Proudly sponsored by PyMC Labs, the Bayesian Consultancy. Book a call, or get in touch!

• Intro to Bayes Course (first 2 lessons free)
• Advanced Regression Course (first 2 lessons free)

Our theme music is « Good Bayesian », by Baba Brinkman (feat MC Lars and Mega Ran). Check out his awesome work!

Visit our Patreon page to unlock exclusive Bayesian swag ;)

Thank you to my Patrons for making this episode possible!

Yusuke Saito, Avi Bryant, Ero Carrera, Giuliano Cruz, Tim Gasser, James Wade, Tradd Salvo, William Benton, James Ahloy, Robin Taylor,, Chad Scherrer, Zwelithini Tunyiswa, Bertrand Wilden, James Thompson, Stephen Oates, Gian Luca Di Tanna, Jack Wells, Matthew Maldonado, Ian Costley, Ally Salim, Larry Gill, Ian Moran, Paul Oreto, Colin Caprani, Colin Carroll, Nathaniel Burbank, Michael Osthege, Rémi Louf, Clive Edelsten, Henri Wallen, Hugo Botha, Vinh Nguyen, Marcin Elantkowski, Adam C. Smith, Will Kurt, Andrew Moskowitz, Hector Munoz, Marco Gorelli, Simon Kessell, Bradley Rode, Patrick Kelley, Rick Anderson, Casper de Bruin, Philippe Labonde, Michael Hankin, Cameron Smith, Tomáš Frýda, Ryan Wesslen, Andreas Netti, Riley King, Yoshiyuki Hamajima, Sven De Maeyer, Michael DeCrescenzo, Fergal M, Mason Yahr, Naoya Kanai, Steven Rowland, Aubrey Clayton, Jeannine Sue, Omri Har Shemesh, Scott Anthony Robson, Robert Yolken, Or Duek, Pavel Dusek, Paul Cox, Andreas Kröpelin, Raphaël R, Nicolas Rode, Gabriel Stechschulte, Arkady, Kurt TeKolste, Gergely Juhasz, Marcus Nölke, Maggi Mackintosh, Grant Pezzolesi, Avram Aelony, Joshua Meehl, Javier Sabio, Kristian Higgins, Alex Jones, Gregorio Aguilar, Matt Rosinski, Bart Trudeau, Luis Fonseca, Dante Gates, Matt Niccolls, Maksim Kuznecov, Michael Thomas, Luke Gorrie, Cory Kiser, Julio, Edvin Saveljev, Frederick Ayala, Jeffrey Powell, Gal Kampel, Adan Romero, Will Geary, Blake Walters, Jonathan Morgan, Francesco Madrisotti, Ivy Huang, Gary Clarke, Robert Flannery, Rasmus Hindström, Stefan, Corey Abshire, Mike Loncaric, David McCormick, Ronald Legere, Sergio Dolia, Michael Cao, Yiğit Aşık and Suyog Chandramouli.

Takeaways:

• Epidemiology requires a blend of mathematical and statistical understanding.
• Models are essential for informing public health decisions during epidemics.
• The COVID-19 pandemic highlighted the importance of rapid modeling.
• Misconceptions about data can lead to misunderstandings in public health.
• Effective communication is crucial for conveying complex epidemiological concepts.
• Epidemic thinking can be applied to various fields, including marketing and finance.
• Public health policies should be informed by robust modeling and data analysis.
• Automation can help streamline data analysis in epidemic response.
• Understanding the limitations of models is key to effective decision-making
• Collaboration is key in developing complex models.
• Uncertainty estimation is crucial for effective decision-making.
• AI has the potential to enhance data interpretation in epidemiology.
• Educational initiatives should focus on understanding exponential growth and lagged outcomes.
• The complexity of modern epidemics requires a deeper understanding from the public.
• Understanding the balance between perfection and practicality is essential in modeling.

Chapters:

00:00 Introduction to Epidemiological Modeling

05:16 The Role of Bayesian Methods in Epidemic Forecasting

11:29 Real-World Applications of Models in Public Health

19:07 Common Misconceptions About Epidemiological Data

27:43 Understanding the Spread of Ideas and Beliefs

32:55 Workflow and Collaboration in Epidemiological Modeling

34:51 Modeling Approaches in Epidemiology

40:04 Challenges in Model Development

45:55 Uncertainty in Epidemiological Models

48:46 The Impact of AI on Epidemiology

54:55 Educational Initiatives for Future Epidemiologists

Links from the show:

• Adam’s website: https://kucharski.substack.com/
• Adam on Google Scholar: https://scholar.google.com/citations?user=eIqfmHYAAAAJ&hl=en
• Adam on Linkedin: https://www.linkedin.com/in/adam-kucharski-1a1b0225b/
• The Rules of Contagion: Why Things Spread - and Why They Stop: https://www.amazon.co.uk/Rules-Contagion-Things-Wellcome-Collection/dp/1788160207
• Adam's next book, The Uncertain Science of Certainty: https://proof.kucharski.io/
• LBS #50 Ta(l)king Risks & Embracing Uncertainty, with David Spiegelhalter: https://learnbayesstats.com/episode/50-talking-risks-embracing-uncertainty-david-spiegelhalter
• LBS #51 Bernoulli’s Fallacy & the Crisis of Modern Science, with Aubrey Clayton: https://learnbayesstats.com/episode/51-bernoullis-fallacy-crisis-modern-science-aubrey-clayton

Transcript

This is an automatic transcript and may therefore contain errors. Please get in touch if you're willing to correct them.

Today, am honored to host Adam Brzezarski, a leading expert in infectious disease modeling

and epidemic forecasting.

2

Adam is a professor of infectious disease epidemiology and co-director of the Center for

Epidemic Preparedness and Response at the London School of Hygiene and Tropical Medicine.

3

His research focuses on harnessing data and analytics to improve epidemic preparedness,

and he has contributed real-time analysis

4

governments and health agencies during major outbreaks, including Ebola, Zika and

COVID-19.

5

In this episode, Adam takes us inside the world of epidemiological modeling, discussing

how Bayesian methods help refine predictions and inform public health decisions.

6

We explore the challenges of modeling infectious diseases from data uncertainty to

real-time forecasting and the importance of communicating findings effectively to

7

policymakers and the public.

8

Adam also highlights common misconceptions about epidemiological data and dives into the

role of automation and AI in epidemic response.

9

This is Learn Invasions Statistics, episode 130, recorded November 26, 2024.

10

Welcome Bayesian Statistics, a podcast about Bayesian inference, the methods, the

projects, and the people who make it possible.

11

I'm your host, Alex Andorra.

12

You can follow me on Twitter at alex-underscore-andorra.

13

like the country.

14

For any info about the show, learnbasedats.com is Laplace to be.

15

Show notes, becoming a corporate sponsor, unlocking Bayesian Merge, supporting the show on

Patreon, everything is in there.

16

That's learnbasedats.com.

17

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.

18

See you around, folks.

19

and best patient wishes to you all.

20

And if today's discussion sparked ideas for your business, well, our team at Pimc Labs can

help bring them to life.

21

Check us out at pimc-labs.com.

22

my dear patients, hope you're doing well.

23

Two main announcements for today.

24

First and foremost, thank you so much to all of you who sent testimonials about learning

vision statistics and intuitive base to support my green card application.

25

I was genuinely touched, surprised and moved by the number, kindness and generosity of

your messages.

26

I am so happy and grateful to be surrounded by truly open-minded, kind

27

and helpful people like you.

28

I received more than 40 testimonials.

29

So thank you so much to all of you who've taken the time out of their busy day to work

about a very nerdy endeavor and how it changed the field of statistics according to you.

30

I will of course let you know what happens with my application and I'm already planning a

very special in-person surprise for all of you for February 20th.

31

6 but more info will come in good time.

32

So again, thank you so much and in the meantime, live long and prosper or shall we say

live base and prosper.

33

And talking about prospering, if you are interested in baseball and patient statistics and

working in an MLP team like let's say the Miami Marlins, so that would mean working with

34

me, if you'd love that, get to Miami, get to meet me and maybe

35

working together in the baseball research and baseball solutions teams.

36

Well, now is the time our teams are growing fast.

37

And this is your chance to get in on something that's very special.

38

Honestly, I absolutely love what's going on there.

39

And we've just opened two brand new roles one baseball analyst for the solutions group.

40

That's more here toward junior applicants and senior one, which is senior data scientist

in research group.

41

It's mine.

42

So

43

If you are passionate about baseball research, you love working on collaborative teams.

44

And well, bonus points, you know your way around advanced modeling approaches like for

instance, patient methods with the PIMC and STAN and neural networks, time series, all

45

that stuff.

46

Well, we definitely want to hear for you.

47

I think what we're doing in Miami is some of the most exciting in the MLB right now.

48

Let me know if you have any questions, if you're a patron of the show, you're on the

discord with me, feel free to send me any of your comments, questions, recommendations of

49

people, send that to your friends and maybe see you soon in Miami.

50

On that note, let's go on with the show.

51

Adam Kucharski, welcome to Learning Vision Statistics.

52

Thank you.

53

Yeah, thank you for taking the time and bearing with me, we've had a few technological

problems to record that episode, folks.

54

That was difficult.

55

But, you know, as they say, the obstacle is the way.

56

we are here.

57

And a big thank you to Chris Wyman.

58

to for putting us in contact.

59

Chris and I met at StunCon 2024 in Oxford.

60

Chris was actually in a panel discussion that I recorded with him and Lisa Semenova that

was episode 120 for people who are curious and want to hear more about how cool

61

epidemiological science and computational biology are.

62

So feel free to

63

to check that out.

64

With Adam today, we're actually going to tackle some similar topics, but you have a very

wide and very broad research.

65

So, yeah, that's why I'm super happy to have you on the show today.

66

You do so many things and you're also a great science communicator.

67

You've written several books, as I've said in the introduction.

68

So, yeah, all of these books will be in the show notes.

69

So people feel free to check them out.

70

Definitely recommend them.

71

are absolutely fascinating.

72

But before we touch on that, can you tell us with your three actually nowadays, how you

ended up working on these?

73

Yeah, sure.

74

So my work kind of bridges really understanding epidemics and getting better at predicting

and responding to them.

75

So there's this mix of aspects to that.

76

So some of it is

77

understanding the drivers of what we see in terms of how things spread.

78

So for something like dengue fever, might be climate influences, accumulation of immunity

in the population.

79

For something like, I say flu or COVID, the implications of vaccination over time, how

those viruses evolve.

80

Alongside that, we're also doing a lot of work to build up the methods and tools that we

need to respond to very quickly.

81

COVID, lot of these things were developed very quickly.

82

often just working over weekends.

83

Can we actually do a lot better, particularly for the predictable questions that we're

going to have to answer?

84

In terms of how got into it, my background's originally in maths.

85

I applied maths to my PhD.

86

But even that was starting to go more in this direction of epidemiology and questions

around how we can understand the process of infectious disease.

87

And for me, as a field, it kind of sits quite nicely between something where having some

mathematical and statistical understanding

88

can give you a lot of value very quickly.

89

But also there's enough unknowns about the underlying rules and processes.

90

It's not something like physics where we have a lot more mechanistic understanding.

91

So it means squeezing a lot more out of the data you have available, particularly under

pressure in a situation like an epidemic.

92

I see.

93

OK.

94

That's very interesting.

95

So it's something that was a long time coming.

96

You've always kind of

97

been interested in these topics, at least since your past high school studies, right?

98

Yeah, I think there's a lot of those questions where it obviously has an enormous impact

on people, but also just from a kind of curiosity point of view, these are quite hard

99

questions.

100

And often the methods that are in the textbooks don't quite work for what you need to do.

101

it's an ongoing, interesting

102

research area to be in because almost every epidemic you work in, what you thought you

knew and what you thought you'd solved, suddenly you happen.

103

And so yes, kept me busy and kept me interested along with many of my colleagues.

104

Yeah.

105

And what about patient stance?

106

Do you remember when you were first introduced to Zain?

107

So I think it was particularly during PhD.

108

I think my kind of undergrad was a lot more

109

traditional maths, so relatively little statistics, a lot more kind of theory.

110

So the sort of thing where you learn measure theory rather than learning the kind of

coming at it from a data point of view.

111

think during my PhD I started working a lot more with data and especially if you start to

look at processes involved in diseases, processes involved in immunity, the natural next

112

question is how can we estimate things meaningfully within that?

113

And then particularly if you have

114

patchy bits of data or different studies that you want to combine in some way, or if you

have some analysis you've put together and then you want to make statements about

115

additional data coming in, Bayesian statistics are a natural framework for doing that.

116

And so really in the mix of my work, in some cases you just get simple probability

problems and Bayes' formula is a nice get out of jail way of solving that problem.

117

And in other cases, it's more of a framework for thinking about the entire analysis.

118

You want to be able to come to the best possible explanation or the best possible evidence

at a point in time.

119

But then you want to update that as you get more data coming in.

120

And again, it gives you very nice toolkit for doing that.

121

Yeah, Yeah, that makes sense.

122

Damn.

123

That's very fascinating.

124

That brings me a lot of...

125

More questions for you, Adam.

126

Yes, one of these questions that I for you is, especially during COVID, Bayesian modeling

became a crucial tool.

127

I remember I did some work at my very low level at that time, but you did a bunch of work

on that.

128

as you were saying, I imagine there were lot of very short nights and very long weeks.

129

Can you share a specific example of how

130

these models were able to inform public health decisions?

131

Yes, I think there's quite a few instances, particularly around scenarios, where models

can be a very useful decision support tool.

132

Because essentially, if you're going to make a decision about what to do in an epidemic,

everyone has a model in their head.

133

Because if you think, let's do this, let's try this, you're making some assumptions about

what you think is going to happen.

134

You make some assumptions about how you think epidemics work.

135

So models.

136

a very nice way of allowing us to really write down and formalize what we think those

processes are and what we think the intervention is going to do.

137

And then we can debate whether that's reasonable, whether that's not.

138

We can see if that generates any counterintuitive effects.

139

But in doing that, you really want to capture the extent of information you have about the

problem you're dealing with.

140

So even, for example, the question of how much effort is it going to take to get

transmission down?

141

Well, one of the key things that

142

going influence that is how much transmission there really is.

143

So a lot of the analysis that we did, being able to pass around uncertainty was really

important.

144

So very early in the pandemic, for instance, because we were relying on quite uncertain

data from China, quite uncertain data about exported cases, we had some sense of what

145

transmission would do in a country without any control measures, but there was quite a

large amount of uncertainty on that, probably reproduction number.

146

somewhere between two and four.

147

And so what we wanted to do is when we generate UK scenarios to present them, we didn't

want to give one number.

148

We wanted to say, look, we don't know actually it's more the upper end or lower end of

this, but we want to kind of define that uncertainty in what we simulate.

149

And then after that wave comes through, we had lockdowns like many other countries, there

was this push to reopen.

150

And then the question is, okay, so we've had that wave.

151

what's the reopening going to look like?

152

And again, that's where these kind of basic methods will be very helpful because we can

take the uncertainty and perhaps we've got more confidence now about what we're dealing

153

with, but then we want to pass that level of confidence into what's going on in future.

154

And then, you know, as variants started emerging, that became even more important because

often for Alpha, we would have some degree of confidence about how much more transmissible

155

it was, how much more severe it was.

156

And in any analysis, for example, we did a lot of work early in 2021 of how quickly you're

to have to vaccinate.

157

And I think there's a lot of pressure to lift lockdown and try to understand that

trade-off.

158

If you're vaccinating at this rate and you're lifting lockdown this quickly, what's that

going to look like?

159

And again, none of those values you had very precisely and you had more epidemic data

coming in all the time.

160

particularly that process of reopening in 2021, there was a very tight relationship

between models and policy because

161

In the UK, was one of the few times where the policy was actually very informed by what

was useful from a technical point of view.

162

So what they did was they had these series of steps in the roadmap.

163

It was designed to give enough signal in the data post each step that if something was

going very wrong, they wouldn't have implemented another step before you had that signal.

164

So they were kind of deliberately spaced out so the models and the epidemiologists would

have enough time to work out what that's.

165

relaxation had done.

166

So again, from a basic point of view, that's really nice because it gives you enough time

to kind of update your posterior sensibly before you do the next step and work out what

167

effect that's going to have.

168

Yeah.

169

Yeah.

170

Yeah.

171

Okay.

172

That's fascinating.

173

I didn't know there was that level of coordination where you could actually do that.

174

There are many examples where it didn't work in such a coordinated fashion, but think that

roadmap reopening was one where there was a much tighter relationship between the

175

questions coming down from policy and what models needed to say something sensible about

the implications.

176

Yeah.

177

Yeah.

178

And I guess, I mean, I was wondering how much do you think the pendulum has swung back

from then?

179

Do you think we'll be faster?

180

to implement these workflows and improve them next time there is a pandemic?

181

Or will we have to work from a blank slate because politics is so short sighted with very

short cycles?

182

Yeah, I think that's a really good question.

183

I think there's...

184

There's some things that have been positive in terms of progress.

185

So I think in terms of what we're doing in others in consolidating a lot of the tools that

are available.

186

So some things now that we, know, looking at quick questions that H5N1, questions that I

just wouldn't have bothered doing previously just as a curiosity, because it would have

187

been three, four hours just for a maybe question.

188

And now in 20 minutes, you can get a rough example.

189

So think that's quite nice is bringing things into reach.

190

But I think in terms of just

191

staff capacity and people to do this work.

192

I think there are a of people who, you know, put a huge amount of time in often as around

the world, getting pulled off different roles and off other projects.

193

And so I think in a way in that sense, we probably don't have the same workforce who could

undergo that amount of pressure for that amount of time.

194

in a way that creates the necessity of we need to get better tools because we don't have,

I think that.

195

volume of people, also just all our projects.

196

we basically took people off a lot of other funded projects.

197

And I think we wouldn't have the resource to do that in the same way.

198

So for me, think particularly what we're interested in at moment is if you imagine the set

of tasks that you might have to do in an epidemic, some are tasks that a lot of other

199

people have to do.

200

And we can really predict ahead of time you're going to need to do that.

201

So stuff like working out the severity of infection or working out the transmission.

202

We know we're going to need that.

203

that done.

204

We know lots of people are going to do that.

205

So that's a really good task for automation and getting really leading edge, know, basic

methods standardized so people can ever want to do that.

206

We don't all have to duplicate effort.

207

There's other things that could be really specific and maybe within the country, there's a

kind of subgroup that's particularly important or there's a certain variant or something

208

you might want to deal with.

209

And that's going to require more domain knowledge.

210

And I think at the moment, even if talk to people

211

at different outbreak organizations around the world, they probably spend six to 70 % of

their time on quite low level, predictable, data wrangling type questions rather than the

212

sort of 30 % of expertise led questions they want.

213

I don't necessarily think in a pandemic people would end up working less hard because

everyone wants to contribute as much as they can.

214

But I think what I'd like to see is a lot more time on that 30 % and going much deeper.

215

into what we can understand rather than spending a lot of time even just trying to get

ahead, a sense of the data and the basic tasks to answer very simple questions.

216

Yeah, yeah, yeah.

217

That makes sense.

218

Not sure I'm really reassured by that answer.

219

But yeah, that's actually a realistic answer.

220

Something I'm also really curious about, because I see that a lot as a modeler myself, is

where the common misconceptions that you've seen the public or elected officials or even

221

professionals have about epidemiological data.

222

and what the scientific community can do to clarify these misunderstandings.

223

Yeah, that's a really good question.

224

I think one common misconception is even just what data are as a thing.

225

I think often when people talk about raw data, they're talking about inferred estimates.

226

So think one example of this is excess stats that I the media often treat as a measured

thing.

227

rather than actually a counterfactual.

228

And I think I remember talking to journalists about the comparison with flu and I think a

lot of them discovered that the flu mortality statistics they've used every year actually

229

involved some quite big modeling assumptions about seasonal patterns behind them.

230

And there isn't just this magic number that we measure.

231

I think similarly with lot of the estimates, for example, in the UK they ran randomized

testing surveys.

232

they randomly tested people in the community each week and reported it.

233

And that number, that wasn't a raw proportion because, of course, you needed to do some

standardization across groups and weighting.

234

And in some cases, because it was quite noisy across multiple panels, they would have a

kind of smooth underlying proportion and then infer that.

235

So it was sensible modeling, but there was modeling behind that.

236

And I think the misconception often is that these things are raw data.

237

often, even a very simple thing, if you calculate

238

If you run a survey and calculate proportion, you're making some modeling assumptions.

239

Or if you know, oh, we're going to adjust for age.

240

Why are you adjusting for age?

241

Why are you not adjusting for something else?

242

You're making assumptions about what you think are important drivers of that thing you're

measuring.

243

So I think that's probably one thing we can get better at communicating is that actually

this distinction between models and data isn't that simple.

244

That a lot of things we think of data are actually modeled estimates.

245

And actually a lot of the things that sometimes people treat as

246

super complex models are actually just quite straight, simple steps for thinking about

data.

247

So even one of, if you've got a bunch of hospitalizations over time and you want to know

when the infections happened, you can use quite simple, like a decomposition model just to

248

take that and work out when the infections are.

249

And I think often in public perception, when people talk about models, what they mean is

very complicated scenario models that people are assuming are used to make forecasts.

250

So I think it's often the of a crystal ball of use this big complex model to say what's

going to happen.

251

And I think in reality, what happens is often the more complex models you use for

scenarios and like what if, because we can't, we can't make forecasts often in pandemics

252

as well, because you'll be forecasting what policymakers are going to do.

253

And if your models are used as tools to support their decisions, it becomes quite odd to

use that as a forecast.

254

I mean, I think the example I sometimes get

255

is that imagine you're playing a game of poker and at any point in time you want to

understand the implications of your decisions.

256

What you don't really want is someone on your shoulder saying, I bet you're going to fold

later this round.

257

What you want is someone who can say, look, if you raise, this is likely the situation

you're going to face.

258

This is the risk you're taking on.

259

If you fold, is potentially what the outcomes are going to be.

260

So that's really how a lot of these models are used in decision support.

261

But I think in public consciousness, they're often seen as this is

262

we're saying this is going to happen in the future.

263

And I think part of it is communication and particularly, I think, getting people to able

to play with those simple models can be very helpful.

264

So they realize it's not this super complex thing.

265

It's probably what they're doing in their head, but just writing it down in a bit more

structured way.

266

Yeah.

267

I find also, indeed, walking through scenarios

268

is something that's really helpful to people because, well, they can imagine the scenarios

and that's much more tangible and concrete than numbers and possible distributions.

269

can see that a lot.

270

mean, sports is a lot like that, especially baseball, lots of different discrete

scenarios.

271

Yeah, and I think often people, particularly with epidemics,

272

are kind doing it in their head all the time.

273

So the epidemic is going up, and you say there's going be a problem, and then you get a

few data points that seem to be tailing off a bit.

274

And basically, everyone's doing that updating in their head of where they think it's

going.

275

I would have loved to have seen more politicians and journalists actually write down

their, you could get them to actually just write down their prediction and their kind of

276

distribution of what they think is going to happen and then see how there's update.

277

And then you could actually

278

give them a better understanding of what's going on in their head relative to actually

what's plausible given the data that's coming in.

279

Yeah.

280

But so I completely agree with that.

281

And I really love that.

282

But I think here the incentives are really bad for politicians, right?

283

then if you like if you publicly and privately, think I that's a lot.

284

I think that's a really good point.

285

I think that's a really good point.

286

I think getting politicians to write down what it's going to happen publicly is very

difficult.

287

But I even privately, think there were a lot of people probably not doing that.

288

So I mean, amongst colleagues and stuff, we sometimes just had just probably more things

that when you think this study comes out, what do you think it's going to show?

289

And I think it's sometimes quite, we like to kid ourselves.

290

So even if it's just writing it down for yourself, I think it's just quite useful to have

that to effect on.

291

Oh, yeah.

292

No, I mean, I

293

I'm like completely on board in helping people develop a more probabilistic thinking.

294

You I think you doing the work you're doing and also the communication work you're doing

is very important.

295

All your books.

296

That's also why I have these podcasts.

297

People like Nate Silver writing books is very important too.

298

I don't know if you read his last book On the Edge, but that's very important too for

that.

299

Right.

300

And I think that'd be awesome if we were gearing towards that direction.

301

But yeah, the problem is like the incentives in the public incentives in politics are so

bad when it comes to that, that you actually have a much better standing if you just, you

302

say anything and everything and just are not

303

held accountable for that, then actually betting on something that would happen and then

changing course if actually what you said would happen did not.

304

Yeah, I think it's partly from the communication.

305

There's also, yeah, what are the things that we can do meaningfully and what are the

things that you can encourage people to do versus just wants to be feasible.

306

And I think also just from a modeling point of view, sometimes there's this idea that

307

for political decisions, you should have this big model with absolutely everything in and

all the kind of the weights of how you do everything.

308

And I can't see any political party wanting to do it because ultimately those things are

going to be weighed, not in a kind of written down, you know, we're going to put 10 % on

309

this and 15 % on this.

310

And so I think it's it's working out where the science can be really informative and where

actually there's more of that kind of human political element and that's not.

311

the best possible to be fighting at this point.

312

Yeah, yeah.

313

Yeah, it's a great point.

314

actually, your talking about your books in the rules of contagion, you explore why things

spread.

315

And I really love that.

316

So can you tell us a bit more about that?

317

And how does patient thinking help in understanding these patterns?

318

especially for diseases.

319

Yeah.

320

So I think that's something striking writing the book is I set out thinking that there

would be these analogies in other fields.

321

And I wasn't sure necessarily how strong they would be.

322

But the more I dug into that, I realized actually they were in many cases very explicit

and very directly informative.

323

For example, after the 2008 financial crisis, there was very

324

clearly described epidemic thinking that drove a lot of the interventions we saw, things

like ring fencing, things like capital requirements for banks at risk in network.

325

It's really about thinking about it like a contagion problem.

326

Similarly, if you dig into the history of companies like Buzzfeed that were very good at

generating viral content, they were actually writing research reports on how you evaluate

327

the reproduction number of marketing campaigns.

328

And actually,

329

I subsequently discovered that Buzzfeed journalists would have a equivalent to the

reproduction number as a metric for their articles.

330

So it wasn't just this quite fuzzy comparison.

331

Actually, this was the same bits of theory that were appearing between these two fields.

332

I think one thing I find quite interesting on the Bayesian angle in how things spread is

particularly some of the debates around how you convince people and how

333

people adopt beliefs and they take off.

334

Because there was quite a popular idea for a while known as the backfire effect, which is

where if you try and convince someone, you can end up basically just strengthening their

335

pre-existing belief.

336

And this idea that attempts to change people's beliefs can kind of backfire, which

obviously doesn't bode well for any kind of social progress, because it's this idea if you

337

try and convince someone that marriage equality or something is good idea, it's just going

to lead to them being entrenched.

338

But what subsequently happened is a lot of the work, both on the applied side of actually

people trying to get these, get support for these kind of this progress, but also on some

339

of the scientific side of people studying them, suggested it's actually much closer to a

Bayesian problem, that it's not that you're leaving people to entrench their beliefs,

340

rather, if you give people weak evidence, you're not going to shift their distribution

much.

341

Which I thought was kind of really

342

Interesting.

343

And I hadn't actually thought about it quite much in that way.

344

And it kind of makes sense that even if you've got like a quite strong prior for

something, if someone gives you evidence that agrees with that prior, it's going to look

345

pretty similar afterwards.

346

And especially because we're not doing all those calculations in our head.

347

We're just sort of seeing the feeling we come away with.

348

And it really struck me that actually in those situations, we're probably much better

evaluating the effects of evidence that we disagree with because the posterior, you know,

349

just

350

Mathematically, you expect the posterior to move more.

351

And so perhaps the situations where what people were thinking was a backlash effect is

more just we're better at critiquing evidence we disagree with rather than the ones that

352

kind of line up.

353

Because if it agrees with us, we're going to walk away with the same opinion afterwards

anyway.

354

So yeah, I think now it's obviously a much harder problem to study in terms of spread of

beliefs and so many factors that can play into that.

355

But yeah, there's still a lot of ongoing debate on the extent to which people's

356

adoption of beliefs and behaviors of these Bayesian versus some other factors that kind of

explain how those are updated over time.

357

Yeah, I see.

358

That's really amazing.

359

I love that.

360

I didn't know that example about...

361

What's the website you were saying?

362

Oh, BuzzFeed.

363

Yes, thank you.

364

It's remarkable.

365

Yeah.

366

And it was coming to it because they did campaigns about Hurricane Katrina.

367

None of these were viral.

368

And this is one of their key findings that this wasn't like COVID where it spread and

spread and spread.

369

But for 10 shares, they might get an extra seven or eight.

370

So if you get it, if you spark lots of little clusters of sharing, you might.

371

actually get quite considerable additional uptake as a result.

372

I think there was one that was marketing campaign for detergent and it just wasn't

transmissible at all basically.

373

So there was quite a nice quantifying that there's certain things that people obviously

want to tell people about and others that even if you got the biggest marketing budget in

374

the world, you're going to struggle to make washing up liquid contagious.

375

I see, see.

376

Yeah, yeah.

377

You know, something I'm curious about and I ask this question to Chris at StancOn is,

concretely, what does it look like for you to work on an epidemiological model?

378

Like, who are you talking to and what's your workflow and technical stack?

379

Yeah, that's a really good question.

380

So I think generally,

381

anything we build starts with a problem and often that problem either comes from someone,

if it's a policy related question, it comes from someone policy side, maybe scientific

382

advisors to certain agencies in the case of like an applied organization like MSF or WHO,

it will come from a representative who we're working on maybe that outbreak or that

383

situation and there'll often be quite specific things that people are interested in.

384

So it might be a forecast of actually what are we dealing with?

385

It might be that there's a plan to implement vaccination or control measure.

386

A lot of the work we did in Ebola, for example, different control measures being proposed

and people wanted some idea of the relative impact that that would have.

387

There's also, of course, just on the scientific side, so the work we've done around

effects of behaviour and immunity, it might be that you've got lab colleagues who've

388

noticed some interesting features and want to work out how can I get

389

sufficient estimates out of my data.

390

mean, I mean, this is, I think another example where Bayesian thinking is very helpful

that you might have, say, lots and lots of antibody responses.

391

And you don't want to analyze them all as like individual data sets, because there's going

to be some commonalities in just how the biology works between individuals.

392

So having this kind of hierarchical models can be very powerful, because you can have

shared information just on the underlying dynamics across the population, but you can also

393

have individual features of

394

what we looked at previously and this sort of thing.

395

In terms of developing the kind of model and the stack, it depends a bit on whether it's

similar to another problem we've seen.

396

So in case of a policy question, often what we'll do is, in the many, many models and

things we've dealt with previously, you'll find the thing that's most similar.

397

And I think increasingly, we're seeing progress in libraries.

398

So a lot of the work we do is in R or in some cases, using kind of Stan.

399

backends.

400

And so you'll take something that maybe is from the library that you think is most

appropriate for that and then a model that's templated up that's closest to that.

401

And then ideally you'll just plug in and work through it away.

402

Often you might either adapt some features of the model process in terms of say how

transmission is happening or what groups are affected or you might bring in data.

403

So you might have a model set up from UK contact structure and you'll just encode data

from some of us so you can adapt it.

404

In other cases, you might get something that's a very specific, quite niche question.

405

you know, for example, I don't know, like the ones we're doing during COVID, testing

people at certain types of gathering or something.

406

And that's not something necessarily you have a model off the shelf, but in some cases,

the equation is quite simple to write down.

407

You if you've got this many people and you're testing this many at this point in time.

408

And in that case, we might just build something more bespoke.

409

And I think that...

410

The challenge always, mean, like with any kind of software development problem is to what

extent do you do something quick for a very specific problem versus take on coding that if

411

you've got to that problem repeatedly.

412

And so I think we're getting a better sense now.

413

There's certain things that are sufficiently complex.

414

There's large enough scope for bugs.

415

It's gonna be useful to enough people where actually that makes sense to package it up

more because this is a library.

416

And there's other things that actually are simple enough and transparent enough and kind

of

417

bespoke enough that you don't want to build a software tool for every single one of those.

418

Some of those you can just do quickly as the problem arises.

419

Yeah, yeah, yeah.

420

That makes tons of sense.

421

once you're like, I'm wondering also that the size of the teams in these cases, because

obviously each time I talk to a modeler in your field, it sounds like the models are

422

really

423

big and huge and take a lot of time to work on because they are so complex.

424

So yeah, I'm wondering how many people does it take to work on a model and how do you

actually do that?

425

Because like, yeah, is everybody working on the model at the same time?

426

Do you have some team for that part of the model, another team for the other part?

427

How does that work?

428

Yeah, so I think, I mean, to give the example of some of the big purpose scenario models

that are used by our group in the UK,

429

I haven't got the github we've off from ever.

430

It's probably at least 10, 15 people who've made substantial contributions to that co-base

at the various points in time.

431

And in some cases, can be, mean, ideally we'd make these things as modular as possible.

432

So early on, for instance, I had some colleagues who were focusing very much on the

transmission dynamics and kind of how people interact, what interventions were going to

433

be.

434

And I worked a lot more on the disease burden module.

435

So once you have transmission infections,

436

you can convert those infections into an estimate of how many are going to be

hospitalized, how long are they going to be hospitalized for.

437

So then you have that kind basic model structure.

438

But then over time, that got expanded because variants came in, vaccines came in, and you

end up with multiple versions of those models.

439

And there's always this kind of challenge of, do you make

440

kind of one core model that has sufficient flexibility to all those problems, or do you

kind of fork a model and use it for a special case and you're not going to use it again?

441

So that model's actually been used in multiple countries.

442

We adapted it to a whole range of different settings, some cases with very unusual

patterns of immunity and infection.

443

And in that case, it didn't make sense just to build that into the original model because

it was just such a specific example.

444

But I think that's one thing where, as well, if we had to go back now, because there's so

many versions of the model and so many applications, because it's real time, we have to

445

deliver that very quickly, it's obviously harder to now say, how would we do that for flu

epidemic?

446

And so I think what we're trying to move more towards is these very modular examples that

you can plug in all the bits you need, but also have that capacity to adapt it.

447

And I think that's just that's kind of an ongoing challenge that you can have these things

that are very stable and structured, but very hard to adapt.

448

Or you can have these things that are very maybe flexible and easy to adapt, but not

necessarily as kind of efficiently structured as you'd like.

449

mean, there are examples as well of modeling where it might be one or two people

developing something quite quickly or just making use of a library.

450

mean, some of the popular methods, for example, for estimating reproduction numbers.

451

that tool would have been used by a huge amount of people, but obviously the active

contributors to the development might be a lot smaller.

452

Okay, okay.

453

Yeah, I see.

454

Yeah, so big diversity in the size of the projects.

455

And do you have a favorite type of models, actually, that you'd like to share with us?

456

So I one, well, I think one of the models that

457

I think it delivered a lot of value in various things that we've worked on over the years.

458

actually it was both on some of the H7N9 analysis we did about 10 years ago, the outbreaks

in China where you had infections coming from poultry and potentially human to human

459

transmission as well.

460

It a big question of looking at that human data, how much was coming from spillover, how

much was coming from human human infections.

461

And we actually got an analogous version of the problem with some of the COVID variants.

462

So for Delta, how much of this was imported cases from India versus transmission

establishing in countries?

463

And in both those situations, so first of all, it's a bit of a modeling headache because a

lot of traditional models are structured in a way that basically says, if you've got your

464

cases over time, the new cases that appear, it has to be one of those past ones that

infected them.

465

So if you look at a lot of the common calculations for how you do reproduction numbers,

it's known as a generative model.

466

So in other words, or renewal equation is the kind of specific version.

467

So the new infections are the products of the infections that come before.

468

Someone has caused that infection.

469

If you've got importations or spillover, that's no longer the case.

470

You've actually got this additional term coming into your equation.

471

So it makes it a trickier inference problem if you've got two routes just to infecting

someone.

472

But it can also be a more powerful one because

473

In the case of avian flu, we knew when the wild lipultary markets were closed.

474

And in the case of Delta, we knew when the travel ban against India was implemented.

475

So what you've got as an estimation problem, you've got two things that are influencing

your infections, but you know the kind of shape of one of these because you know when the

476

markets close, you know when the flight patterns came in.

477

And suddenly that gives you a lot more estimation power on the thing you care about, which

is human-to-human transmission.

478

So I think for me that

479

That's just a really nice example of a model that is not super complicated.

480

It's relatively easy to explain.

481

like there's two things that can affect someone, which is it?

482

But actually, you can squeeze a remarkable amount out of your data as soon as you know the

shape of one of those processes.

483

OK.

484

does that?

485

I mean, it obviously comes with.

486

significant challenges of using these models.

487

Can you detail a bit these challenges that you face when creating such models?

488

Yeah, I think there's...

489

whole mix.

490

mean, in some cases, there's just understanding the process you define in model that for a

lot of outbreaks, there's a lot of things you'd like to know the role of in a process.

491

So, for example, how different age groups are interacting or affected by certain things.

492

But in some cases, you might not have the data that you need to actually

493

So a good example in COVID was a lot of people, know, the epidemic would come down, a lot

of people would argue about why that was.

494

And actually, if you just look at case data or just look at deaths, it could be immunity,

it could be changes in behavior, it could be something to do with the climate, it could

495

be, you know, a whole range of different things that could influence that transmission.

496

And just looking at one time series of outcomes, you can't distinguish what those are.

497

You really need some data on antibodies.

498

You need some data on social mixing to tell you which of those is the most likely

explanation.

499

So I think that's often a big challenge is where you have lots of potential explanations

and you can't actually untangle them.

500

mean, the other, to use H5N1 as a current example, unlike the H5N9 and Delta where we knew

how that shape of the introductions was changing, we don't know that for H5.

501

So now you get cases popping up and they they haven't had a...

502

contact with poultry, maybe it's a wild bird, maybe it's a human, we've got no idea.

503

And I think that's kind of a big challenge.

504

Almost a model can't really tell you anything at this point because the data is so

uninformative about the process.

505

There's lots of questions at the moment, get journalists who ask, you're building models

of H5.

506

Actually, outbreak data is just too bad to say much.

507

We can design some what-if hypothetical models, but I think that's much harder.

508

I think there's also just the technical challenge that

509

you know, in terms of just making sure that the models you build are without bugs.

510

And then also edge cases is another classic one that there's, there can be sometimes some

slight counter-intuitive things, particularly if you're doing with kind of, know, delay

511

processes.

512

So one example, which I think it was sort of a communication one, but even, you know, if

you have an epidemic that's going up and you suddenly stop it and you have a delayed

513

outcome like that.

514

the peak in when the epidemic stops isn't the same as the peak in death because you're

doing a kind of delayed convolution.

515

We're doing a smooth out of lay distribution.

516

And so features like that, that making sure that you've actually got that relationship

defined properly.

517

And again, this is, think, why the move to a lot more established libraries rather than

trying to make sure these things are bug free in real time.

518

OK.

519

Yeah.

520

Yeah.

521

And a question I often get, you

522

personally, and I'm guessing you are getting a lot too, is, okay, cool.

523

I get uncertainty estimation, you know, with the Bayesian models.

524

Why do I care?

525

So I think a lot of what you want to do, particularly decision-making, comes down to

confidence and evidence.

526

So particularly in the middle of an epidemic, there's a lot of things we don't know with

any confidence.

527

So I guess on the one hand, you could just ignore it and just go for a point estimate.

528

Is it going up?

529

Is it going down?

530

But particularly if you want to ask, is the epidemic under control, it's not very helpful

to have a yes, no always.

531

That you might want to say,

532

Yes, it's under control.

533

How confident are you that it is?

534

And again, having that uncertainty in a reproduction number, if you like, well, 100 % of

our density is below one.

535

And that's what we had post lockdowns and social mixing data, for example, in the UK, that

there was uncertainty in that distribution, but all of that density was below one.

536

So the conclusion was we're very confident that transmission is coming down.

537

I think similarly for some of the variants, we found that

538

you'd get uncertainty in the estimates.

539

So maybe it's 30 % more transmissible, maybe it's 40%, maybe it's 50.

540

You can be very confident it is more transmissible.

541

And I think it's a difficult one because policymakers sometimes love single answers.

542

They don't want a kind of vague.

543

But I think particularly if your uncertainty lands either side of a particular threshold

that matters,

544

that can in a way give you more confidence in communicating and saying, yeah, look, this

is definitely going up or this is definitely, almost definitely under control.

545

It's similar, know, the clinical trial, you look at the confidence interval, I think is

the equivalent of that.

546

You want to know, you know, how much can you rely on this estimate?

547

And I think that's where the uncertainty really comes in.

548

Yeah, yeah, so basically, making sure

549

that you're not fooled by the variance of your processes.

550

Yeah, and think especially when you're dealing with exponential processes, errors can

accumulate.

551

There's something that might feel quite small of, oh, the transmission rates this.

552

And actually, if it's slightly higher or slightly lower, you get a rate of an outcome.

553

Even something, say, as a very simple example, each person affects two others, and you cut

transmission in half.

554

So each person affects one other, just another one other.

555

a tiny amount of uncertainty there after two months could be the difference between a

massive epidemic and a handful of cases.

556

And if you don't communicate that, people will be asking, well, why is there a massive

epidemic when you said it would fit it out based on your best estimate?

557

yeah, yeah.

558

That's definitely a good point.

559

I'm curious also to have your

560

to hear your thoughts about the latest advancements that we've seen in the last year in

artificial intelligence and large language models, because I'm guessing this is going to

561

have also an impact, hopefully for the best, on epidemiology.

562

So yeah, with these advancements, how do you see the future of the field of epidemiology

and

563

and the role of patients stance in it.

564

Yeah, I think there's a huge amount of really promising development.

565

I think a lot of it at the moment that we're working on and most of us is trying to find

where do these solutions work.

566

I think essentially we've been given this increasingly amazing toolkit, but in many cases

it wasn't necessarily developed, these are problems we're working on.

567

And so finding where the applications works can be very powerful, where are the ones that

it's going to struggle.

568

And so even, if you look at AI or some more traditional machine learning approaches that

there might be situations where if we're very interested in the mechanism and we've got

569

some process we can find our model, perhaps that estimation or that prediction approach

isn't optimal for actually what we're trying to solve.

570

But think equally, the pandemic we showed, there was a huge amount of data out there that

in many cases we weren't.

571

able to interpret in a meaningful way.

572

you might have some like a social contact and if it drops by a certain amount, you can put

that in a model with a very meaningful as meaningful change.

573

But you might have quite a lot of very noisy data, which is giving some indication about

how people are behaving.

574

But you can't extract those features and weigh them in a useful way in the same sense that

an AI model could do.

575

So I think it's in a way finding a range of problems.

576

think

577

broadly the challenge we have for outbreaks often they're quite rare events.

578

So even near what you validate against and what's your input and what's your kind of

outcome that you're trying to predict.

579

But I think within a epidemic, especially you get larger clinical data sets, behavioral

data sets, a lot of value there.

580

I think also some of the methods like universal differential equations and other things

coming through where it's taking that transmission model structure, but then incorporating

581

things like neural networks to allow for more complexity

582

in understanding the patterns that are kind going on alongside that.

583

think there's, yeah, there's a lot of really interesting progress there.

584

think just also just more generally in the field, there's some of the AI models that

essentially are learning features that can, mean, weather forecasting is one example where

585

it's many, many times faster and less energy intensive for simulations.

586

think again, that kind of, you can find ways of of approximating more complex models.

587

I think just also in terms of the data that comes in, mean, lot of outbreaks often

described in narrative reports where it's kind of like a few paragraphs or so and so did

588

this and went here and did that.

589

And one of my colleagues did some nice prototyping, but he brought in this, that's all

local models that can take those quite difficult to interpret narrative reports and

590

convert them into structured data, which you can then perform standard analysis on.

591

And then you've got the issue of accuracy and validation.

592

But I think there's a lot of

593

work really across the spectrum, both in the models themselves, but also how you're more

efficient in what you're doing.

594

Yeah, yeah, OK, I see.

595

So basically, like in a lot of other fields, you would like to see a better communication

between these models, DLM models, in

596

the humans in the loop, but clearly you don't see these kind of models being made entirely

by artificial intelligence.

597

No, we've tried it actually.

598

even things like co-pilot workspace, which look, you try them out on a simple problem and

they look really, really cool.

599

So the idea that you can just give it a code base and tell it to do things.

600

But often for quite specific things, if you say, build me a model with these features to

do this, I think because just the

601

the training set just doesn't include anything resembling or very little resembling that

kind of problem.

602

So if you want an LLM to build you some JavaScript, it's pretty good because there's just

so much that's trained on it.

603

If you want quite an image, a practical model of an epidemic, it sort of struggles in some

ways.

604

So think finding the shape, it's great for, I've got all these functions, can you package

them up and add the documentation and this kind of stuff?

605

It makes a lot of tasks faster.

606

But I think,

607

Yeah, the idea is just going to do science.

608

think it's mainly put forward by people who've only worked on a narrow set of scientific

problems.

609

And there's probably some science is going to do well.

610

And then there's some that's going to really struggle with, hopefully, not just the boring

bits.

611

I think we need to map out where those gaps are.

612

Yeah, yeah.

613

mean, it's the same in my field, whether that's sports modeling or just

614

statistical modeling in general, where it's really funny because if you ask an LLM what a

hierarchical model is, it can explain it extremely well.

615

That's exactly what I would explain my students.

616

But then if you ask for a pimsy code of that, the model is not at all hierarchical.

617

It's just a classic model.

618

But the LLM will be like, well, this is a hierarchical model.

619

But it's not.

620

Are you sure?

621

Yeah.

622

So I definitely have that back and forth, but it makes you more efficient.

623

Also, it helps me, for instance, to kill my darlings maybe a bit faster, which is very

important in modeling.

624

That's always better to do it, in my experience, with a human.

625

But sometimes you don't have someone at the same stat level as you in your organization or

project.

626

then you have to do that on your own and that can be quite hard.

627

And there's something we're using actually for things like reviewing training materials.

628

Because sometimes you want to throw a human view, but sometimes you want a first pass of

if an applied field epidemiologist is reading this, is there anything in there that just

629

is going to not make sense or really obvious things you want to fix?

630

It can just help accelerate a lot of that kind of production process.

631

OK, OK.

632

Yeah.

633

That's actually, yeah.

634

That sounds very useful.

635

So you've been already very generous with your time.

636

I'm going to play us out here.

637

But I'm also curious to hear your thoughts about more educational perspective, because you

do a lot, as we've heard, lot of public communication.

638

So given your experience, what

639

educational initiatives would you recommend to better prepare the next generations of

epidemiologists?

640

Yes, but also policymakers and citizens in general.

641

Yeah, I think there's probably some specific angle, then it's to also become more general

ones.

642

think it's not realistic to try and get people to have a deeper understanding of epidemics

any more than it's

643

really sit to have a very, very deep understanding of kind of the nuclear threats or

something.

644

But I think there are certain features of epidemics which are very important and very

often misunderstood.

645

So think one aspect is exponential growth is a concept that people find very, very

difficult.

646

You look at who makes and loses money in finance, and there's a definite divide in terms

of understanding that's a concept.

647

And so I think having more intuition of that across more people is very helpful.

648

Similarly,

649

things like having lagged outcomes, but if you have an event you care about and then the

fact that happens later, that's something that's kind of widely misinterpreted during the

650

pandemic.

651

yeah, think it wasn't increasingly as COVID went on, it wasn't just people who hold

academic posts in epidemiology that were making a lot of useful contributions.

652

You had a lot of people who were in adjacent industries who are maybe actuaries or in

finance or in other bits of, you

653

academic work or even just, you know, math teachers, whatever, quite useful stuff because

they hadn't understand those concepts.

654

You get a feel for some of those data problems that just needed more eyes on them.

655

So I think for me, that's where a lot of value is.

656

It's like, how do we have more people who can just not get very basic things wrong and

just have useful eyes on the problem, even if they don't know the ins and outs of exactly

657

how you've captured that specific parameter?

658

I think more generally, though, we're also just seeing, I think it's a challenge for a lot

of fields that with the emergence of AI, with the emergence of climate, sort of more

659

complex understanding of epidemics.

660

I think we are moving into a world that's much harder to sort teach yourself everything

from scratch.

661

I if you think about even statistics as a field, the sort of statistics that practitioners

do relative to what's taught in schools is now very different.

662

You know, it's the sort of doing kind of regression lines and the things that I did at

A-level.

663

And I think we've seen that in a lot of fields now where actually the cutting edge is so

far detached that I think that can create partly challenges in communication because even

664

if you – a very interesting climate, you can't I can't go and run a climate model in the

way that you might be able to sort of do a simple mathematical proof or a simple kind of

665

statistical problem.

666

So think there's that kind of relationship of how do we build

667

trust and enough understanding of how those fields work that people can engage with them,

even though actually the kind of cutting edge is now so computationally intensive, in some

668

cases just so difficult to explain in terms of the algorithms.

669

Yeah, I think that AI is a prominent example.

670

think there's a lot of those situations where it feels like there's a bit of a growing gap

that we need to bridge.

671

Yeah, yeah, yeah.

672

Yeah, definitely.

673

I mean, completely agree with everything you just said.

674

That's a topic that's dear to me, obviously.

675

We've talked about that several times on the podcast.

676

I think one of the best episodes we've done about that was episode 50 with Sir David

Spiegelhalter.

677

Only sir yet to have been on the podcast.

678

But yeah, that was a great episode.

679

David is an awesome communicator also.

680

I'll put that episode in the show notes for

681

people who want to dig deeper because, yeah, definitely one of the episodes I recommend a

lot.

682

And then the next one, episode 51 with Aubrey Clayton about his book, Bernoulli's Fallacy

and the Crisis of Modern Science, really recommend the book and the episode also because

683

these two together make a great combination if you're interested in epidemiology.

684

Okay, well, I am.

685

That was awesome.

686

Really, thank you so much for taking so much time.

687

I knew that was going to be fascinating and it was.

688

Thanks a lot to Chris Wyman again.

689

But before letting you go, of course, I'm going to ask you the last two questions I ask

every guest at the end of the show.

690

One, if you had unlimited time and resources, which problem?

691

would you try to solve?

692

That's a big one.

693

And I think one of the things coming out of COVID that really struck me is with a lot of

infections, we're actually quite unambitious, I think.

694

You know, we put up with a lot of disease in daily life.

695

And even pre-COVID, if you look at lot of adverts for medicine when you're ill, it's, just

keep going, keep going.

696

And I think we saw a lot of indications

697

during COVID for some of the technologies and approaches that weren't lockdowns but were

actually much more efficient and even just like the work that Oxford have done on things

698

like digital tools.

699

So I think we've got these little signals that we could be much, much better in how we

tackle these problems.

700

And I think that's something that would require quite a lot of resource and time to do

effectively.

701

But even though I've got young children and

702

Yeah, there's a lot of kids that kind of hospitalized with a lot of infections that we

could understand, we could test about.

703

I think we rely a lot on we wait for a vaccine to be developed.

704

But I think I've always just wondered, could we do something a bit cleverer?

705

Can we go all this technology?

706

I mean, there's a US colleague who during COVID said, this is our Apollo mission.

707

Can we actually do something extremely innovative and ambitious?

708

So I think vaccines were amazing.

709

some of the treatments coming down, but we didn't, I think, solve that much earlier

problem of what to do about infection.

710

So yeah, I'd love us to live in a different world where we can actually build on all the

tools and technology where we've got emerging.

711

Yeah, yeah, definitely share that passion and objective.

712

And if you could have dinner with a great scientific mind, dead, alive, or fictional, who

would it be?

713

So I think I've recently for a project been reading up a lot about William Gossett, a

student who developed T-Tests and worked at Guinness.

714

But actually digging more into his work, he was a really interesting character because

there was a lot of conflict with Fisher in terms of outlook.

715

And Fisher was very much from the academic focus of, you accrue knowledge and you want

very high confidence in that knowledge.

716

And that's why you have the sort of thresholds and experiments of science he's been

brought up.

717

But Gossett was much more of a pragmatist.

718

He was working for a big business.

719

And I mean, there's one situation where he had a p-value of 0.13.

720

And he said, you know, that's fairly good evidence.

721

If it's a business decision that doesn't cost much and

722

we can explore further, it's worth going forward.

723

Whereas Fisher would have had that, okay, if it's not hitting the 5%, we're not

interested.

724

so I think for me, that's maybe a part of statistics that got suppressed, I think,

probably lot of the 20th century.

725

I think Fisher and Cobe probably won out in many ways in terms of imposing those criteria

and in some cases throwing away a lot of evidence.

726

I mean, although it wasn't all explicitly baked in, was very anti-Basian, but I think that

making use of limited information, the gossip was very adamant.

727

In some cases, I think he had budget to get two data points or something.

728

And was still that I want to do something with that.

729

So I think that would be a very interesting person to dig into the strategy.

730

It also just sounds like he was just a very nice guy relative to Fisher and some of the

others at the time.

731

So yeah, I think that.

732

the kind of different outlook of where do you set the bar and actually, you know, what are

you trying to do with statistics?

733

Are you trying to kind of get perfect knowledge or are you actually just trying to make

better decisions in whatever job or industry you're in?

734

Yeah, yeah, I love that.

735

So yeah, I think from what I understood, wasn't that hard to be a nicer guy than Fisher,

apparently.

736

But yeah, no, I mean,

737

I definitely resonate with what you were saying because all my models, most of my models,

I do them not for academic use, but for companies.

738

yeah, there is a lot of things you have to be able to prioritize.

739

I think that's one of the most important skills to have as a modeler, especially

740

especially if you're in a small team, right?

741

If you're in a big project with a lot of modelers, you can explore a lot of paths at the

same time.

742

But if you're in a small team, then you have to really be able to set the priorities and

see what path you want to explore first.

743

That's often a metaphor I use to explain what the modeling process is.

744

It's like being lost in a desert and you're trying to find your way out.

745

And the most efficient way is usually to explore a lot of paths and see which ones are

successful.

746

There can be several of them.

747

There can be just one of them.

748

There can be zero sometimes.

749

But also exploring a path that ends up not being successful is actually very informative.

750

because then that means the people behind you won't make the same mistakes and they won't

go down that path.

751

So to explore these paths, you can do it alone.

752

If you don't have any more people, that's just going to take you more time, but you have

to do it or you can do it with several people simultaneously.

753

But that's where the idea of priorities also very important because your priorities are

going to dictate which path you go down first.

754

Yeah, and I think it's that, the priorities and the focus that I think is a kind of angle

which, know, maybe in the pursuit of perfection, we don't always get that balance right,

755

particularly when it's kind of academic research interacting with very fast problems.

756

Right, yeah.

757

Awesome.

758

Well, Edem, let's call you to show that this was an absolute pleasure to...

759

have you here on the show.

760

As usual, I will link to your website and your socials and your books in the show notes

for people who want to dig deeper.

761

Thank you again, Adam, for taking the time and being on this show.

762

Yeah, that's not me.

763

Good job.

764

This has been another episode of Learning Bayesian Statistics.

765

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

766

episodes to help you reach true Bayesian state of mind.

767

That's learnbaystats.com.

768

Our theme music is Good Bayesian by Baba Brinkman, fit MC Lass and Meghiraan.

769

Check out his awesome work at bababrinkman.com.

770

I'm your host.

771

Alex and Dora.

772

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

773

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

LearnBasedDance.

774

Thank you so much for listening and for your support.

775

You're truly a good Bayesian.

776

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

than amazing.

777

Let's adjust those expectations.

778

Let me show you how to be a good Bayesian Change calculations after taking fresh baiting

Those predictions that your brain is making Let's get them on a solid foundation