0

I'd like to get your opinions on how to interpret items that had to be eliminated in factor analysis (FA).

I've been researching consumer shopping motivations and ran a survey with 40 items, which included statements from previous research in the field (there is no developed go-to scale). I expected items would be reduced to 10-12 variables. I ran FA in SPSS (extraction method - principal components) and after all items with low factor loadings and communalities were eliminated, I was left with 25 items and 6 underlying variables (criteria for the Number of factors: eigenvalue>1).

Now, what conclusion can I make about statements which dropped-out? Are these (eliminated) motivations not valid for the consumers I ran the survey with?

Thanks a lot on your thoughts!

chl
  • 50,972
  • 18
  • 205
  • 364
desperate-about-statistics
  • 19
  • 1
  • 7
  • Before you draw any conclusions about which items to keep, try running factor analysis per se, instead of principal component analysis. In SPSS it uses the same initial menu but a different extraction method (plus you are better off using oblique rotation). For reasons why, see http://stats.stackexchange.com/questions/1576/what-are-the-differences-between-factor-analysis-and-principal-component-analysi . – rolando2 Jun 25 '14 at 23:27

1 Answers1

3

You are discussing two different issues. The first is what we can call "factorial validity". In other words, your principal component analysis determined how many different components are assessed by your items, and which items may not assess those components very well. You seem to be left with 25 items and 6 different components/scales/factors. The eliminated items are not good measures of these 6 different components; however, this does not mean that they are not "valid" (the second issue). The validity of an item can be somewhat reduced to how useful it is. To determine the utility of an item/scale, you need some sort of real-world dependent variable, or at least other variables that are deemed valid to some extent. What are you trying to predict?

So the eliminated items may not fit well within the scales you seem to have, but that does not mean they are not useful. It depends what you are measuring. If you are only interested in measuring variables associated with your 6 components, then I suppose you could conclude that the remaining items are not valid for your purposes.

Behacad
  • 4,916
  • 8
  • 30
  • 48
  • Thanks a lot Behacad! Well, the thing is, I am not trying to develop scales; rather, I want to use the motivation-variables as clustering variables to determine groups of consumers with similar patterns in motivations (say, group one scores high on motivational var. 1&3, medium on 2&4, low on 5&6 and so on). I used factor scores for cluster analysis, and it worked. But it bothers me that I didn't use the eliminated 15 items for the reasons you mention: they do not assess factors (components) but they describe something else, right? I don't know if/how I could use them in clustering... – desperate-about-statistics Jun 25 '14 at 20:39
  • It seems like not using the eliminated 15 items is not a bad choice since you do not know what they measure. You could interpret individual items and try to come to a conclusion, but otherwise it can be hard to say. – Behacad Jun 26 '14 at 00:21
  • 1
    @desperate, doing cluster analysis on components is wide-spread practice, but a bit risky one: sometimes clusters differ mostly on those junior components that you discard. So, why not cluster on all initial 40 items? – ttnphns Jun 26 '14 at 06:48
  • Thanks much guys! @Behacad, I had an assumption what the items would measure, but they did not prove to have the structure I'd expected, which is a bit frustrating... but data doesn't lie, does it?) – desperate-about-statistics Jun 26 '14 at 13:18
  • @ttnphns, I guess I will try that for comparison, thank you – desperate-about-statistics Jun 26 '14 at 13:19