i've restructured longitudinal data from 3 waves of repeated measures to 15 age repeated measures as per bollen and curran's (2006) synthetic cohort approach. obviously this has created massive missing data and it also has created cells in the covariance matrix with zero coverage (for instance there are no cases with data in the first and last of the fifteen ages). i'm having trouble estimating a latent growth curve model (SEM) from the data. is it, in principle, possible to estimate a SEM with coverage=0 for some cells in the covariance matrix? i'm curious if it is a) not possible due to a limitation of the estimation algorithm, b)statistically impossible or c) possible but i'm missing something (e.g., too much missing data, a problematic pattern of missing data etc...). the functional form of the model i'm trying to fit is fine--i know from fitting it using a mixed model approach.
Zero coverage is not a problem when it is missing be design but low coverage is. This is probably what you are experiencing. You could try the multiple group approach to multiple cohort data as shown in Example 6.18.
thanks for your response, very helpful as always. a follow-up question--i don't think that the multiple group approach is necessary because i am not concerned about cohort differences. having read ex. 6.18, it seems that another way to model this is just to rearrange the data from measurement occasion to age using the DATA COHORT cmd. i've already done this rearranging, so i don't need this cmd. however, as you mentioned, this reorganization results in data which is missing by design. how do i indicate that in the input file? some combo of the the MISSING option of the VARIABLE cmd and PATTERN option of the VARIABLE cmd, correct?
in sum, am i correct in thinking that i should, in principle, be able to estimate latent growth curve (SEM) model that has quite a bit of data missing by design as long as i indicate the characteristics of this missing by design pattern in the input file?
If you take the approach of rearranging the data, there are no extra commands needed. The multiple group approach can be used if you have very low coverage where the missingness is not by design. Example 6.18 will result in the same results as stringing the data out.
Something you might want to think about regarding comparing the cohorts is that if they are different on an important set of background variables, for example, that would imply they do not come from the same population which would make analyzing them together incorrect.
thanks again for the tips. one further question--are you certain i don't need to specify the missing by design patterns in the input file? otherwise i have zero covariance coverage in some cells (by design). should i handle this by setting: ANALYSIS: TYPE IS missing meanstructure; COVERAGE = 0; ?
I’m applying a latent growth model on a longitudinal dataset of 30 waves. In my perspective, the dataset is kind of a „multiple cohort-sequential design“. Therefore, the covariance coverage is very low (not all individuals are assessed at the same time points). When taking the whole dataset, the model does not converge. Is there a statistical rule determining the selection of waves to include in the analysis?