p \simeq 0 < \alpha\)) we reject the Null Hypothesis and conclude that the difference is significant. The mean response for females who receive treatment A is significantly higher than for males receiving treatment A.
B:M-B:F: the difference between both groups is -2.139 which implies that females in group B have a higher response score than males in group B. Since \(0 \notin [-2.525, -1.752]\) (and \(p \simeq 0 < \alpha\)) we reject the Null Hypothesis and conclude that the difference is significant. The mean response for females who receive treatment B is significantly higher than for males receiving treatment B.
B:M-A:M: the difference between both groups is 0.351 which implies that males in group B have a higher response score than males in group A. Since \(0 \in [-0.035, 0.738]\) (and \(p > \alpha\)) we fail to reject the Null Hypothesis and conclude that the difference is not significant. The mean response for males who receive treatment B is not significantly different from the response for males receiving treatment A.

Just as was the case for the Unpaired Two Sample t-Test, the Two Way ANOVA test makes the assumption of equal Variances for each group. This can be assessed by the diagnostic Hypothesis Test called “Levene’s Test for Homogeneity of Variance” (Levene 1960) which is shown in last table of the output. The results show that the Null Hypothesis (i.e. Homogeneity of Variance) is rejected. Hence, the underlying assumption of the Two Way ANOVA test is not satisfied.

For reporting, include effect sizes for each main effect and interaction, e.g. partial eta-squared:

\[ \eta_p^2 = \frac{SS_{effect}}{SS_{effect} + SS_{error}}. \]

Theoretically speaking one might dismiss the results from this Two Way ANOVA procedure (due to the violation of an underlying assumption). In practice, however, the departure from homogeneity of Variance does not have important effects when the groups are well-balanced. In this case we have the same number of females and males in groups A and B (i.e. the four groups have equal size). In other words, we may still be able to use the results from this analysis because the biasing effect of unequal Variances only emerges when the differences in Variance are related to sample size.

To compute the Two Way Analysis of Variance (2-way ANOVA) on your local machine, the following script can be used in the R console.

Note: this local script is a generic template (mtcars) for reproducible syntax. The embedded app example above uses the dedicated interaction dataset and therefore has different numeric output.

library(car)
x = mtcars
par1 = 1 #Response : mpg
par2 = 2 #Factor : cyl
par3 = 10 #Factor : gear
par4 = TRUE #Include Intercept Term
ylab = 'Y Variable Name'
xlab = 'X Variable Name'
main = 'Title Goes Here'
x1<-as.numeric(x[,par1])
f1<-as.character(x[,par2])
f2 <- as.character(x[,par3])
xdf<-data.frame(x1,f1, f2)
names(xdf)<-c('Response', 'Treatment_A', 'Treatment_B')
if(par4 == FALSE) (lmxdf<-lm(Response ~ Treatment_A * Treatment_B -1, data = xdf) ) else (lmxdf<-lm(Response ~ Treatment_A * Treatment_B, data = xdf))
(aov.xdf<-aov(lmxdf))
(anova.xdf<-anova(lmxdf))
(V1<-colnames(x)[par1])
(V2<-colnames(x)[par2])
(V3<-colnames(x)[par3])
#Tukey Honest Significant Difference Comparisons
if(par4==T){
  thsd<-TukeyHSD(aov.xdf)
  names(thsd) <- c(V2, V3, paste(V2, ':', V3, sep=''))
  print(thsd)
} else {
  print('Must Include Intercept to use Tukey Test')
}
(leveneTest(lmxdf))


Call:
lm(formula = Response ~ Treatment_A * Treatment_B, data = xdf)

Coefficients:
              (Intercept)               Treatment_A6  
                   21.500                     -1.750  
             Treatment_A8               Treatment_B4  
                   -6.450                      5.425  
             Treatment_B5  Treatment_A6:Treatment_B4  
                    6.700                     -5.425  
Treatment_A8:Treatment_B4  Treatment_A6:Treatment_B5  
                       NA                     -6.750  
Treatment_A8:Treatment_B5  
                   -6.350  

Call:
   aov(formula = lmxdf)

Terms:
                Treatment_A Treatment_B Treatment_A:Treatment_B Residuals
Sum of Squares     824.7846      8.2519                 23.8907  269.1200
Deg. of Freedom           2           2                       3        24

Residual standard error: 3.348632
1 out of 9 effects not estimable
Estimated effects may be unbalanced
Analysis of Variance Table

Response: Response
                        Df Sum Sq Mean Sq F value    Pr(>F)    
Treatment_A              2 824.78  412.39 36.7770 4.916e-08 ***
Treatment_B              2   8.25    4.13  0.3679    0.6960    
Treatment_A:Treatment_B  3  23.89    7.96  0.7102    0.5554    
Residuals               24 269.12   11.21                      
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
[1] "mpg"
[1] "cyl"
[1] "gear"
  Tukey multiple comparisons of means
    95% family-wise confidence level

Fit: aov(formula = lmxdf)

$cyl
          diff       lwr        upr     p adj
6-4  -6.920779 -10.96399 -2.8775652 0.0007419
8-4 -11.563636 -14.93298 -8.1942913 0.0000000
8-6  -4.642857  -8.51394 -0.7717744 0.0166543

$gear
         diff       lwr      upr     p adj
4-3 0.5599134 -2.678867 3.798694 0.9027767
5-3 1.1092641 -3.209110 5.427638 0.7988777
5-4 0.5493506 -3.901927 5.000628 0.9490971

$`cyl:gear`
                 diff        lwr       upr     p adj
6:3-4:3 -1.750000e+00 -15.689992 12.189992 0.9999560
8:3-4:3 -6.450000e+00 -18.296715  5.396715 0.6507761
4:4-4:3  5.425000e+00  -6.647387 17.497387 0.8319910
6:4-4:3 -1.750000e+00 -14.475413 10.975413 0.9999122
8:4-4:3            NA         NA        NA        NA
4:5-4:3  6.700000e+00  -7.239992 20.639992 0.7778059
6:5-4:3 -1.800000e+00 -17.896516 14.296516 0.9999819
8:5-4:3 -6.100000e+00 -20.039992  7.839992 0.8505445
8:3-6:3 -4.700000e+00 -13.393112  3.993112 0.6587277
4:4-6:3  7.175000e+00  -1.823226 16.173226 0.1962607
6:4-6:3 -3.552714e-15  -9.857063  9.857063 1.0000000
8:4-6:3            NA         NA        NA        NA
4:5-6:3  8.450000e+00  -2.931956 19.831956 0.2698785
6:5-6:3 -5.000000e-02 -13.989992 13.889992 1.0000000
8:5-6:3 -4.350000e+00 -15.731956  7.031956 0.9220157
4:4-8:3  1.187500e+01   6.679872 17.070128 0.0000017
6:4-8:3  4.700000e+00  -1.871375 11.271375 0.3126067
8:4-8:3            NA         NA        NA        NA
4:5-8:3  1.315000e+01   4.456888 21.843112 0.0008236
6:5-8:3  4.650000e+00  -7.196715 16.496715 0.9107420
8:5-8:3  3.500000e-01  -8.343112  9.043112 1.0000000
6:4-4:4 -7.175000e+00 -14.144996 -0.205004 0.0402197
8:4-4:4            NA         NA        NA        NA
4:5-4:4  1.275000e+00  -7.723226 10.273226 0.9998900
6:5-4:4 -7.225000e+00 -19.297387  4.847387 0.5363530
8:5-4:4 -1.152500e+01 -20.523226 -2.526774 0.0055745
8:4-6:4            NA         NA        NA        NA
4:5-6:4  8.450000e+00  -1.407063 18.307063 0.1347959
6:5-6:4 -5.000000e-02 -12.775413 12.675413 1.0000000
8:5-6:4 -4.350000e+00 -14.207063  5.507063 0.8448091
4:5-8:4            NA         NA        NA        NA
6:5-8:4            NA         NA        NA        NA
8:5-8:4            NA         NA        NA        NA
6:5-4:5 -8.500000e+00 -22.439992  5.439992 0.5126836
8:5-4:5 -1.280000e+01 -24.181956 -1.418044 0.0194305
8:5-6:5 -4.300000e+00 -18.239992  9.639992 0.9763412

Levene's Test for Homogeneity of Variance (center = median)
      Df F value  Pr(>F)  
group  7  2.1926 0.07172 .
      24                  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1