

Quadratic Effects and Multilevel Gro... 

Message/Author 


Hi, I have job performance data on individuals with varying numbers of time points due to turnover. Because of turnover, a multilevel random effects approach was more appropriate. I have reason to believe the change over time is nonlinear, but I don't know how to include a quadratic term. When I do the following: S Q  PERFORM ON TIME; I get an error message telling me I can only estimate one random effect at a time. I've tried it a few other ways with no error message, but get very different results. I eventually want to try GMM and survival analysis, but I need to figure the quadratic part out first. Any help would be appreciated. Thanks! 


It sounds like you have your data in the long format as shown in Example 9.16. If this is the case, you need two statements, one for the linear part of the growth and one for the quadratic and you need two time variables, time and time squared: s  perform ON time; q  perform ON timesq; If this does not help, please send your input, data, output, and license number to support@statmodel.com. 


Thanks for the response! I am wondering if you could help me with another issue. I have people in my dataset that only have 3 months of data and others with up to 9 (due to turnover). Within each individual, some missed measurement opportunities and have missing data. I have coded the missing data numerically, and for those that left the job, I left fields in subsequent months blank. I'm unsure if LGM (e.g. example 6.1) can handle the varying number of data points, or if I was correct in my original plan to use a multilevel design with data in the LONG format. How does Mplus handle the system missing data points? I realize there are probably several ways to go about doing this analysis, but considering your expertise in a wide range of growth models, your advice would be greatly appreciated. Thank you! 


I prefer the wide approach because it is more flexible, for example, you can obtain a residual variance for each outcome rather than a single residual variance. The difference in the number of measurement occasions is handled by missing data. Assign the same missing value flag to these occurrences as you do to any other missing data. The results will be the same as the long approach. 


Thank you for the advice. I wasn't sure how to handle the missing data issue. I thought it would be problematic to use FML imputation for measurement instances that occurred after a person was terminated/left the job. I have found that the wide approach is much more flexible for estimating the shape of the growth curve, so thanks again for the suggestion. 

RJC posted on Tuesday, June 04, 2013  9:06 pm



Hi, I have a question about Linda's response to the first post above. Linda explained that to add the quadratic term to a growth model using the long format, the following should be written in the WITHIN section of the model. s  perform ON time; q  perform ON timesq; When I run this, Mplus tells me that I need to define timesq. I tried to define it using the following. However, it doesn't work. DATA: FILE IS data2.dat; DATA WIDETOLONG: WIDE = y1y3; LONG = y; IDVARIABLE = person; REPETITION = time; DEFINE: timesq = time*time; Can you tell me how to define timesq so I can add the quadratic term? Thanks for your help! 


You need to put timesq at the end of the USEVARIABLES list. If you still have a problem, send the output and your license number to support@statmodel.com. 

Back to top 

