Hello,
Does anybody have any statistical suggestions for justify the pertinence to make comparisons between groups with different sample size?
I've conducted latent class analysis and i would like to compare these classes on a distal variable. Chi-square tests were used to compare classes. Results are good but i'm looking for some references that could help me justify the pertinence of these comparisons even though classes differ frome each other in terms of sample size.
Thanks!
Mohammad Ali Koushesh Vatan I want to compare profiles on continuous variables (likert scale 1-7). My hypothesis is that one profile have higher scores on distal variables comparing to the others.
See the categorical book by Agresti in the screenshot. It's probably in the z-library if you want to check. It should be of help. Best wishes, David Booth
OK. In order to compare groups, if your data normally distributed you could use one-way ANOVA. But if it doesn't, you should use kruskal-wallis.
This way, if you come up with normally distributed data, you must take two things into account as follow when you are using ANOVA:
1. See if equal variances assumed.
2. As you said, your groups are different in terms of sample size.
Some just consider first rule for choosing post-hocs (we use them for finding the source of differences). But for your case:
* Non-equal groups: Gabriel’s or Hochberg’s GF2 are most favorites here. Hochberg’s GF2 may be better if there is a larger difference in group sizes.
* Equal variances not assumed: Games–Howell is the most commonly used here.
If you are converting your categorical data into interval data, then you can use Welch's ANOVA, it is suitable for unequal sample sizes. For post hoc test it's equivalent is Games-Howell test. Check for detailed literature on this. Available.
Likert scales cannot ever be considered normal, because the numbers are not equal value. The meaning of the difference between say 2 and 4 is not necessarily the same as the difference between say 5 and 6. The main information content is ordinal. Therefore rank based tests are to be preferred. For instance, a Kruskal-Wallis test does not require equal sample size and it can be followed by subsequent comparisons between treatments using the Wilcoxon test with Bonferroni correction.
Hi Valentin, you can look into:
Gudicha, D. W., Tekle, F. B., & Vermunt, J. K. (2016). Power and sample size computation for wald tests in latent class models. Journal of Classification, 33(1), 30–51. https://doi.org/10.1007/s00357-016-9199-1.
Within the latent variable modelling framework, power estimation is directly related to the parameter values of a specified model, including sample sizes in the different classes. So if you have a very small sample size in one class, this might be a limitation of your study, but if the other parameters (e.g. total sample size, number of classes, number of indicators) are good you can reach a sufficient statistical power, which might justify the relevance of your comparisons.
Good luck !
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size