|
Message/Author |
|
Yuchun Peng posted on Wednesday, February 16, 2011 - 7:48 am
|
|
|
Hi Bengt and Linda, I am running LCGA models with binary variables (call b1-11- 11 time points). As I know, Mplus defaults the thresholds of the binary variable at the 11 time points. My questions is 1) Does it make sense to try time-specfic unequal threshold and cluster-specific unequal threshold? Will it improve the model fit? e.g. time-specific unequal threshold %C#1% [b1$1] (1); %C#2% [b1$1](1); e.g.cluster-specific unequal threshold %C#1% [b1$1] ; %C#2% [b1$1] ; 2) If it does make sense, is there any test I can do to make sure I need this complex model? 3) The linear quadratic growth factors for Cluster 1 are not significant? I know they are meaningless for binary outcomes, but do I still need to specify them as 0? Thanks Vicky Thanks |
|
|
Using time-specific thresholds would not give you a growth model because a growth model assumes invariant measurement across time. Instead of cluster-specific (that is, latent class specific) thresholds you should use cluster-specific growth factor means. |
|
Yuchun Peng posted on Wednesday, February 16, 2011 - 1:26 pm
|
|
|
Hi Bengt, Thanks for your answer. But, two things I am not sure are that 1) do I need to specify a zero cluster growth factor if it is not significant? The second thing I am not sure, how these clusteer-specific growth factor means link to thresholds ? How they derive estimated probability for each category? Thanks for answering my basic questions, Vicky |
|
|
Unless you have theory for it, I wouldn't get rid of say a quadratic growth factor if it is insignificant in one latent class. Just report this. Your last question takes a latent class course to go through well - you should study the literature and our videos. Briefly, take for example the intercept growth factor mean [i]. This influences the probability of say a binary item at the time point where the time score is zero. It does so via the logit as logit = -threshold + [i]. |
|
Back to top |
|
|