Hello! I’m having some problems applying a Structural Equation Model. I have 18 variables and 110 replications. Several variables are non-normally distributed (both endogenous and exogenous), some of which are zero-inflated.
I started using piecewiseSEM, but I had to discard this approach because I need to represent cyclic relationships (i.e., A-B-C-A) which cannot be evaluated through this method. Even without specifying the cycles, and simplifying the model as much as possible (including just 3 variables), I get a poor model fit and a P-value much lower than 0.05. When using Lavaan, although I can include cyclic relationships, I get the same problems with the model fit and P-value. What may be the reason for this?
I cannot change the structure of the model (relationships among variables) without it becoming biologically absurd. I’d like to know (1) if it would be possible to improve the model fit by applying some transformation or standardization?—I’ve already tried standardizing the data, but I get the same problems; (2) Whether it is correct to use zero-inflated and/or Poisson variables with the Lavaan function?—The model accepts the variables as they are, and we get a result, but a poor model fit; (3) Would it be correct to split a model to meet the number of minimum replications versus the number of estimated parameters?
Thanks in advance for your response. Kind regards.
Hello Alba,
there is a lot do untangle here (the model structure, the variables, the measurement design/set up). So I would start with knowing
a) what's the design of the study (cross-sectional?)
b) what's the model structure. There are two possibilities: you can upload/attach a path diagram or you can use the great software dagitty.net to draw a path diagram, then to "publish on dagitty.net" (under "model"). This gives you a link that you can directly copy/paste into here.
Two comments
1) when your DV is a count variable (or any other with non-gaussion errors) or when the effects are nonlinear you cannot use lavaan. Mplus offers options to use other DVs and if the effects are nonlinear, I would use a GAM
2) Cyclic effects are difficult to estimate in case you have cross-sectional designs, as you need to have instruments for the relevant variables.
Best,
--Holger
See these links:
Article Improving Fit Indices in Structural Equation Modeling with C...
https://core.ac.uk/download/pdf/297019805.pdf
Article “How Well Does Your Structural Equation Model Fit Your Data?...
Article Building a Better Model: An Introduction to Structural Equat...
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling