Thank you for your reference to PROC FMM, which is also new to me and therefore very interesting. I looked at your example and also read this paper https://support.sas.com/rnd/app/stat/papers/abstracts/328-2012.html in which the link to the hurdle-model became clearer to me. However, I am not able to include a random effect for my repeated measurements in PROC FMM, since it is only for use with fixed effects, am I right? Furthermore, the ref statements do not work. Is it possible to change the reference categories in PROC FMM? Thirdly, if I just look at the fit statistics of the two models, NB and ZINB, in PROC FMM, then my Pearson statistic suddenly gets larger. Why this? Added: This would be my mixture model proc fmm data=FV1_2016;
class Pool ZP Temperatur;
model Keimzahl = Temperatur|ZP / dist=negbin; /* not truncated, zeros can come from both components */
model Keimzahl = / dist=Constant; /* constant distr for the zero group */
output out = fmm_ZINB residual pred;
run; Unfortunately, the residuals do not look that good either: After some considerations, I came to the following conclusion with my approach in GLIMMIX: I just need a random effect for the repetitions, i.e. a two-way model with paired samples and repeated measurements. This should work with PROC GLIMMIX DATA=FV1_2016;
nloptions maxiter=1000;
CLASS Pool ZP Temperatur;
MODEL Keimzahl = Temperatur|ZP / s dist=negbin;
random _residual_ / subject=Pool type=UN;
LSMEANS ZP|Temperatur;
RUN; Do you think that it is sufficient? Just weird that it only converges with UN or ar(1) as covariance structure, which does not make it any easier for my idea with PROC NLMIXED. 😞 How can I compare both the UN-model and the AR(1)-model in terms of AIC? Actually, I do not have equal spacing, thus I would prefer the second model. I really appreciate your feedback on my thoughts and questions.
... View more
