Message/Author 


In my timeseries data, there are occasional missing data. I am basing my model on example 9.37 (pg. 378). Should I be using TINTERVAL to handle the missing data (and resulting unequal interval sizes between measurements) OR can my TIME variable be missing values (e.g., 2,5,6,7,8,10). Thank you in advance, Brandon 


Use Tinterval. Time must not have missing data. See also the Short Course Topic 12 video and handout, slide 39 in Parts 3 and 4. 


Hi, I have a question regarding the tinterval command with missingness. Specifically, our tinterval command uses a time variable that is subjectdefined (i.e., the time variable indicates how many hours since that subject’s first occasion of measurement). However, this results in a problem for those who have their first occasion of data missing, because the time variable cannot be calculated for any other occasion of measurement for that individual. We thought of two possible solutions to this issue: 1) Subtract the time point for each person with missingness at time 1, making their actual first measurement occasion time 1, even if that occasion was not the first attempt. 2) Take the average time from time 1 to time 2 for those subjects who don’t have missingness at time 1, and manually impute that time at time 1 for subjects missing at that time (e.g., if that average elapsed time from t1 to t2 is 3 hours, we subtract 3 hours from t2 for all of those with missingness at t1). We are leaning towards solution #2, because it seems to presume the least. One important consideration might be the collection method, which is that the occasions of measurement occur at random points within fixed 3hour time slots across the day. We were wondering if there is any precedent for this issue or if there is anything we can do? 


Unless you are using a crossclassified model, both methods 1 and 2 will be equivalent and should yield identical results. Method 1 is what Mplus does automatically for you, however even if you do Method 2, Mplus automatically removes all missing data rows until it reaches a nonmissing observation. Generally we consider the first observed value to be at time t=1. I would recommend looking into the data transformation that Mplus does behind the scenes. Add the command savedata: file is 1.dat; and examine the processed data that is being used in the estimation. The DSEM model is stationary and generally it would not suffer from missing first observation. 

Chang Liu posted on Wednesday, February 13, 2019  12:58 pm



Hi, I have a panel data assessing children from 9 months to 8 years (with unequal time interval). Children were assessed at 0.75, 1.5, 2.25, 4.5, 6, 7, and 8 years. When I use TINTERVAL, should I specify time interval to 0.75 or some other numbers to handle unequal time interval between assessments? Much appreciated! 


0.75 looks reasonable. You can also try 0.25. See also Appendix A http://www.statmodel.com/download/DSEM.pdf 


Hello, In a recent manuscript describing results of an EMA study we include several models using Bayesian estimation and the LAGGED function where predictors at time t predicted an outcome at time t+1. A reviewer suggested we use pattern mixture modeling to determine whether EMA compliance (whether people filled out surveys or provided missing data) affected our findings. I'm not clear about whether running pattern mixture models is actually something that would make sense in the context of the models we ran. What procedure(s) would you recommend for examining whether missing data influenced results in this situation? Thanks in advance for any guidance you might have. 


To explore nonMAR missing data assumptions, I would recommend selection modeling (see page 14 http://www.statmodel.com/download/DecRevPM.pdf) which is easier to incorporate in DSEM. You can add the missing data indicators and then see if there are any significant effects that yield nonMAR models. If you have a few missing data patterns you can run each pattern separately and see if there are substantial differences or via multiple group methodology run as a multivariate parallel. 


Hi, I am conducting a singlelevel timeseries analysis with some x, one mediator and one y. All variables have been measured annually. I have some missing data for x, the mediator and y. I am wondering what's the best strategy to handle missing data  can I use multiple imputation for x variables and then let FIML to handle missingness in the mediator and y? Or should I use TINTERVAL to account for those missing years in my response variables? 


Bring the x's into the model so that FIML can operate on all your variables. See chapter 10 of our RMA book. 


Ok, thanks! Any special reason for this advice (over MI)? Can plausible value imputation be used in DSEM? 


I would use ML whenever possible instead of MI because there are fewer analysis options when using the MIs. Yes, imputation data can be used also with DSEM. 

Back to top 