LIKERT SCALE - Which statistical test to use? Is a test for normal distribution necessary?

[bessie]

Hello!
My questionnaire has 24 likert scale items, asking students if they consider some issues right. The likert scale has 7 responses, as follows: It’s Never right , It’s Seldom right, Sometimes it’s right, It’s Frequently right, It’s Always right, I don’t know if it’s right or not, I don’t know what this means. The answers I have collected are more than 1000.
1. What do you consider the best statistical test to use?
2. Should I do a test for normal distribution, or is it not necessary for likert scale items?

I have read many contradictory information on this, so your expert advice means a lot to me!
Thank you in advance for your time and effort!

 

[spunky]

1. What do you consider the best statistical test to use?

Depends. What do you want to know from this data?

Should I do a test for normal distribution, or is it not necessary for likert scale items?

It is not necessary. Likert items are discrete random variables. The normal distribution is continuous. The normality assumption is already violated.

 

[noetsi]

There is a dispute what Likert is. If you have enough levels and you make the assumptions the difference between each level is the same they are arguably interval like. In which case tests of interval data can be used. And means calculated. I used t tests the last time I did this because it seemed reasonable for the measures I used.

The more I read the less important normality seems if you have a reasonable sized sample.

 

[Miner]

Likert scaled data presents a number of problems. I have analyzed a lot of it for customer surveys (n > 5000) and have concluded that there is a major difference between what is statistically correct and what actually works.

• I have found that nonparametrics based on the median will often not work because the medians of two different groups were often identical despite the fact that the distributions were visually quite different.
• Despite the fact that the distributions were skewed, the mean often gave the only quantitative measure of difference between these distributions.
• In my applications, if you had to resort to a hypothesis test, the results were often of no practical significance. Note: I am in a field were we have no interest in theoretical knowledge, only in what can be leveraged to make improvements.

 

[noetsi]

"In my applications, if you had to resort to a hypothesis test, the results were often of no practical significance. Note: I am in a field were we have no interest in theoretical knowledge, only in what can be leveraged to make improvements."

Me to. And we have the entire population of interest (as you do too I imagine).

I wish I had theory to work from...

 

[Miner]

@noetsi
No, we do not have the entire population. We are probably working with 15-20% of the population, which for surveys, is an excellent response rate.

I worry about theory if the problem is design or manufacturing related, but with customers, it's more about identifying what makes them unhappy. That varies by product line, customer type and customer industry. I often find things that contradict what our 3rd-party survey company says is important, and have been proven correct over time. I think that they analyze at too macro a level and encounter an equivalent of Simpson's Paradox.