Analysis of variance

Packages

library(tidyverse)
library(smmr)
library(PMCMRplus)

Jumping rats

  • Link between exercise and healthy bones (many studies).
  • Exercise stresses bones and causes them to get stronger.
  • Study (Purdue): effect of jumping on bone density of growing rats.
  • 30 rats, randomly assigned to 1 of 3 treatments:
    • No jumping (control)
    • Low-jump treatment (30 cm)
    • High-jump treatment (60 cm)
  • 8 weeks, 10 jumps/day, 5 days/week.
  • Bone density of rats (mg/cm\(^3\)) measured at end.

Jumping rats 2/2

  • See whether larger amount of exercise (jumping) went with higher bone density.
  • Random assignment: rats in each group similar in all important ways.
  • So entitled to draw conclusions about cause and effect.

Reading the data

Values separated by spaces:

my_url <- "http://ritsokiguess.site/datafiles/jumping.txt"
rats <- read_delim(my_url," ")

The data (some random rows)

# rats %>% slice_sample(n=10)
rats

Boxplots

ggplot(rats, aes(y=density, x=group)) + geom_boxplot()

Or, arranging groups in data (logical) order

ggplot(rats, aes(y=density, x=fct_inorder(group))) +
  geom_boxplot()

Analysis of Variance

  • Comparing > 2 groups of independent observations (each rat only does one amount of jumping).
  • Standard procedure: analysis of variance (ANOVA).
  • Null hypothesis: all groups have same mean.
  • Alternative: “not all means the same”, at least one is different from others.

Testing: ANOVA in R

rats.aov <- aov(density~group,data=rats)
summary(rats.aov)
            Df Sum Sq Mean Sq F value Pr(>F)   
group        2   7434    3717   7.978 0.0019 **
Residuals   27  12579     466                  
---
Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
  • Usual ANOVA table, small P-value: significant result.
  • Conclude that the mean bone densities are not all equal.
  • Reject null, but not very useful finding.

Which groups are different from which?

  • ANOVA really only answers half our questions: it says “there are differences”, but doesn’t tell us which groups different.
  • One possibility (not the best): compare all possible pairs of groups, via two-sample t.
  • First pick out each group:
rats %>% filter(group=="Control") -> controls
rats %>% filter(group=="Lowjump") -> lows
rats %>% filter(group=="Highjump") -> highs

Control vs. low

t.test(controls$density, lows$density)

    Welch Two Sample t-test

data:  controls$density and lows$density
t = -1.0761, df = 16.191, p-value = 0.2977
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
 -33.83725  11.03725
sample estimates:
mean of x mean of y 
    601.1     612.5

No sig. difference here.

Control vs. high

t.test(controls$density, highs$density)

    Welch Two Sample t-test

data:  controls$density and highs$density
t = -3.7155, df = 14.831, p-value = 0.002109
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
 -59.19139 -16.00861
sample estimates:
mean of x mean of y 
    601.1     638.7

These are different.

Low vs. high

t.test(lows$density, highs$density)

    Welch Two Sample t-test

data:  lows$density and highs$density
t = -3.2523, df = 17.597, p-value = 0.004525
alternative hypothesis: true difference in means is not equal to 0
95 percent confidence interval:
 -43.15242  -9.24758
sample estimates:
mean of x mean of y 
    612.5     638.7

These are different too.

But…

  • We just did 3 tests instead of 1.
  • So we have given ourselves 3 chances to reject \(H_0:\) all means equal, instead of 1.
  • Thus \(\alpha\) for this combined test is not 0.05.

John W. Tukey

  • American statistician, 1915–2000
  • Big fan of exploratory data analysis
  • Popularized boxplot
  • Invented “honestly significant differences”
  • Invented jackknife estimation
  • Coined computing term “bit”
  • Co-inventor of Fast Fourier Transform

Honestly Significant Differences

  • Compare several groups with one test, telling you which groups differ from which.
  • Idea: if all population means equal, find distribution of highest sample mean minus lowest sample mean.
  • Any means unusually different compared to that declared significantly different.

Tukey on rat data

rats.aov <- aov(density~group, data = rats)
TukeyHSD(rats.aov)
  Tukey multiple comparisons of means
    95% family-wise confidence level

Fit: aov(formula = density ~ group, data = rats)

$group
                  diff       lwr       upr     p adj
Highjump-Control  37.6  13.66604 61.533957 0.0016388
Lowjump-Control   11.4 -12.53396 35.333957 0.4744032
Lowjump-Highjump -26.2 -50.13396 -2.266043 0.0297843
  • Again conclude that bone density for highjump group significantly higher than for other two groups.

Why Tukey’s procedure better than all t-tests

Look at P-values for the two tests:

Comparison        Tukey    t-tests
----------------------------------
Highjump-Control 0.0016     0.0021
Lowjump-Control  0.4744     0.2977
Lowjump-Highjump 0.0298     0.0045
  • Tukey P-values (mostly) higher.
  • Proper adjustment for doing three t-tests at once, not just one in isolation.

Checking assumptions

ggplot(rats,aes(y = density, x = fct_inorder(group)))+
  geom_boxplot()

Assumptions:

  • Normally distributed data within each group
  • with equal group SDs.

Normal quantile plots by group

ggplot(rats, aes(sample = density)) + stat_qq() + 
  stat_qq_line() + facet_wrap( ~ group)

The assumptions

  • Normally-distributed data within each group
  • Equal group SDs.
  • These are shaky here because:
    • control group has outliers
    • highjump group appears to have less spread than others.
  • Possible remedies (in general):
    • Transformation of response (usually works best when SD increases with mean)
    • If normality OK but equal spreads not, can use Welch ANOVA. (Regular ANOVA like pooled t-test; Welch ANOVA like Welch-Satterthwaite t-test.)
    • Can also use Mood’s Median Test (see over). This works for any number of groups.

Mood’s median for multiple groups

  • Find median of all bone densities, regardless of group

  • Count up how many observations in each group above or below overall median

  • Test association between group and above/below

  • Mood’s median_test (over).

Mood’s median test here

median_test(rats, density, group)
$grand_median
[1] 621.5

$table
          above
group      above below
  Control      1     9
  Highjump    10     0
  Lowjump      4     6

$test
       what        value
1 statistic 1.680000e+01
2        df 2.000000e+00
3   P-value 2.248673e-04

Comments

  • No doubt that medians differ between groups (not all same).
  • This test is equivalent of \(F\)-test, not of Tukey.
  • To determine which groups differ from which, can compare all possible pairs of groups via (2-sample) Mood’s median tests, then adjust P-values by multiplying by number of 2-sample Mood tests done (Bonferroni):
pairwise_median_test(rats, density, group)
  • Now, lowjump-highjump difference no longer significant.

Welch ANOVA

  • For these data, Mood’s median test probably best because we doubt both normality and equal spreads.
  • When normality OK but spreads differ, Welch ANOVA way to go.
  • Welch ANOVA done by oneway.test as shown (for illustration):
oneway.test(density~group, data=rats)

    One-way analysis of means (not assuming equal variances)

data:  density and group
F = 8.8164, num df = 2.000, denom df = 17.405, p-value = 0.002268
  • P-value very similar, as expected.
  • Appropriate Tukey-equivalent here called Games-Howell.

Games-Howell

  • Lives in package PMCMRplus. Install first.
# gamesHowellTest(density ~ group, data = rats)
gamesHowellTest(density ~ factor(group), data = rats)
         Control Highjump
Highjump 0.0056  -       
Lowjump  0.5417  0.0120

Careful: explanatory must be factor (so commented-out line does not work).

Deciding which test to do

For two or more samples: