Hello;
I have been working on Theory of Planned Behaviour. My sample size is 750. I have been trying to analyze the TbP and Environmental Values with the Structural Equation Model by AMOS. But I am having problems with Perceived Behavioral Control. There are 4 items under PBC. The factor loadings in the model are low ( ,72 ,53 ,55 55) and the AVE value is very low also. (0.36). CMIN/df, AGFI, CFI and RMSEA are acceptable in spite of this problem. (I have checked the modification indices too.) I have deleted the factors according to the AVE. but no it is not working. And because of this problem, the Structural Equation Model is not working properly in the next level of analys. However, one of the main variables of the TpB is PBC. What would you suggest? Should I remove the Perceived Behavioral Control from the model or?
Since your goal is to test a theory, I would not delete a latent variable just to achieve a better fit. You proposed a model based on a specific theory. I would test that theory and report whether or not it was supported by your data. If you feel a measure was not good, I would discuss that under limitations.
Hi Emre,
what exactly are your 4 indicators for PBC? Depending on how fine grained your indicators are, PBC may either be reflected by two factors (control and self-efficacy) or an overall "feasibility" factor.
Next, how is the fit of the model in terms of the chisquare test? The factor loadings, and especially the descending order could reflect that you try to push causally heteroeneous items in one factor model. I would take a look onto the items and pick those that are the most content valid ones.
Perhaps you can upload a path diagram.
Generally, SEM do not "work" due to three reasons (perhaps more but these are the most prevalent).
a) The imposed structure is wrong. This results in a bad fit (while I define "bad fit" as "the misfit that exceeds that expected by sampling error" and which is the target of the test
b) invalid measures. This may result in a bad fit but not ncessarily. Sometimes models with decent fit contains the wrong latents
c) Latents with no significant contribution/predictive value. This will result in a good fit in terms of chisquare test but low effects, Rsquare and a CFI.
Lot to inspect before you throw PCB out of the window. Even if you end up with a single-indicator-latent PBC variable, keeping this would be better than omitting it as the latter surely results in biased estimates of attitude and subjective norm.
Finally, with 750 individuals I would definitely do some multi-group analyses. Perhaps some relationships differ among subgroups which can cause problems, too.
Best,
Holger
Emre,
Holger already gave you some good advice. Here are my two cents: I cannot see a problem with PBC. You only have 4 indicators (items?), which is a quite small number of items, but the factor loadings are well in line with what you would expect from such a construct. The loading of .72 is quite high, by the way. AVE is quite low, but I would not worry about that. If the 4 items actually measure a single latent construct denoted as PBC, why would you want to throw away important information?
I would rather worry about model misfit. Which estimation method did you use? ML, MLR, or a least squares estimator? If you used ML, were the assumptions of normally distributed and continuous variables fulfilled? Did you check this?
Kind regards,
Karin
Since your goal is to test a theory, I would not delete a latent variable just to achieve a better fit. You proposed a model based on a specific theory. I would test that theory and report whether or not it was supported by your data. If you feel a measure was not good, I would discuss that under limitations.
Mrs. Karin Schermelleh-Engel Thanks for your reply. yes you are right. The PBC is really important factor especially in the Theory of Planned Behavior. I know there is something wrong. But i could not see it yet. I used ML as estimation method. No missing data and normally distributed (I checked it by boxplot.).
Mr. Paul E. Spector thank you very much for your advice. If I have not done something wrong with AMOS or analyzing process, I will probably do what you adviced.
Emre, your problem is not the PBC factor, but the lack of discriminant validity in your model. All factors are highly correlated, which means that they measure about the same construct and not different constructs. It would be necessary to investigate all measurement models from a theoretical point of view: Are these really indicators of the latent constructs that you intend to measure? This does not seem to be the case.
Kind regards, Karin
Hi,
there, beyond what Karin and Paul mentioned, I regard your numbers of indicators as way too many in order to realistically conform to the restrictions implied by the common factor model. If you push many heterogeneous indicators under one factor, not only the factor becomes an unrealistic entity, it overcorrects the latent for the measurement error which inflates the correlations among the latent variables. This may be the explanation for the high correlations.
The sign. chisquare test indicates that. As I said I would start with small subsets of indicators which makes it plausible that they are reflections of the same underlying entity. In my experience (and in this regard, I tend to respectfully disagree with Karin), more than 3-4 indicators almost never confirm to the implications of the factor model. The reason is a simple linguistic one: You cannot construct pure paraphrases of the same psychological entity with more then 3-4 indicators. Beyond that number, you have to necessarily incorporate semantic deviations that will measure something different. People do not function like machines without memory but will interpret something distinct in an item when the predecessors have focused on a similar information.
Best,
Holger
What is your sample size.Try the same model fit by increasing the sample size.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling