Hello,
In my research in SPSS, I use a variable 'quality of friendship'. I will use this variable in a multiple regression analysis. The variable consists of 3 items. 2 items have a 5 point Likert scale, 1 item has a 6 point Likert scale. Can I combine the items or do I need to make sure they all have the same Likert scale first? If so, how do I do that without having a big influence on the data?
I also found out I have a lot of outliers and I think I figured out why. 1 question is as follows: 'My best friend is important in my life'. The Likert scale answers are: 1=very important, 2=important, 3=less important, 4=unimportant, 5=no such person in my life. When I looked at some graphs, I saw that there is a normal distribution for the answers1-4. However, the answer 5 was given a lot and seen as big outliers. Does anybody have an idea what to do with this? Do I just ignore the outliers?
I'd love some tips. Thanks!
Hi Elisa,
For the third question you can merge the 6th and 5th answers into one answer with the code 5.
Also, you need to inverse the codes of answers (e.g. give the answer very important the code 5, important = 4, ...)
The usual approach is to rescale both ratings. Different approaches exist. For example:
i) z transformation: for each Y subtract the mean of all values and divide by the SD of all values
ii) POMP scaling: percentage or proportion of maximum performance. for each Y subtract the minimum and divide by the (max - min). So a rating of 3 on a 1-5 scale is (3-1)/(5-1) = 2/4 = .50 or 50% (which makes sense because it is halfway)
I prefer POMP scaling as z transformation can distort things because of scaling by the SD. (For example if there are ceiling or floor effects on one scale but not the other). POMP scoring is sometimes called normalisation (though that term is used for about a 100 different things).
If the scales are behaving differently I'm not sure it makes sense to combine them (and the resulting scale will have poor psychometric properties). In this case your 1-5 scale is arguably a 1-4 scale with a Not applicable response. I think would be more sensible to handle this as a 0/1 variable (no best friend / best friend) and a 1-4 rating for those with a best friend.
I think it might be more useful to work with the individual ratings directly, though I guess some sort of structural equation model with a latent variable might be possible.
Dear Elisa de Lange,
It is ideal to transform into 5-point Likert's Scale. The measures of the 6-point scale can be transformed by multiplying it by 5/6 and it is a simple way of transforming. For example, if a respondent has responded as 2 to this question, it can be transformed into a 5-point scale as 2*(5/6) = 1.67. You can use this value of 1.67 directly into the analysis, since the mean and standard deviations of each item eventually placed with decimal values.
The suggestion by Thom S Baguley is very good. But, you need to work accordingly. Also, the suggestion by Mohammed Suleiman M. Gibreel is also good to reorganise scale into positive codes, like Very important = 5, Important = 4 and so on.
Hope this will help. Good luck in your studies.