How do you fit a two-part mixed effects model in SAS? I have clustered data (crossed design) and the response variable is semi-continuous with many zeros and is bounded in [0,1). I want to fit a two-part mixed effects model to this data including the assumption that the two parts are correlated. So far I have tried fitting this model using PROC GLIMMIX. The outcome variable is y. First, to prepare the data for Proc Glimmix, I do the following data step to create a new string variable called "distribution" that indicates whether the observation is "Binary" or "normal" so as to fit two parts with different distributions. I have transformed the nonzero part of y using ARSIN(SQRT(.)). The variable "zero" is coded as 0 if y=0 and 1 if y>0. data data2;
length distribution $7;
set data1;
response = (zero = 1);
distribution = "Binary";
output;
response = y;
distribution = "Normal";
output;
keep ID1 ID2 response distribution x1 x2 x3 x4;
run; The variable "response" will then be used as outcome variable in PROC GLIMMIX as below; proc glimmix data = data2 method =rmpl noclprint pconv=1e-10 maxopt=100 ;
class ID1 ID2 distribution x1 x2 x3 x4 ;
model response(event ="1") = distribution distribution*x1 distribution*x2
distribution*x3 distribution*x4 / noint solution dist=byobs(distribution) ddfm=bw;
random distribution/ subject = ID1 type = un solution ;
random distribution / subject = ID2 type =un solution ;
nloptions tech=nrridg maxiter=1000 gconv=0;
ods exclude solutionr ;
run; This model, however, does not converge. Does anyone have an idea on how to fit such models or how to adjust the Proc Glimmix above? Thank you! NOTE: The code above was adapted from the article below but there they do not have crossed random effects: Baldwin, S. A., Fellingham, G. W., & Baldwin, A. S. (2016). Statistical models for multilevel skewed physical activity data in health research and behavioral medicine. Health Psychology, 35(6), 552.
... View more
