4
$\begingroup$

I have to analyse 8 independent variables (IV) toward 8 dependent variables (DV) in my research. The IV is nominal while DV are ordinal (Likert-scale). I run the normality test and the result is zero (p<.05) which lead me to perform non-parametric statistical test to the data. I chose the Kruskal Wallis test at first, because I wanted to test the differences among the groups (involved lots of groups) in influencing the DV. Since my hypothesis is just wanting to know if either of the IVs are significantly associated or not with the DV, I neglected the role of post hoc analysis (for instance Mann-Whitney U test), because my aim is not to know which group causes the differences.

My question:

  1. Is it a must to do post hoc analysis for Kruskal-Wallis test? i'm afraid my analysis go wrong if I just report the significant value of Kruskal Wallis result. I found some article that just report the result of mean rank, rather using Mann Whitney U test. Is it acceptable if I just report the significant value and mean rank of the result? If I proceed with Mann-Whitney U test, I am gonna end up producing lots of result (possibly more than 1000 tables for that purpose).

  2. I consult with several people and they suggest that I use MANOVA because my study involved lots of variables. Therefore, they suggest MANOVA which assist in producing result based on the whole data, not segregated like Kruskal Wallis analysis. However, I found that the data shall be in normally distributed in order to proceed with it. My data are totally non-normally distributed. How can I proceed with it? Is it wrong if I used MANOVA for non-normally distributed data?

I confuse in choosing the appropriate analysis. Which one should I use for my research? Thank you so much for your help! Please let me know if there is any other information I should provide.

$\endgroup$
  • $\begingroup$ Welcome to CV, farah! The Mann Whitney $U$ test is inappropriate as a post hoc test following rejection of a Kruskal-Wallis null hypothesis, because (1) the Mann Whitney $U$ test does use the same rankings as K-W (it uses pairwise rankings), and (2) the Mann Whitney $U$ does not incorporate the pooled variance estimate implied by the K-W test. Dunn's test, and the (less well know, but strictly more powerful) Conover-Iman test are appropriate post hoc tests folling the rejection of K-W. $\endgroup$ – Alexis Apr 17 at 4:19
  • $\begingroup$ The general sentiment on Cross Validated is that a formal test of normality is unhelpful. You got a small p-value...okay, but p-values do not just measure how large the deviation from the assumed null hypothesis is. Sample size plays a huge role, and fairly large sample sizes can result in normality testing rejecting even though the difference is not enough for us to care to change our methodology. Graphical examination of your data, such as a normal quantile-quantile plot, is likely to be more helpful. $\endgroup$ – Dave Apr 17 at 7:09
0
$\begingroup$

For 8 input variables and 8 outcome variables, you need multivariate multiple regression or MANCOVA.

MANOVA is used in case of one input and multiple outcomes.

| cite | improve this answer | |
$\endgroup$
  • 2
    $\begingroup$ It might be worth making a comment about the normality requirements, since that is a large part of the OP's concern. $\endgroup$ – Silverfish Mar 16 '16 at 11:29
0
$\begingroup$

Can I ask what your sample size is? Are you planning on running all those tests?

Is there no other way of reorganizing your data?

If not, why not use the following multinomial model:

DV = IV1 + IV2 + IV3 + IV4 + IV5 + IV6 + IV7

where each of your nominal categories are dummy variables (1,0) and your IV 8 is your baseline?

| cite | improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.