WEBVTT

00:00:00.000 --> 00:00:02.550
♪ [music] ♪

00:00:03.800 --> 00:00:05.800
- [Narrator] Welcome 
to Nobel Conversations.

00:00:07.040 --> 00:00:08.100
In this episode,

00:00:08.100 --> 00:00:11.570
Josh Angrist and Guido Imbens
sit down with Isaiah Andrews

00:00:11.570 --> 00:00:14.600
to discuss how the field 
of econometrics is evolving.

00:00:16.100 --> 00:00:18.750
- [Isaiah] So, Guido and Josh, 
you're both pioneers

00:00:18.750 --> 00:00:21.500
in developing tools for
empirical research in economics.

00:00:21.500 --> 00:00:22.930
And so I'd like to explore

00:00:22.930 --> 00:00:25.300
where you feel like 
the field is heading --

00:00:25.709 --> 00:00:28.079
economics, econometrics,
the whole thing.

00:00:28.510 --> 00:00:31.302
To start, I'd be interested to hear

00:00:32.171 --> 00:00:35.200
about whether you feel
the way in which

00:00:35.200 --> 00:00:38.510
the local average treatment 
effects framework took hold

00:00:38.800 --> 00:00:42.187
has any lessons for how 
new empirical methods in economics

00:00:42.187 --> 00:00:44.300
develop and spread
or how they should.

00:00:44.560 --> 00:00:45.960
- [Josh] That's a good question.

00:00:46.610 --> 00:00:47.790
You go first.

00:00:47.790 --> 00:00:49.460
[laughter]

00:00:49.700 --> 00:00:52.940
- [Guido] Yeah, so I think 
the important thing

00:00:52.940 --> 00:00:58.550
is to come up 
with good convincing cases

00:00:58.550 --> 00:01:02.207
where the questions are clear

00:01:02.400 --> 00:01:05.720
and where the methods
apply in general.

00:01:06.253 --> 00:01:07.560
One thing I --

00:01:08.192 --> 00:01:12.000
looking back
at the subsequent literature.

00:01:12.200 --> 00:01:16.700
So I really like the regression
discontinuity literature

00:01:16.700 --> 00:01:19.670
where there were clearly a bunch
of really convincing examples

00:01:19.670 --> 00:01:23.378
and that allowed people
to think more clearly,

00:01:23.378 --> 00:01:27.200
look harder 
at the methodological questions.

00:01:27.400 --> 00:01:28.800
Do clear applications

00:01:28.800 --> 00:01:30.600
that then allow you 
to kind of think about,

00:01:30.600 --> 00:01:33.600
"Wow, does this type of assumption
seem reasonable here?

00:01:33.600 --> 00:01:38.000
What kind of things do we not like
in the early papers?

00:01:38.500 --> 00:01:39.802
How can we improve things?"

00:01:39.802 --> 00:01:44.210
So having clear applications 
motivating these literatures

00:01:44.210 --> 00:01:46.400
I think is very helpful.

00:01:46.800 --> 00:01:48.050
- I'm glad you mentioned

00:01:48.050 --> 00:01:49.382
the regression 
discontinuity, Guido.

00:01:49.382 --> 00:01:53.300
I think there's a lot of 
complementarity between IV and RD,

00:01:54.700 --> 00:01:57.060
Instrumental Variables
and Regression Discontinuity.

00:02:00.506 --> 00:02:03.260
A lot of the econometric
applications

00:02:03.260 --> 00:02:04.520
of regression discontinuity

00:02:04.520 --> 00:02:07.230
are what used to be called 
"fuzzy" RD,

00:02:07.230 --> 00:02:11.620
where, you know, it's not discrete 
or deterministic at the cutoff

00:02:11.620 --> 00:02:14.900
but just the change
in rates or intensity.

00:02:14.900 --> 00:02:17.737
And the LATE framework
helps us understand

00:02:17.737 --> 00:02:18.740
those applications

00:02:18.740 --> 00:02:21.140
and gives us a clear interpretation

00:02:21.140 --> 00:02:25.000
for something like
in my paper with Victor Lavy,

00:02:25.000 --> 00:02:28.100
where we use Maimonides'
rule, the class size cutoffs,

00:02:28.430 --> 00:02:30.030
and what are you getting there?

00:02:30.290 --> 00:02:31.820
Of course, you can
answer that question

00:02:31.820 --> 00:02:33.900
with a linear 
constant effects model,

00:02:34.200 --> 00:02:36.310
but it turns out 
we're not limited to that,

00:02:36.310 --> 00:02:39.889
and RD is still very powerful 
and illuminating,

00:02:40.630 --> 00:02:43.092
even when the correlation

00:02:43.092 --> 00:02:45.866
between the cutoff
and the variable of interest,

00:02:45.866 --> 00:02:49.133
in this case class size, 
is partial,

00:02:49.133 --> 00:02:51.000
maybe even not that strong.

00:02:52.000 --> 00:02:54.999
So there was definitely kind
of a parallel development.

00:02:54.999 --> 00:02:56.400
It's also interesting --

00:02:56.600 --> 00:02:59.780
nobody talked about
regression discontinuity designs

00:02:59.780 --> 00:03:01.220
when we were in graduate school.

00:03:01.220 --> 00:03:05.300
It was something that other social
scientists were interested in,

00:03:05.800 --> 00:03:09.507
and that grew up 
alongside the LATE framework,

00:03:09.507 --> 00:03:11.927
and we've both done work

00:03:11.927 --> 00:03:14.565
on both applications
and methods there,

00:03:14.565 --> 00:03:18.377
and it's been very exciting 
to see that develop

00:03:18.377 --> 00:03:19.800
and become so important.

00:03:20.000 --> 00:03:21.767
It's part of a general evolution,

00:03:21.767 --> 00:03:26.086
I think, towards credible
identification strategies,

00:03:26.086 --> 00:03:27.441
causal effects...

00:03:29.393 --> 00:03:30.642
making econometrics

00:03:30.642 --> 00:03:33.300
more about causal questions
than about models.

00:03:33.640 --> 00:03:34.650
In terms of the future,

00:03:34.650 --> 00:03:37.660
I think one thing that LATE 
has helped facilitate

00:03:37.660 --> 00:03:42.008
is a move towards
more creative, randomized trials

00:03:42.008 --> 00:03:44.400
where there's
something of interest.

00:03:45.500 --> 00:03:48.460
It's not possible 
or straightforward

00:03:48.460 --> 00:03:50.700
to simply turn it off or on,

00:03:51.000 --> 00:03:54.584
but you can encourage it 
or discourage it.

00:03:54.584 --> 00:03:58.200
So you subsidize schooling 
with financial aid, for example.

00:03:59.000 --> 00:04:02.080
So now we have a whole 
framework for interpreting that,

00:04:03.600 --> 00:04:07.113
and it opens the doors
to randomized trials

00:04:07.113 --> 00:04:10.300
of things that maybe would

00:04:10.300 --> 00:04:12.471
not have seemed possible before.

00:04:14.500 --> 00:04:17.864
We've used that a lot in the work
we do on schools in our --

00:04:17.864 --> 00:04:21.160
in the Blueprint Lab at MIT.

00:04:22.360 --> 00:04:26.600
We're exploiting random assignment
in very creative ways, I think.

00:04:28.100 --> 00:04:31.395
- [Isaiah] Related to that,
do you see particular factors

00:04:31.395 --> 00:04:34.445
that make for useful research
in econometrics?

00:04:34.445 --> 00:04:38.290
You've alluded to
it having a clear connection

00:04:38.290 --> 00:04:40.300
to problems 
that are actually coming up,

00:04:40.300 --> 00:04:42.650
and empirical practice 
is often a good idea.

00:04:43.290 --> 00:04:45.000
- Isn't it always a good idea?

00:04:45.700 --> 00:04:47.292
I often find myself sitting

00:04:47.292 --> 00:04:50.100
in an econometrics
theory seminar,

00:04:50.700 --> 00:04:52.500
say the Harvard MIT seminar,

00:04:53.400 --> 00:04:56.350
and I'm thinking, "What problem
is this guy solving?

00:04:56.350 --> 00:04:57.960
Who has this problem?"

00:04:57.960 --> 00:04:59.800
And, you know...

00:05:01.600 --> 00:05:04.700
sometimes there's an
embarrassing silence if I ask

00:05:04.900 --> 00:05:08.300
or there might be
a fairly contrived scenario.

00:05:08.800 --> 00:05:11.600
I want to see 
where the tool is useful.

00:05:12.500 --> 00:05:14.765
There are some 
purely foundational tools.

00:05:14.765 --> 00:05:16.250
I do take the point.

00:05:16.250 --> 00:05:21.735
There are people who are working
on conceptual foundations of ...

00:05:22.600 --> 00:05:25.300
it becomes more like
mathematical statistics.

00:05:25.800 --> 00:05:27.653
I mean, I remember 
an early example of that

00:05:27.653 --> 00:05:29.920
that I struggled to understand

00:05:29.920 --> 00:05:32.500
was the idea
of stochastic equicontinuity,

00:05:32.500 --> 00:05:35.070
which one of my thesis advisors,
Whitney Newey,

00:05:35.070 --> 00:05:36.479
was using to great effect,

00:05:36.479 --> 00:05:38.821
and I was trying
to understand that.

00:05:40.600 --> 00:05:42.034
It's really foundational.

00:05:42.034 --> 00:05:45.200
it's not an application
that's driving that --

00:05:45.890 --> 00:05:47.300
at least not immediately.

00:05:48.600 --> 00:05:53.200
But most things are not like that,
and so there should be a problem.

00:05:53.800 --> 00:05:59.100
And I think it's on the seller 
of that sort of thing,

00:06:00.480 --> 00:06:02.250
because there's opportunity cost,

00:06:02.250 --> 00:06:05.295
the time and attention,
and effort to understand things.

00:06:05.980 --> 00:06:07.200
It's on the seller to say,

00:06:07.400 --> 00:06:08.900
"Hey, I'm solving this problem,

00:06:09.400 --> 00:06:12.900
and here's a set of results
that show that it's useful,

00:06:12.900 --> 00:06:15.200
and here's some insight
that I get."

00:06:16.200 --> 00:06:18.280
- [Isaiah] As you said, Josh,
there's been a move

00:06:18.280 --> 00:06:20.700
in the direction of thinking 
more about causality

00:06:20.700 --> 00:06:22.900
in economics and empirical
work in economics.

00:06:22.900 --> 00:06:24.800
Any consequences of the --

00:06:24.800 --> 00:06:26.570
the spread of that view 
that surprised you

00:06:26.570 --> 00:06:28.347
or anything that you view
as downsides

00:06:28.705 --> 00:06:31.400
of the way that empirical
economics has gone?

00:06:31.500 --> 00:06:34.322
- Sometimes I see 
somebody does IV,

00:06:34.322 --> 00:06:38.304
and they get a result 
which seems implausibly large.

00:06:38.800 --> 00:06:40.200
That's the usual case.

00:06:42.500 --> 00:06:45.220
So it might be
an extraordinarily large

00:06:45.220 --> 00:06:48.600
causal effect of some 
relatively minor intervention,

00:06:49.100 --> 00:06:52.260
which was randomized
or for which you could make a case

00:06:52.260 --> 00:06:54.490
that there's a good design.

00:06:54.900 --> 00:06:57.205
And then when I see that,

00:06:57.944 --> 00:07:00.101
I think it's very hard 
for me to believe

00:07:00.101 --> 00:07:02.030
that this relatively 
minor intervention

00:07:02.030 --> 00:07:03.720
has such a large effect.

00:07:04.100 --> 00:07:06.277
The author will sometimes resort

00:07:06.277 --> 00:07:08.690
to the local average 
treatment effects theorem

00:07:08.690 --> 00:07:11.066
and say, "Well, these compliers,

00:07:11.066 --> 00:07:12.700
they're special in some way."

00:07:13.300 --> 00:07:15.800
And they just benefit
extraordinarily

00:07:15.800 --> 00:07:17.600
from this intervention.

00:07:18.100 --> 00:07:21.175
I'm reluctant to take that
at face value.

00:07:21.175 --> 00:07:23.750
I think often when effects
are too big,

00:07:24.300 --> 00:07:26.780
it's because the exclusion
restriction is failing,

00:07:26.780 --> 00:07:29.456
so you don't really have the right
endogenous variable

00:07:29.456 --> 00:07:31.380
to scale that result.

00:07:32.000 --> 00:07:35.700
And so I'm not too happy to see

00:07:36.937 --> 00:07:40.022
a generic heterogeneity argument

00:07:40.022 --> 00:07:41.760
being used to excuse something

00:07:41.760 --> 00:07:43.800
that I think might be 
a deeper problem.

00:07:45.190 --> 00:07:47.358
- [Guido] I think it played 
somewhat of an unfortunate role

00:07:47.358 --> 00:07:50.083
in the discussions
between reduced form

00:07:50.083 --> 00:07:51.700
and structural approaches,

00:07:51.700 --> 00:07:55.510
where I feel 
that wasn't quite right.

00:07:56.090 --> 00:07:58.810
The instrumental 
variables assumptions

00:07:58.810 --> 00:08:03.483
are at the core, structural 
assumptions about behavior --

00:08:03.483 --> 00:08:05.200
they were coming from economic...

00:08:07.100 --> 00:08:09.900
thinking about the economic
behavior of agents,

00:08:10.300 --> 00:08:15.100
and somehow it got pushed
in a direction

00:08:15.100 --> 00:08:17.600
that I think wasn't 
really very helpful.

00:08:20.426 --> 00:08:21.663
I think, initially,

00:08:22.800 --> 00:08:24.067
we wrote things up,

00:08:24.067 --> 00:08:26.480
it was describing
what was happening.

00:08:26.480 --> 00:08:29.783
There were a set of methods 
people were using.

00:08:29.783 --> 00:08:32.111
We clarified what 
those methods were doing

00:08:32.811 --> 00:08:38.361
and in a way that I think
contain a fair amount of insight.

00:08:39.100 --> 00:08:42.050
But it somehow 
got pushed into a corner

00:08:42.050 --> 00:08:45.379
that I don't think 
was necessarily very helpful.

00:08:45.379 --> 00:08:48.604
- In just the language
of reduced form versus structural,

00:08:48.604 --> 00:08:50.306
I find kind of funny in the sense

00:08:50.306 --> 00:08:52.985
that the local average
treatment effect model,

00:08:52.985 --> 00:08:54.154
the potential outcomes model

00:08:54.154 --> 00:08:56.110
is a nonparametric
structural model,

00:08:56.110 --> 00:08:58.600
if you want to think about it, 
as you suggested, Guido.

00:08:58.600 --> 00:09:01.129
So there's something a little funny

00:09:01.129 --> 00:09:03.505
about putting these 
two things in opposition when --

00:09:03.505 --> 00:09:05.116
- [Guido] Yes.
- [Josh] That language, of course,

00:09:05.116 --> 00:09:08.371
comes from the simultaneous
equations framework

00:09:08.371 --> 00:09:09.641
that we inherited.

00:09:10.400 --> 00:09:11.440
It has the advantage

00:09:11.440 --> 00:09:14.160
that people seem to know 
what you mean when you use it,

00:09:14.160 --> 00:09:16.240
but that might be that people 
are hearing different --

00:09:16.240 --> 00:09:18.300
different people
are hearing different things.

00:09:18.300 --> 00:09:20.480
- [Guido] Yeah. I think
reduced form has become

00:09:20.480 --> 00:09:22.200
used in a little bit 
of the pejorative way...

00:09:22.200 --> 00:09:23.540
- [Josh] Sometimes.

00:09:25.104 --> 00:09:28.250
...which is not really quite what 
it was originally intended for.

00:09:30.100 --> 00:09:33.090
- [Isaiah] I guess something else
that strikes me in thinking about

00:09:33.090 --> 00:09:35.645
the effects of the local average
treatment effect framework

00:09:35.645 --> 00:09:37.676
is that often folks will appeal

00:09:37.676 --> 00:09:40.000
to a local average treatment
effects intuition

00:09:40.000 --> 00:09:42.358
for settings well beyond ones

00:09:42.358 --> 00:09:44.963
where any sort of formal result
has actually been established.

00:09:45.440 --> 00:09:49.180
And I'm curious, given all the work
that you guys did

00:09:49.180 --> 00:09:52.390
to establish LATE results 
in different settings,

00:09:52.390 --> 00:09:54.415
I'm curious, any thoughts on that?

00:09:55.360 --> 00:09:57.306
- I think there's going
to be a lot of cases

00:09:57.306 --> 00:10:02.200
where the intuition 
does get you some distance,

00:10:02.800 --> 00:10:04.989
but it's going to be 
somewhat limited,

00:10:04.989 --> 00:10:07.600
and establishing 
formal results there

00:10:08.400 --> 00:10:09.490
may be a little tricky

00:10:09.490 --> 00:10:12.700
and then maybe only work 
in special circumstances,

00:10:14.600 --> 00:10:16.540
and you end up
with a lot of formality

00:10:16.540 --> 00:10:19.500
that may not quite 
capture the intuition.

00:10:19.900 --> 00:10:21.550
Sometimes I'm somewhat 
uneasy with them,

00:10:21.550 --> 00:10:24.438
and they are not necessarily
the papers I would want to write,

00:10:25.148 --> 00:10:27.819
but I do think intuition

00:10:27.819 --> 00:10:31.217
often does capture 
part of the problem.

00:10:33.100 --> 00:10:36.300
I think, in some sense,
we were very fortunate there

00:10:36.900 --> 00:10:39.250
in the way that the LATE paper 
got handled at the journal,

00:10:39.250 --> 00:10:41.766
so that, actually, the editor,
made it much shorter

00:10:42.100 --> 00:10:46.300
and that allowed us to focus
on very clear, crisp results.

00:10:49.924 --> 00:10:51.770
There's a somewhat
unfortunate tendency

00:10:51.770 --> 00:10:53.118
in the econometrics literature

00:10:53.118 --> 00:10:55.100
of having the papers 
get longer and longer.

00:10:55.100 --> 00:10:56.550
- Well, you should be able
to fix that, man.

00:10:56.550 --> 00:10:58.915
- I'm trying to fix that.
[laughter]

00:10:58.915 --> 00:11:01.625
But I think this is an example 
where it's very clear

00:11:01.625 --> 00:11:03.097
that having it be short
is actually --

00:11:03.097 --> 00:11:04.750
- You should have imposed
that no paper

00:11:04.750 --> 00:11:06.802
can be longer than the LATE paper.

00:11:07.269 --> 00:11:09.617
- That... wow! That may be great.

00:11:09.617 --> 00:11:11.685
- At least no theory, 
no theory paper.

00:11:11.892 --> 00:11:14.300
- Yeah, and I think...

00:11:14.500 --> 00:11:16.800
I'm trying very hard to get
the papers to be shorter,

00:11:16.800 --> 00:11:19.514
and I think there is a lot
of value today

00:11:19.514 --> 00:11:21.573
because it's often
the second part of the paper

00:11:21.573 --> 00:11:25.049
that doesn't actually
get you much further

00:11:25.049 --> 00:11:26.395
in understanding things,

00:11:27.000 --> 00:11:29.870
and it does make things 
much harder to read.

00:11:32.426 --> 00:11:36.179
It goes back to how I think
econometrics should be done.

00:11:36.179 --> 00:11:38.070
You should focus on --

00:11:38.700 --> 00:11:41.300
It should be reasonably
close to empirical problems.

00:11:41.500 --> 00:11:43.900
They should be very clear problems.

00:11:44.800 --> 00:11:48.900
But then often the theory
doesn't need to be quite so long.

00:11:48.900 --> 00:11:50.010
- [Josh] Yeah.

00:11:51.100 --> 00:11:54.670
- I think things have gone
a little off track.

00:11:56.260 --> 00:11:57.750
- [Isaiah] A relatively 
recent change

00:11:57.750 --> 00:12:00.230
has been a seeming
big increase in demand

00:12:00.230 --> 00:12:03.765
for people with econometrics
causal effect estimation skills

00:12:03.765 --> 00:12:04.994
in the tech sector.

00:12:04.994 --> 00:12:07.563
I'm interested, 
do either of you have thoughts

00:12:07.563 --> 00:12:09.840
of how that's going to interact

00:12:09.840 --> 00:12:11.600
with the development 
of empirical methods

00:12:11.600 --> 00:12:13.950
or empirical research
in economics going forward?

00:12:14.600 --> 00:12:16.770
- [Josh] Well, there's
sort of a meta point,

00:12:16.770 --> 00:12:21.000
which is, there's this new 
kind of employer,

00:12:21.800 --> 00:12:27.530
the Amazons and the Uber,
and the TripAdvisor world,

00:12:28.000 --> 00:12:29.300
and I think that's great.

00:12:29.300 --> 00:12:32.030
I like to tell my students
about that.

00:12:32.600 --> 00:12:35.500
At MIT we have a lot
of computer science majors --

00:12:35.500 --> 00:12:37.000
that's our biggest major.

00:12:37.400 --> 00:12:42.362
I try to seduce some of those folks
into economics by saying

00:12:43.228 --> 00:12:46.700
you can go work for these companies

00:12:46.700 --> 00:12:49.250
that people
are very keen to work for

00:12:49.250 --> 00:12:50.800
because the work seems exciting,

00:12:52.000 --> 00:12:54.250
that the skills 
that you get in econometrics

00:12:54.250 --> 00:12:56.100
are as good or better

00:12:56.100 --> 00:12:59.736
than any competing 
discipline has to offer.

00:12:59.736 --> 00:13:01.100
So you should at least

00:13:01.400 --> 00:13:04.200
take some econ, take some
econometrics, and some econ.

00:13:04.800 --> 00:13:07.000
I did a fun project with Uber

00:13:07.600 --> 00:13:09.920
on labor supply of Uber drivers,

00:13:09.920 --> 00:13:12.805
and it was very, very exciting
to be part of that.

00:13:13.100 --> 00:13:15.400
Plus I got to drive 
for Uber for a while,

00:13:15.900 --> 00:13:17.730
and I thought that was fun too.

00:13:17.730 --> 00:13:19.231
I did not make enough

00:13:19.231 --> 00:13:22.616
that I was tempted
to give up my MIT job,

00:13:22.616 --> 00:13:25.100
but I enjoyed the experience.

00:13:25.230 --> 00:13:27.534
I see a potential challenge

00:13:27.534 --> 00:13:30.900
to our model
of graduate education here,

00:13:31.700 --> 00:13:36.068
which is, if we're training people
to go work at Amazon,

00:13:37.900 --> 00:13:41.190
it's not clear why
we should be paying

00:13:41.190 --> 00:13:42.900
graduate stipends for that.

00:13:43.200 --> 00:13:47.280
Why should the taxpayer
effectively be subsidizing that.

00:13:47.280 --> 00:13:51.180
Our graduate education
in the US Is generously subsidized,

00:13:51.180 --> 00:13:53.160
even in private universities,

00:13:53.160 --> 00:13:56.477
it's ultimately -- there's a lot of
public money there,

00:13:56.477 --> 00:13:59.300
and I think the traditional
rationale for that is,

00:13:59.643 --> 00:14:01.992
we were training 
educators and scholars,

00:14:01.992 --> 00:14:05.657
and there's a great externality
from the work that we do,

00:14:05.657 --> 00:14:07.910
it's either 
the research externality,

00:14:07.910 --> 00:14:09.557
or a teaching externality.

00:14:10.100 --> 00:14:13.389
But if many of our students
are going to work

00:14:13.389 --> 00:14:14.688
in the private sector --

00:14:16.300 --> 00:14:17.414
that's fine,

00:14:19.000 --> 00:14:21.700
but maybe their employers
should pay for that.

00:14:22.120 --> 00:14:23.370
- For me, it's [just] so different

00:14:23.370 --> 00:14:26.780
from people working 
for consulting firms.

00:14:27.200 --> 00:14:28.780
It's not clear to me

00:14:28.780 --> 00:14:32.836
that the number of jobs
in academics has changed.

00:14:33.370 --> 00:14:36.325
- I feel like this is
a growing sector,

00:14:36.325 --> 00:14:39.289
whereas consulting -- 
you're right to raise that,

00:14:39.289 --> 00:14:42.100
it might be the same 
for consulting.

00:14:44.846 --> 00:14:47.500
I'm placing more and more 
students in these businesses,

00:14:47.500 --> 00:14:49.467
so it's on my mind, in a way,

00:14:49.467 --> 00:14:53.960
that I've not been attentive 
to consulting jobs.

00:14:53.960 --> 00:14:56.920
Consulting was always important,

00:14:56.920 --> 00:14:58.950
and I think also 
there's some movement

00:14:58.950 --> 00:15:01.140
from consulting back
into research --

00:15:01.140 --> 00:15:02.723
it's a little more fluid.

00:15:03.900 --> 00:15:07.630
A lot of the work in both domains,

00:15:07.630 --> 00:15:09.430
I have to say, 
it's not really different

00:15:09.430 --> 00:15:12.730
but people who are working
in the tech sector

00:15:12.730 --> 00:15:15.480
are doing things that are 
potentially of scientific interest,

00:15:15.480 --> 00:15:16.800
but mostly it's hidden.

00:15:17.100 --> 00:15:18.550
Then you really have to say

00:15:18.550 --> 00:15:20.900
why is the government
paying for this?

00:15:22.477 --> 00:15:23.732
I mean to Guido's point,

00:15:23.732 --> 00:15:26.783
I guess there's 
a data question here of,

00:15:26.783 --> 00:15:32.772
has the total [no-neck] say
for-profit sector employment

00:15:32.772 --> 00:15:35.595
of econ Ph.D. program
graduates increased

00:15:35.595 --> 00:15:38.290
or has it just been
a substitution from finance

00:15:38.290 --> 00:15:40.200
and consulting towards tech?

00:15:40.300 --> 00:15:42.300
- I may be reacting to something

00:15:42.300 --> 00:15:44.300
that's not really happening.

00:15:44.400 --> 00:15:45.890
- I've actually done some work

00:15:45.890 --> 00:15:48.200
with some of these tech companies.

00:15:49.100 --> 00:15:52.200
I don't disagree with Josh's point 
that we need to think

00:15:52.200 --> 00:15:53.675
a little bit about 
the funding model

00:15:53.675 --> 00:15:56.390
who is, in the end, paying 
for the graduate education.

00:15:56.913 --> 00:15:59.400
But from a scientific perspective,

00:15:59.980 --> 00:16:02.840
not only do these places
have great data

00:16:02.840 --> 00:16:05.112
and nowadays they tend to be
very careful with that

00:16:05.112 --> 00:16:07.100
for privacy reasons,

00:16:07.380 --> 00:16:08.900
but they also have great questions.

00:16:10.200 --> 00:16:13.929
I find it very inspiring to listen
to the people there

00:16:13.929 --> 00:16:15.950
and see what kind
of questions they have,

00:16:15.950 --> 00:16:17.330
and often they're questions

00:16:18.200 --> 00:16:21.241
that also come up
outside of these companies.

00:16:21.241 --> 00:16:27.430
I have a couple of papers 
with Raj Chetty and Susan Athey,

00:16:27.430 --> 00:16:31.600
where we look at ways 
of combining experimental data

00:16:31.600 --> 00:16:33.274
and observational data.

00:16:35.500 --> 00:16:38.600
Raj Chetty was interested
in what is the effect

00:16:38.600 --> 00:16:42.893
of early childhood programs 
on outcomes later in life,

00:16:42.893 --> 00:16:46.330
not just kind on test scores
but on earnings and stuff,

00:16:46.330 --> 00:16:48.300
and we kind of developed methods

00:16:48.600 --> 00:16:51.500
that would help you shed light
on that under some --

00:16:52.718 --> 00:16:53.868
in some settings,

00:16:53.868 --> 00:16:56.920
and the same problems came up

00:16:56.920 --> 00:17:00.533
in these tech company settings.

00:17:00.800 --> 00:17:03.240
And so, from my perspective,

00:17:03.240 --> 00:17:05.420
it's the same kind of --

00:17:05.420 --> 00:17:07.600
I was talking to people 
doing empirical work.

00:17:07.600 --> 00:17:09.700
I tried to kind of look
at these specific problems

00:17:09.700 --> 00:17:13.370
and then try to come up
with more general problems,

00:17:15.110 --> 00:17:18.230
reformulating the problems
at a higher level,

00:17:18.500 --> 00:17:22.900
so that I can think about solutions
that work in a range of settings.

00:17:23.400 --> 00:17:24.933
And so from that perspective,

00:17:24.933 --> 00:17:27.570
the interactions 
with the tech companies

00:17:27.570 --> 00:17:30.300
are just very valuable
and very useful.

00:17:31.700 --> 00:17:35.204
We do have students now
doing internships there

00:17:35.204 --> 00:17:38.516
and then coming back 
and writing more interesting theses

00:17:38.516 --> 00:17:43.400
as a result of
their experiences there.

00:17:44.600 --> 00:17:47.020
- [Narrator] If you'd like to watch
more Nobel Conversations,

00:17:47.020 --> 00:17:48.200
click here.

00:17:48.200 --> 00:17:50.500
Or if you'd like to learn
more about econometrics,

00:17:50.500 --> 00:17:53.100
check out Josh's 
"Mastering Econometrics" series.

00:17:53.700 --> 00:17:56.720
If you'd like to learn more 
about Guido, Josh, and Isaiah,

00:17:56.720 --> 00:17:58.300
check out the links 
in the description.

00:17:59.036 --> 00:18:01.495
♪ [music] ♪