Artwork for podcast Learning Bayesian Statistics
#127 Saving Sharks... with Python, Causal Inference and Aaron MacNeil
Behavioral & Social Sciences Episode 1275th March 2025 • Learning Bayesian Statistics • Alexandre Andorra
00:00:00 01:04:08

Share Episode

Shownotes

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

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 and Michael Cao.

Takeaways:

  • Sharks play a crucial role in maintaining healthy ocean ecosystems.
  • Bayesian statistics are particularly useful in data-poor environments like ecology.
  • Teaching Bayesian statistics requires a shift in mindset from traditional statistical methods.
  • The shark meat trade is significant and often overlooked.
  • Ray meat trade is as large as shark meat trade, with specific markets dominating.
  • Understanding the ecological roles of species is essential for effective conservation.
  • Causal language is important in ecological research and should be encouraged.
  • Evidence-driven decision-making is crucial in balancing human and ecological needs.
  • Expert opinions are crucial for understanding species composition in landings.
  • Trade dynamics are influenced by import preferences and species availability.
  • Bayesian modeling allows for the incorporation of various data sources and expert knowledge.
  • Field data collection is essential for validating model assumptions.
  • The complexity of trade relationships necessitates a nuanced approach to modeling.
  • Understanding the impact of management interventions on landings is critical.
  • The role of scientists in informing policy is vital for effective conservation efforts.

Chapters:

00:00 Introduction to Marine Biology and Statistics

04:33 The Role of Bayesian Statistics in Marine Research

10:09 Challenges in Teaching Bayesian Statistics

21:58 The Importance of Sharks in Ecosystems

26:35 Understanding Shark Meat Trade and Conservation

32:09 The Trade in Ray and Shark Meat

36:18 Modeling Landings and Trade

42:56 Challenges in Data Integration

44:50 Running Complex Models

51:57 Expert Elicitation and Prior Construction

55:52 Future Directions and Research

56:46 Reflections on Science and Policy

Links from the show:

Transcript

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

Transcripts

Speaker:

Welcome to another live episode of Learning Basics and Statistics, recorded at Pandetta

New York,:

2

:

Today, I'm excited to host a colleague who's become a friend, the great Aaron McNeil, a

professor of marine biology at Dalhousie University and head of the Integrated Fisheries

3

:

Lab.

4

:

Aaron shares the fascinating ways patient statistics are transforming marine research,

particularly in data-borne environments.

5

:

We dive into the complexities of shark and ray conservation, exploring the often

overlooked trait dynamics of their meat and the ecological importance of these species in

6

:

maintaining healthy ocean ecosystems.

7

:

Aaron also discusses the challenges of teaching patient statistics to ecologists

8

:

highlighting the mindset shift required to embrace patient methods.

9

:

We further explore the role of expert elicitation, field data collection, and

evidence-based decision-making in ecological modeling.

10

:

From coral reef ecosystems in the Indo-Pacific to fisheries' livelihoods globally, Aaron's

work underscores the importance of balancing human and ecological needs through rigorous

11

:

data-driven approaches.

12

:

This is Learning Vision Statistics, episode 127, recorded live.

13

:

on November 8, 2024.

14

:

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

projects, and the people who make it possible.

15

:

I'm your host, Alex Andorra.

16

:

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

17

:

like the country.

18

:

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

19

:

Show notes, becoming a corporate sponsor, unlocking Beijing Merch, supporting the show on

Patreon, everything is in there.

20

:

That's LearnBasedStats.com.

21

:

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.

22

:

See you around, folks.

23

:

and best patient wishes to you all.

24

:

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

help bring them to life.

25

:

Check us out at pimc-labs.com.

26

:

Hello my dear patients, I wanted to let you know that I will be at a fun new conference on

,:

27

:

It's gonna be in Manchester, UK and we're gonna talk about Sports Analytics.

28

:

So if you're into that, please join us.

29

:

I am not gonna give a talk this time, much better, our dear Chris Fonsbeck, Pi MC's BDFL,

that you

30

:

heard of very recently in the podcast, well, Chris will be talking about baseball, of

course.

31

:

And there are speakers from top Premier League football clubs, from Olympic winning teams,

from academia.

32

:

It's going to be fun.

33

:

So feel free to get your tickets at fieldofplay.co.uk.

34

:

The link is in the show notes.

35

:

Okay, on to the show.

36

:

Aaron McNeil, welcome to Learning Vision Statistic.

37

:

Thank you.

38

:

Good to be here.

39

:

Yeah, thank you.

40

:

We're live again from Piedetta.

41

:

Yeah, with a very quiet audience.

42

:

I was testing, know, like just I'm not putting you in competition, but at StandCon, people

were like, you know, shouting and so on.

43

:

know, think you might, you might want to, you know.

44

:

They're on the edge of their seat.

45

:

Increase the energy, you know.

46

:

No, so thanks a lot to everybody for being here.

47

:

It's going to be super fun.

48

:

I'm really happy to have you on the show, Aaron, because I've been working with you on the

project we are going to talk about today.

49

:

I've learned a lot.

50

:

There are still a lot of things I don't know.

51

:

So I'm really happy to be able to ask you all these questions and also talk about your

career because you have like a very interesting career.

52

:

think a very inspiring one for people who are, you know, wondering what they should do

with their lives.

53

:

So that's going to be super fun.

54

:

So let's start.

55

:

Yeah.

56

:

Who are you?

57

:

Aaron McNeil, what's your job?

58

:

What are you doing nowadays?

59

:

Yeah, so I'm a professor of marine biology in Dalhousie University.

60

:

That's in Halifax in Nova Scotia, so that's east of Maine.

61

:

And I run something called the Integrated Fisheries Lab, and we're called that because

we're kind of a shop that solves kind of bespoke problems in

62

:

conservation and management typically for fisheries, but we work on other species as well.

63

:

And basically the idea is that, you know, if people have a problem that they need some

sort of inference made on or some sort of a projection or a decision, we're available to

64

:

try to put something sensible together to help them solve problems basically.

65

:

Yeah.

66

:

And that's like everything fish or

67

:

A lot of fish, but other things as well.

68

:

mean, I have a student who's working up in Haida, that's in Northern British Columbia on

the coast.

69

:

And there we're looking at the reintroduction of sea otters, which were decimated by the

fur industry in the 19th century.

70

:

And they're returning.

71

:

And that's actually a problem because the way that laws are written for endangered species

say you can't touch them, but the Haida have been living.

72

:

with those species for 40,000 years.

73

:

And part of the way that they live with them is to harass them or kill them in certain

days so that they can maintain clam beds that that society depends on.

74

:

And so there's a real conflict there between the legislation that comes from the federal

government and the traditional ways that those native people would have regulated their

75

:

ecosystem.

76

:

our job is to go in there and show that actually they can keep doing what they always did,

which is...

77

:

easy.

78

:

They did it for 40,000 years.

79

:

It's probably going to work again.

80

:

But to be able to make that visible to something like a federal government.

81

:

that's, you know, those types of problems, they do make you feel good because you do feel

like you're helping people.

82

:

Yeah.

83

:

No, that's also something I really like about your job is that it, you know, it sounds on

paper, yeah, you just like you write papers.

84

:

And then what happens with that?

85

:

But no, actually, you have a lot of, you know, exchanges with

86

:

the government in Canada and local or in federal, guessing.

87

:

Yeah, local, federal, international.

88

:

There's projects that we have, like the one we'll talk about today is an international

project.

89

:

and, you know, I think one of the great things about doing Bayesian statistics is that,

it's, fits.

90

:

It's a, know, everything looks like a nail, but there it's a really big hammer and you can

hit particularly in data poor environments, which ecology is, it's probably the weakest.

91

:

form of science that you can do, inferentially.

92

:

But Bayesian methods really fit well with a huge range of problems.

93

:

Yeah.

94

:

And I think that's like super inspiring in your work that we can take the results of the

models and enlighten the policy making thanks to that.

95

:

And the feedback loop is not that long.

96

:

That's really great.

97

:

Something I learned a lot during the project and the project was about sharks.

98

:

so like me who didn't know anybody about anything about sharks before I was like, okay, I

don't know my priorities that sharks are endangered, you know, so you don't want to kill

99

:

them.

100

:

It's also because that's the predominant thing you hear on, on lots of most media, you

know?

101

:

So, for us, I'm French, there is this French Island, beautiful in the Indian.

102

:

ocean called La Reunion between Mauritius and Madagascar.

103

:

Yeah.

104

:

If you didn't get the opportunity to go, definitely recommend.

105

:

great for hiking.

106

:

There is an incredible volcano.

107

:

You cannot go to swim though, because there a lot of sharks.

108

:

Bull shark.

109

:

Yeah.

110

:

And bull sharks, which are, they're like the, they, they are not very kind to, to humans.

111

:

correctly.

112

:

Right?

113

:

Yeah.

114

:

Yeah.

115

:

Bull sharks are, they're a really amazing species because they,

116

:

They can live in freshwater.

117

:

If you go to Lake Nicaragua, there are bull sharks in the lake, which is fresh water.

118

:

Bull sharks can regulate the salt balance in their tissues more than other species.

119

:

And the reason they can do that is because they're basically evolutionarily designed to go

in and out of rivers and take things that are fee or drinking and feeding by the river

120

:

mouth.

121

:

they're particularly aggressive because of that kind of life history and then reunion.

122

:

It's insane what's happened in Reunion and one of the great surfing destinations of the

world.

123

:

And it's all been decimated by the Bull shark population.

124

:

Yeah.

125

:

Yeah.

126

:

And, so that kind of work is really interesting because that enlightens, well, sometimes

you do want to touch some species because they are completely invasive and they're

127

:

actually, decimating the rest of the, of the ecosystem.

128

:

that's, and also life.

129

:

as you were talking, we still, your, your, students, projecting in Canada.

130

:

but before that, before we dive into the shark project, as usual, can you tell us how you

came to do Bayesian stance?

131

:

I know the answer to that question.

132

:

like all biologists, marine biologists, I avoided statistics.

133

:

like the plague, in fact, I failed my first year statistics class.

134

:

In the department in which I'm now cross appointed as a professor.

135

:

So I like to say it's because I knew that the frequentist methods were, were BS, but, so I

avoided it like the plague.

136

:

And then, I did a masters in the university of Georgia and there was a PhD student there

called, called Chris Fonsbeck and he was creating a package to do Bayesian analysis.

137

:

That was called PI MC.

138

:

Never heard of it.

139

:

Yeah.

140

:

I, was new to me at the time.

141

:

And, so he started me in on kind of showing me what it was.

142

:

And through a comedy of errors, when I went to do my PhD in Newcastle in England, I made

quantitative methods, the subject of my PhD, because it was intimidating.

143

:

And I knew if I made it the subject of what I was doing, I wouldn't be intimidated by it

at the end.

144

:

And Chris had sort of, convinced me that, this was a thing worth pursuing the Bayesian

methods.

145

:

And so I learned it through.

146

:

Pi MC and I'm probably the longest user of Pi MC in the world as a result of that.

147

:

I've seen the evolution.

148

:

and, it's a pretty interesting because, you know, learning, by doing, think is a really

important way to teach statistics.

149

:

it's generally really badly taught.

150

:

the analogy I use is it's badly taught in the same way that French is badly taught in

Canada, which is you learn a bunch of rules and you never learned to speak.

151

:

Whereas.

152

:

you know, the software that we have now, PyMC in particular, the way that you can kind of

write down various distributions and put it all together.

153

:

Yes, you may not understand likelihood theory or various other things, but provided you're

given a few guardrails, you can start implementing Bayesian models easily.

154

:

And then as I did, I filled in the background knowledge afterwards.

155

:

So I really learned using

156

:

PMC and applying it to problems in fisheries through my PhD and beyond.

157

:

Yeah, that's, yeah, that's amazing because you're like the, used PMC one.

158

:

Zero.

159

:

wasn't even the beta version.

160

:

It was terrible.

161

:

Chris is going to hear that, you know.

162

:

It's okay.

163

:

He knows it was terrible.

164

:

Well, the difference, it's a thing that's hard for people who are just coming to it now to

understand, but the pre

165

:

Hamiltonian sampling with Metropolis.

166

:

Like there was one workshop I had in Zanzibar where I was getting results.

167

:

So I was in this sort of sweltering hotel room and my computer was just melting.

168

:

Like it was so hot.

169

:

couldn't touch it because I had to run a million iterations of some Hamilton, some

Metropolis algorithm, know, and so I'm, you know, forever indebted to Chris for basically

170

:

giving me my career through, through Pi MC.

171

:

But,

172

:

It was a lot harder to use.

173

:

So that way.

174

:

Yeah.

175

:

Like, so as you can hear, also Aaron is doing workshops in very awful places in world,

like Zanzibar, the Galapagos, you know, that's like, that's hot.

176

:

But, and actually I'm curious.

177

:

So HMC for you was like a leap in the ability to use these methods.

178

:

Is there another, you know, advance in PIMC since then that you've

179

:

being blown away by.

180

:

Yeah.

181

:

So it's funny.

182

:

So when I teach Richard McElwain's book as a course now, I we were going to say I teach

Richard McElwain.

183

:

Yeah, everything you know.

184

:

And, you know, I used to tell students they can use Stan or PMC, you know, agnostic.

185

:

And I used to tell them they're basically the same.

186

:

Just use whatever.

187

:

Since the introduction of X-Array and particularly with our project.

188

:

which is basically a fancy label or that has completely changed my life because I'm a

really bad for making careless mistakes.

189

:

And while I was, when I would index things with numbers and the whole rest of it, and now

to be able to just pull out things directly, is insane how good X array is when it's

190

:

paired with, with PMC.

191

:

That said, PMC state space.

192

:

which is just starting, needs to get better because in fisheries, that's like one of the

major places that state space models were developed and are used all the time.

193

:

And so it's a place that we're tinkering with now and we need to get better at.

194

:

That's good to know.

195

:

Yeah.

196

:

got, so state space is a special sub-module of Pines experimental that Jesse Grabowski

developed.

197

:

All on his own.

198

:

Jesse was on the show.

199

:

So I don't know which episode that's going to be, but it's going to be out in a few weeks.

200

:

So if you're interested in that, definitely recommend listening to that.

201

:

I talked a bit about state space models and Gaussian processes with Chris Fonspec.

202

:

We recorded two days ago, so it's going to be out before your episode.

203

:

Definitely recommend listening to that.

204

:

When would you use a Gaussian processes versus state space models?

205

:

But yeah, I see you up to date with Pimc News because StaySpace is pretty new.

206

:

And that's really amazing.

207

:

I've been also playing a bit with that at the Marlins.

208

:

I have some stuff I want to contribute, hopefully that will help you.

209

:

Especially vectorizing the time series because right now you cannot do that.

210

:

If you wanted to do that for...

211

:

100,

212

:

Really want to learn something that seems tangential to that because X-rays basically

multidimensional pandas.

213

:

But that's one of the best investment you can make.

214

:

Hands down.

215

:

afterwards, because actually most of your time in the model development is not really

writing the model.

216

:

It's diving into the model before when you have to do prior predictive sampling.

217

:

And then afterwards, when you have to confirm that the model is not nonsense or

218

:

more precisely to take aware of the model in nonsense.

219

:

And for that, X-rays are absolutely amazing.

220

:

Yeah, and we, in my lab in particular, but in science in general, make, or in ecology in

general, we make pretty bespoke figures using the posteriors and the ability to go in and

221

:

pull a species out or to put extra labels on groups of species and then pull out the

posterior for the group is just, it's incredible.

222

:

Yeah, because to make that clear to people,

223

:

Like basically what you get, for instance, let's say you have a parameter in your model

that's like the, call that the shark effect.

224

:

That's a cool name.

225

:

And so you have 10,000 sharks, let's say.

226

:

And you have that through time.

227

:

Let's say you have 10 years of data on that one.

228

:

So that's already without anything else in the model, that's four chains, 1000 draws per

chain,:

229

:

So that's four dimensional.

230

:

vector before when you were using NumPy, you had to isolate and then call it call it.

231

:

And then you were indexing with a number.

232

:

Whereas right now you can just sell dot cell.

233

:

You say dot cell shark equals dog shark.

234

:

And then boom, you have what you care about.

235

:

that's like, that's amazing.

236

:

But I know that's a bit frustrating because that's not sexy at all.

237

:

Yeah.

238

:

But the results are.

239

:

It's totally worth it.

240

:

Yeah, for sure.

241

:

Yeah.

242

:

Yeah.

243

:

Custom plots.

244

:

That's, that's the thing to do.

245

:

And actually since we're talking a bit about teaching before diving into the shock stuff,

I'm curious if there is a common mistake or hurdle that you see your students having when

246

:

you teach them patient stance that if they knew about that, that would

247

:

accelerate, jumpstart their learning?

248

:

Yeah, that's a good question.

249

:

mean, McElwraith's book is called Statistical Rethinking for a reason, and it's basically

requiring people to somewhat discard the rules that they were taught in their

250

:

undergraduate course.

251

:

I think for people to come in with a fresh mind, I think is the biggest thing.

252

:

The hurdle that I find, even sort of halfway through the course, you still find people

253

:

you know, benchmarking the results against stuff that they know.

254

:

And that's natural to do that.

255

:

But I think the mentality about explaining variances and the obsession with that as if

that's the only thing that's worth doing.

256

:

it pervades ecology.

257

:

don't know about other disciplines, but that's a really big problem.

258

:

so, getting people to discard that sort of thinking and then move into a sort of causal

mentality, is a big deal.

259

:

know, that causal words are kind of.

260

:

a taboo in ecology.

261

:

you know, when in reality it's kind of what everybody's doing.

262

:

So it's this sort of Victorian behind the scenes, there's people that are doing something

and they're something out front that's different.

263

:

and, so one of my biggest jobs teaching people is to give people permission to use causal

language and Bayesian models really help with that because, you know, we're able to say,

264

:

well, look, this, you know, priors are about your belief and, and what you think.

265

:

And then we're going to confront that with some data.

266

:

We're going to see what the conclusions are and just changing the mentality is a big deal.

267

:

So I think if somebody's starting out in a Bayesian context, getting them to think in

really general terms about what they're trying to do when they're running models and

268

:

trying to, particularly in science and ecology, where we're often doing inference, it's

okay to be doing that, you know, to give themselves permission to do that and to kind of

269

:

discard the

270

:

the sort of rules that are prescribed when you do undergraduate statistics.

271

:

think it's a big deal.

272

:

Yeah, it makes sense.

273

:

So because like your students are like, are coming from undergrad, so they already have a

bit of knowledge about stats, right?

274

:

Yeah, in a way that's the worst you can get, right?

275

:

Because you'd prefer students who don't know anything, that way it's just a blank page,

right?

276

:

Yeah, a little bit.

277

:

Yeah.

278

:

And I mean, what's shocking about undergraduate statistics is that people don't, they'll

come through that discipline and have no idea what a likelihood is.

279

:

you think, well, it's not actually that complicated to know what a likelihood is.

280

:

You know, it's just numbers.

281

:

You substitute them into an equation, you get a number at the end.

282

:

Like that should be the foundation of what they're taught.

283

:

And yet it's all about, you know, working through these sort of golems that McElroy talks

about working through these rules and

284

:

And that's where they get reassurance is in the rules.

285

:

They don't get it in the knowledge of the system itself.

286

:

And that's, it's a mentality problem.

287

:

Yeah.

288

:

No, they completely agree.

289

:

so sharks, Aaron, like, why?

290

:

No, that's probably quite the question.

291

:

I get that question quite a lot when I talk about that project.

292

:

Yeah.

293

:

It's like, why do you care about sharks?

294

:

Yeah.

295

:

I mean, so I started in.

296

:

looking at them because, they were cool and I was an undergrad and I wanted to be cool.

297

:

So I started working on sharks.

298

:

Yeah.

299

:

and my career, my sort of reputation is actually made it primarily in coral reef

fisheries, which is much more sort of important for people in their, in their lives.

300

:

but the sharks came back to me later on through some, colleagues and collaborators.

301

:

Uh, who are doing conservation work.

302

:

And the reason that it's important is because, uh, you know, on average sharks in the open

ocean declined around 90 to 95 % in the last half of the 20th century.

303

:

And that was due to overfishing from longlining.

304

:

So going for things like tunas and billfish and things.

305

:

And,

306

:

And you know 90 % declines in any species is a big deal and the question that I motivated

me in in shark research is does it matter?

307

:

You know if they're just up there at the top of the food chain just picking the cream off

the top then maybe it doesn't matter We know from sort of lions on the Serengeti that

308

:

that's probably not true You know lions if they're not there they fail to control the

antelope populations the grass ecosystem collapses so predators in ecosystems have these

309

:

functional roles where they actually are a key to how the whole ecosystem works.

310

:

So if we want healthy oceans and healthy oceans are the ones that are going to be most

productive for human society, then we want to have them functioning really well.

311

:

so sharks are an important component of that.

312

:

And they're a fascinating, beautiful 65 million year old species.

313

:

they're pretty cool to work on.

314

:

Their evolution is just...

315

:

Yeah.

316

:

Incredible.

317

:

all the time when we have the workshops, like I learned new facts about shark evolution

from you and Chris and the whole team is just like, these animals are incredible.

318

:

Like, do I remember correctly that they don't have any skeleton?

319

:

Yeah.

320

:

So they don't have bones.

321

:

have cartilage and cartilage as we, you may or may not know, but it has a lot of spring to

it.

322

:

It's a way for them to be very efficient in the ocean in terms of how they're moving

around.

323

:

Yeah.

324

:

But yeah, so what I really like also in that, how you motivate, like how you were

motivating natural, so in the whole group you have working on that is that you guys are

325

:

very evidence driven.

326

:

So that's recall because I don't want people listening to that thinking that, yeah,

basically like you're here to...

327

:

protect sharks no matter what.

328

:

And even if the evidence tells you that, here that's fine.

329

:

You know, it's like, it's normal.

330

:

You want kind of a Goldilocks zone, as you were saying, not too many top predators, but

also not too little.

331

:

So that I learned from interacting with you.

332

:

And I think that's very important because in these fields, you kind of have also these

extremist view of it's like, no, you cannot touch anything.

333

:

And actually that's not always the optimal solution to do.

334

:

Well, and I think if you ran a hotel in Reunion, that's a really good case.

335

:

Where was it Brigitte Bardot that came there and just said, you can't touch sharks and

their majesty and stuff.

336

:

And, know, bull sharks are not threatened anywhere in the world.

337

:

The fact that people can't recreate in the ocean has completely decimated the tourism

industry in Reunion.

338

:

I think it's rational to call those sharks

339

:

for a good that has to do with the population that lives there.

340

:

And that may be a popular opinion among shark conservationists.

341

:

But yeah, we try to walk the line as much as possible of rational decision-making and

trying to understand that, you know, other people have different values than we do.

342

:

And how can we make decisions that are going to see the most benefit for people and for

ecosystems?

343

:

Yeah.

344

:

And that's also why I think it's very interesting because your Bayesian models really

345

:

feed into that, how to make the most optimal decision based on what the evidence say.

346

:

So yeah, because that being said, there are a lot of species of sharks and some of them

are really endangered, but some of them are not.

347

:

and I think that's where the project we worked on, you know, like is interesting to talk

about because yeah, like, can you explain this project?

348

:

Like give us the elevator pitch, what...

349

:

we're doing in the paper and what this is all about.

350

:

Yeah.

351

:

So at an international level, way that shark conservation operates is often through

agreements that are made among countries, particularly around trade.

352

:

So you have CITES, which is the Convention on International Trade in Endangered Species.

353

:

And that regulates, for example, you can't take ivory across a border.

354

:

so CITES is the instrument that regulates that trade.

355

:

so, with this project is about looking at, meat that gets traded for sharks and rays

around the world.

356

:

And right now that trade happens just as meat.

357

:

There's no species labels on it at all.

358

:

And we know just because we know what gets landed that a huge proportion of that trade is

actually in species that are listed under CITES and are not reported in the way that

359

:

they're legally required to be.

360

:

So it's an illegal trade in.

361

:

in shark meat.

362

:

And so the question that we had or we were presented with is, can we figure out a way to

decompose that trade into species?

363

:

And, you know, it kind of was brought to me as a problem, and I'm not somebody that knew

anything about compositional data.

364

:

And so I thought, but I thought, well, it's a decent amount of work to do this, but it's

an interesting problem.

365

:

And so we took it on and it was a

366

:

It was stressful because it was beyond my ability, just beyond my ability to do so.

367

:

But as a result, I've learned a lot about how to decompose things and how to build up

compositions in a sensible way.

368

:

it's worked out.

369

:

Yeah, that was a very hard project.

370

:

I know you've lost a bit of sleep from time to time because the model wasn't running.

371

:

Wasn't yeah, I wasn't running.

372

:

mean, there's there there are just so many hidden dimensions to this because you know when

when things get reported to international authorities, first of all, you have landings of

373

:

species and those get some reported to species, but they also get reported at these kind

of aggregated taxonomic levels.

374

:

So you may get reported at the genus level or the family level or the class or whatever.

375

:

And so how do you sensibly decompose?

376

:

those aggregations into the constituent species because it's going to vary depending on

what the aggregated category is.

377

:

Some species could be in that category and a bunch of species are not in that category.

378

:

There's a taxonomic hierarchy that we had to take into account.

379

:

Of course, the whole thing is probabilistic.

380

:

It's all about, well, probably it's this or that.

381

:

We'll dive into the nitty-gritty in the model in a minute.

382

:

depending on where you're from in the world, you could be completely surprised about the

fact that there is a trade of shark meat because that means people eat shark.

383

:

Yeah.

384

:

Which is like, least to me as a French educated person was really surprising.

385

:

so yeah, can you

386

:

Well, because you're not from the coast.

387

:

You're not from Normandy.

388

:

Yeah, I know.

389

:

So France is one of the largest consumer of ray meat in the world.

390

:

Yeah, ray that.

391

:

Yeah, yeah, yeah.

392

:

And so, you know, and I guess one of the most important species in this whole thing is

spiny dogfish.

393

:

spiny dogfish is fished in the North Atlantic and fished in New Zealand as well.

394

:

It occurs all over the place.

395

:

And it's the number one species for fish and chips.

396

:

So if anybody's been to Europe, Belgium, the UK, even Australia to some extent, if you're

eating fish and chips, you're mostly eating shark.

397

:

And that's the number one species or group that is in the fish and chip trade.

398

:

So there are really meat focused species that we do know something about and we can

incorporate that sort of information into the models.

399

:

Yeah.

400

:

But that was like really mind blowing to me to learn that beast.

401

:

Lots of shark is actually eaten.

402

:

So trade of shark meat is not negligible.

403

:

But often, at least depending on where you are in the world, it's not marketed as shark

meat because it's not right.

404

:

It's clearly not attractive.

405

:

Yeah.

406

:

So most of the time you don't know fish and chips.

407

:

The fish and fish and chips is actually shark and chips.

408

:

Yeah.

409

:

Well, a cool name.

410

:

That's right.

411

:

Well, and I think I think somewhat because I had the same.

412

:

reaction somewhat that our own ignorance culturally.

413

:

mean, and so completely.

414

:

Yeah.

415

:

Yeah.

416

:

And so, you know, in, in South American Brazil, you know, there's very specific dishes

there.

417

:

They're clearly identified as shark in India.

418

:

there's a certain size range of coastal sharks that are very important culturally and

something that say pregnant women are culturally eating.

419

:

So I think, you know, one of the things about this project is a lot of work.

420

:

in the past 20 years has focused on shark fins and shark fin soup and how that's driven

over exploitation, but to the detriment of people thinking about meat and the fact that

421

:

actually the original reason people were catching these animals was for meat.

422

:

so let's take a look at that.

423

:

And it's worth, you know, a billion dollars more than the fin trade industry.

424

:

it's meat is a big deal.

425

:

Yeah.

426

:

Yeah.

427

:

I for sure.

428

:

And so like, do you already had some, some results from the model net, how the trade

429

:

between countries and across species is done to share with us today?

430

:

Yeah, for sure.

431

:

I mean, I think the biggest thing to note is that the trade in ray meat is just as large

as the trade in shark meat.

432

:

It's actually a little bit larger.

433

:

And that's surprising to people, but it's much more concentrated.

434

:

So, for example, there is a species of skate, which is a kind of ray that is in Argentina.

435

:

Enormous volumes of this are being traded to South Korea, which is the largest importer of

ray meat in the world.

436

:

So that ray meat is going to very few markets, whereas the shark meat is a much more

global market and is much more diverse in terms of the species that are involved.

437

:

I think, you know, another surprising aspect is, so about 40 % of the meat trade is blue

shark.

438

:

Do you remember the species name?

439

:

Is that pre-United States Glaciers?

440

:

Well done.

441

:

Well done.

442

:

He's got it.

443

:

But the rest of the trade is composed of some 90 to 100 other species.

444

:

And so it's a really diverse trade that's happening and it rotates depending on how things

get exploited or over exploited.

445

:

So we're seeing, you know, a lot of species will be fished.

446

:

they'll get over exploited and then fishers will move on to other species.

447

:

so there's a lot of exchangeability of species within the market that is really important.

448

:

Yeah.

449

:

so how did you end up having these insights, right?

450

:

Like what does the model look like?

451

:

Were you able to feed the model on sharks and rays at the same time?

452

:

Do you have any hierarchy on that?

453

:

What does the model look like?

454

:

Yeah, all of that.

455

:

So sharks and rays are

456

:

as a group called the Lasma branks.

457

:

the, the, the cartilaginous fishes, and they'll sometimes get traded as a Lasma branks.

458

:

So we had to have a model that could jointly estimate both groups.

459

:

And then the other thing I think, which is particularly basing here is that, we have, you

know, in order to figure out what's in the trade, we first had to figure out what was in

460

:

the landings and landings are reported sometimes to the species level and sometimes to

these aggregated levels.

461

:

And in order to figure that out,

462

:

in any given country, we had about 65 countries that we were looking at and we don't know

everything in every country.

463

:

And so the first thing we had to do is to get a sense from experts about what the

composition of the species are in each country in terms of landings.

464

:

so that's an elicited prior.

465

:

And this estimation that we're doing, you couldn't do without those priors because in

places where, say for China, for example,

466

:

doesn't report anything to anybody except at an aggregated level, it's impossible to know

what the species composition is without some expert opinion about what the composition

467

:

would be in the landings.

468

:

And so really the anchor of this whole project was to be able to figure out what the

priors are in the landings and to have something sensible to say there and then propagate

469

:

that information through the model.

470

:

And the great thing about this sort of hierarchical decomposition is that

471

:

Sometimes you do actually know what the species are, and so your priors are overwhelmed in

those cases.

472

:

But in places where you don't know, this is what you have.

473

:

And I think when the results come out, a country like China could say, you have it totally

wrong in terms of this composition.

474

:

And we could say, great, then start reporting the composition of the species in your

landings and we'll correct this thing.

475

:

We're not in the work we're doing, we're not finger pointing and saying you're doing this

or you're doing that.

476

:

We're saying, what is the current state of knowledge of this system for us?

477

:

And Bayesian modeling really is the only way that we would be able to do that.

478

:

Yeah.

479

:

Yeah.

480

:

That's, that's what I find fascinating in this model is that you have so many pieces

together.

481

:

Like first, yeah, as you were saying, the priors are very important here because there are

some combinations of countries and species that you don't have.

482

:

data on, and also you have two models in one, right?

483

:

Like you have the model for the landings and then you have the model for the trade and the

results of the model for landings are fed into the other model that tries to infer the

484

:

trade between countries.

485

:

how does that work?

486

:

Yeah.

487

:

How does that work?

488

:

How do you do this?

489

:

So basically what the way it's set up is that we have a kind of latent state, which is

490

:

you know, a species by country matrix for the landings.

491

:

And so, that is the thing we're really trying to estimate.

492

:

And, you know, the certain cells of that matrix are hard zeros.

493

:

We know that they're not, that it's an impossible combination of things.

494

:

And so we developed a technique that is throughout the model of masking where we are using

either one zero or negative number masking, depending on the context to actually wink out.

495

:

certain nodes within that latent matrix so that none of the catch can be allocated to that

cell and it gets allocated only to the cells that have some prior information involved.

496

:

And so through the landings model where we estimate this latent landings thing, we then

have the problem of, we know that there's this much of these species in this country.

497

:

How do you then trade that out into the world?

498

:

Because as a basic

499

:

Assumption you could say well equally, know, whatever the distribution is by species in a

country in its landings It just equally trades that out into the world, but that's not

500

:

particularly satisfying it doesn't seem right probably the only clever insight I had in

this whole project was I Woke up in a yurt in Cape Breton in about 2 30 in the morning one

501

:

winter.

502

:

It's right on yeah

503

:

With it with it was a thing that was bothering me obviously and I woke up in the middle of

the night.

504

:

This is about two years into the project and we were on a ski trip with my kids and I

realized that in that my subconscious realized that the composition of species that any

505

:

given importer wants to trade is going to imply some sort of correlation in the exporters

that they trade with.

506

:

So if I'm France and I want

507

:

Raja clavata is a really common European species.

508

:

I'm going to have larger volumes of trade, which we know, with countries that have that

species.

509

:

So there's an implied correlation there for any given importer in terms of who they trade

with and the amount of volume that those exporters will all have correlations in terms of

510

:

their species that they care about.

511

:

So this is a variable that we call import preferences.

512

:

And it works.

513

:

So countries that we know have a lot of fish and chips, European countries are trading

more with countries that have a lot of dogfish.

514

:

Right.

515

:

And, and, you know, with the rays and so forth.

516

:

and then we have other parts of the trade model that deal with, you know, trade volumes of

seafood that, you know, countries are more likely to trade with each other if they have

517

:

high seafood volumes and so forth.

518

:

But basically we tried to come up with some covariates for trade as well as

519

:

potentially more expert information that would help us take the landings and send them out

into countries in a sensible way.

520

:

Yeah, that's fascinating.

521

:

And how do you handle that correlation, that matrix of correlation, you know, like where?

522

:

you estimate that correlation at inference time or you have hard-coded the correlation

between the trade preference of

523

:

to given countries.

524

:

So given how big this thing is already, I didn't go a kind of multivariate normal way.

525

:

We might do later on.

526

:

But right now what we have is for any given species, for an importer, for all of the

various species, we simply estimate an effect size so that if that country, if that

527

:

species is available with countries that they trade with, you're going to see a larger

trade volume with that effect, basically.

528

:

Okay.

529

:

Yeah.

530

:

Yeah, so it's like you have a slope per combination, per pair.

531

:

Right.

532

:

Okay.

533

:

Yeah.

534

:

So that's a lot of parameters.

535

:

Brutal.

536

:

Because also this is not symmetric, right?

537

:

So it would have one parameter for France and Spain and another for Spain and France.

538

:

Well, for the species, the desirability of any given species, it's really species by

importer.

539

:

So France cares about...

540

:

A particular array of species and Spain cares about a different array of species,

basically.

541

:

okay.

542

:

Yeah.

543

:

Yeah.

544

:

Yeah.

545

:

That's, that's super.

546

:

That's amazing.

547

:

Yeah.

548

:

Yeah.

549

:

And without X array keeping track of all this would be impossible.

550

:

Yeah.

551

:

Yeah.

552

:

Yeah.

553

:

Yeah.

554

:

I remember.

555

:

But no, I mean, that's amazing.

556

:

Well done.

557

:

And like getting all of these together.

558

:

That's just incredible.

559

:

So you've talked about.

560

:

about one of these things already, but I'm curious, given the side of the model and the

project, what is the hardest obstacle you had to overcome?

561

:

What was the biggest thing that kept you up at night?

562

:

Well, delivering the project.

563

:

Definitely.

564

:

It's funny, at some point during this process, I may have said even to you at one of our

workshops,

565

:

I feel like we're just doing sums, right?

566

:

This is just fancy adding and subtracting.

567

:

And you kind of said, yeah, but it's pretty fancy.

568

:

And so, you know, I think making sure that we get things added up correctly in a project

that has so many dimensions and those dimensions are not symmetric.

569

:

So the combination of species in a country.

570

:

is just varies by every single country and making just making all the addition of those of

those distributions add up correctly.

571

:

Yeah, it was just mind blowing and and one of the problems that we had is when you're

doing compositional data, you know, some countries would aggregate, they'd have one

572

:

aggregated category or they would have two aggregated categories.

573

:

Most countries would have three or more aggregated categories.

574

:

So depending on which country you fall into, that implies either a soft max or some sort

of a diraclée breakdown with some multinomial, or it implies a binomial, or it implies

575

:

just a one zero flag of possible.

576

:

And so doing that automatically and having it all sum up correctly, you you just have

these, it's just the detail, which is just extraordinary.

577

:

And it's crazy that it can take years to make that detail.

578

:

Correct.

579

:

But you'd see the results and you'd realize, no, we have a leakage somewhere in the system

and you'd have to go back and really think about those details.

580

:

So it was a lot of really just buttoning up all the leakages so that the distribution of

species could flow through the system.

581

:

Yeah.

582

:

That sounds a lot like what I'm doing at the model with baseball players.

583

:

And what's amazing also I found is that you have several likelihoods in the model.

584

:

Yeah.

585

:

And that's a question I often have from people like, how do I do if I have several

likelihoods, you know, just send them to the model?

586

:

Well, it's amazing.

587

:

It's a thing I haven't even mentioned is we have a whole field program where we went to

countries that had the most landings in the world and we sampled their landings and

588

:

actually had field data that we incorporated.

589

:

And that field data, as you're saying, all it is, is it feeds back into this latent

landings.

590

:

matrix that we're estimating and it's just more information.

591

:

so the ability to take gain for Bayesian analysis, the ability to take expert information,

field data, the structure of how trade happens, all of it gets integrated together with

592

:

however many likelihoods you need.

593

:

It's unbelievable.

594

:

Couldn't do it any other way.

595

:

And so before opening to some questions from the audience, how is the model running?

596

:

How do you run such a model?

597

:

What's the running time?

598

:

Did you have difficulties running such Yeah, we have.

599

:

And it's a thing where we had to go in and figure out how to send all the samples to a

backend because basically the computers that we have were sort of melting.

600

:

I don't have it on some gigantic machine.

601

:

I think we will.

602

:

That'll be a next step with our computer science department is to...

603

:

is to get it happening altogether.

604

:

yeah, basically we had to chunk, we had to take the landings model and run that and then

take the posteriors for the latent landings and apply it to the trade model separately.

605

:

And that worked pretty well and we can run a few hundred, 500 samples or something with a

decent tuning period in a day or something like that.

606

:

it's not too bad.

607

:

things blow up quickly.

608

:

How did you handle that?

609

:

know, like if it takes one day to get an iteration of the model, how do you practically do

model development, you know, without...

610

:

Yeah, it's a good, mean, you're chunking things up is a big thing.

611

:

And then also, you know, again, Hamiltonians are so insane that, you know, I could run 50

samples and I'm not getting that much of a posterior difference to when I run a thousand.

612

:

Yeah.

613

:

It's absolutely incredible.

614

:

It just goes straight to the peak or the valley.

615

:

mean, tuning is good.

616

:

Yeah, exactly.

617

:

it went well.

618

:

Tuning goes well.

619

:

Yeah.

620

:

Yeah.

621

:

So that's been a big advantage.

622

:

Yeah.

623

:

So yeah, running small and then going to sleep with it running in the background,

basically.

624

:

Yes.

625

:

Yeah.

626

:

And so I think you might want to try running that on a GPU, Basically, you a huge matrix.

627

:

Yeah.

628

:

So yeah, you could try.

629

:

sparse matrix, you would have to, I mean, I need to see the latest code, but then you

would probably have to modify the code so that you could try, you could use PyTensors

630

:

sparse operations and then run that.

631

:

But even without that, you could try that on a GPU already.

632

:

with PyMC now it's super simple ways.

633

:

could either specify PM dot sample.

634

:

NetSampler equals NumPyro.

635

:

And if you're on a GPU, it's going to detect it automatically and use that.

636

:

And that should be way faster.

637

:

Otherwise you can also use NetPy.

638

:

So I don't know if you followed that development.

639

:

So NetPy is the re-implementation of Hamiltonian Monte Carlo in REST.

640

:

Adrian is able to need that as a hobby.

641

:

So Adrian is the one who came up with the zero-sum novel.

642

:

We use a lot in the project.

643

:

Yeah.

644

:

So he did that and that's like, that's way faster.

645

:

Also he's got some new fancy tuning adaptation algorithm.

646

:

Nice.

647

:

So tuning is faster and better.

648

:

So you need only like 300 tuning, tuning draws.

649

:

Whereas in the, in the default Pinty semper it's 1000.

650

:

So that's way faster.

651

:

And you can use NutFi with the Jax backend.

652

:

And so if you're on a GPU, that's going to send it over.

653

:

Yeah.

654

:

Yeah.

655

:

So if you can get access to GPU at the HUSI, should be efficient and worth your time.

656

:

Nice.

657

:

Any questions already from the audience?

658

:

Yeah.

659

:

Sir.

660

:

What's your name?

661

:

So, beautiful name.

662

:

like that.

663

:

So, Dante is asking about the different likelihoods you're having the model.

664

:

basically, I understand correctly, how...

665

:

how you could data basically how that looks like is one feeding into the other.

666

:

Yeah, there's so they're jointly because our the inference we're trying to make is kind of

a joint inference like we care about combining the knowledge together with the data to get

667

:

our best estimate of what this latent matrix is.

668

:

And so that is the composition of things is really the key part to this whole thing.

669

:

And

670

:

that composition applies to both parts of the likelihoods.

671

:

so essentially we're using the same latent part of the model to indicate the species

composition that could be in the trade data and in our field data.

672

:

And so it's simply a question of coding it up that way where the parent distributions in

the hierarchy are going to be the same parent distributions basically.

673

:

So yeah, follow up question from Denti.

674

:

Basically the observations, do you observe the landings?

675

:

Do you observe the trade?

676

:

How does that work?

677

:

Yeah.

678

:

the trade is just meat.

679

:

So we have the volume of trade, but we don't have anything about the composition.

680

:

So the compositions are informed by in three main ways.

681

:

One is the priors from experts.

682

:

The second is in field data that we went out and surveyed.

683

:

And the third one is in reported data.

684

:

The Food and Agricultural Organization of the United Nations has a fish group and every

country that lands fish reports to them every year what they've landed.

685

:

That reporting can happen at different levels.

686

:

Sometimes it's at the species level, you know, the United States, Canada, we report to

these things.

687

:

Vietnam and China report nothing.

688

:

And so, in every country in between.

689

:

And so that data...

690

:

Some of it's at the species level, some at the genus level, some at the class level, like

it goes up and up and up.

691

:

And the structure of that reporting implies what species the composition could be.

692

:

So there's volumes of elopidae, which is the thresher sharks, there's only three species.

693

:

So the model, what it has to do is based on our field data and expert opinion, it has to

allocate that tonnage of reported elopidae to those three possible species.

694

:

depending on the priors and on the data itself.

695

:

it's all, it's kind of all together, but it happens in the landing side of the model

because the trade data doesn't have any species composition in it.

696

:

Yeah, that's right.

697

:

So yeah, the results are, yeah, deal with those discrepancies, deal with the hierarchy.

698

:

They also, which I didn't talk about, we have from, there's of course, French economists.

699

:

who give reliability data for things that are reported to international organizations.

700

:

And so we have a score that they've created for each country, each exporter.

701

:

And we actually model the uncertainty in their reporting based upon this score.

702

:

countries that are more reliable would have more certainty about what they're reporting.

703

:

Yeah, that's true.

704

:

That's a cool part of the model.

705

:

We're modeling the standard deviation.

706

:

You know, like your observational noise, for instance, in a normal likelihood, we're not

only modeling the mu, we're modeling the sigma and that sigma is a regression.

707

:

It's like the sigma depends on the reliability score of the countries.

708

:

that's really good.

709

:

And that's like another really cool stuff of Bayesian model where often you have people

like, how do you deal with your heuristicity?

710

:

So it's like if your sigma depends on features of the data, you just model it.

711

:

That's a really cool thing.

712

:

Great.

713

:

Happy Dente?

714

:

Yeah.

715

:

Other question?

716

:

Yeah, Jeff, of course.

717

:

Yeah.

718

:

So Jeff asks, how do you do the prior elicitation Aaron with so many parameters?

719

:

Right to the heart of the matter.

720

:

What we're eliciting is, so you have fisheries experts in every country who have some

sense of what the catch composition is for these species.

721

:

There aren't that many shark experts in the world and

722

:

So I sit on something called the shark specialist group of of, excuse me, of, what is it

called for the UN and the IUCN, the international union, conservation union of the UN.

723

:

And so we have experts in each country that know what the composition is.

724

:

And so we're asking them a pretty simple question, which is how much of these things do

you get?

725

:

Do you get never see it rarely see it?

726

:

You know, it's on a sort of Likert scale.

727

:

And then we've translated that into a log odds basically.

728

:

And that's what goes into the prior.

729

:

Now the real problem with that process is of course, how certain are they about their

scores?

730

:

And we don't have that in there yet.

731

:

It's actually a hard problem.

732

:

I haven't found a really good example of people doing this kind of elicitation.

733

:

I've seen some stuff in the computer science literature, related to like market surveys

with products, but

734

:

In expert elicitation, need, the mean, but then you need the uncertainty about that means

and you need to construct a good prior in that way.

735

:

And we have not done that.

736

:

We just have these log odds and we have basically a standard deviation of three that goes

into a multinomial.

737

:

And there's been enough trial and error over three or four years that we know that we're,

you know, we've tried bigger or smaller and we've kind of seen what the consequences are

738

:

for the way the model is.

739

:

so we've

740

:

Just by iteratively going through, we've kind of tuned that uncertainty, but it's one of

the things that we need to work on for sure.

741

:

Any other questions?

742

:

Okay, great.

743

:

still have some.

744

:

yeah, one I have Aaron's that so well, we'll put in the show notes any link we can at the

time we have the episode out.

745

:

So hopefully the pre-print very probably also the GitHub repo because all the code is

going to be available in open source.

746

:

So that's the great thing with a huge Pimacy model for you guys to check out.

747

:

Are you going to be able to share the data too?

748

:

Yeah.

749

:

It's all open data.

750

:

Yeah.

751

:

Yeah.

752

:

Awesome.

753

:

Yeah.

754

:

Yeah.

755

:

What a great project.

756

:

Yeah.

757

:

So yeah, like natural question, what's the next step in that project for you?

758

:

Yeah.

759

:

So we have a project that our postdoc Chris Mull has got put together to look at the ways

in which different countries are exporting their ecological footprint related to sharks.

760

:

So that's upcoming in January.

761

:

And then over the next sort of four or five years, our long-term goal is to look at

762

:

how various management interventions have changed the landings composition.

763

:

So in other words, thinking causally, know, have things like fin retention bands or just

retention bands in general actually had an effect on the landings.

764

:

does international conservation work is kind of the question we have.

765

:

And that's a really big one and it's going to have implications for any wildlife trade.

766

:

So that's super exciting.

767

:

Yeah.

768

:

It's going to be a fun one.

769

:

Probably.

770

:

interrupted time series, analysis, stuff like that.

771

:

Cause the inference coming your way.

772

:

Nice.

773

:

And other question is you're a very curious guy.

774

:

So I always love asking you that question, but yeah, what are you learning these days?

775

:

know, what are you curious about?

776

:

What are you, you know, reading about things like that?

777

:

I guess.

778

:

Yeah, what am I reading right now?

779

:

Well, the book I'm reading right now on the plane here is about, it's called The Honest

Broker, and it's about the role of scientists in informing policy.

780

:

And so particularly at universities, there is often a real appetite to be an advocate, to

really advocate hard for conservation or hard for something.

781

:

And having spent 10 years with the Australian government,

782

:

I kind of see the value of not doing that and, trying to, you know, provide people,

particularly politicians, people who are making decisions with an understanding of the

783

:

field and what the options are and what the consequences of various decisions are.

784

:

And that to me is a very important area that science can inform, policy.

785

:

so the, book, the honest broker is really about that.

786

:

And so I think as I'm sort of becoming more senior in my job.

787

:

I want to make sure I'm doing a good job advising people who are coming to us for advice.

788

:

And that's a really big one.

789

:

And I guess the other thing is I just read a book on Dirac's life.

790

:

yeah?

791

:

totally fascinating.

792

:

it's very hard.

793

:

You know, I give a talk about the conservation consequences of the atomic bomb.

794

:

And one of the dimensions of that is of course the methods that we use, right?

795

:

A lot of stuff in MCMC of course came out of the Manhattan Project.

796

:

But also if you go to Bikini Atoll, right, where the French did their nuclear tests, it's

one of the most beautiful coral reefs in the world.

797

:

And it's because after those tests finished, nobody went there.

798

:

And it's a bit like Chernobyl is the same thing.

799

:

If you go to Chernobyl, it's just a wildlife paradise.

800

:

And yet it's on the back of this kind of nuclear disaster.

801

:

So there's all these weird kind of conservation benefits of war.

802

:

So just thinking about all the things that happened, you know, what a fertile time in

science, even though it's horrific what was created.

803

:

It's incredible also at the same time.

804

:

And that's also a good reminder that you need cultural thinking in your model, right?

805

:

Because otherwise just an AI model will tell you, well, actually to preserve wildlife,

need to bomb them up.

806

:

So that's, that's, that's good to think positively, right?

807

:

Awesome.

808

:

to close up the show, Aaron.

809

:

As usual, the last two questions, ask every guest.

810

:

So first one, if you had unlimited time and resources, which problem will you try to

solve?

811

:

I mean, it's impossible not to say climate change.

812

:

mean, so, and I say that everybody's kind of says this, but I do it because a lot of my

work has been on coral reefs and coral reefs are like the Arctic are at the bounds of

813

:

temperature and coral reefs are spilling into temperatures that

814

:

haven't existed for a long time on the earth.

815

:

And so they're the most threatened by climate change alongside Arctic ecosystems, which of

course they're moving in a different direction away from freezing.

816

:

so seeing the consequences of climate change with coral bleaching and the effects on

people and communities, it's just enormous.

817

:

And so it has to be the thing that you work on and from a people perspective, not just

from an environment perspective.

818

:

Yeah.

819

:

And second question, if you could have dinner with any great scientific mind, dead or

alive, fictional, who would it be?

820

:

Dennis Lindley.

821

:

Okay.

822

:

That was prepared.

823

:

Well, I've thought about it.

824

:

Yeah.

825

:

And it's because he was an early, early Bayesian and one, you know, he came through the

war, became a statistician through the war.

826

:

And again, all the war people, mean, Feynman and...

827

:

Von Neumann and all these people would be really interesting.

828

:

What I really would like to hear about his war experience, but also he confronted Fisher.

829

:

And Fisher was an asshole.

830

:

not enough people know that because as Lindley said, he was a brilliant man, but he was an

awful person.

831

:

And I think it's really important to understand that.

832

:

know, Aubrey Clayton illustrates this beautifully in his book.

833

:

Bertoulli's, Bernoulli's fallacy.

834

:

Bernoulli's fallacy.

835

:

That episode 50 of of the podcast.

836

:

Yeah.

837

:

And, know, he outlines not only the fallacy, but also the context under which, you know,

Galton and Pearson and Fisher were operating was not neutral.

838

:

And the, and their dominance and arrogance over everybody in statistics, which is what I

live now with the students that I teach, you know, back to talking about students and

839

:

trying to have a fresh mind.

840

:

Everything is so prescribed in statistics because of these people and not only because of

them, but because of their bad actions and their bad intentions.

841

:

I think talking to Lindley about that idea and how he kind of fought Fisher and what it

was like to be in that position would just be super fascinating.

842

:

And from all intents that I can tell, he was a really nice man as well.

843

:

Yeah.

844

:

Yeah.

845

:

Great answer.

846

:

And you're the first one to answer that.

847

:

Awesome.

848

:

Well, thank you very much, Aaron, for taking the time.

849

:

That was as awesome as I expected.

850

:

So yeah, thank you so much.

851

:

Learned a lot.

852

:

And well, you're welcome anytime on the show for the next paper.

853

:

All right.

854

:

And thank you everybody for joining us live here.

855

:

Thanks everybody.

856

:

Yeah.

857

:

So actually we had a great question afterwards, after the show, that I should have asked,

but didn't.

858

:

So we're recording that bonus.

859

:

So Shalom actually asked Aaron if there were any counterintuitive finding in the results

from the study.

860

:

Yeah, and I think the biggest thing is just the diversity of the trait itself.

861

:

if you look at the literature right now on sharks that are caught, it's mostly pelagic

species like blue sharks, thresher sharks, things like that, that are out in the open

862

:

ocean caught on long lines.

863

:

And those are what you see in the fin trait a lot.

864

:

But once we kind of dealt with those species,

865

:

It's the diversity of other species.

866

:

The fact that the meat trade basically permeates the whole world and anywhere that people

are catching these animals, they're eating them.

867

:

And then that, but that that is then ending up in international trade is what's really

surprising the diversity.

868

:

This has been another episode of Learning Bayesian Statistics.

869

:

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

870

:

episodes to help you reach true Bayesian state of mind.

871

:

That's learnbaystats.com.

872

:

Our theme music is Good Bayesian by Baba Brinkman, fit MC Lars and Meghiraam.

873

:

Check out his awesome work at bababrinkman.com.

874

:

I'm your host,

875

:

Alex and Dora.

876

:

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

877

:

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

LearnBasedDance.

878

:

Thank you so much for listening and for your support.

879

:

You're truly a good Bayesian.

880

:

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

than amazing.

881

:

Let's adjust those expectations.

882

:

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

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

Chapters

Video

Watch

More from YouTube

More Episodes
127. #127 Saving Sharks... with Python, Causal Inference and Aaron MacNeil
01:04:08
121. #121 Exploring Bayesian Structural Equation Modeling, with Nathaniel Forde
01:08:12
120. #120 Innovations in Infectious Disease Modeling, with Liza Semenova & Chris Wymant
01:01:39
119. #119 Causal Inference, Fiction Writing and Career Changes, with Robert Kubinec
01:25:00
117. #117 Unveiling the Power of Bayesian Experimental Design, with Desi Ivanova
01:13:11
114. #114 From the Field to the Lab – A Journey in Baseball Science, with Jacob Buffa
01:01:31
113. #113 A Deep Dive into Bayesian Stats, with Alex Andorra, ft. the Super Data Science Podcast
01:30:51
112. #112 Advanced Bayesian Regression, with Tomi Capretto
01:27:18
109. #109 Prior Sensitivity Analysis, Overfitting & Model Selection, with Sonja Winter
01:10:49
102. #102 Bayesian Structural Equation Modeling & Causal Inference in Psychometrics, with Ed Merkle
01:08:53
94. #94 Psychometrics Models & Choosing Priors, with Jonathan Templin
01:06:25
92. #92 How to Make Decision Under Uncertainty, with Gerd Gigerenzer
01:04:45
89. #89 Unlocking the Science of Exercise, Nutrition & Weight Management, with Eric Trexler
01:59:50
84. #84 Causality in Neuroscience & Psychology, with Konrad Kording
01:05:42
83. #83 Multilevel Regression, Post-Stratification & Electoral Dynamics, with Tarmo Jüristo
01:17:20
19. #19 Turing, Julia and Bayes in Economics, with Cameron Pfiffer
01:00:26
27. #27 Modeling the US Presidential Elections, with Andrew Gelman & Merlin Heidemanns
01:00:52
28. #28 Game Theory, Industrial Organization & Policy Design, with Shosh Vasserman
01:03:56
31. #31 Bayesian Cognitive Modeling & Decision-Making, with Michael Lee
01:09:18
34. #34 Multilevel Regression, Post-stratification & Missing Data, with Lauren Kennedy
01:12:39
40. #40 Bayesian Stats for the Speech & Language Sciences, with Allison Hilger and Timo Roettger
01:05:32
52. #52 Election forecasting models in Germany, with Marcus Gross
00:58:07
71. #71 Artificial Intelligence, Deepmind & Social Change, with Julien Cornebise
01:05:07
53. #53 Bayesian Stats for the Behavioral & Neural Sciences, with Todd Hudson
00:56:12
57. #57 Forecasting French Elections, with… Mystery Guest
01:21:48
77. #77 How a Simple Dress Helped Uncover Hidden Prejudices, with Pascal Wallisch
01:09:00