Episode

Today, we're stepping into a world where statistics directly shape patient outcomes.

My guest is Scott Berry, statistician and co-founder of Berry Consultants, one of the most influential firms in modern clinical trial design.

Over the past two decades, Scott has helped push the field from static one-shot trial designs to what Bayesian adaptive and platform trials that learn as data arise.

In this episode, we talk about why traditional clinical trials so often move too slowly, how adaptive designs change that equation, and what the COVID-19 pandemic taught us about

speed, uncertainty, and decision-making under pressure.

Scott and I dig into the real trade-offs between frequencies and patient frameworks, the communication challenges statisticians face with regulators and clinicians,

and why adaptive trials require deeper collaboration, not less.

This is Earning Vision Statistics, episode 148, recorded October 21, 2025.

Let me show you how to be a good peasy and change your predictions after taking information in and if you think it now be less than a

Welcome to Learning Bayesian Statistics, a podcast about Bayesian inference, the methods, the projects, and the people who make it possible.

I'm your host, Alex Andorra.

You can follow me on Twitter at alex-underscore-andorra, like the country.

For any info about the show, learnbaytestats.com is the place to be.

Show notes.

becoming a corporate sponsor, unlocking Bayesian Merge, supporting the show on Patreon, everything is in there.

That's learnbasedats.com.

If you're interested in one-on-one mentorship, online courses, or statistical consulting, feel free to reach out and book a call at topmate.io slash alex underscore and dora.

See you around, folks, and best Bayesian wishes to you all.

And if today's discussion sparked ideas for your business, well, our team at Pimc Labs can help bring them to life.

Check us out at timez-lamps.com

Hello, my dear patients!

Today, before the show with Scott, I would like to thank uh a dear patron anonymously called SL on the Discord.

Thank you very much because you're the one who recommended me, Scott Berry.

I hope you will enjoy the episode.

And if you or other patrons have recommendations for other guests, please feel free to let me know in the dedicated channel on the Discord.

And by the way, talking about patrons,

We have a bunch of you to welcome today on the show and I'm talking about the beautiful Avinicio Baca, Spencer Boucher, Christophe Leschowski, Danny Mohr, Jacint Juhasch, Sander

and Philippe.

Yes, some of you just have first names.

Thank you very much to all of you for contributing to the show's Patreon on the full posterior tier or higher.

Again,

This is really what makes the shop possible.

I pay for all the editing and hosting and publishing and marketing of the episodes with your contributions.

So thank you so much everyone.

I will see you in the Discord.

Please come in the introductions channel and give us your introduction to us.

you're from, what you do and why you learn patient stats, and if you have questions, modeling questions and so on, well, we have all the dedicated channels for that.

Now, let's talk about trial designs with Scott Berry.

And again, folks, thank you so much.

Got the Berry, welcome to Learning Vision Statistics.

Thank you, Alex.

I appreciate the invite and excited to be here.

Yeah, that's a real pleasure, a real treat to have you on because you do very interesting things yourself, but also all your colleagues at Berry Consultants.

So we're going to dive into that.

That's going to be, em I am guessing, a very dense episode.

strap on folks.

uh But first I want to thank Stéphane Lawrence, one of the patrons of the show for recommending me to have you on the show, em Scott.

So thank you, Stéphane.

um Thank you for your support.

And Scott, let's start before we dive into all the cool technical things you do.

I always love to get an idea of where my guests come from.

So first, can you tell us what you're doing nowadays and how did you end up working on that?

Like what's your origin story?

ah So it's a long origin story ah in it.

I am currently working at Barry Consultants.

We are a scientific consulting company working in clinical trial design.

So we design...

a range of trials and we really entered this from the idea of designing smarter, better clinical trials, more efficient clinical trials and we have about 30 scientists here at

Barry.

So that's currently here.

Getting here is a huge part of this story is certainly that my father is a statistician and the family has a statistical component to it.

So I started as a PhD statistician, certainly went into it knowing my father was a statistician.

So I knew about statistics and I was pretty good at math.

I entered a lot of my interest in statistics through sports, enjoyed playing sports, watching sports, the quantitative aspects of sports, the questions.

uh which has certainly since exploded.

Sports analytics has certainly exploded.

uh So I started after getting my PhD at Carnegie Mellon.

I went to Texas A and I was on faculty there, research, teaching, and I did quite a bit in sports actually.

It was solving puzzles.

And I started working on clinical trial designs.

And my father at the time was at MD Anderson Cancer Center.

And he was leading the bioinformatics, biostatistics group there.

And they were doing uh enormous number of trials.

And so was working on some consulting projects that came to him and designing trials.

it was, I found it quite exciting and impactful.

And each one of these had a problem to solve.

And I sort of enjoy that strategy problem solving part.

Started to work on some of those designs and decided to create a business out of that.

So in 2000 we created a business to design better, smarter clinical trials and hence to where we are now with 30 scientists working on such trial designs.

Damn yeah, that's a, that's a big team.

30, 30 people is so, I remember when we started PaimCLS, we were four and then we started growing.

I remember already 10, 15 people that starts to be, you know, you can't, you start needing some documentation, some processes and things like that.

I, I definitely resonate with.

uh, managing 30 people that's like, I'm guessing your, you know, GitHub and documentations must be, must be on point.

and actually I forgot to mention that you have a podcast also.

as a fellow podcaster, I should definitely mention that.

So, um, go listen to Scott's podcast, people in the interim, it's in the show notes.

I also added episode 45 of learning vision statistics.

where I received Frank Harrell, which I'm sure Scott, you know.

um And in there, he talks about your father actually, in very, very good terms.

And yeah, I really recommend listeners to listen to that because Frank is an extremely good um teacher, let's say.

Like he communicates extremely well about biostats, clinical trial design, why that's...

important how to do that, why the Bayesian framework is very important in there.

So definitely recommend to give that a listen.

It's going to compliment what Scott is going to talk about today.

So yeah, like your background is, yeah, you are already familiar with stats.

You entered through that actually more with sports than with bio stats.

And then in 2000, you and your father, so Donald Berry, you founded Berry Consultants.

So what was the founding vision and how has it evolved over time?

Because we're 25 years later now.

Yeah.

So in 2000, I had worked on several design projects and I very much was trained as a Bayesian.

I went to Carnegie Mellon, my advisor, Jay Cadane is very much a Bayesian.

So I, and this was right at the advent of Markov chain Monte Carlo and uh being able to actually do computation in this and very much enjoyed the modeling part of it, but I think

as a Bayesian and the interesting part of the world of biostatistics, the world of clinical trials is

at that time and even now it's a very frequentist dominated world.

ah And you could almost say it's the last bastion of frequentist.

um It's kind of the, if frequentist is going to be successful, that's where it's going to be successful.

It's set up for it and all of that.

I think that's coming down and we're finding that they're becoming much more complicated, much more interesting.

uh A clinical trial is not just the data in the clinical trial, which frequentist needs that for its inference and it needs this sort of closed space and that whole world is

changing.

But in 2000, incredibly dominated by the frequentist world, and I think it's limiting in the trial designs actually.

It is strangely taking this sort of approach, limits the adaptations, limits the design.

And you can imagine when I tell people what I do that aren't statisticians or aren't in this world is we take what the old model was of a trial where you say go enroll 200

patients and then in three years you look at the data and invariably say, shoot, I wish we'd have done a different trial.

I wish we'd had a bigger sample size, a smaller sample size.

We'd have changed the randomization.

We'd have changed the patients.

There's a whole aspect.

to this that the idea that you run this trial for two or three years and only then do you get insights into the data, they look at you like, that really is what's happening?

So we thought the idea that you do multiple interims during the trial, you make sure it's progressing the right way, you allow it to change to answer the relevant questions.

are better trials, better use of patients, better answers.

And it was really jumping in in 2000, having no idea whether this would stick.

And all of a sudden this is my way of income and I three small children, but was willing to take the risks and go into this.

And hence you talked about 30 people and now...

ah There's a great deal of innovation in this space, but it was initially a lot of medical device trials, which has a certain synergy to Bayesian to then drugs uh within that

situation jumping on and really doing that.

So it is something that hit a sweet spot and there was a great deal of interest and still is today in alternative called novel approaches to clinical trials.

Yeah.

That, that makes sense.

And so actually before diving a bit more into that, can you give listeners an idea of what that means concretely?

Like what is, what does your job look like?

You know, we're on a, we're on a Tuesday uh in the morning.

So right now you're recording, but usually what do you do Tuesdays at 10 AM?

What does the work you're doing look like concretely?

Yeah.

So it's typically that we are working with an organization that's interested in designing a trial.

And that organization can be a pharmaceutical company, could be a medical device company, it could be the NIH, it could be a federal government, uh it could be a patient

organization, where they're interested in a trial design.

They're interested in a question or...

usually multiple questions and they want to design an experiment to answer those questions.

So we will work with them on these designs.

And so for example, there may be a particular drug and they want to get it approval by the FDA and they want to run a trial to do this, but they have questions.

Maybe it's there's three potential doses and they don't know the right dose.

Can we run a trial where we explore the three doses, answer the question, figure out the right dose, have it advance into a confirmatory stage and rolling against a placebo, and

that demonstrates safety and efficacy in the same trial?

Maybe even a question about do we have the right patients?

So the alternative would be you run a trial to investigate dose, you wait a year, you analyze the data, you run another trial to then

compare, can we do this in one?

Can we do this more efficiently?

That might be a standard and we work with them on the trial design, what that looks like.

We're simulating different trial designs, comparing them, building a better design and not infrequently within those designs.

The thing that's driving those, the dose decisions, the sample size decisions, our Bayesian uh models are driving those decisions.

And so we are

building trial designs for a wide range of groups interested in building really a smarter, more efficient design than a simple, fixed, let's look at the data in two to five years.

Yes, and actually, so I think it's a good segue to explain what the main issues are with the classic way of designing trials.

You've mentioned already a couple of times that uh there is, well, the problem of picking, the problem of how to determine the sample size you need to get to the power uh that you

want, things like that.

So can you give...

listeners a rundown of how these things are usually done.

And I'm guessing a lot of listeners will also be familiar with these terms, not only if they work in biostats, but also if they work like in any field where you do a flavor of

A-B testing, where you have a placebo and you are testing if something, if a treatment is more efficient than the placebo.

Well, you're doing basically what, what stock is going to...

Scott is going to talk about right now.

So maybe give us the rundown of what the current way, the classical way of doing things is and where the main pitfalls and problems are.

And then we'll enter into how you guys solve that and how base is actually helping a lot in the end.

the, now trials are, there's a, there's a huge range of them and they

And they're typically grouped in phases where phase one, you're really worried about safety, you're giving a new novel drug to, and it might be first in human kind of stuff,

through phase two where you're really trying to use multiple doses, frequencies, patients, it's more learning about it to phase three, which is confirming safety and efficacy.

So all of this, this can be different, but the old model, and it sort of ties to a frequentist approach is,

You enroll some number of patients on a control, it might be a placebo or on a control, and then some number of patients on a treatment and you're addressing the question of are

patients doing better on treatment than they are on control?

And yes, many experiments are set up in the same particular way.

The frequentist approach to this is asking,

typically using a p-value and doing a hypothesis test of our, you know, testing, these treatments the same or are we seeing statistically significant improvements in the

treatment arm?

So if you think about that, the p-value saying, assuming that there's no difference, what's the probability we see as big an advantage in the data as we see?

And this could be, uh

time of survival in an oncology trial, it could be weight loss, what's the amount of weight loss in a weight loss treatment.

ah It could be the rate of decline in Alzheimer's, it could be the neurological status in stroke.

So it's creating these measures we're looking at and our patients better off within that.

If you think about the frequentist approach and the p-value, it's okay.

Assuming some hypothesis, what's the probability of data in this experiment as extreme as it is?

Now, this is actually an ideal case for frequentist.

I think like Bayesian and all, because we write down the design and we prospectively do everything.

And so when we say, what's the probability this experiment would create data like that,

We can actually calculate that reasonably well because it's all prospective.

If you're out there doing experimentation out in many circumstances, don't have a design.

P values are really hard in this thing.

So this is a case where it actually applies really well.

Now there's two things that this starts to break down is one is we start to make the experiment more complicated where there's multiple.

uh

tentacles to this, it could go down multiple routes.

We could increase the dose in the middle of the trial.

We could do response adaptive randomization.

The trial itself is becoming more complicated.

So if you say in the frequentist, what's the probability we see data as extreme or more extreme?

That's a sample space thing.

And that's really hard if there's lots of moving parts to the design.

So it becomes really hard to be a frequentist when the design gets complicated.

So if you have to be a frequentist, it's hard to make a complicated design.

You're sort of stuck with simple analytical things, which is really limiting.

The other thing that's changing in the world of experimentation clinical trials is we don't just analyze the data in a trial in many circumstances.

We're getting incredible amounts of data.

external to the trial.

So if you start to say, what's the probability of getting data in my experiment as extreme as it is, it's sort of not defined very well anymore because we're using multiple sources

of data and all of that.

It's much easier to be a Bayesian in these circumstances because you only condition on the data you've seen.

and you don't have to worry about all the things that didn't happen in the world.

So it's something that with more complicated design or data external to the trial, it's really hard actually to be a frequentist where 25 years ago with simple designs, it was

quite easy.

Yeah.

Okay.

Okay.

Yeah.

That's super interesting because I was always aware of, well, these classical issues that you have in the Frequentis framework where um basically one of the main issues is peaking.

If you look at your test results before you reach the sample size that you wanted to reach the statistical significance, then you...

You increase the probability of false positive.

Also the issue of multiple comparisons.

This is very hard to do in the Frequentius framework.

All these problems don't really appear in the Bayesian world because well, you're not estimating the probability of the data and knowing H0, you're estimating the probability

of H0 knowing the data.

so looking, peaking is not really a problem.

You're just updating your posterior.

Right now, by the way, this is the part you brought up Frank Carroll.

So if you go to in the interim, I do a podcast with Frank Carroll and you're welcome to do.

So we're both very strongly Bayesian, but this is the part where, where as two Bayesians, we, have sort of the conflict is yes.

What's beautiful in an adaptive design is if I do an interim analysis,

And then I get to a final analysis.

That interim analysis doesn't affect the final.

Looking at the data doesn't affect that analysis.

And there's a beauty to that in the Bayesian approach where it does in the frequentist approach in that.

there's still, we live in a world where people are interested in the experiment you're running.

What's the chance that that experiment creates a false positive?

Now that's place where Frank thinks we shouldn't even calculate that, but the world still is interested in that.

In evaluating a design you're interested in, does it get the right answer?

Frequent characterization of if you ran this design multiple times is not an irrelevant thing to evaluate a design.

Now, it is when you're analyzing the data, and we agree on that.

but there's still sometimes you still have to characterize the look.

So there's no free lunch to looking all the time and finally saying, yes, we're successful in this world still.

Yeah, yeah.

Thanks.

This is a very important nuance.

And so what I was saying also, my main thought was, okay, so I'm aware of these issues, these differences where, I appreciate what you're saying, which is, well,

This is interesting because in theory that would be paradise for frequentist methods, but actually de facto it's not because of all these theoretical issues, picking multiple

comparisons, things like that, that we just talked about.

But also what I wasn't aware of is what you were saying, which is, well, we actually have other data that's coming from outside our experiment.

And so now the idea of

If we repeated this experiment multiple times becomes very blurry because well, are you able to repeat actually the same experiment?

This is more and more doubtful.

And so, yeah, that is actually a very important point I wasn't aware of.

And so here, think it's interesting to actually try and get an example for listeners.

So, Viz, you've been involved.

in designing a lot of Bayesian adaptive clinical trials.

And you also wrote platform trials.

So could you first define for people what these are?

What's an adaptive clinical trial?

em What is a platform trial?

And then walk us through a recent example that you found, especially in illuminating em and what made it interesting, see, you know, like was it its complexity?

regulatory challenges, technological innovation, things like that.

Yeah.

So an adaptive trial is one where during the course of a trial, an analysis of the data alters important parts of the trial.

And I don't want to say it changes the trial because from the beginning you design it that way.

You design it that at 50 patients, we're gonna do an analysis of the data and certain characteristics in that data will change the trial.

We stop enrolling the low dose if it appears to be ineffective.

We drop a patient population if it appears the drugs not benefiting those patient populations.

And models will drive those decisions.

So it's a trial that during the course of the trial, analyses of the data in the trial can change important characteristics.

The patients enrolled, the arms given, the ratios of them, whether you keep enrolling, whether you stop enrolling, shifting to a new trial.

There's lots of things to that.

In some ways, a way to explain an adaptive trial is it's not one that you set it up, you run, and you only analyze the data at the end.

Within it and you can imagine the potential positives of this is You're trying to answer questions by this trial.

We're asking mother nature questions does this drug work is this device safe and During the course of the trial if you're able to look at the data you learn things that you

didn't know when the trial started You would have designed a different trial if you knew that

And so you start to learn that.

So can you learn efficiently to make a better design to answer the questions more efficiently and maybe perhaps answer more questions uh within it?

And you can imagine this can be a very, very powerful experiment to have this flexibility to it uh in that.

those are what adaptive designs are.

Platform trials are relatively new concept where

We, you can imagine we run a lot of trials where let's just take Alzheimer's for example, where company A is running a trial of 500 patients on drug to 500 on placebo.

Company B is running a trial of 500 on drug to 500 on placebo.

Company C is running 500 on drug to 500 on placebo.

As an ecosystem, we're enrolling 1500 patients on placebo and

These in A aren't used for B and Bs aren't used for A and all of that.

So you bring them all into one trial.

You enroll A of 500, B 500, C 500 and 500 placebo's that they all share and compare to.

We've just eliminated a thousand placebo's, a thousand patients taking nothing.

to compare to these arms by working together in a scenario like that, we can be already much more efficient and still each one of A, B and C get exactly the power, the

inferential strength they needed, but by working together as an ecosystem, we're doing much better.

So this is a trial that had some initial designs.

This was being done in cancer where multiple arms would come in.

But it was really a pretty new type of design.

And then COVID hit and the pandemic hit.

And so now we have many possible ways to treat COVID.

This is separate from the vaccines.

But to treat COVID, there's all kinds of repurposed drugs to this.

Do we want to do antivirals, steroids, IL-6 receptor agonists, all kinds of things that we're interested in treatments.

We don't have the ability in two ways to go out and start a new trial.

It takes a year to start a trial and we're enrolling and we've got way more questions than we do ability run trials.

So lots of platform trials were created.

The US worked on Operation Warp Speed where their active program were all platform trials where they had common controls and they brought multiple drugs in.

This was done in recovery in the UK.

This was done the together trial.

The remap cap trial was a global trial.

It was this immediate, my God, this is the only way to answer this question in COVID.

And they were adaptive.

So they were doing analysis.

Remap cap was doing analysis largely every two weeks saying, do we know the answer?

Do we know the answer?

Stop enrolling that drug.

doesn't work.

Let's focus on this drug or that drug.

So it was really a, an incredible opportunity where almost everything we learned about how to treat COVID came from platform trials.

This incredibly efficient way to investigate multiple arms simultaneously.

And they were adaptive.

We stopped them when the drug wasn't working.

We stopped them when the drug was working and we announced to the world, if you could announce that steroids were benefiting patients a month earlier.

You save thousands of lives.

And so all of these things came together as a demonstration of how incredibly valuable all of these innovations in the experiments uh that we're using had in the treatment of COVID.

Damn.

Okay.

Yeah.

This is super clear.

Thanks, Scott.

And yeah, like this is impressive to see as you were saying that really if you can, this is also

almost always a run against the clock, right?

Where if you can have that better solution, even just one month earlier, you can save many more lives.

So this is, this is also something that can be extremely impactful.

so by the way, there are many diseases that we always have that urgency.

So, um for example, and there was this great learning of the power of these platform trials.

so a really good example of this is Sean Healy was somebody who had ALS.

Um, and he funded the start of a platform trial for the treatment of ALS to accelerate, creating treatments for ALS.

And now this is being run out of Mass General Hospital.

There's a network global, uh, network in the U S.

ah the Niels network that runs these trials.

And it was because of this same urgency in ALS.

We have patients, you know, to them, this is incredibly urgent.

Now, COVID was sort of this global thing, we understand, but people who have pancreatic cancer, who have ALS, who have many other ailments have this same incredible urgency.

And we have an ecosystem that's actually slow.

It's lumbering.

It's not good in the setting to tackle that problem.

Yeah.

something I want to, I want to dive in also is when you do that.

when you work on these, on these adaptive trials and these platform trials, as you were saying, you folks use a lot of the Bayesian frameworks because it comes with a lot of

bells and whistles and solve a lot of the issues that the classic framework has.

However,

What do you see as the biggest statistical or logistic hurdles when implementing adaptive Bayesian trials in practice?

Um, what are the hurdles?

So, so good question.

ah the, what's harder in it is it takes longer to design.

So you can design a fixed trial in 10 minutes.

Uh, you write down a sample size, you do a calculation and that's a little unfair, but, but largely it's pretty straightforward to design one and the interaction with the

statistician has been one almost at arm's length.

Here's the design, here's the effect size, give us a sample size sort of thing.

The creation of an adaptive design takes longer because you're thinking about what kinds of things do you want to adapt, what are our learning mechanisms for it?

ah And now to evaluate a design that's adaptive, you have to do trial simulation.

We can't calculate on pen and some paper that.

So the design is longer.

The design of a platform trial takes longer.

to create than for a single drug.

So that's the part that's kind of changing the system is this isn't a arm's length reaction to the statistician.

The statistician is a huge part of the team, the questions and doing that.

There are some implementation hurdles to this that now historically

Yeah, let's wait three years and look at the data.

that's not, know, we have data safety monitoring board looking at them, but largely this is true that we, we analyze the data three years out or when the trial's over and now

we're, doing interims monthly.

We're doing interims every couple months and the data needs to be available and all of the parts to it.

So it's a bit of a change to the system that, and there's a lot of, there's a lot of, um,

This is the way we do things that you're changing.

You're changing uh operations to that.

has been the, what, there's a lot of Bayesian things that have been done over COVID in the last 25 years.

What hasn't been a hurdle is explaining what it means.

It, and it's kind of interesting because people,

think they know what a p-value is, but they actually, it's really they're interpreting like a Bayesian.

So being able to say the probability this treatment is effective is 99.1 % is a really, really valuable thing rather than this opaque p-value thing that people are just used to

it.

So they kind of know they're used to it.

They don't really know what it means, but they're used to it.

So that's not.

been a hurdle in all of this, but yes, adaptive trials were changing the ecosystem a little bit platform trials, uh, within it.

I think largely the, the creation of these has been longer than for a simple trial.

Hmm.

Okay.

Yeah.

Yeah.

Um, does it resonate with also the interpretation part that you were saying?

and I guess because

So what's interesting to me also is that I see a of people who interpret the p-values or the frequent disconfidence interval the Bayesian way, and they know it's not right, but

they say, well, that's fine.

So for that's always like, for me, this is this, this terrible.

But I guess if you're not working in a field where it's literally a matter of life and death, that's fine.

I would not do that when you're.

um

designing clinical trials and working on some new drugs as you guys are doing first.

and, and second, then, okay.

So as you were saying, this is, this can be a harder to set up, but then you get a lot of, um, lot of bells and whistles for your effort.

I'm wondering.

Do you have, do you recommend using that almost in all the cases or you see some cases where you can get away with a more approximate um frequentist framework when it comes to

clinical trials like you're doing?

So we still design trials where we do frequentist tests at the end of the trial.

And it's usually at the end of development, this confirmatory trial.

where there's still a good bit of regulation where they're looking for an experiment with a 2.5 % type one error that demonstrate this treatment's effective.

If we're going in and we were gonna use a flat prior in the scenario and it's only data in that experiment rather straightforward, we would do a frequentist test of that in part

because it's still different to do Bayesian and you're probably not

getting any value above and beyond the frequencies, you're probably reaching the same conclusion for the same data set.

so generally that's an area where, you know, it doesn't bring anything extra.

would, we would go ahead with and, do frequences stuff.

And a lot of that is, you know, um, are we bringing value?

Are we doing it just purely because we want it to be Bayesian sort of thing.

And if it's not bringing any additional value, we'll do the

the frequentist approach to that.

So each trial you sort of evaluate, uh are we getting better answers?

And it might be by doing a Bayesian design, we're able to do a better experiment.

And at the end, we still do a t-test at it.

it can vary.

Yeah, that makes sense.

And something I was also very curious when I was preparing for the episode is

I saw your team develop some simulation tools.

um I think uh I saw one that's called FACST.

FACST, that's an acronym for Fixed and Adaptive Clinical Trial Simulator.

um Can you tell us what this is about and how do you balance the technical software innovation with delivering usable tools for clinical trials, uh teams and regulators?

Yeah, so when we started designing uh adaptive trials, we might do a trial where it's a phase two trial with seven doses.

And during the course of the trial, we want to increase the randomization to doses that have better therapeutic benefit.

We did this in a diabetes trial, for example, where we want to increase the patients on doses where

HbA1c change is better, weight loss is better, and lessen the randomization to those that aren't doing as well.

And even lessen the randomization if we're starting to see increased heart rate and blood pressure, which we're trying to avoid.

And so during the course of the experiment, we might do interims at multiple times.

In order to create and evaluate that design, saying does the design work well?

Are we getting better answers than an alternative design?

The only way to answer that is through trial simulation.

So we set up a simulation of the trial and we plug in scenarios where dose four is the best dose, dose seven is unsafe, and we simulate and say, how often do we get the right

answer?

And then we look at different designs.

We look at, should we do interims monthly?

twice a month, and we evaluate multiple characteristics to it, and simulation was the only way to do this.

So for years, we would write code to do this simulation, and we became somewhat experts in this to design a more complicated trial.

It was a necessity to that.

We were doing it enough, and the pharmaceutical companies wanted to be able to do this more, and we were the rate-limiting step.

At the time the six statisticians that we had were writing code in C or R or S plus or Fortran and we were the rate limiting step.

And so they said, can we put some of this in a package that we can use and do this exploration and not have to write code and, you know, take three months to write code and

debug it and all that.

So we created the package.

that under it, has a good bit of flexibility where you can set up interims and analysis methods and adaptations to it.

And it's all compiled code.

So it's compiled our code.

It's very, very fast.

Simulations of a thousand trials can be reasonably slow if you're doing 10 Bayesian analysis with MCMC in each one of them.

ah That can be reasonably slow.

And so this was a very fast tool that enabled people to explore adaptive trials.

Kind of a necessity once you get into more complicated designs and all of that.

And it was something, and now we've continued to expand it.

We use it internally when we're building trials.

We license it to a number of pharmaceutical companies, to ah CROs, so they can use it to design trials.

So now it's a...

simulation to a much like simulation to build a house, to build an airplane.

It's a very common technique to create more complicated designs.

Yeah, yeah, yeah, this is, this is super interesting.

And I love how, how the product came to be, right?

It came out of a necessity on your end, but actually something that's usable for, for everybody.

And yeah, it shows again, the power of simulated data.

This is so, so important.

And uh I know Andrew Gelman very often talks about that and says how important it is.

I use that all the time also in my own modeling, especially when the models are extremely complex.

Or, and or you need to use some other approximation algorithms that are less guaranteed to converge than MCMC.

Let's say you need to use some, some version of a variational inference then testing the model on simulated data before fitting on real data is extremely helpful.

um Gives you some

give you a lot more guarantees.

And it's also something that I find is a bridge between the veteran framework and the frequentist framework.

Where actually when the structure of the problem becomes way too complicated for the normality assumptions and the central limit theorem to apply, then the frequentist

framework has also to rely on simulated data.

compute the p-values, things like that, because otherwise it's like you don't have any guarantee that your tests are actually telling you something meaningful.

They will run, but they don't tell you what you're interested in.

And the part of the simulation in what we do that I didn't appreciate at first.

So simulation is sort of the way statisticians think.

We think about simulation and when we do...

P-value calculations and all that.

I mean, it's very, very natural to statisticians.

It's not a very natural thing for other people, but it's become a really valuable tool to show them an example of a trial.

Okay, we're about to run this trial.

Here's what the data will look like at interim one, interim two.

Here's the decision the trial makes, and it demystifies something that can be a black box to them.

So it's an incredibly valuable teaching tool or to make this, you can show somebody exactly what's gonna happen in the trial, what the data may look like, what the trial

does.

And then they kind of understand the simulation.

we're gonna simulate another trial.

And then we simulate a thousand and here's how often this thing happens or that happens.

So it's been an incredible tool actually to making this very approachable.

to everybody in the space, regulators, clinicians, operations.

So it's been more than just a statistical simulation tool.

Yeah, yeah, that's very true because in your, in your line of work and in a lot of, oh of lines of work, you're not doing modeling in a vacuum and you have to explain that to a

different array of people with different backgrounds.

And so as you were saying, models and parameters being interpretable, that's extremely helpful.

And that's also something you get from the Bayesian framework.

But as you were saying,

Being able to communicate that back to people, showing them the data, what this would look like beforehand.

This is extremely, extremely valuable.

found that working in sports actually is also something that's extremely helpful because when you talk to a coach ah or someone who is an expert in baseball, but not in modeling,

they don't really care about what's in the model.

But the input and the output, ah they...

can relate to that and that's something then you do your statistician translation in your head where okay what he told me here I can use that in my model in that way with this

prior or that distribution.

um But this is a way of bridging the gap between the different departments you have to interact with.

Talking about that actually, I'm like you...

In the, like when you do adaptive platform trials, uh I saw when preparing the episode that, and as you were saying, this is something that got prominent attention, especially

in pandemics and high urgency settings.

I'm wondering what you have learned about working that way under faster timelines or high-stakes conditions, uh whether they are regulatory, ethical, operational.

What is something that maybe now you've changed in your processes that you learned during the last pandemic?

Yeah, it's a good question.

don't know if there's anything that we've changed.

It was certainly a, I've always thought, you know, when Armageddon comes, I have no skill sets.

You know, I can't build things.

I can't help somebody who's sick.

You know, what am I going to do when Armageddon comes?

And all of a sudden a form of

Uh, of the pandemic hits a form of Armageddon hits, and all of a sudden statisticians played an incredibly important role.

We were doing modeling of rates of hospitalizations in LA County.

And the same times we're designing experiments to treat people better.

Oh, then there's a huge role of statistician.

So there were, it was us.

was, it was, it was the community.

Um, just.

did everything they possibly could at breakneck speed within it to answer these questions.

And there was something actually uh incredibly ah wonderful about that, the whole scientific community of that.

I think there was sort of this common aspect of we can help kind of thing within it and this urgency that went on with it.

Now, if not sustainable,

within the setting of that with the multiple trials to that.

ah But it was a sense of what was kind of interesting is we did things that we wouldn't have done at the time in other diseases.

And it's a little bit of this, there's huge benefits that people took risks, they did things differently, there was no other way to do it.

And we learned a tremendous amount with that.

So one of the sort of lessons is the, you know, the experiment with experiments, the innovations disrupting people, we had to there that we should continually be doing this in

what we're doing.

I think that there's a stasis that grows a little bit.

So it gave us power to continue to do novel things, I think, and be able to push.

the envelope a little bit.

don't know whether, I don't know whether any of us wants to go back to that time period and work under the time constraints we had in the hours.

And one of our trials, the remap cap trial is a global trial and there was always somebody awake.

And so things were moving and you go to bed, you wake up in the morning and it's moved and it's your job to move it forward.

You know, this way we'd be on calls and the only time, the only way you knew what time it was, was who was drinking wine and who was drinking coffee in which part of the world they

were on.

So I'm not sure that was something we, anybody wants to go back to, but there were a lot of just incredibly positive things and innovation that happened because of the need.

Yeah, yeah, Yeah, for sure.

Like these, these kinds of very urgent situations you,

things happen way faster than it would have been otherwise, for sure, because it streamlines the process in a way.

What's not really important gets out of the way.

But yeah, I agree, this is definitely not sustainable.

On the long term, I'm guessing you didn't sleep enough during that period.

So I feel better that you guys can sleep on a better schedule now.

So as you've talked about already earlier in this episode, when you design an adaptive trial, you have to deal with many moving parts, whether that's interim analysis,

predictive probabilities, simulating data, stopping rules, blah, blah, So what do you, I'm wondering concretely, what do you consider when choosing between a classic randomized

design?

versus an adaptive patient design?

uh So there's no doubt.

Our trials that are adaptive are still randomized.

this is not...

This is the gold standard.

There's huge value to that.

So it may be having multiple parts.

And that's the part that takes a little bit longer in the evaluation.

Are we getting better answers?

Are we using patients...

resources more efficiently and we're comparing different designs.

So we compare a fixed 600 patient trial and then we'll evaluate what happens if we add these stopping rules, what happens if we add these adaptations, the frequency of it and

we're always sort of comparing the important part is you're comparing to different things and you're hopefully getting to the most efficient scientific experiment.

uh at that.

And so it's this continued iteration to get to it.

Meanwhile, you're also asking the question, okay, if we're going to do an interim analysis daily, can we do that?

Can we operate it?

So we need to design something that operationally is going to function well.

We're going to ask a data safety monitoring board to review the analyses.

uh So a machine could run these trials and

It could be run by machine, we, it's it's a automatic pilot flying a plane and we build the models that, that, runs it by automatic pilot, but we want a pilot sitting there

watching it, making sure that, you know, it's doing right things and it's reasonable.

we need to, we need to create something that can be operational.

So it's sort of that, talked about the need of simulation.

So when we went into this world of let's design better trials, it was bringing those designs.

The two other things that we found that we had to do as a company in order to change the world to do this was the clinical trial simulation.

You can't just write down a design and hope it's a good one that you've got to evaluate it, compare it and create it.

And then you've got to actually run the trial.

So we have a group here that helps companies make sure they run the trial right.

And they handle all of the weird things that happen with the trial, data issues, working with the CRO, because we would design trials and all of a sudden they wouldn't be able to

implement it.

We were doing stuff so different than they were used to that it didn't no good to design a better trial if we couldn't implement it.

So that's a huge part of the design part, but also something that

A lot of times it's just getting in and say, we can do this.

Sure, we can do this.

We can update randomization every month.

We can do this.

So those are really the three parts of it that we've had to focus on in order to bring the better trial to that.

So that's absolutely at the forefront of making sure we're building a better design.

Okay, yeah, yeah, yeah.

Okay.

This is really interesting to see how in your field you have so many different parts in different stakeholders that you have to...

You're really at the center of it and you have to coordinate everything and make sure everybody feels confident about each step, which I think is also extremely...

interesting because you have a big part of communication and teaching in a way in these kind of roles, which I found super, super interesting.

um Actually, how do you communicate the statistical complexity of your design features to non-statistician stakeholders so they feel confident?

Yeah, I mean, it's a huge part of...

uh

of being a statistician, being a scientist in this to communicate it.

If you create something really nice but can't explain it to all of the stakeholders, to regulators, to the clinicians, to the journals at the end that are publishing the data, to

the data safety modern board doesn't do any good.

So it's a huge part of it.

uh

Simulations have been incredibly valuable because we can pull out example trials and say, here's how it works.

Here's an example.

Here's another example of how it works.

Here's what you're going to see here.

Here's what you're going to see here.

And then the simulations, here's why it's better.

So you sit down with the FDA and you say, we could do this other trial, but we're proposing this one and here's why.

And we can walk through and explain, here's why it's better in language that isn't statistical.

Almost never does it, does it, does it work to say power is higher type one errors lower because these are terms we kind of understand really well, but others don't.

So what are the average number of patients that we're going to use in the trial?

What are the proportion of patients that go on an ineffective dose?

How are we going to treat patients when it's over?

What's the average time of the trial?

How often do we get the right answer or wrong answer?

These are all things that we need to be able to communicate to the different stakeholders because we're doing something different.

In the sports world, for example, it's easy to do what everybody's done before, but to do something different, you've got to bring on all kinds of change.

so we've got to be able to communicate to every one of these stakeholders because if any one of them say, no, it doesn't happen.

uh

typically.

So it's probably the most important thing we do.

So it may take two days to run the simulations, and then it takes a week to put the material together to show and explain and compare.

And that's the huge part that we go back and forth.

that can be the huge challenge of it.

So in some ways, it's the most important thing we do.

Yeah, yeah.

And it takes a lot of your time also as the modeler.

There is something very important, I think, for practitioners to know is that modeling is actually a small part of your job.

And a lot of the time is spent talking with stakeholders before and after modeling, um doing a lot of customized plots.

Also, I know in spots it's very, very important because you can have very complicated models.

Usually showing the outputs of the models, looking at different players is much more important than, you know, showing, uh, showing anything that's inside the model.

That's usually not, not helpful.

Yeah.

So if you were going to explain to a, an NFL coach, why going for it on fourth and two from the 40 yard line is beneficial.

You can't just say.

Your problem, your probability of winning the game is better.

So I mean, you've, I think you've got to be able to explain that, or if you're going to explain to a golfer why hitting a driver or hitting a four iron off a tee is going to

minimize your stroke.

I think you got to be able to show them why and show them that.

and, and it's, it's, it's a hard and sometimes that's a hard skill set for us.

We love the math problem.

We solve it.

ah

you know, kind of thing.

And then we move on to the next problem and, and that's, you know, it's hard to change the world that way.

Yeah, yeah, completely.

That's also why it's something I'm very um passionate about, especially with these podcasts where I think there's, are so many people doing um great things in advancing

science.

And in the end, you also have to

tell a story about how that's done.

um The persons who do that, the stories behind the formulas, because otherwise it's way too dry for people and we're a species that thrives on stories.

We need stories, we love stories and that's how we pass information, how we've been doing that for centuries, know, like these legends you hear and so on.

That's basically how we pass the stories before we could write.

So.

em

Yeah.

As you're saying, this is an extremely important part of the job.

I'm curious, know, like beyond, beyond drugs, do you see patient adaptive trial methods becoming standard in other fields?

Whether that's uh public health or any other fields actually that you're curious about and what would need to change for that to happen widely?

Yeah.

uh

It's certainly true of public health.

uh Now, what's going to happen in public health is the explosion of available data and uh making sense of that data, whether it's the complexity of the data.

uh I wear a loop band, for example, on my arm or people wear a watch, a wearable, and you can imagine the enormity of the data that's going to come out.

How are we going to make sense of that?

Um, and so that's, I think that's going to explode and it's already there.

And we see, we see bad conclusions being done in public health right now, but, but that that's a huge part of it within other fields.

I there it's a question of experimentation and use.

I think it is being used.

I mean, sports is an example of that.

in interesting application that you wouldn't necessarily think of is,

My father, Don, his PhD dissertation was on bandit problems.

Do you want to pull arm A or do you want to pull arm B, which is much like a clinical trial within that?

It turns out that they're not used a whole lot in clinical trials and there's sort of reasons to that, but it's used incredibly by companies like Google in advertising.

So suppose you're on a website and you have

several different ads you could give to somebody to click on something.

All you're trying to do is get them to click to go to that website.

So you show them an ad with different components and you throw it out there and they don't click on it.

That's much like a clinical trial.

And so let's try a different ad.

And we've got 14 different arms and we optimize the messaging that's given to people and you get trial after trial after trial.

They're the ones that are going to my father and saying, hey, we want to learn more about the bandit problem stuff because they do this and they do this on such large numbers and

they want to optimize the marketing aspect of the presentation of it is much like a clinical trial in that sense.

And so they're doing Bayesian adaptive stuff all the time.

Now in ways we hate because it's in our face sort of thing.

But it's incredibly potent in the marketing world.

And my brother's a statistician that lives in that space and they're doing these experiments all the time.

ah I don't know if they're doing Bayesian adaptive things, but I know certainly ah internet, Google and all of that are certainly doing stuff like that.

So I think there's a number of places that this is certainly being done.

Yeah, this is fantastic.

If you have any...

um

links that you can add to the show notes about that, please feel free to do so because I think uh listeners will be willing to dig into that.

um if it turns out your brother is doing some fun-basian things, we should definitely have him on the show because that will be another arm of the field that we have to pull.

So definitely let me know.

That'd be great.

Actually, so, well, I'm going to start playing a South European piece.

want to be respectful of your time, but I'm also curious, you know, with, so you've talked about increasing data in a lot of fields already.

We also have more computational power and as you were saying, more interesting platform designs.

Where do you see that taking clinical trial designs?

in a few years.

I think the complexity of outcomes is certainly going to expand that we don't just do a single test at six months and how out of 10, what's your score out of 10 at a single point

in time in six months, but we're going to have this continuous data on the way people behave in the act.

In the sports world, for example, I think they all of sudden take a snapshot, like 10 snapshots every second of where everybody is on the ice in a hockey game.

I don't know if we have any idea what to do with that.

How does that help us?

And I think the same thing's going to happen in health that we've got these biological measurements of our gait, of our behavior, of all these things.

And it's going to be, how do we put that together in a story to that?

But the other part of it is we learn in what we give therapies from less than 1 % of people that have conditions.

And they're only people that go into this experiment.

And meanwhile, somebody else has pancreatic cancer, they have acute stroke and they're being treated.

And we learn nothing from those people.

And 99.9 % of people fall into that.

I think we're going to get to the point where that data and learning healthcare systems and can we tap into that 99 % to figure out better ways to treat people and all of that.

And that's again, this external data to just a single trial and that there's so much valuable data and quality data that wasn't there 10 years ago.

mean, we didn't have these data sources, medical records were terrible and all that.

In 10 years, we're going to have this just phenomenal amount of data and we're going to be addressing questions about how to treat people, what work in all that.

And it's going to get more and more complicated.

So more and more Bayesian approaches are going to be hugely valuable and statisticians are going to be hugely valuable.

But think about another circumstance here where why do we do randomized trials?

We give treatment to somebody

We wanna know the counterfactual of what would have happened to that patient if we didn't give them the treatment, which is why we have people that we don't give a treatment to.

We give them a placebo.

So we have 100 here and 100 here, and we compare them and they play the counterfactual.

Suppose we get to the point that this patient, this Jane Doe with Alzheimer's or with ALS, we know what's gonna happen to them if we don't give them the treatment.

We have this incredible amount of data, know, hey Alexa, what's going to happen to Jane Doe in six months if we don't give them a treatment.

Now, are we going to have to keep continually putting people under control when we know the behavior of it?

We have incredibly good modeling.

So as this modeling gets better and better and better, we're going to be much more efficient because we don't need to enroll a hundred people on a counterfactual.

We know what's going to happen.

Is it AI?

Is it Bayesian modeling that does that?

Now we need to do that well and there's still always going to be value of randomization, but maybe we can do four to one randomization and evaluate that, yeah, what we thought was

going to happen on the placebo is exactly what would happen and we learn this.

So the world of health and wellness and decision-making and learning, I think is going to be

very different in 25 years.

It has to be very different in 25 years.

Yeah, yeah, that makes sense.

yeah, as you were saying, the more complex it's going to get, the more um statisticians and expert statisticians you need to make sure we're not getting fooled by the

interpretations of these very complicated models.

um Having that in mind, what do you recommend to

junior statisticians or students who are interested in this field?

Like what skills do you believe are most critical to thrive in the next generation of clinical trial design?

it's great question.

And I think some of these skills will be skills I don't have for certain, and it's moving in this.

uh

I know nothing about large language models and AI and all of this sort of thing.

I think that the ability, the statistical inference is still going to be hugely valuable and statistical learning.

So I worry a little bit if this shifts to the handling of data, the data science part without the science part.

So that's still hugely valuable.

ah

in it.

ah The modeling part's important, so the statistical part.

Handling of data, so the data science is going to be hugely valuable.

The ability to handle huge amounts of data, control it, use it appropriately in that for sure.

The science of this, you can learn the science kind of as you're going, but the background in science and medicine and chemistry and genetics and all of this.

I think is going to be hugely valuable within that.

So these are some of the skill sets to that.

But the interesting thing is that's sort of the tools to do it.

Really the two most important things when we look for somebody at Barry are communicate.

Can you communicate what you're doing?

And if you can't, it's limiting ah in that.

And it's a huge sort of challenge to it.

The other one is the passion for solving it.

It's that, you know, I don't sleep well when, okay, how am going to do this?

How am I going to do that?

So this, this passion to answer these questions.

And what's great in my field is people, you know, I use the example of you can do data science and modeling to optimize the size of a toothpaste hole to optimize profit.

You know, the bigger of it, maybe it doesn't really work smaller of it.

You know, we want to, you do you have passion to answer that question?

But if it's bringing better treatments to people, we're all patients.

You know, we're all, we're all patients.

And so there's sort of that natural passion to answer those questions.

So the passion, the communication, and then the various tools we talked about, and then there's these emerging tools about understanding AI, harnessing AI.

And the parts to this, it's a little scary in some circumstances, but understanding and utilizing it, you have to be able to do that going forward.

But it's hard to know where that's going.

It's such a new sort of science and its role in healthcare.

Yeah.

And I think this is fascinating.

This is going to be extremely interesting.

And actually I resonate with all, with all you've been saying.

So I've worked in a lot of different fields when I was at Palimpsil Labs.

Now I'm focused more on spots modeling.

Who knows what I'll do in the future, but yeah, at least in sport.

And I've seen that also in marketing in...

with the biostats clients I've worked with also everything you've just said where, well, you need to be, you need to be passionate about what you're doing because it's going to be

hard.

That interest and that drive to actually learn something new is going to, what's going to keep you awake at night at times.

I do agree with that, uh unfortunately, but that's also what's going to

give you the drive to keep going and actually solve the problem.

Ability to communicate, extremely important and going to be, I think even more important now that we have more data, we have more tools, we have generative AI tools that can help

us.

And then also getting familiar with generative AI tools, not only using them, but I think also understanding how they work um themselves inside.

is extremely important, at least if you're the model or if you're the statistician.

I've been doing that for the last few months and that's really helpful because that helps you debug and be the bridge and sometimes even, know, fine tune your model if you need.

um And I think fine tuning is going to be extremely important where now we can't even see like these LNMs become less and less.

Generally, then it's much more a matter of, okay, now we have a general list LLM, but we need to fine tune it to what we want it to do right now on these use case.

So I think this is also going to be extremely important.

And, just as you were saying, just one part of the skill set.

Well, Scott, that was, that was really fascinating.

Thank you so much for, taking the time.

Is there anything, any topic you wanted to mention that I, I didn't?

No, think all things I'm excited about and uh absolutely looking for people to jump in and join the quantitative fight, if you will.

Yeah, yeah, yeah, yeah.

And this is what we do here.

So definitely agree with that.

Awesome.

Well, Scott, before closing up the show, I'm going to ask you the last two questions.

I'll every guest at the end of the show, of course.

em

If you had unlimited time and resources, which problem would you try to solve?

Yeah, boy.

uh The question is would I jump into something sports related or would I jump into sort of science related?

I'm not sure I have a great response to that.

I'm not sure that there's, I think about if I retired, what would I do?

And I'm not excited about retiring.

I think there's a lot of...

sort of problems to solve within that.

ah What problems would I jump into?

There are sort of some sports things I'd like to come back to, some decision making uh within that.

ah But I think I'm too invested in designing better experiments.

uh The thing I'd be most passionate about is a learning healthcare system.

I think it's sort of the...

while we don't have that, that could make massive differences.

So if I had unlimited time and effort, I would create a learning healthcare system.

And here, what do you mean by learning healthcare systems?

So when you go in and you, go see the doctor and you say, if you've been thinking about all your problems, I'm not sleeping well.

Um, you know, the doctor might give you something to take, in that we don't learn.

about that patient.

Suppose at that point where we have seven different options we can give somebody, oh we randomize and we look to see how well that patient does.

We have data now about that patient and how well that patient does.

And now the next one that comes in and we're learning from that, and maybe we're learning about the sequence of things you try.

Maybe you try A and then you go to B.

This could be in mental health.

This could be in weight loss.

This could be in diabetes.

This could be in wellness care.

That right now we don't learn any of that.

So the doctor's vision of this is reading a journal publication of an incredibly expensive trial or the 14 patients they treated in the year or that.

What if this all came together and this was an integrated effort that everybody treated for this question goes into a data set and now

The 99.9 we don't learn from, now learning.

And that, by the way, the healthcare is gonna treat patients better.

Patients are gonna have better outcomes.

It's gonna be cheaper ah in that.

So there's this enormous potential for us to treat patients better and cheaper, lower the costs by learning from everybody that gets touched by the healthcare.

Okay.

Yeah, that would be fascinating.

ah What would be the first step?

Like if you could do that, if you could start this project today, what would be the first step needed to start that?

Yeah.

So we need the ability to get the data.

So that's outside of, so when we do an experiment, we create a huge database that we collect and we make people come in at three months and we collect all this stuff.

That needs to be naturally collected.

So we've got to be able to receive that data.

when we're getting better and better with electronic health records.

So jump into say the VA, jump into Medicare where we're treating all kinds of people.

And there are multiple ways to do this.

And in our system, we show that something's effective, but we never compare it.

You know, is A better than B?

That's never done.

But now...

in the VA or in Medicare or within some of the large healthcare systems, start with several questions that's relatively common.

uh That and start this real world randomized embedded experimentation where we can draw the data out and we can probably do this for pennies on the dollar compared to a true

experiment.

So

Do you have the data?

Do you have the patients?

Are there multiple options would be a very natural place to start.

Okay.

Yeah.

Yeah.

Damn.

Yeah.

That'd incredible.

You said unlimited time.

so I'll boil the ocean by learning healthcare.

That sounds good to me.

And second question.

If you could have dinner with any great scientific mind, dead, alive, fictional, who would be?

Yeah, you sent me that question.

So it's interesting.

I think the choice, and it's sort of an interesting story, is uh my father, his dissertation, he was at Yale, his thesis advisor was Jimmy Savage.

He was one of Jimmy Savage's last students before he passed.

And uh Jimmy Savage didn't see well and didn't write well sort of at this time.

And so unbeknownst to me, and I didn't find out to this week, he used to record on a tape recorder.

He would read my father's dissertation and he would talk into this tape recorder and uh he would send it.

My father was working in Washington, DC at the time.

This is 1970.

And he would send it to him as he was finishing up his dissertation.

Those audio tapes exist and still to this day they've been digitized and my father gave them to me a week ago.

So I've been listening to tape recordings of Jimmy Savage talk about my father's thesis ah within that.

And it's such a fascinating thing.

So I would love to have dinner with Jimmy Savage and the incredible things he did, the incredible sort, his sort of story that

That sort of uh front and center to me.

So he would be the first and I would love to have dinner with and hear his story.

sounds like a great dinner for sure.

I'd be great to arrange that.

I'd love to join.

Do let me know.

Fantastic.

Well, Scott, think we can call it a show.

Thank you so much for taking the time today.

um First, check out the show notes for this one if you want to dig deeper.

as usual.

And thanks again to Stefan for your recommendations.

If you folks have other guest recommendations, please let me know.

If you're a patron, you have access to Discord.

Otherwise, contact me on mostly on LinkedIn or email.

The other things I don't check.

But yeah, Scott, thank you again for taking the time and being on this show.

enjoyable.

Thank you for the invitation.

This has been another episode of Learning Bayesian Statistics.

Be sure to rate, review, and follow the show on your favorite podcatcher, and visit learnbaystats.com for more resources about today's topics, as well as access to more

episodes to help you reach true Bayesian state of mind.

That's learnbaystats.com.

Our theme music is Good Bayesian by Baba Brinkman.

Fit MC Lass and Meghiraam.

Check out his awesome work at bababrinkman.com.

I'm your host.

Alex and Dora.

can follow me on Twitter at Alex underscore and Dora like the country.

You can support the show and unlock exclusive benefits by visiting Patreon.com slash LearnBasedDance.

Thank you so much for listening and for your support.

You're truly a good Bayesian.

Change your predictions after taking information and if you're thinking I'll be less than amazing.

Let's adjust those expectations.

Let me show you how.

Let's get them on a solid foundation