I am trying to estimate the mediation effects of multiple parallel mediators on a dichotomous outcome using complex survey data. Some mediators are continuous, and some are categorical. The predictor is dichotomous. The sample size is 1565.
I am using the WLSMV estimator. Because the mediators are not all normally distributed and due to the complex nature of the data, I am also bootstrapping using replicate weights. I also ask for model fit indices in the output.
When I run the model, I get the warning that the MOD option in the OUTPUT command is not available with BOOTSTRAP.
But I do get RMSEA=0.092, which I understand is outside the acceptable range of 0.06-0.08.
When I run the model without any bootstrapping, I get RMSEA=0.069, which seems ok, but CFI=0.445 and FLI=-0.428, which I understand in not good at all.
So my questions are: 1. Am I correct in using replicate weights and bootstrapping? And if so, do I just use the RMSEA value and not worry about the other fit indices?
2.Or do I use the not-bootstrapped fit indices and the bootstrapped standard errors? And if so, does it matter that the RMSEA values are so different between the bootstrapped and non-bootstrapped models?
This is my first foray into mplus-land and mediation analysis, so I am very grateful for any help you can offer me. Thank you very much for your assistance.
There is no assumption of normality for the mediator that you have to be resolving here. If the mediators are declared as categorical the mediation occurs via the underlying continuous variable - not directly through the observed value.
Using the non-bootstrap estimation you can get Mod to see where the issue in the model is. You can also use residual output to see that.
The difference in the bootstrap v.s. non-bootstrap shouldn't be much in terms of model estimation and fit. In principle it should not have changed RMSEA - it should only change the SE only, but the reality is that it does change it a bit because the weights (in the fit function for the different sample quantities) we use for the estimation changes. In the bootstrap case we use weights that do not reflect the complex sample (cluster and strata only sample weights are used). In general when the replicate weights are generated outside Mplus the strata and cluster are not even known.
Finding the model misfit is the way to go rather than rely on the estimator switch.