Hi all,
I hope anyone can help me and save my life!
I have been trying to gather as much information as possible to decide which analysis method I should use for my collected data, but I can’t make a decision.
In particular, I am struggling to decide whether I should see my data as non-normal or normal, then I can decided which statistical method I can adopt.
For my thesis, I am currently working on analyses with the collected primary data (n=110) based on a cross-sectional, non-probability and self-reported survey for a theory-testing purpose.
All constructs are reflective measures and multi-item scales (between 3 to 6 items) across 2 IVs, 2 DVs, 2 moderators. .
1 IV is a higher-order construct (3 dimensions), and another IV is unidimensional (5 items).
All of them are seven points Likert scale that is previously developed and validated in other papers
Positive sides:
After the preliminary analysis, reliability (Cronbach’s a >.7), correlation, and regression analyses (OLS) show a good sign and are mostly in line with the conceptual and theoretical models from the role model papers.
Negative sides:
· The results of EFA (to check the unidimensionality) and CFA analysis (to test the model) were horrible. The EFA results were messy as per image below. Many of them are cross-loaded or loaded on unmeaningful components.
· KMO is >.850, suggesting that a sample size is not an issue?
· Basically, many OLS assumptions are violated because of leptokurtic distribution, heteroscedastic, and non-random sampling.
· Normality tests (e.g. Shapiro–Wilk test ) says all of them are nonnormally distributed. Besides, across all constructs, the Skewness and Kurtosis values are around -2 and +6.5, respectively. However, although I am aware of the threshold line, +- 2 for BOTH skewness and Kurtosis, I can chose to follow the Hair’s (2010) guideline, which suggests that +- 2 for skewness and +- 7 for Kurtosis for data to be considered as “normal”.
· Lastly, Common method bias is checked through Harman's single factor test (34% of total variance)
Based on this information, I have questions.
(1) Which regression model should I use? (OLS, PLS and so on). Please provide me with some justifications of choice (e.g. Sample size).
(2) Then, which statistical tool do you recommend? ( I am currently using SPSS and AMOS, but I can use R, SmartPLS or MPlus if needed.
(3) Should I consider my data as normal or non-normal?
Thank you so much for your help in advance!!
Ted
Hello Ted,
That's a lot of questions! I'll start with one of my own...
When you say, "reflective-based," do you mean "self-report?" I ask because the PLS-SEM folks talk about formative and reflective relationships among manifest and latent variables.
The EFA:
If your model posits that three factors (latent variables) are contributed to by a second order factor, then I don't understand why you would use in your EFA: (a) principal components extraction; and (b) varimax or any orthogonal rotation method. That doesn't make a lot of sense to me.
In the context of EFA, the KMO of .85 simply suggests that the data are reasonably good for factoring. By itself, it tells you nothing about whether the sample size was too small, sufficient, or excessive for the task at hand.
CFA/SEM:
At first blush, it sounds as if you would be better served by simply applying CFA to your data to see whether the proposed model works adequately to reproduce the observed relationships in your measured variables.
PLS-SEM does make somewhat weaker demands on the nature of the scores and is claimed to be more robust with smaller sample sizes. The usual issue of recommended sample size in SEM work depends on the number of observed variables, the number of latent variables, and the proposed number of relationships in the model. Probably, the highest confidence with respect to justifying a given sample size comes from having run simulations of the proposed model.
There are a number of software packages capable of working with the model you propose (whether via PLS-SEM, CB-SEM, or an ordinal-based SEM, as you indicate the manifest variables are ordinal).
Good luck with your work.
David Morse
Thank you so much for your comment!
To respond to your questions, the only reason I run the EFA with Varimax is that the original paper (the paper that developed the scales) did so when they were in the process of scale development. Thus, basically, I was copying their procedure. I am a beginner, so please excuse my ignorance.
As for the term “reflective-based”, it simply means it is a reflective measurement. Sorry for the confusion, and I have corrected it.
As you mentioned, I saw the argument on RG (the PLS-SEM folks talk about formative and reflective relationships among manifest and latent variables). But, I will see how my data goes with PLS-SEM!!
Again, thank you so much!
Ted
@Ted I would recommend to go with PLS approach in Smart-PLS as it doesn't require normality assumption and allows to add latent variables add as moderator and test the model as a whole. AMOS won't allow you to add latent variables as moderator and you would have to run path analysis for moderation taking composite variables and interaction variables (IV*moderator).
Imran Anwar
Thank you so much for your suggestion!
I have tried Smart-PLS and it's working great so far.
Best regards!
Ted
First identify which statistical Package is comfortable to you. Learn the Pros and Cons and then decide on the mechanism to follow. With regards to the normality test you can very well depend on the Scaling and reliability measure of SPSS or histogram with normal curve in case of individual variables. Get the basic clarity and then proceed.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis