Part of my dissertation looks at a replication of the NEP study conducted in 2000. I'm attempting to compare the two (my repsonses and the responses received in 2000) and see whether my survey garnered a higher frequency for any of the 5 likert-scale answers. However, from looking at the distribution, it appears the GLM that I had originally ran would not be appropriate as the data is not normally distributed.
I have organised the data be researcher, question, Likert rating and percentage response. I'm unsure how to test the difference between our participants and ensure validity as I only have the frequency of responses for Dunlap et al. I have looked at several tests but I am unsure which is the best option?
Any help would be greatly appreciated
It makes a difference whether you Likert-scored items are used individually or added to form a scale. Because there have been so many questions here about Likert-scored items, I have compiled a set of resources on this topic:
https://www.researchgate.net/post/How_can_one_help_people_create_simple_scales_from_Likert-scored_items
Thank you for your help! I’ve been getting a bit lost in some of the terminology some responses involve. Hopefully the link you provided will supply some answers. It seems that advice often seems to contradict at times but hopefully I’ll be able to interpret the best course of action 🙂
.
if i understand correctly, the only information you have from the previous study is the frequency of each Likert item or, to put it simply, the frequencies of the answers for each question, those answers taking values in the ordered set {1, 2, 3, 4, 5}
if by "frequency" you mean "percentage" and you do not have the population total of the previous study (unlikely ...) this is bad news ; otherwise, for each question, you have the number of respondents for each value in {1, 2, 3, 4, 5}
obviously, you have the same data for your own study as you asked the same questions but to different people
first things first, il would just contemplate the histograms for each question and look at some rough summaries such as mode, median, 1st and 3rd quartiles, IQR ...
then, for each queston, i would run a simple Mann Whitney test (your two samples are unpaired)
https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test
remember that Mann Whitney tests for "stochastic dominance" (its usual interpretation as testing for the median only holds if both distributions are continuous and only differ by the median)
this is where contemplating basic statistics and histograms can be useful ; for instance MW will not reject the null hypothesis for the following two distributions (listing frequencies for {1, 2, 3, 4, 5}) : {1000, 0, 0, 0, 1000} and {0, 2000, 0, 2000, 0}, although they are indeed quite different.
.
if the above just added to your confusion, just ignore it and accept my apologies !
.
That actually is very helpful thank you! They give their data in % but they do say how many people answered each question as you mentioned.
.
ok, then you can proceed for a simple MW test for each question
the next question is what kind of difference do you want to test : difference on a question per question basis or difference on the complete set of questions
if this sounds unfamiliar, have a look at :
https://en.wikipedia.org/wiki/Bonferroni_correction
.
of course, when you only have the populations for each answer to all questions, you cannot reconstruct the Likert score of the previous study and, unless the authors give some information about their score distribution, you cannot test anything about the Likert score
.
Fabrice Clerot ideally I would like to look at both? However, the most important would be the difference on a question per question basis for my study
- Badges
- Science topic
- Similar topics
- Medical Science Nursing
- Dissertations