Hi I have 5 groups and there are huge differences in the sample sizes :
Grp 1 n=99, grp2 n=261, grp3 n=107, grp4 n=43 and grp5 n=12.
And I read somewhere that it's better to use Games-Howell test in case of large differences in sample sizes?
So it that correct?
Generally speaking you are right, the Games-Howell procedure was designed to tackle sample size differences between groups, but I would assume that such huge differences as in your data set lead to other problems, such as power and external validity. I am not sure if GH can fix this.
Hey.. thanks Rainer Duesing.
So can you suggest some recommendations/solutions to fix such problems.?
No, not really. I do not know how the GH reacts to different sample sizes, but t-Test typically lose power, when sample sizes differ. Such things should be considered before you start your research. Since you have such massive differences, I assume that you did not conduct an experiment, which also limits the interpretation and validity. Statistical procedures cant fix bad data.
Maybe I can merge/bifurcate the groups..
I was sceptical with such huge differences.. so thank you
I am not entirely sure about this but I think the Tukey-Kramer method is suitable for unequal sample size.
You are right Maria, but according to Howell (2013) - Statistical Methods for Psychology, the GH is preferrable over the TK method (p.395).
Just to follow up on Rainer's point: tests that assume equal variances pool the error term and hence end up with a potentially more accurate estimate of the error term. Tests that don't will have to handle the error terms differently - probably by pooling across only the groups being compared - GH does this with the Welch correction. These pooled tests will have lower apparent power than the pooled test when n is small for one group. This is however not particularly problematic as the traditional pooled test would have spurious power in this situation (e.g., inflating type I error in some situations).
Less information means less statistical power but otherwise I don't think there is a particular problem with small n here.
Hey.. thank you Thom S Baguley..
so i understand that in such situations i may end up with no statistical differences..?
You can go ahead with Games-Howell test. The Games-Howell test is more flexible than other nonparametric approaches such as Tukey’s test.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size