

GMM with missing/truncated data 

Message/Author 

Adam Slez posted on Monday, October 01, 2007  6:29 pm



I am interested in using growth mixture modeling to identify types of career trajectories. Is this possible given that career lengths vary significantly across individuals? I'm not certain whether it is appropriate to think of variation in career length as either a missing or truncated data issue. Any suggestions on whether it is feasible to use GMM? If so, what is the best way to do so? What I would like to be able to do is to be able to identify career types defined by both the shape and the duration of the trajectory. 


Can you tell me a bit about the outcome as well as the different shape classes that you expect? 

Adam Slez posted on Tuesday, October 02, 2007  2:10 am



The outcome variable is a hierarchy measure which varies between 0 and 1. Essentially, what the measure does is rank all jobs or positions that an individual might occupy and then assign each job a score between 0 and 1 which is proportionate to its position within the overall ranking. The top job is assigned a hierarchy score of 1, the lowest a score of 0; all jobs in between receive a score proportional to their distance from the top. If there is a strict career ladder in play, we might expect a linear relationship between time and job rank. Alternatively, I think that it would be reasonable to observe stalled careers in which their is a nonlinear relationship such that there is a an early increase and then a levelingoff. The other dimension in play here is career duration. Some people have long careers and some people have short careers. Most of what I have read on the use of growth curve models assumes an ideal situation in which there are observations for all individuals at all points in time. In this case, the fact that there aren't observations at all points in time (i.e. short careers) is part of what needs to be explained. I wish I could be off more help, but I am still trying to figure out whether this can even be done with a GMM or LGM. 


It's an interesting question that seems to connect to both growth modeling and survival analysis. I don't think I have a final answer, but here are some thoughts. One approach would be growth modeling, either with one or several latent trajectory classes, where a short career is simply treated as missing data in line with "MAR" (see missingness lit.). MAR says that missingness is predicted by previous observed outcomes  for instance, a leveledoff development might predict missingness (leaving the career). Given great variation in career length, perhaps the growth modeling should not be done as a singlelevel multivariate approach but as a twolevel approach. Another approach would be survival analysis where you model the time to leaving the career, i.e. the career length. A third approach combines the above two. For example, certain early career shapes predict not surviving. Such modeling is not always, however, straightforward. Some related modeling ideas are given in the MuthenMasyn (2005) JEBS article on the web site under Papers, Survival Analysis. 

Adam Slez posted on Wednesday, October 03, 2007  5:31 pm



Much thanks! 

Back to top 

