StATS: When does heterogeneity become a concern? (created 2008-06-05).
Dear Professor Mean, I have an ANOVA model and I am worried about heterogeneity--unequal standard deviations in each group. How should I check for this?
There's no perfect answer to a question like this and certainly no consensus in the research community on how to handle heterogeneity. There is no magic formula and not very much empirical data on this question. As a general rule a lack of research consensus is good news in that it allows you more latitude to analyze the data you see best. Just be ready to apply the opposite approach when a peer-reviewer insists that his/her way is the only way to handle this problem.
Here are some general guideline that I use.
First, avoid formal significance tests like Levene's test. There are a wide range of reasons why significance tests should be avoided for checking some of the underlying assumptions in a model. They tend to have too much power and detect trivial deviations from the assumptions when the sample size is large and tend to have too little power and miss critical deviations when the sample size is small. Also when you choose between one test statistic and another on the basis of a statistical test can sometimes destroy nice properties of you hypothesis testing framework.
I prefer to use a criteria based on the magnitude of the degree of heterogeneity, rather than the statistical significance for heterogeneity. Look at the ratio of standard deviations in each group. Do any differ by more than a factor of 2 or 3? Only then do I start to worry about heterogeneity.
Is there a "dose response" pattern to the standard deviations? A common one is that the larger standard deviations are associated with the groups with the larger means and the smaller standard deviations are associated with the groups with the smaller means. That makes me sit up and take notice more than if the group with the largest standard deviation is not one with an extremely large (or small) mean.
Is there a theoretical or empirical basis for heterogeneity? A theoretic basis might occur, for example, if you are comparing two ethnic groups, for example, and one is known to be tightly knit with highly similar dietary and behavioral patterns. There may be an empirical basis for heterogeneity because previous studies using your outcome measure showed evidence of unequal variation across groups. With theoretical or empirical support for heterogeneity, you should be more open to this possibility with your data, possibly to the point of pre-specifying an analysis model that incorporates unequal variation across groups.
Is there imbalanced data? Heterogeneity is more of a concern if the sample sizes are not the same in each group. The combination of imbalance and heterogeneity can distort the Type I error probability in your model. With balanced data (same or nearly the same sample size in each group), there is less need to worry about heterogeneity because it is unlikely to have a major impact.
This page was written by Steve Simon while working at Children's Mercy Hospital. Although I do not hold the copyright for this material, I am reproducing it here as a service, as it is no longer available on the Children's Mercy Hospital website. Need more information? I have a page with general help resources. You can also browse for pages similar to this one at Category: Analysis of variance.