I have attempted to run latent interactions with the XWITH command, but this gives me a fatal error saying I do not have enough memory on my system to run it. I am unable to obtain a new computer in my workplace as of yet. As an alternative, I have done the following:
1) Exported the factor scores from my model without the interaction effect in the ON statement (Model 1), and manually created the interaction term from the factor scores.
2) Created a new MPlus input file including the "hand-calculated" interaction term.
3) Re-run the model, and include the interaction term in the ON statement. (Model 2) I have run this model both letting the factor loadings be freely estimated, and constraining the factor loadings to be equal to Model 1.
Both of the new models with the interaction term run, and fit well. The model with factor loadings constrained to be the same as Model 1 fit better than when factor loadings are freely estimated.
What I need is your expert opinion on whether this method is mathematically sound (i.e. no significant error introduced and parameter estimates reliable), as an interim method of modelling an interaction effect to present preliminary findings while I am searching for a machine capable of running it with the XWITH command.
Thank you very much for your time and input on this.
Thank you for your reply Dr. Muthen. I have managed to find a machine which I am able to run the interaction effects on. It runs fine at the between level. However, I am unable to run the XWITH command for the within level of my data.
It gives me an error "The resulting covariance matrix cannot be inverted. Please check your data or change your model." What would be the likely cause of this problem? My model looks something like this at the within level:
int | x1 XWITH x2 ; y ON x1 x2 int ;
It gives me the error even if I run it as a single level model only at the within level.
Dear Prof. Muthen, I am running a latent cross-lagged model with an interaction term. This interaction term is between a latent factor (W1_faci) and a time-invariant dichotomous variable (fintervention_master). This is done on two waves of data. The outcome variable is hope at two waves (W1=wave1; W2=wave2). The latent factors (W_Faci; W_hope) are full metric invariant over time. W1_supp_faci | W1_faci XWITH fintervention_master; W2_supp_faci | W2_faci XWITH fintervention_master; W1_hope on W1_supp_faci W1_faci fintervention_master; W2_hope on W1_supp_faci W2_supp_faci W1_faci W2_faci fintervention_master ; --> The SEM model without interaction showed a good fit. In the model with interaction term using the XWITH option, W2_supp_faci was significant on W2_hope.
As a second strategy, we tried to apply the same statistical procedure as aforementioned (with factor scores calculating an interaction term). The interaction term W2_supp_faci was also significant on W2_hope. However, the model fit was not good. The same model without interaction has a good model fit. How could it be possible that adding a significant path (interaction term W2_supp_faci ) decreases the fit of the model with factorscores a lot?
Is it better to use the XWITH statement or the factorscores method? What is the difference between both methods? Thank you very much in advance, Caroline