Gustaf,

the use of parcels is highly debated and honestly, I think you are not only fooling yourself but the reader of your work. Validity and measurement are based on rigour and transparency--both of which do not exist when using parcels. As initial scales are in 90% of the cases "multidimensional" --which more precisely means (and clouds) that the items measure different attributes--so will the parcels be. You may end up with a fitting measurement model but in that case, it will be unclear what the latent variable will actually represent. IMHO it is even preferable to add all the items together in one composite score and noting that the variable is an index of potentially different entities as this makes that at least clear.

When you strive for valid measurement and reasonable latent variables, however, there are two options:

a) You start with thinking clearly about theoretically wanted latents and their meaning and select those 2-3 indicators which almost perfectly fit these assumed latents (in terms of exchangeable reflections of the same underlying entity). Even these 2-3 items will allow a rigid test and give you at least a chance of having a causally correctly specified model. Further, if that works, you will end up with crystal clear indicators, strong loadings and the best chance to get what you wanted to measure.

b) The more tedious "bottom-up" approach would start with using all available items and then using CFA or EFA to explore the likely latent structure. This will require some kind of replication--preferably an "enriched replication" with the addition of additional validation criteria that add new testable implications but also (by means of their relations to the explored latents) provide some validation. I would argue (however having only modest statistical background) that this enriched replication does not require new data but could be done in your exploration data as these new variables were not involved in the exploration. But reviewers will expect that hence better be prepared.

If this is unclear, here I presented this idea in more detail:

https://www.researchgate.net/post/Can_I_use_EFA_before_SEM_analysis

I personally would go for the first approach as this reflects my identification of theory-based measurement.

HTH

--Holger

Holger Steinmetz thank you for your response. You've given me something to think about. For me, and I might be coming at it from the wrong angle, but the reason I wanted to use a SEM approach as opposed to a regression model, was that I wanted to take into account somehow the fact that items are better or worse indicators of the theoretical construct (whereas regression models treats all items are equally representative of the theoretical construct). Although not perfect, I figured a parceling method would be a suitable compromise given that using all items would be untenable in SEM, but you make some good points there.

If I were to go with your option a) and extract the 2-3 most theoretically valuable indicators, are there any publications yet to support such a method?

Also, I do not really have any items that are more or less "theoretically" precise or meaningful, since the scales I need to shorten are scales which measuring awareness as well as fear of a list of anxiety-related bodily sensations, none of these sensations being more relevant than the other (at least, I don't have a theory about it). Would it make sense to perhaps split the sample in two, using one half with an EFA/CFA to extract the 2-3 indicators with high factor loadings, and the other half in SEM using these indicators?

Hi Gustaf,

the gist is that you have to be able to estimate the error variances and, thus, correct for measurement error. This however (and of course) requires that all items measure the same latent variable (I recommend to avoid the term "construct" or at least use it with some kind of internal alarm bell as constructs can be anything from simple to complex). Reflective indicators of the same latent are exchangeable and if each indicator is a valid indicator each subset of several indicators is also a set of valid indicators. Often, the call for more indicators clouds the multidimensionality of the construct (sic!)--that is the presence of more than one latent that represent the construct. This, however, is conceptual issue---no measurment issue.

A source for this view is

Bollen, K. A., & Lennox, R. (1991). Conventional wisdom on measurement: A structural equation perspective. Psychological Bulletin, 110(2), 305-314. doi:10.1037/0033-2909.110.2.305

Further, if you allow a self-citation, we did that in study 1 of a validation paper:

I described the paper here:

https://www.researchgate.net/post/Model_fit_measures_in_CB_SEM

However, please that reviewers will be upset as using only 2-3 indicators out of scale of 10 or so will trigger them. I tend to simply note that I used 2-3 out of the instrument by.....This always works.

Beyond the 2-3 indicators issue, there is a discussion whether even only 1 key indicator is enough although this would use some additional assumption (which I still see as advantageous in contrast to misfitting multiple-indicator models:

Hayduk, L. A., & Littvay, L. (2012). Should researchers use single indicators, best indicators, or multiple indicators. BMC Medical Research Methodology, 12(159), 1-17. doi:10.1186/1471-2288-12-159

Savalei, V. (2019). A comparison of several approaches for controlling measurement error in small samples. Psychological Methods, 24(3), 352–370. doi:10.1037/met0000181

Onyskiw, J. E., & Hayduk, L. a. (2001). Processes underlying children's adjustment in families characterized by physical aggression. Family Relations, 50(4), 376-385. doi:10.1111/j.1741-3729.2001.00376.x

Bergkvist, L., & Rossiter, J. R. (2007). The predictive validity of multiple-item versus single-item measures of the same constructs. Journal of Marketing Research, 44(2), 175-184.

Holger Steinmetz if I wanted to follow your suit and choose to extract 2-3 indicators of particular relevance from each scale, but from an a priori standpoint I have no good theoretical reason for choose one over the other (like you said, my indicators are reflective indicators of the latent variable). What other criteria would you apply in choosing these indicators in this case?

Hi Gustaf,

if you have a clear understanding /hypothesis about the latent variable you measure, there should be a possibility to select those indicators which most closely reflect the latent. Where's the problem?

If you like you could post it here

a) select one latent (just as an exercise)

b) describe the meaning of the latent

c) post the reflective indicators that should be appropriate candidates.

Best

--Holger