Hi,

overall the pattern of results strongly indicates that your hypothesis of unidimensionality is wrong. Covarying erorrs my lead to better fit (as does elimination) but will not make the overall (wrong) structure valid. Get a step back and simply take a look at the items. Then re-consider your picture of a possible underlying structure. Items varying in their question wording or connotation often measure different things. Laymen (participants) do not think in terms of "overall constructs" - they answer specific questions. One example that once strongly influenced by thinking occurred in a pretest when a respondent gave different responses to two items ("i feel tired" and "i feel exhausted") that - I thought - would measure the same latent variable (fatigue). When I asked about whey she gave different responses, she said "its obvious. Feeling tired is something completely different than feeling exhausted"...

The point is: the "devil is in the detail". Select indicators of a supposed underlying variable when they are psychologically so similar that it makes sense that they share the same cause (=latent variable).

HTH

Holger

I found that the scholar's opinion is divided on deleting an offending item from a measurement model. Some argue against deleting an item, citing the content validity of the whole measurement. I would be very careful in deleting key items that may change the meaning of the construct.

I also agree with Prof. Steinmetz that adding error covariances does not make an invalid measurement valid. However, I would like to highlight that it is acceptable to add an error covariance, if the error covariance is between items of the same construct (variable or dimension). Yet, I would always take a measurement model with a grain of salt whenever I see lots of error covariances.

Another point that was often missed is the reliability of the response itself, rather than the items. How consistent is the respondent? I usually include a consistency check in my questionnaires (e.g. reverse items), and exclude inconsistent responses (between the original and the reverse). I found it helps achieving model fit. Mind you, I am not encouraging you to cherry pick your data. The key point is reliability of your response (i.e. whether they really read the items when giving the response).

Thankyou Imam Salehudin for clarifying my query.U got my point.

In my research model I have 10 constructs and 48 items, there are a minimum of 4 and maximum of 6 items per  construct.

As of now I have added 4 covariances of error term and deleted 3 items to get a better model fit. 

Out of deletion and relation which approach is better to obtain a good model fit?

Personally, I would think deletion as the very last resort. I would not touch covariates between errors from different constructs. I may consider adding error covariates to items in the same construct only if: 1) The modification index identify a significant improvement to model fit, AND 2) the items' content are theoretically and logically related.

Before all that, though, I would check the data for inconsistent or patterned responses. I often found some zigzag or straight line response which obviously failed the consistency check. That kind of data really does ruin the measurement model.

Thanks for the insight. Assumptions before applying sem for normality,multicollinearity,and homoscedasticity were done before applying sem.I had covaried items of same constructs only.

.

We have just published an article about scale purification, i.e. the process of eliminating items from multi-item scales. We have used the example of SCM, but our framework can be applied to any other discipline. Download: https://doi.org/10.1108/SCM-07-2016-0230 (or request via my ResearchGate page).