Scale Issue - MANOVA vs Kruskall Wallis

#1
Hello,

In my dissertation, I asked people in the general population to complete a battery of team process measures based on an effective team (condition 1) or ineffective team (condition 2) that they had participated on.

I asked each participant to describe the pattern of technology use (e.g. virtuality) by completing the following item:

Please indicate what percentage of your teamwork was conducted via the following platforms (answers must total 100):

• Videoconferencing (WebEx, Skype Video)
• Audioconferencing (Phone, Skype without Video)
• Emails (Gmail, Hotmail)
• Project Management Platforms (Basecamp)
• Instant Messanging (Chat, SMS)
• Face-to-Face
• Other (enter response)

Participants had to allocate exactly 100 percentage points across the tools.

I'm trying to compare the pattern of technology use across the effective team sample and the ineffective team sample (I have no specific hypothesis, I just want to see if the proportion of tool use is the same or not for good vs bad teams).

Initially, I thought of doing a t-test for each tool across the two samples (e.g. t-test comparing videoconferencing in effective vs ineffective teams, t-test for audioconferencing etc.). This seems problematic though since the percentage allocation is dependent across items (e.g. putting 100% on videoconferencing automatically means you can't put points on any other tool).

I also explored MANOVA. Conceptually, this makes sense because I'm combining my DVs to create one virtuality factor and examining the difference in this factor across effective/ineffective teams. But, analytically the dependence across my items seems to be problematic here too.

Finally, I also thought about using some non-parametic testing akin to a chi-square test but haven't found anything that fits well yet. Maybe Mann-Whitney U or Kruskall Wallis?

Do you have any ideas or experience with similar scales? I really appreciate it.
 

Karabiner

TS Contributor
#2
MANOVA won't be of much use here, I suppose, because
in case of a significant result you will want to know where
in particular groups differ. So you'd have to perform
6 or 7 comparisons (t-test or U-tests) anyway. I would
just start with t-tests (or Welch-corrected t-tests, probably)
and think about protection against false-positive results
due to multiple testing (for example, by using Bonferroni-
Holm correction or something).

With kind regards

K.