Dear Gro,

Try this

http://www.talkstats.com/showthread.php/4478-Checking-procedure-for-collapsing-5-point-likert-scale-to-3-point

Dear Gro,

Try this

http://www.talkstats.com/showthread.php/4478-Checking-procedure-for-collapsing-5-point-likert-scale-to-3-point

Another option!

https://www.researchgate.net/post/How_can_I_convert_different_point_Likert_scales_for_all_questionnaires_in_a_survey_so_that_they_are_the_same

+

http://www.talkstats.com/showthread.php/27379-How-to-Convert-a-5-Point-Likert-Scale-into-a-7-Point-Likert-Scale

+

http://www.talkstats.com/showthread.php/1256-Converting-from-a-7-to-a-5-point-scale

Thanks a lot - very helpful!

Ellen

It depends on how these different Likert scales were worded. They could have been worded by simply giving the extremes (1 is "don't do this activity" to 5 "I do this very often" vs. 1 for no activity to 7 "I do this very often"), without category labels. They could have been worded with every category labeled. And so on. Can you provide the full text of both questions?

One solution might be to transform the items or item mean scores to POMP scores via POMP = 100*(old - min)/(max-min) whereby min and max are the endpoints of your Likert scale (e.g. 1 and 5 or 1 and 7), not the actual minima or maxima in your data. See: Cohen, P., Cohen, J., Aiken, L. & West, S. (1999). The problem of units and the circumstance for POMP. Multivariate Behavioral Research, 34(3), 315-346. [ http://www.informaworld.com/smpp/content~content=a785043286~db=all~order=page ]. Of course, there are psychometic issues, especially if the wording of the extremes differ or if not only the extremes were labeled.

The five point scale: 1 = strongly disagree, 2= disagree, 3 = neutral, 4=agree, 5=strongly agree

The seven point scale: 1 strongly disagree, 2, 3, 4 , 5, 6, 7 strongly agree (that is, no explicit descriptions of 2-6)

A few options as I see them:

1. Compare proportions in the categories that are directly comparable: strong opinionos "SD" Old 1 vs. New 1, "SA" Old 5 vs. New 7, intermediate Old 2 through 4 vs. New 2 through 6.

2. Convert everything to [0,1] interval or [-1,1], so that -1 = Strongly Disagree and 1 = Strongly Agree. Treat the scale as continuous for comparison purposes, however, keep in mind that the data are effectively heteroskedastic, and if say t-tests of differences are used, they should be of Welch's form.

3. Assume an underlying normal continuous trait and categorization into ordinal responses a la ordered probit (http://en.wikipedia.org/wiki/Ordered_probit) and convert everything to the conditional mean of the resulting truncated normal distribution (http://en.wikipedia.org/wiki/Truncated_normal_distribution). Treat the scale as continuous for comparison purposes. I implemented the latter in my polychoric Stata package.

The high brow discussions in technical literature may all be fine, but in all likelihood they are an overkill. We can argue ad infintium whether a scale like this should be analyzed using a partial credit model, and whether increasing the range of the scale leads to violation of residual invariance, or some other stuff like that. My take on this variety of approaches is practical: if you have large differences, any method can show them. If you have no differences, most methods would say, inconclusive/statistically insignificant. It may well be that you have weird differences that some methods detect but others don't, but frankly I would not expect to stumble upon that very often.

If you want to have a really fancy approach, you can probably utilize a structural equation model with multiple groups (two periods) if you have common demographic variables and other external drivers for your Likert scales. If you are not familiar with SEM, then of course this would push your limits way too far. I am not going to try to dig your profile to figure that out though :).

I would also advice that the likert scales are reduced to two levels i.e. either agree/disagree. This was the results/responses would be comparable. I am currently working on a study that requires comparing responses obtained before the project and those after the project with the same respondents involved. This is what I have done in the analysis.