Dear all,
I've encountered a problem while doing SEM and I can't find a solution anywhere. Hope that someone here can help me with this. The problem is, that some standardised estimates in regressions are greater than 1 (part of output below) in my model. The sample isn't low (551 participants), but the data is non-normal. There is no multicollinearity. All the variances are above 0. Although, z-value for sex variable variance is above 6000, and I don't know what to make of it. I've been conducting the analyses in R (lavaan). Moreover, I have 3 observable variables (2 categorical) and 6 latents (see attached simplified model - we also assumed that sex/age/know are predictors for social/control/utility, but I did not include it in the drawing). Due to all that I have decided to use the WLSM estimator.
Now, I'm wondering if those values are ok, or if maybe I forgot to check for something or miss-specified my model.
Part of the regression output (with the values I'm not sure about in bold):
Estimate Std.Err z-value P(>|z|) Std.lv Std.all
social ~
age (d) -0.056 0.008 -6.921 0.000 -0.038 -0.521
reactance (e) -0.332 0.073 -4.527 0.000 -0.249 -0.249
severity (f) 1.299 0.111 11.667 0.000 1.061 1.061
sex (g) 1.411 0.174 8.123 0.000 0.961 0.481
know (x) 0.362 0.168 2.152 0.031 0.247 0.113
utility ~
year (l) -0.063 0.010 -6.072 0.000 -0.035 -0.479
reactance (m) -0.459 0.086 -5.360 0.000 -0.282 -0.282
severity (n) 1.588 0.129 12.319 0.000 1.062 1.062
sex (o) 1.590 0.211 7.532 0.000 0.887 0.444
know (z) 0.251 0.202 1.245 0.213 0.140 0.064
Thank you! I would be really grateful if someone could provide me some insight into what might have went wrong.
All the best
Radosław Trepanowski
Radoslaw,
please attach code or output as a .txt file with courier new as the font style. RG makes a mess out of it :)
And please attach a full path diagram with latent variables and their associated measured indicators and some background information on constructs, question wordings. All this helps and saves us time.
Best,
Holger
Holger,
I didn't notice that! I attach 2 files - full model + output (+ model specification in the file).
As for the measured constructs:
Reactance - psychological trait reactance measure with Hong Psychological Reactance Scale (HPRS; Hong & Faedda, 1996). 7-point Likert scale. Example item “I resist the attempts of others to influence me.”
Utility (attitude) - Beliefs about likely consequences of taking an action to counter a sickness (7-point bipolar scale ex. “harmful - beneficial")
Control - Beliefs about own control over taking an action to counter a sickness (ex. "It is only my decision to get vaccinated"). 7-point Likert scale
Social - Beliefs about how (important) others will perceive the participant in case of taking an action to counter a sickness (ex. "Most of the people I know would commend me for getting vaccinated). 7-point Likert scale
Severity - perceived severity of a sickness (ex. "If I ever got sick with it, it could be dangerous to my life"). 7-point Likert scale.
Intention - intention of taking an action to counter a sickness. (ex. "When the vaccine becomes available I will get vaccinated"). 7- point Likert scale
Sex - 0/1 female/male
Know - "Do you know anyone with this sickness?" 0/1 no/yes
+ Age
Hope that helps,
Thank you!
Radosław
Hi, thanks.
two direct thoughts.
1) The model does not fit the data. I would start with reducing the numbers of indicators for the reactance latent. These are too much and you have no chance that the rigidness of the model (all indicators only measure one variable) is correct. I would pick the 3 essential items. This provides validity chances (as you select those that match your understanding of reactance and those that plausibly match the common factor implications
2) The severity latent seems to be the problem. But we get to that later. In addition i would start with a CFA with freely correlating variables.
3) Estimate error covariances among control, utility, and social as otherwise you assume that all common causes are in the model. It may well be that this is the problem as severity is one of the common causes and the algorithm will try to upward bias its effects to account for the substantial correlations among control, utility, and social.
HTH
Holger
Holger!
Thank you! I did as you wrote, and it worked perfectly! Reducing the number of indicators in reactance latent helped a bit, but estimating the error covariances among control/utility/social changed all the values to the ones I expected.
The fit of the model also impoved by a lot.
Thanks again!
Radosław
lavaan is just in beta version, and complex, hierarchical model are sometimes reported to cause some issues (I've seen at least few similar topics on stat and stackoverflow, but I can't find them at the moment). Perhaps Yves Rosseel could be contacted himself?
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling