Hello,
My dissertation uses a mediator variable to explore the relationship between three latent insecure attachment styles (preoccupied, fearful, and dismissive) and social media addiction. My survey used complete scales to measure attachment (RSQ Scale with 30 items, but only 4-5 items measure each attachment style), social media addiction (BFAS with 18 items that measure six dimensions of social media addiction), and the mediator has 17 items.
My questions include the following:
1. Can pre-existing scales be reduced to a few indicators to run the SEM analysis? It's my understanding that latent variables should have 3-4 indicators. The scale for the mediator variable has 17 items, which seems quite large to run the CFA.
2. What specific steps would I take to reduce my data before running the SEM?
Any help or guidance regarding where I might find more resources on this topic would be greatly appreciated!
Reducing scale items for Structural Equation Modeling (SEM) analysis is a common practice to create more parsimonious models and improve model fit. Here are some steps and considerations to help you reduce your scale items before running SEM:
Remember that reducing scale items should be done thoughtfully and transparently. Clearly report the items retained and provide justification for their inclusion in your SEM analysis. Additionally, it's always a good practice to consult with your dissertation advisor or a methodologist to ensure that the decisions you make are appropriate for your research.
To the best of my knowledge, reducing or removing any item from a validated scale is not acceptable because it affects a part of the theory, which means that you are losing a portion of the developed theory for that particular scale. However, in some fields, especially in psychology researchers reduced a large scale to a new short version (For example: Health Survey (SF-36) to Health Survey (SF-12) , which is done based on a similar process for developing a new scale.
Yes, having 17 items on a single factor is probably not going to result in a well-fitting model. What you could do is either select only the best (most valid and most reliable) indicators ("marker" variables) or think about an item parceling strategy in which you would distribute the 17 items across several (e.g., 3) homogeneous subscales (parcels). For example, you can try to randomly assign the items to three parcels and examine how homogeneous the parcels are.
Christian Geiser just wanted to follow up with you regarding your answer. When you mentioned parceling items into homogenous subscales, this would mean running a single-factor EFA to obtain each item's factor loading. Based on those factor loadings, would a parcel be created by combining parcels with similar factor loadings?
I'm asking for clarification because I've read up on other strategies recommending combining the highest with the lowest factor loadings into a parcel.
For example, below is the factor loading for the 17 items that I would like to parcel into:
Component Matrix
- PSR10 .821
- PSR7 .795
- PSR9 .793
- PSR11 .784
- PSR13 .723
- PSR12 .721
- PSR17 .703
- PSR16 .685
- PSR6 .678
- PSR2 .666
- PSR1 .639
- PSR15 .635
- PSR4 .610
- PSR8 .590
- PSR14 .574
- PSR5 .552
- PSR3 .409
If I understand your answer correctly, I would create the first parcel by combining PSR 10 through PSR 17 (e.g., factor loadings between .7 - .8), the second parcel by combining items with factor loadings between .61 - .68 (e.g., PSR16, PSR6, PSR2, PSR1, PSR15, PSR), and the third parcel with the remaining items.
Thanks again for your input!
Running an EFA with a single factor and assigning items to parcels according to their loadings is only one possibility. Another approach is to simply randomly assign items to parcels.
When using the single-factor EFA approach, you would try to maximize parcel homogeneity by distributing items with similarly large loadings as evenly as possible across different parcels. For example, according to your ranking, you could assign items as follows:
PSR10 -> Parcel 1
PSR7 -> Parcel 2
PSR9 -> Parcel 3
PSR11 -> Parcel 3
PSR13 -> Parcel 2
PSR12 -> Parcel 1
PSR17 -> Parcel 1
PSR16 -> Parcel 2
PSR6 -> Parcel 3
and so on. That way, all parcels contain items with high, medium, and small loadings. Again, this is only one option. There are pros and cons to different parceling strategies as well as pros and cons associated with item parceling in general.
Thanks Christian Geiser for showing how this parceling option is be done in practice. Can you probably suggest a good paper were this and/or other parceling strategies (and their pros n cons) are described? So far I know the chapter of Sterba/Right in actual SEM handbook edited by Hoyle.
I am asking because I am doing an analysis of a already validated 11-item scale in a pre-post-test design and this can get easily complicate, eg because of error-covariances that need to be specified across the 2 measurement points. Reducing items to eg 3 parcels would make SEM-life much easier.
Gert Lang Take a look at the following paper:
Article To Parcel or Not to Parcel: Exploring the Question, Weighing...
Starting on p. 165, the authors describe different parceling techniques. The paper also contains a thorough discussion of the pros and cons of the item parceling approach in general.
See also:
Little, T. D., Rhemtulla, M., Gibson, K., & Schoemann, A. M. (2013). Why the items versus parcels controversy needn’t be one. Psychological Methods, 18(3), 285-300.
Little, T. D., Rioux, C., Odejimi, O. A., & Stickley, Z. L. (2022). Parceling in structural equation modeling: A comprehensive introduction for developmental scientists. Elements in Research Methods for Developmental Science.
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers