

Longitudinal analysis with varying sa... 

Message/Author 


Hi, I am doing some analyses involving crosslag regression and latent growth models. We have converted the data from wave to age so that each variable is separated by what age the participant was when they took it. As a result, we have some variables that have a very low N, as well as in correlations involving these variables and other variables. I am suspecting this might be a problem in analyses. I know that MPLUS uses FIML by default, but in some of these cases, the original correlation has an N of 15, so I wonder if FIML is much help (our sample size is almost 500). So, when estimates are given, does it account for the fact that some correlations likely are quite off and weight them lower in minimizing chi squared? Or when calculating indices of fit (i.e. SRMR), does it weight all of the elements in the covariance matrix equally? We have seen some really large SRMR values (about .15) in models with nonsignificant chisquared. If not, is there any way to run the model so that it accounts for this? Or would it be better just to drop these timepoints? 


It sounds like you represent the outcome for each age as a column in your data using a wide approach. It doesn't matter that some coverage entries are low for some pairs of variables because your growth model parameter estimation is based on covariances between many pairs (some of which have high coverage). I don't think this is related to your SRMR value. 


Yeah, we have the data as wide. SRMR is basically calculated by averaging the residuals (well technically squaring them and then square rooting the whole thing), right? So, for one of the models, our SRMR is. 096. Most of the elements in the residual matrix are between .02 and .02, with a few around + .05. However, for the elements involving covariances where our sample has very low N, the elements are quite high, ranging from + 0.15 even up to + 0.5. However, when I take out the latest age (20) and just run crosslag from 15 to 19, the SRMR is only .043. The other fit indices really improve as well. Looking at the residual matrix, there are no longer a bunch of residuals >+ .1. So taking that age out clearly effects our model. However, we would like to use all ages if possible. Is there a way to fix this? 


You may want to modify the model part for that age. Request Modindices to see which adjustment is needed  this is more to the point than going by residuals. 

Back to top 

