Hi All,
I am testing a series of CFA models (single factor, correlated traits factor, higher-order factor and bifactor model) using AMOS. This test was conducted to validate a translated questionnaire.
There are three items that should be reverse scored. However, whenever I insert the reversed items in any models, Amos output always showed 'iteration limit reached...' and the standardized factor loadings cannot be retrieved. When I used the original items (without reversing score), the Amos showed positive output which is 'minimum was achieved'.
My question is:
1) Are the items actually not representing the measurement scale?
2) Can we still report the model fit indices from the output that says 'iteration limit reached'? even some models showed acceptable fit?
2) If I decided to use the original items (without reversing the scores) because the Amos can provide the output 'minimum was achieved' (regardless the outcome of fit indices), am I doing it correctly? I decide to interpreted it as the need to use the original items in the future studies.
I attached the pictures of a single factor model with original items and reversed score items. Hope someone can enlighten me, thanks so much for the help!
This is somewhat unusual. I would have expected easier convergence after recoding (rather than the opposite). In principle, recoding should not matter, except for the sign of the loadings (positive vs. negative). Do the three items in questions have loadings that are of the expected sign (e.g., negative loadings when all other loadings are positive) in the original (non-recoded) solution that did converge? If so, then that original solution could make sense and could be retained if it also looks appropriate otherwise.
Maybe a simpler question (but in my experience not uncommon): are you sure that the reverse coding was correctly executed?
Rainer Duesing , thanks for the suggestion! By following Pallant's (2020) suggestions in her book (SPSS Survival Manual), I reversed the scores again which was taken from a raw data. I re-ran the analysis and obtained the same results.
Christian Geiser , thank you so much for the information. For the three items (without reversing the scores) do not demonstrate negative loadings. All of them have positive loadings.
I checked for reliability. The results indicated that the three items with reversing scores have negative signs (-) and below than 0.3 on a corrected item total correlation. Without reversing the scores, the items have positive signs (+) and above than 0.3 on the corrected item total correlation.
Would it be safe to interpret that these three items without reversing the score suit my sample (and the targeted population) best? I'm just thinking may be it is due to age difference or cultural difference?
Or should I consider to delete the items? Some of previous studies suggested to remove the items. I haven't decided due to I found the items without reversing the score showed positive results when conducting the CFA analysis.
Nur Afrina Rosharudin What you describe appears odd to me in multiple ways. I would recommend finding a statistician with expertise in measurement who can look over the results in detail with you.
Christian Geiser , I'll find the statistician.
Thank you so much for the suggestion and all the informative guidance.
Here's my (late) addition to this conversation:
I agree this seems very odd. Like @Rainer Duesing I have to wonder if reversing the coding might have introduced an error. It could be as basic as the missing data values getting reversed without then being redefined as missing. I would do some very simple univariate and bivariate comparisons for diagnosis. Is the variance (or std. dev.) of every original item identical to the variance (or sd) of the corresponding reversed item? Are the correlations of every original item with all other variables identical, except for sign reversal, to the correlations of the corresponding reversed item? If the answer to either question is No, then the reversed coding was done incorrectly. But if the answers are both Yes, then you'll need a statistician's help as suggested by Christian Geiser .
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size