0:04

Hi everyone. My name is Michael, this is Lindsey and

we're glad to have Amy here with us today for this brief interview.

We're going to be asking her a couple questions and

hopefully we have a lot of fun.

>> Thanks so much, Michael.

I'm delighted to be here.

>> Thank you.

>> First I wanted to ask if you could tell us some applications of Bayesian methods

you've worked on that deal with analysis of biomedical or public health data.

>> Sure.

So one area where we're doing a lot of Bayesian analysis that is really in large

part driven by some of our students in this school of public heath over at UNC

are shrinkage methods where we'll use a prior

to think about an effect of an entire class of chemicals, say phthalates.

And then, each individual member of that class might have an affect that

deviates with the overall class affect.

This allows us to gain a little bit of power while also handling the fact that

a lot of these chemicals tends to be highly correlated.

1:02

>> What is examples of success stories of Bayesian statistics in public health?

>> So there are a lot of interesting success stories and

one of them that I like to think about is an example of a chemical risk assessment.

So the Environmental Protection Agency assesses risks of exposures to different

types of chemicals and all sorts of products.

And one of these products was perchlorate.

Which is used in fireworks and in rocket fuel.

And so, a number of studies were carried out what perchlorate might do and

the hypothesis was that it might effect the thyroid function and

animal studies were used for this purpose.

They were the standard in that area, and what one of the studies found

was that there were two tumors and an exposure group that were thyroid tumors.

And so, for a frequent analysis if you had to expose the animals and

no control animals that have a tumor,

you don't really have a lot of power usually to say much about that.

Two versus zero.

But what was really interesting in this example was that the type of thyroid tumor

was extremely rare and in years and years and years and years of similar animal

studies with the same species, they had never seen these tumors.

And so, the researchers were able to use evasion analysis to bring in these

historical controls.

To show even though you might only see two teamers, you never see these teamers.

In fact, in the historical database,

they'd never seen this type of teamer ever before.

And so, with the frequencies analysis,

that would have been really difficult to do while a vision framework is a nice way

to bring in that historical information about the rarity of that teamer type.

2:34

What are some of the challenges in teaching Bayesian methods compared to

classical methods?

>> I guess, I see challenges a little bit on both sides, so I teach, for example,

an introductory frequentist course, and I find that students don't really understand

what P values are They're really complicated to explain and

they really want them to be posterior probability.

So I certainly see on the one hand challenges in teaching

frequentist methods.

On the other hand, some of the challenges I run into with Bayesian methods are more

to do with after you've taught someone and you run into a journal editor,

who just doesn't believe in Bayesian statistics at all.

And so, we had a silly paper actually a few years ago,

where we used a well known national study.

It was a survey, and we were in the study,

we had people that said they'd had babies but they never had sexual intercourse and

they hadn't used any assisted reproduction.

And so, this was a paper for a Christmas issue of a prestigious medical journal.

The Christmas issue sort of joke papers, and so this was one of them.

We estimated a prevalence of half a percent of virgin births in the US.

And so, what we did was we thought this was a great teaching example,

we'll use a Bayesian analysis, we'll use a prior distribution.

And so, we got on Wikipedia and estimated how many people lived on planet Earth.

And we had a couple different priors.

We had a prior where we said, we have our Christian prior where Jesus is the only

person whose been born of a virgin.

But then if you look at other traditions there’s some Greek gods,

there’s some Indian deities.

Maybe 100 others who were also purported to be born of virgins.

And use that as a prior distribution and we thought this is great,

this shrinks our estimated way off half a percent which no one

believes down to reasonable number very close to zero.

But the journal actually refused to publish it.

So they said that the Bayesian analysis was not rigorous science.

And our responses was really think half a percent

of purse to purse is rigorous science.

But they really refused to budge so I think that's one of the biggest challenges

is thinking about what a reviewer says who's maybe a little old fashioned and

doesn't understand what the methods would do,

if we had a really complicated frequentist method that took into account.

Measurement error and bias and response, I'm sure we would had an easier time of

the paper even though it wasn't quite as elegant as the solutions.

>> Yeah.

Great.

Well, thank you so much for being here.

>> Thank you. >> Thank you.

[LAUGH]