Likert Scale as dependent variable

#1
Hi all,

I have a question regarding the analysis of my survey data. I have created one variable of seven likert scale variables after conducting the Cronbach's alpha. Now I like to use this 'merged' variable as my dependent variable and analyse it with a number of descriptive variables which are also both likert scale and 'merged' likert scale variables. The likert scales are measures in 7 levels from strongly disagree to strongly agree.

Can I apply Ordered Logistic Regression and does someone know how to apply it? I am using STATA since that is mandatory for my thesis.

Thanks in advance!
 

noetsi

Fortran must die
#2
I am not sure what you mean by "one variable of seven likert scale variables..." How many distinct levels of the dependent variable do you have? If your final scale combines 7 variables with 7 levels and this results in say 49 distinct levels you probably would be better off running linear regression than logistic regression.

You can use ordered logistic regression with 7 levels (although the assumptions inherent in ordered might not be met forcing you to use multinomial that is unordered logistic regression instead). As always you have to test the assumptions including the ones specific to ordered logistic regression.

I don't know Stata unfortunately (sort of strange your professor required a specific software).
 

CowboyBear

Super Moderator
#3
I think by a 'merged' variable you mean the sum of responses to a set of items in a Likert scale. (Likert scale = multiple items. Likert item = one item. Score on Likert scale = a score obtained by aggregating responses to all items in the scale).

I don't think ordered logistic regression is feasible here. Remember that the number of possible values on the scale score is equal to the number of response options multiplied by the number of items. (So there are lots of possible score values). And by summing item responses you've effectively already assumed that the data is quantitative rather than ordinal.
 

noetsi

Fortran must die
#4
It is a good point that likert scale, as compared to a likert variable is commonly treated as a continuous variable. The real key is how many distinct levels you have and the assumption (difficult to justify, but commonly made) that the distance between all levels of all the variables in your likert scale are the same. As CWB noted if you are combining variables you have already made this assumption.

The rule of thumb I heard from my professors is that once you have 12 or more distinct levels in a variable, given equal spacing, your data is "interval like" and you are probably better off to use a linear method such as regression or ANOVA. The more levels the harder it is to make comparisons in logistic regression.
 
#5
Thanks for your quick reply. As you explain it this way, I indeed have 49 distinct levels in my dependent variable. I will try and conduct the regression and ANOVA test. What would you recommend if I want to regress a 7 distinct level likert scale on this dependent variable?
 

Karabiner

TS Contributor
#6
What would you recommend if I want to regress a 7 distinct level likert scale
Please do not confuse a single Likert-item with a Likert-scale. As was already mentioned in this thread, a Likert-scale consists of several Likert-items.

Regarding your last question, as noetsi already told you,
You can use ordered logistic regression with 7 levels (although the assumptions inherent in ordered might not be met forcing you to use multinomial that is unordered logistic regression instead). As always you have to test the assumptions including the ones specific to ordered logistic regression.

With kind regards

K.
 

noetsi

Fortran must die
#7
The real key is not if its likert or not, but how many distinct levels you have. In practice if you have 49 you should run linear regression or ANOVA. If you have 7 you probably should run ordered logistic regression first and test the assumption of ordered logistics. If it is not met then run multinomial logistic regression. One of these is the proportional odds assumption, unique I believe to ordered logistic regression.
 

CowboyBear

Super Moderator
#8
Thanks for your quick reply. As you explain it this way, I indeed have 49 distinct levels in my dependent variable. I will try and conduct the regression and ANOVA test. What would you recommend if I want to regress a 7 distinct level likert scale on this dependent variable?
I think there is some confusion here about terminology again. Karabiner has mentioned a couple of issues, but I think another one that's causing some confusion is your use of the term "regress on". When we say "regress A on B", A is the dependent variable. I suspect here that you actually mean that you want to use a single Likert item as an independent variable predicting your multi-item Likert scale dependent variable?

If so, it is possible to treat an independent variable as ordinal but this is a fairly complex and reasonably unusual approach. I suspect that this may not be the best thing right now. Instead you could either treat the Likert item as nominal or as continuous/metric. (I'd edge toward the latter since it's simpler and more consistent with your treatment of the DV).
 

noetsi

Fortran must die
#9
In theory it does not matter what the scale is for an independent variable (there are no distribution assumptions). Not all agree with that in practice. When I use 5 point ordinal predictors I treat them as interval because that is simplest to do. Essentially I assume that they are interval like with the difference between each level the same.

But of all the questions I have over the years, that is the one I found the least on. A professor at my defense said that in his opinion true ordinal predictors could make the slopes impossible to interpret. I don't know.
 

CowboyBear

Super Moderator
#10
You're quite right about there being no distributional assumptions on the DV. There are measurement-theory reasons why people might be uncomfortable treating an ordinal IV as interval, but that's probably a rabbit hole not to dive down in this thread :)