Dear Steve, thanks for your reply. I have a very large sample (n range from 5K to 32K for each country). So the hypothesis of being underpowered seems very unlikely. I have looked at the y distributions in each country. Both the average and the % of zeros is different across countries. In one or two countries the % of zero is not so big while in some other countries the % of zero is very big. There seem not to be overlap between countries that would justify the warning. I tried to run a similar model using proc FMM. title "Finite Mixture Model - Gamma, zero inflated model" ; proc fmm data =[DATA] gconv=0; class country; model Y=country/dist=gamma ; model Y=/dist=constant ; probmodel country; run; The model failed to compute the mixing probabilities This was not the case after running a zero inflated, log-normal model. title "Finite Mixture Model - Lognormal, zero inflated model" ; proc fmm data =[DATA] gconv=0; class country; model Y=country/dist=Lognormal; model Y=/dist=constant ; probmodel country; run; The logs were the same for both models: NOTE: Convergence criterion (FCONV=2.220446E-16) satisfied. NOTE: At least one element of the gradient is greater than 1e-3. NOTE: PROCEDURE FMM used (Total process time): real time 5.43 seconds cpu time 19.15 seconds This is strange because, I looked at the distributions of the Y using proc univariate and the gamma distribution fitted the positive part of the Y distribution quite well (at least as well as the lognormal, in any case). proc univariate data=[DATA]; var Y; histogram Y/ midpoints=uniform lognormal (theta=est) weibull (theta=est) gamma (theta=est) vaxis = axis1 name = 'Histogram'; inset n mean(5.3) std='Std Dev'(5.3) skewness(5.3) / pos = ne header = 'Summary Statistics'; axis1 label=(a=90 r=0); where Y ne 0; run; Any thought?
... View more