I have used 3 pre-established (4 points, 5 points and 7 points) scales to develop and validate a new comprehensive scale. I collected data using original point scales. How ever I need to do variable redundancy and I just wonder whether I can put all these 3 scales into same PCA analysis. Your guidance highly appreciated
kanthi
Dear Kanthi,
You can put all the items in a PCA... but some suggestions:
(a) Try other extraction methods (rather than PCA). I also would like to encourage you to explore factor analysis methods for ordinal data.
(b) Put (a) under suspicion :-) I think factor analysis doesn't replace the need for a content analysis - in your case - . If the scales are similars, factor solution will not give you much more informations to avoid redundance. Perhaps you need to compare item by using other item-analysis options (e.g., item-discrimination indexes, etc.).
(c) Last and more important, I suggest integrate your theoretical background in the item selection process (I think your objective will not be achieved by just applying statistical analysis).
Hope it helps!!
Hi Kanthi,
For PCA, the data should be numerical. Since your data is categorical/ordinal in nature, you may apply other statistical technique that would depend on your goals.
For example: if you want to develop a statistical model and your response variable is ordinal in nature (Likert scale) , you may apply logistic regression ordinal/binary. If you can clarify I may provide more input.
I have been doing PCA and Logistic regression modeling so if you need any support let me know.
Thank you.
Produce z-scores. This will standardise different scales.
And one more thing- use not the Cronbach's alphas, but composite reliability, since the scales are different.
Best,
Eugene
Thank for all valuable answers. I will certainly use theoretical background to decide which items to keep. main purpose of PCA to reduce item form the tool. I am not familiar with Logistic regression for item redundancy. If Dr. Singh can help me that is appreciated
Thanks
Kanthi
Hi Kanthi,
If I understand correct,ly you have intention to develop a statistical model. PCA applies when you have several variables and you are looking for the variables that have the maximum contribution to the model based on the eigen values. Other suggestions may also work but I would recommend only if I have more clarity on your goals. If you can clarify: your goals and type and number of variable. I may provide some robust suggestions. Thank you and good luck.
Best,
Sushant
Hello Kanthi,
That the variables have different potential ranges isn't a cause for worry. Internally, both principal components and common factor analysis work with correlations, which automatically standardizes all variables. So, converting all individual scores to z-scores in advance might feel good, but won't change the results.
What can be a cause for concern is that individual 3-, 4-, or 7-point response scale variables are not great representations of a continuous variable (and they seldom manifest normal distributions, individually or jointly). As a result, some analysts will choose polychoric correlations to represent relationships among variables rather than Pearson product-moment correlations. In the R statistical package, the psych library includes a very simple polychoric computation which you could use as a pre-processor to the factoring.
Good luck with your study!
Hi Dr. Sushant
yes I need to use PCA to identify the variables which make maximum contributions based on eigen value. I have 88 items in three different scales (4 point, 5 point and 7 point) I need to reduce the number of items in my final tool which will be using at military recruitment.
I hope this make sense to you about my work.
Hi Kanthi,
pca requires continuous variables; the Likert scales produce a maximum ordinal variables . My advice is to analyze separately the three scales to eliminate variables less associated in each scale . Then you can make a last pca on selected variables in the previous analyzes ,
I suggest you investigate this process in my book .
http://www.francoangeli.it/Ricerca/Scheda_libro.aspx?ID=21599&Tipo=Libro&titolo=Factor+analysis+and+principal+component+analysis++
Dear Dr. Franco
This is so useful. Thank you very much. I will look at the book. Hope it is in English
Thanks again
Kanthi
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis