04-21-2017 11:09 AM
I have monthly claims data with 27% having zero value, and some monthly values exceeding $1M. Members can have 12 months of data. As a repeated measure, I'm using GLIMMIX to predict the monthly value where, in a data step to prepare the data, I'm using the statement:
if COST = 0 then dist = "POISSON"; else dist = "GAMMA";
This simulates a Tweedie distribution.
Then in the model, I have:
proc glimmix data = DSN method = quad;
class month dist plan;
model COST = month plan / solution DIST=byobs(dist) ;
random intercept / subject = member TYPE = Toeplitz ;
estimate "Plan vs No Plan" intercept 1 plan 0 1 / E ilink;
estimate "Month 1: Plan vs no plan" intercept 1 month 1 0 0 0 0 0 0 0 0 0 0 0 0 0 plan 0 1 / E ilink;
E and ILINK provides me with a mean estimate but it is not in scale of the COST. What is the back transformation from the estimated mean to recover the actual cost?
04-22-2017 01:51 AM
Try not use ilink. estimate "Plan vs No Plan" intercept 1 plan 0 1 ; or try lsmeans lsmeans plan/ilink diff exp cl;
04-24-2017 11:08 AM
I switched EXP for EST in the LSMEANS statement and it worked. Thanks for the tip.