Hello everyone,
I'm using traj command in Stata to do group-based trajectory modelling based on Najin's method. Here is the code:
traj, var(V1 V2 V3) indep(time1 time2 time3) model(cnorm) min(1) max(3) order(2 2 2 2 2)
trajplot, xtitle(Time of Measurement) ytitle(YY)
In the above code, V1 to V3 mean three repeated measurements at three different timepoints, all of them are discrete variables with values 1, 2 and 3. The timepoints of measurement are time1 time2, and time3, respectively. However, there are some participants who only have two measurements in my sample (~ 10% of the entire sample). I tried from class = 2 to class = 6 while assumed all the shape of trajectories are quadratic, the BIC are gradually increasing with the increase of the number of the class. But when I tried order(2 2 2 2 2), there was one trajectory looked quite linear, which is different from what I assumed. Moreover, there were two groups looked quite quadratic, but while looking at the table of Maximum Likelihood Estimates, the p values of both the linear and quadratic terms of these two gourps are almost 1. There are similar situations in order(2 2 2 2 2 2). Does this make sense? Is anyone who has ever encountered similar problems?
Question 2: Does the sequency in order() matter? Because I tried different combinations with same number of class and shape of trajectories, like order(2 2 2 1 1), order(1 1 2 2 2), order(2 2 1 2 1), order(2 1 2 2 1), etc. The results of the models were different, some model could converge while others couldn't.
Question 3: I also tried to decrease the complexity of the model by assuming two trajectories were linear "order(1 1 2 2 2)", same situation happened. Although the BIC was better than order(2 2 2 2 2), there were two groups, one looked quite quadratic and another looked like linear. While both of their p values of both the linear and quadratic terms were almost 1. Why this could happen?
I would be more than appreciated if anyone can share their experience. Discussions are highly welcomed in this post!
I'm using traj command in Stata to do group-based trajectory modelling based on Najin's method. Here is the code:
traj, var(V1 V2 V3) indep(time1 time2 time3) model(cnorm) min(1) max(3) order(2 2 2 2 2)
trajplot, xtitle(Time of Measurement) ytitle(YY)
In the above code, V1 to V3 mean three repeated measurements at three different timepoints, all of them are discrete variables with values 1, 2 and 3. The timepoints of measurement are time1 time2, and time3, respectively. However, there are some participants who only have two measurements in my sample (~ 10% of the entire sample). I tried from class = 2 to class = 6 while assumed all the shape of trajectories are quadratic, the BIC are gradually increasing with the increase of the number of the class. But when I tried order(2 2 2 2 2), there was one trajectory looked quite linear, which is different from what I assumed. Moreover, there were two groups looked quite quadratic, but while looking at the table of Maximum Likelihood Estimates, the p values of both the linear and quadratic terms of these two gourps are almost 1. There are similar situations in order(2 2 2 2 2 2). Does this make sense? Is anyone who has ever encountered similar problems?
Question 2: Does the sequency in order() matter? Because I tried different combinations with same number of class and shape of trajectories, like order(2 2 2 1 1), order(1 1 2 2 2), order(2 2 1 2 1), order(2 1 2 2 1), etc. The results of the models were different, some model could converge while others couldn't.
Question 3: I also tried to decrease the complexity of the model by assuming two trajectories were linear "order(1 1 2 2 2)", same situation happened. Although the BIC was better than order(2 2 2 2 2), there were two groups, one looked quite quadratic and another looked like linear. While both of their p values of both the linear and quadratic terms were almost 1. Why this could happen?
I would be more than appreciated if anyone can share their experience. Discussions are highly welcomed in this post!
