I have a questions measured through Likert scale (1-5)

Independent variables: I believe xx is a problem strongly disagree(1) disag neutral agree s-agree(5)

I believe yy is a problem strongly disagree(1) disag neutral agree s-agree(5)

Dependent variable xyz (1,2,3) ---not a binary

I am planning to run some tests.

I found there are lots of debate on parametric vs non parametric on it.

Some test needs normality. What about normality test in this case? Do I need to carry out normality test on Likert scale.Or is it logical?

but lets suppose that in question 1, most of the respondents agree that xx is a problem(4 or 5), so how would data be a normal distributed? obviously it will be skewed(right).

So does normality make sense in it? (unless the result is kind of neutral). I saw lots of debates going on but for an academic research what can i do? (Likert scale score vs category score or continuous and so on)

Since my dependent variable is not binary, I am planning use multiple regression(cannot use logistic regression, but log regression is much lenient on assumptions), but problem is for some reasons my all questions replies are not meeting the criteria for multiple regression (especially normality of data, and standard residual).

so what kind of do? Is it needed to carry out all assumption test to run multiple regression on likert scale question?

What may be my options(to predict or test the hypothesis)?

thank you

  • 1
    $\begingroup$ 1. Likert items are plainly not normal stats.stackexchange.com/search?q=normality+likert -- there would be no point in testing something you already know for sure... however 2. IVs are not assumed to be normal in regression. 3. Your DV is not normal either (though it's not the raw/marginal DV that is assumed to be normal), not even close. You don't state what the 1,2,3 values represent -- is that categorical? Ordered categorical? A count? Likely-to-be-suitable models for each of those cases are discussed in posts on site already. Please clarify the nature of your DV. $\endgroup$
    – Glen_b
    Oct 17 '17 at 22:54
  • $\begingroup$ thanks for your comments. DV is categorical, 0=not present,1=low presence,2=high presence. I also found that there is something called orginal regression, what may be good in my case. I am trying to take IV1,IV2 and trying to test effect on DV(hypothesis). $\endgroup$ Oct 18 '17 at 2:05
  • $\begingroup$ Please edit that additional (critical) information into your question (and also change "screwed" to "skewed" in the question). That would be ordinal regression rather than orginal, and yes, that might be a good choice for a DV with ordered categories, depending on what you're trying to find out. What is the actual question you're trying to answer? (in plain words, avoid any hint of statistical jargon please) $\endgroup$
    – Glen_b
    Oct 18 '17 at 2:14
  • $\begingroup$ my hypothesis is: A factor(IV-multivariate) is positively related to intention to use something(DV). $\endgroup$ Oct 18 '17 at 3:08
  • $\begingroup$ HOW many IVs or factors you have ? What are your question items? $\endgroup$ Oct 18 '17 at 5:26

Important issues have been answered by Glen-b. I suggest that you may reclassify responses to your dependent variable in order to meet your objective of study. An option is - classify your data on intention to use (DV) into binary data i.e. intention to use =1 and no intention to use = 0. Further, if you may classify data into high or low presence ignoring no- intention cases and look for factors - what you term as IVs with the help of Anova - that are key determinants. You can do so for intention to use or no intention to use as well by using Anova. Apparently application of multiple regression may not be a valid idea to handle the data for DV and factor or factors as Independent variables. Further querries are solicited.

  • $\begingroup$ My problem with having binary DV in this case is, that would put people using it in same category either they may be using 99% or 1%, i.e. 1% and 99% percent in same bag does not sound logical.Suggestions would be appreciated. thanks $\endgroup$ Oct 18 '17 at 12:31
  • $\begingroup$ Never dichotomize a continuous or ordinal variable. This results in huge information/power/precision loss and is arbitrary. The problem is handled ideally by ordinal regression, e.g., proportional odds model as discussed at length on this site. $\endgroup$ Oct 18 '17 at 12:35
  • $\begingroup$ Likert-scaling is bipolar and does not in any case , generate ranks ! I am not able to comprehend - ordinal scoring for responses to Likert items and ordinal-regression. $\endgroup$ Oct 18 '17 at 16:18
  • $\begingroup$ Thanks. I did some homework.I found that ordinal regression is what I should do in my case. I tried(SPSS) but problem is:a warning about too many cells with zero frequencies as there was too many IVs but sample size is low(n=200) and one category is too small.Since it is kind of academic research, conducting more survey to add data is not feasible. So,maybe I will have to contend with changing to high and low(binary) and use other form of regressions(e.g. Logistic regression). Off course, some information will be lost but that looks the most logical thing to do for now. Any suggestions? $\endgroup$ Nov 22 '17 at 2:42
  • $\begingroup$ If you indicate scoring apprach for question items and nature (positive or negative) of question items for your independent variables, I should be able to respond.If you have too many IVs, you can reduce them by grouping some question items into specific domains.I need to see your database and objectives/ hypotheses you want to test to give you a solution. best wishes . $\endgroup$ Nov 22 '17 at 6:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.