Hello,

your question is a question, many scientist from non social science are asking.

A: If you follow the strict statistical requirements, you should follow the first group.

B: It is not absolut wrong what the second group is doing. Because your data is a part of social science (because of questionnaire). So you have to regard your results more carefully. This means you can't interpret the absolut number of an correlation coefficient. E.g.: if you have two different simmilar coefficients (a) 0.565 and (b) 0.621. You should not interpret that correlation a is 0.056 less than b. Maby you should not even say that a is less than b (because the different ist so small). But you can say ther is a correlation in both cases. This carful interpretion approach is important.

C: Often Persons and Spearman correlation came to similar results. So if you have simmilar results, but you can't interpret the absoulut number (because of B) ... it becomes less important which method you choose.

D: It is common in social science that so called likert-scales could be interpretet as metric scales (not only ordinal). Behind this praxis is a lot of psychological findings: people regarde such scales as an continuum. And so you could trade such a scale as an metric scale.

So group one should not be so confident about there exact methods. But it is up to you. If your environment don't aceppt social approach (often in economic) you should follow group number one. But if you like to argue with the social science tradition, you can follow group number two. It is also possible to use the methods of both groups: when the results are very simmilar, this could mean the results are also "robust".

But I'm sorry now I have no references yet (in english).

Does this help to answer your question?

I agree with the answer of @Maria Neubauer .

But to make one point clear: There is a difference between analyzing individual Likert items (e.g. a single question), and analyzing a _scale_ composed of many Likert items.

Responses from a single item are clearly ordinal in nature.

Response data from a scale are often analyzed as a continuous. It may not be really appropriate to assume a gaussian conditional distribution (t-test, anova, Pearson correlation, linear regression), but sometimes it's a reasonable approximation. Otherwise, some other distribution could be used (beta regression), but this is rather rare.

I prefer to use the first one. If the second one is followed, assumptions of linear regression are violated. However, many researchers do use the second one also assuming continuity in likert scale data. If the re are many items then total scores / or average scores can also be used so that continuity can be assumed more easily. In this context researcher also use exploratory factor analysis.