Qiwen Zheng posted on Thursday, June 16, 2016 - 1:40 pm
Anyone know how to specify a log-normal distribution instead of the default normal distribution of the within-person residual?
My model is a two-part random intercept model, the data is longitudinal with each person with 8 repeated measures. I want to compare model-data fit under normal v.s. log-normal distribution for the within-person residual e_ij.
Below is the model for the continuous part: m_ij=a_0i+a_1X1_ij+a_2X2_i+e_ij a_0i=a_0+b_0i where i is person and j is time.
My Mplus code: u is a binary dependent variable and m is a continuous dependent variable. Both X1 and X2 are time-varying covariates.
VARIABLE: Names are ID u m X1 X2; Missing are all (-999) ; usevariables are ID u m X1 X2; cluster = ID; within = X1 X2; categorical=u; ANALYSIS: type = twolevel random; estimator=ml; MODEL: %within% u on X1 X2; m on X1 X2; m*; %between% u with m*; u*; m*; [u$1*];[m*];
You can perform the log transformation in the define command and then compare the likelihoods. They are on different scale however - you need to add Sum Log (data) to put the likelihoods on the same scale. It is not possible to explicitly model only the residual using log-normal.
Some other possibilities are available through the constraint = command as in user's guide example 5.23.
Qiwen Zheng posted on Thursday, June 16, 2016 - 8:03 pm
Thank you Dr. Asparouhov! I now understand that I should do log transformation to the dependent variable.