REML estimator of variance components in frailty modeling

[Mehdi_R]

Thanks for your invaluable comments

[Rick_SAS]

Another thought: Replace the logsig parameter by s>0 and use

random W ~ normal(0, s) subject = cluster_id ;

Do you get equivalent results? Does the bias change?

[Mehdi_R]

Thank you Rick. Actually I got similar results. The relative bias is still between 20 and 70 percent!! (while it should be at most around 5%)