0:00

Let me show you how to be a

0:02

good piece. Today I am honored to host

0:04

Adam Bucharski, a leaning expert in

0:06

infectious disease modeling, and epidemic

0:09

forecasting. Adam is a professor

0:11

of infectious disease epidemiology and

0:13

co-director of the Center for

0:16

Epidemic Preparedness and Response at

0:18

the London School of Hygiene

0:21

and Tropical Medicine. His research

0:23

focuses on harnessing data and

0:26

analysis to improve epidemic preparedness,

0:28

and he has contributed real-time

0:31

analysis to governments and health

0:33

agencies during major outbreaks, including

0:35

Ebola, Zika and COVID-19.

0:37

In this episode, Adam

0:40

takes us inside the

0:42

world of epidemiological modeling,

0:44

discussing how these methods

0:46

help refine predictions and

0:48

inform public health decisions.

0:50

We explore the challenges

0:52

of modeling infectious diseases.

0:54

from data uncertainty to

0:56

real-time forecasting, and the

0:58

importance of communicating findings

1:00

effectively to policymakers and

1:02

the public. Adam also

1:04

highlights common misconceptions about

1:06

epidemiological data and dies into

1:09

the world of automation and

1:11

AI in Epidemic Response. This

1:13

is an early invasion statistics,

1:16

episode 130, recorded November 26,

1:18

2024. Let me show you

1:20

how to be a good

1:22

Bayesian, change your predictions after

1:25

taking information. And if you're

1:27

thinking I'll be less than

1:29

amazing, let's adjust those expectations.

1:31

What's a Bayesian? It's someone

1:34

who cares about evidence. Welcome

1:36

to Learning Bayesian statistics, a

1:38

podcast about patient inference, the

1:40

methods, the projects, and the

1:43

people who make it possible.

1:45

I'm your host, Alex Andorra.

1:47

like the country for any

1:49

info about the show. LearnBatesat.com

1:52

is labless to be. Show

1:54

notes, becoming a corporate sponsor,

1:56

unlocking vision merge, supporting the

1:58

show on Patreon. is

2:01

in there. That's learn base stats.com.

2:03

If you're interested in one-on-one mentorship,

2:05

online courses, or statistical consulting, feel

2:08

free to reach out and book

2:10

a call at topmate.io/Alex underscore and

2:12

Dora. See you around, folks, and

2:15

best patient wishes to you all.

2:17

And if today's discussion sparked ideas

2:20

for your business, well, our team

2:22

at Poymce Labs can help bring

2:24

them to life. Check us out

2:27

at Pymce Dash Labs.com. Hello

2:30

my dear patients, hope you're doing

2:33

well. Two main announcements for today.

2:35

First and foremost, thank you so

2:37

much. to all of you who

2:40

sent testimonials about learning vision statistics

2:42

and intuitive base to support my

2:44

green card application. I was genuinely

2:47

touched, surprised, and moved by the

2:49

number, kindness, and generosity of messages.

2:51

I am so happy and grateful

2:54

to be surrounded by truly-minded, kind,

2:56

and helpful people. like you, I

2:58

received more than 40 testimonials. So

3:01

thank you so much to all

3:03

of you who've taken the time

3:06

out of their busy day to

3:08

work about a very nerdy endeavor

3:10

and how it changed the field

3:13

of statistics according to you. I

3:15

will of course let you know

3:17

what happens with my application and

3:20

I'm already planning a very special

3:22

in-person surprise for all of you

3:24

for February. 2026 but more info

3:27

will come in good time. So

3:29

again, thank you so much and

3:31

in the meantime, leave long and

3:34

prosper or shall we say leave

3:36

base and prosper and talking about

3:38

prospering if you are interested in

3:41

baseball and patient statistics and working

3:43

in an MLP team like let's

3:45

say the Miami Monins so that

3:48

would mean working with me if

3:50

you'd love that get to my

3:52

me get to meet me and

3:55

maybe working together in the baseball

3:57

research and baseball solutions teams. Well,

4:00

now is the time. Our teams

4:02

are growing fast and this is

4:04

your chance to get in on

4:07

something that's very special. Honestly, I

4:09

absolutely love what's going on there.

4:11

And we've just opened two brand

4:14

new roles, one baseball analyst for

4:16

the Solutions Group that's more here

4:18

to our junior applicants and a

4:21

senior one, which is senior data

4:23

scientist in the research group. It's

4:25

mine. So if you are passionate

4:28

about baseball research, I love working

4:30

on collaborative teams and well bonus

4:32

points, you know your way around

4:35

offense modeling approaches like for instance

4:37

patient methods with Pymc and Stan

4:39

and neural networks time series all

4:42

that stuff. Well we definitely want

4:44

to hear for you. I think

4:46

what we're doing in Miami is

4:49

some of the most exciting in

4:51

the MLB right now. Let me

4:54

know if you have any questions,

4:56

if you're a patron of the

4:58

show, you're on the discord with

5:01

me. Feel free to send you

5:03

any of your comments, questions, recommendations

5:05

of people. Send that to your

5:08

friends and maybe see you soon.

5:10

Miami. On that note, let's go

5:12

on with the show. Adam Kucherski.

5:15

Welcome to Learning Vision Statistics. Thank

5:17

you. Yeah, thank you for taking

5:19

the time and bearing with me

5:22

we've had a few technological problems

5:24

to record that episode folks that

5:26

was That was that was difficult,

5:29

but you know as they say

5:31

The obstacle is the way so

5:33

we are here And a big

5:36

thank you to Chris Wyman to

5:38

for putting us in in contact

5:41

Chris and I met at Stankon

5:43

2024 in Oxford. Chris was actually

5:45

in a panel discussion that recorded

5:48

with him. and Lisa Simonova, that

5:50

was episode 120 for people who

5:52

are curious and want to hear

5:55

more about how cool epidemiological science

5:57

and computational biology are. So feel

5:59

free to take that out. With

6:02

Adam today, we're actually going to

6:04

tackle some similar topics, but you

6:06

have a very wide and very

6:09

broad research. So that's why I'm.

6:11

I'm super happy to have you

6:13

on the show today. You do

6:16

so many things and you are

6:18

also a great science communicator. You've

6:20

written several books as I've said

6:23

in the introduction. So yeah, all

6:25

of these books will be in

6:27

the show notes. So people feel

6:30

free to check them out. I

6:32

definitely recommend them. They are absolutely

6:35

fascinating. But before we touch on

6:37

the end, can you tell us

6:39

with your tree? Actually nowadays, Adam

6:42

and how you ended up working

6:44

on this. Yeah, sure. So, I

6:46

mean, my work kind of bridges

6:49

really understanding epidemics and getting better

6:51

at predicting and responding to them.

6:53

So there's a mix of aspects

6:56

to that. So some of it

6:58

is understanding the drivers of what

7:00

we see in terms of how

7:03

things spread. So for something like

7:05

thank you fever, that might be

7:07

climate influences, accumulation of immunity in

7:10

the population for something like, say

7:12

flu or COVID. the implications of

7:14

vaccination over time, how those viruses

7:17

evolve. Alongside that we're also doing

7:19

a lot of work to build

7:21

up the methods and tools that

7:24

we need to respond to very

7:26

quickly during COVID. A lot of

7:29

these things were developed very quickly,

7:31

often just working over weekends and

7:33

can we actually do a lot

7:36

better particularly for the predictable questions

7:38

that we know we're going to

7:40

have to answer. In terms of

7:43

how I got into it, my

7:45

backgrounds originally in maths, have applied

7:47

maths as my PhD. But even

7:50

that was starting to go more

7:52

in this direction of epidemiology and

7:54

questions around. how we can understand

7:57

the process of infectious disease. And

7:59

for me, as a field, it

8:01

kind of sits quite nicely between

8:04

something where having some mathematical and

8:06

statistical understanding can give you a

8:08

lot of value very quickly. But

8:11

also, there's enough unknowns about the

8:13

underlying rules and processes. It's not

8:16

some like physics where we have

8:18

a lot more mechanistic understanding. So

8:20

it means. sort of squeezing a

8:23

lot more out of the data

8:25

you have available, particularly under pressure

8:27

in a situation like an epidemic.

8:30

I see, okay, that's very interesting.

8:32

So it's something that, yeah, was

8:34

a long time coming, right? You've

8:37

always kind of been interested in

8:39

these topics, at least since your

8:41

past high school studies, right? Yeah,

8:44

I think there's a lot of

8:46

those questions where, yeah, obviously has

8:48

an enormous impact on people, but

8:51

also just from a kind of

8:53

curiosity point of view, these are

8:55

quite hard questions and often the

8:58

methods that are in the textbooks

9:00

don't quite work for what you

9:02

need to do. So it's an

9:05

ongoing interesting research area to be

9:07

in because almost every epidemic you

9:09

work in, what you thought you

9:11

knew and what you thought you

9:13

had solved. Suddenly you happen. And

9:15

so yeah, it's kept me busy

9:17

and kept me interested along with

9:20

many of my colleagues. Yeah. And

9:22

what about patient stance? Do you

9:24

remember when you were first introduced

9:26

to them? So I think it's

9:28

particularly doing PhD. So I think

9:30

my kind of undergrad is a

9:32

lot more traditional math. So it's

9:34

a relatively little statistics, a lot

9:36

more kind of theory. to the

9:38

sort of thing where you'd learn

9:40

measure theory rather than learning the

9:42

kind of coming at it from

9:44

a data point of view. I

9:46

think during my PhD, I started

9:49

working a lot more with data,

9:51

and especially if you start to

9:53

look at processes involved in diseases

9:55

versus involving immunity, then that's the

9:57

next question is, well, how could

9:59

we estimate things meaning? meaningfully within

10:01

that. And then particularly if you

10:03

have patchy bits of data or different

10:05

studies that you want to combine in

10:08

some way or if you have some

10:10

analysis you've put together and then you

10:12

want to make statements about additional

10:15

data coming in, basic statistics are

10:17

a natural framework for doing that.

10:19

And so really in the mix

10:21

of my work in some cases

10:23

you just get simple probability problems

10:25

and base formulas are nice. get

10:27

out of jail way of solving

10:30

that problem. And in other cases,

10:32

it's more of a framework for

10:34

thinking about the entire analysis that,

10:36

you know, you want to be able

10:38

to come to the best possible explanation

10:40

or the best possible evidence at

10:42

even point in time, but then you

10:45

want to update that as you get

10:47

more data coming in and again it

10:49

gives you a very nice toolkit for

10:51

doing that. Yeah, yeah. Yeah, that makes

10:53

sense. Damn. That's very fussy to

10:55

you here. brings me a lot

10:57

of more questions for you

10:59

Anna. Yes, one of these

11:01

questions that I have for

11:04

you is, especially during COVID,

11:06

Bayesian modeling became a crucial tool.

11:08

I remember I did some work

11:11

at my very low level at

11:13

that time, but you did a

11:15

bunch of work on that and

11:18

as you were saying, I mentioned

11:20

there were a lot of very

11:23

short nights and very long weeks.

11:25

Can you share specific examples of

11:27

how these models were able to

11:29

inform public health decisions? Yes,

11:31

I think there's quite a

11:34

few instances, particularly around scenarios

11:36

where models can be a

11:38

very useful kind of decision

11:40

support tool. Because essentially if you're

11:42

going to make a decision about

11:44

what to do in an epidemic,

11:47

everyone has a model in their

11:49

head because if you think let's do

11:51

this, let's try this. you're making some assumptions

11:53

about what you think is going to happen,

11:55

you make some assumptions about how you think

11:57

epidemics work. So models are a very nice way

11:59

of it. allowing us to really write down

12:01

and formalize what we think those processes

12:04

are and what we think the

12:06

interventions are going to do. And then

12:08

we can debate whether that's reasonable,

12:10

whether that's not, so we can see

12:12

if that generates any counterintuitive effects.

12:15

But in doing that, you really want to

12:17

capture the extent of information you have

12:19

about what you're dealing with. So even,

12:21

for example, the question of how much

12:23

effort is it going to take to

12:25

get transmission down? One of the key

12:28

things that is going to influence that

12:30

is how much transmission there really is.

12:32

So a lot of the analysis that we did,

12:34

being able to pass around uncertainty is

12:36

really important. So very early in

12:38

the pandemic, for instance, because we

12:40

were relying on quite uncertainty from

12:43

China, quite uncertain data about

12:45

exported cases, we had some sense

12:47

of what transmission would do in

12:49

a country without any control measures,

12:51

but there was quite a large

12:53

amount of uncertainty on that probably

12:55

reproduction number. somewhere between two and

12:58

four. And so what we wanted to do

13:00

is when we generate UK scenarios to

13:02

present them, we didn't want to

13:04

give one number, we wanted to

13:06

say, look, we don't know actually it's more

13:08

the upper end or lower end of this,

13:11

but we want to kind of define

13:13

that uncertainty in what we simulates.

13:15

And then after that wave comes through, you

13:17

know, we had a lockdown as like many

13:19

other countries, there was this push to

13:21

reopen and then the question is okay, so

13:24

we give we have that wave. what's

13:26

the reopening going to look like? And

13:28

again, that's where these kind of basic

13:30

methods we have very helpful, because

13:32

we can take the uncertainty and

13:35

perhaps we've got more confidence now

13:37

about what we're dealing with, but then

13:39

we want to pass that level of

13:41

confidence into what's going on in the

13:43

future. And then, you know, it's very

13:45

inside emerging, that became even more important

13:47

because often the alpha we would

13:49

have some degree of confidence about

13:51

how much more transmissible it was,

13:53

And in any analysis, for

13:56

example, we did a lot of

13:58

work early in 2021, how... quickly you're

14:00

going to have to vaccinate. And I

14:02

think there's a lot of pressure to

14:04

lift lockdown and try to understand that

14:07

trade-up. If you're vaccinating at this rate

14:09

and you're lifting lockdown this quickly, what's

14:11

that going to look like? And again, none

14:13

of those values you had very precisely and

14:15

you had more epidemic data coming in

14:17

all the time. And as particularly that

14:20

process of reopening in 2021, there's a

14:22

very tight relationship between models and policy

14:24

because in the UK more of the few

14:26

times where the policy was actually very informed

14:29

by what was useful from a technical point

14:31

of view. So what they did was they

14:33

had these series of steps in the road

14:35

madness. It was designed to give enough signal

14:38

in the data post each step that

14:40

if something was going very wrong, they

14:42

wouldn't have implemented another step before

14:44

you had that signal. So they

14:46

would have deliberately spaced out so

14:49

the models and the epidemiologist would

14:51

have enough time to work out what that

14:53

relaxation had done. So again, from a

14:55

basic important... point of view that

14:57

that's really nice because it gives

14:59

you enough time to kind of

15:02

update your posterior sensibly before you

15:04

do the next step and work out

15:06

what effect that's going to have. Right, yeah,

15:08

yeah, yeah, yeah, okay, that's, that's, I

15:10

didn't know, I didn't know there

15:12

was that level of coordination where

15:14

you could actually do that, you know, go

15:16

there. There were many examples where,

15:19

yeah, it didn't, it didn't work

15:21

in such a coordinated fashion, but

15:23

I think that was that that

15:25

roadmap reopening is one where there

15:27

was a much tighter relationship

15:29

between the questions come

15:31

down from policy and what you

15:34

know models needed to say something

15:36

sensible about the

15:38

implications. Yeah, yeah, and I guess, I mean,

15:41

I was so wondering, you know, how

15:43

much do you think the pendulum

15:45

has swung back from from then,

15:47

you know, like do you think will

15:50

be faster to implement

15:52

these workflows and improve them

15:54

next time there is a

15:57

pandemic? Or will we have

15:59

to? work from a blank slate

16:01

because politics is so short-sighted

16:04

with very short, you know,

16:06

cycles? Yeah, I think that's

16:08

a really good question. I

16:10

think there's some things that

16:12

have been positive in terms

16:15

of progress, so I think in

16:17

some people we're doing others in

16:19

consolidating a lot of the tools

16:21

that are available. So some things

16:23

now that we, you know, look

16:25

at quick questions that

16:28

age 5 anyone. Questions that

16:30

I just wouldn't have bothered doing

16:32

previously just as a curiosity because

16:34

it would have been three four

16:36

hours just for maybe question

16:38

and now in 20 minutes you can

16:40

get a rough example. So I think

16:42

that's quite nice is bringing things into

16:44

reach. But I think in terms of

16:46

just staff capacity and people to do this

16:48

work, I think there are a lot of people

16:51

who put a huge amount of time in

16:53

often as around the world getting pulled

16:55

off different roles and off other projects. And

16:57

so I think in a way in that

17:00

sense, we probably don't have

17:02

the same workforce who could undergo that

17:04

in our pressure for that a matter

17:06

of time. So in a way that

17:08

creates a necessity, we need to get

17:10

better tools because we don't have, I

17:12

think that volume of people, but

17:15

also just all our projects, we

17:17

basically took people off a lot

17:19

of other funded projects, and

17:21

I think we wouldn't have the results

17:23

to do that in the same way.

17:25

Now, so for me, I think particularly

17:27

what we're interested in the moment is

17:29

if you imagine the set of tasks you might

17:31

have to do in an epidemic. Some are tasks

17:34

that a lot of other people have to

17:36

do, and we can really predict ahead of

17:38

time you're going to need to do that.

17:40

So stuff like working out the severity of

17:42

infection or working out the transmission. We know

17:45

we're going to need that done. We know lots

17:47

of people who are going to do that.

17:49

So that's a really good task for automation

17:51

and getting really leadingage, basic

17:53

methods. standardized if people can ever want

17:55

to do that and we don't all have to

17:57

duplicate effort. There's other things that would be really

17:59

specific. Maybe within a country there's a

18:02

kind of subgroup that's particularly important

18:04

or there's a certain variance or

18:06

something you might want to deal

18:08

with and that's going to require

18:10

more domain knowledge. And I think

18:12

at the moment, even if you

18:14

talk to people at different outbreak

18:16

organizations around the world, they probably

18:19

spend six or 70% of their

18:21

time on quite low level, predictable

18:23

data ranking type questions rather than

18:25

30% of expertise led to questions

18:27

they'd want. I don't necessarily think

18:29

in a pandemic people they're not

18:31

working less hard because everyone wants

18:33

to contribute as much as they

18:35

can, but I think what I'd

18:38

like to see is a lot

18:40

more time on that 30% and

18:42

going much deeper into what we

18:44

can understand, rather than spending a

18:46

lot of time, you know, even

18:48

just trying to get ahead the

18:50

sense of the data and the

18:52

basic tasks to answer very simple

18:55

questions. Not sure I'm

18:57

really reassured by that answer, but

18:59

yeah, that's absolutely realistic, the realistic

19:01

answer. Some things I'm also really

19:03

curious about as a, because I

19:05

see that a lot as a

19:08

modeler myself, is where the, you

19:10

know, the common misconceptions that you've

19:12

seen the public or elected officials

19:14

or even professionals have about epidemiological

19:16

data in what the scientific community

19:18

can do to clarify these misunderstandings.

19:20

Yeah, I think that's a really

19:23

good question. I mean, I think

19:25

one common misconception is even just

19:27

what data are as a thing.

19:29

You know, I think often when

19:31

people talk about law data, they're

19:33

talking about inferred estimates. So I

19:36

think one example of this is

19:38

excess stats, that I think the

19:40

media... often treat as a measured

19:42

thing rather than actually a counterfactual.

19:44

And I think I remember talking

19:46

to journalists about the comparison with

19:49

food. I think a lot of

19:51

them discovered that the flu mortality

19:53

statistics they've used every year actually

19:55

involved some quite big modeling assumptions

19:57

about seasonal patterns behind them. And

19:59

there isn't just this magic number

20:01

that we measure. Right. I think

20:04

similarly with a lot of the

20:06

estimates, for example, in the UK

20:08

they ran randomized testing surveys surveys.

20:10

randomly tested people in the community

20:12

and reported it. And that number

20:14

that wasn't a raw proportion because

20:17

of course you needed to do

20:19

some standardization across groups and waiting

20:21

and in some cases because it

20:23

was it was quite noisy across

20:25

multiple panels they would have you

20:27

know a kind of smooth underlying

20:30

proportion and then in further. So

20:32

it was sensible modeling but there

20:34

was modeling behind that and I

20:36

think the misconception often is that

20:38

these things award data, but often,

20:40

yeah, even a very simple thing

20:42

if you calculate, if you run

20:45

a survey and calculate a proportion,

20:47

you're making some modeling assumptions, or

20:49

if you know, are we going

20:51

to adjust for age, why are

20:53

you adjusting page, why are you

20:55

not adjusting to something else, you're

20:58

making assumptions about what you think

21:00

are important drivers of that thing

21:02

you're measuring. So I think that's

21:04

probably one thing we can get

21:06

better of communicating is that actually

21:08

model estimates. And actually a lot

21:10

of the things that sometimes people

21:13

treat as super complex models are

21:15

actually just quite straight simple steps

21:17

for thinking about data. So even

21:19

one of, you know, if you've

21:21

got a bunch of hospitalizations over

21:23

time and you want to know

21:26

when the infections happened, you can

21:28

use quite simple like a decomposition

21:30

model just to take that and

21:32

work out when the infections are.

21:34

And I think often in public

21:36

perception when people talk about models,

21:39

what they mean is very complicated.

21:41

scenario models that people are assuming

21:43

are used to make forecasts. I

21:45

think it's often a kind of

21:47

a crystal ball of uses big

21:49

complex models to say what's going

21:51

to happen. And I think in

21:54

reality what happens is that often

21:56

the more complex models are used

21:58

for scenarios and like what if

22:00

because we can't we can't make

22:02

forecasts often in pandemics as well

22:04

because you'll be forecasting what policy

22:07

makers are going to do. And

22:09

if your models are used as

22:11

tools to support their decisions, it

22:13

becomes quite odd to use that

22:15

as a forecast. I mean, I

22:17

think the example I sometimes get

22:20

is like, yeah. And at the

22:22

point in time you want to

22:24

understand the implications of your decisions.

22:26

What you don't really want is

22:28

someone on your shoulder saying, I

22:30

bet you're going to fold later

22:32

this round. What you want is

22:35

someone who can say, look, this

22:37

is likely, this is likely the

22:39

situation you're going to face, this

22:41

is the risk you're taking on,

22:43

if you fold, this is potentially

22:45

what the outcomes are going to

22:48

be. So that's really how a

22:50

lot of these models are used

22:52

in kind of decision support. But

22:54

I think in public consciousnessness, they're

22:56

often, this is going to happen.

22:58

in the future. And I think

23:00

part of it is communication and

23:03

particularly I think getting people to

23:05

be able to play with the

23:07

simple models can be very helpful.

23:09

So they realize it's not this

23:11

super complex thing. It's probably what

23:13

they're doing in their head, but

23:16

just writing it down in a

23:18

bit more restructured way. Yeah, yeah.

23:20

I find also indeed, you know,

23:22

walking through scenarios is something that's

23:24

really helpful to people. because, well,

23:26

they can imagine the scenarios and

23:29

that's much more, you know, tangible

23:31

and concrete than numbers and posterior

23:33

distributions. I can see that all.

23:35

I mean, sports is a lot

23:37

like that, especially baseball, lots of

23:39

different discrete scenarios. Yeah, and I

23:41

think, you know, often people, particularly

23:44

epidemics, are kind of doing it

23:46

in their head all in the

23:48

time that, you know, The epidemic's

23:50

going off because it's going to

23:52

be a problem. and then you

23:54

get a few data points that

23:57

seem to be tailing off a

23:59

bit. And basically everyone's doing that

24:01

updating in their head of where

24:03

they think it's going. But they

24:05

often, I would have loved to

24:07

have seen more politicians and journalists

24:09

actually write down their, yeah, you

24:12

could get them to actually just

24:14

write down their prediction and their

24:16

kind of distribution of what they

24:18

think is going to happen and

24:20

then see how there's update and

24:22

then you could actually. give

24:25

them a better understanding of what's

24:27

going on in their head relative

24:29

to actually what's possible given the

24:32

data was coming in. Yeah, yeah,

24:34

but So I completely agree with

24:36

that and I really love that

24:38

But I think here the incentives

24:40

are really bad for politicians might

24:42

be is then if you like

24:44

if you publicly and privately I

24:47

think I think yeah, I think

24:49

that's I think that's really that's

24:51

I think that's a really good

24:53

point. Yeah, I think getting politicians

24:55

to write that what it's going

24:57

to have in public is very

24:59

difficult. But I think even privately,

25:01

I think there were a lot

25:04

of people probably not doing that

25:06

that would. Yeah, so I, yeah,

25:08

I mean, amongst colleagues and stuff,

25:10

we sometimes just had, just probably

25:12

more things, when you think this

25:14

study comes out, what do you

25:16

think is going to show? And

25:19

I think it's sometimes quite, we

25:21

like to kid ourselves. So even

25:23

if it's just writing. I'm like

25:25

completely on board in helping people

25:27

develop a more probabilistic thinking. You

25:29

know, I think you doing the

25:31

work you're doing and also the

25:34

communication work you're doing is very

25:36

important. All your books, that's also

25:38

why I have these podcasts, people

25:40

like Nate Silver. and writing books

25:42

is very important too. I don't

25:44

know if you read his last

25:46

book on the edge, but that's

25:49

very important too for that, right?

25:51

And I think that'd be awesome

25:53

if we were gearing towards that

25:55

direction. But yeah, the problem is

25:57

like the incentives in. The public

25:59

incentives and politics are so bad

26:01

when it comes to that that

26:04

you actually have a much better

26:06

Standing if you just you know

26:08

say Anything and and everything and

26:10

just are not held accountable for

26:12

that then actually You know betting

26:14

on something that would happen and

26:16

then and then changing course if

26:19

actually what you said would happen

26:21

did not and then you know

26:23

I think partly from the communication,

26:25

there's also the, yeah, what are

26:27

the things that we can do

26:29

meaningfully and what are the things

26:31

that you can encourage people to

26:34

do versus just want to be

26:36

feasible. And I think also just

26:38

from a modeling point of view,

26:40

yeah, sometimes there's this idea that

26:42

for political decisions, you know, you

26:44

should have this big model with

26:46

absolutely everything in and all the

26:49

kind of the weights of how

26:51

you do everything. And I can't

26:53

see any political party wanting to

26:55

do it because ultimately those things

26:57

are going to be weighed, not.

26:59

in a kind of written down,

27:01

you know, we're going to put

27:04

10% on this and 15% on

27:06

this. And so I think it's

27:08

working out, yeah, where the science

27:10

can be really informative and where

27:12

actually there's more of that kind

27:14

of human political element and that,

27:16

you know, that's not the best

27:19

battle to be through fighting at

27:21

this point. Yeah, yeah, yeah, it's

27:23

a great point. And actually in

27:25

your... Talking about your books in

27:27

the rules of contagion, you explore

27:29

why things spread. And I really

27:31

love that. So can you tell

27:34

us a bit more about that?

27:36

And how does patient thinking help

27:38

in understanding these patterns, especially for

27:40

diseases? Yeah. So I think that's,

27:42

um, striking what in the book

27:44

is a set out. thinking that

27:46

there would be these analogies in

27:48

other fields and I wasn't sure

27:51

necessarily how strong there would be.

27:53

But the more I dug into

27:55

that, I mean, there's actually... there

27:57

were in many cases very explicit

27:59

and very generally informative. For example,

28:01

after the 2008 financial crisis, there

28:03

was a very purely described epidemic

28:06

thinking that drove a lot of

28:08

the interventions resource, things like ring

28:10

fencing, things like capital requirements of

28:12

banks that were risk in network.

28:14

It's really about thinking about it

28:16

like a contagion problem. Similarly, if

28:18

you dig into the history of

28:21

companies like Buzzfeed that were very

28:23

good at generating viral content. They

28:25

were actually writing research reports on

28:27

how you evaluate the reproduction number

28:29

of marketing campaigns. And actually, that's

28:31

what we discovered, that Buzzfeed journalists

28:33

would have a measure equivalent to

28:36

the reproduction number as a metric

28:38

for their articles. So it wasn't

28:40

just this quite fuzzy comparison. Actually,

28:42

this was the same bits of

28:44

theory. that were appearing between these

28:46

two fields. I think one thing

28:48

I find quite interesting on the

28:51

Bayesian angle in how things spread

28:53

is particularly some of the debates

28:55

around how you convince people and

28:57

how people adopt beliefs and they

28:59

take off. Because there was quite

29:01

a popular idea for a while

29:03

known as the backfire effect, which

29:06

is where if you try and

29:08

convince someone, you can end up

29:10

basically just strengthening their existing belief.

29:12

and this idea that attempts to

29:14

change people's beliefs can kind of

29:16

backfire, which also doesn't bode well

29:18

for any kind of social progress

29:21

because it's this idea if you

29:23

try and convince someone that marriage

29:25

equality or something is a good

29:27

idea, it's just going to need

29:29

them to be entrenched. But what

29:31

subsequently happened is a lot of

29:33

the work, both on the applied

29:36

side of actually people kind of

29:38

get these, get support for these

29:40

kind of, this progress. but also

29:42

on some of the scientific side

29:44

of people studying them, suggested it's

29:46

actually much closer to a Bayesian

29:48

problem, that it's not that you're

29:51

leaving people. to entrench their beliefs.

29:53

Rather, if you give people weak

29:55

evidence, you're not going to shift

29:57

their distribution much. I thought it

29:59

was kind of really interesting. And

30:01

I hadn't actually thought about it

30:03

quite much in that way. And

30:06

it kind of makes sense that

30:08

even if you've got like a

30:10

quite strong prior for something, if

30:12

someone gives you evidence that agrees

30:14

with that prior. it's going to

30:16

look pretty similar after us, and

30:18

especially because we're not doing all

30:21

those calculations in our head. We're

30:23

just sort of seeing the feeling

30:25

we've come away with. And it

30:27

really struck me that actually in

30:29

those situations, we're probably much better

30:31

evaluating the effects of evidence that

30:33

we disagree with, because the posterior,

30:36

you know, just mathematically expected the

30:38

posterior to move more. And so

30:40

perhaps the situations where what people

30:42

were thinking was about flash effects,

30:44

it's more just we're better. critiquing

30:46

evidence we disagree with rather than

30:48

the ones that kind of line

30:50

up because if it agrees with

30:53

us we're going to walk away

30:55

the same opinion afterwards anyway. So

30:57

yeah I think that was it's

30:59

obviously a much harder problem to

31:01

study in terms of spread of

31:03

relief and there's so many factors

31:05

that can play into that but

31:08

yeah there's still a lot of

31:10

ongoing debate on the extent to

31:12

which people's adoption of beliefs and

31:14

behaviors, these Bayesian versus you know

31:16

some other factors that kind of

31:18

explain how those are updated over

31:20

time. Yeah, I see. That's, that's

31:23

really amazing. I love that. I

31:25

didn't know that, that example about,

31:27

what's the, what's the, what's the

31:29

website you were saying? Ah, well,

31:31

Bosfield, yeah, yes, thank you. I

31:33

was, it's remarkable, yeah, and it's,

31:35

it was coming to it because

31:38

they did, you know, campaigns about

31:40

a hurricane Katrina with. I mean,

31:42

none of these were viral and

31:44

this one they keep findings that

31:46

this wasn't like COVID where it

31:48

just spends, spends, spends, but you

31:50

know, for 10 shares they might

31:53

get an extra seven or eight.

31:55

So, you know, if you get

31:57

it, if you spark lots of

31:59

little clusters of sharing, you might

32:01

actually. yet quite quite considerable additional

32:03

uptake as a result. I think

32:05

there was one that was a

32:08

marketing campaign for detergent and it

32:10

wasn't accountable at all basically. So

32:12

there was quite a nice quantifying

32:14

that there's certain things that people

32:16

obviously want to tell people about

32:18

and others that even if you

32:20

got the biggest marketing budget in

32:23

the world you're going to struggle

32:25

to make washing up liquid contagious.

32:28

You know, I'm something I'm I'm curious

32:30

about and I asked this question to

32:32

Chris to at Stankon is Confrictly what

32:35

does it look like for you to

32:37

work on an epidemiological model? Like

32:39

who are you talking to and what's

32:42

your workflow and technical stack? Yeah, that's

32:44

really good question. Generally, anything we build

32:46

starts with the problem and often that

32:49

problem either comes from someone, if

32:51

there's a policy where a question comes

32:53

from, so on policy sites, so maybe

32:55

it's scientific advisors to certain agencies in

32:58

the case of like an applied organization

33:00

at MSF or WHO, it will

33:02

come from a representative who we're working

33:04

on maybe that outbreak. or that situation.

33:07

And there'll often be quite specific things

33:09

that people are interested. Yes, it

33:11

might be a forecast of actually that.

33:13

What are we dealing with? It might

33:16

be that there's a plan to implement

33:18

vaccination or control measure. A lot of

33:20

the work within the BOLA, for

33:22

example, different control measures being proposed and

33:25

people wanted some idea of the relative

33:27

impact that that would have. There's also,

33:29

of course, just on the scientific side,

33:32

so the work we've done. round

33:34

effects of behavioral immunity, it might be

33:36

that you've got lab colleagues who've noticed

33:38

some interesting features and want to work

33:41

out. How can I get sufficient

33:43

estimates out of my data? I mean,

33:45

I mean, this is, I think another

33:47

example where Bayesian thinking is very helpful

33:50

that you might have say lots and

33:52

lots of antibody responses. And you

33:54

don't want to analyze them all as

33:56

like individual data sets because there's going

33:59

to be some commonalities in just how

34:01

the biology works between individuals. So having

34:03

this kind of hierarchical models can

34:05

be very powerful because you can have

34:08

shared information just on the underlying dynamics

34:10

across the population, but you can also

34:12

have individual features of what we

34:14

look at previously and this sort of

34:17

thing. the kind of model was that

34:19

it depends a bit on whether it's

34:21

similar to another problem you've seen. So

34:24

a case of policy question, often

34:26

what we'll do is, you know, in

34:28

the many, many models and things we've

34:30

dealt with previously, you'll find the thing

34:33

that's most similar. And I think increasingly

34:35

we're seeing progress in libraries. So

34:37

a lot of the work we do

34:39

in is in our or in some

34:42

cases you can kind of stand back

34:44

ends. And so you'll take something

34:46

that maybe is from the library that

34:48

you think is most appropriate for that

34:51

and then a model that's templated up

34:53

that's closest to that. And then ideally

34:55

you'll just plug in and work

34:57

straight away. Often you might either adapt

35:00

some features of the model process in

35:02

terms of say how transmissions happen or

35:04

what groups are affected or you

35:07

might bring in different data. You might

35:09

have a model set up for UK

35:11

contact structure and you'll just import data

35:14

from somewhere else so you can adapt

35:16

it. In other cases, you might

35:18

get something that's a very specific, quite

35:20

neat question. So, you know, for example,

35:23

I don't know, like the ones doing

35:25

COVID, testing people at certain types of

35:27

gathering or something, and that's not

35:29

something necessarily you have a model off

35:32

the shelf, but in some cases, the

35:34

equation is quite simple to write down.

35:36

You know, if you've got this

35:38

many people and you're testing this many

35:41

at this point in time, and in

35:43

that case, we might just build something

35:45

more to smoke. The challenge always, I

35:48

mean like with any kind of

35:50

software development problem, is to what extent

35:52

do you do something quick for a

35:54

very specific problem versus take on coding

35:57

debt if you've got to do that

35:59

problem repeatedly. And so I think

36:01

we're getting a better sense now. There's

36:03

certain things that are sufficiently complex. There's

36:06

large enough script for bugs. It's going

36:08

to be useful for enough people.

36:10

Well, actually, that makes sense to package

36:12

up more consistent library. And there's other

36:15

things that actually are simpler enough and

36:17

transparent enough and kind of bespoke enough

36:19

that you don't want to build

36:21

a software tool for every single one

36:24

of those. Some of those you can

36:26

just do quickly as the problem arises.

36:28

Yeah. Yeah. That makes tons of sense.

36:31

And once you're, I'm wondering also

36:33

the size of the of the teams

36:35

in these, in these cases, because obviously

36:37

each time I talk to a, to

36:40

a modeler in your field, it

36:42

sounds like the models are really big

36:44

and huge and and take a lot

36:46

of time to work on because they

36:49

are so complex. So yeah, I'm wondering

36:51

how many people does it take

36:53

to work on a model and how

36:55

do you actually do that? Because Like

36:58

yeah, is everybody working on the model

37:00

at the same time? Do you have

37:02

some team for that part of

37:04

the model, another team for the other

37:07

part? How does that? Yeah, so I

37:09

think, I mean, to give you an

37:11

example of some of the big

37:13

kerbic scenario models that are used by

37:16

our group in the UK, I haven't

37:18

got the get have worked on it.

37:20

It's probably, you know, at least 10,

37:23

50 people who made it's a

37:25

central contributions to that co-based at their

37:27

various points in time. And in time.

37:30

And in some case, can be, I

37:32

mean, ideally we'd make these things as

37:34

module as possible. So early on,

37:36

for instance, I had some colleagues who

37:39

have focusing very much on the transmission

37:41

dynamics and kind of how people interact,

37:43

what interventions were going to be,

37:45

and I worked a lot more on

37:48

the sort of disease burden module. So

37:50

once you have transmission infections, you can

37:52

convert those infections, the investment of how

37:55

many are you going to be

37:57

hospitalized for. So then you have that

37:59

kind of basic model structure, but then

38:01

over time that got... expanded because variance

38:04

came. in, vaccines came in, and

38:06

you ended up with multiple versions of

38:08

those models. And there's always this kind

38:10

of challenge of, you know, do you

38:13

make kind of one core model that

38:15

has sufficient flexibility to do all

38:17

those problems? Or do you kind of

38:19

fork a model and use it for

38:22

a special case? And you're not going

38:24

to use it again. So that model

38:26

has actually been used for multiple

38:28

countries. We adapted it to a whole

38:31

range of different settings, I think. patterns

38:33

of immunity and infection. And in that

38:35

case, it didn't make sense just

38:37

to build that into the original model

38:40

because it was just such a kind

38:42

of specific example. But I think that's

38:44

one thing where as well, if we

38:47

had to go back now, because

38:49

there's so many versions of the model

38:51

and so many applications because it's real

38:53

time we had to deliver that very

38:56

quickly, it's obviously harder to now say

38:58

how we would do that for

39:00

flu epidemic. And so I think what

39:02

we're trying to move more towards is

39:05

these very modular examples that you can

39:07

plug in all the bits you

39:09

need but also have that capacity to

39:11

adapt it and I think that's just

39:14

that's that's kind of an ongoing challenge

39:16

that you can have these things that

39:18

very stable and structured but very

39:20

hard to adapt or you can have

39:23

these things that very maybe flexible and

39:25

easy to adapt but not necessarily as

39:27

kind of efficiently structured as you like.

39:30

I mean there are examples as

39:32

well modeling where it might be one

39:34

or two people developing something quite quickly

39:36

or just making use of a library.

39:39

I mean some of the popular

39:41

methods for example for estimating reproduction numbers

39:43

that tool will have been used by

39:45

a huge amount of people but obviously

39:48

the active contributors to the development might

39:50

be a lot smaller. Okay, yeah

39:52

I see. Yeah so a big big

39:55

diversity in the in the size of

39:57

the other projects. And do you have

39:59

a... Do you have a favorite type

40:02

of models actually that you'd like

40:04

to share with this? So I think

40:06

one of, well I think one of

40:08

the models that I think has delivered

40:11

a lot of value in various

40:13

things that we've worked on over the

40:15

years is actually both on some of

40:17

the H7 and 9 analysis we did

40:20

about 10 years ago for the efforts

40:22

in China where you had infections

40:24

coming from poultry and potential human transmission

40:26

as well. There's a big question of

40:29

looking at that human data, how much

40:31

was coming for human human infections. And

40:33

we actually fat. you've got an

40:35

analogous version of the problem with some

40:38

of the COVID variants. So for Delta,

40:40

how much of this was imported cases

40:42

from India versus transmission establishing in

40:44

countries. And in both those situations, so

40:47

first of all, it's a bit of

40:49

a modeling headache because a lot of

40:51

traditional models are structured in a way

40:54

that basically says if you've got

40:56

your cases over time. the new cases

40:58

that appear have to have been one

41:00

of those past ones that infected them.

41:03

So if you look at a

41:05

lot of the common calculations for how

41:07

you do reproduction numbers, it's known as

41:09

a generative model. So in other words,

41:12

when your equation is the kind of

41:14

specific version. So the new infections

41:16

are the products of the infections that

41:18

come before. Someone has caused that infection.

41:21

If you've got importations or spillover, that's

41:23

no longer the case. You've actually got

41:25

this additional term coming into your

41:27

equation. So it makes it a trickier

41:30

inference problem if you've got two groups

41:32

just to infecting someone. But it can

41:34

also be more powerful because in

41:36

the case of avian flu we knew

41:39

when the wild like poultry markets were

41:41

closed and in the case of Delta

41:43

we knew when the travel ban against

41:46

India was implemented. So what you've

41:48

got as a kind of estimation problem

41:50

you've got two things that influence your

41:52

infections but you know the kind of

41:55

shape of one of these, because you

41:57

know when the market goes, you

41:59

know when the flight buttons came in.

42:01

And suddenly that gives you a lot

42:04

more estimation power on the thing you

42:06

care about, which is humans' human

42:08

transmission. So I think for me that

42:10

that's just a really nice example of

42:13

a model that it's not super common.

42:15

It's very relatively easy to explain. It's

42:17

like there's two things that can

42:20

affect someone which is it. But actually,

42:22

you can squeeze a remarkable amount out

42:24

of your data as soon as you

42:27

know the shape of one of those

42:29

processes. Okay. And does that... I

42:31

mean, it obviously comes with significant challenges.

42:33

I'm using these models. detail of these

42:36

changes that you face when creating such

42:38

models? Yeah, I think there's a

42:40

whole mix. I mean, in some cases,

42:42

there's, you know, just just understanding the

42:45

person you define a model that for

42:47

a lot of outbreaks, there's a lot

42:49

of things you'd like to know

42:51

the role of in in a process.

42:54

So for example, how different age groups

42:56

are interacting or affected by certain things,

42:58

but in some cases you might not

43:01

have the data that you need

43:03

to actually pick that apart. So a

43:05

good example in COVID was a lot

43:07

of people, you know, the epidemic would

43:10

come down, a lot of people

43:12

would argue about why that was. And

43:14

actually if you just look at case

43:16

data or just look at deaths, it

43:19

could be immunity, it could be changing

43:21

in behavior. it could be something

43:23

to do with the climate, it could

43:25

be, you know, a whole range of

43:28

different things that could influence that transmission

43:30

and just looking at one time series

43:32

of animals, you can't distinguish what

43:34

those are. You really need some data

43:37

on antibiotics, you need some data on

43:39

social mixing to tell you which of

43:41

those is the most likely explanations.

43:43

I think that's often a big challenge

43:46

is where you have lots of potential

43:48

explanations and you can't actually untangle them.

43:50

I mean the other... to use H5N1

43:53

as a current example, unlike the...

43:55

H5N9 and Delta where we knew how

43:57

that shape of the introductions was changing.

43:59

We don't know that for H5. So

44:02

now you get cases popping up

44:04

and they say they haven't had a

44:06

contact with poetry, maybe it's a wild

44:08

word, maybe it's a human, we've got

44:11

no idea. And I think that's that's

44:13

a kind of big challenge. Almost

44:15

a model can't really tell you anything

44:17

at this point because this is the

44:20

data is so uninformative about the process

44:22

that we, yeah, there's lots of question

44:24

at the question at the moment

44:26

at the moment, and I get, and

44:29

I get, and I get, and I

44:31

get, I get, and I get, I

44:33

get, and I get, and I

44:35

get, and I get, and I get,

44:38

the output data is just too bad

44:40

to say much. We can design some,

44:43

you know, what if hypothetical models, but

44:45

I think that's much harder. I

44:47

think there's also just the technical challenge

44:49

that, you know, in terms of just

44:52

making sure that models you build are

44:54

without bugs, and then also edge cases,

44:56

another classic one that there's, there

44:58

can be sometimes some psychicalters, just things,

45:01

particularly if you're doing the kind of

45:03

delay processes. So one example, which I

45:05

think it was a sort of

45:07

communication one, but even if you have

45:10

an epidemic that's going up and you

45:12

suddenly stop it and you have a

45:14

delayed outcome like deaths, the peak in

45:17

when the epidemic stops isn't the

45:19

same as the peak in death because

45:21

you're doing a kind of delayed convolution,

45:23

we're doing it smooth out of lay

45:26

distribution. As a feature like that, you're

45:28

making sure that you've actually got

45:30

that relationship defined properly. And again, this

45:32

is I think why the move to

45:35

a lot more established libraries rather than

45:37

trying to make sure these things

45:39

are sort of bug-free in real-time. Okay,

45:41

yeah, yeah. And a question I often

45:44

get, you know, personally, and I'm guessing

45:46

you are getting a lot too, is,

45:48

okay, cool. I get uncertainty estimation,

45:50

you know, with the Bayesian models. Why

45:53

do I care? So

45:55

I think a lot of what

45:57

you want to do, particular decision

46:00

making. comes down to confidence and

46:02

evidence so particularly in in the

46:04

middle of an epidemic there's a

46:06

lot of things we don't know

46:08

with any confidence. So I guess

46:10

on the one hand you could

46:12

just ignore it and just go

46:14

for you know the point estimate

46:16

is it going up is it

46:18

going down but particularly if you

46:20

want to say you want to

46:22

ask is the epidemic under control

46:25

it's not very helpful to have

46:27

a yes-no always that you might

46:29

want to say yes is under

46:31

control how confident are you for

46:33

it is and again having that

46:35

uncertainty in a reproduction number, if

46:37

you're like, well, 100% of our

46:39

density is below one. And that's

46:41

what we had post-lockdowns and social

46:43

mixing day, for example, in the

46:45

UK, that there was uncertainty in

46:48

that distribution, but all of that

46:50

density was below one. So the

46:52

conclusion was we're very confident that

46:54

transmission is coming down. I think

46:56

similarly for some of the variants

46:58

we found that we found that

47:00

some of the variants we found

47:02

that you'd get uncertainty in the

47:04

estimates. So maybe it's 30% maybe

47:06

it. be very confident it is

47:08

more transmissible and I think it's

47:11

a difficult one because policy makers

47:13

sometimes love you know single answers

47:15

they don't want a kind of

47:17

vague but I think particularly if

47:19

your uncertainty lands either side of

47:21

a particular threshold that matters that

47:23

can in a way give you

47:25

more confident communicating of saying yeah

47:27

look this is definitely going up

47:29

or this is definitely almost definitely

47:31

on a different crop. Just it's

47:33

similar, you know, the clinical problem,

47:36

you look at the conference into

47:38

book, I think it's the equivalent

47:40

of that, and you want to

47:42

know, you know, how much can

47:44

you rely on this estimate? And

47:46

I think that's where the uncertainty

47:48

really comes in. Yeah, yeah, so

47:50

basically making sure that you're not

47:52

fooled by the variance of the

47:54

processes. Yeah, and I think especially

47:56

when you're dealing with exponential processes,

47:59

that you know, exponential processes, that

48:01

you know, exponential processes, that you

48:03

know, that you know, I was

48:05

going to accumulate. So there's something

48:07

that might feel quite small of

48:09

the transmission rates this. And actually

48:11

if it's slightly higher or slightly

48:13

lower, you get a very different

48:15

outcome. Even something, you know, say

48:17

it's a very simple example, each

48:19

person affects two others and you

48:22

cut transmission in half. So, you

48:24

know, each person affects one other,

48:26

just on one other, a tiny

48:28

amount of uncertainty there after, you

48:30

know, two months could be the

48:32

difference between a massive epidemic and

48:34

a handful of cases. you know

48:36

if you don't communicate that people

48:38

be asking well was a massive

48:40

epidemic when you said it would

48:42

it would be allowed based on

48:44

the best estimate. Yeah yeah yeah

48:47

that's definitely a good point in

48:49

I'm curious also to have your

48:51

to hear your thoughts about you

48:53

know the latest advancement step we've

48:55

seen in the last year in

48:57

artificial intelligence and large language models

48:59

because I'm guessing this is going

49:01

to have also an impact hopefully

49:03

for the best on epidemiology. So

49:05

yeah with these advancements how do

49:07

you see the future of the

49:10

field of epidemiology and and the

49:12

role of patient stance in it?

49:14

Yeah I think it's a huge

49:16

amount of They're really promising development.

49:18

I think a lot of it

49:20

at the moment that we're working

49:22

on in most others is trying

49:24

to find where do these solutions

49:26

work. I think essentially we've been

49:28

given this increasingly amazing toolkit, but

49:30

in many cases it wasn't necessarily

49:32

developed exact problems we're working on.

49:35

And so finding where the applications

49:37

work is going to be very

49:39

powerful, where are the ones that

49:41

is going to struggle more. And

49:43

so even... It's a bit of

49:45

AI or some more traditional machine

49:47

learning approaches that there might be

49:49

situations where if we're very interested

49:51

in the mechanism and we've got

49:53

some process we define our model,

49:55

perhaps that estimation or that prediction

49:58

approach. isn't optimal for actually what

50:00

we're trying to solve. But I

50:02

think equally, the pandemic re-showed, there

50:04

was a huge amount of data

50:06

out there that in many cases

50:08

we weren't able to interpret in

50:10

a meaningful way. So you might

50:12

have something like a social contact

50:14

and if it got by a

50:16

certain amount, but you put that

50:18

in a model and a very

50:21

meaningful change. But you might have

50:23

quite a lot of very noisy

50:25

data, which is giving some indication

50:27

about how people are behaving. But

50:29

you can't extract those features and

50:31

weigh them in a useful way

50:33

in the same sense that an

50:35

AI model could do. So I

50:37

think it's in a way to

50:39

find the range of problems. I

50:41

think broadly the challenge we have

50:43

for our works often that they're

50:46

quite rare events. So even what

50:48

do you what do you validate

50:50

against and what and what's your

50:52

input and what you kind of

50:54

outcome you kind of predict. But

50:56

I think within a epidemic especially

50:58

you get larger clinical data sets,

51:00

larger behavioral data sets of a

51:02

lot of value there. I think

51:04

also some of the the methods

51:06

like universal differential equations and other

51:09

things coming through where it's taking

51:11

that that sort of transmission model

51:13

structure but then incorporating things like

51:15

neural networks to to allow for

51:17

more complexity in understanding the patterns

51:19

that are kind of going on

51:21

alongside that I think there's a

51:23

lot of really interesting progress there

51:25

I think just also just more

51:27

generally in the field there's some

51:29

of the AI models that essentially

51:32

learning features that can I mean

51:34

by the forecasting is one example

51:36

where it's many many times faster

51:38

and less energy intensive for simulations.

51:40

I think again that kind of

51:42

you can find ways of approximating

51:44

more complex models. I think just

51:46

also just in terms of the

51:48

data that comes in I mean

51:50

a lot of outbreaks often describe

51:52

the narrative reports where it's kind

51:54

of like a few paragraphs and

51:57

so did this and went here

51:59

and did that. And one of

52:01

my colleagues did this nice prototyping

52:03

but equally on this. That's all

52:05

local models that can take those

52:07

quite difficult to interpret. Now it's

52:09

important to convert them to structured

52:11

data which you can then... Yeah,

52:13

basically, you know, like in a

52:15

lot of other fields, you would

52:17

like to see kind of a

52:20

better communication, but I think there's

52:22

a lot of work, really across

52:24

the spectrum, both in a lot

52:26

of other fields, you would like

52:28

to see kind of a better

52:30

communication between these. kind of models,

52:32

the LA models in the humans

52:34

in the loop, but clearly you

52:36

don't see these kind of models

52:38

being made entirely by artificial intelligence.

52:40

No, but we've tried it actually,

52:42

so even things like Copilot workspace,

52:45

which look, you try them out

52:47

on a simple problem and they

52:49

look really, really cool. So, yeah,

52:51

the idea that you can just

52:53

give it a code base and

52:55

tell it to do things. But

52:57

often for quite specific things, if

52:59

you say build me a model

53:01

with these features to do this,

53:03

I think because just the training

53:05

set, just doesn't include anything with

53:08

them, or very little with them,

53:10

that kind of problem. So if

53:12

you want an LLEM to build

53:14

you, to build you some JavaScript,

53:16

it's pretty good, because there's just

53:18

so much to train on. If

53:20

you want quite an inch compatible

53:22

model of an epidemic, it's struggles

53:24

in some way. So I think

53:26

finding the shape, you know, it's

53:28

great for, I've all these functions.

53:31

package them up and add the

53:33

documentation and this kind of stuff.

53:35

It makes a lot of tasks

53:37

faster. But I think, yeah, the

53:39

idea is she's going to do

53:41

science. I think it's maybe put

53:43

forward by people who've only worked

53:45

on a narrow set of scientific

53:47

problems. And there's probably some science

53:49

that's going to do well and

53:51

then there's some that's going to

53:53

really struggle with, hopefully not just

53:56

the boring bits. But we'll, I

53:58

think we need to, yeah, map

54:00

out where those gaps are. You

54:02

know, it's the same in my

54:04

feel, whether that's sports modeling or

54:06

or just, you know, statistical. modeling

54:08

in general where it's really funny

54:10

because if you ask an LLLM

54:12

what a hierarchical model is it

54:14

can explain into training well you

54:16

know that's like that's exactly what

54:19

I would explain my students but

54:21

then if you ask for Pymc

54:23

code of that the code the

54:25

model is not at all hierarchical

54:27

it's just a classic model you

54:29

know but the the L& will

54:31

like well this is a hierarchical

54:33

model but it's not. So I

54:35

definitely have to drive that back

54:37

and forth but it makes you

54:39

it makes you yeah more more

54:42

efficient also you know it can

54:44

like it helps me for instance

54:46

to kill my darlings maybe a

54:48

bit faster which is very important

54:50

in modeling that's always better to

54:52

do it in my experience with

54:54

a human but sometimes you don't

54:56

have someone at the same stat

54:58

level as you in your organization

55:00

or project so then you have

55:02

you have to do that on

55:04

your own and that can be

55:07

quite hard. I think there's something,

55:09

we're using it actually for things

55:11

like reviewing training materials, because sometimes

55:13

you want to throw at human

55:15

view, but sometimes you want the

55:17

first part of, you know, if

55:19

an applied field ecologists is reading

55:21

this, is there anything in there

55:23

that just is going to not

55:25

make sense or really, really obvious

55:27

things you want to fix. It

55:30

can just help accelerate a lot

55:32

of that kind of production process.

55:34

Okay. Okay. Yeah. That's actually, yeah.

55:36

That sounds very useful. So you've

55:38

been already very generous with your

55:40

time. So we're gonna I'm gonna

55:42

play us out here, but I'm

55:44

also curious to hear Your thoughts

55:46

about more educational perspective because you

55:48

do a lot as we've heard

55:50

a lot of public communication. So

55:52

Given your experience what educational initiatives

55:55

would you recommend to better prepare

55:57

next? generations of epidemiology. Yes, but

55:59

also policymakers and citizens in general.

56:01

Yeah, I think there's probably like

56:03

some specific angle within the audience,

56:05

which has become more general ones.

56:07

I mean, I think it's not

56:09

realistic to try and get people

56:11

to have a deep understanding of

56:13

epidemics any more than it's realistic

56:15

to have a very, very deep

56:18

understanding of kind of the unique

56:20

threats or something. But I think

56:22

there are certain features of epidemics

56:24

which a very important and very

56:26

often misunderstood. So I think one

56:28

aspect is exponential growth is a

56:30

concept that people find very difficult.

56:32

You look at who makes and

56:34

loses money in finance and there's

56:36

a definite divide in terms of

56:38

understanding that's a concept. And so

56:41

I think having more interest of

56:43

that across more people is very

56:45

helpful. Similarly, things like having lagged

56:47

outcomes that, you know, if you

56:49

have an event you care about

56:51

and then an effect that happens

56:53

later. That's something that's kind of

56:55

widely misinterpreted during the pandemic. And

56:57

yeah, I think it wasn't increasingly,

56:59

as code went on, it wasn't

57:01

just people who hold academic posts

57:03

in epidemiologists that were making a

57:06

lot of useful contributions. You had

57:08

a lot of people who were

57:10

adjacent industries who maybe, you know,

57:12

actually doing finance and other bits

57:14

of, you know, academic work or

57:16

even just, you know, mass teachers,

57:18

whatever doing. quite useful stuff because

57:20

they hadn't understand those concepts when

57:22

you get a feel for some

57:24

of those data problems that just

57:26

needed more eyes on them. So

57:29

I think for me that's where

57:31

a lot of value is. It's

57:33

like how do we have more

57:35

people who can just not get

57:37

very basic things wrong and just

57:39

have useful eyes on the problem

57:41

even if they don't know the

57:43

ins and outs exactly how you

57:45

calculate that specific prancer. I think

57:47

more generally though we're also just

57:49

seeing I think it's a challenge

57:52

for a lot of fields that

57:54

with the emergency AI with the

57:56

mode of flying the sort of

57:58

more complex understand of epidemics. I

58:01

think we are moving into a

58:03

world that's much harder to sort

58:05

of teach yourself everything from scratch.

58:07

I mean, you know, if you

58:10

think about even statistics as a

58:12

field, the sort of statistics that

58:14

practitioners do relative to what's important

58:16

schools is now very different. You

58:19

know, it's the sort of doing

58:21

kind of regression lines and the

58:23

things that, you know, I did

58:25

at a level. And I think

58:28

we're seeing that in a lot

58:30

of fields now, where actually the

58:32

cutting edge is so far detached.

58:34

partly challenges in communication because even

58:37

if you, a very interesting climate,

58:39

you can't, I can't go and

58:41

run a climate model in the

58:44

way that you might be able

58:46

to do a simple mathematical proof

58:48

or a simple statistical problem. So

58:50

I think there's that kind of

58:53

relationship of how do we build

58:55

trust and enough understanding of how

58:57

those fields work that people can

58:59

engage with them even though actually

59:02

that the kind of cutting edge

59:04

is now. so computationally intensive, in

59:06

some cases just so difficult to

59:08

explain in terms of the algorithms.

59:11

Yeah, I think like AI is

59:13

a problem example, but I think

59:15

there's a lot of those situations

59:17

where it feels like there's a

59:20

bit of a growing gap that

59:22

we need to bridge. Yeah, yeah,

59:24

yeah. Yeah, definitely. I mean, completely

59:27

agree with everything you just said.

59:29

That's a topic that's the other

59:31

to me, obviously, and... We've talked

59:33

about that several times on the

59:36

podcast. I think one of the

59:38

best episodes we've done, but that

59:40

was episode 50 with Sir David

59:42

Piegel Halter, only sir, he had

59:45

to have been on the podcast.

59:47

But yeah, that was a great

59:49

episode. David is an awesome communicator

59:51

also. So I'll put that episode

59:54

in the show notes for people

59:56

who want to dig deeper. one

59:58

of the episodes I recommend a

1:00:00

lot. And then this, the next

1:00:03

one, episode 51, with Aubrey Clayton

1:00:05

about his book. and the crisis

1:00:07

of modern science really recommend the

1:00:10

book and end the episode also

1:00:12

because these two together make a

1:00:14

great combination if you're interested in

1:00:16

epidemiology. Okay, well, I am. That

1:00:19

was awesome. Really, thank you so

1:00:21

much for taking so much time.

1:00:23

was going to be fascinating and

1:00:25

it was. So, thanks a lot.

1:00:28

Thanks a lot to Chris Wyment

1:00:30

again. But before letting you go,

1:00:32

of course, I'm going to ask

1:00:34

you the last questions. I ask

1:00:37

every guest at the end of

1:00:39

the show. One, if you had

1:00:41

unlimited time and resources, which problem

1:00:43

would you try to solve? That's

1:00:46

a big one. One of the

1:00:48

things coming out of COVID that

1:00:50

really struck me is with a

1:00:53

lot of infections, we're actually quite

1:00:55

unambitious, I think. You know, that

1:00:57

we put up with a lot

1:00:59

of disease in day life and

1:01:02

even free COVID, if you look

1:01:04

at a lot of adverse for

1:01:06

medicine when you're ill, it's, you

1:01:08

know, just keep going, keep going.

1:01:11

And I think we saw a

1:01:13

lot of indications during COVID for

1:01:15

some of the sort of technologies

1:01:17

and approaches. it weren't lockdowns but

1:01:20

were actually much proficient and even

1:01:22

just like the work that I've

1:01:24

sort of done on things like

1:01:26

digital tools. So I think we've

1:01:29

got these little signals that we

1:01:31

could be much much better in

1:01:33

how we tackle these problems and

1:01:36

I think that's something that would

1:01:38

require quite a lot of resource

1:01:40

and time to do effectively. But

1:01:42

even like I've got young children

1:01:45

and yeah there's a lot of

1:01:47

kids that kind of hospitalized with

1:01:49

a lot of infections that We

1:01:51

could understand, we could test about,

1:01:54

and I think we rely a

1:01:56

lot on, we wait for a

1:01:58

vaccine to be developed, but I

1:02:00

think I've always just wondered, yeah.

1:02:03

could we do something a bit

1:02:05

clever? Can we go to all

1:02:07

this technology? I mean, there's a

1:02:09

US colleague who, during COVID, said,

1:02:12

you know, this is our Apollo

1:02:14

mission. Can we actually do something

1:02:16

extremely innovative and ambitious, COVID? I

1:02:19

think vaccines were amazing, some of

1:02:21

the treatments coming down, but we

1:02:23

didn't, I think, solve that, that

1:02:25

much earlier, a problem of what

1:02:28

to do around infection. So yeah,

1:02:30

I'd love us to live in

1:02:32

a different world where we can

1:02:34

actually. Yeah, yeah, yeah, definitely share

1:02:37

that passion and objective. And if

1:02:39

you had, if you could have

1:02:41

dinner with a great scientific mind,

1:02:43

dead alive or fictional, who would

1:02:46

it be? So I think recently

1:02:48

for a project, I mean, reading

1:02:50

up a lot about William Gossett

1:02:52

called AK Student. who develops tests

1:02:55

and works at Guinness, but actually

1:02:57

digging more into his work, he

1:02:59

was a really interesting character because

1:03:02

there was a lot of conflict

1:03:04

with Fisher in terms of outlook.

1:03:06

And Fisher was very much from

1:03:08

the kind of academic focus of,

1:03:11

you know, you recruit knowledge and

1:03:13

you want very high confidence in

1:03:15

that knowledge and that's why you

1:03:17

have the sort of thresholds and

1:03:20

experiments of science that he's been

1:03:22

bored. Gossett was much more of

1:03:24

a pragmatist. Yeah, he was working

1:03:26

for a big business and I

1:03:29

mean, there's one situation where he

1:03:31

had a P value of point,

1:03:33

um, point one three and he

1:03:36

said, you know, that's very good

1:03:38

evidence. You know, if it's a

1:03:40

business and it doesn't cost much

1:03:42

and we can explore further, it's

1:03:45

worth going forward. Well, as you

1:03:47

know, Fisher would have had that.

1:03:49

Okay, if it's, if it's not

1:03:51

hitting the five percent, we're not

1:03:54

interested, we're not interested. And we're

1:03:56

not interested. maybe a part of

1:03:58

statistics that got suppressed I think

1:04:00

probably a lot of the 20th

1:04:03

century. I think Fisher and Co.

1:04:05

probably one hour. in many ways

1:04:07

in terms of imposing those those

1:04:09

criteria and in some cases throwing

1:04:12

away a lot of evidence and

1:04:14

I mean although it wasn't so

1:04:16

it was basically this was very

1:04:19

anti-basian but I think that that

1:04:21

making use of limited information the

1:04:23

gossip was you know it was

1:04:25

very adamant in some cases I

1:04:28

think you had budget to get

1:04:30

two data points or something and

1:04:32

it was still that I want

1:04:34

to do something like that. So

1:04:37

I think that would be a

1:04:39

very interesting person to dig into

1:04:41

the strategy. It also just sounds

1:04:43

like he was just a very

1:04:46

nice guy relative to Fisher and

1:04:48

some of the others at the

1:04:50

time. So yeah, I think that

1:04:52

the kind of different outlook of

1:04:55

where you set the bar and

1:04:57

actually, you know, what are you

1:04:59

trying to do with statistics? Are

1:05:02

you trying to kind of get

1:05:04

perfect knowledge or are you actually

1:05:06

just trying to make better decisions?

1:05:08

Wasn't that hard to be a

1:05:11

nicer guy than Fisher apparently? But

1:05:13

yeah, no, I mean, I definitely

1:05:15

resonate with what you were saying

1:05:17

because all my models, most of

1:05:20

my models, I do them not

1:05:22

for academic use, but for companies.

1:05:24

And yeah, like, there is a

1:05:26

lot of things you have to

1:05:29

be able. to prioritize. I think

1:05:31

that's one of the most important

1:05:33

skills to have as a modeler,

1:05:35

especially if you're in a small

1:05:38

team, right? If you're in a

1:05:40

big project with a lot of

1:05:42

modelers, you can explore a lot

1:05:45

of path at the same time.

1:05:47

But if you're in a small

1:05:49

team, then you have to really

1:05:51

be able to set the priorities

1:05:54

and see what path you want

1:05:56

to explore first. That's often a

1:05:58

metaphor used to explain what the

1:06:00

modeling process is. It's like being

1:06:03

lost in a desert and you're

1:06:05

trying to find your way out

1:06:07

in the most efficient way is

1:06:09

usually to explore a lot of

1:06:12

path and see which ones are

1:06:14

successful. There can be several of

1:06:16

them. There can be just one

1:06:18

of them. There can be zero

1:06:21

sometimes. But also exploring a path

1:06:23

that ends up not being successful

1:06:25

is actually very informative. because then

1:06:28

that means the people behind you

1:06:30

won't make the same mistakes and

1:06:32

they won't go down that path.

1:06:34

So to explore these paths you

1:06:37

can do it alone if you

1:06:39

don't have any more people that's

1:06:41

just going to take you more

1:06:43

time but you have to do

1:06:46

it or you can do it

1:06:48

with several people simultaneously. But that's

1:06:50

where the idea of priorities, also

1:06:52

very important because your priorities are

1:06:55

going to dictate which path you

1:06:57

go down first. Yeah,

1:07:00

and I think it's that

1:07:02

that I said that the

1:07:04

priority is in the focus

1:07:06

that I think to kind

1:07:08

of angle which you know

1:07:10

maybe in the kind of

1:07:13

the pursuit perfection We don't

1:07:15

always get that balance right

1:07:17

in it particularly when it's

1:07:19

kind of academic academic research

1:07:21

interacting with very fast problems

1:07:23

Right. Yeah. Yeah. Awesome. Well

1:07:26

Adam Let's call you to

1:07:28

show that was this was

1:07:30

in Absolutely pleasure to have

1:07:32

you here on the show

1:07:34

as usual. I will link

1:07:36

to your website and your

1:07:39

socials and your books in

1:07:41

the show notes for people

1:07:43

who want to dig deeper.

1:07:45

Thank you again Adam for

1:07:47

taking the time and being

1:07:49

on this show. Yeah, that's

1:07:51

not me good shot. Be

1:07:54

sure to rate, review and

1:07:56

follow the show on your

1:07:58

favorite. put catcher and visit

1:08:00

run-based stats.com for more resources

1:08:02

about today's topics as well

1:08:04

as access to more episodes

1:08:07

to help you reach true

1:08:09

patient state of mind. That's

1:08:11

run-based stats.com. Our theme music

1:08:13

is good-basin by Baba Brittman,

1:08:15

fit M.S. and Megaran. Check

1:08:17

out his awesome work at

1:08:20

Baba Brittman.com. I'm your host

1:08:22

Alex Andorra. You can follow

1:08:24

me on Twitter at Alex

1:08:26

Undor andora like the country.

1:08:28

You can support the show.

1:08:30

and unlock exclusive benefits by

1:08:32

visiting patron.com/learn base stats. Thank

1:08:35

you so much for listing

1:08:37

and for your support, you're

1:08:39

truly a good lazy and

1:08:41

change your predictions after taking

1:08:43

information and if you're thinking

1:08:45

I'll be less than amazing.

1:08:48

Let's adjust those expectations. Let

1:08:50

me show you how to

1:08:52

be a good lazy. Change

1:08:54

calculations after taking fresh data.

1:08:56

that your brain is making.

1:08:58

Let's get them on a

1:09:01

solid foundation.