Take the 2-minute tour ×
Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. It's 100% free, no registration required.

I have a three point Likert scale question:

How happy are you?
    1= low levels of happiness 
    2= medium levels 
    3= high levels 

I want to do multiple linear regression on the variable. I am making the assumption that that it has the same difference between low and medium and medium and high. I want to treat it as a continuos dependent variable.

I know some people would see this as inappropriate, and there is the age old issue about whether to treat survey data as continuous.

  • But are there any further problems applicable to my dependent variable because it is only on a 3 point scale as opposed to 4 or 5?
share|improve this question
1  
Although explicitly not asked for, this question on when Likert scales should be treated as ordinal or interval scales may be of interest: stats.stackexchange.com/q/10/442 –  Henrik Sep 27 '11 at 11:37

3 Answers 3

What I like to do in situations like this is run ordinal logistic, multinomial logistic and linear regression and compare results. I compare results by looking at predicted values from the different models. This is easy for multinomial vs. ordinal, because both yield probabilities. To compare with linear, I sometimes see what the highest probability response is from logistic, and how close it is to the predicted value from linear.

share|improve this answer
1  
Thanks for this. I need to justify what i have done is right. So it is for my thesis, i cannot do more stats and do a small section on methodogy i.e. different ways of running the models produces the same/different answers. I dont have the word space! –  Laila Sep 27 '11 at 13:40
    
@Laila In the cases that I've seen, thesis appendices do not count towards the word count. Appendices are a good place to report supplementary analyses. –  Jeromy Anglim Sep 28 '11 at 0:41

The issue is not about "treating it as a continuous variable", but rather, as you point out, the assumption that 1v2 is the same as 2v3.

There's no real issue about the 3 categories (vs more). If there were just two categories, it would be equivalent to a 0/1 binary variable. All that would change (for different encodings) would be the scale of the coefficient.

A full model would contain two columns (say, indicators for medium and high). Using just the one column is perfectly reasonable but the assumption deserves to be checked.

share|improve this answer
    
My answer is wrong, as I was thinking of it as the independent variable not the dependent variable. I guess I should delete it? Let's just say I'm answering a different question. –  Karl Sep 27 '11 at 14:54

A bit on terms first. Your observable variable isn't continuous, it is descrete. Instead of "continuous" you should have better use word "scale" (or "metrical" or "interval"+"ratio"), as opposed to "ordinal". And you plan to treat the variable as scale rather than ordinal.

Formally logically you have the right to treat any Likert variable as scale variable if it has 3 or more levels. Unless you suspect that the intervals are psychometrically uneven. For example, in many countries they use 1 through 5 scale to assess performance at schools. This scale is not equiinterval because difference in knowledge appeares to be greater between 3 (satisfactory) and 4 (good) than between 4 and 5 (excellent). So, the school gauge is ordinal.

If you are right deciding to treat a Likert variable as metrical then the more it has levels the beter it is statistically. But if the variable is rather ordinal and has many levels it may be better to roughen it into less levels. One example is awful 10-point (1 to 10) scale frequently used in marketing to assess satisfaction. I usually recode it into 3-level "1 through 4", "5 through 8", "9 through 10".

share|improve this answer
    
Hi @ttphns Have you got any links or further comments about why the 10 level scale is "awful" and why roughening it makes it better? –  Peter Flom Sep 27 '11 at 9:32
2  
First, because it looks trivial it exaggerates focal effects on "magic" numbers 5, 7, 10 - you always encounter this 3-modality in the distribution. Second, looking at "intrapsychic scale" or usual respondent of satisfaction question, 9/10 almost indistinquishable, but stand apart from 5-7 or 5-8 range, and this is in turn apart from 1-3 or 1-4 range. Third, 10 points is just too much for a usual human differentiation capacity. 6 or 7 Likert scales are optimal for an adult, and 3 to 4, for a child. –  ttnphns Sep 27 '11 at 10:01
    
Interesting. Long ago in grad school, I did some work on alternate scorings of a depression inventory - either 4 point or essentially infinite point (we let the respondents check a spot on a line). Results were highly correlated, but the factor structure was different. I guess you could always check trimodality. Have you got any references for the stuff about intrapsychic scales or human differentiation capacity? Thanks –  Peter Flom Sep 27 '11 at 13:06
    
Thanks Ttphns....are there any tests or assumptions i need to test before i can say that treating it as an 'interval' scale is correct? –  Laila Sep 27 '11 at 13:43
2  
The idea that "happiness" occurs on a discrete scale is amusing. –  whuber Sep 27 '11 at 14:11

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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