I am interested in the growth of reading between first grade and middle school. Unfortunately, the measure changes its metric reflecting the age or school year specific range of the reading level. Would it be appropriate to standardize each score at each time point (eg., z-transformation) or are there better ways of dealing with that problem.
Dear Linda and/or Bengt, I am estimating a linear growth curve model over 4 time periods tracking development by age in middle-aged and older adults. Some data are missing by design (over the age of 76 people are skipped out of the survey) and some data are missing unintentionally (death, refusal, etc.). Using MCOHORT to rearrange the data causes deletion of cases to be listwise resulting in no valid cases (since all subjects will be missing data for years post-76). The COPATTERN option however seems to use a cohort-specific listwise deletion. Consequently, I lose quite a few cases that have some data, though not all possible data for their cohort. Does MPLUS allow me to construct a full information file with multiple cohorts? I am currently working with Version 2. Would V3 better handle the type of design I want?
bmuthen posted on Monday, August 08, 2005 - 8:23 pm
I would suggest using a multiple-group approach to handle the multiple-cohort structure. This way, you would use all available cases.
Heather See posted on Wednesday, April 16, 2008 - 4:57 pm
I'm conducting a latent growth model looking at development from preschool through first grade. In my dataset, three-year-olds and four-year-olds were sampled separately (with threes being in preschool for 2 years, fours for one year). Thus, for the threes I have 5 waves of data and for the fours, I have 4 waves of data. If I do a two-group model, can I have uneven numbers of waves of data? Or is it better to drop the last wave for the threes, and have 4 waves of data for all cases? Or, combine data and add a covariate for cohort/number of years in preschool?
In multiple group analysis, both groups are required to have the same number of observed variables so you would have to drop one variable to do this. I would suggest first analyzing each group separately to see if the same shape growth model fits for each group. If so, you could consider stringing the data out by age using DATA COHORT and use cohort as a covariate.
I am exploring a dataset which observes repetitive behaviours in children at 14 months, 24 months and 72 months (6 years)of age.
I also use gender and SDQ total score at 72 months (overall emotional and behavioral difficulties) as time invariant predictors. Although model seems to fit well(CFI=0.97, TLI=.98 RMSEA=0.05)I get the message: "WARNING: THE RESIDUAL COVARIANCE MATRIX (THETA) IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL VARIANCE FOR AN OBSERVED VARIABLE... PROBLEM INVOLVING VARIABLE REPS72W5[observation at last timepoint]."
Despite this the results make sense, except there is no R squared for the repetitive behaviour rating at the last timepoint. Is this a problem resulting in t2 being so distal from t0 and t1, issues relating to having only 3 timepoints or something else? Can it be solved by constraining the model in some way? Your advice would be much appreciated.
It sounds like you have a negative residual variance at the last time point. This makes the results inadmissible. You can try holding the residual variances equal across time. This is what is done in most multilevel programs.