Is non-parametric test necessary for independent Likert-type items?

#1
Hi everyone I am having a difficulty in extending my survey analysis containing Likert-type item/questionnaires. I'll try my best to give a detailed overview of how my findings and analysis have been doing up to this point.

The purpose of my survey is to determine the most important signage in our surroundings seen by Deaf people. To succeed with this, I decided to create a 5-point Likert scale (5=very important, 4=important 3=moderately important 2=less important 1=unimportant). After tallying the results, I proceeded with the commonly used descriptive statistics which take the ff:

1. Median or mode of each Likert response
2. Upper and lower quartile (range) in case no.1 doesn't suffice the tallied result of an item.
3. Bar chart for summary of tallied results

My problem and/or concern is proceeding with an inferential technique to further support my data such as that of non-parametric tests that are suitable for ordinal data (I treat them as ordinal since each item is completely independent and doesn't affect one another). I would like to ask if descriptive statistics will do to suffice my data, or is it strongly necessary that I proceed with a non-parametric test? If yes, kindly recommend me one that suits my nature of data.
 

CB

Super Moderator
#2
Welcome to TalkStats.

It isn't clear what you would be trying to achieve with an inferential analysis. You may need to describe your data and specific goals in more detail - that way we can give better help. E.g., what would be the question that you would be trying to answer with a non-parametric hypothesis test?
 
#3
First of all, thank you for your warm welcome and acknowledging my concern.

Honestly, I am unfamiliar with the usage of these non-parametric tests but as for the nature of my survey I'm going to show a screenshot of the first five items.



What I am trying to achieve is to sort the items which scored the highest so I would be able to identify which are the most important signs among the 75 Likert-items that I have inserted in the survey. My main problem is that, there are inconsistencies with my mode results (for example: in item 1, I get a bimodal for 1,1,3,5,5 which is inconsistent because 1=unimportant, 5=very important) and it really confuses me. I've researched that to simplify this scenario, one can merge very important, important, moderately important to "agree" (1) and less important, unimportant to "disagree" (0), which can be done by either a Cochran Q test or McNemar test. And that's my problem, I lack the proficiency to pursue this non-parametric test. I would like to know if I can support my data just by depending on descriptive statistics such as interquartile range, median, frequency distribution tables or such that gets the central tendency of a variable? Thank you.
 

CB

Super Moderator
#4
What I am trying to achieve is to sort the items which scored the highest
Probably the median is the main thing you're looking for then (and the other descriptives you mention could help too). The mode is rarely the focus in this kind of analysis so don't worry about it.

You haven't really mentioned any research questions that call for an inferential test, such as a non-parametric test. Statistical tests are used to answer specific research questions. If your research questions change, maybe you'll need to change your tests. But don't think about data analysis as "I have this data, therefore I must do test Y". What analysis you use depends on what you're trying to find out. :)
 

rogojel

TS Contributor
#5
hi,
if the bimodality is significant, could this be a sign of inhomogenity in your population? These are preferences after all, maybe different subgroups have different preferences? In this case investigating this could help you to understand the situation a lot better then doing a test, IMO.

regards
rogojel
 
#6
Probably the median is the main thing you're looking for then (and the other descriptives you mention could help too). The mode is rarely the focus in this kind of analysis so don't worry about it.

You haven't really mentioned any research questions that call for an inferential test, such as a non-parametric test. Statistical tests are used to answer specific research questions. If your research questions change, maybe you'll need to change your tests. But don't think about data analysis as "I have this data, therefore I must do test Y". What analysis you use depends on what you're trying to find out. :)
Thank you, seems clear to me that I don't have to completely rely on mode alone. This was a huge clarification to me. :yup:
 
#7
Perhaps a small sample size? Because I only had 5 Deaf respondents for conducting this survey. That is why I also included median and interrange quartile in getting my results. Thanks!
 

noetsi

No cake for spunky
#8
With five respondents you have two issues. First, some statistics are asymptotically correct (that is they are unbiased only with a large sample size which five certainly is not). Also power will be low - which means any test of the null hypothesis will be very doubtful. For pure descriptives that may not matter, but it will for most statistical tests and I doubt non-parametrics is going to solve that (in fact I think they have weaker power).

A second issue, outside of statistics per se, is that your ability to generalize to the larger population with five people is very doubtful.

I note in passing that this is intended for observational studies in the social sciences. Many disagree that you need a specific size in controlled experiments.
 
Last edited: