One way ANOVA vs Kruskal-Wallis

[stat_18]

Hi Guys,

I have three grous and their distribution is not normaly distributed so I decided to use Kruskal-Wallis rather than One way ANOVA. The annoying thing is that Kruskal-Wallis tells if there is/isn't a significant difference between the three groups but does not tell from where the difference come if there is. In other words, is the difference (if there is) is between group 1 and 2 or 1 and 3 or 2 and 3. Any ideas how to know these info?

Many thanks in advance
Ali

 

[Dragan]

Hi Guys,

I have three grous and their distribution is not normaly distributed so I decided to use Kruskal-Wallis rather than One way ANOVA. The annoying thing is that Kruskal-Wallis tells if there is/isn't a significant difference between the three groups but does not tell from where the difference come if there is. In other words, is the difference (if there is) is between group 1 and 2 or 1 and 3 or 2 and 3. Any ideas how to know these info?

Many thanks in advance
Ali

In terms of a post-hoc analysis for the Kruskal-Wallis test, have a look at this:

http://www.talkstats.com/showthread.php?t=4894

 

[stat_18]

Thanks Dragan. You mean if |RBari - RBarj| >= the right side term, then the two groups are significantly different. is this what you meant to say?

Many thanks
Ali

 

[gianmarco]

Hi,
every stat pack should routinely provide a post-hoc test performing pair-wise comparisons to' assess where the significant difference is.

I am familiar with software like xlstat, statistica, sigmaplot, and they all provide such comparisons. If you want to try a free one, go and search for the program PAST.

Regards.
Gm

 

[emre_dunder]

Primarily, you should use ANOVA, because the distribution is normal,then if you want to compare the means between these groups,you should use post hoc comparisons, you can occur that by many package programme like SPSS,you can use Tukey and Scheffe test for this

 

[Dragan]

Thanks Dragan. You mean if |RBari - RBarj| >= the right side term, then the two groups are significantly different. is this what you meant to say?

Many thanks
Ali

Yes, that's right.

 

[stat_18]

Thanks Guys for all of your suggestions.

I have one more question please. I have set of data that compares 5 measures between three group of patients. I need to compare measurement A between the three groups. Then do the same thing for all other measuemnts. So which test I should use. is there any restriction regarding the normality (ie if measurement A is normally distributed use x test, if not normally distributed use y test).

Many thanks
Ali