Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable. ======================================================================== # Long-Run Effects of H-1B Immigration on the U.S. Economy Authors: Ran Abramitzky, Leah Platt Boustan, Ahmet Gulek, Jens Hainmueller Discussant: None Video: https://www.youtube.com/watch?v=7LDwpGoPNFc&t=7467s ## Talk (02:04:27 – 02:23:25) [02:04:27] >> Thank you, Glenn. Um, so, first of all, I speak fast and in a 20-minut presentation, I speak even faster. If you don't hear, if you don't understand anything, please stop me to like ask questions. All right, with that, um, I'm Amit G. I'm a postto at Stanford. Um, um, this is joint work with Rena Bitzky, [02:04:41] Leustan, and Yan Simuler. The big question that we want to understand is how do highkll imig immigrants affect firms, um, industries and regions that they work at, and how do these effects propagate throughout the economy? [02:04:52] There's a large literature studying the first of these questions, the direct effects of im immigrants on several outcomes, including m wages and employment, innovation, and tax specialization. [02:05:02] More specifically, there's a growing literature um studying the effects of H1 policy, the primary uh visa program for high school immigration to the United States. There are three main designs in this literature that I will shortly uh go over, providing very valuable [02:05:15] insights, but some with somewhat mixed evidence. The first of these approaches is the standard cross city shift share design where you essentially compare regions that receive H1B workers to regions that don't. And using this design, Kur and Lincoln find that H1B immigrants increase patents in patenting [02:05:30] in the host regions and Perry and co-authors find that HB immigrants increase wages of high-skilled natives there. These um results later are later challenged by Doran and co-authors who exploit the lottery induced firm firm variation. So the when the cap when the [02:05:45] H1 demand exceeds the cap the which firms acquire H1 workers are determined through a lottery. This provides a great identification design. Comparing the firms that randomly acquire H1B workers, they show that the H additional H1B worker results in significant native [02:05:59] displacement. Um their results themselves are um challenged by majors who show that using very similar design but using more recent um lotteryies they find no displacement of native born workers. The third approach in this [02:06:14] literature is structural. um you ban and quarters. So Goka and the Monles showed that they using partial and general equilibrium approaches. Um im H1B immigrants can lower the wages in um computer science. But these decrease in [02:06:28] wages themselves lower the prices of goods which then results in benefits acquiring in complimentary occupations or ini final goods um final good consumers. Now the reason why I want to tackle this literature is the whereas these approaches provide very good [02:06:42] valuable insight um in my job paper I showed that the when immigrants change the price of goods they the effects become global not local effects propagate along supply network and unfortunately the cross city and firm level designs uh partial out these indirect effects and in a structural [02:06:57] approach takes you far but it's relying on strong um structural assumptions and you cannot emply measure these spillovers and we want to do do that. So essentially in this paper what we're going to do is we're going to ask whether the uh what are the direct and indirect effects of the H1B policy on [02:07:11] the US economy. To do that we're going to rely on industry level shift share approach. Why industries? Two reasons. [02:07:16] In two slides I will show you that first H1B policy is essentially an industry policy. Year after year it's the same industry that acquire H1B workers. And secondly in the US we actually observe the input output network at the industry level. So any kind of industry shock you [02:07:29] can pro you can measure their IO input output spillovers. We want to do do that as well. The ship that we're going to rely on is the uh the relaxation of the cap between 1999 and 2004 during which approximately 490,000 more visas were [02:07:43] were allowed. The sh the share is going to be the indust the pre pre-shock industry's demand for H1B workers. Now as we know from recent literature in shift shares you can get identification from either share exogenity or ship exogenity. We're going to rely on share exogenity and we will relax that [02:07:58] assumption by a bit through applied scientific controls using a method that I developed in a different paper. I will what I will use in this paper. I will go over the method in like couple of slides but um so as data it's we're going to rely on every publicly available data source that we can find. So census ACS [02:08:13] CPS QCW patent data from Kogan and coers and input output network data from BEA. [02:08:18] I warned you I speak fast. So here's a a couple of like summary results. First we find that the um H1B immigrants increase the wages of natives and pre-existing immigrants in non- STEM occupations. [02:08:32] Um we find absolutely almost like no losses among any type of native workers or high skill low skill STEM versus non- STEM. Some of our estimates are small like negative but very small and pre and imprecise enough that we cannot statistically conclude that there any [02:08:46] natives are hurt. Um we find that effects propagate forward to customers and not back not much backward to to suppliers. What does it mean? For example, workers in insurance industry which rely a lot on computer computer science inputs those those uh workers [02:09:00] benefit from the H1B workers arriving in into computer science industry. Now the forward effects being positive is interesting in itself that the welfare effects uh for welfare purposes more people benefit from H1Bs but the forward effects being much bigger than the [02:09:14] backward effects is theoretically interesting. It highlights that the H1B immigration is not only a labor supply shock. I will explain it during the model section but we can actually estimate through through the model how much of a productivity gain the H1B immigration must have caused to be able [02:09:29] to create those kind of strong forward but not much back backward effects and we find that a 1% immigration like shock of of H1B workers is associated around like 4% increase in honor productivity uh where is this productivity gains [02:09:42] coming from we find no no effect on patents so no increase in patterns which causes us to conclude that the effects come productivity gains come from better execution of skilled tasks. And what we are thinking about is um a superstar software engineer in 2000, they can [02:09:56] improve the software. They're not going to patent it, but that impro that software being much better is going to allow like um improve the productivity on those industries and benefit all the consumer industries that rely on computer inputs. [02:10:10] Um so this is a again if you have any questions feel free to stop me, but otherwise I'm just going to rush forward. Um this is showing the U H1B caps across here. So the program starts in 1990. The cap was around 65,000 until 1999. In the eve of the.com bubble, [02:10:25] Congress relaxes the the capacity from line 65,000 to 115,000 and then to 195,000 before letting it drop back to 65,000 and adding another 20,000 for masters students and stay 85,000 since 2005. [02:10:39] Now I told you that the this the um HRB is mostly an industry shock. Let me provide some evidence to that. This is showing the total number of H1Bs acquired by um highest sponsoring firms [02:10:52] since 2009 the earliest beginning period of for the publicly available data. So as you can see the Cognizant Tata and Infosys they were by far the highest beneficiaries of the H1B workers in early 2010s but later they were [02:11:05] overturned by the Amazon Google Facebook and Apple. So which firms acquire H1Bs over time changes. But if you look at which industries acquire H1Bs year after year, it's going to be one a few industries that dominate the system. [02:11:18] Computer systems is design acquire between like at worst 20 at best 50% of all H1B H1Bs given in any year. They are followed by universities which are capex exempt and then a bit of engineering and consulting and then they dwarf everyone [02:11:32] else. But this is just total number of H1Bs. What economic instance is based on among 100 workers how many are H-1B workers. If you look at that exposure, this is the Y-axis is the H1B per 100 [02:11:46] number of H1B uh H1Bs per 100 existing workers and the X axis is the share of college graduates like skilled intensity of the industry. [02:11:56] As you can see that a couple of industries that are heavily exposed the computer and data processing services, radio, TV and communication are heavily exposed to H1Bs. The computers engineering etc are also exposed uh but but they are clearly developed by the computer and data processing services. [02:12:10] These are mostly high skill intensive industries acquire H1B workers naturally and in our identification we're going to compare high skill high skill intensive industries with each other to get it to get at the effects of H1B workers. So let me give you what kind of identification we're going to rely on. [02:12:24] Oh I have time. Great. Speaking too fast. So the um expo we're going to rely on a standard event study design right the so we're going to um interact the exposure with time time dummies controlling for industry fix effects um [02:12:38] skill by quart um skill quantile by time fix effects always shows pre-trends I won't show you the those for the sake of time but we're going to apply a what we call a scientific control defin approach to adjust for those pre-trends. So how does it work for each we're going to do [02:12:53] this for each industry in the algorithm for each industry I in any skill group S we're going to define a donor pool as every other industry in that skill group right once you you solve the scientific center control problem by um solving on [02:13:05] the u by matching on the trends of your key outcomes for us this going to be non- stem and stem trends trends once you have your synthetic control weight you the unique part is you generate your synthetic data which is consists which consist of not only the outcome but also [02:13:20] the treatment because we are in the world of continuous treatment not in the simple world of the binary treatment. [02:13:25] Once you have your scientific data you difference it out from your from your actual data and it's granting you what we call a device data. Once you have your device data you apply OS. Now we didn't invent this approach. This is the reduced form approach of the synthetic [02:13:38] IV estimator that I developed with with Jamaida. This was his job mark paper from two years ago. [02:13:44] Um there are a couple of there are only few effective clusters. So cluster centers don't work. So instead we're going to rely on randomization inference following FM and Pinto to for inference. [02:13:54] Now let me highlight a couple of results. So I'm going to show you the results for native college um for for four groups. Native high skill college graduates, native uh less than college and pre-existing immigrant college versus non college. And within each group I'm going to look at the non- STEM and STEM occupation separately. So let [02:14:08] me have a couple of things. So then high school native we u their incomes go up and we can across all of our robustness checks the effects remain more or less significant. [02:14:18] non stem we the coefficient is negative but super super small and quite imprecise. This is not a precise null. [02:14:26] So we cannot um we cannot rule out say minus 1% uh decrease in wages. The senators are large enough that we can only rule out like large large negative effects but we cannot detect any significant decrease. We thought perhaps [02:14:40] the story is that im new immigrants substitute existing immigrants which complement natives but we don't think that's that's the story either. [02:14:47] Pre-existing immigrants are not substituted their incomes if anything are are definitely not going down. If anything they are going up seems like the overall non- steam occupations workers are benefiting from the H1B shock and like long run effects are very similar to short run effects in in this [02:15:02] design. Yeah. Um red is the non- stem and blue is the stem. Oops. [02:15:12] Yes. So the first top two are natives. The the bottom two are pre-existing immigrants. So immigrants who arrived in the US before 2000 wages. Well, annual income from in the ICS. But then we we use the CPS to look [02:15:26] at hourly wages exactly the same. So it's not the week's work, it's not the hours work, it's it's hourly wages. [02:15:34] Now there are several things that can go wrong in an industry level design. We cannot rule everything out but we rule rule out a large number of checks that we at least we try to do our best. Um for example the across industries we we are worried about contemporaneous [02:15:47] productivity shocks. We compare industries with similar patent growth with similar industry productivity growth and with similar exposure to to China shock which is happening in contemporaneously during this time. In all of those results we find the exact same results. Uh we purge out regional [02:16:02] shock. So you can be you know the Silicon Valley is being produced during during this time. We can assume that the regional trends are projecting towards the industry in our design. We we partial regional regional effects out drop hubs the not results change. Uh we [02:16:16] drop the cab industries out results become stronger. We drop sway fails later because I will show you in two slides that there are downstream effects. We take out those industries results become stronger. Um we also have a triple defendive design in the [02:16:30] background where we compare not only the high skill intensive like the H1B exposing industries to other industries but also look at the fact that the in in redwood city there's like a lot more H1B H1Bs arrived in that commuting zone comp or San Jose gets a lot more H1Bs [02:16:43] compared to Utah and you compare the industries in those regions you get acquire the exact same results and there are a bunch of econometric like the checks the randomization inference has good coverage we apply back testing we we we we don't find any placebo [02:16:57] place effects. So I think our results are more or less robust. Now I want to go over some of the indirect effects which I think are super interesting for this paper. So the forward effect is going to be the how much industries buy from H1B exposing industries. Sorry and [02:17:12] backward is how much you industries sell to H1B exposed industries and we look at them by dropping out the most directly exposed industries from the sample and running the regression on forward and backward. What we find here is the red is the forward effect. That's like I'm [02:17:26] buying from computer science. Um and backward is the blue I'm selling to science. The top is what we get from QCW. The bottom is what we get from ACS. [02:17:34] Different different data sets we acquire the same results which means that the forward effect is significant in the QCW. The um customers benefit suppliers not so much. So insignificant and substantially smaller. Now this is I think interesting in itself but I will [02:17:49] explain why. I think the we we argue that the forward effect being much bigger than the backward is is theoretically super interesting. So to do that I'm going to show you a couple of um couple of figures. So imagine three industries supplier sells to J and [02:18:02] JS to their customers and suppose like for for now that the H1 H1B is purely a labor supply shock. Labor comes in production costs go down prices go down. [02:18:12] When labor and intermediate goals are gross compliments it creates interesting um G G effects. So consu from the consumer's perspective in in input costs are going down and if when input costs are going down if inputs and labor are [02:18:26] gross compliments the customers industries are going to spend more on more on labor increasing the wages in the in the for in the customer industries that's the positive forward effect but the same but the same complimentarity will result in a positive [02:18:41] very strong backward backward effect as well why because from J's perspective wages are becoming cheaper and J would spend then more on input goods input goods coming from the suppliers. So we should have observed under a pure labor supply shock we should have observed [02:18:54] exactly or similarly significant and strong backward effect and in the data we can actually reject a pure labor supply story from from the model. So what's going on? [02:19:05] We argue that the what's the missing part is the productivity. Let me see. [02:19:09] Yep. The missing part is the productivity gain. So what's happening? [02:19:13] Imagine now there's no techn for a second. We we hit the industry J with a productivity shock. From the consumer's p perspective, it's the same effect because from the consumer's perspective, whether the prices are going down because labor is cheaper or productivity is better, it's the same. [02:19:28] The input is going down. I if there are compliments, I'm going to spend more on labor, wages are going up and there will be no direct backward effects. That's that's the model 2012 paper. Um but there's an there's a sec not a second order but a indirect effect coming from [02:19:42] the fact that as the input costs of consumer C is becoming cheaper they spend more more on labor so they reduce their spending on the inputs overall that reduces J's demand from consu the supplier S which would in general [02:19:56] equilibrium would result in a decrease in demand for S. So technology shock by itself would have lowered the wages of the of the backward of of the suppliers. [02:20:05] Now we argue that the two effects of of the backwards will are canceling each other out in the data. We can this is there's a single crossing argument with no productivity shock we we would have found the backward effect to be much bigger than the forward effect in the [02:20:19] data. um with the productivity with the with the productivity effect being at particular level we that that ratio would be going down the particular using the our estimates from the QCW we would estimate that a the immigration the H1B immigration beyond increasing the labor [02:20:34] supply by 1% should be should have also increased the productivity by around 47% 47% to create that back forward versus backward effect that we observe in the data and then using the CSS we would have found a similar 0.33 take the [02:20:48] averages of two different samples 0 0.4 is our estimate for how much of a productivity gain H1Bs are causing. Now with that I want to say like there's a reason why we wanted to work on this paper in the first place. We are living [02:21:01] in a very interesting times. um the with with the additional 800,000 fee for currently for the out for um workers coming applying from outside of the US but there's are um it's unclear whether those fees will be applied to people [02:21:15] within the US or like how the how how the uh what the future of the H1B pro uh program is is going to be like what we know is that the right now the demand of H1B H1B demand is dropping by a large [02:21:28] amount it's a $100,000 a big cost for most firms Um so let's see based on our um estimates what if the program ends what what can we say based on our estimates today. So let me give you a couple of back of the envelope [02:21:41] calculations. The 85,000 few H1B workers today would result in a uh for every year a decrease in labor supply for the computer services by around 1.1%. That alone uh would reduce the productivity in computer sciences by.5%. their [02:21:55] productivity um decrease would result in the income of non- STEM native college graduates to decrease by around 1.2%. [02:22:02] The incoming customer industries would also decrease um based on 1997 production network they would have decreased by around.5% but computer science today is a lot more central than it used to be based on 2017 numbers they [02:22:16] they would decrease by 7%. Today we are living in the AI world. How the AI the computer science is becoming even more central. Therefore negative effects would be even stronger. Real GDP would decrease by around 13% of which the half of it would come from productivity [02:22:29] effect alone. Um so and these effects will accumulate year after year based on the like more people like more people didn't come in year after year these effects would accumulate more or less linearly. Now the main takeaway of our paper is that H1B is discussed as an [02:22:43] immigration policy which at its at its core is is an immigration policy but from our estimates is also a technology policy because H1B provided them with some of the highest talent that the US US has providing when they entered more the productivities have increased when [02:22:56] they enter less productivities will go down. [02:23:00] Restricting it is like is less like shielding workers from cheap labor and more like flowing technology the whole economy runs on. Um to to close um speaking of H-1B H1B visa and talent um excuse me but I I am on the short market [02:23:13] looking for an H1B visa. Um and if you're hiring um Josh is one of my adviserss um and um he can he can give you hopefully a good word about uh my portfolio. Um this is me. Um thank you so much for listening to me and if you have any comments we are more than happy ## Q&A (02:23:25 – 02:25:02) [02:23:25] to listen to you. So thank you. I do need an H1B >> and so we can take a couple bits of [02:23:38] questions. Um, Alex, I expect set up the slides while uh on it starting questions. [02:23:48] >> So I have I have a question. Seems that the shop that you're studying is basically an IT industry shop, right? [02:23:56] would we call it like a basically if you exclude the IT industry the results would just basically not be there is that like I'm correct prior so in the in the simple effects we dropped the computer science industry to begin with [02:24:10] because it's mostly directly exposed right so the the forward versus backward effects are coming straight from uh the nondirectly exposed industries when we look at the direct effects you are absolutely right the the moment we we eliminate that one industry it's heavily [02:24:23] exposed the confidence intervals become so large that the coefficients are like relatively similar but still you cannot um you cannot rule out a null effect and you cannot rule out that without the computer science the effects are different than the effects of computer [02:24:36] science. So that's why the the uh cluster standards don't work for example that's why we need the randomization randomization inference. [02:24:44] >> Yeah. >> Are you able to distinguish between H1B for people who are already here? [02:24:54] >> Great question. Absolutely not. I wish >> Yeah. [02:24:59] >> Yeah. >> All right.