YT YouTube2Text

YouTube2Text → Transcripts → The Political Economy of Policy Experimentations in China with Prof. David Yang

The Political Economy of Policy Experimentations in China with Prof. David Yang — Transcript

by Stanford Center on China's Economy and Institutions · 14,329 words · 2,477 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:02welcome back
0:03david uh i don't think uh i need to
0:05introduce david because
0:06i think everyone here knows who he is so
0:09he's our own product
0:11and we are so proud i mean he's teaching
0:13at
0:14howard at the moment he is the young
0:16star i think real star
0:18in the whole field of china studies i
0:20mean in the younger generation he's
0:22really young i think
0:23but he has published widely i mean
0:26especially on
0:27political economy i behave in
0:31experimental economics history and
0:33cultural economics
0:35and i think he's also a
0:39a risk fellow at mber and global global
0:42scholar at the canadian institute
0:44for advanced research has published i
0:48think in
0:48other top econ journals like jpe query
0:52ar i think he's he was student here
0:55actually
0:57so i think uh uh without further ado
1:01i think i'll turn the stage to save it
1:03uh it's all yours
1:05all right thank you so much for for
1:07having me i would very very much
1:09sort of would like to join you guys in
1:10california but until then i
1:12would like to entertain you with with
1:14some of our some of our latest sort of
1:16uh research project
1:18i'm very much looking forward to your to
1:19your thoughts and feedback
1:21okay great um so so this is a a very
1:23much work in progress
1:24uh join with chava uh from from chicago
1:27and we'll try to
1:29sort of describe and and understand uh
1:31the
1:32sort of how policy experimentation has
1:34have been conducted in china over the
1:36last couple of decades
1:37and then hopefully sort of this is very
1:38much a disruptive work and try to
1:40generate some new
1:40facts and new insights uh that can then
1:43help us understand you know to what
1:44extent china's is
1:45growth and reform is attributed to the
1:48to the systematic policy experimentation
1:50okay uh so let's start by by stepping
1:53back a little bit
1:54and thinking about sort of policy
1:57learning
1:57and policy explanation more broadly
1:59speaking now determining which policy to
2:01implement is one of the most important
2:03role that governments around the world
2:05are
2:05are trying to deal with and many
2:07governments exclusively or inclusively
2:10are engaged in various forms of policy
2:11experimentations
2:12uh by that i mean sort of trying things
2:14out uh
2:15and sort of you know some might involve
2:17some very rigorous sequence of
2:19tries and errors and coupled with
2:21evaluation and so on
2:22some might even involve randomized
2:23control trials at the government level
2:25uh but by and large sort of try to
2:27figure out in a very uncertain space of
2:28policies to figure out which works
2:30is one of the key functions of good
2:32governance
2:34now the systematic policy expectations
2:36that's happened in china since 1980s
2:38are quite unparalleled in the breadth
2:40and depth of duration regarding sort of
2:42this
2:42this endeavor the government is the
2:44chinese government has been taking and
2:45many
2:46sort of in the field of political
2:47economy or among
2:49scholars of chinese economy has even
2:51sort of attributed
2:52to the the fact that sort of the china's
2:55continued
2:55growth and smooth transition and reform
2:58is due to
2:59the various systematic past explanation
3:01the government has been measuring
3:02okay so what we want to do in this
3:04project is try to bring some data to to
3:06to to this issue uh try to describe and
3:09understand the positive communication
3:10that china has been contacted over the
3:11last four decades
3:13so so i want to be clear up front that
3:14sort of this is very much
3:16sort of a project descriptive in nature
3:18and
3:19what we want to do is try to try to try
3:21to look at sort of all the policy
3:23explanation that that took place
3:25and hopefully to be able to highlight
3:26some of the features uh that
3:28sort of even in a very successful case
3:30such as china
3:32policy completion is not something
3:33that's that easy to do
3:35because it oftentimes faces complicated
3:37political uh
3:39environment that can generate frictions
3:40that can generate constraints
3:42okay that will become hopefully a bit
3:43more uh clear in
3:45a second so so let me give you a big
3:47picture of what we want to do in this
3:49in in this particular project so part of
3:51the disruptive
3:52uh sort of exercise that we will embark
3:54on is
3:56to bring some new data so we want to
3:58collect sort of as comprehensive as we
4:00can
4:01on all the positive expectations that
4:03initiated by the chinese
4:05government over the last four decades
4:07and this is going to cover about 139
4:09ministries and commissions okay so
4:11pretty much
4:11all the uh the ministries uh at the
4:13central government level we're going to
4:15hopefully capture those okay i'm going
4:17to start the the
4:18the talk with yeah so a lot of the
4:22experiments were conducted at the local
4:24level
4:25and started from village township county
4:27right but this is our central
4:29uh policy experiment so so so so yes and
4:33no so so what we're going to be able to
4:34see is that
4:35as soon as a policy gets to the level of
4:37if the central government say
4:38this is something we want to try and
4:40they'll start to sort of
4:42playing the coordinated role of these
4:43are the places that that should try it
4:45out
4:46this will be included now what's not
4:47going to be included you're going to see
4:49in a second our essential process that
4:51local you know village county labor
4:53government try
4:53try it out and sort of get killed before
4:56you ever reach the central government
4:57level
4:58uh that's not going to be captured here
4:59but if the local government tried out
5:01and the central government see this is a
5:02good idea and they say okay the rest of
5:04the
5:04half the country tried it out those are
5:06will be included so so
5:08we don't take a particular stance on the
5:10very you know the origin of those pause
5:12ideas
5:13where does that come from uh so yeah so
5:15i'll be i'll be a bit more clear about
5:17what's included uh here and
5:19yeah just i know you're probably gonna
5:21say it but what's the
5:23what's the key word i mean i don't see
5:26the word experimentation very much
5:28um is the key uh it's one word
5:32so you know in chinese there's there's
5:35it's a blend of words like experiment
5:37experimentation
5:38trials it's oftentimes blending in in
5:40one word uh but but we would
5:42we would try our best to to cover
5:44anything that's had the idea of
5:46shitty and pilots and those things or we
5:48can check here
5:49okay okay so i want to i want to start
5:51with after i sort of briefly show you
5:53the data
5:53and what's covered and and also
5:55importantly what's not covered
5:57here i want to start by by just fixing
5:59some ideas uh to go through one
6:01particular experimentation
6:02which is a fairly large one an important
6:04one which is expectation on county
6:06physical empowerment
6:07so i want to use this as a case to to
6:09demonstrate typically what the structure
6:10of
6:11of experimentation looks like in the
6:12space and this is also one of the neat
6:15uh case where we can look and we can we
6:17can try to
6:18closely trace the draw out and and sort
6:21of and impact the expectation
6:23uh and this is what generally some of
6:24the facts that we're going to use to
6:25motivate some of the
6:26more broad exercises we're going to do
6:28so in this particular explanation
6:30what i'm going to show you is that the
6:32early waves of the
6:34of the countries this this is
6:35expectation at the county level the the
6:37early waves of the countries who joined
6:39the experimentation
6:40experience very large positive effects
6:42from from experiencing this physical
6:44empowerment reform
6:45that's the the new policy that's been
6:47tried okay and then the
6:48the treatment effects if you if you will
6:50flipped towards the lateral part of the
6:52experimentation that the experiment if
6:54anything actually hurt the county would
6:55join the new
6:56exact same policy towards the latter
6:58part of the reform okay and i'm going to
6:59show you that
7:00this flipping or very large
7:02heterogeneity in the in the
7:04in the new policy treatment effects is
7:05quite largely due to the underlying
7:07slack
7:09the experimentation size is very
7:10possibly selected in the beginning and
7:12it's become marginally negative selected
7:13towards the end
7:14okay so that's gonna sort of you know
7:16fixing our ideas and one of those
7:18then the rest of the paper is gonna zoom
7:19out to the rest of the experimentation
7:20and we're gonna ask three questions
7:22first is that if we look at all the
7:23policy conditions that's conducted in
7:25the last 40 years in china
7:27on average are the experimentation sites
7:29representative in their selection
7:31uh and i'm going to show you some
7:33evidence that there's
7:35widespread possible selection of
7:36experimentation sites
7:38even if you use very conservative tests
7:40we can very confidently reject
7:42at least 40 percent of the
7:43experimentation that's conducted
7:45significantly deviates away from
7:46representatives and they're they're
7:48deviating typically more in the direct
7:50possible selection
7:51so david you're probably going to tell
7:53us this too
7:54are you surprised that it's only forty
7:57percent
7:58or do you think that that's really high
8:01at forty percent i mean i
8:03i you would have asked me i would have
8:04said ninety percent
8:06but uh yeah um so so it depends on so
8:08this is what i mean by conserving
8:10in the sense of you know almost all
8:12excellent patients
8:13uh and ninety percent is actually
8:14remarkably correct to the actual number
8:16almost ninety percent exponentially we
8:17can we can see their possible selected
8:19it's just that you know some of them you
8:20can you can you can attribute that to
8:22noise or randomness and want to
8:24put a confidence interval on it and then
8:26for about 50
8:27of those we cannot really tell whether
8:29it's just by chance but
8:31the rest of the if you see an
8:32explanation with the t-stats of between
8:34treatment control t
8:35slash equal to 10 uh it's unlikely to
8:38use due to chance
8:39so so that's what i mean by
8:40conservatives okay but by and large
8:42you just think about i'm going to show
8:43you some graphs you should think about
8:44almost
8:45every single experimentation that is
8:47possibly selected now that might not be
8:49an issue that might be optimal
8:50uh so i'm going to talk our second
8:52question is like what drives this this
8:53underlying deviation from developers
8:55and and the first actors i want to do is
8:57that can we justify this positive
8:59selection
9:00using recent experimental literature to
9:03think about if we incorporate learning
9:04related sort of
9:06objectives and very sophisticated
9:08optimal experimentation
9:09models can we justify the degree of
9:12possible selection that we see
9:14okay and then the answer pretty much is
9:15it's no the sort of a lot of these
9:17recent models
9:18will predict what suggests possible
9:20selection is something that's actually
9:21optimal but very rarely you would
9:23you would be able to see optimal design
9:24that's ex that's possible selected at
9:26the level that we see
9:28empirically okay and then in fact i'm
9:30going to show you some evidence that
9:32almost
9:32half of the possible selection that we
9:33see can be attributed to some
9:36type of misalignment of incentives in
9:37the sense of central government to some
9:39extent really wants something closely
9:40representative
9:41and as they move down to the political
9:43hierarchy it gets more and more
9:45deviating from policy selection that the
9:46local government
9:47provincial government who handle the
9:48experimentation starts to do the
9:50explanation only in the richest counties
9:52for example
9:53okay so i'm going to show you a couple
9:54of examples uh
9:56of such such political misalignment
10:00and finally the last question i want to
10:02ask is why do we care what's the
10:04consequence of such deviation from
10:05brilliantness if
10:06if anything if central government can
10:08can fully account for the possible
10:10selection
10:10uh it may not be a big issue but also
10:14the first thing i want to show you is
10:15that the type of signals the type of
10:17information you can gather from
10:18experimentation
10:19really hinges on the underlying
10:21selection of the sites so to extend that
10:23the the experimentation sites are
10:25massively possibly selected
10:26that's that's going to be that's going
10:27to bias the signals into government gets
10:30and central government is not sort of
10:32fully sophisticated about about devising
10:34that
10:34so i'm going to show you some evidence
10:36that the experimentation that are
10:37possibly
10:38that are more possibly selected which
10:40sort of almost by definition is going to
10:42generate more possible signal
10:43are more likely to end up becoming
10:44national policies that get rolled out to
10:45the entire country
10:47and this also gonna generate a set of
10:50sort of patterns that
10:51those policies and sometimes when they
10:53try it out in rich places and that they
10:55work
10:55the central government tried it out
10:56those are going to be to have those
10:58policy potentially tailored to the
10:59richer places that that
11:01that starts to to make the policy space
11:03to be more
11:04uh favored uh towards developing regions
11:06because because these are the the powers
11:07that
11:08end up getting getting blocked and
11:09didn't specifically sufficiently try
11:11to try those policy up in the poor
11:13places before growing up
11:15okay so so this is the the the the
11:18the road map uh any other questions
11:21before i
11:22delve right in to to show you the data
11:25oh
11:25hi debbie yeah yeah i wonder uh
11:29what do you think of the following
11:30explanation for the division from
11:32representatives
11:33so supposed to improve economy we need
11:36those three elements
11:37to be good so now we want to test
11:41whether one reform in whether whether
11:44reform in one element works
11:46so we need to make sure that the other
11:49two elements are good
11:50right so we go to the area
11:54where the the other two elements are
11:56good so in that sense we
11:58make the experimentation not
12:02representative right but it is
12:05for me to for us to do the experiment
12:08yeah yeah no i i think that a slightly
12:10different way of asking your question is
12:11that
12:12which i will get to it in a little bit
12:14you know maybe you just get a very noisy
12:16signal of trying out the policy in a
12:17poor place because they couldn't even
12:19implement the reform and then what's the
12:20point of trying it
12:22you just don't learn much and maybe the
12:23policy only works when you have a good
12:25sort of
12:26local state capacity and so on so that's
12:29that's fine and then
12:30we're going to sort of consider that a
12:31little bit but what becomes sort of
12:32problematic is that
12:33if if the whole new experimentation
12:36space is conducted
12:37in which places with the precondition
12:40that the new policy has proposed to work
12:42then and those costs are more likely to
12:43become national policy because the
12:45positive signal the central government
12:46gets
12:46that still means that you know poor
12:48places get policy that's not suitable
12:50for them
12:51and and that could hurt uh uh sort of
12:53you know poor regions in china and sort
12:55of
12:55in something like simple simplest
12:57takeaway is that
12:58almost every new policies might work in
13:00shanghai it might not work for gansu
13:02but given that the process is trading
13:04shanghai gets real to the respite
13:05contribution games will get a lot more
13:07positive taylor for shanghai but not
13:08taylor for the poor places
13:10and that can be problematic okay and
13:12i'll get i'll get to some of this
13:13towards that towards that
13:15so david let me talk about this but how
13:18is this related to
13:19the rcts which is not very public in
13:22china right
13:23yeah deposit making yeah
13:26so so i so you know we it's it's a
13:29broadly speaking you can you can i'm
13:31going to talk a little bit about about
13:32how this relates to to rct there are
13:34some similarities of this look
13:36one of the biggest critic varsity is
13:38external validity uh
13:40uh uh of some sort and here sort of in
13:42the for this
13:43i'm not saying that this plus condition
13:44the chinese government are doing is an
13:46rct
13:47but you can think about rct or our test
13:49for rendering this is essentially a
13:50benchmark in them against rcgs
13:52uh but we're still going to come back to
13:53the point of accidentality for the
13:55experimentation because even if you
13:56achieve
13:57internal validity actually let me hold
13:59off on outlands for a little bit because
14:00i want
14:01i'm going to get through this for the
14:02case study of what what exactly is the
14:05issue here
14:05regarding actionability that even if you
14:08run an rct
14:09uh uh at the local level it might not be
14:11able to solve unless you sample
14:12uh your experimentation and sample is
14:15the whole country
14:16otherwise you're still going to get to
14:17the issue of accidental ability
14:20hopefully you'll become more clear in
14:21two slides when i want to show you the
14:23case study
14:24okay so so let me let me show you very
14:27briefly
14:27about the data and how to chat more
14:29about the about the data source uh sets
14:31uh uh later if anything's unclear
14:35so as i said we our goal is to to try to
14:38measure as as comprehensive as we can
14:41about policy competitions
14:43and we cover sort of the time ranges
14:45that cover from
14:461980 to 2020. uh our primary data
14:49sources built on a database on on the
14:51government's official documents
14:53uh uh so overall sort of you know we
14:56screen out
14:56uh sort of partially manually imparted
14:58automatically on all the keywords that
15:00we think are associated with
15:01experimentation
15:02studio and shield or whatever and that
15:04give us about 20 000
15:06documents among those 20 thousand
15:08documents but 4 000 of them
15:09are documents at the central government
15:11level now those are important in the
15:12sense of those are documents that
15:14essentially
15:14tell from a central government
15:15perspective what's the goal of the
15:16experimentation
15:17uh what are the ministries that are
15:19involved in the expectation how it will
15:21be rolled out
15:21uh if it doesn't involve multiple ways
15:24which countries get selected
15:25does it allow for voluntary
15:27participation in expectation
15:28house gets uh uh evaluated on all that
15:31stuff at the meta level
15:32for expectation okay and then we're also
15:35going to be able to match those with
15:36about 15
15:37000 local government documents uh for
15:40those who are familiar with
15:41it is essentially that if a local
15:42government is if a local region is
15:44called to be a part of temptation
15:45local municipal government for example
15:47is going to respond to the central
15:48government's calling by issuing their
15:50own
15:50uh policy document to say this is how
15:52we're going to do the explanation
15:54after we've been called to be part of
15:55the experimentation sites and then from
15:57the local government
15:58documents we're going to see the
15:59implementation details uh
16:01and sort of essentially also as a as a
16:02crosstrack of of of
16:05the sites that get selected on the
16:06central cover okay
16:08so we're going to be very conservative
16:10about sort of to the extent that there
16:11if there are minor policy evolutions of
16:13the policy changes name but it's
16:15still wrapped about the same content
16:16we're going to call them the same policy
16:18even if we evolve over time
16:19so this gives us about roughly 850
16:22policy expectations initiated by 139
16:25ministries
16:25and commissions over this period now
16:27importantly and in relation to the home
16:29business question earlier
16:30very early stage bottom-up
16:32experimentations that started in some
16:33counties
16:34and they got killed very quickly and
16:36never reached the central government
16:37level because it's just a bad idea
16:38those will never show up in our datasets
16:40and we're sort of trying to find some
16:41way to at least get a sense of what
16:43those space would look like
16:44now the policy that initially by the
16:46local government and central government
16:47say this is a great idea let's let's try
16:49it out
16:49those will be part of our but but why
16:51didn't you and
16:53it seems to me i mean when we you know
16:56our group does a lot of rcts
16:58right and and we'll do rcts that totally
17:02fail
17:03and we learn a lot about those
17:06right they don't get upscaled but we
17:09don't upscale them right because
17:11so is there any reason why you didn't
17:14include those
17:16so the simple unsatisfying answer is
17:19that
17:19it's just hard to observe them because
17:21local government when they first try
17:22something they might not
17:24put that into it into into official
17:25document uh so given that's the space
17:27that we can observe so
17:29so in fact what we're going for right
17:31now is that if the local government
17:32tried it and there's they're not
17:34sufficiently ashamed of it to put it in
17:36a in an official document that's to give
17:39us some hope to
17:40want to capture those but for now it's
17:42just that if
17:43if those trials don't exist in the
17:45policy document space
17:47it's hard for us to know unless we sort
17:48of go out also to all the newspapers and
17:51and so on so so that so that's another
17:53way to answer the question
17:54i think for sure that for the policy
17:57that enter into this level of central
17:59government issue a document to say this
18:00is something i want to try it out
18:02the the average prior of how likely the
18:04policy is going to success uh to this
18:06to be successful is already very high uh
18:08they already killed off a lot of the
18:09crazy ones or the ones that clearly
18:11are are not going to work so so so
18:14that's the level of selection
18:15uh well we're not going to be dealing
18:16with now there's another whole set
18:18that eventually will hopefully compare
18:20it against is there's also tons of
18:21houses in china that
18:23go out to the whole country without
18:24going through any experimentation uh
18:26that potentially is the questions
18:27regarding the very consequences
18:28something you want to do without any
18:29uncertainty
18:30uh so how that compared to the to this
18:32middle range of
18:34sufficient uncertainty for them to try
18:35out uh this is something that's more
18:37doable for us to do with some benchmark
18:39uh against uh there it but this uh i
18:41think this is
18:42also important for people to know china
18:45as well
18:46most policies starting from bottom up
18:49are top down in china
18:51oh it's a mix of both uh so so
18:54so my view is it's a mixture of both in
18:57the sense of
18:58the very so it gets back to the question
19:00of how the central government gets the
19:02idea of these policies to try
19:04and i think a lot of these ideas are
19:05sort of initiated by the local
19:07government
19:07and then they start pushing very
19:09strongly and then but
19:10the speed of you know when uh when when
19:13the local government tried out a new
19:14idea
19:15and then it get raw combat central
19:17government that period typically is less
19:18than one year
19:19so then once they get to central
19:20government level you get this massive
19:22sort of
19:22top-down operation of okay this is
19:24something i want to start to try out in
19:26these places
19:27so i think it's it's a combination of of
19:29of top-down versus bottom-up
19:31and some of the recent actually quality
19:32work in this literature actually sort of
19:34put it feels like too much weight on the
19:35bottom
19:36on the on on the bottom up part of this
19:38where i think sort of bottom up
19:40reach to the top down version and very
19:41quickly uh and at the whole process i
19:44think but but
19:45also to be to be clear i think what we
19:47are so far to observe is is
19:49is selecting on a set of positive reach
19:51the top down
19:52level so how many of those top-down
19:55version of this designation looks like
19:57uh oh and also hopefully i will by the
19:59end of the movie i'm going to
20:01discuss a little bit about how this
20:03compared to policy experimentation
20:05in the federal system where almost
20:08everything is from bottom up
20:09uh and how the chinese case compared to
20:12that
20:12okay so what i'm showing you in this
20:14figure is just the kind of policy
20:16experimentation
20:17that the central government initiated
20:19each year for the last 40 years
20:21okay so so you see a pattern that's just
20:24that's
20:24that's important uh in the sense of like
20:26the population
20:28sort of rolled out gets increasingly so
20:32and reach the peak in the early 2010s
20:35and it's not increasing fairly rapidly
20:37for by about 50 okay uh and then this
20:40sort of english skill pattern is
20:41actually fairly robust regardless of how
20:43you how you measure them and various
20:44sort of assumptions you can we can take
20:46it's going to be a little bit off the
20:47scope for for today's talk but we would
20:49do a little bit more about this in
20:51in in the paper like how to explain this
20:53this this straw
20:54it's your first extinction uh sort of
20:57instinct might be this is a
20:58season painting effect which you know
21:00part of the part of the general trend
21:02maybe
21:02but i think at the ministry level what
21:04we see is that this coincides with the
21:06with the vertical reform of state
21:08bureaucracies in china by by each
21:10ministry that when they're essentially
21:11moving from a horizontal flag was
21:13flexible and from organism structure to
21:16a vertical
21:16structure which is much more efficient
21:18and that makes it less flexible to
21:20conduct experimentation
21:21and we see that as each ministry is
21:23going through this virtualization reform
21:25the number of experimentation that get
21:26initiated by the ministry start to drop
21:28and i think part of this job is it's due
21:30to sort of they're essentially weeping
21:31the efficiency of ministry
21:32but letting go of some of the
21:33flexibility that the early state
21:35bureaucracy had
21:36okay so let me letting this sort of
21:39inverse you aside
21:40uh what i want to what i want to do now
21:42is to show you a um a case
21:44about a particular experimentation and
21:46show you some of the stylus facts that
21:48that we really would like to
21:49then generalize to the rest of the
21:50expectations so
21:52so this can be fiscal empowerment reform
21:54uh for those of you who are
21:56familiar with is it's a at least an econ
21:59is actually
21:59a fairly uh explanation
22:02it becomes a very large one uh the
22:05status quo is that
22:06provincial government sort of you know
22:08administrative professional government
22:09professional government administrator uh
22:11uh county government and many sort of
22:13has attribute the county
22:14government sort of failure to deliver a
22:16lot of the growth potential because
22:18because of its lack of autonomy for
22:19these physical resources because there
22:21are two layers above them
22:22and and to partially to guide the
22:24national policy to foster country level
22:26growth
22:26since the since 2002 actually sorry a
22:29type of year 2002
22:30the central government has begun
22:32experimentation on giving
22:34county government more physical autonomy
22:36sort of in a nutshell that reform is
22:38essentially letting the country
22:39government bypass the
22:40the provincial uh bypass the municipal
22:42government and the provincial government
22:44can directly allocate fiscal resources
22:45to the county
22:46uh so that it doesn't get rid of the
22:48middlemen of potential general
22:49distortion so that's that's reform it's
22:51delivery meant to increase the county's
22:54works okay
22:55it's a large explanation so over the
22:57course of two to ten years roughly
22:59about half the countries in china
23:00participate in this expectation
23:02by by various ways okay in the typical
23:04sort of nature of exploitation uh
23:06whereas different batch of concept
23:09warnings throughout the period
23:10now we don't know exactly how central
23:13government evaluates this policy but as
23:15an applied economist
23:16no uh we we can we can estimate sort of
23:18event study specifications to look at
23:20what this particular reform did to each
23:22country who joined it
23:24uh and we can control for country fix
23:25facts we can we can look at the company
23:27before and after to join the forum
23:29controlling for provincial uh by by year
23:31fixed fast to get rid of the time
23:33the promise specific time trend and this
23:35is sort of one of
23:36you can call it this is the best thing
23:38that the applied economist can do to
23:40look at this
23:41so i'm gonna i'm gonna estimate this uh
23:43event study uh design
23:44to look at what this particular county
23:47uh empowerment reform did
23:48uh to the to the to the to the local
23:50account for those who joints
23:52okay and i'm going to do this accident
23:54separately first
23:55for the countries will join the reform
23:57in the first half of the batch so those
23:59who join the particular experimentation
24:01before 2007
24:02and then i'm going to do it separately
24:03for those who join it in the second
24:05ladder batch
24:06like after 27. okay so i'm probably here
24:08the aesthetic coefficient
24:10which you see here is a fairly standard
24:11plot you see typically in the events
24:13like design where there's no pre-trend
24:15uh uh
24:16and then as soon as the country joined
24:17the reform in the first year and so on
24:19uh it started to receive a boost in the
24:21local growth uh just as the central
24:23government should intend to
24:24to further further reform to generate uh
24:27to the to the
24:28to the particular county okay so that's
24:30our good news for the central government
24:32if we do if we repeat the exact same
24:34exercise on the latter batch
24:36of the after of the contact with 204 you
24:38start seeing
24:39no pre-trend before but the content for
24:41who are not experiencing the exact same
24:42policy change
24:43start to get hurt uh by the by the
24:46fiscal uh in
24:47empowerment towards towards the to the
24:50towards end
24:51okay in fact if anything you know i can
24:54do this a bit more so
24:56uh continuously so we can call each
25:00each year a bachelor of currency join
25:02the experimentation
25:03and we can call that sort of a cohort of
25:05constitutional exploitation as
25:07before 2005. and i'm going to include
25:10one experimentation plus times one year
25:12and then we can still estimate even
25:14study in those moving narrow windows
25:16and showing you here the average
25:17treatment effects of this reform
25:19for for the for the cohorts of the
25:21entire reform uh
25:22expansion that naturalized what you see
25:24here is that the reform really had a
25:25positive effect on those counties at the
25:27very beginning
25:28of their form and then the effects of
25:29the reforms start to decrease
25:31over time and it starts to become even
25:33negative towards the very end of the
25:35reform
25:35so that's just another way to say there
25:36is very strong heterogeneity in the in
25:38in the exact same policy
25:40and we see sort of almost a linear
25:41decrease in in a
25:43in an effective effectiveness reform to
25:45the policy to the content that joins
25:47it now what i'm going to show you next
25:49is that
25:51we can look at for each year look at the
25:55pre-experimentation characteristics of
25:57those counties either their
25:58pre-experimentation gdp per capita or
26:01any measurement of socio-economic status
26:02and we can compare the
26:04the countries who join the reform versus
26:06our experimentation versus the county
26:08who are part of the control group whose
26:09central government didn't select which
26:10one in reform
26:11and then we can compare those two and
26:13then we can we can now give us a
26:16t stats uh and you can think about the t
26:18size as a summary stats of how deviating
26:20from representing this or randomness
26:22the the size section is what i'm talking
26:24about here
26:25is the t-stats for each of the cohort of
26:27the console join the reform
26:28over time okay and you see a pattern
26:30that's actually fairly similar to the
26:32treatment effects that you showed
26:33earlier where
26:34in the very beginning of the reform you
26:35have a t-stats of 10. that essentially
26:37is giving out the reform
26:38or the explanation only to the richest
26:40places in china
26:41and you get gradually lower the degree
26:43of decreasing degree of positive
26:44selection
26:45uh still on average more positive
26:47effective and and towards the very end
26:48it starts to get marginally
26:50selected okay so
26:53david what's the end for each of those
26:56uh
26:56years i mean is is 2011 is that like
27:012 300 and 2003 is one
27:05i mean so they're fairly balanced uh i
27:07don't know the number on top of my head
27:08so there are a thousand of them over the
27:10course of 10 years on average is about
27:11100 but they're not sort of easily
27:13distributed so
27:13on average it's about sort of a little
27:15less than 100 or around that for
27:17each so most most of these national
27:20policies are
27:21are rolled out in a hundred counties not
27:23not really so so this is one of
27:25the the reason why i want to generate
27:26this so we couldn't do this very sort of
27:28sort of sharp kind of cohort by cohort
27:31analysis
27:32for many of the policies because they
27:33they try out in 10 places and then
27:35there's not much we can do for two stats
27:36year by year but this is one of the big
27:38ones where
27:38this part is particularly uh uh sharp
27:41okay
27:42so that large heterogeneous treatment
27:44effect if in fact
27:46if we're just imposing a linear
27:47structure on this to think about sort of
27:49you know that
27:50we can estimate how heterogeneous the
27:52treatment effects of this
27:53policy with respect to the
27:55pre-experimentation economic conditions
27:57and then we can project the policy
28:00effect
28:00on the places that didn't get reform had
28:03the reform being drawn to the entire
28:04country
28:05and showing you the simulated sort of
28:07policy effect
28:08for the entire country for this verb
28:10this particular policy
28:11and what you see here is that large
28:13heterogeneity
28:14half the country would benefit from this
28:16reform have to come and some of them
28:18benefit very large
28:19from this form the other half would lose
28:21so the net effect of the cost is
28:22actually if anything very close to zero
28:24but there's very large
28:25distributional effects okay so so this
28:28is something this
28:29is something that i hope for hopefully
28:30sort of know i i will be able to sort of
28:32take
28:32take take uh uh with you and then
28:35convince you for the for the rest of the
28:37expectation i want to look at
28:38systematically is that
28:41the type of given the large car sample
28:44heterogeneity in a lot of these policies
28:47the type of signals that central
28:49government are able to get from
28:50instrumentation
28:51is very much depend on the underlying
28:54sort of site selection
28:55uh you know you can get for this
28:56particular cause you can get positive
28:58average treatment effects of the
28:59experimentation but net zero
29:01uh when it gets royalty to the entire
29:02country and then to back to sort of
29:04hombing's question uh earlier
29:06this would suggest econometrics sort of
29:09site selection bikes
29:10uh uh and and in particular what makes
29:13important
29:14important is that this is abstract away
29:16from even from
29:17internal validity because when we use
29:18event society to evaluate the policy
29:20there might be additional biases if the
29:22central government are
29:23less sophisticated about the way to
29:24analyze the policies
29:26so even if you just do everything you
29:28can uh the best you can i guess in this
29:30case to analyze the positive effect
29:32uh the the type of signals you can get
29:34it it can still be biased just because
29:36the the
29:37the underlying sense like site selection
29:39uh and that's coupled with
29:40the the heterogeneous treatment effects
29:42okay so that's sort of part of what i
29:44mean by axiomality
29:45that that's hard to get around with uh
29:47and even even sort of
29:49thinking about sort of in the rct
29:51environment
29:52okay so so so those sort of no what i
29:55want to do next is
29:56no do this pattern that show you in the
29:58content physical empowerment reform
30:00uh holds more generally uh for the for
30:02the for the rest of the policy that we
30:04uh we uh we we see in
30:09so do you know if the government if the
30:11central government
30:12finds the result of the experimentation
30:15is not good
30:16would they reverse uh the reform
30:20so so the typical so there so so so yes
30:23and no
30:24so very very very central government
30:26come out to declare this is a big
30:27failure and let's scratch it
30:29so the typical way of a failed reform is
30:31that they
30:32should re reveal the reform every five
30:35years
30:36and the if the fifth year after the
30:38reform the reform didn't get renewed
30:40uh uh that's essentially sort of
30:42implicitly the failure for the reform
30:43because it didn't get well to the entire
30:45country
30:45and and oftentimes it means it'll go
30:47back to the status quo uh
30:48for a lot of these policies okay you see
30:51some
30:52very small number of cases where if the
30:54central government certainly sometimes
30:55for political sort of persuasion reasons
30:57i think want the policy to fail because
30:58they can persuade the
31:00the political opponents they roll the
31:02policy out and get into just one year
31:03and they realize the policy is gonna
31:04fail
31:05and and then the end of the end the
31:07whole thing uh party i think just to
31:08show some evidence for the opponents of
31:10the policy but that's a very small share
31:11of the policy we saw
31:13okay so let's let's continue with with
31:15the with the exercise
31:17uh i'm gonna now zoom out to this
31:19particular experimentation and look at
31:20all the policy conditions the
31:21the 850 of them uh and then i'm going to
31:24repeat the exercise the first exercise
31:26to look at
31:26whether the size selection is
31:27representative on average okay
31:29so the way i'm going to do it is that we
31:32can do it multiple ways but this is one
31:33particular way of doing this
31:35for each experimentation we can compare
31:38the
31:38the the treatment sites or the sites
31:40that are being part of the condition
31:41with the control sites uh thinking that
31:43in the language of
31:44rtt compare a range of
31:46pre-experimentation characteristics
31:51as a graph but it doesn't matter which
31:53one these are all very correlated okay
31:56then i'm going to use the the simple
31:57t-stats for the comparison as a summer
31:59stance the bigger the t stats
32:01adjusted for the underlying sort of
32:02degree of freedom because different
32:03explanation has different sort of
32:04treatment and control they're very
32:05rarely balanced
32:06in terms of the size but it wouldn't but
32:08but after all those adjustments you can
32:10you can interpret
32:11the t stats as a bigger the number means
32:13the more positive
32:14selected explanation is that zero means
32:17they're exactly identical between the
32:18between your treatment and control sites
32:20okay so this is the distribution of the
32:24t stats
32:25across 850 expectations that we see okay
32:28this is back to scott's question uh
32:30earlier like
32:31vast majority of the expectation uh has
32:33t-stats above
32:34zero so they're possibly selected uh a
32:37fairly long tail
32:38on either ends but they're they're very
32:40right share and it will be very
32:42conservative look at the vertical dots
32:43here that's that's the confidence
32:44interval where we can reject
32:46by but with ninety percent confidence
32:48interval that this is not due to random
32:49chance
32:50then we can write at least forty percent
32:51of experimentation to be to be
32:53to be sort of uh sort of
32:55representatively
32:56allocated and this share uh uh you know
32:59the
33:00vast majority of being possibly centered
33:02and and we can roll about 40 percent of
33:04the the size being processed
33:05this is actually very robust uh using
33:07various test procedures
33:08uh that we have and and and
33:14david uh there's a question that just
33:16came came into the chat room
33:17um and it's um that
33:21are they using one or two or
33:24you know uh 50 or 75 counties over and
33:27over and over and over again
33:29and um so is shanghai part of you know
33:32uh
33:32750 of these experiments but you know
33:35that's that's an exaggeration but but
33:38do you see these experiment counties
33:40being used over and over
33:42so we so so so yes and no so by and
33:44large we don't see that
33:46except for those uh in chinese that's
33:49the the that's the sri and
33:50or the essentially the reform sort of
33:52demonstration
33:53that just gets you know shanghai gets a
33:55lot of the policy exploration of various
33:57sorts
33:58but those are sort of you know a handful
34:00of cities essentially and apart from
34:01those who are just
34:02constantly getting new policies uh
34:04there's essential
34:06revolving door of different countries
34:08get different versions of the policies
34:10you can just give a map of all the i
34:12mean where it started right
34:15whole country yes so if we want to do
34:18this a bit more dynamically and then and
34:20then
34:20trace and then that will be something
34:22that's similar to what i just showed you
34:24uh with the fiscal reform and so on yeah
34:27okay a couple of additional facts i want
34:30to show you for the
34:31development is that we don't see this
34:33change over time so
34:34it's equally positively selected in 1980
34:36then 2010.
34:38there is however sort of fairly large
34:39heterogeneity by ministry or by sort of
34:42policy domain as
34:43you wish but what's interesting is that
34:45you actually you know
34:46one way to get around sort of you know
34:48to justify possible selection is that if
34:50their policy that you know it's not
34:51meant to
34:52learn a national policy you just meant
34:53to target and for the possible mean that
34:56you have more room to have
34:57regional policies than the national
34:59uniform policies that may be more
35:00justifiable to have a positive selection
35:02uh because you're you're not meant to
35:04have randomness in in
35:05in the first place now if if anything
35:08you see that those
35:09national policies the policies that only
35:11reasonable at the national level are
35:12actually more possibilities like
35:14tax policies education policies and so
35:15on and for agricultural policy you
35:17oftentimes see regional policies and
35:18targeted actually might be justifiable
35:20uh you get less of the positive
35:22selection actually if anything on
35:24average but still sort of you know uh
35:25the t stats is still still one they're
35:27not trying even agricultural policy
35:29in the poor uh regions uh still richer
35:32agricultural resources in china okay so
35:35any questions on this before i
35:36move to the second part of the can we
35:38explain the the deviation from the
35:40positive selection
35:41okay so as i said earlier i want to
35:43start i want to do two things the first
35:44is that
35:45sort of if we if we take the optimal
35:48design perspective
35:50what does it take for the first to
35:51justify uh the possible selection that
35:53we see
35:54uh can it be optimal or at least can
35:56optimal experiment design rationale
35:58justify at least part of this
36:00and then we're going to go out to the
36:01optimal sort of uh
36:03rom and to see other distortions
36:05essentially in the political hierarchy
36:07that generates possible sections
36:08okay so so for the for the optimum
36:11design expectation
36:13so the flavor of things that that we do
36:15and i want just to give you a sort of
36:17a glimpse of of of the headline
36:19conclusion is that we're going to take
36:21some of the recent this is
36:22actually a fairly recent literature take
36:24some of the recent literature and
36:26and take some of the modification of
36:28additional sort of
36:29learning related motivation that that
36:31you can incorporate into into designing
36:33the expectation
36:34and oftentimes they predict possible
36:35selection and we can essentially
36:36simulate
36:37uh given all the parameters that we can
36:39estimate from that from our empirical
36:41data
36:45selection kind of level under this uh
36:47motivation
36:48one such motivation is that it's besides
36:50standard learning if the central
36:51government cares about the
36:53subjective expectation and the central
36:55government's prior or their
36:56uh beliefs is very asymmetric uh and it
36:59puts a lot of weight on maximizing
37:01subjectivity residential persuading or
37:03learning about
37:04the the policy per se it can justify a
37:07division from from realization
37:09okay so even if you push this model all
37:11the way to the extreme
37:12if the central government cares a
37:14hundred percent about subjective
37:16expectation and have very extreme uh uh
37:18asymmetric priors
37:20this can only justify a a
37:22experimentation optimal explanation of t
37:24stats on the magnitude of 0.76
37:26okay so that's positive selection but
37:29really sort of much
37:30magnitude smaller almost and then the
37:32possibility that we observe in our data
37:34okay the second type of sort of you know
37:37uh just location process selection is
37:39that
37:39you know we're not in the in in the case
37:41of infinite uh
37:42uh observations every place you try the
37:44policy it's it's
37:46it's a sample that you probably care
37:47about is welfare so it's not about so
37:49you can just hold up that sample and
37:50then learn from the infinite world
37:52so if you if you if the central
37:54government cares about the warfare of
37:55the experimentation sites
37:56that could justify for duration from
37:58possible deviation from randomness
38:01i i guess i'm just um
38:04a little puzzled by considering all of
38:07this under the rubric of experimentation
38:10in other words if the government decides
38:13that it wants to do this
38:16wouldn't it isn't it very smart and
38:18close to optimal for the government to
38:19say
38:20well we're going to select a few uh
38:23key points for trial implementation and
38:26we want
38:27and we we don't really want them to be
38:29representative because our objective is
38:31not to elicit information about
38:33whether the policy works or not is to
38:35elicit information about how effectively
38:37we can roll it out
38:38so does how much does it change things
38:41if we think of this
38:42as a staggered implementation
38:46yeah yeah so the so so so that's that's
38:49a that's a very important question and
38:51part of the
38:52part of the sort of the the core
38:54challenge of of of this project is that
38:56we don't know exactly what the objective
38:58function for this regard we can only
38:59make inference and then see
39:00you know given the various assumptions
39:02it's can you justify that
39:04now the two main issue that that related
39:07to your comments is that
39:09i think it's very likely that especially
39:11for the policies that the central
39:13government has a very high priority that
39:14this is something what we will do
39:15regardless we just need to figure out
39:17the implementation details and
39:18potentially it's better to figure out
39:20those details
39:20in shanghai or in a few places that we
39:23are comfortable trying those out
39:24with and this is sort of a relative
39:26question that come up earlier that's
39:27totally fine
39:28it's but it's but it still is
39:30problematic if that if the policy
39:33i guess i give the same answer as before
39:34if the policy that you end up
39:36tailored in all the information detail
39:38in shanghai
39:40end up sort of hurt then so and those
39:42are the policies that
39:43actually get more likely to go out to
39:45the whole country and if the central
39:47government don't take that into account
39:48that becoming an an issue if the central
39:50bank is fully sophisticated
39:52i i know that i'm trying not to the
39:54house only in rich places and then
39:56when i wrote it out i'm going to de-bias
39:58in all this and then
39:59tailor the policy also a little bit for
40:01the poor places then we would maybe
40:02worry less about that
40:04and hopefully i have some evidence to
40:05show you that they're at least they're
40:06not doing 100
40:07of of those and which which can become
40:11an issue
40:12look i look forward to the evidence but
40:14at the same time just in response to
40:16that i mean if you pick
40:18if you say we're going to roll it out in
40:19shanghai and cue joe
40:21uh that's not going to be representative
40:23but it might tell you what you want to
40:24know
40:25about different paths yeah i agree yeah
40:28so so to something i'm so we're
40:30definitely not suggesting that
40:31expectations should always be successful
40:33you know a field experiment is just
40:34it's equally as a form clip to the
40:36central government now the point is like
40:38central government seems to be just
40:39disproportionately get
40:41success cases because this because they
40:43place the try
40:44uh another way to put it is that maybe
40:46trying policies in shanghai is not
40:48as important as intercom and thought
40:49because almost everything hype will work
40:52and try not any more violence because
40:54this might might be more valuable and in
40:56fact sort of
40:56this is the quote that i'm giving you
40:58know do central government actually
40:59understand representativeness
41:01at some level and that that seems to be
41:03the case the national development and
41:05reform commission the
41:06the the fact that way for example which
41:08which is the the bureau that guides a
41:09lot of the expectation
41:11very exclusively said getting something
41:13close to balanced
41:14sort of sample being nationally
41:16representative is something
41:17they're also of course they're they're
41:19open about potential alternative
41:20objective which i'll talk about in a
41:22second but at least on some level
41:23understand
41:24representativeness is something that
41:25they're they're describing for and they
41:26understand this is something that
41:28that's valuable for explanation now that
41:31should it become a question of
41:33when the national
41:37development if for an experimentation uh
41:39how can end up getting a key stats of 10
41:41for some of these trials okay now the
41:45to give you some of the smoking gun of
41:47what might be going going on here
41:48is that the further you're moving away
41:51from the central government's direct
41:52control over which sites get selected
41:55the the further away you are from uh
41:57from from representatives
41:58so for the future examination where
42:00central government select the province
42:02and then the provincial government
42:03selects the county you get very close to
42:05representatives at the provincial level
42:07and you get massive positive effects
42:09selection at the uh within problems at
42:11the company
42:12well that suggests so the central
42:13government might be wanting to represent
42:14this but that's not the case for the
42:15provincial government
42:17and also sort of for examination that
42:19that allow for
42:20uh for for uh for uh
42:24uh voluntary participation essentially
42:26letting local government have a more say
42:27of which
42:28provinces or which local size guessing
42:30transformation they're much more
42:31possibly selected
42:32than the tradition that's quickly top
42:33down and the central government laid out
42:35the entire plan
42:36uh up front
42:40so for i mean for i mean officials in
42:43beijing
42:44if they want to do something you always
42:47pick
42:47the places that can have an initial
42:50successful experiment right
42:52and then they can report to the top
42:54leader oh this is really good which
42:56i mean square out the whole country yeah
42:59i mean they always have an agenda right
43:00i mean
43:00they won't now it's like a place
43:03randomly
43:05yeah no i completely agree so so so 2010
43:09that's part of the you know i'm going to
43:11tell the story i think
43:12both directions central government has
43:14some incentive to do that using
43:15experimentation as
43:16as as essentially sort of a you know i
43:19can i can essentially manufacture
43:20whatever i want to get
43:21uh from the place i tried the expedition
43:23and use this as my political sort of
43:24evidence for to convince my
43:26my opponents and local government might
43:28do the same i'm going to show you some
43:29evidence that local government
43:31once they know that this is something
43:32central government might want to be
43:33successful local government
43:35really wants to show that it's
43:36successful and that's going to also
43:37generate biases in learning so that's a
43:39very different way of thinking about
43:40central government just want to learn
43:42and then they just literally want to be
43:44very scientifically try things out
43:45uh uh and there's very sort of you know
43:48part part of the the key point
43:50if if you take away from from this talk
43:52is that once explanation are done
43:54in a complex clinical environment uh
43:56you're very likely gonna deviate from
43:58this very scientific modes of just
43:59learning
44:00uh and so on okay so we have a bit more
44:03to say on local side just because we
44:04just don't know
44:05what exactly central government is
44:06thinking uh directly
44:08so from the local government side i'm
44:10going to show you uh in this case
44:12actually it's
44:12not local government per se but anything
44:14below the very top of the of the global
44:16hierarchy
44:17i can show you two particular case about
44:19the particular missile
44:20line incentive uh just to demonstrate
44:22they're not meant to be exhaustive but
44:24they're just to show you a flavor of
44:25of where the margins where versus can
44:28generate this deviation from the
44:30representativeness political patronage
44:32is one so so there
44:33you can think about the story of do the
44:36ministries for example of agriculture
44:38more likely to assign explanation to the
44:41places the province for example that
44:42they're connected to that that have been
44:44previously worked or have
44:45have like plus with high schools okay so
44:47we can use the the one that's used uh
44:49the specification used a lot by the
44:51literature
44:52uh uh looking at sort of the likelihood
44:53of a place guest experimentation
44:55uh uh the the regressive on whether that
44:58particular problem is connected to a
45:00particular ministry
45:01controlling for the province by ministry
45:03of fixed effects that means that all of
45:04our issues come from
45:05a province gets connected to a ministry
45:09not the year one but in year two after a
45:12new minister
45:13gets into the office and gets that
45:14particular promise gets connected
45:16okay so we see a fairly positive and
45:18very large magnitude in the sense of
45:20the the ministry is very much more
45:23likely to sign an explanation
45:24to the provinces that there are
45:26previously have worked really enjoyable
45:27political ties
45:29and to the extent that you get more of
45:30the administrators of china coming from
45:32russian places the richer places end up
45:33then on average more likely to get
45:35explanation
45:36uh than in the poorer places that's one
45:38source of deviation
45:39from from from representing this uh so
45:41can local governments reject
45:44uh the central garments request of uh
45:47experimentation we think yes uh not
45:50exclusive to people but i think in some
45:52sense
45:52there has to be sort of implicit consent
45:54uh in this process
45:55of central government i don't think they
45:57can force maybe at the provincial level
45:59they can exert
46:00more more effort but i don't think they
46:01can sort of literally just say you have
46:03to be part of this convention
46:04and so on typically if there's enough
46:07sort of local resistance they can they
46:08can find the substitutes
46:11so i'm not going to show you be able to
46:13show you much of the resistance i know
46:14here
46:15actually for the for the misaligned
46:16incentive number two i can show you the
46:17opposite side some local government are
46:19particularly eager to join the
46:20expectation because of
46:21for their for the ability to generate
46:23political credits uh uh
46:25you know we can we can look at sort of
46:26whether the current incentives of local
46:28government and korean stand to come
46:31using the standard thing has been using
46:34which is
46:34you know how much more time in your
46:37career
46:37that for the next term uh before you had
46:40to to retire
46:41and and and the longer you have the more
46:43coincidence you have to be able to
46:44generate
46:44political sort of uh essentially uh
46:47positive cases
46:48uh uh and we can use the version just
46:50coming from a discontinuous decrease in
46:52political
46:53incentive because if you're you're
46:54getting very close to to the retirement
46:57and then we can again regress that on on
47:00on
47:00on on on the likelihood of a particular
47:02local government joins the expectation
47:04and you see very large uh uh magnitudes
47:07in this in
47:08in in the sense of for a local
47:10government who have
47:11very high political and local incentive
47:13innocence if they have more room to go
47:15uh in their political career they're
47:17much much more likely to join
47:18expectations especially for the ones
47:19that allow for voluntary participation
47:21and given that on average they're more
47:24likely to come from rural places that's
47:25another source where
47:27central government may at some point
47:28want to be representative but but
47:30any fluctuate that you open for local
47:32government to potentially participate
47:33uh is going to generate more of those
47:36who are particularly motivated
47:37and local engineers to join yeah but
47:40you know if if if it's the incentive for
47:43the
47:44central government person to uh to
47:47have his experiment successful like
47:49homebeam was talking about
47:50maybe they go they only go choose
47:54uh younger more potentially more
47:57motivated people so
47:58is that a is that a selection at the
48:00local level or just a
48:02i know lots of times these top-level
48:04leaders say i want this guy in this area
48:06um
48:07so where is the decision making there
48:10yeah
48:10so so qualitatively
48:14you know where we're with i don't have
48:16too much to say we're sort of in the
48:18process of
48:19interviewing some of the some of the
48:20former sort of you know way
48:23folks to see how the how they actually
48:24do this some of the things that we we
48:27we observe in our data that it's not
48:28just a generic thing about about
48:30korean incentive in the sense of if you
48:32have very high current incentive
48:33and you're actually coming from a very
48:34poor places that doesn't quite flip the
48:36thing
48:37so in some sense it's like korean
48:39incentive interacted with you are coming
48:40from
48:41average virtual places that gets you
48:42into the expectation uh then sort of
48:45a very clear motivated person in guys
48:47who didn't quite do much
48:49in terms of expectation maybe because of
48:51the some kind of screening from the from
48:52the from the from the top level
48:54uh that's unclear uh for us the last
48:56thing i want i want to show you
48:58uh besides this is not meant to be
49:01again exhaustive uh but just purely as
49:04accounting exercise if we if we take
49:06into account of the two
49:07misalignment of incentive sources that i
49:09just show you patronage and
49:11very politically motivated local
49:12politicians just taking these two
49:14accounts
49:14can already explain the way about 50 of
49:17the observed
49:18selection that that that we see in the
49:19data sets okay
49:22now there are there are other objectives
49:23the central government might have that
49:25that that's
49:26on top of representatives and that may
49:28actually counteract representatives for
49:29example they might want political
49:31stability
49:31they don't want to try things out in a
49:33place that just had a protest because
49:35any policy this uh sort of change might
49:37generate this further disturbance
49:39and in fact we see that you know the
49:41places the companies for example that
49:43just experience political unrest
49:44don't get experimentation in the next
49:46three years uh potentially as a veto
49:48essentially just say that you know
49:50you you know they need political safe
49:51environment for that
49:53and to extend that political unrest more
49:54likely to happen in the slightly richer
49:56places that's going to veto out a lot of
49:58the
49:58sort of more likely happening in in in
50:00the places that are extremely important
50:02to economic hardship
50:03and that's going to veto out a lot of
50:05the all of the progress to be part of
50:07communication
50:09so the first couple of evidence i want
50:10to show you is is that
50:13they don't seem to to to fully take that
50:15into account so the underlying sort of
50:16logic of this particular exercise is
50:18that
50:18to extend there's a large heterogeneous
50:20treatment effects
50:21and to the extent that there is on
50:23average positive selection
50:24experimentation sites
50:26that means that the more possibly
50:28selected a particular explanation is
50:30the more likely it's going to generate
50:32positive policy outcome
50:33or signals to the central government and
50:35if that's the thing that central
50:36government is evaluating that
50:37policy on trial on that would suggest
50:39that those policy gets rolled at a
50:41higher
50:42frequency than the same exploitation
50:44that are slightly less like
50:45possibly selected okay so then we can
50:47look at this non-representative
50:50on the right-hand side is essentially
50:51the t-stats of the examination so each
50:52explanation
50:53we can predict whether that that that
50:56the pulse gets raw to the entire country
50:58as a function of the underlying positive
51:00selection of the policy the t-stats
51:03and this still a large coefficient
51:07meaning that for every one unit
51:09increases
51:09t stats uh that they increase almost
51:12like two thirds of the
51:13of in terms of the base rate of the
51:15likelihood of the cost gets wrong to the
51:16entire country
51:17okay so another way to put this is more
51:19positive psychopathies and they're more
51:21likely to
51:22deem as successful but a central
51:23government in relation to the entire
51:25country
51:25and note that you know even for the
51:27policy communication space
51:29that's already selected on central
51:30government having a very high prior
51:32very not all of them go out to entire
51:34country if anything you know only a
51:35quarter
51:36of them become national policies two you
51:38know
51:39three quarters of them get killed after
51:41the experimentation and never become a
51:42national policy
51:43and and then the next thing i would like
51:47to show you is that
51:48you know there are a couple of other
51:50pieces of evidence that we can gather to
51:51say that
51:52the central government just don't
51:53systematically sort of you know
51:55ignore osagonal signals when they when
51:57they when when they value a policy
51:59when they evaluate a experimentation
52:02outcome so another access we can do is
52:04that we can look at
52:05say agricultural policies agricultural
52:06experimentation and then
52:08that experimentation with the policy
52:10that's under trial should the
52:12effectiveness of the policy should have
52:13nothing to do with
52:14the idiosyncratic natural disasters such
52:16as draw the flood that happened during
52:17the period of the exhibition for the
52:19sites that are undergoing
52:19experimentation
52:20okay so then we can look at sort of the
52:22just the agricultural policy
52:24and look at the percentage of the
52:25expansion sites that actually gone
52:27through
52:27or experienced natural disaster during
52:29the period of simulation
52:31which would generate better bad outcomes
52:33for the local economy of philadelphia
52:34outbox but it has nothing to do with the
52:35policy
52:36do those natural disasters actually make
52:38that policy on trial less likely to draw
52:40to the entire country
52:42and that that's what we see here which
52:43suggests that you know central
52:45government are not taking into account
52:46of those orthogonal signals that
52:48we think it's they should get rid of
52:50when
52:51when it when they value the policy uh uh
52:53but but that gets sort of rolled in
52:55uh to the to the overall evaluation of
52:58the policy
52:59okay so if the the policy experiment
53:01just get unlucky and then
53:02get rid of in a few places that
53:03experience uh a natural disaster
53:05the central government seems to to
53:06consider those policies as a failure uh
53:09and and that that stopped the party from
53:10growing to the whole country
53:13okay the final two uh uh
53:16set of facts or results i want to show
53:19you is that
53:20now among the policies that africa
53:21rolled out to the entire country
53:23what happened to those policies and what
53:25happened to the places that
53:26experience the new policy that wasn't
53:28being part of the trial
53:30so one thing we can do is that so
53:32consider you know a policy that you know
53:34take the
53:34for example the fiscal reform policy
53:36that we just had it gotta go out to half
53:37the country
53:38we can live for each of the remaining
53:40half of the country that didn't with the
53:42wooden part of expectation
53:44we can calculate how close they are to
53:46the expectation size in terms of their
53:48pre-examination characteristics
53:51and then we can see you know whether the
53:53closeness of that particular count
53:54quantities to the explanation sites
53:56it's related to how positive your your
53:58ex in terms of effect you're
53:59experiencing from the policy
54:01okay so this is what i'm showing you in
54:03in this particular table
54:05uh uh the f the end distance is for each
54:07of the policy
54:08we can calculate the non-expectation
54:10size distance
54:11to the expectation sites among the
54:13policies that get rid of the whole
54:14country
54:15okay so the takeaway here is that for
54:17the policies that get brought to the
54:18entire country
54:20the new places that get to experience
54:22the policy that wasn't part of the
54:23expansion sites
54:24the further away they are in terms of
54:27their characteristics to the expansion
54:28sites
54:29the less beneficial their policy is to
54:31them okay the negative here is actually
54:33relative so it doesn't mean that
54:34necessarily hurt but they're just
54:35they're benefiting less foreign policy
54:37from the places if another way to put
54:39this if shanghai tried to policy the
54:40parts get relative to the rest of the
54:42country
54:42shouldn't benefit equally but not forget
54:44to uh uh
54:46given sort of how close they are to the
54:47shanghai's pre-explanation
54:49characteristics
54:50okay finally what we can show and you
54:54sort of you take this as
54:55a grain of thought given sort of you
54:56know the the the the degree of
54:59aggregation here so so this is very much
55:00just
55:01correlational and i want to claim
55:03anything causal here
55:04for each of the policies that roll out
55:06for the foot of the entire country in
55:08china for the last
55:0940 years we can trace back how possible
55:12selected that policy was
55:13at the time of the expectation okay and
55:16then we can count
55:17in year 2010 how many policies that get
55:20rolled in that particular year
55:22that are massively possibly selected and
55:24how many positives get rolled in that
55:26year that are somewhat representative
55:28uh uh and so on we can zoom into just
55:30the economic space because we're looking
55:32at sort of the the economic in default
55:33and then we can see
55:34what's the share of the policies that
55:36are deviating from the expectation and
55:38in terms of making to the final policy
55:40on the intercounty inequality that will
55:42be measured into revolutionary
55:43inequality
55:44in china okay and what you see here is
55:46that
55:47the more policies that get rolled out in
55:48a particular year that's originated from
55:50under representative expectation
55:52the more subsequent intercounty
55:54inequality
55:55it's associated with again nothing
55:57causal here but there's a very strong
56:00association this is not entirely driven
56:02by just policies
56:04that coming from experimentation because
56:05you don't get to you don't you don't see
56:06this possible correlation from policies
56:08to coming from
56:09fairly representative expectation so
56:11it's only the non-representative
56:12experimentation
56:13once it brought the whole country in the
56:14next period it sort of starts to blow up
56:17the inter-county
56:17uh uh uh inequality in a subsequent
56:20period
56:21all right i'm just thinking about my
56:24example
56:24if social economic zone a national
56:26policy in the end
56:29so so so experimentation zone it's not
56:31it's it
56:32we don't count that as the price like
56:34local
56:35exactly so so these are just a subset of
56:37policies
56:38about a quarter of the whole explanation
56:41that eventually become a policy that we
56:43so actually we're fairly strict in in
56:45essence like if a policy that
56:47ended up having a central government
56:48document to say this is the policy
56:50that's originally from the exhibition
56:51and would not want to do it for the
56:52entire country and wouldn't call that a
56:54a national policy
56:57and it they can change the name or not
56:59that's fine we can still trace them but
57:00but it has to come with the central
57:02government documents that this is
57:03something
57:03the whole country should have so i will
57:06not count
57:07but some of the policies are education
57:09some are pollution right i mean there
57:10are different
57:12uh but this this will affect income and
57:14increase differently right
57:16so yes exactly i'm going to increase
57:17something different right so very right
57:19so so so so
57:20given that we in the inequality here is
57:22measured just in terms of local gdp
57:24uh we zooming into the to just the
57:27economic relation policies
57:29uh and then we can we can sort of column
57:30two and column four is zooming into
57:33i can call them sort of the more
57:34important economic policies those are
57:35mentioned in the five-year plan so those
57:37are policies
57:37including policies the pressure have a
57:38bigger base in changing local econom
57:41economic performance than a random
57:44average policy
57:46is there in all of these policies in
57:49here
57:49are rolled out nationally or so in this
57:52in this regression yes oh got it okay
57:55yeah
57:56so so so so the way to understand this
57:58question is that in 2010
58:0010 policies get through all of the whole
58:01country uh
58:03these ten policies are what has been
58:05experimented over the past
58:07sort of a decade and then we look at the
58:09latin policy get roller in 2010
58:11at the time of the their initial
58:13expectation
58:15how how big the division that they had
58:17from from the events
58:18and then we look at 2011's coefficient
58:22and some of the places already had the
58:24policy for for like two or five years uh
58:26if they're
58:27early part of the expectation so so i'm
58:29coming to the end so
58:30let me quickly summarize and i'm sort of
58:32you know eager to hear some additional
58:34comments
58:35that you have uh hopefully i i i i
58:38shall show you some of the some set of
58:40interesting facts
58:42about china's policy competition which
58:44you know is is one of the larger the
58:45scale of such plastic conditions
58:47in recent history potentially possibly
58:50selected
58:51uh part of the positive selection is not
58:54driven by central government's intention
58:56but driven by some misaligned incentive
58:58down the line in the political hierarchy
59:00and that has consequences to the extent
59:02of central government don't fully take
59:03the positive structure into account
59:05and that's not the bias in the process
59:07the pause space
59:08towards more uh towards the ones that
59:10are that are more
59:11uh uh favorable to the ritual
59:15okay i want to link this off a little
59:17bit to the to the
59:18to the literature in economics that to
59:21contrast this with
59:23exploitation on the federal level which
59:25is what is which is
59:26one kind of citation that literature has
59:28successfully studied theoretically
59:29which oftentimes predicts
59:31experimentation because there are free
59:33writing that
59:33of of the information that i don't have
59:35to try the uncertain policy i can just
59:37wait for
59:37my nearby countries or provinces to try
59:39it uh and here sort of a top-down
59:42expectation might actually generate over
59:43expectation because
59:44if that that sort of induce a lot of the
59:46local incentive to participate
59:50and but i think a more potentially sort
59:52of important
59:53lesson from this is that for a large
59:55scale plus expectation at
59:57the level that we're talking about here
1:00:00this has to
1:00:01inevitably conduct it in a very complex
1:00:04environment
1:00:05with various layers of bureaucracy bot
1:00:07and some of them might involve
1:00:08the very sort of frictions or
1:00:10distortions that the explanation want to
1:00:12get rid of
1:00:13but those complex environment could
1:00:15generate system
1:00:16systematic misaligned incentive of the
1:00:18way the explanation is being conducted
1:00:20and to some extent this is it's yet
1:00:22another form of policy capture but this
1:00:25is capture of the policy at the learning
1:00:26stage that
1:00:27the information that government is able
1:00:29to collect are systematically biased
1:00:31more favorably biased towards the
1:00:32virtual places and that should
1:00:34capture the policy towards favoring more
1:00:36rich places because that's the that's
1:00:37the signal that this
1:00:38government are evaluating on uh and
1:00:41central government
1:00:42doesn't seem to be uh fully
1:00:44sophisticated of
1:00:45devising uh this person okay i sure am
1:00:48here and uh i'm happy to
1:00:50uh open up more discussions uh thank you
1:00:53so much
1:00:54yeah david and just a question the
1:00:58is it is that convenient to share the
1:01:00descript
1:01:02summary of your example
1:01:05uh in terms of attributes being captured
1:01:08so we could understand a little bit more
1:01:10because depends on type of experiments
1:01:13you have
1:01:14i wonder if we will have different
1:01:16objective functions
1:01:19and if you have fiscal policy versus
1:01:23education policy
1:01:25you might measure the performance in
1:01:27different
1:01:28metrics yeah yeah i know so so i mean
1:01:31eventually our hope is to make the whole
1:01:33sort of explanation database public uh
1:01:36uh
1:01:36once once the the publication process or
1:01:40it goes through uh and hopefully though
1:01:42that will help many many others to to
1:01:44look at this
1:01:44and and in fact sort of you know you're
1:01:46exactly right sort of the the outcome
1:01:48space is complicated
1:01:50uh that's why i was focusing on the
1:01:52physical reform because you know the gdp
1:01:54local 3d performance is exactly what the
1:01:55reform is meant to induce and some of
1:01:57the education policy
1:01:58uh without measuring the sort of the the
1:02:00the particular dimension that is meant
1:02:03to change
1:02:04uh it's it's it's hard to it's hard to
1:02:06map
1:02:07directly into the outcome space so
1:02:08pollution policy uh tax policy often
1:02:11is a few other domains where we can very
1:02:14so narrowly map into the outcome space
1:02:16uh but but you know you're exactly right
1:02:18it's at least
1:02:20to to fully estimate the heterogeneous
1:02:22treatment effects
1:02:23we can only do this for policy by policy
1:02:25or domain by domain because there's
1:02:26certain domains we just have no hope of
1:02:29seeing the data the central government
1:02:30might have access to
1:02:32thank you uh
1:02:35as a non-economist looking at the the
1:02:38graph that shows the sharp
1:02:39increase in uh
1:02:43experimentation uh beginning
1:02:46in the latter part of the fujian tower
1:02:48regime
1:02:50that is concern
1:02:54about slowing growth a driving factor
1:02:58here for experimentation or is it
1:03:02explicable by increased capabilities
1:03:05within the system
1:03:07uh as a part of the policy design
1:03:10process to conduct experimentations or
1:03:14something else
1:03:15yeah so that's a good question so the
1:03:18answer is i i don't know the only thing
1:03:19i can tell you is that
1:03:22it doesn't matter whether you look at
1:03:23the absolute count of the expectation or
1:03:25the share of policies that wins the
1:03:27expectation
1:03:27out of out of all the policy that's get
1:03:30rolled out to the whole country
1:03:31it's rapidly increasing very sharply
1:03:33over that period
1:03:34so so it's not just purely because you
1:03:36know in the 2000s the central government
1:03:38are doing more policies than drive up
1:03:39the share it's like for every single
1:03:40policy that
1:03:41they will end up rolling there's a
1:03:42higher shelf than winston
1:03:43experimentation stage
1:03:45uh i don't have a good take on what
1:03:46drives the the increase i sort of
1:03:49we have some sense of like what drives
1:03:50the drop a part of the verticalization
1:03:52but the
1:03:54that increase for the two for the two
1:03:55decade uh uh i don't have a direct
1:03:58uh uh i before barry let me let me just
1:04:01follow up on that
1:04:03it who implemented that m versus u-form
1:04:07regime change uh sort of uh
1:04:10um management uh change
1:04:14was that she or was that done before she
1:04:16got on
1:04:17so it started in the very early 2000s
1:04:21uh in the very first ministry actually i
1:04:23should have known the num
1:04:25which ministry got it first and then it
1:04:27was a long lag
1:04:28and then there's a batch of ministries
1:04:30getting that reform or
1:04:31whether that reform uh uh in in in a
1:04:34right before this year
1:04:35so so so many of the ministries that you
1:04:37know imagine so
1:04:38so what that reform means is that it
1:04:41used to be that each local
1:04:42each provincial government has its own
1:04:44for example environmental
1:04:45ministry that's the previous m form and
1:04:48then after the ufo reform then
1:04:50the national sort of sort of
1:04:52environmental production ministry
1:04:54sort of you know had direct restriction
1:04:55over the provincial one and provincial
1:04:57government don't have that uh
1:04:58control anymore that takes years for
1:05:00many of the ministry to to reform over
1:05:02and i think a lot of them started before
1:05:03c
1:05:04uh that's why i was thinking the whole
1:05:05drop is not attributed to this example
1:05:07contribute to this this this ministry
1:05:09reform
1:05:09it's not a tribute to she per se
1:05:13all right yeah if so if the um
1:05:17positive selection is much stronger in
1:05:19uh
1:05:20policies that are rolled out nationwide
1:05:24just for reference what does the sample
1:05:26of policies
1:05:27that weren't rolled out nationwide look
1:05:30like
1:05:31does the positive selection disappear
1:05:34have you ever
1:05:37[Music]
1:05:41policy never end up becoming a national
1:05:43policy and then the average
1:05:45average t stats is above one skill uh so
1:05:48they're still
1:05:48quite far from some representatives but
1:05:50i can't i can get you the the more
1:05:52precise number
1:05:53because i well i think that would be
1:05:54interesting so i mean this is a great
1:05:56talk it's really super interesting
1:05:58um and and i just wanted to deal with
1:06:00the question of sort of staggered
1:06:02implementation versus experimentation
1:06:04um that would show shed light on it you
1:06:07know if
1:06:08if the positive selection went away in
1:06:11the in the subsample that was not
1:06:13adopted
1:06:15uh that would really strengthen your
1:06:17argument and kind of it wouldn't
1:06:19disprove the alternative view but it
1:06:21sure would make it less likely
1:06:23so i think that'd be interesting to
1:06:24report as a benchmark okay yeah no
1:06:26that's
1:06:27that's a great idea uh and and the
1:06:29related one is
1:06:31at some point we're also implicitly
1:06:32taking the stems on
1:06:34many of the policies the central
1:06:35government is trying out are meant to be
1:06:37national policies
1:06:38because it will be a very different sort
1:06:40of conceptual framework if the whole
1:06:41thing is just try to target
1:06:43that i i try out things out and i stop
1:06:46uh
1:06:46the policy when i see an active effect
1:06:48and then i only target the the places
1:06:50that
1:06:50this possible fight will become
1:06:52favorable so so again we don't think
1:06:54that's
1:06:54that's particularly the case uh for
1:06:56example for tax policy for a lot of
1:06:58education policies
1:06:59but there are certain possible domains
1:07:00which targeting might be more
1:07:02suitable to to describe it than sort of
1:07:05from them they're actually sort of
1:07:06figure out a national uniform policy per
1:07:07se
1:07:08but but i agree with you we can we can
1:07:10we can sort of further
1:07:11segment the essentially the comparison
1:07:14but different policy types and uh
1:07:17hi uh thank you thank you david i think
1:07:19it's very interesting
1:07:21i think the unrepresentative selection
1:07:23of the reform
1:07:24size is undoubtedly a reason behind the
1:07:27heterogeneous effective reforms
1:07:29assume as it was illustrated in the in
1:07:31the presentation
1:07:32but i was wondering if the effects also
1:07:34due to some sort of like a general
1:07:35equilibrium outcome
1:07:36like the early reformers like got some
1:07:39market advantages and so they perform
1:07:41well
1:07:42and by once away from growth out to the
1:07:44whole country the followers
1:07:45don't have these advantages and they
1:07:47suffer a little bit more
1:07:48because i think you talk about this
1:07:50somewhere in the presentation
1:07:53yeah yeah right but i i want to leave
1:07:55kind of elaborate a bit more
1:07:57yeah thank you okay yeah so i can't tell
1:08:00i can't tell you sort of definitively
1:08:02for all the policies but for a few ones
1:08:03again again because it matters for
1:08:05what kind of outcome we can match it to
1:08:06for the for the fiscal reform policy for
1:08:08example we don't think it's all driven
1:08:10by
1:08:11or first mover advantage because for a
1:08:13few counties who get
1:08:14the reform very early on that are poor
1:08:17they are also
1:08:18lost from the reform so it doesn't if
1:08:20you're if you're a poor place relatively
1:08:22poor place
1:08:22you lose from this particular physical
1:08:24reform regardless of when you get it
1:08:26and same for sort of the ritual place
1:08:27you'll get a policy later on they're
1:08:28they're also on average more likely to
1:08:30be benefiting so there i don't think
1:08:31it's all
1:08:32coming from a ge effect but i can't say
1:08:34this to be
1:08:35true for all the policies out there the
1:08:36there's certainly definitely going to be
1:08:37certain
1:08:38you know industrial policy for example
1:08:40there's there should be a huge
1:08:41first mover advantage uh and
1:08:44and we don't yeah that'll be another
1:08:48that will be sort of exaggerating the
1:08:50essential incentive for a possibly
1:08:52selected local government to
1:08:53want to join experiments early on and
1:08:55further biasing the signal that they get
1:08:58uh in the in the first wave of the
1:08:59explanation uh to be positive
1:09:01so so so let me actually give let me
1:09:04this is somewhat related to bearish
1:09:05question earlier is that
1:09:07from a theoretical perspective this
1:09:09won't be as big of an issue
1:09:11if the trial error is just sufficiently
1:09:13long you know if they do
1:09:15if they try every single policies like
1:09:17for 10 years and eventually reach the
1:09:19to rf poor places and get a negative
1:09:20signal and they decide this is a bad
1:09:22policy as they actually
1:09:23end up realizing for the fiscal reform
1:09:25the physical reform policy didn't go out
1:09:27for the country because the last three
1:09:29years it was very
1:09:30negative the effect and that's okay to
1:09:33something it's just that a lot of the
1:09:34policy gets
1:09:35sort of get rolled to the whole country
1:09:36too early they do the expectation for
1:09:38one year or two year
1:09:39only in shanghai and they realize it's
1:09:41good and then
1:09:42they didn't try out in any of the poor
1:09:44places and that massively positive
1:09:45signal the central government get take
1:09:47uh
1:09:48as face value and then it goes to the
1:09:50whole country and that's i think it's
1:09:52part of the key the key uh friction
1:09:55is in this in this in this process
1:09:59i think you try to compare i mean china
1:10:02to
1:10:02the literature about other countries
1:10:04right but you didn't
1:10:06uh say too much about different
1:10:09countries
1:10:10yeah well i mean the political system
1:10:13is totally different i mean how the
1:10:16center controls the bureaucrats
1:10:18of the country this is
1:10:21very different right even in china i
1:10:24mean even tom noticed that there's a
1:10:26third right he expressed experimentation
1:10:29during who
1:10:29period right also depends on the
1:10:31leadership style right
1:10:33some experiments some doesn't right
1:10:36so this will change too right so in the
1:10:38end it's
1:10:40about political systems so how do any so
1:10:44so the way i think about and we're sort
1:10:46of in the process of collecting so two
1:10:48sets of data that we're collecting right
1:10:49now
1:10:50try to make some benchmark on is you
1:10:51know u.s will be a perfect
1:10:53uh perfect suitable example to look at
1:10:55for bottom-up experimentation on
1:10:57of massachusetts have one policy
1:10:58california have one policy that
1:11:00possibly selected but being rich places
1:11:02uh
1:11:03and and and and you know federal
1:11:05government might decide to make this
1:11:07sort of a national policy and so on so
1:11:08one of we'll try to benchmark with sort
1:11:10of the bottom of experimentation
1:11:12or possible innovation in in a federal
1:11:14system in the u.s
1:11:16another version of this is we can look
1:11:18at for example all the
1:11:19rcts that get rolled out by j-pop uh for
1:11:22example
1:11:23that they have internal validity
1:11:25conditioning on extension sites
1:11:26everything's eternally valid
1:11:28but then they don't try it out in the
1:11:29random places for the whole country they
1:11:31you know
1:11:31the compliance sort of set it's it's
1:11:33likely also possible factory because
1:11:35because you don't you don't go to the
1:11:36poorest places
1:11:37or even in representative places in
1:11:38india to do towards rtt so that's
1:11:40another thing
1:11:41we can we can compare now that what come
1:11:44back to with with this
1:11:45very large theoretical literature that
1:11:47actually hasn't been much empirical work
1:11:49on this is that
1:11:50the typical sort of uh uh uh
1:11:53sort of conclusion from that literature
1:11:54is that in the bottom up environment
1:11:58unless you have massive first mover
1:12:00advantage that you want to jump on the
1:12:01the policy early on otherwise everyone
1:12:04just wait
1:12:05uh because because of the risk and they
1:12:07wait for the first person to try it and
1:12:09then they decide whether it's good or
1:12:10not and that's gonna
1:12:11have to generate under experimentation
1:12:13that's kind of that can be a bad
1:12:14actually
1:12:15so here in china's case i think which is
1:12:17i think is quite representative of a bot
1:12:19of a top-down version is that if the
1:12:21local government have any
1:12:23prior that this is something that the
1:12:24central government would like to do to
1:12:26succeed because they
1:12:27they did a whole operation uh they're
1:12:29going to overly participate overly
1:12:31zealous about even a
1:12:33not so good policy and this over
1:12:35exploitation
1:12:36because of the proof incentive of
1:12:38potentially getting promoted and so on
1:12:40uh is the exact opposite of what a
1:12:42bottom line expectation would predict
1:12:43because that will predict under
1:12:45explanation that the information coming
1:12:47too slowly in those environments great
1:12:49but i think
1:12:50i think time's up i think thanks david
1:12:52for a really stimulating
1:12:54talk and we hope to see you here soon

About this transcript

This page contains the full transcript of The Political Economy of Policy Experimentations in China with Prof. David Yang by Stanford Center on China's Economy and Institutions, generated from the public captions YouTube serves with the video. The transcript has 14,329 words across 2,477 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.