2
$\begingroup$

I came across the following toy example and am lacking a final answer/step to finish the analysis.

Imagine a surgery or medical procedure where we don't know the success-probability. It can be thought of as a Bernoulli processes (hit or miss) with unknown theta.

Now for the first patient the first try of the surgery is a success. I know how I can calculate the posterior when I start with a uniform prior.

But know I want to include a 2nd patient. For him the surgery only works at the tenth trial. So he has 9-misses before the first hit. I also know how I would get a likelihood function from that and with a uniform prior that would be my posterior.

But what if my prior is not uniform, but I have strong belief that for parts of the population the operation has a high success-probability but for another part of the population it only works in one out of ten or even twenty attempts.

I know how to produce a prior distribution (or grid) that represents that success-probability beliefe (and bimodal distribution, as a mix of the population of easy and hard patients).

What I don't know is how to calculate the likelihood function that strengthens that prior. I have two independent observations here (1,1) for instant hit and (9,1). Both have equal weight.

If I update my bimodal prior with the first patient the bump of the pdf to the left (low-success-theta) would decrease and the posterior would increase in the middle and to the right. If I then use that posterior as a basis for the second patient the right-hand side of the distribution (high success) would decrease and the middle again would grow.

So I was wondering if I actually need a bimodal likelihood function to update my belief.

Am I making a mistake in my thinking here or do I miss some Bayesian fundamentals?

$\endgroup$
  • $\begingroup$ While it's possible to construct and use a bimodal prior, rather than try to construct an explicitly bimodal prior, why not set up a hierachical prior that represents your understanding of the circumstances - that is, have a prior that is itself a finite mixture; or better still, if possible have the entire model based off that understanding. $\endgroup$ – Glen_b Jul 29 '13 at 22:23
  • $\begingroup$ Hi Glen, thanks for the suggestion. I actually came across that post before, when I was looking for bimodial/baysian threads. I was just wondering, since I have only two observations (1/1 hit) and (1/10 hit) I am not sure if I can estimate 5 parameters that way. I guess I have to read up a bit more on the fundamentals on likelihood functions and definitely have a look at hierachical priors - coz I never heard of those before. $\endgroup$ – user28607 Jul 30 '13 at 9:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.