Thanks a lot… I was trying to test these using various methods. My original code with your modifications gives sort of correct values which I tested by using simple tobit-reg function and doing loglike minimization in scipy. Their fitted values for A and B matches within error bars. However, both these methods fail if the number of lower-limits data points are more than the number of uncensored data-points. Any suggestion for this situation?
Coming to using pymc.censored (the code I attached in my previous reply), it is not working even for fixing the lower and upper censored values. I do not understand what is wrong.