5
$\begingroup$

I have a hypothesis that a particular intervention/treatment will cause more variation in participant responses to a particular question.

The intervention variable is categorical, with five different treatment groups. The response variable (the participant responses to a question) is a continuous variable.

I don't necessarily expect the means to differ, I just expect greater variation in responses in certain groups as compared to others.

Can anyone advise me on a method to test the difference in the variance of different treatment groups? To be clear, this is not for the purpose of checking the assumption of homogeneity of variance for statistical tests like ANOVA, rather I am interested in specifically if and to what extent the variation differs between the different levels of my categorical intervention/treatment variable.

I thought to run Bartlett's test for homogeneity of variance but (in R at least) I only get an output that shows me whether the variances are homogeneous or not - it doesn't tell me where these differences lie between the different levels of the categorical variable (i.e. is it between group 1 and 5, or group 2 and 3 etc).

I thought also to try to calculate a coefficient of variation for each group and compare these, but I was not sure of a method of how to do this statistically ...

I am probably missing something very obvious, but I cannot find information on how to proceed. Any advice would be much appreciated.

$\endgroup$
1
  • $\begingroup$ You could calculate the variance in each treatment group. Then you get the point estimates (of the variances). Where you see the largest differences, those should be driving the Bartlett. $\endgroup$
    – Student
    Nov 6 '19 at 13:25
2
$\begingroup$

This is an interesting question! Post-hoc tests of variances after a test of unequal variances do not seem to be a much studied topic, I was not able to find any published papers. One similar question with some ideas is Post-hoc test to determine difference in variance. Another approach is the following.

But first, note that tests of variances are typically not very robust, so distributional assumptions do matter. Maybe you could include a plot of your data in the post. I will illustrate my methods with some simulated, normal data, if your data is far from normal be careful.

One test of variances which is somewhat robust is Levene's test. As it is based on an analysis of variance, but using absolute residuals as response, it is useful in this situation, we can construct Tukey HSD intervals based on those absolute residuals. The Levene's test is often nowadays used with median as location estimator in place of the mean, but here I will use the mean for illustration. The code for simulation the example data is at the end.

 with(mydf, lawstat::levene.test(X, Group, location="mean"))
oneway.test(absres ~ Group, data=mydf, var.equal=TRUE)

    Classical Levene's test based on the absolute deviations from the mean
    ( none not applied because the location is not set to median )

data:  X
Test Statistic = 4.2954, p-value = 0.003079

    One-way analysis of means

data:  absres and Group
F = 4.2954, num df = 4, denom df = 95, p-value = 0.003079

showing the equivalence. Then the post-hoc test:

TukeyHSD(aov(absres ~ Group, data=mydf))
  Tukey multiple comparisons of means
    95% family-wise confidence level

Fit: aov(formula = absres ~ Group, data = mydf)

$Group
          diff         lwr      upr     p adj
B-A  0.9578979 -0.87239851 2.788194 0.5937356
C-A  1.9853491  0.15505269 3.815646 0.0265586
D-A  1.8196928 -0.01060365 3.649989 0.0521132
E-A  2.4268567  0.59656026 4.257153 0.0033957
C-B  1.0274512 -0.80284522 2.857748 0.5258877
D-B  0.8617949 -0.96850157 2.692091 0.6860887
E-B  1.4689588 -0.36133766 3.299255 0.1770429
D-C -0.1656563 -1.99595277 1.664640 0.9990998
E-C  0.4415076 -1.38878886 2.271804 0.9622041
E-D  0.6071639 -1.22313251 2.437460 0.8875472

The same approach is used here, but with medians in place of means. Code for simulating the data:

set.seed(7*11*13)# My public seed
N <- 20; k <- 5
Group <- rep(LETTERS[1:k], rep(N, k))
X <- rnorm(k*N, rep(rnorm(k, 10, 1), rep(N, k)),
           rep(2*sqrt(1:k), rep(N, k)))
mydf <- data.frame(Group=factor(Group), X)
rm(X, Group)
mydf$absres <- abs(resid(lm(X ~ Group, data=mydf)))

This could at least serve as a starting point.

With unequal variances, it could also be interesting to ask about reasons for it, specifically if the treatment could have anything to do with it. If so, this post could be of interest.

$\endgroup$
3
  • 1
    $\begingroup$ Thank you so much for this very detailed response! It is incredibly useful. Can I ask why one might choose to base the residuals on the median vs the mean (or vice versa)? Would one be better suited to skewed data? This may end up all being hypothetical in my case as my data are extremely skewed, but it still would be good to know! Thank you again! $\endgroup$
    – Sarah
    Nov 6 '19 at 19:15
  • 1
    $\begingroup$ If distributions in each group is close to normal, mean should be best. Else use median, or maybe trimmed mean. Maybe deserves better investigations. $\endgroup$ Nov 6 '19 at 19:17
  • $\begingroup$ Thank you. Median would seem sensible in my case then! I wonder if instead of performing an ANOVA/Tukey HSD analysis on the residual data I could instead use a non-parametric Kruskal-Wallis with the Dunn test as a post hoc? This would make the analysis rank based and so deal better with skew. Although I am not sure how much skew even rank-based tests can deal with and still be reliable... $\endgroup$
    – Sarah
    Nov 6 '19 at 19:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.