I have 5 means from 5 distributions: Mean Group A 33 Group B 5500 Group C 33 Group D 32223 Group E 80
I want to determine if the difference in means is significant so I run an anova and the p-value < .05 so there are at least differences in 2 means.
n=500
value<- stack(data.frame(x= rnorm(n,33,7),y=rnorm(n,5500,5), z=rnorm(n,33,7) , a=rnorm(n,32223,7) , b=rnorm(n,80,4) ) )
ex = rep(LETTERS[1:5],each=n)
dat = data.frame(value= value$values,ex)
results = aov(value ~ ex, data=dat) #NULL is EQUAL MEANS FOR ALL GROUPS, alternative is at least 2 means different. p-value < .05 means reject null and have difference in means
summary(results)
Then I want to determine which differences are significant so I run the TukeyHSD test and it report these results
t=TukeyHSD(results, conf.level = 0.95) #p-value<.05 means difference are significant
t
Tukey multiple comparisons of means
95% family-wise confidence level
Fit: aov(formula = value ~ ex, data = dat)
$ex
diff lwr upr p adj
B-A 5467.316931 5.466280e+03 5468.353394 0.0000000
C-A 0.591297 -4.451667e-01 1.627761 0.5251299
D-A 32190.195837 3.218916e+04 32191.232301 0.0000000
E-A 47.576884 4.654042e+01 48.613347 0.0000000
C-B -5466.725634 -5.467762e+03 -5465.689170 0.0000000
D-B 26722.878907 2.672184e+04 26723.915370 0.0000000
E-B -5419.740047 -5.420777e+03 -5418.703583 0.0000000
D-C 32189.604540 3.218857e+04 32190.641004 0.0000000
E-C 46.985587 4.594912e+01 48.022050 0.0000000
E-D -32142.618953 -3.214366e+04 -32141.582490 0.0000000
My qurestion is how do you report the results of the TUKEYHSD to an audience. There are 10 differences and only C-A is not significant but what I report to my audience is just the means
Mean
Group A 33
Group B 5500
Group C 33
Group D 32223
Group E 80
In reality I have 50 means not 10 so the Tukey HSD test would return (50^2-50)/2 = 1225 differences!!! How do I report on those 1225 differences?
I know this question is more on reporting but it seems like a real problem. How should one communicate that some of the differences are significant while others are not when the # of means tested is large?
Thank you.
Consider using a heat map:
you could also use
heatmap(mat)
, but assigning color according to significance level becomes dificult.