Tim Bednall posted on Thursday, September 06, 2018 - 11:04 am
I'm trying to estimate a ridiculously complicated longitudinal model, with 8 latent variables measured at 4 occasions. (It is a scalar invariance model, with corresponding factor loadings / intercepts constrained to be equal over time.)
The longitudinal model that I am using is a random intercepts cross-lagged panel model (Hamaker et al., 2015). I am trying to estimate a 3-way interaction between three of the time-invariant (trait-like) variables using the XWITH command. After what seemed like a promising start, the model failed to converge.
I considered a couple of workarounds, but I'm not sure how legitimate they are. The first was to calculate plausible values to represent the latent variables, and to use these in a follow-up analysis.
The second workaround was to estimate the model without the latent interactions, and to retain the model parameter estimates (using the OUTPUT: SVALUES command). I then ran the XWITH model using these SVALUES as fixed parameters, with the exception of the main effects, interactions, and residual variances/covariances of the variables predicted by the interactions. This enabled the model to converge, and the estimated main effects / interactions seemed to be sensible.
I would be keen to hear whether other people have encountered similar difficulties, and how best to deal with it.