I have altered two instruments and collected a sample of n=60.
1.
For the first instrument, I changed the wording of the items to refer to life instead of work. I believe checking the cronbach alpha's for the entire scale and per sub scale should be sufficient. Doing a correlation could be beneficial to see how weak or strong the items are related but not essential.
I do not believe a factor analysis is required.
I do not believe I need to test for a normal distribution.
Is all of the above correct?
2.
For the second instrument, I also changed the wording of the items to look at a different type of the same construct. For example, if the current items refer to happiness, the set of items I created I substituted happiness with sadness. The difference between this instrument and the one mentioned in point 1, is that I have added items (10 items per scale when I only need 5 items per scale) and therefore I need to choose which five items best represent each scale.
I believe checking the cronbach alpha's for the entire scale and per sub scale should be sufficient.
I think doing a correlation is also required.
Do I need to test for normal distribution? If yes just histograms?
Do I need to run a factor analysis? If yes which method (Principal Components Analysis or Maximum Liklihood or Principal Axis Factoring and why?) AND what rotation?
Many thanks.
All instrumental studies must have factorial analysis, although for this type of study a minimum sample of 200 is required, the normal distribution (Shapiro wilk) is evaluated and descriptive data of the indicators, it must have a theoretical support for the proposal of the instrument. You must perform exploratory factorial analysis, because PCA reduces the number of factors to a m
Hello Daniel,
Your intended use(s) of scores from a measure is what governs the types of psychometric analyses that would be suitable. In the absence of this, the best I can offer is some general suggestions:
1. Even a simple change in attribute of interest (from work to life) can cause a distinction in underlying structure of a measure (especially one with 60 items!). I'm not so confident that you can safely avoid factoring. The normal distribution assumption (I presume you mean of individual responses vs. total scores) is less critical. As well, in correlational based methods, it would be bivariate normality that is under consideration.
2. There's plenty of research to indicate that swapping item polarity (e.g., from rating of happy to rating of sad) does not yield responses that are arithmetic complements. So, again, you'd want to check the behavior of the modified items (especially if measure 2 contains both kinds: it's not uncommon for folks to conclude that there are two "factors" in such situations--positive and negative--when others would argue that the correct model should be bipolar).
3. I agree with Cristian that common factor analysis would be preferable to principal component analysis.
Good luck with your work.
It is worth saying what you had hoped to learn from this pilot and how it will inform your main study. For the first instrument, I assume the main study will be large and that will be the study for you to really understand the underlying structure of the items, so with an n of 60 you are presumably looking to make sure that there is some variability for the items.
On the second instrument, if I understand correctly, you want to use the results to remove half of the items. From a statistical standard small SDs would be bad, but I would think getting subject matter experts to choose which are measuring what you want to measure will be helpful (or think alouds and see which are confusing, or response times). Maybe if you want everything in the scale to be measuring the exact same thing you could look at median correlations, but you probably want the item to measure the construct in a few ways, and not all the same ways. With big samples you could do things, but ...
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running