Using a Fligner test to infer about the respect of the assumption of homoscedasticity is not very smart given that the Fligner test tests to the null that there is no difference of variance between the groups. This will wrongly favors small sample size. As it has been said by @Michael Mayer here.

How can we further investigate if the assumption of homoscedasticity is respected?

Is it worth plotting the model's residuals versus the fitted values? Below lines are R coded:

m = aov(myFormula, myData)
plot(residuals(m), m$fit)

I don't have much experience in statistics and it seems rather hard for me to decide from this plot whether the assumption of homoscedasticity is respected. What else can I do?

  • 1
    $\begingroup$ If you plot(m) you should get 4 plots, including a scale-location plot, which should be easier to interpret - if the mean of the scale location plot clearly changes, you don't have homoskedasticity. $\endgroup$
    – Glen_b
    Commented Nov 20, 2013 at 1:37

1 Answer 1


You might consider Levene's test, and this discussion of how to use it in R. It may be less sensitive to non-normal distributions than Bartlett's test, according to the following reference. There's also the Brown-Forsythe test for ANOVAs on transformed response variables, though this may be another version of Levene's test, judging from this question.

It sounds like your independent variable (IV) has discrete categories, but if this isn't the case, this answer provides loads of good recommendations for testing homoscedasticity across a continuous dimensional IV.


Bartlett's test. NIST/SEMATECH e-Handbook of Statistical Methods, Available online, URL: http://www.itl.nist.gov/div898/handbook/eda/section3/eda357.htm. Retrieved December 31, 2013.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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