Kolmogorov and Levene
Kolmogorov and Levene
I check Kolmogorov for all data and are >0.05 after for repeated measures anova, for Levene test, 3 of my 4 variables are <0.001. Should I change for non parametric anova?
-
reclusivestupid
- Posts: 2
- Joined: Mon Sep 18, 2023 4:12 am
Re: Kolmogorov and Levene
If you did the Kolmogorov-Smirnov test and found that the data for your dependent variable in the repeated measures ANOVA are normally distributed (p > 0.05), this means that the assumption of normality is met for that variable. This is proof that parametric analysis, like the repeated measures ANOVA, should be used connections puzzle
But if the Levene's test for homogeneity of variances shows that the variances of your groups are significantly different (p 0.05) for three out of the four variables, it suggests that the assumption of homogeneity of variances has been broken. In this situation, you might want to use something other than the repeated measures ANOVA.
But if the Levene's test for homogeneity of variances shows that the variances of your groups are significantly different (p 0.05) for three out of the four variables, it suggests that the assumption of homogeneity of variances has been broken. In this situation, you might want to use something other than the repeated measures ANOVA.
Re: Kolmogorov and Levene
What type of test would be the best option in this case? Everything I've reviewed so far suggests using non-parametric tests. The data transformation has been tested with several options and the result for the Levene test remains the same.reclusivestupid wrote: ↑Mon Sep 18, 2023 4:16 am If you did the Kolmogorov-Smirnov test and found that the data for your dependent variable in the repeated measures ANOVA are normally distributed (p > 0.05), this means that the assumption of normality is met for that variable. This is proof that parametric analysis, like the repeated measures ANOVA, should be used connections puzzle
But if the Levene's test for homogeneity of variances shows that the variances of your groups are significantly different (p 0.05) for three out of the four variables, it suggests that the assumption of homogeneity of variances has been broken. In this situation, you might want to use something other than the repeated measures ANOVA.
Re: Kolmogorov and Levene
I could see your answer but I dont understand it. It´s simple my question.
Re: Kolmogorov and Levene
gvt,
My short, simply answer is:
"No," Given your design, you should not attempt to find a non-parametric substitute for the ANOVA. Why? Because there's no broadly accepted non-parametric test.
Note that my longer answer continues to be:
Just because the data/residuals are significantly non-normal and the variances are significantly unequal doesn't necessarily mean the violations are large enough to matter. Some judgment is required to make those determinations. It's not unusual to see people report that their analyses violate one or more assumptions while they nevertheless argue that the results are still meaningful because the violations aren't too severe.
If you're dealing with the kind of data-set you've talked about previously (multiple and possibly interacting factors, with some repeated-measures and some non-repeated-measures factors),there's no standard, widely-accepted 'non-parametric' test for that.
Alternatively, if you want to deal with just a single, repeated-measures factor there's Friedman's test. However, it needs to be kept in mind that, as a rank-based measure, Friedman's test can tell you only whether the distributions are non-identical. The test can't tell you for example that the means are significantly different unless you have good reasons to assume that the distributions have identical shapes.
While there are sophisticated approaches such as permutation testing and bootstrapping, it takes considerable effort and time to figure those out--and even then you'll likely run into disagreements about which is the best specific strategy to use.
Personally, I would run the repeated-measures ANOVA and acknowledge any assumption-violations. Then to address remaining concerns about the validity of the conclusions, I would follow-up with particular, select, non-parametric tests on the most important pairs of means and/or individual means. I think the easiest, most widely acceptable tests would be the Mann-Whitney rank-based test for pairs of non-repeated-measures means, or the Wilcoxon signed rank test for a pair of repeated-measures means or for a single mean.
My short, simply answer is:
"No," Given your design, you should not attempt to find a non-parametric substitute for the ANOVA. Why? Because there's no broadly accepted non-parametric test.
Note that my longer answer continues to be:
Just because the data/residuals are significantly non-normal and the variances are significantly unequal doesn't necessarily mean the violations are large enough to matter. Some judgment is required to make those determinations. It's not unusual to see people report that their analyses violate one or more assumptions while they nevertheless argue that the results are still meaningful because the violations aren't too severe.
If you're dealing with the kind of data-set you've talked about previously (multiple and possibly interacting factors, with some repeated-measures and some non-repeated-measures factors),there's no standard, widely-accepted 'non-parametric' test for that.
Alternatively, if you want to deal with just a single, repeated-measures factor there's Friedman's test. However, it needs to be kept in mind that, as a rank-based measure, Friedman's test can tell you only whether the distributions are non-identical. The test can't tell you for example that the means are significantly different unless you have good reasons to assume that the distributions have identical shapes.
While there are sophisticated approaches such as permutation testing and bootstrapping, it takes considerable effort and time to figure those out--and even then you'll likely run into disagreements about which is the best specific strategy to use.
Personally, I would run the repeated-measures ANOVA and acknowledge any assumption-violations. Then to address remaining concerns about the validity of the conclusions, I would follow-up with particular, select, non-parametric tests on the most important pairs of means and/or individual means. I think the easiest, most widely acceptable tests would be the Mann-Whitney rank-based test for pairs of non-repeated-measures means, or the Wilcoxon signed rank test for a pair of repeated-measures means or for a single mean.