I am currently testing a structural moderation model in AMOS using the orthogonalization procedure as advocated by Little et al. (2006). I am using the Big Five personality traits as moderators, and I have built five different models, i.e., one model for each Big 5 trait.
Everything is working out fine, BUT something strange is happening for one of the Big 5 traits, Neuroticism. When I run the model with 2.000 bootstraps, the estimates and SE's of neuroticism on the outcome variables are extremely high. Neuroticism was measured with a 5-point scale, but the estimates range from 24-35, and the SE's from 97-147! To make it even more confusing, the Beta values are very high, ranging from .36 to .75 (!). The normal correlations in SPSS are fine, and an SPSS regression also shows much more logical outcomes.
I have checked everything regarding the parcels, etc., and everything has been calculated correctly. So, there is nothing wrong with the actual variables and the latent factors. These high estimates and SEs are not occurring with the other Big 5 traits, only with this one. Something seems to be happening in this structural model that influences the results. I have done a quick Google search, but was unable to come up with a good explanation of what is happening heren.
Does anyone have any idea what the problem is and how I could solve this? Many thanks in advance for your responses!
Jos,
Can you identify anything else different between Neuroticism and the other moderators? For example, substantially higher or lower correlations for Neuroticism with your other predictors or outcomes?
Also, can you tell us what the estimates are like for the other four so we have a standard of comparison?
Pat
Thank you for your response, Patrick!
The correlations between neuroticism, IVs, and DVs are very comparable for all the Big 5 variables, and are not clearly higher or lower than the others. These correlations between neuroticism and the other variables are very normal and logical, as they are all negative and ranging between .20 and .45.
The regression coefficients in the AMOS model for the other Big 5 variables are normal. Only Openness shows a similar problem, with unstandardized estimates and SEs around 20 (not .20, but ranging between 18 and 24). However, for Extraversion the estimates are very normal. Unstandardized estimates range between .8 and 1.2, with SEs of approximately .10-.15. Conscientioussness and Agreeableness show a similar pettern without issues. Just to be clear: I'm talking about the direct effects of the Big 5 on the outcome variables here, not the moderations.
When I look closely at the standardized estimates, I see that they are actually very high. Even though the unstandardized estimates for extraversion are normal, beta values range between .50 and .93 (!). When I run a regression in SPSS, beta values range between .10 and .35. Also, these direct effects do not have anything to do with the orthogonalization procedures as they are part of a latent factor model where the Big 5 factors are represented by 3 parcels (in a seperate AMOS model for each Big 5 factor).
Does that give you any lead on what could be going on here?
Hmm. Have you looked at the raw distributions of your Big Five measures? I can imagine a weirdly-distributed Neuroticism variable (e.g., a floor effect and a long right tail) that might result in something strange with the bootstrap. Bivariate scatterplots might also be informative.
That's all I can think of at the moment.
Can you look at the results from individual runs within your bootstrap? Maybe you have outliers there, which are pulling mean estimates into extreme territory. Also, look at the variance of the Neuroticism factor. A beta slope is basically covariance (Y,X) / variance(X) partialed, so if the variance of the predictor became very small, you might get some funny values. I also endorse Patrick's suggestion to look at the distribution of your observed variables. If you have a concentration of responses plus a thin tail, a given bootstrap re-sampling of the data might miss the tail altogether, leaving you with almost no variance in the observed variable.
Thanks for your suggestions! I will look into this very shortly and I'll update the situation then. Hopefully I will find the solution.
Big SE and very variable B parameters sometime referred as jumping bees (look for the haywood case allso) are usualy caused by high collinearity. If you can't find bivariate collinearity you might have multivariate collinearity so check VIF (should be les than 10) or tolerance (should be more than 0.1). Multivariate collinearity happen when you have reasonable bivariate correlations but high correlations between the linear combinations of the variables in the model.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling