Kruskal-Wallis = significant but post-hoc doesn't show any sig results...

#1
Hi everybody, really hoping somebody can tell me if I'm missing something obvious here.

I've run a Kruskal-Wallis test on SPSS using the Analyze-Nonparametric-Independent Samples route and the Hypothesis Test Summary box tells me to reject the null hypothesis (sig=.003). However when I double click the hypothesis summary box to have a look at the post-hoc pairwise comparisons, none of the 'Adj. Sig' values are highlighted (I've run a few of these recently and all adjusted significant values have helpfully been highlighted in orange) and none of them are significant at the 0.05 significance level.. There is a .051 and a .056 in there, so am I to assume that the significant Kruskal-Wallis result is due to these two? I'm very confused at the moment on how to report the KW and post-hoc test results!:confused:
 

Karabiner

TS Contributor
#2
I do not know how SPSS actually performs post-hoc tests
for the K-W test. Maybe you search for it and post the result
here. But that post-hoc tests adjusted for 10 (?) pairwise
comparisons are much more conservative than the global test,
is not surprising. I suppose this problem has been dealt with
here three dozen times, maybe they should improve the forum's
search function a bit.

With kind regards

K.
 
#3
Hi Karabiner, thanks for the quick reply.
Maybe this problem has been dealt with 'three dozen times' but I have yet to find a specific answer to my question so I figured I'd ask in the hope somebody could help me out. All I want to know is how to report these results, I don't really want to have to write that the K-W found significant differences but the post-hoc couldn't determine where those differences were (of course I will if I have to, I just wondered if anybody had any advice).
 

Karabiner

TS Contributor
#4
(of course I will if I have to,
As indicated before, you could try to find out what
your software actually does. Maybe there are alternatives.
And any sensible (or mean) reviewer should ask you what
post-hoc adjustements you have performed.
 
#5
I don't know what you mean by 'what post-hoc adjustments you have performed' - I'm new to this, hence me coming to ask for help! As I said in my original post I went through Analyze - Nonparametric tests - Independent samples, then chose 'Kruskal-Wallis' and 'Multiple comparisons - all pairwise'. From what I understand after spending all evening trying to get my head around it, this route on SPSS uses Mann-Whitney U with a Bonferroni correction as a post-hoc.. please correct me if I'm wrong!
 

Karabiner

TS Contributor
#6
From what I understand after spending all evening trying to get my head around it, this route on SPSS uses Mann-Whitney U with a Bonferroni correction as a post-hoc
If this is true, then essentially every p value from pairwise
comparisons was multiplied by 10 (I suppose there were
10 pairwise comparisons). Therefore, it is not extremely
surprising that no pairwise comparison became statistically
significant on the 0.05 level. I suppose you can't do much
more than to report that you cannot provide evidence between
which groups significant differences exist.

By the way, seemingly you have compared age groups
with respect to your dependent variable (DV). Was it necessary
to categorize age, and why did you perform H-test, was
your DV rank scaled or ordinal scaled? Maybe another
data analytic approach would be more powerful.

With kind regards

K.