03-30-2017 10:45 PM - edited 03-30-2017 10:51 PM
Hello Everyone, I am trying to fit a finite mixture model using proc fmm for a continuous dependent variable that is left at zero. Should I use dist=truncnormal(0,.) for this in the model statement.
When I used dist=truncnormal(0,.) for a single component model and compare results from proc qlim (for a left censored regression) the results do not match. If I use the estimates from proc qlim as the starting value for proc fmm and force no optimization (tech=none), then the log likelihood is different than under qlim. The code I used is below and data is attached (data is simulated - qlim recovers correct parameters).
proc qlim data=tobit plots=none; model y = x; endogenous y ~ censored(lb=0); run; proc fmm data=tobit; model y = x / dist=truncnormal(0,.); run; proc fmm data=tobit tech=none; model y = x / dist=truncnormal(0,.) parms(-10.27 1.05 3.57); run;