Hi Jun Yin,

testing and interpreting structural models makes only sense when the measurement models are correct. Whether the low CFI is based on a misspecification or simply the fact that there are no substantial correlations in your data, cannot be said. What was the chisquare test and df?

And perhaps, you could present more details (the path diagram of the model along with info about constructs and indicators).

Best,

Holger

CFI is one of the fit indices to consider when estimating a structural model. I assume you did a confirmatory factor analysis (CFA), to establish the validity of your construct. In doing so then you looked at the RMSEA, CFI, and TLI, as well as the standardized residuals in the covariance matrix and the correlations between the variables.

If you did so, you could rely on the following criteria with the reference I've mentioned.

- In the CFA, the standardized factor loading of items should be significant (p

Holger Steinmetz Thank you a lot! And the attachment is my data output.

my research aims to investigate how individual factors (intrinsic motivation, paradox mindset, task significance, resource available, proactive personality) in total five independent variables impact employee creativity mediated by unlearning(the process employees abandon old out-dated knowledge and skills to learn new ones).

Thank you a lot if you could comment some advice on my output!

Holger Steinmetz Thank you a lot for your reply!

My research aims to investigate how individual factors (intrinsic motivation, paradox mindset, task significance, resource available, proactive personality) in total five independent variables impact employee creativity mediated by unlearning(the process employees abandon old out-dated knowledge and skills to learn new ones).

The attachment is the diagram of model.

And about the construct

IVS:

Intrinsic motivation: 5 items

Cronbach's alpha: .888

paradox mindset: 9 items

Cronbach's alpha: .873

task significance: 4 items

Cronbach's alpha: .724

resource available: 6 items

Cronbach's alpha: .837

proactive personality: 17 items

Cronbach's alpha: .959

Mediator:

unlearning: I have two questions designed for unlearning.

The first question includes 11 items and I did reverse coding of 6 items.

Cronbach's alpha: .350

The second question includes 5 items and I did reverse coding of 2 items.

Cronbach's alpha: .389

DV:

creativity: 9 items

Cronbach's alpha: .905

All the items and scale were adopted from previous scholar's article.

And what I tried to do was adjust the items of unlearning. The reverse coding seemed can not work and if I just tested the items without the ones need to be reversed, the Cronbach's alpha could be better about .678. But will this really correct?

Thank you a lot in advance if you could give me some advice! Thanks!

Dear Jun Yin,

as the chisquare test (and the indices) show, the model is very misspecified. You have a ton of indicators for each latent (esp. the dependent variable). This is likely to be plausible as it implies (again) a ton of restrictions, namely that every relationship / correlation between each pair of indicators (within or across the latents) "runs" across the loading and latent variable (like a channel) or in other words: that there are not "left over" routes from one indicator to another. For instance, the model implies that there is a conditional independence between two indicators of the DV when you control for the joint influence by the underlying latent. When you'd list all of those impilcations, the list would surely go over 3-4 pages!. All of these implications are tested by the chisquare test at once and the estimations of factor loadings and effects are based on these restrictions.

There are the following options.

a) Forget the CFA, go to EFA and go the exploratory route. This, however, would imply to draw another sample and do the full SEM (preferable with some enrichments (see here a discussion about the role of retest/cross-validation: https://www.researchgate.net/post/Can_I_use_EFA_before_SEM_analysis )

b) Stay on the confirmatory road I: If you still have a clear unidimensional / singular imagination of your construct (1 construct = 1 singular latent entity) than pick 2-3 items that EXACTLY AS POSSIBLY fit your imagination of the latent of interest as a reasonable assumption of the latent being a common cause of the selected variables. To the model with those indicators. 2-3 Indicators embedded in a model still implies a huge number of testable implications. Still you can list the implications and you get a huge list. Hence, still a solid ground for the structure and validity of the model and measurement models

c) Stay on the confirmatory road II: You are not sure whether the constructs are one- or better multidimensional. Either you have some ideas and can enlarge the model (i.e., breaking a certain latent in to pieces based on theoretical considerations) -- then proceed as in "b" or you get the idea by inspecting the item formulations. Either way, selected 2-3 focus indicators and test the model again.

Each of these routes more or less implies incorporating some exploratory aspects which should be transparent and critically discussed.

HTH

Holger

Holger Steinmetz Thank you so much for your help! I will have a detailed a look and thinking. Thanks again!

Sincerely.

Jun Yin

Holger Steinmetz Thank you a lot for your help! And I'm so sorry to bother you again and would like to consult more from you.

When the mediation failed by using Barron&Kenny regression mediation test, I was doing two things: for my sample size is 107 and I had so many items in my questionnaire, so I collected more questionnaire. Now the sample size is 333.

And another thing I do is to do the CFA(sorry to bother you a lot for I was just started to learn how to do SEM, so I might lack of understanding.).

I tried Barron&Kenny again and the mediation seemed to be okay. And the attachment is the result. The first picture is using sample size of 107 and the mediation seemed to fail and the second picture is using sample size of 333.

Can I say that the mediation now works with sample size 333 and the reason could be the sample size simply? I might be wrong..and I will continue doing the suggestions you have given on EFA and CFA relevant issues.

Thank you a lot in advance.

Yours,

Sincerely.

Jun Yin

Hi Jun Yin,

For doing the mediation analysis, I guess you created composites of the items, right? If the CFA failed (which it did), the composites are sums of different things meshed together. This means, that--even if your proposed causal structure (the mediation structure) holds, the estimates do not represent effects but averages of those effects implied by the hidden things you meshed together. Therefore, I would try (as aforementioned) go result in crystal clear sets of items that represent a single meaningful latent variable.

Don't try to achieve this with following the modification indices (as I saw in a different thread that you created). The MI are blind with regard to errors in the structure and their proposals only concerns parameters fixed to zero (loadings and error variances). If this is not the problem you will end in a fitting, albeit even more false model.

There is no other way than to reflect on your imagination of the intended latent variables (=what EXACTLY do they represent in the real world) and re-consideration of those indicators that best reflect these latents.

Best,

Holger

Holger Steinmetz Thank you a lot for your detailed reply.

My items are adopted from the previous published items from other scholars that they created in their articles to measure variable. And the Cronbach's alpha are good of each variable.

The model fit by using AMOS with larger sample size 333 is better compared with before, but still the key factors such as CFI, RIF are about .831 and .754.

And I will keep on thinking all the issues. Thanks.

the fitness index may not be achieved because of the low factor loading items, or the redundant of items in CFA process (showed by the value of modification indices, >15). To solve this, we have to delete any low factor loading item, or constraint the redundant items. Normally, we run EFA before run CFA to avoid the low factor loading items in CFA process.

Best regard

Dear Norliana Abd Majid ,

if you would allow me to politely disagree :)

A model misfit cannot be explained by low factor loadings. Misfit occurs because the model makes predictions of the data (due to its structure) that the data does not match with.

For instance, a pure CFA model has the essential assumption of "local independence" which means that all correlations among the indicators disappear once you control the latent variable. The often occurring modification indices pointing to error covariances reflect failure to meet this assumption.

Low loadings per se do not imply that local independence will not hold. In the extreme, you could have totally uncorrelated indicators which would lead to a FITTING model with zero factor loadings. The same happens in full fitting SEM with low or zero R-square. Note that prediction (in terms of variance explanation) is not the same as model fit (which implies that the models implications fit with the data).

However, I agree that low factor loadings will often reveal that the factor structure is wrong (which implies misfit). That is why an theory-based interpretation of loadings based on your personal measurement theory is so important (in contrast with rules of thumb as "loadings have to be " (insert some arbitrary and varying number that you got from the literature).

Best,

Holger