01-16-2017 06:50 AM
I am fitting a bayesian piecewise exponential model to survival data. I am using the in-built gamma prior. I specify the (piecewise) intervals on which to get the lambdas . In the example below, the intervals are [0,10),[10,Infty). I run the same type of model for many segments in my data, using differnet intervals. For all my datsets, the number of "events" overall is very low. I am facing some issues...
For all my models, I find that the initial values for the lambdas are exactly the same as the final MLE estimates. Is it because of the low-event nature or is there something else wrong - should I experiment with more seed values- say 5 different ones- and then take the average of the MLEs for my lambdas? How does SAS choose the initial values for the lambdas ?
|Initial Values of the Chain|
|Maximum Likelihood Estimates|
|95% Confidence Limits|
Here is the ESS diagnostic: I read somewhere that ESS should be around 1000.
|Effective Sample Sizes|
proc phreg data=seg1;
bayes seed=1 piecewise=hazard (INTERVAL=( 10) prior=gamma );
Note: I don't use any explanatory variables in my model.
2) I am not sure how to specify a Gamma(a=0.001,0.001) prior --- the syntax I used below resulted in error messages: piecewise=hazard (INTERVAL=( 10,20) prior=gamma(a=0.001 b=0.001 ))