Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # A Quarter of US-Trained Scientists Eventually Leave. Is the US Giving Away Its Edge? Authors: Lee Fleming, Daniel P. Gross, Dror Shvadron, Hansen Zhang Discussant: None Video: https://www.youtube.com/watch?v=7LDwpGoPNFc&t=9867s ## Talk (02:44:27 – 02:58:59) [02:44:27] >> All right. Hi everyone. I'm my name is D. I'm presenting a paper co-authored with Hansen Zang and Daniel Gross at Duke and Lee Fleming at UC Berkeley. So [02:44:38] about 40% of PhD STEM PhD trained in the US are foreigners. [02:44:48] descriptively where do they go after they graduate? Do they stay in the US? [02:44:53] Do they go back to their home country? Do they go somewhere else? What about American students? Do they leave after they graduate? [02:45:04] If we are to to rely on uh some you know foreign statistics for example as the economist report reported from the Chinese ministry of education 94% of Chinese students return to their [02:45:17] home country after their their studies. If that's the case then we might as well ask you know why are we funding them? [02:45:25] Why are we training them in the US? uh maybe we want to kind of put a cap on foreign students in the US. We do need to take into account that if that's what we're doing then other countries will [02:45:38] respond and for example China might you know invest in creating their own national scientific ecosystem and we will contribute to the disintegration of the global scientific community. So big [02:45:52] questions descriptively we wanted answers to what do we know about PhD graduates? Uh we know quite a lot. We know that in the US we are training about 35,000 uh [02:46:06] students per year in STEM fields. We know that they have crit critical influence on the rate of scientific and technological advance. [02:46:14] Uh but given that all the rapid changes that we're seeing in government support in immigration policy in science becoming more important for sec uh for national security in uh how AI is uh [02:46:28] changing the landscape of science we want to know more. [02:46:35] The problem is that we lack evidence on scientists career trajectories. We have a lot of good um administrative data coming from uh from the government from the NSF. We have the SED survey that [02:46:48] looks at population at the year of graduation. So we know what happens with people uh once they graduate. We have a smaller scale survey the SDR that follows a small sample over time. But we [02:47:01] lack a longitudinal population level data set that can give us the granularity that we are looking for to get a sense of people's career trajectories. [02:47:12] So that's what we were set out to create in this project. [02:47:17] We started from the SED as a benchmark. That's the orange line here which is basically the population of uh of uh STEM PhDs uh by graduation year. We collected data from ProQuest which is [02:47:30] the main repository for dissertations and seems that uh the repository followed the SECD quite well until the early 2000s where they started to u uh u diffuse because of uh uh institutions [02:47:45] leaving kind of their ties with with with ProQuest. So we uh added additional information from scraping institutional online repositories and we got quite good coverage also for more recent [02:47:57] years. Overall uh from this we have a list of about 1.3 million graduates by year uh in STEM fields where they graduated from what was their topic we have we have their dissertations uh text [02:48:11] for many of them and so that was uh the first paper in this project which was recently published. So a little bit of self-promotion. We have uh indicating uh indicators for uh who funded each uh graduate based on the acknowledgement [02:48:25] section in their dissertation. So that's available also with data if anybody is interested. But here we're looking at their career trajectories. What happens after they graduate? [02:48:38] Now where can we get data about this? If we are looking at um uh what scientists do, we can look at their publications data from open Alex. We can look at their affiliations and get a sense of where they are based on the institution [02:48:52] they're affiliated with. That's selected on being a scientist. So we complement this with data from uh reveling career profiles to cover also people who are not uh publishing. [02:49:06] As you can imagine, lots of problems with creating such a data set. Just uh kind of trying to create a a name blocking filter between dissertations and open Alex authorships. We have about [02:49:20] a billion uh possible pairs that we need to kind of filter out. Uh a lot of this is given uh is is caused by you know um uh names that are uh hard hard to kind [02:49:33] of tell apart especially uh East Asian names which is exactly the population that we're after in this project right so uh to uh to account for this we are [02:49:46] treating this as a as a predictive uh uh task and or uh training a machine learning model to kind of give us a probability of a match between a person's dissertation and a given paper. [02:49:58] Uh fortunately we have a ground truth a big ground truth to to rely on which is the orchid records. Here is my own example where I listed where I graduated for from and also my my papers that I [02:50:11] published. So I have uh true positive matches. I can also have uh all the non-matches based on uh same name. And then the first step is to train a model based on everything we have. We have [02:50:24] text similarity. We have authorship with with um advisors. We have a bunch of stuff that we can throw into a predictive model and find the anchor papers. Those that I know with high certainty uh is connected to the [02:50:38] dissertation. Once I have these anchors, I can then expand. I have much more features to throw into the model. Think of two papers that are site that are cited by a third paper or these two papers are citing something else. a [02:50:52] bunch of other features that make the model more predictive and gets me farther uh to to a far f far farther coverage of of publications. Overall during this process I got uh we got to [02:51:04] about 17 million matches to about 80% of graduates uh in our sample with very high precision and reasonable recall. [02:51:15] Um here we see uh the performance over different ethnicities uh that are derived from people's names and we can see at the top that uh the hard hardest um uh ethnicity the East Asian names [02:51:28] with very low rates of uh of baseline true matches the model performs uh uh well in terms of precision. We are paying some uh some you know recall but overall we're doing well across [02:51:41] different uh ethnicities. So let's uh let's look at an example. We have here a graduate from Stanford 2005. [02:51:51] His name is Ian Buck. He invented CUDA which is the software that is behind all the kind of AI revolution driving uh allowing us to run uh calculations on [02:52:03] GPUs. Uh we have his dissertation and through all the models we can match his uh publications. We see here stuff that is related to uh parallel programming and GPUs and so forth. We can kind of [02:52:16] clear out other papers by other Ian bucks that are doing other things. Uh we also have a false positive that uh is incorrectly matched to this person. [02:52:29] However, by using kind of network analysis, we can further remove these. [02:52:35] Matching to Revelo is uh a little bit easier because although we have many people uh people usually list their uh where they studied. So we can use that to create a match. And fortunately we [02:52:48] also have undergrad location which we are going to use as a proxy for ethnicity. So for about twothirds of the data we have uh not ethnicity the country of origin. So for about two twothirds of the data we have the country of origin proxied by where they [02:53:03] did their undergrad. we have the match either to publications um or to uh LinkedIn career profiles. So our final sample is about uh 520,000 [02:53:16] uh graduate students over the period of the last uh four four century uh four decades uh which covers about 60% of the uh of the graduates. We have more coverage. We have the full coverage is [02:53:29] about 80%, but we need uh at least five publications or at least some employment history from Revelio to get to get an indication for uh for migration. [02:53:42] So what do we find? About a quarter of uh STEM PhDs in the US eventually leave. That has uh slightly increased uh from uh previous [02:53:54] decades, but has been relatively stable over the last years. However, as you can see here, it's not immediate. People stay in the US for several years before they leave. Even those who leave stay in [02:54:07] the US for several years before leaving. And we also see that compared to the foreign national share of graduate students, there is a big gap. Most graduate students that are foreigners [02:54:18] stay for at least 15 years uh in the US. If we drill down uh beyond these kind of aggregates, we see more action. We see in the '9s kind of the uh rise of [02:54:31] migration to South Korea and Taiwan. uh possibly um you know uh people who return to their uh home countries and in more year in more recent years we see the rise of Chinese uh students u um im [02:54:46] immigrating to to China. However, if we look at Chinese students uh who uh who did their undergrad in China is do all Chinese students return to their home country not uh that doesn't show up in [02:55:00] our statistics. we see that about a third leave the US but only 18% are um uh returned to China. This is rising in more uh recent years but it's definitely [02:55:12] not u the majority of students. Next we can look at where people who are who is employing these people. So we we have the we have their first employers [02:55:27] coming either from LinkedIn or from the affiliations on their publications. What we find is that immigration is mostly driven by university and research positions. People who move usually don't [02:55:40] move because they are employed by a a company abroad. They have a position at a university. We see that uh there is um this is driven mostly by top talent at elite institutions. So top ranked [02:55:55] programs, top 10 programs uh see higher shares of migration compare compared to lower ranked programs. And among the top programs, we see uh this kind of mobility of students who move to elite [02:56:08] institutions worldwide. Even among people who go to work at firms about a quarter of positions about a third of of positions are positions [02:56:20] that are um related to US headquartered firms. So if we think people are moving and are going to work for foreign firms well a lot of these positions are actually American companies. With this [02:56:34] granular data we can actually see uh what these institutions are. So on the left here we see universities and research institutions. Max plank leading the the the leading the list uh followed by my [02:56:47] friends at the University of Toronto and uh and other other institutions down the list. In terms of companies, well Samsung is at the top, IBM. So these are uh international international centers [02:57:01] of IBM and then a long list of other companies. And as you can see here, many of them are US headquartered. [02:57:09] We can also be a bit more fancy with kind of sank keys that look at where people are coming from, where do they go, do they live to other countries? Uh, and we have several of these analyses as well that can kind of break down the the [02:57:22] data farther and farther. So, what are the takeaways? immigration rates of PhD STEM uh uh graduate students that are trained in the US [02:57:34] about 25% 15 years after graduation. Increasingly in recent years, Chinese students uh on the rise. However, most foreign origin graduates stay in the US [02:57:47] for many years and immigration is not is not immediate. [02:57:52] the most of the of of these of these movements are are kind of um related to universities and research institutions and even though those people who go to work for firms many of them are US [02:58:04] headquartered. So overall the uh the main takeaway is that the US is still benefiting from uh for people either because they stay in the US for many years or they leave but it takes them some time and even if they leave uh if [02:58:19] they leave uh the science that are producing at uh foreign uh research uh universities and institutions will still benefit uh uh US innovation. So this is a long-term project uh following kind of [02:58:33] the training of the of the uh US PhD workforce. Um coming up in a couple of days is a NBR working paper on AI diffusion among doctoral students and a couple of other projects that are in the pipeline looking at the [02:58:46] commercialization of doctoral science and scientific entrepreneurship. Thank you very much. ## Q&A (02:58:59 – 03:03:14) [02:58:59] >> Yes. score >> since you have the publication data you know this thing about US benefiting you could dig deeper into that by looking at kind of crossber collaborations like you know my co-authors in China have access to data and equipment that I don't and [02:59:12] they've gone back but like we publish together and so in the publication data can you see like uh do do say people going back to China continue to publish with uh university us [02:59:26] >> uh that's a great idea we haven't done this but we are yeah That's something that we can definitely look into. Yes. [02:59:36] >> China was also post. [02:59:45] >> So yes. So um LinkedIn is quite unique in that we have um uh kind of this you know this type of data um open Alex is is is [03:00:00] possibly has higher coverage. So if somebody publishes in China we will pick it up from open Alex. If somebody leaves and work for a company and doesn't publish we are considering ways to extend this to patents. Uh so maybe we'll be able to pick up pick people [03:00:15] from there. Uh however um you know there are you there there are limitations to the extent that we can we can pick up people if they're not in the data >> more like this [03:00:31] >> in in LinkedIn >> that that that's that's that [03:00:54] >> in in recent years. So in the last so in the last very recent years uh it's very hard to to know even in terms of you know publications are not yet there the revel data is not yet there uh to get at [03:01:07] what happened in last two years uh I I can come back to you in a couple years and I'll have have an answer take a while to to to produce it. Yeah. Did you try to predict any observe bubble to [03:01:19] heavy or is more like the US 100 to get 75 or could you target at the beginning of your studies or like something that they found? [03:01:31] >> Uh so we did a little bit of this uh I presented a little bit of that and there's more more on that in in the um in the in the paper. um we were looking at kind of ranks of universities and [03:01:43] programs and uh kind of uh measures of productivity but um yeah that's that's kind of like the product it seems that to be uh higher here I showed higher ranked uh programs are more likely to [03:01:57] improve >> so we also just saw a presentation about how PhD slots are decreasing particularly at top programs it I know most people stay even if they come from [03:02:10] foreign countries. But how does that compare to the alternative where the slots that are going to those people go instead to people from who are even more likely to stay? [03:02:21] >> Um that's fair enough. That's I mean that's a question to to explore and I don't have an answer >> just but you show this [03:02:33] >> almost 95% of Chinese go to M that's really not in your data but these are undergrad students >> so yeah so this statistic I tried to figure out what is the it's a different population it's not only the US I think it's uh different types of students but [03:02:48] you know there there's no data other than what we here >> as far as I know >> that people get sometime China. [03:03:01] >> I think the Chinese government tracks. >> All right. Thank you very much.