Learning Bayesian Statistics

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 😉

Takeaways

  • Bayesian methods align better with researchers’ intuitive understanding of research questions and provide more tools to evaluate and understand models.
  • Prior sensitivity analysis is crucial for understanding the robustness of findings to changes in priors and helps in contextualizing research findings.
  • Bayesian methods offer an elegant and efficient way to handle missing data in longitudinal studies, providing more flexibility and information for researchers.
  • Fit indices in Bayesian model selection are effective in detecting underfitting but may struggle to detect overfitting, highlighting the need for caution in model complexity.
  • Bayesian methods have the potential to revolutionize educational research by addressing the challenges of small samples, complex nesting structures, and longitudinal data. 
  • Posterior predictive checks are valuable for model evaluation and selection.

Chapters

00:00 The Power and Importance of Priors

09:29 Updating Beliefs and Choosing Reasonable Priors

16:08 Assessing Robustness with Prior Sensitivity Analysis

34:53 Aligning Bayesian Methods with Researchers’ Thinking

37:10 Detecting Overfitting in SEM

43:48 Evaluating Model Fit with Posterior Predictive Checks

47:44 Teaching Bayesian Methods

54:07 Future Developments in Bayesian Statistics

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 and Francesco Madrisotti.

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.

Transcript
Speaker:

Priors represent a crucial part of the

Bayesian workflow, and actually a big

2

00:00:10,854 --> 00:00:14,154

reason for its power and usefulness.

3

00:00:14,154 --> 00:00:15,414

But why is that?

4

00:00:15,414 --> 00:00:18,794

How do you choose reasonable priors in

your models?

5

00:00:18,794 --> 00:00:21,954

What even is a reasonable prior?

6

00:00:22,054 --> 00:00:26,914

These are deep questions that today's

guest, Sonja Winter, will guide us

7

00:00:26,914 --> 00:00:27,694

through.

8

00:00:27,694 --> 00:00:31,654

an assistant professor in the College of

Education and Human Development of the

9

00:00:31,654 --> 00:00:37,114

University of Missouri, Sonia's research

focuses on the development and application

10

00:00:37,114 --> 00:00:41,954

of patient approaches to the analysis of

educational and developmental

11

00:00:41,954 --> 00:00:47,314

psychological data, with a specific

emphasis on the role of priors.

12

00:00:47,314 --> 00:00:48,994

What a coincidence!

13

00:00:48,994 --> 00:00:54,574

In this episode, she shares insights on

the selection of priors, prior sensitivity

14

00:00:54,574 --> 00:00:57,454

analysis, and the challenges of working

15

00:00:57,454 --> 00:00:59,534

with longitudinal data.

16

00:00:59,534 --> 00:01:05,514

She also explores the implications of

Bayesian methods for model selection and

17

00:01:05,514 --> 00:01:10,834

fit indices in structural equation

modeling, as well as the challenges of

18

00:01:10,834 --> 00:01:13,374

detecting overfitting in models.

19

00:01:13,374 --> 00:01:19,054

When she's not working, you'll find Sonja

baking delicious treats, gardening, or

20

00:01:19,054 --> 00:01:21,274

watching beautiful birds.

21

00:01:21,274 --> 00:01:24,526

This is Learning Bayesian Statistics,

episode.

22

00:01:24,526 --> 00:01:50,746

Welcome to Learning Bayesian Statistics, a

podcast about Bayesian inference, the

23

00:01:50,746 --> 00:01:54,382

methods, the projects, and the people who

make it possible.

24

00:01:54,382 --> 00:01:56,582

I'm your host, Alex Andorra.

25

00:01:56,582 --> 00:02:00,922

You can follow me on Twitter at alex

.andorra, like the country.

26

00:02:00,922 --> 00:02:05,222

For any info about the show, learnbasedats

.com is Laplace to be.

27

00:02:05,222 --> 00:02:09,722

Show notes, becoming a corporate sponsor,

unlocking Bayesian Merge, supporting the

28

00:02:09,722 --> 00:02:12,352

show on Patreon, everything is in there.

29

00:02:12,352 --> 00:02:14,232

That's learnbasedats .com.

30

00:02:14,232 --> 00:02:18,702

If you're interested in one -on -one

mentorship, online courses, or statistical

31

00:02:18,702 --> 00:02:21,646

consulting, feel free to reach out and

book a call.

32

00:02:21,646 --> 00:02:27,646

at topmate .io slash alex underscore and

dora see you around folks and best patient

33

00:02:27,646 --> 00:02:29,190

wishes to you all.

34

00:02:32,494 --> 00:02:39,254

Hello my dear patients, today I want to

thank the fantastic Jonathan Morgan and

35

00:02:39,254 --> 00:02:43,794

Francesco Madrisotti for supporting the

show on Patreon.

36

00:02:43,794 --> 00:02:49,358

Your support is invaluable and literally

makes this show possible.

37

00:02:49,358 --> 00:02:52,698

can't wait to talk with you guys in the

Slack channel.

38

00:02:52,698 --> 00:02:57,858

Second, with my friends and fellow PymC

core developers, Ravin Kumar and Tommy

39

00:02:57,858 --> 00:03:03,138

Capretto, we've just released our new

online course, Advanced Regression with

40

00:03:03,138 --> 00:03:04,598

Bambi and PymC.

41

00:03:04,598 --> 00:03:10,078

And honestly, after two years of

development, it feels really great to get

42

00:03:10,078 --> 00:03:16,418

these out into the world, not only because

it was, well, long and intense, but mainly

43

00:03:16,418 --> 00:03:19,312

because I am so proud of the level of

44

00:03:19,312 --> 00:03:24,012

of details, teachings, and exercises that

we've packed into this one.

45

00:03:24,012 --> 00:03:29,412

It's basically the course I wish I had

once I had gone through the beginner's

46

00:03:29,412 --> 00:03:33,262

phase when learning patience tests, that

moment when you're like...

47

00:03:33,262 --> 00:03:38,512

Okay, I know how to do basic models, but

where do I go from here?

48

00:03:38,512 --> 00:03:42,982

I remember feeling quite lost, so we

wanted to give you a one -stop shop for

49

00:03:42,982 --> 00:03:48,742

such intermediate models with the most

content possible, as evergreen as it gets.

50

00:03:48,742 --> 00:03:53,942

If that sounds interesting, go to

intuitivebase .com and check out the full

51

00:03:53,942 --> 00:03:54,782

syllabus.

52

00:03:54,782 --> 00:03:58,926

We're enrolling the first cohort as we

speak!

53

00:03:58,926 --> 00:04:03,286

Of course, you get a 10 % discount if

you're a patron of the show.

54

00:04:03,286 --> 00:04:07,786

Go to the Patreon page or the Slack

channel to get the code.

55

00:04:07,786 --> 00:04:13,366

Okay, back to the show now and looking

forward to seeing you in the intuitive

56

00:04:13,366 --> 00:04:14,726

base discourse.

57

00:04:18,350 --> 00:04:22,530

Sonia Winter, welcome to Learning Bayesian

Statistics.

58

00:04:22,890 --> 00:04:23,590

Thank you.

59

00:04:23,590 --> 00:04:24,660

Thanks for having me.

60

00:04:24,660 --> 00:04:27,570

I'm really excited to talk to you today.

61

00:04:27,570 --> 00:04:28,360

Same, same.

62

00:04:28,360 --> 00:04:29,630

That's a treat.

63

00:04:29,710 --> 00:04:31,330

I have a lot of questions.

64

00:04:31,330 --> 00:04:32,526

I really love.

65

00:04:32,526 --> 00:04:35,396

like everything you're doing in your

research.

66

00:04:35,396 --> 00:04:39,286

We're going to talk a lot about priors

today, folks.

67

00:04:39,646 --> 00:04:42,586

So yeah, like get ready.

68

00:04:42,586 --> 00:04:49,186

But first, can you provide a brief

overview of your research interests and

69

00:04:49,186 --> 00:04:52,646

how patient methods play a role in your

work?

70

00:04:52,966 --> 00:04:54,326

Yeah, sure.

71

00:04:54,346 --> 00:04:58,766

So my background is actually in

developmental psychology.

72

00:04:58,766 --> 00:05:02,526

I did a bachelor or master's degree at

Utrecht University.

73

00:05:02,542 --> 00:05:07,282

And during that time, I really realized

that a lot of work needed to be done on

74

00:05:07,282 --> 00:05:09,982

the analysis part of social science

research.

75

00:05:09,982 --> 00:05:14,482

And so I switched and got really into

structural equation models, which are

76

00:05:14,482 --> 00:05:19,642

these big multivariate models that include

latent variables.

77

00:05:19,642 --> 00:05:22,102

I'm sure we'll talk more about that later.

78

00:05:22,942 --> 00:05:26,882

But those models can be hard to estimate

and there are all these issues.

79

00:05:26,882 --> 00:05:30,628

And so I was introduced to Bayesian

statistics.

80

00:05:30,670 --> 00:05:35,670

right after my master's degree when I was

working with Rens van der Schoot, also at

81

00:05:35,670 --> 00:05:37,230

Utrecht University.

82

00:05:37,650 --> 00:05:42,650

And he asked me to do this big literature

review about it with him.

83

00:05:42,650 --> 00:05:44,060

And that really introduced me.

84

00:05:44,060 --> 00:05:49,290

And so now I focus a lot on Bayesian

estimation and how it can help us estimate

85

00:05:49,290 --> 00:05:50,850

these structural equation models.

86

00:05:50,850 --> 00:05:55,030

And then specifically more recently, I've

really been focusing on how those priors

87

00:05:55,030 --> 00:05:57,358

can really help us.

88

00:05:57,358 --> 00:06:02,118

both with estimation and also just with

understanding our models a little bit

89

00:06:02,118 --> 00:06:03,238

better.

90

00:06:03,778 --> 00:06:06,952

So yeah, I'm really excited about all of

that.

91

00:06:09,870 --> 00:06:13,490

Yeah, I can guess that sounds awesome.

92

00:06:14,490 --> 00:06:19,030

So structural equation modeling, we

already talked about it on the show.

93

00:06:19,030 --> 00:06:24,030

So today we're going to focus a bit more

on priors and how that fits into the SEM

94

00:06:24,030 --> 00:06:25,030

framework.

95

00:06:25,030 --> 00:06:30,530

So for people who don't know about SEM, I

definitely recommend episode 102 with Ed

96

00:06:30,530 --> 00:06:31,210

Merkel.

97

00:06:31,210 --> 00:06:36,710

And we talked exactly about structural

equation modeling and causal inference in

98

00:06:36,710 --> 00:06:37,550

psychometrics.

99

00:06:37,550 --> 00:06:39,022

So that will be a.

100

00:06:39,022 --> 00:06:48,022

a very good introduction i think to these

topics for people and what i'm curious

101

00:06:48,022 --> 00:06:56,442

about sonia is you work a lot on priors

and things like that but how how did you

102

00:06:56,442 --> 00:07:00,262

end up working on that was something that

you were always curious about or that

103

00:07:00,262 --> 00:07:06,342

something that appeared later later on in

your in your phd studies

104

00:07:08,206 --> 00:07:12,906

I would say definitely something that

started or piqued my interest a little bit

105

00:07:12,906 --> 00:07:14,086

later.

106

00:07:14,146 --> 00:07:19,606

I think so after I first got familiarized

with Bayesian methods, I was excited

107

00:07:19,606 --> 00:07:25,546

mostly by how it could help, like priors

could help us estimate, like avoid

108

00:07:25,546 --> 00:07:27,746

negative variances and those types of

things.

109

00:07:27,746 --> 00:07:33,906

But I saw them more as a pragmatic tool to

help with that.

110

00:07:33,906 --> 00:07:36,270

And I didn't really focus so much on that.

111

00:07:36,270 --> 00:07:40,870

I feel like I also was a little bit afraid

at the time of, you know, those

112

00:07:40,870 --> 00:07:45,330

researchers who talk a lot about, well, we

shouldn't really make our priors

113

00:07:45,330 --> 00:07:48,230

informative because that's subjective and

that's bad.

114

00:07:48,230 --> 00:07:53,930

And so I really typically use like

uninformative priors or like software

115

00:07:53,930 --> 00:07:56,590

defaults for a lot of my work in the

beginning.

116

00:07:56,590 --> 00:08:01,440

But then during my PhD studies, I

actually.

117

00:08:01,486 --> 00:08:06,426

Well, first of all, I worked with another

researcher, Sanaa Smith, who was also a

118

00:08:06,426 --> 00:08:08,086

PhD student at the time.

119

00:08:08,086 --> 00:08:13,086

And she was really intrigued by something

she found that these software defaults can

120

00:08:13,086 --> 00:08:18,266

really cause issues when you're,

especially when your data is like very

121

00:08:18,266 --> 00:08:22,106

small, it can, it can make your results

look wild.

122

00:08:22,106 --> 00:08:26,486

And so we worked on this paper together

and created a shiny app to demonstrate all

123

00:08:26,486 --> 00:08:26,846

of that.

124

00:08:26,846 --> 00:08:31,083

And that made me realize that maybe

uninformative priors.

125

00:08:31,181 --> 00:08:33,281

are not always the best way to go.

126

00:08:33,281 --> 00:08:38,281

And also a prior that looks informative in

one scenario might be relatively

127

00:08:38,281 --> 00:08:39,601

uninformative in another.

128

00:08:39,601 --> 00:08:46,761

And so I really started shifting my, my

perspective on priors and focusing more on

129

00:08:46,761 --> 00:08:52,661

how ignoring them is kind of like ignoring

the best part of Bayesian in my opinion,

130

00:08:52,661 --> 00:08:54,021

at this point.

131

00:08:54,221 --> 00:08:59,441

and so now I really want to look at how,

how they can help us and how we can be

132

00:08:59,441 --> 00:09:00,615

thoughtful.

133

00:09:00,781 --> 00:09:05,381

We don't want to drive our science by

priors, right?

134

00:09:05,381 --> 00:09:10,341

We want to learn something new from our

data, but we find that balance is really

135

00:09:10,341 --> 00:09:12,421

what I'm looking for now.

136

00:09:13,341 --> 00:09:19,621

Yeah, well, what a fantastic application

of updating your belief, right?

137

00:09:19,621 --> 00:09:25,121

From a meta standpoint, you just like

updated your priors pretty aggressively

138

00:09:25,121 --> 00:09:27,551

and also very rationally.

139

00:09:27,551 --> 00:09:28,731

That's really impressive.

140

00:09:28,731 --> 00:09:29,921

Well done.

141

00:09:29,933 --> 00:09:32,013

Because that's hard to do also.

142

00:09:32,013 --> 00:09:35,153

It's not something we like to do.

143

00:09:35,393 --> 00:09:37,313

So that's great.

144

00:09:37,313 --> 00:09:39,973

Well done on doing that.

145

00:09:39,973 --> 00:09:48,373

And actually now that you're on the other

side, how do you approach the selection of

146

00:09:48,373 --> 00:09:55,393

priors in your research and what advice do

you have for people new to Bayesian

147

00:09:55,393 --> 00:09:56,453

methods?

148

00:09:57,253 --> 00:09:59,015

Yeah, great question.

149

00:09:59,181 --> 00:10:07,121

I think at least within structural

equation modeling, we as like applied

150

00:10:07,121 --> 00:10:13,641

researchers are helped somewhat because

distributions, at least for priors, are

151

00:10:13,641 --> 00:10:15,801

sort of clear.

152

00:10:15,801 --> 00:10:18,841

Like you don't have to think too much

about them.

153

00:10:18,841 --> 00:10:24,141

And so you can immediately jump into

thinking about, okay, what level of

154

00:10:24,141 --> 00:10:27,297

information do I want to convey in those

priors?

155

00:10:27,309 --> 00:10:32,069

And I think whenever I'm working with

applied researchers, I try to strike a

156

00:10:32,069 --> 00:10:37,429

balance with them because I know they are

not typically comfortable using like super

157

00:10:37,429 --> 00:10:39,909

informative priors that are really narrow.

158

00:10:39,909 --> 00:10:43,699

And so I just asked them to think about,

well, what would be a reasonable range?

159

00:10:43,699 --> 00:10:49,469

Like if we are estimating a linear

regression parameter, what would that

160

00:10:49,469 --> 00:10:50,869

effect size look like?

161

00:10:50,869 --> 00:10:51,139

Right.

162

00:10:51,139 --> 00:10:55,349

It might be zero or it might be two, but

it's probably not going to be 20.

163

00:10:55,349 --> 00:10:56,653

And so we can.

164

00:10:56,653 --> 00:11:02,273

sort of shape our prior to align with

those sort of expectations about how

165

00:11:02,273 --> 00:11:04,353

probable certain values are versus others.

166

00:11:04,353 --> 00:11:10,413

It's a really, I don't know, interactive

process between me and the researcher to

167

00:11:10,413 --> 00:11:14,193

get this right, especially for those types

of parameters that they are really

168

00:11:14,193 --> 00:11:15,973

interested in.

169

00:11:15,973 --> 00:11:21,073

I think another type of parameter that is

more challenging for applied researchers

170

00:11:21,073 --> 00:11:24,793

are those that are placed on residual

variances, for example.

171

00:11:24,793 --> 00:11:26,157

Like people typically don't...

172

00:11:26,157 --> 00:11:30,597

think about the part of the outcome that

they can't explain that much.

173

00:11:30,597 --> 00:11:39,397

And so that's where I do rely a bit more

on sort of, I don't know, industry

174

00:11:39,397 --> 00:11:43,597

standard choices that are typically not

super informative.

175

00:11:43,597 --> 00:11:48,617

But then once we pick our like target

priors, I always advise the researcher to

176

00:11:48,617 --> 00:11:51,841

follow it up with a sensitivity analysis

to see.

177

00:11:51,917 --> 00:11:55,537

like how robust their findings are to

changes in those priorities, either making

178

00:11:55,537 --> 00:11:58,177

them more informative or less informative.

179

00:11:58,677 --> 00:12:01,457

And so yeah, that's really the approach I

take.

180

00:12:01,457 --> 00:12:06,317

Of course, if someone wants to go full

base and full informative and they have

181

00:12:06,317 --> 00:12:12,117

this, this wealth of previous research to

draw from, then I'm all for going, going

182

00:12:12,117 --> 00:12:13,317

that route as well.

183

00:12:13,317 --> 00:12:15,397

It's just not as common.

184

00:12:15,797 --> 00:12:16,257

Hmm.

185

00:12:16,257 --> 00:12:16,757

Hmm.

186

00:12:16,757 --> 00:12:18,117

Yeah, I see.

187

00:12:18,777 --> 00:12:21,677

in what, what are the...

188

00:12:21,677 --> 00:12:28,137

main difficulties you see from people that

you advise like that?

189

00:12:28,157 --> 00:12:35,937

Where do you see them having more defenses

up or just more difficulties because they

190

00:12:35,937 --> 00:12:41,117

have a hard time wrapping their head

around a specific concept?

191

00:12:42,357 --> 00:12:47,757

I think just all over, I think if anyone

has ever tried to do like a power analysis

192

00:12:47,757 --> 00:12:51,837

working with researchers, it's sort of a

similar concept because

193

00:12:51,917 --> 00:12:58,177

It is not, at least in my field or the

people I work with are not very typically

194

00:12:58,196 --> 00:13:02,957

already thinking about the exact parameter

estimates that they are expecting to see,

195

00:13:02,957 --> 00:13:03,357

right?

196

00:13:03,357 --> 00:13:06,457

They are just, they just go with the

hypothesis.

197

00:13:06,457 --> 00:13:11,037

I think these two things are correlated

and they might not even go as far as to

198

00:13:11,037 --> 00:13:12,777

think, is it positive or negative?

199

00:13:12,777 --> 00:13:19,237

So then once you ask them those questions,

it really forces them to go much deeper on

200

00:13:19,237 --> 00:13:22,245

their theory and really consider like.

201

00:13:22,861 --> 00:13:24,791

What is, what am I expecting?

202

00:13:24,791 --> 00:13:29,741

What is reasonable based on what I know

from, from previous studies or just

203

00:13:29,741 --> 00:13:31,081

experience.

204

00:13:31,081 --> 00:13:35,301

And that can be kind of challenging.

205

00:13:35,781 --> 00:13:40,541

It's, it's kind of, I think sometimes the

researchers might feel like I'm

206

00:13:40,541 --> 00:13:44,871

criticizing them for not knowing, but I

think that's perfectly normal to not know.

207

00:13:44,871 --> 00:13:47,741

Like we already have so many other things

to think about.

208

00:13:47,741 --> 00:13:50,029

But it definitely.

209

00:13:50,221 --> 00:13:51,391

is kind of a hurdle.

210

00:13:51,391 --> 00:13:56,981

Also the time commitment, I think, to

really consider the priors, especially if

211

00:13:56,981 --> 00:14:00,181

you're coming from a frequentist realm

where you just say, okay, maximum

212

00:14:00,181 --> 00:14:01,401

likelihood go.

213

00:14:01,401 --> 00:14:06,001

Not only do you not have to think about

the estimation, but then also your results

214

00:14:06,001 --> 00:14:07,461

are almost instant.

215

00:14:07,461 --> 00:14:10,981

And so that's always kind of a challenge

as well.

216

00:14:14,221 --> 00:14:15,181

I see.

217

00:14:15,181 --> 00:14:15,381

Yeah.

218

00:14:15,381 --> 00:14:15,621

Yeah.

219

00:14:15,621 --> 00:14:19,441

Definitely something also I seen, I seen

beginners.

220

00:14:19,921 --> 00:14:23,681

yeah, it, it really depends on also where

they are coming from, as you were saying.

221

00:14:23,681 --> 00:14:24,131

Yeah.

222

00:14:24,131 --> 00:14:24,991

I did.

223

00:14:24,991 --> 00:14:27,481

Your advice will depend a lot on that.

224

00:14:27,881 --> 00:14:28,641

yeah.

225

00:14:28,641 --> 00:14:29,441

Yeah.

226

00:14:29,441 --> 00:14:34,281

And actually you work also a lot on prior

sensitivity analysis.

227

00:14:34,841 --> 00:14:38,241

can you, can you tell people what that is?

228

00:14:38,241 --> 00:14:43,373

And the importance of it in your, in your

modeling workflow and.

229

00:14:43,373 --> 00:14:46,773

how you incorporate it into your research.

230

00:14:47,393 --> 00:14:48,253

Yeah.

231

00:14:48,253 --> 00:14:54,573

So a sensitivity analysis for priors is

something that you typically do after you

232

00:14:54,573 --> 00:14:56,293

run your main analysis.

233

00:14:56,293 --> 00:15:02,553

So you come up with your target set of

priors for all your parameters, estimate

234

00:15:02,553 --> 00:15:05,533

the model, look at the results, look at

the posteriors.

235

00:15:05,533 --> 00:15:10,553

And then in the next step, you think

about, well, how can I change these

236

00:15:10,553 --> 00:15:11,661

priors?

237

00:15:11,661 --> 00:15:17,761

in sort of meaningful ways, either making

them more informative, perhaps making them

238

00:15:17,761 --> 00:15:21,931

represent some other theory, making them

less informative as well.

239

00:15:21,931 --> 00:15:26,121

So making the influence of the prior

weaker in your results.

240

00:15:26,221 --> 00:15:31,721

And then you rerun your analysis for all

of those different prior scenarios, and

241

00:15:31,721 --> 00:15:35,361

then compare those results to the ones

that you actually obtained with your

242

00:15:35,361 --> 00:15:38,081

target analysis and your target priors.

243

00:15:38,081 --> 00:15:40,509

And the idea here is to see,

244

00:15:40,973 --> 00:15:46,293

how much your results actually depend on

those prior beliefs that you came into the

245

00:15:46,293 --> 00:15:47,413

analysis with.

246

00:15:47,413 --> 00:15:52,233

If you don't find any differences, then

you can say, well, my results are mostly

247

00:15:52,233 --> 00:15:57,993

influenced by my data, by the new evidence

that I obtained.

248

00:15:57,993 --> 00:16:01,253

They are robust to changes in prior

beliefs, right?

249

00:16:01,253 --> 00:16:04,393

It doesn't really matter what beliefs you

came into the analysis with.

250

00:16:04,393 --> 00:16:07,913

The results are going to be the same,

which is great.

251

00:16:08,141 --> 00:16:11,981

In other cases, you might find that your

results do change meaningfully.

252

00:16:11,981 --> 00:16:15,841

So for example, in effect that was

significant with your priors is no longer

253

00:16:15,841 --> 00:16:23,061

significant using a frequentist term here,

but hopefully people will understand once

254

00:16:23,061 --> 00:16:24,551

you change your priors.

255

00:16:24,551 --> 00:16:28,821

And that's, of course, is a little bit

more difficult to handle because what do

256

00:16:28,821 --> 00:16:30,061

you do?

257

00:16:30,101 --> 00:16:32,877

I want to say that the goal is not to

258

00:16:32,877 --> 00:16:36,877

use the sensitivity analysis to then go

back and change your priors and run the

259

00:16:36,877 --> 00:16:38,867

analysis again and report that in your

paper.

260

00:16:38,867 --> 00:16:42,537

That would be sort of akin to p -hacking.

261

00:16:43,377 --> 00:16:46,607

Instead, I think it just contextualizes

your findings.

262

00:16:46,607 --> 00:16:52,217

It's showing that the knowledge you came

into the analysis with is partially

263

00:16:52,217 --> 00:16:53,327

driving your results.

264

00:16:53,327 --> 00:16:58,413

And that probably means that the evidence

in your new data is not super strong.

265

00:16:58,413 --> 00:17:03,352

And so it may indicate some issues with

your theory or some issues with your data.

266

00:17:03,352 --> 00:17:10,153

And you have to collect more data to

figure out which of those it is basically.

267

00:17:10,853 --> 00:17:16,213

And so it's, it's kind of helping you also

figure out the next steps in your

268

00:17:16,213 --> 00:17:18,953

research, I feel, which is helpful.

269

00:17:18,953 --> 00:17:27,113

But it can be frustrating, of course, and

harder to convince maybe co -authors and

270

00:17:27,113 --> 00:17:28,173

reviewers to.

271

00:17:28,173 --> 00:17:30,633

move forward with a paper like that.

272

00:17:30,633 --> 00:17:35,329

But to me it is very interesting these

results from sensitivity analyses.

273

00:17:36,877 --> 00:17:39,857

Yeah, yeah, completely agree in that.

274

00:17:39,857 --> 00:17:46,417

That's very interesting to see the, yeah,

if the results differ on the priors, and

275

00:17:46,417 --> 00:17:51,457

that can also help, you know, settle any

argument on the choice of prior.

276

00:17:51,457 --> 00:17:56,857

You know, if people are really in

disagreement about which priors to choose,

277

00:17:56,857 --> 00:18:01,857

well, then you can run the model with both

sets of priors, and if the results don't

278

00:18:01,857 --> 00:18:04,737

change, it's like, well, let's stop

arguing.

279

00:18:04,737 --> 00:18:05,805

That's kind of...

280

00:18:05,805 --> 00:18:06,685

It's kind of silly.

281

00:18:06,685 --> 00:18:08,585

We just lost time.

282

00:18:08,585 --> 00:18:11,105

So let's just focus on the results then.

283

00:18:11,105 --> 00:18:14,565

I think it's a very interesting framework.

284

00:18:16,265 --> 00:18:18,265

And then there is another.

285

00:18:18,265 --> 00:18:23,945

So that is like that entails running the

model, running MCMC on the model.

286

00:18:23,945 --> 00:18:28,745

But there are some checks that you do

before that to ensure the robustness of

287

00:18:28,745 --> 00:18:30,065

your patient models.

288

00:18:30,065 --> 00:18:32,653

And one of that step is.

289

00:18:32,653 --> 00:18:36,093

very crucial and called primary predictive

checks.

290

00:18:36,353 --> 00:18:39,413

Can you talk about that to beat Sonja?

291

00:18:39,973 --> 00:18:45,593

Yeah, so as you said, these checks happen

before you do any actual analysis.

292

00:18:45,593 --> 00:18:48,933

So you can do them before you collect any

data.

293

00:18:49,053 --> 00:18:54,533

In fact, one reason for using them is to

figure out whether the priors you came up

294

00:18:54,533 --> 00:19:00,877

with results in sensible ranges of

possible parameter estimates, right?

295

00:19:00,877 --> 00:19:05,377

In some cases, especially with these

complex multivariate models, your priors

296

00:19:05,377 --> 00:19:12,417

may interact in unexpected ways and then

result in predictions that are not in line

297

00:19:12,417 --> 00:19:17,157

with what your theory is actually telling

you you should expect.

298

00:19:17,297 --> 00:19:23,477

And so prior predictive checks basically

until you specify your priors for all your

299

00:19:23,477 --> 00:19:24,357

parameters.

300

00:19:24,357 --> 00:19:27,097

And then you generate

301

00:19:27,213 --> 00:19:31,813

parameter values from those priors by

combining it with your model

302

00:19:31,813 --> 00:19:33,313

specification.

303

00:19:33,753 --> 00:19:38,553

And then those combinations of parameter

estimates are used to generate what are

304

00:19:38,553 --> 00:19:40,553

called prior predictive samples.

305

00:19:40,553 --> 00:19:47,733

So these are samples of some pre

-specified size that represent possible

306

00:19:47,733 --> 00:19:52,893

observations that align with what your

priors are conveying combined with your

307

00:19:52,893 --> 00:19:53,793

model.

308

00:19:53,793 --> 00:19:56,045

And so ideally,

309

00:19:56,045 --> 00:20:02,205

those prior predictive samples look kind

of like what you would expect your data to

310

00:20:02,205 --> 00:20:03,125

look like.

311

00:20:03,125 --> 00:20:07,705

And sometimes for researchers, it is

easier to think about what the data should

312

00:20:07,705 --> 00:20:10,605

look like compared to what the parameter

estimates can be.

313

00:20:10,605 --> 00:20:16,785

And so in that sense, prior predictive

checks can be really helpful in checking

314

00:20:16,785 --> 00:20:20,925

not just the priors, but checking the

researcher and making sure that they

315

00:20:20,925 --> 00:20:24,965

actually convey their knowledge to me, for

example, correctly.

316

00:20:25,549 --> 00:20:27,749

Yeah, did that answer your question?

317

00:20:27,749 --> 00:20:33,709

Yeah, yeah, I think that's a great

definition and definitely encourage any

318

00:20:33,709 --> 00:20:39,049

Bayesian practitioner to include prior

predictive checks in their workflow.

319

00:20:39,049 --> 00:20:42,729

Once you have written a model, that should

be the first thing you do.

320

00:20:42,729 --> 00:20:46,369

Do not run a CMC before doing prior

predictive checks.

321

00:20:46,369 --> 00:20:50,809

And recently, I feel like a lot of the

software packages for Bayesian methods

322

00:20:50,809 --> 00:20:51,437

have...

323

00:20:51,437 --> 00:20:56,537

included very simple ways of running these

checks, which when I first started looking

324

00:20:56,537 --> 00:21:00,427

at them, it was kind of more of a niche

step in the workflow.

325

00:21:00,427 --> 00:21:04,737

And so it required a few more steps and

some more like coding, but now it's as

326

00:21:04,737 --> 00:21:09,957

easy as just switching like a toggle to

get those prior predictive samples.

327

00:21:09,957 --> 00:21:11,597

So that's great.

328

00:21:12,617 --> 00:21:14,177

Yeah, yeah, yeah, completely agree.

329

00:21:14,177 --> 00:21:19,017

That's also, yeah, it's definitely

something that's, that's been more and

330

00:21:19,017 --> 00:21:21,761

more popular in the different

331

00:21:21,869 --> 00:21:28,469

classes and courses that I teach, whether

it's online courses or live workshops,

332

00:21:28,609 --> 00:21:32,549

always show prior predictive checks almost

all the time.

333

00:21:32,869 --> 00:21:38,529

So yeah, it's becoming way, way more

popular and widespread.

334

00:21:38,529 --> 00:21:43,149

So that's really good because I can tell

you when I work on a real model for

335

00:21:43,149 --> 00:21:46,969

clients, always the first thing I do

before running MCMC is prior predictive

336

00:21:46,969 --> 00:21:47,349

checks.

337

00:21:47,349 --> 00:21:50,573

And actually there is a fantastic way

of...

338

00:21:50,573 --> 00:21:55,633

you know, doing prior predictive checks,

like kind of industrialized and that's

339

00:21:55,633 --> 00:21:58,753

called simulation based calibration.

340

00:21:58,753 --> 00:22:01,033

Have you heard of that?

341

00:22:01,153 --> 00:22:06,813

No, I mean, maybe the term, but I have no

idea what it is.

342

00:22:06,813 --> 00:22:14,573

So that's just like making prior

predictive checks on an industrialized

343

00:22:14,573 --> 00:22:15,853

scale.

344

00:22:16,333 --> 00:22:19,853

Basically now instead of just

345

00:22:19,853 --> 00:22:25,713

running through the model forward, as you

explained, and generate prior predictive

346

00:22:25,713 --> 00:22:36,573

samples, what you're doing with SPC, so

simulation -based calibration, is

347

00:22:36,573 --> 00:22:41,593

generating not only prior predictive

samples, but prior samples of the

348

00:22:41,593 --> 00:22:43,173

parameters of the model.

349

00:22:43,173 --> 00:22:49,053

You stock these parameters in some object.

350

00:22:49,549 --> 00:22:54,729

but you don't give them to the model, but

you keep them somewhere safe.

351

00:22:55,769 --> 00:23:03,229

And then the prior predictive samples, so

the plausible observations generated by

352

00:23:03,229 --> 00:23:10,929

the model based on the prior samples that

you just kept in the fridge, these prior

353

00:23:10,929 --> 00:23:14,809

predictive samples, now you're going to

consider them as data.

354

00:23:14,809 --> 00:23:19,437

And you're going to tell the model, well,

run MCMC on these data.

355

00:23:19,437 --> 00:23:23,377

as if we had observed these prior

predictive samples in the wild, because

356

00:23:23,377 --> 00:23:25,387

that's what prior predictive samples are.

357

00:23:25,387 --> 00:23:31,637

It's possible samples we could observe

before we know anything about real data.

358

00:23:31,637 --> 00:23:33,027

So you feed that to the model.

359

00:23:33,027 --> 00:23:35,017

You make the model run MCMC on that.

360

00:23:35,017 --> 00:23:36,877

So that means backward inference.

361

00:23:36,877 --> 00:23:41,997

So now the model is going to find out

about the plausible parameter values which

362

00:23:41,997 --> 00:23:43,997

could have generated this data.

363

00:23:43,997 --> 00:23:49,165

And then what you're going to do is

compare the posterior

364

00:23:49,165 --> 00:23:54,765

distribution that the model inferred for

the parameter values to the true parameter

365

00:23:54,765 --> 00:23:58,765

values that you kept in the fridge before.

366

00:23:58,765 --> 00:24:01,245

You're going to get, so these parameter

values are true.

367

00:24:01,245 --> 00:24:05,885

So you just have one of them, because it's

just one sample from the prior parameters.

368

00:24:05,885 --> 00:24:11,665

And you're going to compare these value,

these value to the distribution of

369

00:24:11,665 --> 00:24:15,785

posterior parameters that you just got

from the model.

370

00:24:16,105 --> 00:24:17,947

And based on that,

371

00:24:18,061 --> 00:24:26,641

and how far the model is from the true

parameter, you can find out if your model

372

00:24:26,641 --> 00:24:30,101

is biased or if it's well calibrated.

373

00:24:31,201 --> 00:24:42,781

And that's a really great way to be much

more certain that the model is able to

374

00:24:42,781 --> 00:24:45,581

recover what you want it to recover.

375

00:24:47,949 --> 00:24:52,229

basically playing God, and then you're

trying to see if the model is able to

376

00:24:52,229 --> 00:24:55,099

recover the parameters that you use to

generate the data.

377

00:24:55,099 --> 00:25:01,389

And not only will you do that once, but

you want to do that many times, many, many

378

00:25:01,389 --> 00:25:07,249

times, because, well, the more you do it,

then you enter a kind of a frequentist

379

00:25:07,249 --> 00:25:07,919

realm, right?

380

00:25:07,919 --> 00:25:10,589

Where you're like, you just repeat the

experiments a lot.

381

00:25:10,589 --> 00:25:14,669

And then that's how you're going to see

how calibrated the model is, because then

382

00:25:14,669 --> 00:25:17,057

you can do some calibration plots.

383

00:25:17,133 --> 00:25:22,273

there are a lot of metrics around that

it's a kind of a developing area of the

384

00:25:22,273 --> 00:25:25,813

research but there are a lot of metrics

and one of them is basically just plotting

385

00:25:25,813 --> 00:25:33,973

the true parameter values and well for

instance the mean posterior value from the

386

00:25:33,973 --> 00:25:40,413

parameter and then if this mean is most of

the time along the the line x equals y

387

00:25:40,413 --> 00:25:44,573

well that means you are in pretty good

shape you are but I mean it's the mean

388

00:25:44,573 --> 00:25:45,101

here you

389

00:25:45,101 --> 00:25:48,561

So you have to look at the whole

distribution, but that's to give you an

390

00:25:48,561 --> 00:25:49,301

idea.

391

00:25:49,301 --> 00:25:53,141

And so the bottleneck is you want to do

that a lot of time.

392

00:25:53,141 --> 00:25:55,291

So you have to run MCMC a lot of times.

393

00:25:55,291 --> 00:25:59,111

Most of the time, if you're just doing a

regression, that should be okay.

394

00:25:59,111 --> 00:26:03,721

But sometimes it's going to take a lot of

time to run MCMC and it can be hard.

395

00:26:03,721 --> 00:26:12,481

In these cases, you have new algorithms

that can be efficient because there is one

396

00:26:12,481 --> 00:26:13,165

called

397

00:26:13,165 --> 00:26:17,025

amortized Bayesian inference, a method

called amortized Bayesian inference.

398

00:26:17,025 --> 00:26:21,205

We just covered that in episode 107 with

Marvin Schmidt.

399

00:26:22,005 --> 00:26:27,605

And basically that's exactly a use case

for amortized Bayesian inference because

400

00:26:27,605 --> 00:26:33,485

the model doesn't change, but the data

changes in each iteration of the loop.

401

00:26:33,545 --> 00:26:39,725

And so what amortized Bayesian inference

is doing is just, well, just is training a

402

00:26:39,725 --> 00:26:42,669

deep neural network on the model.

403

00:26:42,669 --> 00:26:44,289

as a first step.

404

00:26:44,529 --> 00:26:49,229

And then the second step is the inference,

but the inference is just instantaneous

405

00:26:49,229 --> 00:26:52,559

because you've trained the deep neural

network.

406

00:26:52,559 --> 00:26:58,569

And that means you can do, you can get as

black, almost as many poster samples as

407

00:26:58,569 --> 00:26:58,908

you want.

408

00:26:58,908 --> 00:27:01,209

Once you have trained the deep neural

network.

409

00:27:01,209 --> 00:27:03,869

And so that's why it's almost all

inspection inference.

410

00:27:03,869 --> 00:27:10,549

And that's a perfect use case for SBC

because then like you can just like you

411

00:27:10,549 --> 00:27:12,529

get new, a new,

412

00:27:12,589 --> 00:27:14,289

new samples for free.

413

00:27:14,569 --> 00:27:17,449

And actually, so I definitely encourage

people to look at that.

414

00:27:17,449 --> 00:27:18,949

It's still developing.

415

00:27:19,208 --> 00:27:24,989

So right now you cannot, for instance, use

Baseflow, which is the Python package that

416

00:27:24,989 --> 00:27:32,819

Marvin talked about in 1 .07 with PIMC,

but it's something we're working on.

417

00:27:32,819 --> 00:27:35,989

And the goal is that it's completely

compatible.

418

00:27:36,809 --> 00:27:42,445

But yeah, like I'll link to the tutorial

notebook in the show notes for people.

419

00:27:42,445 --> 00:27:48,945

who want to get an idea of what SPC is

because even though you're not applying it

420

00:27:48,945 --> 00:27:53,065

right now at least you have that in mind

and you know what that means and you can

421

00:27:53,065 --> 00:27:56,005

work your way to out that.

422

00:27:56,005 --> 00:27:59,245

Yeah that's amazing.

423

00:27:59,245 --> 00:28:05,345

I feel like one of the biggest hurdles in

the structural equation modeling approach

424

00:28:05,345 --> 00:28:09,045

with using Bayesian is just the time

commitment.

425

00:28:09,045 --> 00:28:09,861

I'm

426

00:28:09,997 --> 00:28:15,657

There is one analysis I was running and it

takes, I think for one analysis, it takes

427

00:28:15,657 --> 00:28:20,837

almost a week to run it because it's a big

sample and then it's a complicated model.

428

00:28:20,837 --> 00:28:27,377

And so if I would have to rerun that model

a thousand times, it would not be fun.

429

00:28:27,917 --> 00:28:33,717

so knowing that there's maybe some options

on the horizon to help us speed along that

430

00:28:33,717 --> 00:28:38,989

process would be, I think that would

change our field for sure.

431

00:28:38,989 --> 00:28:40,789

So that's very exciting.

432

00:28:40,869 --> 00:28:41,629

Yeah, yeah, yeah.

433

00:28:41,629 --> 00:28:43,669

That's really super exciting.

434

00:28:43,669 --> 00:28:48,569

And that's why I'm also super enthusiastic

about the desalmatized Bayesian infant

435

00:28:48,569 --> 00:28:52,649

stuff, because I discovered that in

episode 107, so it's not a long time ago.

436

00:28:52,649 --> 00:28:59,509

But as soon as I heard about that, I dug

into it, because that's super interesting.

437

00:28:59,569 --> 00:29:00,549

Yeah.

438

00:29:00,549 --> 00:29:04,829

I'm going to read about it after we finish

recording this.

439

00:29:04,849 --> 00:29:06,549

Yeah, yeah, for sure.

440

00:29:06,549 --> 00:29:09,257

And feel free to send me any questions.

441

00:29:09,421 --> 00:29:17,841

And I find it's also a very elegant way to

marry the Bayesian framework in the deep

442

00:29:17,841 --> 00:29:20,801

neural network methods.

443

00:29:20,801 --> 00:29:21,961

So I really love that.

444

00:29:21,961 --> 00:29:26,881

It's really elegant and promising, as you

were saying.

445

00:29:27,161 --> 00:29:33,461

Talking about SCM, so structural equation

modeling, do you find that Bayesian

446

00:29:33,461 --> 00:29:35,441

methods help?

447

00:29:36,493 --> 00:29:41,933

in for these kind of models and especially

when it comes to educational research

448

00:29:41,933 --> 00:29:44,673

which is one of your fields?

449

00:29:46,113 --> 00:29:54,153

Yes, I think Bayesian methods can sort of

help on both ends of the spectrum that we

450

00:29:54,153 --> 00:30:00,053

see with educational data which is either

we have very small samples and so

451

00:30:00,053 --> 00:30:04,461

researchers still have these ambitious

theoretical models that they want to test.

452

00:30:04,461 --> 00:30:07,861

but it's just not doable with frequentist

estimators.

453

00:30:07,861 --> 00:30:14,001

And so based with the priors, it can help

a little bit to boost the information that

454

00:30:14,001 --> 00:30:16,141

we have, which is really nice.

455

00:30:16,141 --> 00:30:21,801

And then on the other side, ever since

starting this position and moving into a

456

00:30:21,801 --> 00:30:28,781

college of education, I've been given

access to many large data sets that have

457

00:30:28,781 --> 00:30:30,931

very complicated nesting structures.

458

00:30:30,931 --> 00:30:33,281

That's something you see all the time in

education.

459

00:30:33,281 --> 00:30:33,901

You have

460

00:30:33,901 --> 00:30:38,161

schools and then teachers and students and

the students they change teachers because

461

00:30:38,161 --> 00:30:42,721

it's also longitudinal so there's a time

component and all of these different

462

00:30:42,721 --> 00:30:47,641

nested structures can be very hard to

model using estimators like nextman

463

00:30:47,641 --> 00:30:52,141

likelihood and bayesian methods not

necessarily structural equation modeling

464

00:30:52,141 --> 00:30:58,141

but maybe more a hierarchical linear model

or some other multi -level approach it can

465

00:30:58,141 --> 00:31:03,365

be super flexible to handle all of those

466

00:31:03,565 --> 00:31:10,355

structures and still give people results

that they can use to inform policy.

467

00:31:10,355 --> 00:31:14,805

Because that's something in education that

I didn't really see when I was still in

468

00:31:14,805 --> 00:31:18,925

the department of psychology before is

that a lot of the research here is really

469

00:31:18,925 --> 00:31:23,485

directly informing what is actually going

to happen in schools.

470

00:31:23,485 --> 00:31:28,145

And so it's really neat that these

Bayesian methods are allowing them to

471

00:31:28,145 --> 00:31:31,845

answer much more complicated research

questions and really make use of all of

472

00:31:31,845 --> 00:31:32,909

the data that they have.

473

00:31:32,909 --> 00:31:34,409

So that's been really exciting.

474

00:31:34,409 --> 00:31:38,949

And actually, I wanted to ask you

precisely what the challenges you face

475

00:31:38,949 --> 00:31:45,729

with longitudinal data and how do you

address these challenges because I know

476

00:31:45,729 --> 00:31:47,989

that can be pretty hard.

477

00:31:48,009 --> 00:31:53,069

I think with longitudinal data, the

biggest challenge actually doesn't have

478

00:31:53,069 --> 00:31:54,679

anything to do with the estimator.

479

00:31:54,679 --> 00:31:59,357

It is more just inherent in longitudinal

data, which is that we will always...

480

00:31:59,597 --> 00:32:03,417

unless you have a really special sample,

but we will always have missing data.

481

00:32:03,417 --> 00:32:08,217

Participants will always drop out at some

point or just skip a measurement.

482

00:32:08,417 --> 00:32:14,257

And of course, other estimation methods

also have options for accommodating

483

00:32:14,257 --> 00:32:17,637

missing data, such as full information

maximum likelihood.

484

00:32:17,737 --> 00:32:22,037

But I find that the Bayesian approach

where you can do imputation while you're

485

00:32:22,037 --> 00:32:27,717

estimating, so you're just imputing the

data at every posterior sample, is very

486

00:32:27,717 --> 00:32:29,809

elegant, efficient.

487

00:32:30,085 --> 00:32:33,485

and easy for researchers to wrap their

minds around.

488

00:32:33,485 --> 00:32:37,405

And it still allows you just like with

other multiple imputation methods to

489

00:32:37,405 --> 00:32:43,505

include an sort of auxiliary model

explaining the missingness, which helps

490

00:32:43,505 --> 00:32:47,555

with the like missing at random, type data

that we deal with a lot.

491

00:32:47,555 --> 00:32:51,685

And so I feel that that is especially

exciting.

492

00:32:51,685 --> 00:32:56,385

I honestly started thinking about this

more deeply when I started my position

493

00:32:56,385 --> 00:32:58,565

here and I met my new colleague.

494

00:32:58,861 --> 00:32:59,271

Dr.

495

00:32:59,271 --> 00:33:04,841

Brian Keller, he is working on some

software, it's called BLIMP, which I think

496

00:33:04,841 --> 00:33:11,021

it stands for Bayesian Latent Interaction

Modeling Program, I want to say.

497

00:33:11,301 --> 00:33:16,341

So it's actually created for modeling

interactions between latent variables,

498

00:33:16,341 --> 00:33:17,721

which is a whole other issue.

499

00:33:17,721 --> 00:33:22,741

But within that software, they actually

also created a really powerful method for

500

00:33:22,741 --> 00:33:26,061

dealing with missing data, or not

necessarily the method, but just the

501

00:33:26,061 --> 00:33:27,001

application of it.

502

00:33:27,001 --> 00:33:28,325

And so...

503

00:33:28,333 --> 00:33:32,313

Now that I've met him and he's always

talking about it, it makes me think about

504

00:33:32,313 --> 00:33:33,013

it more.

505

00:33:33,013 --> 00:33:34,973

So that's very exciting.

506

00:33:35,153 --> 00:33:37,353

Yeah, for sure.

507

00:33:37,353 --> 00:33:42,923

And feel free to add a link to this

project to Blimp in the show notes,

508

00:33:42,923 --> 00:33:46,873

because I think that's going to be very

interesting to listeners.

509

00:33:48,753 --> 00:33:56,335

And how, I'm wondering if patient methods

improve...

510

00:33:56,461 --> 00:34:00,681

the measurement and the evaluation

processes in educational settings, because

511

00:34:00,681 --> 00:34:02,221

I know it's a challenge.

512

00:34:02,221 --> 00:34:09,981

Is that something that you're working on

actively right now, or you've done any

513

00:34:09,981 --> 00:34:12,981

projects on that that you want to talk

about?

514

00:34:14,721 --> 00:34:18,441

Well, I teach measurement to grad

students.

515

00:34:18,441 --> 00:34:22,061

So it's not necessarily that I get to talk

about Bayes a lot in there.

516

00:34:22,061 --> 00:34:24,069

But what I'm realizing is that

517

00:34:24,301 --> 00:34:30,341

When we talk about measurement from a

frequentist standpoint, we typically start

518

00:34:30,341 --> 00:34:32,411

with asking students a bunch of questions.

519

00:34:32,411 --> 00:34:34,961

Let's say we're trying to measure math

ability.

520

00:34:34,961 --> 00:34:37,201

So we ask them a bunch of math questions.

521

00:34:37,201 --> 00:34:43,181

Then if we use frequentist estimation, we

can use those item responses to generate

522

00:34:43,181 --> 00:34:48,261

some sort of probability of those

responses giving some underlying level of

523

00:34:48,261 --> 00:34:49,041

math ability.

524

00:34:49,041 --> 00:34:53,869

So how probable is it that they gave these

answers given this level of math?

525

00:34:53,869 --> 00:34:59,429

But actually what we want to know is what

is the student's math ability, given the

526

00:34:59,429 --> 00:35:01,249

patterns of observed responses.

527

00:35:01,249 --> 00:35:05,269

And so Bayes theorem gives us a really

elegant way of answering exactly that

528

00:35:05,269 --> 00:35:06,569

question, right.

529

00:35:06,569 --> 00:35:08,039

Instead of the opposite way.

530

00:35:08,039 --> 00:35:14,889

And so I think in a big way, Bayesian

methods just align better with how people

531

00:35:14,889 --> 00:35:19,229

already think about the research that

they're doing or the thing, the questions

532

00:35:19,229 --> 00:35:20,989

that they're, they want to answer.

533

00:35:20,989 --> 00:35:22,511

I think.

534

00:35:22,541 --> 00:35:26,741

This is also a reason why a lot of

researchers struggle with getting the

535

00:35:26,741 --> 00:35:29,801

interpretation of things like a confidence

interval correct, right?

536

00:35:29,801 --> 00:35:31,821

It's just not intuitive.

537

00:35:31,821 --> 00:35:33,921

Whereas Bayesian methods, they are

intuitive.

538

00:35:33,921 --> 00:35:38,201

And so in that sense, I think not so much

like estimation wise, but just

539

00:35:38,201 --> 00:35:42,081

interpretation wise, Bayesian methods can

help a lot in our field.

540

00:35:42,081 --> 00:35:48,653

And then in addition to that, I think when

we do use Bayesian estimation,

541

00:35:48,653 --> 00:35:53,953

those posterior distributions, they can

give us so much more information about the

542

00:35:53,953 --> 00:35:57,212

parameters of interest that we are

interested in.

543

00:35:58,113 --> 00:36:03,633

And they can also help us understand what

future data would look like given those

544

00:36:03,633 --> 00:36:04,313

posteriors, right?

545

00:36:04,313 --> 00:36:08,813

If we move from like prior predictors to

posterior predictors, which are these

546

00:36:08,813 --> 00:36:13,573

samples generated from the posteriors,

that should look like our data should look

547

00:36:13,573 --> 00:36:14,613

like that data, right?

548

00:36:14,613 --> 00:36:17,653

If our model is doing a good job of

representing our data.

549

00:36:17,653 --> 00:36:18,285

And so,

550

00:36:18,285 --> 00:36:21,585

I think that's an exciting extension of

Bayes as well.

551

00:36:21,585 --> 00:36:26,065

It gives us more tools to evaluate our

model and to make sure that it's actually

552

00:36:26,065 --> 00:36:30,185

doing a good job of representing our data,

which is especially important in

553

00:36:30,185 --> 00:36:35,445

structural equation modeling, where we

rely very heavily on global measures of

554

00:36:35,445 --> 00:36:35,845

fit.

555

00:36:35,845 --> 00:36:39,605

And so this is a really nice new tool for

people to use.

556

00:36:39,605 --> 00:36:39,995

I see.

557

00:36:39,995 --> 00:36:40,315

Okay.

558

00:36:40,315 --> 00:36:40,884

Yeah.

559

00:36:40,884 --> 00:36:41,825

I am.

560

00:36:41,825 --> 00:36:43,765

I need to know about that in particular.

561

00:36:43,765 --> 00:36:44,237

That's...

562

00:36:44,237 --> 00:36:45,267

That's very interesting.

563

00:36:45,267 --> 00:36:45,677

Yeah.

564

00:36:45,677 --> 00:36:50,716

So I mean, I would have more questions on

that, but I want to ask you in particular

565

00:36:50,716 --> 00:36:55,047

on a publication you have about under

-fitting and over -fitting.

566

00:36:55,047 --> 00:37:03,177

And you've looked at the performance of

Bayesian model selection in SEM.

567

00:37:03,417 --> 00:37:05,657

I find that super interesting.

568

00:37:05,657 --> 00:37:10,597

So can you summarize the key findings of

this paper and...

569

00:37:10,669 --> 00:37:15,069

their application, their implications for

researchers using SEM?

570

00:37:15,069 --> 00:37:16,129

Yeah, for sure.

571

00:37:16,129 --> 00:37:21,349

This is a really fun project for me to

work on, kind of an extension of my

572

00:37:21,349 --> 00:37:21,919

dissertation.

573

00:37:21,919 --> 00:37:27,649

So it made me feel like, I'm really moving

on, creating a program of research.

574

00:37:27,649 --> 00:37:30,669

So yeah, thanks for asking about the

paper.

575

00:37:31,829 --> 00:37:36,333

So yeah, as I already kind of mentioned,

within structural equation modeling,

576

00:37:36,333 --> 00:37:41,313

Researchers rely really heavily on these

model selection and fit indices to make

577

00:37:41,313 --> 00:37:44,713

choices about what model they're going to

keep in the end.

578

00:37:44,713 --> 00:37:48,833

A lot of the times, researchers come in

with some idea of what the model would

579

00:37:48,833 --> 00:37:51,873

look like, but they are always tinkering a

little bit.

580

00:37:51,873 --> 00:37:55,873

They're ready to know that they're wrong

and they want to get to a better model.

581

00:37:55,873 --> 00:38:00,453

And so the same is true when we use

Bayesian estimation and we have sort of a

582

00:38:00,453 --> 00:38:03,133

similar set of indices to look at.

583

00:38:03,725 --> 00:38:08,585

in terms of the fit of a single model or

comparing multiple models and selecting

584

00:38:08,585 --> 00:38:10,065

the best one.

585

00:38:11,105 --> 00:38:18,505

And so very typically those indices are

tested in terms of how well they can

586

00:38:18,505 --> 00:38:20,045

identify underfit.

587

00:38:20,045 --> 00:38:24,255

And so underfit occurs when you forgot to

include a certain parameter.

588

00:38:24,255 --> 00:38:29,445

So your model is too simple for the

underlying data generating mechanism.

589

00:38:29,445 --> 00:38:31,085

You forgot something.

590

00:38:31,125 --> 00:38:33,421

And so all of these indices generally

work.

591

00:38:33,421 --> 00:38:38,101

pretty well, and that's also what we found

in our study in terms of selecting the

592

00:38:38,101 --> 00:38:44,301

correct model when there are some

alternatives that have fewer parameters or

593

00:38:44,301 --> 00:38:49,341

picking up on the correct model fitting

well by itself versus models that forget

594

00:38:49,341 --> 00:38:50,701

these parameters.

595

00:38:50,701 --> 00:38:55,301

But what we were really interested in is

looking at, OK, how well do these indices

596

00:38:55,301 --> 00:38:57,141

actually detect overfitting?

597

00:38:57,141 --> 00:38:59,751

So that's where you add parameters that

you don't really need.

598

00:38:59,751 --> 00:39:02,435

So you're making your model overly

complex.

599

00:39:02,445 --> 00:39:06,525

And when we have models that are too

complex, they tend not to generalize to

600

00:39:06,525 --> 00:39:07,405

new samples, right?

601

00:39:07,405 --> 00:39:12,685

They're optimized for our specific sample

and that's not really useful in science.

602

00:39:13,245 --> 00:39:18,045

So we want to make sure that we don't keep

going and like adding paths and making our

603

00:39:18,045 --> 00:39:19,905

models super complicated.

604

00:39:20,085 --> 00:39:26,965

And so surprisingly what we found across

like a range of over fitting scenarios is

605

00:39:26,965 --> 00:39:31,801

that they do not really do a good job of

detecting any of this.

606

00:39:31,801 --> 00:39:37,081

Most indices, if anything, just make the

model look better and better and better.

607

00:39:37,081 --> 00:39:42,821

Even some of these indices, like model

selection indices, will have a penalty

608

00:39:42,821 --> 00:39:47,301

term in their formula that's supposed to

penalize for having too many parameters,

609

00:39:47,301 --> 00:39:47,461

right?

610

00:39:47,461 --> 00:39:50,341

For making your model too complex.

611

00:39:50,341 --> 00:39:52,861

And even those were just like, yeah, this

is fine.

612

00:39:52,861 --> 00:39:54,701

Keep going, keep going.

613

00:39:55,061 --> 00:39:57,371

And so that's a little bit worrisome.

614

00:39:57,371 --> 00:39:58,961

And I think...

615

00:40:00,141 --> 00:40:05,701

We really need to think about developing

some new ways of detecting when we go too

616

00:40:05,701 --> 00:40:07,021

far, right?

617

00:40:07,081 --> 00:40:11,121

Figuring out at what point we need to stop

in our model modification, which is

618

00:40:11,121 --> 00:40:15,481

something that researchers really love to

do, especially in structural equation

619

00:40:15,481 --> 00:40:15,881

modeling.

620

00:40:15,881 --> 00:40:18,781

I won't speak for any other areas.

621

00:40:19,421 --> 00:40:23,141

And so, yeah, I think there's a lot of

work to be done.

622

00:40:23,141 --> 00:40:27,081

And I was very surprised that these

indices that are supposed to help us

623

00:40:27,081 --> 00:40:29,581

detect overfitting also didn't really do.

624

00:40:29,581 --> 00:40:30,541

a good job.

625

00:40:30,541 --> 00:40:35,861

And so I'm excited to work more on this.

626

00:40:35,861 --> 00:40:41,161

I would say in general, if people want an

actionable takeaway, it is always helpful

627

00:40:41,161 --> 00:40:45,731

when you have multiple models to compare

versus just your one model of interest.

628

00:40:45,731 --> 00:40:52,561

It will help you tease, sort of figure out

better, which one is the correct one

629

00:40:52,561 --> 00:40:55,101

versus just is your model good enough?

630

00:40:55,101 --> 00:40:58,421

And so that would be my, my advice for

researchers.

631

00:40:58,421 --> 00:40:59,405

Yeah.

632

00:40:59,405 --> 00:41:00,705

Yeah, definitely.

633

00:41:01,225 --> 00:41:09,905

I always like having a very basic and dumb

looking linear regression to compare to

634

00:41:09,905 --> 00:41:17,205

that and build my way on top of that

because you can already do really cool

635

00:41:17,205 --> 00:41:23,245

stuff with plain simple linear regression

and why making it harder if you cannot

636

00:41:23,245 --> 00:41:28,973

prove, you cannot discern a particular

effect of...

637

00:41:28,973 --> 00:41:31,913

of the new method you're applying.

638

00:41:32,773 --> 00:41:33,473

Yeah.

639

00:41:34,153 --> 00:41:44,233

And so do you have then from from your

dive into these, do you have some fit

640

00:41:44,233 --> 00:41:46,193

indices that you recommend?

641

00:41:46,193 --> 00:41:49,653

And how do they compare to traditional fit

indices?

642

00:41:50,413 --> 00:41:57,533

So I think for model

643

00:41:57,901 --> 00:42:03,401

fit of a single model within structural

equation modeling.

644

00:42:03,401 --> 00:42:08,081

The most popular ones are called

comparative fit index, the Tucker Lewis

645

00:42:08,081 --> 00:42:11,241

index, and then the root mean square error

of approximation.

646

00:42:11,241 --> 00:42:14,101

You'll see these in like every single

paper published.

647

00:42:14,101 --> 00:42:20,601

And so there are Bayesian versions of

those indices, but based on all my

648

00:42:20,601 --> 00:42:22,873

research using those so far,

649

00:42:23,085 --> 00:42:27,685

I would actually not recommend those at

all for evaluating the fit of your

650

00:42:27,685 --> 00:42:29,445

specific model.

651

00:42:30,825 --> 00:42:35,825

It seems from at least my research that

they are very sensitive to your sample

652

00:42:35,825 --> 00:42:41,345

size, which means that as you get a larger

and larger sample, your model will just

653

00:42:41,345 --> 00:42:45,485

keep looking better and better and better

and better, even if it's wrong.

654

00:42:45,585 --> 00:42:49,445

So something that would be flagged as like

a...

655

00:42:49,453 --> 00:42:53,653

a misspecified model with a small sample

might look perfectly fine with a large

656

00:42:53,653 --> 00:42:54,113

sample.

657

00:42:54,113 --> 00:42:57,253

And so that's not what you want, right?

658

00:42:57,253 --> 00:43:01,053

You want the fit index to reflect the

misspecification, not your sample size.

659

00:43:01,053 --> 00:43:06,653

And so I was really excited when these

were first introduced, but I think we need

660

00:43:06,653 --> 00:43:12,553

a lot more knowledge about how to actually

use them before they are really useful.

661

00:43:12,853 --> 00:43:18,153

And so my advice for researchers who want

to know something about their fit is

662

00:43:18,153 --> 00:43:19,341

really to look at

663

00:43:19,341 --> 00:43:22,041

the posterior predictive checks.

664

00:43:22,241 --> 00:43:27,361

And within structural equation modeling,

I'm not sure how widespread this is for

665

00:43:27,361 --> 00:43:32,021

other methods, but we have something

called a posterior predictive p -value,

666

00:43:32,021 --> 00:43:38,301

where we basically take our observed data

and evaluate the fit of that data to our

667

00:43:38,301 --> 00:43:40,541

model at each posterior iteration.

668

00:43:40,541 --> 00:43:44,861

For example, using a likelihood ratio test

or like a chi -square or something.

669

00:43:44,861 --> 00:43:48,877

And then we do the same for a posterior

predicted sample.

670

00:43:48,877 --> 00:43:53,257

using this in within each of those samples

as well.

671

00:43:53,257 --> 00:44:00,257

And the idea is that if your model fits

your data well, then about half of the

672

00:44:00,257 --> 00:44:04,627

predictive samples should fit better and

the other half should fit worse, right?

673

00:44:04,627 --> 00:44:07,977

Yours should be nicely cozy in the middle.

674

00:44:08,497 --> 00:44:14,397

If all of your posterior predictive

samples fit worse than your actual data,

675

00:44:14,397 --> 00:44:17,567

then it's an indication that you are

overfitting, right?

676

00:44:17,567 --> 00:44:18,317

Like,

677

00:44:18,317 --> 00:44:22,057

the model will never fit as well as it

does for your specific data.

678

00:44:22,057 --> 00:44:27,737

And so I think in that sense, that index

could potentially give some idea of

679

00:44:27,737 --> 00:44:32,377

overfitting, although again, in our study,

we didn't really see that happening.

680

00:44:33,057 --> 00:44:39,997

But I think it's a more informative method

of looking at fit within Bayesian

681

00:44:39,997 --> 00:44:41,737

structural equation modeling.

682

00:44:41,737 --> 00:44:46,893

And so even though it's kind of old

school, I think it's still probably the...

683

00:44:46,893 --> 00:44:49,773

the best option for researchers to look

at.

684

00:44:49,773 --> 00:44:51,653

Okay, yeah, thanks.

685

00:44:51,813 --> 00:44:53,273

That's like, I love that.

686

00:44:53,273 --> 00:44:54,933

That's very practical.

687

00:44:55,573 --> 00:44:59,433

And I think listeners really appreciate

that.

688

00:45:01,593 --> 00:45:10,333

I have like, I was wondering about SEMs

again, and if you have an example from

689

00:45:10,333 --> 00:45:15,853

your research where Bayesian SEM provided

significant insights that

690

00:45:15,853 --> 00:45:18,095

traditional methods might have missed.

691

00:45:19,629 --> 00:45:27,909

Yeah, so some work I'm working on right

now is with a group of researchers who are

692

00:45:27,909 --> 00:45:34,269

really interested in figuring out how

strong the evidence is that there is no

693

00:45:34,269 --> 00:45:35,149

effect, right?

694

00:45:35,149 --> 00:45:39,889

That some path is zero within a bigger

structural model.

695

00:45:40,129 --> 00:45:44,969

And with frequentist analysis, all we can

really do is fail to reject the known,

696

00:45:44,969 --> 00:45:45,349

right?

697

00:45:45,349 --> 00:45:48,269

We have an absence of evidence.

698

00:45:48,493 --> 00:45:51,513

but that doesn't mean that there's

evidence of absence.

699

00:45:52,093 --> 00:45:57,413

And so we can't really quantify like how

strong or how convinced we should be that

700

00:45:57,413 --> 00:45:59,993

that null is really a null effect.

701

00:45:59,993 --> 00:46:03,953

But with Bayesian methods, we have base

factors, right?

702

00:46:03,953 --> 00:46:09,853

And we can actually explicitly test the

evidence in favor of the estimate being

703

00:46:09,853 --> 00:46:13,233

zero versus the estimate being not zero,

right?

704

00:46:13,233 --> 00:46:15,913

Either smaller or larger than zero.

705

00:46:15,913 --> 00:46:17,805

And so that's really...

706

00:46:17,805 --> 00:46:21,754

When I talked to the applied researchers,

once they came to me with this problem,

707

00:46:21,754 --> 00:46:26,165

which started as just like a structural

equation modeling problem, but then I was

708

00:46:26,165 --> 00:46:28,895

like, well, have you ever considered using

Bayesian methods?

709

00:46:28,895 --> 00:46:31,785

Because I feel like it could really help

you get at that question.

710

00:46:31,785 --> 00:46:37,225

Like how strong is that evidence relative

to the evidence for an effect, right?

711

00:46:37,225 --> 00:46:42,985

And so we've been working on that right

now and it is very interesting to see the

712

00:46:42,985 --> 00:46:46,829

results and then also to communicate that

with them and see.

713

00:46:46,829 --> 00:46:48,589

They get so excited about it.

714

00:46:48,589 --> 00:46:51,409

So that's been fun.

715

00:46:52,129 --> 00:46:53,069

Yeah, for sure.

716

00:46:53,069 --> 00:46:54,249

That's super cool.

717

00:46:54,249 --> 00:46:58,789

And you don't have anything to share in

the show notes yet, right?

718

00:46:58,789 --> 00:47:00,149

Not yet.

719

00:47:00,609 --> 00:47:02,569

No, I'll keep you posted.

720

00:47:02,569 --> 00:47:03,489

Yeah, for sure.

721

00:47:03,489 --> 00:47:09,429

Because maybe by the time of publication,

you'll have something for us.

722

00:47:09,429 --> 00:47:10,389

Yes.

723

00:47:12,669 --> 00:47:16,109

And now I'd like to talk a bit about

your...

724

00:47:16,109 --> 00:47:20,369

your teaching because you teach a lot of

classes.

725

00:47:20,709 --> 00:47:25,609

You've talked a bit about that already at

the beginning of the show, but how do you

726

00:47:25,609 --> 00:47:31,309

approach teaching Bayesian methods to

students in your program, which is the

727

00:47:31,309 --> 00:47:35,349

statistics measurement and evaluation and

indication program?

728

00:47:35,789 --> 00:47:43,589

Yeah, so I got to be honest and say I have

never taught an entire class on Bayesian

729

00:47:43,589 --> 00:47:44,845

methods yet.

730

00:47:44,845 --> 00:47:52,165

I'm very excited that I just talked with

my colleagues and I got the okay to

731

00:47:52,165 --> 00:47:54,355

develop it and put it on the schedule.

732

00:47:54,355 --> 00:47:55,965

So it's coming.

733

00:47:56,645 --> 00:48:02,245

But I did recently join a panel

discussion, which was about teaching

734

00:48:02,245 --> 00:48:03,585

Bayesian methods.

735

00:48:03,585 --> 00:48:09,505

It was organized by the Bayesian Education

Research and Practice Section of the ISBA

736

00:48:09,525 --> 00:48:10,825

Association.

737

00:48:11,345 --> 00:48:14,477

And so the other two panelists, I was

really starstruck.

738

00:48:14,477 --> 00:48:16,877

to be honest, were E .J.

739

00:48:16,877 --> 00:48:24,197

Wagemakers and Joachim van de Kerkoven,

which are like, to me, those are really

740

00:48:24,197 --> 00:48:25,477

big names.

741

00:48:25,817 --> 00:48:30,087

And so talking to them, I really learned a

lot during that panel.

742

00:48:30,087 --> 00:48:36,157

I felt like I was more on the panel as a

as an audience member, but it was great

743

00:48:36,157 --> 00:48:37,557

for me.

744

00:48:38,117 --> 00:48:42,747

And and so from that, I think if I do get

to teach a class on Bayesian methods,

745

00:48:42,747 --> 00:48:44,589

which hopefully will be soon.

746

00:48:44,589 --> 00:48:51,309

I think I really want to focus on showing

students the entire Bayesian workflow,

747

00:48:51,309 --> 00:48:51,589

right?

748

00:48:51,589 --> 00:48:56,129

Just as we were talking about, starting

with figuring out priors, prior predictive

749

00:48:56,129 --> 00:49:00,389

checks, maybe some of that fancy

calibration.

750

00:49:01,489 --> 00:49:06,569

And then also doing sensitivity analyses,

looking at the fit with the posterior

751

00:49:06,569 --> 00:49:09,029

predictive samples, all of that stuff.

752

00:49:09,549 --> 00:49:10,949

I think...

753

00:49:11,789 --> 00:49:16,089

For me, I wouldn't necessarily combine

that with structural equation models

754

00:49:16,089 --> 00:49:18,619

because those are already pretty

complicated models.

755

00:49:18,619 --> 00:49:22,449

And so I think within a class that's

really focused on Bayesian methods, I

756

00:49:22,449 --> 00:49:27,489

would probably stick to a simple but

general model, such as a linear regression

757

00:49:27,489 --> 00:49:30,929

model, for example, to illustrate all of

those steps.

758

00:49:31,869 --> 00:49:37,409

Yeah, I've been just buying, like I have a

whole bookshelf now of books on Bayesian

759

00:49:37,409 --> 00:49:38,469

and teaching Bayesian.

760

00:49:38,469 --> 00:49:41,449

And so I'm excited to start reading those.

761

00:49:42,125 --> 00:49:50,164

developing my class soon yeah that's super

exciting well done congrats on that i'm

762

00:49:50,164 --> 00:49:57,085

glad to hear that so first eg vagon makers

was on the show i don't remember which

763

00:49:57,085 --> 00:50:05,745

episode but i will definitely link to it

in the show notes and second yeah which

764

00:50:05,745 --> 00:50:11,725

books are you are you gonna use well

765

00:50:11,725 --> 00:50:13,085

Good question.

766

00:50:13,545 --> 00:50:22,185

So there's one that I kind of like, but it

is very broad, which is written by David

767

00:50:22,185 --> 00:50:25,485

Kaplan, who's at the University of

Wisconsin Madison.

768

00:50:25,485 --> 00:50:31,525

And it's called, I think, vision

statistics for the social sciences.

769

00:50:31,525 --> 00:50:35,505

And so what I like about it is that many

of the examples that are used throughout

770

00:50:35,505 --> 00:50:38,665

the book are very relevant to the students

that I would be teaching.

771

00:50:38,665 --> 00:50:41,369

And it also covers a wide range.

772

00:50:41,389 --> 00:50:45,809

of models, which would be nice.

773

00:50:45,889 --> 00:50:50,529

But now that I've like philosophically

switched more to this workflow

774

00:50:50,529 --> 00:50:56,069

perspective, it's actually a little bit

difficult to find a textbook that covers

775

00:50:56,069 --> 00:50:56,849

all of those.

776

00:50:56,849 --> 00:51:01,049

And so I may have to rely a lot on some of

the online resources.

777

00:51:01,049 --> 00:51:07,519

I know there's some really great posts by,

I'm so bad with names.

778

00:51:07,519 --> 00:51:10,549

I want to say his name is Michael

something.

779

00:51:11,565 --> 00:51:13,625

Where he talks about workflow.

780

00:51:13,625 --> 00:51:14,525

Yes, probably.

781

00:51:14,525 --> 00:51:16,225

Yes, that sounds familiar.

782

00:51:16,225 --> 00:51:20,945

His posts are really informative and so I

would probably rely on those a lot as

783

00:51:20,945 --> 00:51:21,525

well.

784

00:51:21,525 --> 00:51:26,045

Especially because they also use

relatively simpler models.

785

00:51:26,645 --> 00:51:30,525

I think, yeah, for some of the components

of the workflow that they just haven't

786

00:51:30,525 --> 00:51:33,525

been covered in textbooks as much yet.

787

00:51:33,525 --> 00:51:40,025

So if anyone is writing a book right now,

please add some chapters on those lesser

788

00:51:40,025 --> 00:51:40,525

known.

789

00:51:40,525 --> 00:51:43,345

components, that would be great.

790

00:51:43,605 --> 00:51:43,885

Yeah.

791

00:51:43,885 --> 00:51:50,945

Yeah, so there is definitely Michael

Bedoncourt's blog.

792

00:51:52,125 --> 00:51:58,965

And I know Andrew Gelman is writing a book

right now about the Bayesian workflow.

793

00:51:58,965 --> 00:52:01,125

So the Bayesian workflow paper.

794

00:52:01,545 --> 00:52:03,205

Yeah, that's a good paper.

795

00:52:03,205 --> 00:52:05,185

Yeah, I'll put it in the show notes.

796

00:52:05,185 --> 00:52:08,644

But basically, he's turning that into a

book right now.

797

00:52:09,185 --> 00:52:10,125

Amazing.

798

00:52:10,125 --> 00:52:12,885

Yeah, so it's gonna be perfect for you.

799

00:52:12,885 --> 00:52:19,085

And have you taken a look at his latest

book, Active Statistics?

800

00:52:19,525 --> 00:52:26,745

Because that's exactly for preparing

teachers to teach patient stats.

801

00:52:27,345 --> 00:52:31,365

Yes, he has like an I feel like an older

book as well where he has these

802

00:52:31,365 --> 00:52:35,985

activities, but it's really nice that he

came out with this newer, more recent one.

803

00:52:35,985 --> 00:52:38,689

I haven't read it yet, but it's on my

804

00:52:38,829 --> 00:52:40,729

on my to buy list.

805

00:52:40,729 --> 00:52:45,689

I have to buy these books through the

department, so it takes a while.

806

00:52:45,749 --> 00:52:49,739

Yeah, well, and you can already listen to

episode 106 if you want.

807

00:52:49,739 --> 00:52:55,689

He was on the show and talked exactly

about these books.

808

00:52:56,129 --> 00:52:57,109

Amazing.

809

00:52:57,109 --> 00:53:00,669

I'll put it in the show notes.

810

00:53:01,029 --> 00:53:02,289

And what did we talk about?

811

00:53:02,289 --> 00:53:07,159

There was also Michael Betancourt, E .G.

812

00:53:07,159 --> 00:53:08,881

Wagenmarkers,

813

00:53:09,549 --> 00:53:15,689

Active statistics, microbed and code,

yeah, and the Bayesian workflow paper.

814

00:53:15,969 --> 00:53:19,729

Yeah, thanks for reminding me about that

paper.

815

00:53:19,829 --> 00:53:21,439

Yeah, it's a really good one.

816

00:53:21,439 --> 00:53:23,089

I think it's going to be helpful.

817

00:53:23,089 --> 00:53:26,279

I'm not sure they cover SBC already, but

that's possible.

818

00:53:26,279 --> 00:53:31,929

But SBC, in any case, you'll have it in

the Bayes flow tutorial that I already

819

00:53:31,929 --> 00:53:33,649

linked to in the show notes.

820

00:53:34,909 --> 00:53:37,029

So I'll put out that.

821

00:53:38,477 --> 00:53:43,597

And actually, what are future developments

in Bayesian stats that excite you the

822

00:53:43,597 --> 00:53:47,797

most, especially in the context of

educational research?

823

00:53:48,897 --> 00:53:54,617

Well, what you just talked about, and this

amortized estimation thing is very

824

00:53:54,617 --> 00:53:56,257

exciting to me.

825

00:53:56,257 --> 00:54:01,757

I think, as I mentioned, one of the

biggest hurdles for people switching to

826

00:54:01,757 --> 00:54:06,617

Bayesian methods is just the time

commitment, especially with structural

827

00:54:06,617 --> 00:54:07,915

equation models.

828

00:54:07,917 --> 00:54:12,957

And so knowing that people are working on

algorithms that will speed that up, even

829

00:54:12,957 --> 00:54:16,537

for a single analysis, it's just really

exciting to me.

830

00:54:16,537 --> 00:54:22,317

And in addition to that, sort of in a

similar vein, I think a lot of smart

831

00:54:22,317 --> 00:54:27,537

people are working on software, which is

lowering barriers to entry.

832

00:54:28,217 --> 00:54:32,457

People in education, they know a lot about

education, right?

833

00:54:32,457 --> 00:54:36,749

That's their field, but they don't have

time to really dive into.

834

00:54:36,749 --> 00:54:37,849

Bayesian statistics.

835

00:54:37,849 --> 00:54:40,549

And so for a long time, it was very

inaccessible.

836

00:54:40,549 --> 00:54:46,909

But now, for example, as you already

mentioned, Ed Merkel, he has his package

837

00:54:46,909 --> 00:54:51,049

Blavan, which is great for people who are

interested in structural equation modeling

838

00:54:51,049 --> 00:54:54,089

and Bayesian methods.

839

00:54:54,089 --> 00:54:59,609

And sort of similarly, you have that

Berkner has that BRMS package.

840

00:54:59,609 --> 00:55:04,669

And then if you want to go even more

accessible, there's JASP.

841

00:55:04,877 --> 00:55:10,957

which is that point and click sort of

alternative to SPSS, which I really enjoy

842

00:55:10,957 --> 00:55:17,057

showing people to let them know that they

don't need to be afraid that they'll lose

843

00:55:17,057 --> 00:55:19,917

access to SPSS at some point in their

life.

844

00:55:20,477 --> 00:55:23,677

So I think those are all great things.

845

00:55:23,677 --> 00:55:28,205

And in a similar vein, there are so many

more online resources now.

846

00:55:28,205 --> 00:55:32,345

Then when I first started learning about

base, like when people have questions or

847

00:55:32,345 --> 00:55:37,485

they want to get started, I have so many

links to send them of like papers, online

848

00:55:37,485 --> 00:55:41,705

courses, YouTube videos, podcasts like

this one.

849

00:55:42,725 --> 00:55:46,865

and so I think that's, what's really

exciting to me, not so much what we're

850

00:55:46,865 --> 00:55:48,955

doing behind the scenes, right?

851

00:55:48,955 --> 00:55:53,285

The actual method itself, although that's

also very exciting, but for working with

852

00:55:53,285 --> 00:55:54,309

people.

853

00:55:54,509 --> 00:55:56,349

in education or other applied fields.

854

00:55:56,349 --> 00:56:00,609

I'm glad that we are all working on making

it easier.

855

00:56:01,809 --> 00:56:03,989

So, yeah.

856

00:56:04,529 --> 00:56:05,589

Yeah.

857

00:56:05,609 --> 00:56:09,269

So first, thanks a lot for recommending

the show to people.

858

00:56:10,009 --> 00:56:11,869

I appreciate it.

859

00:56:12,849 --> 00:56:17,249

And yeah, completely resonate with what

you just told.

860

00:56:18,249 --> 00:56:21,517

Happy to hear that the educational efforts

are.

861

00:56:21,517 --> 00:56:25,877

useful for sure that's something that's

very dear to my heart and I spend a lot of

862

00:56:25,877 --> 00:56:33,177

time doing that so my people and yeah as

you are saying it's already hard enough to

863

00:56:33,177 --> 00:56:38,217

know a lot about educational research but

if you have to learn a whole new

864

00:56:38,217 --> 00:56:43,737

statistical framework from scratch it's

very hard and more than that it's not

865

00:56:43,737 --> 00:56:49,597

really valued and incentivized in the

academic realm so like why would you even

866

00:56:49,597 --> 00:56:51,149

spend time doing that?

867

00:56:51,149 --> 00:56:55,689

you'd much rather write a paper.

868

00:56:56,049 --> 00:57:00,149

So that's like, that's for sure that's an

issue.

869

00:57:00,309 --> 00:57:05,349

So yeah, definitely working together on

that is definitely helping.

870

00:57:05,749 --> 00:57:11,009

And on that note, I put all the links in

the show notes and also Paul Burkner was

871

00:57:11,009 --> 00:57:13,989

on the show episode 35.

872

00:57:14,049 --> 00:57:19,969

So for people who want to dig deeper about

Paul's work, especially BRMS, as you

873

00:57:19,969 --> 00:57:21,165

mentioned Sonia.

874

00:57:21,165 --> 00:57:29,585

definitely take a well give a give a

listen to that to that episode and also

875

00:57:29,585 --> 00:57:35,205

for people who are using Python more than

are but really like the formula syntax

876

00:57:35,205 --> 00:57:42,405

that BRMS has you can do that in Python

you can use a package called BAMI and it's

877

00:57:42,405 --> 00:57:47,909

basically BRMS in in Python in the

878

00:57:48,013 --> 00:57:53,273

that's built on top of PimC and that's

built by a lot of very smart and cool

879

00:57:53,273 --> 00:57:56,433

people like my friend Tomica Pretto.

880

00:57:57,353 --> 00:57:59,493

He's one of the main core developers.

881

00:57:59,493 --> 00:58:05,393

I just released actually an online course

with him about advanced regression in

882

00:58:05,393 --> 00:58:06,813

Bambi and Python.

883

00:58:06,813 --> 00:58:08,053

So it was a fun course.

884

00:58:08,053 --> 00:58:12,793

We've been developing that for the last

two years and we released that this week.

885

00:58:12,793 --> 00:58:15,793

So I have to say I'm quite relieved.

886

00:58:16,133 --> 00:58:17,332

Congratulations.

887

00:58:17,645 --> 00:58:18,985

Yeah, that's exciting.

888

00:58:18,985 --> 00:58:20,345

Yeah, that was a very fun one.

889

00:58:20,345 --> 00:58:26,405

It's just, I mean, it took so much time

that because we wanted something that was

890

00:58:26,405 --> 00:58:31,045

really comprehensive and as evergreen as

it gets.

891

00:58:31,045 --> 00:58:36,425

So we didn't want to do something, you

know, quick and then having to do it all

892

00:58:36,425 --> 00:58:37,485

over again one year later.

893

00:58:37,485 --> 00:58:43,185

So I wanted to take our time and basically

take people from normal linear regression

894

00:58:43,185 --> 00:58:45,295

and then okay, how do you generalize that?

895

00:58:45,295 --> 00:58:46,711

How do you handle?

896

00:58:46,861 --> 00:58:52,281

non -normal likelihoods, how do you handle

several categories?

897

00:58:52,281 --> 00:58:59,041

Because most of the examples out there in

the internet are somewhat introductory.

898

00:58:59,181 --> 00:59:03,601

How do you do Poisson regression and

binomial regression most of the time?

899

00:59:03,601 --> 00:59:06,281

But what about the most complex cases?

900

00:59:06,281 --> 00:59:09,601

What happens if you have zero inflated

data?

901

00:59:09,601 --> 00:59:14,321

What happens if you have data that's very

dispersed that a binomial or a Poisson

902

00:59:14,321 --> 00:59:15,437

cannot handle?

903

00:59:15,437 --> 00:59:19,637

What happens if you have multi -category

called data?

904

00:59:19,637 --> 00:59:20,927

More than two categories.

905

00:59:20,927 --> 00:59:22,317

You cannot use the binomial.

906

00:59:22,317 --> 00:59:25,797

You have to use the category called all

the multinomial distributions.

907

00:59:25,797 --> 00:59:27,597

And these ones are harder to handle.

908

00:59:27,597 --> 00:59:32,597

You need another link function that the

inverse logit.

909

00:59:32,597 --> 00:59:34,637

So it's a lot of stuff.

910

00:59:35,077 --> 00:59:38,277

But the cool thing is that then you can do

really powerful models.

911

00:59:38,277 --> 00:59:44,877

And if you marry that with hierarchical

models, that is really powerful stuff that

912

00:59:44,877 --> 00:59:45,325

you can do.

913

00:59:45,325 --> 00:59:49,075

So yeah, that's what the whole course is

about.

914

00:59:49,075 --> 00:59:54,365

I'll have Tommy actually on the show to

talk about that with him.

915

00:59:54,365 --> 00:59:57,325

So that's going to be a fun one.

916

00:59:57,425 --> 01:00:00,425

Yeah, I'm looking forward to hearing more

about it.

917

01:00:00,425 --> 01:00:04,405

Sounds like something I might recommend to

some people that I know.

918

01:00:04,405 --> 01:00:05,285

Yeah, yeah.

919

01:00:05,285 --> 01:00:06,665

that's exciting.

920

01:00:06,665 --> 01:00:07,965

Yeah, yeah, for sure.

921

01:00:07,965 --> 01:00:09,165

Happy to.

922

01:00:09,265 --> 01:00:11,545

Happy to.

923

01:00:12,013 --> 01:00:16,233

like send you send you the link I put the

link in the show notes anyway so that

924

01:00:16,233 --> 01:00:23,913

people who are interested can can take a

look and of course patrons of the show

925

01:00:23,913 --> 01:00:28,553

have a 10 % discount because they are they

are the best listeners in the world so you

926

01:00:28,553 --> 01:00:35,693

know they deserve a gift yes they are well

Sonya I've already taken quite a lot of

927

01:00:35,693 --> 01:00:39,395

your time so we're gonna we're gonna start

closing up but

928

01:00:39,565 --> 01:00:44,805

I'm wondering if you have any advice to

give to aspiring researchers who are

929

01:00:44,805 --> 01:00:49,665

interested in incorporating Bayesian

methods into their own work and who are

930

01:00:49,665 --> 01:00:52,525

working in your field, so educational

research?

931

01:00:52,525 --> 01:00:58,045

Yeah, I think the first thing I would say

is don't be scared, which I say a lot when

932

01:00:58,045 --> 01:00:59,665

I talk about statistics.

933

01:00:59,905 --> 01:01:02,845

Don't be scared and take your time.

934

01:01:03,305 --> 01:01:04,237

I think...

935

01:01:04,237 --> 01:01:09,197

A lot of people may come into Bayesian

methods after hearing about frequentist

936

01:01:09,197 --> 01:01:10,997

methods for years and years and years.

937

01:01:10,997 --> 01:01:15,837

And so it's going to take more than a week

or two to learn everything you need to

938

01:01:15,837 --> 01:01:16,637

know about Bayes, right?

939

01:01:16,637 --> 01:01:17,717

That's normal.

940

01:01:17,717 --> 01:01:22,467

We don't expect to familiarize ourselves

with a whole new field in a day or a week.

941

01:01:22,467 --> 01:01:25,277

And that's fine.

942

01:01:25,557 --> 01:01:27,537

Don't feel like a failure.

943

01:01:28,877 --> 01:01:29,997

Then.

944

01:01:29,997 --> 01:01:34,026

I don't know, I would also try and look

for papers in your field, right?

945

01:01:34,026 --> 01:01:40,097

Like if you're studying school climate, go

online and search for school climate base

946

01:01:40,097 --> 01:01:45,117

and see if anyone else has done any work

on your topic of interest using this new

947

01:01:45,117 --> 01:01:46,437

estimation method.

948

01:01:46,437 --> 01:01:51,457

It's always great to see examples of how

other people are using it within a context

949

01:01:51,457 --> 01:01:53,337

that you are familiar with, right?

950

01:01:53,337 --> 01:01:56,397

You don't have to start reading all these

technical papers.

951

01:01:56,397 --> 01:02:00,597

You can stay within your realm of

knowledge, within your realm of expertise,

952

01:02:00,597 --> 01:02:03,957

and then just eke out a little bit.

953

01:02:04,157 --> 01:02:08,317

And then after that, I mean, as we just

talked about, there are so many resources

954

01:02:08,317 --> 01:02:14,117

available that you can look for, and a lot

of them are starting to become super

955

01:02:14,117 --> 01:02:15,877

specific as well.

956

01:02:15,877 --> 01:02:19,897

So if you are interested in structural

equation models, go look for resources

957

01:02:19,897 --> 01:02:22,285

about Bayesian structural equation

modeling.

958

01:02:22,285 --> 01:02:26,385

But if you're interested in some other

model, try and find resources specific to

959

01:02:26,385 --> 01:02:27,425

those.

960

01:02:27,665 --> 01:02:32,705

And as you're going through this process,

a nice little side benefit that's going to

961

01:02:32,705 --> 01:02:36,645

happen is that you're going to get really

good at Googling because you've got to

962

01:02:36,645 --> 01:02:38,485

find all this information.

963

01:02:38,485 --> 01:02:41,425

But it's out there and it's there to find.

964

01:02:41,785 --> 01:02:44,345

So, yeah, that would really be my advice.

965

01:02:44,845 --> 01:02:46,785

Don't be scared.

966

01:02:46,985 --> 01:02:48,505

Yeah, it's a good one.

967

01:02:48,505 --> 01:02:51,557

That's definitely a good one because

then...

968

01:02:51,885 --> 01:02:57,065

Like if you're not scared to be

embarrassed or fail, you're gonna ask a

969

01:02:57,065 --> 01:03:01,265

lot of questions, you're gonna meet

interesting people, you're gonna learn way

970

01:03:01,265 --> 01:03:02,685

faster than you thought.

971

01:03:02,685 --> 01:03:05,565

So yeah, definitely great advice.

972

01:03:05,565 --> 01:03:07,025

Thanks, Sonja.

973

01:03:07,025 --> 01:03:10,645

And people in our field, Invasion Methods,

they are so nice.

974

01:03:10,645 --> 01:03:13,465

I feel like they are just so excited

when...

975

01:03:13,465 --> 01:03:17,625

I'm so excited when anyone shows any

interest in what I do.

976

01:03:17,709 --> 01:03:21,809

Yeah, don't be scared to reach out to

people either because they're going to be

977

01:03:21,809 --> 01:03:23,608

really happy that you did.

978

01:03:24,469 --> 01:03:25,069

True, true.

979

01:03:25,069 --> 01:03:26,769

Yeah, very good point.

980

01:03:27,429 --> 01:03:36,269

Yeah, I find that community is extremely

welcoming, extremely ready to help.

981

01:03:36,689 --> 01:03:42,749

And honestly, I still have to find trolls

in that community.

982

01:03:42,769 --> 01:03:45,463

That's really super value.

983

01:03:45,485 --> 01:03:50,525

I feel like it helps that a lot of us came

into this area through also kind of like a

984

01:03:50,525 --> 01:03:51,815

roundabout way, right?

985

01:03:51,815 --> 01:03:55,885

I don't think anyone is born thinking

they're going to be a Beijing statistician

986

01:03:55,885 --> 01:03:58,505

and so we understand.

987

01:03:58,645 --> 01:03:58,985

Yeah, yeah.

988

01:03:58,985 --> 01:04:00,015

Yeah, well, I did.

989

01:04:00,015 --> 01:04:03,845

I think my first word was prior.

990

01:04:03,845 --> 01:04:05,745

So, you know, okay.

991

01:04:05,745 --> 01:04:07,495

Well, you're the exception to the rule.

992

01:04:07,495 --> 01:04:08,015

Yeah, yeah.

993

01:04:08,015 --> 01:04:10,385

But you know, that's life.

994

01:04:10,385 --> 01:04:13,345

I'm used to being the black sheep.

995

01:04:13,345 --> 01:04:14,457

That's fine.

996

01:04:14,873 --> 01:04:22,473

no, I think I wanted to be a football

player or something like that.

997

01:04:22,473 --> 01:04:25,193

no, also I wanted to fly planes.

998

01:04:25,393 --> 01:04:32,273

I wanted to be a fighter pilot at some

point later after I had outgrown football.

999

01:04:32,553 --> 01:04:35,313

You're a thrill seeker.

Speaker:

01:04:36,493 --> 01:04:43,353

I wanted to be a vet or something, but

then I had to take my pets to the vet and

Speaker:

01:04:43,353 --> 01:04:44,013

they were

Speaker:

01:04:44,013 --> 01:04:47,493

bleeding and I was like, no, I don't want

the event anymore.

Speaker:

01:04:48,293 --> 01:04:56,733

Well, it depends on the kind of animals

you treat, but veterinarian can be a

Speaker:

01:04:56,733 --> 01:05:00,053

thrill seeking experience too.

Speaker:

01:05:00,053 --> 01:05:07,513

You know, like if you're specialized in

snakes or grizzlies or lions, I'm guessing

Speaker:

01:05:07,513 --> 01:05:12,827

it's not all the time, you know, super,

super easy and tranquil.

Speaker:

01:05:13,645 --> 01:05:14,825

no.

Speaker:

01:05:16,545 --> 01:05:17,005

Awesome.

Speaker:

01:05:17,005 --> 01:05:19,865

Well Sonia, that was really great to have

you on the show.

Speaker:

01:05:19,865 --> 01:05:22,285

Of course, I'm going to ask you the last

two questions.

Speaker:

01:05:22,285 --> 01:05:24,625

Ask every guest at the end of the show.

Speaker:

01:05:24,625 --> 01:05:29,725

So if you had unlimited time and

resources, which problem would you try to

Speaker:

01:05:29,725 --> 01:05:30,341

solve?

Speaker:

01:05:32,045 --> 01:05:37,185

I thought about this a lot because I

wanted to solve many problems.

Speaker:

01:05:37,525 --> 01:05:41,805

So when I give this answer, I'm hoping

that other people are taking care of all

Speaker:

01:05:41,805 --> 01:05:43,005

those other problems.

Speaker:

01:05:43,005 --> 01:05:48,505

But I think something that I've noticed

recently is that a lot of people seem to

Speaker:

01:05:48,505 --> 01:05:56,505

have lost the ability or the interest in

critical thinking and being curious and

Speaker:

01:05:56,505 --> 01:05:59,245

trying to figure out things by yourself.

Speaker:

01:05:59,245 --> 01:06:02,093

And so that's something that I would like

to.

Speaker:

01:06:02,093 --> 01:06:04,533

solve or improve somehow?

Speaker:

01:06:04,533 --> 01:06:10,193

Don't ask me how, but I think being a

critical thinker and being curious are two

Speaker:

01:06:10,193 --> 01:06:17,173

really important skills to have to succeed

in our society right now.

Speaker:

01:06:17,173 --> 01:06:23,513

I mean, there's so much information being

thrown at us that it's really up to you to

Speaker:

01:06:23,513 --> 01:06:26,633

figure out what to focus on and what to

ignore.

Speaker:

01:06:26,633 --> 01:06:30,061

And for that, you really need this

critical thinking skill and...

Speaker:

01:06:30,061 --> 01:06:33,461

and also the curiosity to actually look

for information.

Speaker:

01:06:33,461 --> 01:06:38,681

And so I think that's, it's also a very

educational problem, I feel.

Speaker:

01:06:38,681 --> 01:06:43,161

So if it's where I am right now in my

career, but yeah, that would be something

Speaker:

01:06:43,161 --> 01:06:44,521

to solve.

Speaker:

01:06:44,641 --> 01:06:45,601

Yeah.

Speaker:

01:06:45,861 --> 01:06:49,181

Completely understand that was actually my

answer also.

Speaker:

01:06:49,181 --> 01:06:50,511

So I like, really?

Speaker:

01:06:50,511 --> 01:06:51,261

Yeah.

Speaker:

01:06:51,261 --> 01:06:51,641

Yeah.

Speaker:

01:06:51,641 --> 01:06:51,921

Yeah.

Speaker:

01:06:51,921 --> 01:06:53,261

I completely agree with you.

Speaker:

01:06:53,261 --> 01:06:53,621

Yeah.

Speaker:

01:06:53,621 --> 01:06:55,041

These are topics I found.

Speaker:

01:06:55,041 --> 01:06:56,161

I find them.

Speaker:

01:06:56,161 --> 01:06:57,241

I find super interesting.

Speaker:

01:06:57,241 --> 01:06:58,181

How do you.

Speaker:

01:06:58,221 --> 01:07:02,581

do we teach critical thinking, how do we

teach the scientific methods, things like

Speaker:

01:07:02,581 --> 01:07:02,821

that.

Speaker:

01:07:02,821 --> 01:07:07,401

It's always something I'm super excited to

talk about.

Speaker:

01:07:07,561 --> 01:07:13,941

Yeah, I also hope it will have some sort

of trickle down effect on all the other

Speaker:

01:07:13,941 --> 01:07:14,721

problems, right?

Speaker:

01:07:14,721 --> 01:07:18,661

Once the whole world is very skilled at

critical thinking, all the other issues

Speaker:

01:07:18,661 --> 01:07:22,001

will be resolved pretty quickly.

Speaker:

01:07:22,101 --> 01:07:27,149

Yeah, not only because it's directly

solved, but...

Speaker:

01:07:27,149 --> 01:07:34,729

I would say mainly because then you have

maybe less barriers.

Speaker:

01:07:35,809 --> 01:07:39,789

And so yeah, probably coming from that.

Speaker:

01:07:40,429 --> 01:07:46,709

And then second question, if you could

have dinner with any great scientific

Speaker:

01:07:46,709 --> 01:07:50,869

mind, dead, alive or fictional food.

Speaker:

01:07:51,929 --> 01:07:54,285

So I ended up

Speaker:

01:07:54,285 --> 01:07:59,485

Choosing Ada Lovelace who's like one of

the first or maybe the first woman who

Speaker:

01:07:59,485 --> 01:08:02,525

ever worked in computer programming area.

Speaker:

01:08:02,525 --> 01:08:06,785

I think she's very interesting I also

recently found out that she passed away

Speaker:

01:08:06,785 --> 01:08:12,505

when she was only like 36 Which is like

I'm I'm getting at that age and she

Speaker:

01:08:12,505 --> 01:08:16,665

already accomplished all these things By

the time she passed away and so now I'm

Speaker:

01:08:16,665 --> 01:08:22,145

like, okay I gotta I gotta step it up, but

I would really love to talk to her about

Speaker:

01:08:22,145 --> 01:08:24,027

just her experience.

Speaker:

01:08:24,077 --> 01:08:30,657

being so unique in that very manly world

and in that very manly time in general, I

Speaker:

01:08:30,657 --> 01:08:36,017

think it would be very interesting to hear

the challenges and also maybe some

Speaker:

01:08:36,017 --> 01:08:40,577

advantages or like benefits that she saw,

like why did she go through all this

Speaker:

01:08:40,577 --> 01:08:42,597

trouble to begin with?

Speaker:

01:08:42,597 --> 01:08:47,237

Yeah, I think it would be an interesting

conversation to have for sure.

Speaker:

01:08:48,017 --> 01:08:49,757

Yeah, yeah, definitely.

Speaker:

01:08:49,757 --> 01:08:50,757

Yeah, great choice.

Speaker:

01:08:50,757 --> 01:08:53,549

I think, I think somebody already

Speaker:

01:08:53,549 --> 01:08:54,289

had answered.

Speaker:

01:08:54,289 --> 01:08:59,269

I don't remember who, but yeah, it's not a

very common choice.

Speaker:

01:08:59,269 --> 01:09:02,709

We can have a dinner party together.

Speaker:

01:09:03,049 --> 01:09:04,129

Yeah, exactly.

Speaker:

01:09:04,129 --> 01:09:05,329

That's perfect.

Speaker:

01:09:05,449 --> 01:09:06,649

Fantastic.

Speaker:

01:09:06,929 --> 01:09:07,689

Great.

Speaker:

01:09:07,689 --> 01:09:08,289

Thank you.

Speaker:

01:09:08,289 --> 01:09:09,989

Thank you so much, Sonja.

Speaker:

01:09:09,989 --> 01:09:11,829

That was a blast.

Speaker:

01:09:11,889 --> 01:09:13,889

I learned so much.

Speaker:

01:09:14,609 --> 01:09:15,909

Me too.

Speaker:

01:09:16,509 --> 01:09:18,049

You're welcome.

Speaker:

01:09:18,849 --> 01:09:23,245

And well, as usual, I put resources and a

link to a website.

Speaker:

01:09:23,245 --> 01:09:26,685

in the show notes for those who want to

dig deeper.

Speaker:

01:09:26,685 --> 01:09:30,165

Thank you again, Sonia, for taking the

time and being on this show.

Speaker:

01:09:30,465 --> 01:09:31,765

Yeah, thank you.

Speaker:

01:09:31,765 --> 01:09:33,701

It was so much fun.

Speaker:

01:09:37,869 --> 01:09:41,609

This has been another episode of Learning

Bayesian Statistics.

Speaker:

01:09:41,609 --> 01:09:46,569

Be sure to rate, review, and follow the

show on your favorite podcatcher, and

Speaker:

01:09:46,569 --> 01:09:51,489

visit learnbaystats .com for more

resources about today's topics, as well as

Speaker:

01:09:51,489 --> 01:09:56,229

access to more episodes to help you reach

true Bayesian state of mind.

Speaker:

01:09:56,229 --> 01:09:58,149

That's learnbaystats .com.

Speaker:

01:09:58,149 --> 01:10:02,989

Our theme music is Good Bayesian by Baba

Brinkman, fit MC Lass and Meghiraam.

Speaker:

01:10:02,989 --> 01:10:06,149

Check out his awesome work at bababrinkman

.com.

Speaker:

01:10:06,149 --> 01:10:07,309

I'm your host.

Speaker:

01:10:07,309 --> 01:10:08,289

Alex Andorra.

Speaker:

01:10:08,289 --> 01:10:12,549

You can follow me on Twitter at Alex

underscore Andorra, like the country.

Speaker:

01:10:12,549 --> 01:10:17,629

You can support the show and unlock

exclusive benefits by visiting Patreon

Speaker:

01:10:17,629 --> 01:10:19,809

.com slash LearnBasedDance.

Speaker:

01:10:19,809 --> 01:10:22,249

Thank you so much for listening and for

your support.

Speaker:

01:10:22,249 --> 01:10:24,489

You're truly a good Bayesian.

Speaker:

01:10:24,489 --> 01:10:28,039

Change your predictions after taking

information in.

Speaker:

01:10:28,039 --> 01:10:34,149

And if you're thinking I'll be less than

amazing, let's adjust those expectations.

Speaker:

01:10:34,605 --> 01:10:40,065

Let me show you how to be a good Bayesian

Change calculations after taking fresh

Speaker:

01:10:40,065 --> 01:10:46,045

data in Those predictions that your brain

is making Let's get them on a solid

Speaker:

01:10:46,045 --> 01:10:47,845

foundation

Previous post
Next post