Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # The Return(s) of Superstar Scientists Authors: Michela Giorcelli, Yujing Huang, Bo Li Discussant: None Video: https://www.youtube.com/watch?v=7LDwpGoPNFc&t=14235s ## Talk (03:57:15 – 04:32:34) [03:57:15] >> Okay. >> Okay. Great. Thanks for including us on this great program. I'm Eugene from UCLA. Today I'm presenting this working paper the returns of superstar scientists. This is joint work with [03:57:29] Michaela from UCLA and Bo from Pecking University. So boy is also here with us. [03:57:34] We are very happy to take any comments feedback as we're still developing the paper. So to start we all know or acknowledge that human capital especially the upper tail human capital is very crucial for the scientific [03:57:48] knowledge creation which ultimately matters for innovation and long-term growth. However, the process of knowledge formation is often indogenous and it is empirically difficult for researchers to disentangle the effect of [03:58:02] human capital due to various selection issues. On the other hand, existing work on high-skll migration has primarily focused on brain drain and its impact for the sending side or the impact of sudden inflows of foreign scientists on [03:58:16] the receiving end. So both are one direction. So what we know much less about is the return migration which is the brain circulation and another empirical challenge is to follow these individuals across different contexts [03:58:30] different countries and different corporas. So to um fill the gap in the literature in this paper we ask the following two questions. How does the return migration of these top talents shape their individual scientific [03:58:44] trajectories? And second, what are the long-term effects of their return on the knowledge production and innovation in their home country? [03:58:54] And to answer yes, >> sorry. [03:59:00] >> No, sorry. So that we can see the title of your slides. I mean, hide. [03:59:06] >> Hide these ones. Okay. >> Okay. [03:59:10] >> Okay. Perfect. Oops. Great. [03:59:16] Let me pull this here. Yep. So in order to understand the returns to human capital in this paper, we look into a very unique historical episode of [03:59:29] the great return of US trained Chinese scientists roughly from the end of the World War II and lasted for a decade. So after the world war II there was a surge in Chinese students who went abroad to study in the US and also who returned [03:59:44] from US to China and by the time when people's republic of China was established around 3,500 Chinese students were studying the US 70% of them were pursuing STEM majors and after [03:59:57] communist party took power the diplomatic ties were cut between the two countries and study abroad were restricted in the western countries. So in this period around 1300 Chinese students returned from the US and they [04:00:11] became important scientific main power for new China. So in order to trace the individual scientific output across decades, we first did extensive data work to compile comprehensive records of [04:00:25] US trained Chinese scientists tracing back from the very first wave in 1872. [04:00:31] and then we match their um names with their publications both in English and Chinese journals. As for identification, we exploit the quasi random variation in the barriers for these students to [04:00:45] return to China driven by this US policy in 1951 to prevent the departure of aliens with scientific and technical training. And we match each returne with a stayer that shares very similar [04:00:59] observational characteristic and also very similar um very comparable pre- returnturn uh outcomes and then see the effect of return on their individual trajectories and we also compare the [04:01:13] fields in China that received a higher dose versus a lower dose of these returnees and what's more interesting on the China side after 1957 the country entered a 20-year episode of domestic political turmoil which [04:01:28] prevented further returns and this gives us a natural experiment to see the effect of these returnees in a quite isolated environment. [04:01:38] So what we've found so far is in uh three parts. First, at the individual level, we found that those returnees redirect their output from international publications to domestic journals, whereas their total publication remains [04:01:53] almost unchanged. And interestingly, there's no loss of research quality measured by important score or frontier alignment. But if we zoom in to see the exact topics they were studying, we find [04:02:06] that they shift toward more applied um missionoriented work in order to comply with the government priorities at that time. And second, at the aggregate level, we find that these returnees [04:02:19] generate important domestic spillovers. in the fields that received a higher dose of these returnees. There was an increase of output by their colleagues, including more work closer to the [04:02:32] research of the returnees and also global research frontier. And finally from a global perspective we quantify the economic value of science via a network approach and we calculate the [04:02:46] social value for each returnnee and find that China gains around only a third against the US laws and this is primarily due to China's low absorptive capacity at that time which I will [04:02:58] elaborate in the final section. So our findings are broadly related to these four strands of literature. First on the determinance of knowledge production in silververse we study the effects of [04:03:12] these highskll return migration especially for an economy at the very early stage of industrialization and also scientific development. And second, we also speak to the literature on the accumulation of upper tail human [04:03:26] capital. And to borrow the words from Mokir, the diffusion of useful knowledge, we find that these returnees can actually see the field. They provide the foundation of China's nowadays [04:03:39] scientific infrastructure. And third, we complement Reason's theoretical work on the global allocation of talent by adapting their framework to calculate the social value [04:03:51] of uh these reallocation of talent. And we find that the social value of these um scientists really depend on their individual positions in the knowledge network. And finally, a broader literature discussing the political [04:04:05] economy of science. We're going to show that China was able to jumpstart because the state directed these top talents into technologies it designated as crucial says allowing the state to build [04:04:20] capacity under central planning. So here is an outline of my talk today. [04:04:26] In the first part I will briefly introduce the historical background just to bring everyone on the same page. [04:04:33] So the people's republic of China was founded in 1949. At that time the country was pretty backward after decades of war and the government in order to modernize and industrialized the economy called for the overseas [04:04:48] students to return and serve motherland. And at that time around 3,500 Chinese students were in the US and they were like 70% of total overseas students. [04:05:00] However, due to the Korean war and also amid anti-communist sentiment, the US government feared that the loss of these human capital would contribute to the building of the communist regimes and [04:05:12] then they issued this policy in 1951 to prevent departure of these scientists especially working in the fields of military value. So even if they really want to go back home, they were effect effectively barred from exit and the [04:05:27] scenario lasted for around four years until in Geneva talks the two countries reach a consensus that the nationals residing in each other's home each other's country can return home if they [04:05:40] wish. However, a lot of these stranded scholars already turned into permanent residents because otherwise they don't have work authorization and uh thanks to this 1953 [04:05:54] refugee relief act. So the main key key takeaway from this slide is that these cold war barriers really generate quadrandom variation in who return conditional on their intent. [04:06:08] So here is in the time series of the number of returnees over half a decade over half um a century and what we see is that there's really a surge of these [04:06:20] returnees in this great return wave that's why we're especially interested in this period and these US exiban really effectively control the departure of these people and after like Geneva [04:06:35] talks the restrictions were lifted there another small peak followed by almost nobody going back to China due to these like cultural revolution etc. [04:06:46] So here are some examples of these famous scientists in this great return wave. Probably for every Chinese the textbook example would be Tenu who obtained his PhD in physics in Caltech [04:06:58] and returned to China in 1955 eventually pioneered China's missile and space program. Along with him were who were both studying nuclear physics. [04:07:10] They obtained their PhD around 1950, returned home right after their graduation. One designed China's first atomic bomb and the other directed China's nuclear weapons program. And [04:07:24] equally famous are those scientists who eventually stayed in the US like who obtained their PhD in University of Chicago and eventually won the Nobel Prize in physics in 1957. [04:07:39] So we're not say saying that every individual in our sample are as prominent as these guys but they are really sitting at the top of the human capital distribution even from today's perspective especially for China at that [04:07:53] time. So in order to compile a very comprehensive data we really did expend a lot of time um first um trying to get this scholar level panel of US French [04:08:06] Chinese scientists. So we emerged from four historical archival sources to obtain information on these uh 4600 uh US trained Chinese returnees [04:08:19] especially on their education fields departure and returneers and along with their post return life outcomes. We also obtained over 9,400 US French Chinese stairs from these two [04:08:33] resources two sources. Uh similarly collecting their individual characteristics and this is a structure of biography data for returnees in 1950s. [04:08:46] We have very rich biographical information and this allows us to match each individual correctly to their publications and in order to link the scholars to individual level scientific [04:09:00] publications um we combine two complimentary sources one is the standard openal and the other is specific to China let's call it China obanalex CNKI it provides nearly [04:09:14] complete record of all Chinese language academic publications and overall we want to get the scholar level productivity measures which is a panel of their publications and citations [04:09:26] every year. Yes, correct. Seems like publications might be a big deal in the US. But >> yeah, >> maybe you know the examples you gave those three people who went back to China. [04:09:36] >> Yes. >> Started missiles and nuclear weapons. [04:09:39] They were not theor back in the US became theoretical physicists. I mean >> yes >> that in terms of outcome measures there's a there might be less of an incentive to publish in China but more of an incentive to [04:09:53] >> publishing technologies and things like that. So as an outcome measure, we do are we missing >> kind of all the stuff that they're creating in China even though they're not publishing. [04:10:04] >> That's a great question. So scientific output is the most measurable output for China implanted economy. We don't have the patents and a lot of these scientists were involved in those secret [04:10:18] national projects which we cannot measure. Um however we can observe what happens to their colleagues in order to document the domestic spillovers. Yes. [04:10:28] But in any in either case we're providing a lower bound of our estimate if they are even contributing to other non um publication contribution. Yes. [04:10:41] >> Yeah. is that the first one of the earliest the social welfare producing the aggregate welfare in fact so you know there he was saying that >> the loss to ch the gain to China is only about one quarter of the loss to the US [04:10:56] >> yes >> my guess is that's true of the publication >> exactly yes >> we might be missing the fact that the gate to China >> in other yes totally yeah we take it so [04:11:09] we also went beyond the standard are biblometric measures to perform textual analysis on all these papers. We get the titles and abstract for each one of the paper embed each paper via sentence [04:11:23] transformer so that it become a point on the semantic space and then we take the average of that scholar that year to create this agenda vector for that scholar and compare this agenda vector [04:11:37] to uh the reference vectors using cosine similarity. So eventually we're interested in their sp the stability of their research agenda whether they converge to their host institution whether they have a very close [04:11:51] relationship still with their US co-authors and particularly the frontier alignment and policy alignment of their research in China. So the output is a dynamic scholar level metrics capturing the direction the diversity and [04:12:06] innovativeness of their research. So the first step is to validate um our text embedding. We plot it on this text embedding space. The key takeaway is so [04:12:18] each point okay it's not very easy to see but each point on these graphs represent a paper and these papers actually clusters by field. So the embeddings actually capture meaningful [04:12:31] research content for each paper. it's 75% of its nearest ne 25 nearest neighbor shear is filled. We also compare different models and eventually [04:12:43] use the multilingual one as our primary measure. [04:12:48] And next the natural question to ask is how do we construct the frontier alignment? So to do that we adapt cali at all from patent to publications in the embedding space. So the key [04:13:01] intuition is that a paper is important if it is novel meaning that it shares very low backward similarity but it's influential meaning that it shares very high forward similarity and we [04:13:16] residualize Q on the year effect to get the final important scores. So to visualize this, so every point on this uh first graph is a paper and then we [04:13:29] embed each paper, pick the breakthrough papers which are sitting at the top 10th um percent of that important score distribution and call them the set of breakthrough papers for that subfield [04:13:42] that year. And then we take the average embedding to become the frontier and we measure the frontier alignment to be the cosign similarity between the research agenda of that scholar that year and the [04:13:56] frontier. So after um we plot the frontier alignment and important index in the English journal which is the open Alex [04:14:08] the first row and CNKI which is Chinese journals. We find that they vary across the sematic space. Importantly, frontier alignment captures the direction of the research. Whereas important score [04:14:22] captures the quality of it, their research. So these two have actually very low correlation. Direction doesn't imply quality and they both serve as our main outcome variables. [04:14:35] So after introducing the data and our measures in the first section, we try to understand how the productivity and research agenda evolve conditional on the return. And to do this, we use this [04:14:50] matched sample of returnee and stairs. Basically for each returnnee we match with a stayer on their research specialization their PhD cohort first publication year their pre-return [04:15:02] productivity and their PhD institution so we are regressing their publication citation or cosign similarity measures on the interaction between uh the year [04:15:16] dummies and returnee indicator controlling for scholar year and fillby-year fix effect. So essentially I'm comparing you and your cohortmate who stayed in the US and what we find is [04:15:30] that first the returnees actually redirect output from the international journals into domestic journals. [04:15:38] However, their total publication almost remains unchanged. If anything, there's even a bit higher after the return. And because you were publishing much more in the domestic journals, they [04:15:53] have fewer audience. So their total citations went down. But this doesn't mean that their research quality goes down. And to do that, we use the measures I just introduced, the [04:16:06] important score and frontier alignment. We pull all of the scholars output in that year both in English and Chinese and scored against the same global reference and neither of these move [04:16:20] after their return. So they actually produce very similar work Santiago. [04:16:26] Yeah, >> I was wondering about the speed over from treatment control. We are too power similar topics. If my colleague left now I have more chances [04:16:40] to publish to get conference spots >> etc. I might benefit from reduce competition in London. So I don't know if you consider that possibility and also if you are showing different maybe you could show us the the raw data so we [04:16:54] don't see that. >> Ah yes >> I see. So interesting the interesting feature of this context is because China [04:17:07] was quite isolated at that moment after the returnees went back home. I do not think they have very frequent communication with the stairs in the US. [04:17:18] It's very hard for them. >> Yes. [04:17:24] >> Ah, with less. Ah, I see. I see. Yes. Yeah, we can show that by just showing probably the raw trends, right? Yeah. [04:17:31] Great. Great point. One question on the left panel slide >> just to make sure I interpret correctly the red one saying like the coefficients around 20% difference just noisy or how [04:17:45] should I interpret this >> the same speed as the coefficient >> so great point so you're so we are basically comparing the returnees and stayers and their important score went [04:17:59] up um the outcome is the raw important score and um I don't think it's the percentage percent of uh the increase [04:18:12] this is actually the raw >> yeah points >> but this is the important no this is the how we construct the important score it's the raw important score not like [04:18:24] the log of important score that would be 20% more does that make sense yes Okay. [04:18:33] Yes. But I should have uh economic magnitude stated on this side. Yeah. [04:18:38] Okay. So after we zoom in to see the research content, takeen as an example, although he studied in the same sub field, but he moved his topics from more specific [04:18:52] aerodynamic programs to more applied problems in general. So we test our hypothesis in the universe of all the returnees and find that they did change [04:19:05] their research more and shifted more towards the applied work and the natural question to ask is to what applied toward what and to on this oops so to [04:19:19] this end we use textual anal analysis on the 12-year science and technological development plan and see that they actually aligned much more with the government policies in order uh [04:19:33] especially these 12 priority areas. So after having a sense of how the individual Yes. [04:19:44] So when we're comparing my cohortmates to me and what do we have a sense of why I chose to return like is it because I wanted to shift my you know what I was working on and >> yes >> institutionally it was more conducive to [04:19:58] do that in China to build bombs or Yeah. >> Yeah. So actually that's why in the historical background I was trying to explain the prevention of departure of [04:20:09] aliens. So basically the US generates the quadrand variation in the actual return. It's not like I want to return and then I can go back. It's due to US [04:20:22] control. It's more or less exogenous to individual's preferences. [04:20:28] >> See that was like at a time that was like the variation over time, right? But then the variation once the you know when that variation kicks in which of us [04:20:40] in the same cohort chooses to return. >> There is selection into that chart right? [04:20:48] >> Oh I see. So you mean the in the people who have applied for exit they are already different from those who yeah who stayed >> who chose to return >> and the cohort chose to stay [04:21:02] >> it could be that the person who chose to return >> just wanted to change like you know I wanted to publish in general you know like where it's not successful publishing in the US you know >> just ride based selection essentially [04:21:16] >> yes that's a good point Yes, Fabian which we guys who stay are the Chinese nationals. [04:21:25] >> Yes. >> Also >> also so there are US trained Chinese scientists who eventually stayed in the US. [04:21:34] >> Yeah. >> Okay. So I think so we are actually applying for a classified data on the full list of individuals who were prevented from departure. So we know who actually we will know who actually [04:21:48] applied for exit and within that sample we see the individuals who prevent who were prevented and uh uh went back eventually but because it's classified it's taking them some time to get back [04:22:00] to us. Yes. Yes. So this is so the current design is what we are using now on the universe of returnees but later on we would refine the identification. [04:22:12] So in the second part we tried to identify the effects of returnees on the domestic scientific publication. [04:22:19] Basically trying to understand whether the subfields receiving more US trained scholars could would not only produce more output scientific output but also more frontier approximate research and [04:22:33] we did that by estimating this stacked event study. We regress the domestic publication or frontier approximate papers on the interaction between year dummies and the standardized number of [04:22:45] returnees which is the basically the continuous measure of returnees controlling for subfield and year fix effect. [04:22:55] Oops. So okay and then we find that in the fields that received more returnees there is an increase of the total domestic [04:23:10] scientific output and interestingly it is driven by the output of the non- returnturnees who were their colleagues and this implies that the frontier knowledge effectively diffuses into the [04:23:23] domestic community and to quantify the magnitude of suff field receiving one standard deviation more returnees sees actually 70% more of the non- returnturn output over the first decade which is [04:23:38] about 25 extra papers a year and then also in these subfields there is an increase of um the others frontier [04:23:50] approximate papers and also papers around the research returnees agenda so you will See there is a weird pretrend here and this is because what we just said the returnees themselves shifted [04:24:03] their research agenda and this in yeah >> well already a lot better >> a lot better yes [04:24:17] >> they are closer to >> yes I I guess a better way is for us to show the summary statistics of the people who stayed in the Yes, I don't [04:24:30] have that but um I would say that >> from these uh we can back it up um yes so basically there are more productions after they arrive near the global [04:24:45] frontier so the quality increase so it must be the case that these returnees are were having better skills in general compared to the local trained ones Can you tell us? [04:24:57] >> Yes. >> How much of that is because it's your colleagues versus your students? [04:25:02] >> Yes, this is um we actually haven't done that yet, but this is what we plan to do to disentangle the network effects either through their colleagues or through the intergenerational channel. [04:25:17] The Yeah. So a natural question is Yes. So what are the mechanisms and for now we haven't done the network analysis but we um show the summary statistics we see [04:25:32] the returnees actually ran the institutes and taught the cohorts 58% of them took university post 26% actually lead uh were leaders and seven headed [04:25:45] the institutes and 17 of the 23 two bombs once satellite honored niece were these US trained Chinese scientists. [04:25:55] And then if we look at these strategic fields, we see very meaningful catchup with convergence toward the return frontier and also in volumes across [04:26:08] subfields in over decades. So they actually really set the foundation of scientific infrastructure for China. And in the final 10 minutes, I would briefly [04:26:21] go through how we actually quantify the welfare especially for China's gain versus the US laws. And to this end, we [04:26:32] um borrow and adapt the these two recent working papers on innovation networks and the economic value of science. So the idea is uh paper matters um a [04:26:45] scientist matters not only because of his papers their papers but also because what is built upon it. So a field's value is defined as the sum of its total [04:26:59] output and the discounted chain of spillovers it sets across the innovation network. And for these fields that sit at the center of the network because they generate more spillovers, they're [04:27:14] actually worth much more than their raw output. [04:27:19] So each field would keep a share of its spillovers at home, which we call the self-sufficiency rate and leaks the rest to the word frontier. [04:27:29] And in order to calculate China's game versus the US laws, we need to know first the returnees of course first the network structure and centrality structure and second um each field's [04:27:43] output share. Third the returnee output share in the field F and finally how much each country keeps. So for us we calculate this from domestic citation [04:27:56] sheare where in China we set field by field by how much each China's stuff field sits to the global frontier which is US. [04:28:06] So this is our first ingredient omega. This omega is just a matrix and we see that these darker dots darker boxes are the citations concentrated in these [04:28:19] subfields. So losing scientists in these con highly concentrated subfields has very important consequences. [04:28:29] And on the right hand side this is the China US innovation network. [04:28:34] The circle size represents the field's share of research output. Whereas the width of the bar uh indicates China's reliance on the US frontier at that time especially for aerospace engineering and [04:28:47] mechan mechanical engineering. Yes, there's okay. [04:28:55] >> Okay. And then for this innovation centrality, we plot the centrality against subfield. [04:29:03] And we find that the cit centrality is actually highly skewed. These top fields actually warrant four to 18 times the R&D of the lower rank ones. So which field that the scientists eventually [04:29:17] worked in dominates the social value. And then if you plot the returnee output share against the subfield network centrality we can find a very positive correlation around 6. This means that [04:29:32] the returnees were actually going to these central high value fields. And this is eventually why a large value of these uh a large amount of these social value were moved from country to [04:29:46] country. And finally why China gained less. We calculate um the China's subfield self self um sufficiency rate in every [04:29:59] subfield. And it's it's far oh sorry oops and and it's far below the US in every field. So China at that time was only able to absorb a fraction of each [04:30:14] returne spillovers. So to give an example for Chinese scientists producing work in the US subsequently their American colleagues can actually utilize their papers to generate new ideas. But [04:30:28] in China it they don't have very good peers to actually utilize their ideas. [04:30:34] So the rest leaked to a frontier it could not yet use. [04:30:38] Yes. So eventually we c after having these ingredients we calculate per returne social value and find that China's gain offsets only a third of the US loss and eventually this is a net worth loss in [04:30:53] every subfield ordered by the top returnees subfield the US loss is larger than Chinese gain everywhere so the scientific value was actually destroyed not just transferred so despite the net [04:31:08] worth loss what China gained So eventually for these fields that received more returnees it become less reliant on the US science eventually building the scientific and technical power for the country. [04:31:22] So um to conclude, we build a very novel large scale data set to link the mid-century scientific migration education and their scientific output especially looking at this very unique [04:31:36] historical episode of the great return of US ranked Chinese scientists and we are able to trace how the frontier knowledge were diffusing from uh from a frontier to a less uh developed country. [04:31:51] And all of our findings um you can read it. But eventually we showed that these talent mobility policy shapes the global allocation of knowledge and more importantly the scientific capacities [04:32:04] behind the international power and we do it using very novel measures that uh and our results cannot be captured by the standard metric measure. Yes, that's all I have. [04:32:18] Thank you for your feedback comment. Um yes, happy to talk during the break. ## Q&A (04:32:34 – 04:33:22) [04:32:34] I think we have time for one question if anyone has anything still. [04:32:44] >> Yeah, go ahead. Did you think at all at how like China's proximity to the global frontier has changed and whether we should believe that now there's like net global loss to return? [04:32:55] >> That's a good point. So now the China's self-sufficiency rate has um increased a lot compared to like 1950s. So the analysis the results from our network analysis would be very different for [04:33:10] this because that's a very big But you can take more questions. [04:33:20] >> Good. >> Okay. Yeah. Go ahead. Thank you.