i have a three time point longitudinal data.All students in time point 1 are in time point 2 and and all students in time point 2 are in time point 3. but in time point 2 and 3 more students were included, so there are students in time point 2 and 3 not in time point 1 and students in time 3 not in time point 2 and 1. WHAT IS THE REQUIREMENT FOR CROSS-LAGGED MODELS IN MY CASE.Should i ONLy use the students who were involve in all the three time points for a cross-lagged model or i can use the data as it stands.
A cross-lagged model requires a minimum of two constricts. You can have more. General questions like this are more appropriate for a general discussion forum like SEMNET.
hazel liao posted on Thursday, May 21, 2015 - 8:22 am
Hi~ I want to use cross-lagged panel to analysis my data. There is tow point time longitudinal data. One of the variable data is normal, so I used ML estimator to conduct CFA to determine the latent variable. The other data is not normal distributed, so I used MLM estimator to conduct CFA to determine the latent variable. However, I want to use these tow latent variables which are used different estimator in the CFA to conduct cross-lagged panel. The question is what estimator should I use when I use cross-lagged panel to analysis my data? ML? MLM?
ps. I had used the ML to conduct cross-lagged but the model fit are really poor. Then, I changed the estimator to MLM, the model fit are much better.
MLR is robust to non-normality of continuous variables. Categorical data methodology takes care of any floor or ceiling effects of categorical variables. Using WLSMV or MLR and the CATEGORICAL option takes care of this.
hazel liao posted on Wednesday, May 27, 2015 - 8:16 am
Floor effects means almost item's response are 1? Ceiling effect means almost item's response are 4?
hazel liao posted on Wednesday, May 27, 2015 - 10:35 am
Thank you ~~~
1. You said MLR is robust to non-normality of continuous variables. However, when I use MLR and the CATEGORICAL option at the same time, in this way categorical data could be analyzed?
2. I want to use cross-lagged panel to analyze data, however a latent variable is from CFA using MLR estimator, the other latent variable is from CFA using WLSMV estimator. In this case, what estimator should I use to conducting cross-lagged?
1. I am not sure what you are asking, but when declaring your variables as Categorical the non-normality robustness is not relevant. And you don't want to replace the Categorical statement with asking for MLR. Asking for Categorical and MLR is fine.
2. Either is fine.
hazel liao posted on Saturday, May 30, 2015 - 12:03 am
Thank you for your response!
I have try the WLSMV estimator to analyze the categorical data, but why there is no residual variance in the model result?
And where could I find the reference to interpretation of the Threshold?
I am running the following cross-lagged model and in the output Mplus is automatically providing covariances between the outcomes variables (Pos_T2, Neg_T2, and Dest_T2). Is there any way to prevent Mplus from automatically specifying these covariances?
Pos_T2 on Pos_T1 Neg_T1 Dest_T1; Neg_T2 on Pos_T1 Neg_T1 Dest_T1; Dest_T2 on Pos_T1 Neg_T1 Dest_T1;
I need to run two identical but separate cross-legged models - one for mothers and one for fathers from the same family and I want to see if their paths are significantly different from one another.
Ideally, I would use the grouping option, and then fix the parameters of interest to be equal across groups, then compare a model with the parameters constrained and then with them free using an equal fit test.
However, the two groups are dependent - mothers and fathers from the same family.
Is there any way to take this dependency into account?
You can put mothers and fathers in the same model to account for them coming from the same family. So if you have T times points and 2 repeated measures outcomes, you will have a data set with 2*T + 2*T columns (mother + father). So end up with 4 processes, one for mother and one for father.
You can then decide how the 2 processes should be correlated. And you can easily test equalities.
Hi, Thank you for your reply. I think I didn't explain myself - I have a mother-child model, and a father-child model, and I want to see whether the cross-lagged paths in the mother-child model are significantly different from the paths in the father-child model.
For each parent I have 2 variables (the same variables) and one child variable. I want to see whether there are bidirectional associations over time between each parent and the child, and to see whether these associations differ between the parents.
Unless the child variable has the same values for the mother and father model, that sounds like you observe 4 different (different in values) processes over time. But where this is structured as 2 model parts, each with 2 processes. This can be done in an extended cross-lagged model. If I now understand you correctly, we can discuss how.
Hi, The child variable does have the same values for fathers and mothers model - they are observations of the child and are separate from the parents variables which are parents self-reported questionnaires.
Ok, so then you have 3 processes: mother, child, father. So you have 3*T columns in your data. You can think of how to specify it in Mplus by first drawing a model diagram where in the top row, from left to right, you have the mother variable at the different time points, below that you have the row of child variables, and below that a row of the father variables. So 3 rows of variables. You have a cross-lagged model for row 1 and 2, and another cross-lagged model for row 2 and row 3. If you go by the RI-CLPM, you have 3 random intercept factors which are all correlated. This means that the dependence between mother and father variables is accounted for. This also makes it easy to test for equalities across mother-father.
Hi, I conducted a cross-lagged path analysis in Mplus
1. A reviewer is saying that, when testing a model without latent variables (path analysis), fit indices are about useless. What do you think? Is it supported by research?
2. Also, we wanted to control for the effects of demographics variables. We included pathways between the demographics and T1 variables, but the reviewers says we should have included the links from demographics to T2 variables instead. What do you think?
I'm working with a relatively small dataset of 296 individuals to look at cross-lagged effects across two waves of data between two constructs.
Given that my sample is large but not 'super' large, I decided to use a path analysis approach rather than a full structural equation model. The reviewers are commenting that I should have used a full model to account for measurement error.
My worry related to using a full model is sample size. The full model would include about 220 free parameters, while the path analysis model would include less than 50 free parameters (including several control variables in both cases).