Comparing Likert scales of different lengths

S

Statischmics

New member
Joined
Jan 15, 2013
Messages
3
Hi there math and statistics experts! For my thesis I am doing research about listener's attitudes towards 4 different Dutch accents. Next to that I also do research about the lengths of Likert scales. I did 2 surveys with the same questions but one with a likert scale length of 6 and one of 7 answer options. I'd like to compare the 2 in a statistical useful way. Also, statistics isn't my strength. I use SPSS. What do you guys think that I should do?


(My solutions are: -Multiplying the answers of the 6 scale survey by 7, and the 7 scale survey by 6. This might be hard to defend, given that the intervals between the answer options weren't the same to begin with. -Using z-scores to compare the two surveys. This is quite helpful, and both my survey answer data show quite a high degree of normality. But I cannot really tell which survey of the 2 gets rated the highest, for example.)

Thanks for your help in advance!
 
Statischmics said:
Hi there math and statistics experts! For my thesis I am doing research about listener's attitudes towards 4 different Dutch accents. Next to that I also do research about the lengths of Likert scales. I did 2 surveys with the same questions but one with a likert scale length of 6 and one of 7 answer options. I'd like to compare the 2 in a statistical useful way. Also, statistics isn't my strength. I use SPSS. What do you guys think that I should do?


(My solutions are: -Multiplying the answers of the 6 scale survey by 7, and the 7 scale survey by 6. This might be hard to defend, given that the intervals between the answer options weren't the same to begin with. -Using z-scores to compare the two surveys. This is quite helpful, and both my survey answer data show quite a high degree of normality. But I cannot really tell which survey of the 2 gets rated the highest, for example.)

Thanks for your help in advance!
Click to expand...
Is SPSS the current name for my favorite data analysis program, SigmaPlot? (I use version 8.0.)

The interval of a Likert scale is arbitrary, so you are justified in rescaling. I would be tempted to put zero in the middle. Perhaps let both scales run from -3 to +3. The 7-point scale would be integers, but the 6-point scale would be
(-3, -11/6, -7/12, 7/12, 11/6, 3)

[At this point I usually but in the disclaimer that I am not a statistician per se, but that as an experimental physicist I do have a strong reputation for proper application of statistical analysis to experimental data. Come to think of it, wearing a different hat, I currently use something a little bit like a 5-point Likert scale (perhaps I should say "range" because fractions are allowed) with values -1, 0, 1, 2, 3 meaning poor, neutral, good, very good, and "perfect." The survey in that case has 4 items, and a weighted average is used to calculate an overall score. My goal here is to generate a number from qualitative input.]

The statistical weights to use when combining or comparing data should "always" be the numbers of people who took each survey.

I generally favor using mean and standard deviations as estimators of central tendency and of spread - but for a Likert scale it might be better to use the median and inter-quartile distance. Are you looking for the difference between even- and odd-scales?
 
DrPhil said:
Is SPSS the current name for my favorite data analysis program, SigmaPlot? (I use version 8.0.)

The interval of a Likert scale is arbitrary, so you are justified in rescaling. I would be tempted to put zero in the middle. Perhaps let both scales run from -3 to +3. The 7-point scale would be integers, but the 6-point scale would be
(-3, -11/6, -7/12, 7/12, 11/6, 3)

[At this point I usually but in the disclaimer that I am not a statistician per se, but that as an experimental physicist I do have a strong reputation for proper application of statistical analysis to experimental data. Come to think of it, wearing a different hat, I currently use something a little bit like a 5-point Likert scale (perhaps I should say "range" because fractions are allowed) with values -1, 0, 1, 2, 3 meaning poor, neutral, good, very good, and "perfect." The survey in that case has 4 items, and a weighted average is used to calculate an overall score. My goal here is to generate a number from qualitative input.]

The statistical weights to use when combining or comparing data should "always" be the numbers of people who took each survey.

I generally favor using mean and standard deviations as estimators of central tendency and of spread - but for a Likert scale it might be better to use the median and inter-quartile distance. Are you looking for the difference between even- and odd-scales?
Click to expand...

Thanks for your reply! You were right about the goal of that particular part of my research. The difference between even and odd scales, (and to a lesser extent the length of scales in general). The way you are presenting the solution sounds quite straightforward and also statistically justifiable. Maybe I'll combine that with the multiplying part of my original post, so I'll only leave integers. I also like the fact that I don't have to use Z-scores in that case. This still leaves me with one last question: Can you think of a way in which I can statistically prove that one of the 2 surveys yields more positive/negative responses than the other one? Cause that would be pretty cool if that's possible.

Btw, SPSS is probably one of the most well known statistics program in the social sciences. Not sure if it has anything to do with Sigmaplot...

Thanks again Phil!
 
Statischmics said:
This still leaves me with one last question: Can you think of a way in which I can statistically prove that one of the 2 surveys yields more positive/negative responses than the other one? Cause that would be pretty cool if that's possible.
Click to expand...
The standard deviation of the distribution would probably miss the effect, because its computation depends a lot on the tails. But the inter-quartile range should be smaller if there is a significant central (neutral) peak. To emphasize that effect further, could look at octiles or deciles... trying finer divisions depends on how many data you have in each survey.

[The web site for SigmaPlot is spss.com - either two products merged, or one spun off the other!]
 
DrPhil said:
The interval of a Likert scale is arbitrary, so you are justified in rescaling. I would be tempted to put zero in the middle. Perhaps let both scales run from -3 to +3. The 7-point scale would be integers, but the 6-point scale would be
(-3, -11/6, -7/12, 7/12, 11/6, 3)
Click to expand...

Good idea about using octiles and deciles. I'll look into that.

Are you sure about those numbers? (-3, -11/6, -7/12, 7/12, 11/6, 3)
If I do the math, there's not a constant number in between those fractions. Between 0,583333 (7/12) and 1,83333 (11/6) is 1,25. But between 3 and 1,83333 (11/6) is 1.1667. You probably see that my math skills are very bad, but shouldn't there be the same number between those scale points?

Thanks again for your quick response!
 
You must log in or register to reply here.
Top