Some help needed in analyzing Likert-type data (using Repeated Measures ANOVA?)

#1
First of all hello to everyone, since I'm new to this forums.

I'm afraid statistics is not my strongpoint (to put it mildly), and thus apologize that I have to ask your help on what I fear is quite a 'beginner-level' type of question. Any help you can give me is greatly appreciated though!

First let me describe the methods of the research project I'm to analyse.

In the experiment, subjects watch several short video clips, and after each clip they're asked to give their responses to a couple of likert-type questions. 3 seperate factors are varied in the different video clips, giving 3 independent variables. The first (consisting of 2 levels) is varied between subjects, the other 2 (one consisting of 2, and one of 3 levels) are varied within subjects. The questions are in the form of 'I found the video to be disgusting', to which responses are given on a 5-point Likert-scale (strongly disagree, disagree, neutral, agree, strongle agree). 4 categories of questions exist, each one consisting of 3 different questions describing the same concept in different words.

So basically we have 12 different conditions, and the dependent measure is the responses to questions of 4 different categories. I will need to compare responses in each of those 4 categories with each other, for each of the 12 conditions, and include interaction effects etc.

There's a whole range of questions I could ask, but I'll try to spare you and limit it a bit.

My first question is obvious and cliche: which analysis should I use? I'm aware of the ongoing debate between people arguing that Likert-data can only be analysed non-parametrically, and those who say parametric analysis is just fine. At first the non-parametric argument made the most sense to me, but after reading several articles I have learned that apparently the non-parametric view is right when you analyse single likert-type items, since you're dealing with ordinal data. But that likert-type scales, consisting of several items, can be seen as providing interval data and thus can (or even should) be analysed with parametric measures.

Therefore, am I right to conclude that I can analyse my data using Repeated Measures ANOVA? I'm not quite sure if I don't violate any assumptions. I can't find a definite answer yet as to how many items a scale should consist of at minimum (since my 3 items per scale seem a bit meagre). Some state it should be 8, others say 4, and again others say it's fine as long as it's more than 1... I'm also not sure if I violate the assumption of Sphericity, and am not sure how to check for it either.

To make it more clear how I view my data, I have added a picture of what my SPSS file looks like when I try to analyse my data using the Repeated Measures analysis mentioned above (mind you the actual data is purely fictional, I've made them up just to try out how to get my analysis to work).


In the left-most column you see that this is the combined data of 3 subjects. The second column indicates the nature of the between-subjects independent variable for each subject. Finally, there are 24 'W-named' columns, which indicate the different conditions. The first two letters indicate the nature of the 2 within-subjects idependent variables (e.g. the differences in the video clip), while the final number (which follows 'Q') indicates the category of the Likert-question. As you can see, this means I have concluded that in order to be able to compare the answers to the different question categories with eachother, I have to threat them as being a 4th independent variable.

- Am I right in this conclusion that 'question category' is just another independent variable, and is this thus the right way to organize my data?

- Am I right to include all the trials of a single subject (i.e. 3 trials per condition, originating from the 3 different questions describing each question category) in the overall analysis (as depicted in the picture)? Or should I first do a kind of 'pre-analysis' per subject, and only include the mode or the median (I guess the mean makes no sense at all in this case) for each condition in the final analysis?


Again sorry for my 'newbie'-type of questions and also my long ramblings. I hope I've made my problems clear to you, and that someone can help me. A push in the right direction would do wonders, because at the moment I can't really see the trees through the forest.
 

noetsi

Fortran must die
#2
If you are using scales which combine likert questions (that is more than one question) then virtually everyone agrees there is no problem using a parametric method. The debate is over using individual likert questions. Also ANOVA is seen as far more robust to violations of normality than say regression because the CLM applies. Central limit theorem. Even with regression simulation studies suggest that the method is robust to all but extreme violations.