In my model, I have a construct Environmental Concern (EC) with five items. During CFA, All these five items are poorly loading on EC (EC1 = .05, EC2 = .42, EC3 = -.07, EC4 = -.14, EC5 = .08). Can someone guide how to proceed further? I can not drop this construct as being the most important one in my study. Should I recollect the data? Any other way to improve these factor loadings? Also, what are the possible reasons for such poor loading?
Assuming that this finding is not due to a data error or a small sample, the most likely reason for these very weak loadings (only one is > .4) are covariances/correlations close to zero between the five EC items. Have you looked at the inter-item correlations? How large are they?
Low inter-item correlations could indicate heterogeneity of the indicators, that is, that they do not measure a single common factor and/or that they contain a lot of measurement error and/or specific variance. If that is the case, then the only way to solve the issue is to replace the items with better ones.
Another reason could be if the items have few response categories (e.g., 3- or 4-point Likert scale) with heavily skewed distributions. Skewed distributions can lower the inter-item (Pearson) correlations. You do not mention the response format or what type of CFA you used (i.e., for continuous or ordinal variables?). Ordinal factor analysis based on polychoric correlations addresses this issue to a certain extent. But I doubt that this would solve the issue completely--the loadings appear to be extremely low.
Christian Geiser thank you for the detailed response. I just checked you are right the inter-item correlations are close to zero (.036, .020, .010, -.059, -.008).
The response format that I have used is a five-point Lickert scale. All the items were taken from the same scale. The sample size is almost 1200 responses.
Hi Abdul Rehman Shaikh
given both the insufficient factor loadings and less than substantial inter-item correlations, I see but two options:
- discard the factor entirely, which is what I would prefer, or
- use the "best" single item from that factor in further calculations
Of course, there is always the option of starting from scratch, but oftentimes that option is not feasible.
Best of luck
Marcel
I agree with Marcel Grieger . These inter-item correlations provide clear evidence against the use of these items as measures of a single common factor. If you cannot redo the study with different items, your only choice would be to pick the theoretically "best" single-item measure from the current study. Perhaps there is an item that can be seen as a marker indicator based on its wording or prior studies?
By the way, these correlations seem extremely low (they are essentially zero). Are you sure that there is not an error in your data or analysis? The results seem odd. One would expect measures created to reflect the same attribute to be at least somewhat positively correlated. What correlations have been reported in prior studies for these items?
I agree with Christian Geiser last answer: in your case the more important question is "why do they not correlate" and not how to proceed. You constructed the items to reflect different aspects of the same factor, therefore it is totally unexpected and odd that they do not correlate in any way (btw, you should have 10 inter-item correlations with 5 items, you only showed 5....). So, either something went wrong in your analyses or you really have no idea how your construct looks like, if you assume that some aspects go together where they clearly have no correlation.
It would make it considerably easier if you shared the actual items of the problematic factor and also elaborated on the underlying construct itself. What is it supposed to capture? Come to think of it, I now tend to believe there must be something wrong with the calculations, not necessarily with the items, but we will see.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs