Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
========================================================================

# Borrowing for Breakthroughs: Student Loans and Innovation

Authors: Alex Bell, Vikram Jambulapati  
Discussant: None  
Video: https://www.youtube.com/watch?v=7LDwpGoPNFc&t=8702s


## Talk (02:25:02 – 02:43:12)

[02:25:02] >> Okay. Thank you very much. Let's move on to our next talk. uh the uh Alex on borrowing for breakthrough.  
[02:25:09] >> Thanks Glenn. Thanks. Excited to be here and sharing my joint work with Vicram Jimati who couldn't be here today unfortunately on uh borrowing for breakthroughs student loans and innovation. Uh the motivation for this talk really comes from two  
[02:25:22] distinct sources. The first motivation is um sort of my lost Einstein's line of work. I think it's an observation that many of us have made in many different contexts. some of it within the econ literature, some of it outside the economic literature, that systemic barriers seem to stifle the innovation  
[02:25:36] of many talented youth. Um, so the way that I think about this graphically is just shown in the slide on the on the right, which is from my previous work.  
[02:25:44] Um, looking at how even as young as third grade, math test scores really have a strong predictive power of who goes on to innovate, but that's not actually true by different races, different socioeconomic statuses. So there seems to be something about a misallocation of talent here that we  
[02:25:57] maybe don't quite know all the the the factors that characterize that misallocation of talent, all the factors that lead to it. But we know it's big.  
[02:26:05] We know that if you could um close uh disparities in the United States by gender, race, and parent background, back of the envelope calculation would suggest that could lead to as many as four times as many inventors. And when you think about it, that's sort of leaving a lot of innovation on the table. We could maybe be growing the pie  
[02:26:20] bigger than we are now. So that's the first motivation. And the second motivation, sort of somewhat orthogonal, but I'm going to try to connect them for you. Um, is there's been a rapidly evolving policy landscape uh on student  
[02:26:32] loans in the last 5 10 years. Um, a lot of changes to qualification standards, talked about, income driven repayment plans, uh, always in the news, potential sort of blanket onetime um, uh, debt forgiveness, things like that.  
[02:26:45] there has been a a sort of a nice literature emerging around um student loan policy but the contrast I want to draw is a lot of that literature has been focused on the student um it our changes to student loan policy helping students succeed um so the black at all  
[02:27:00] paper I think does a really nice job of that I'm I'm going to borrow um some research design from that in a few slides there's also some sort of um there's a bit of a tension in the literature that we see that student debt sometimes pushes students away from  
[02:27:14] public interest jobs. So in that way there could be sort of like benefits and and drawbacks to policies that change debt. Um so I want to ask the question here of what does all this mean for the size of the economic pie? Can we  
[02:27:27] understand the impacts of student loan policy not just on the well-being of students and maybe not necessarily just on the career choice but on a an outcome that we care about for social spillover reasons like innovation.  
[02:27:41] So the question I want to put forward here is how does student loan policy affect innovation? And there are sort of a lot of channels that I think smart people might might really think about.  
[02:27:49] Um some of them go in the positive direction. Some of them go in the negative direction. There's a sense in which you know expanding student loan policy is going to enable access to education that's going to lead to all sorts of good human capital formation and we'd see more innovation coming from  
[02:28:03] that. But I think there's a real serious um sort of counterveailing threat here which people sometimes call debt overhang in the literature that as you expand people's access to to debt, not only will it finance their education, but it's going to make them say like,  
[02:28:17] you know, hey, when I graduate, should I really be going into, you know, innovation, the sort of risky career with social spillovers, or do I want to go into banking where I can actually make enough money to to repay my student loans? So, I think there's competing hypotheses here about how student loan  
[02:28:31] policy um might theoretically uh affect innovation.  
[02:28:36] And I want to try to answer this question because it's sort of a a new question in some sense. I want to use as many sort of known good building blocks in the literature as I can. So, let me walk you through the the data and the methods that I'm going to try to sort of  
[02:28:50] cobble together from existing literature um to answer this new question.  
[02:28:55] Um so data set I think you know the most familiar data set I'll use to to people here probably the US patent and trademark office data. I'm going to subset for the purposes of this paper to um about onethird of US inventors that ever lived in California. So we're going  
[02:29:10] to be focused on California because that's where our student loan data comes from. It's only one of 50 states but it's actually a really big state when it comes to innovation. So it's not that restrictive fortunately. Um you know sort of characteristic of the literature here. I'll think of patents as a proxy  
[02:29:25] for innovation. Sort of a rough proxy, but not necessarily the whole story, but it's a way to operationalize that innovation channel as an outcome.  
[02:29:33] And then I'll link it through what's called a a hashed merge technique. And if you don't know what a hash merge technique is, good, stay that way. But it's a very difficult, cumbersome way to to do a merge when you can't really see clear text of either data set for for  
[02:29:46] for privacy reasons. So, so we link the the patent data to the UC University of California consumer credit panel um which gives us 20 years of complete credit history for all the borrowers in  
[02:29:58] California as well as a 2% uh US sample. So, this is coming from one of the three credit bureaus. Um so, they'll have information on on various types of loans, but we're going to be focused on student loans for this project.  
[02:30:13] So, that's a little bit of the data. The design is going to be following very closely um to a recent AER paper um from uh black at all called taking it to the limit. I'll walk through the identification strategy in a few slides  
[02:30:28] with you. Um but the identification intuitively is going to come from these large ad hoc increases in the Stafford loan limits um that occurred in the early 2000s.  
[02:30:38] As we walk through the identification strategy, it's important to think about what's going to come out of this identification strategy as a local average treatment effect um in a couple different dimensions. And I think the way that I've thought about this, all that's probably going to lead you to find a little bit of a larger effect  
[02:30:53] than some other counterfactuals that you might have in mind here. So the first way that it might lead to a larger effect is that constrained students might have larger responses. If you think about sort of like a Roy model type logic, there might be something a little bit different about the student  
[02:31:08] who kind of can't afford college but decides to go to college anyway. They might have private information that the returns to college and education are are really high for them because they could do something really great with it.  
[02:31:19] The second reason that I think the late is likely to be a little bit inflated relative to other policy counteracts you have in mind is it's occurring in the early 2000s. And this depends how sort of optimistic you are about how much change there's been since the early  
[02:31:33] 2000s, but I I I have in mind sort of like the Sha Hurst Jones cleanout kind of model of misallocation of talent.  
[02:31:40] Maybe that's getting better over time as we see more people from underrepresented groups becoming represented in top professions. So to the extent that that's true, going back in time, we might see relatively large effects on policies that remedy misallocation  
[02:31:53] relative to what we might see on similar policies today. So we want to keep that kind of late interpretation in mind as we go.  
[02:32:05] So this is the the research design in a nutshell.  
[02:32:09] These are the first year Stafford loan ceilings. So if you go to college today as a freshman, you can borrow uh about $5,500, but that wasn't always true. And going back in time in the early 2000s, it would have been half that amount. So the  
[02:32:24] loan limit ceilings have increased a lot and we're going to ask sort of what's happened to the patenting rates among the cohorts that have been imp impacted by these increases.  
[02:32:33] You can see that that that sort of bunching response in a nutshell here also. So here I have a histogram of the amount that people have borrowed in the data set and you see the sort of first big bunching point. That's the indivi individual Stafford loan limit here. In  
[02:32:47] the early 2000 cohorts it was something like $2,600.  
[02:32:52] And then in the next cohort, oh sorry, and then you you see other bunching points. So we're going to mainly use the the um we'll only use the dependent loan limit, but you see other bunching points at for instance the non-dependent loan limit here. And then for the next  
[02:33:06] cohort, you see that limit has moved again. So they've increased that loan limit by about $1,000 per year. Now you see people bunching at the new bunching point. So it's all these students who are constrained, but the constraint has has lifted somewhat. So that's the sort  
[02:33:19] of quasi experimental variation in the X variable of of student loans.  
[02:33:25] Built a little bit of graphical intuition for the research design before we move to the regression specification.  
[02:33:30] Here I have first year borrowing on the x-axis and just kind of like an OS type representation graphically of innovation rates. And what you see here is if you look at that that last point of first year borrowing the students who in the  
[02:33:44] early cohorts the blue is the the earliest cohorts we can observe the students at that bunching point have a lower innovation rate and then the bunching point in the next year moves up then the bunching point in the next year moves up. So as the loan limit is increasing there is some sort of  
[02:33:57] graphical evidence that the students who are constrained are bar are um innovating at higher rates.  
[02:34:05] With that graphical intuition, I want to move to regression specification. Um, so it's trying to follow black at all as closely as we can in our data set. Um, we're going to uh regress the outcome uh  
[02:34:17] in this case of innovation rates on uh cohort fixed effect. So that'll account for cohort level fluctuations. The constrained indicator for the early cohorts that's borrowing at the limit and for subsequent cohorts borrowing at at the earlier limit through through  
[02:34:31] their new limit. And then the way to think about this coefficient of interest here, it's the coefficient on the interaction between whether you're in the constrained group and you're in one of those years after which the constraint has been lifted. The way I think about that, how do the outcomes of  
[02:34:43] the constrained students change when the policy exogynously lifts their constraints? Um, so that's the kind of like diff and diff style of identification. And there's also an event study version where instead of that one interaction, we're going to have the interaction with all the  
[02:34:58] periods here. So let's see what the first stage looks like. Here I have the cohort from 2002 to 2013 with that sort of like policy change region in the middle and we have the effect of first year debt going  
[02:35:12] here. So the baseline is normalized to to zero and we see increases in firstear debt around the time of the policy changes. So this makes sense as the loan limits are are increasing students are taking out more debt.  
[02:35:27] So you can think of that graph as the first stage and you think about this as a reduced form result. So the same x-axis, same sample, but instead of putting um loan limits on the vertical axis, I'm looking at the effects on patenting rates. So we see as the loan  
[02:35:41] limits are are are getting raised when that constraint is lifted, there's an increase in the patenting rates among the cohorts that are impacted too.  
[02:35:51] to think about the magnitudes a little bit more. It's it's kind of useful to move to an IV specification. So in the IV specification, you think about, you know, the usual we're going to regress patents on on on loan rates, but in this case, we're going to instrument for  
[02:36:04] loans with this cohort level variation and constraints.  
[02:36:10] So here I I have the OS specifications and IV and OS kind of consistent with the graphical uh evidence. OS doesn't show any real relationship between borrowing and future innovation, but when we move to the IV specification in the last column, we're finding a  
[02:36:25] significant increase in patenting rates among the students that are impacted by the lifted constraint. Um, the number here is a little bit difficult to interpret, but let me walk you through a little bit about how to think about that magnitude and try to build up some  
[02:36:38] intuition relative to to other effects you might have in mind.  
[02:36:42] So the point estimate says that a $1,000 more loans moves innovation rates basically from a very small number to another very small number. That's actually bigger in percentage terms, but these are small numbers overall, of course. Um so it's a large percentage  
[02:36:56] increase. It's a 74% in increase, but it's um uh imprecise. So the lower bound, we can't rule out a 25% increase from $1,000 loans. Also, what do you make of $1,000 loans? it's probably easier to think in terms of elasticities  
[02:37:11] here. Um, and as we go, we also want to bear in mind, again, like I prefaced earlier, these are local estimates.  
[02:37:18] They're among students who we think might be lo likely to to be most responsive here. But in order to make the point estimate portable in relation to other kind of literatures here, it's useful to think about an elasticity of  
[02:37:31] of patents to loans, basically just dividing through by the mean loan rate.  
[02:37:36] So, we end up with an elasticity of 1.9. So it seems like a relatively high elasticity as far as elasticities go. Um it's imprecise. So it's a got a confidence interval that sort of can't reject par or or some lower numbers.  
[02:37:48] Also it's relatively similar to the elasticity of stem degree completion from the original blackall paper that they observe. Also let me think let me tell you how I'm thinking about putting those magnitudes in perspective. One way I I think about  
[02:38:02] it is you know what are other margins that we know seem to move patenting also. So I had that original motivation of uh sort of the lost Einstein's kind of literature and one thing that came out of that literature is it does seem like exposure to uh adult innovation  
[02:38:17] whether it's through the childhood commuting zone or through parents in particular um seems to to to cause kids to innovate more also. So we can go back and look back at that literature and calculate an elasticity from that. And  
[02:38:30] what we find here is that the elasticity of patenting to student loan policy is about six times the elasticity of patenting to childhood exposure when measured at the CZ level. So it's it's a a more responsive margin that than that  
[02:38:44] type of childhood exposure measure at least when measured at the place um uh uh level.  
[02:38:51] I I think a productive sort of direction to extend this in and this is why I have a crowd sourcing call for for everyone here to weigh in on also is you know I really want to hear from you on on what do you think are the similar reference points based on your own experiences uh  
[02:39:04] your own research also um you know what are the ways that people think about programs and policies either to get more innovation or to get more patents in particular so I I I have you know sort of a couple reference points here that I think about the sort of summary I'll  
[02:39:18] I'll put out there for you is our point estimate looks somewhat high relative to these benchmarks of other programs and policies that seem to either get more innovation, more patents, more R&D. Um although it's an imprecise estimate with a wide confidence interval and it's also  
[02:39:32] an estimate in a case where we might expect the late to be a little bit larger than than other average treatment effects that we might be able to get. Um so I want to put a pin in this and I'll tell you at the end about sort of like what my takeaways are and whether my priors are moved and and you'll be thinking about that too. But before we  
[02:39:47] get there, I want to try to unpack the result a little bit by telling you a little bit about heterogeneity and uh about mechanisms here in terms of heterogeneity. I think it it shouldn't be that surprising that we just don't have a whole lot of power to say much.  
[02:40:00] And I think there's really interesting questions here. You know, is the loan policy uh changing what margin someone innovates on in different technology classes or something like that? Um I want to just say, you know, we we've checked heterogeneity by gender, race, and technology category. we just can't get a lot of significant differences  
[02:40:14] there which I think is not not heavily surprising. But I do have something a little bit more concrete to say about mechanisms. And I also want to sort of loop back to that kind of lost Einstein's motivation. I want to think about who is the marginal wouldbe  
[02:40:27] inventor who we'd be losing if they if not for the student uh student debt policy. You know, who's the kind of student that that says that's too much debt to go to college. I don't want to do it. You know, I don't I don't want to  
[02:40:40] go on to innovate. One hypothesis is these could just be marginal lowquality innovations that we're losing. The alternative hypothesis is sort of lost Einstein's hypothesis. There sort of may missing at random. Maybe there could be some breakthroughs that we're we're  
[02:40:53] losing through this type of uh margin. And I want to test this hypothesis a little bit using detailed citation data as a sort of a rough proxy for the quality of innovation. I'll rank patents by citations within a year to account for censoring.  
[02:41:07] Let me show you what I find. So here the first column is our baseline result that there's sort of this uh the 74% increase in patenting rates overall. When I make that increasingly more granular, I say well what if the outcome is not any patent? What if the outcome is having a  
[02:41:21] patent that's been cited or having a patent that's in the top quartile of citations or the top 5% of citations?  
[02:41:28] What we see here is uh there's no if if you if you if you sort of doubted the lost Einstein's hypothesis, you would expect that all these elasticities would be falling. Um but we what we don't see is what what we see is there's  
[02:41:42] no evidence that these uh marginal innovations are of any lower quality.  
[02:41:48] And if anything, the point estimates seem to suggest they're higher quality, but there's no real statistical significance behind that statement.  
[02:41:56] So I want to wrap up there. Um just to recap, you know, I laid out some competing hypotheses on how student loan policy impacts innovation and we brought in a large scale linkage of admin patenting data and credit records and a quasi experimental expansion design  
[02:42:10] following a recent AER paper by Blackall.  
[02:42:13] So how do I update? I think the clearest takeaway for me is that the student loan reforms of the early 2000s probably look like they have uh had quite a lot of value in uh producing innovation among the students that were impacted. Um what  
[02:42:27] does it mean for today? I think it's a little murkier for me and I want to be be clear about that. I would say my prior on the role that higher education financing can play in innovation probably has updated somewhat. Um there's likely an important role that that that uh loan policy for higher  
[02:42:42] education can play in innovation. Quantitatively, I think I have a a more diffuse prior on this. Whether we'd see as large effects of increases in access to credit today, I think that depends on how we think about who the marginal  
[02:42:55] borrower is today. Um I think what would help most now in advancing this literature and advancing this question is more studies and especially on more margins of education financing so we can understand how different margins shape  
[02:43:08] the innovation decision of students. So I'm going to leave it there for today.  

## Q&A (02:43:12 – 02:44:27)

[02:43:12] Thanks again. We time for a couple questions while sets up this slides.  
[02:43:28] What do you think of the do you think is like students going to STEM instead of some other discipline and uh that's why or do they increase studying at all?  
[02:43:40] >> Yeah, I think it's really hard in this data to get into the mechanisms. I can tell you a lot about sort of the mechanism that we hypothesized going into it is that if they're not borrowing for school um they might be free to do something else. So basically if they don't have to sort of work during school  
[02:43:53] to repay their loans, they might have more time to be sort of creative, flexible, explore the topic in a way that they wouldn't be able to do otherwise if they were working to repent already while that's something that you  
[02:44:07] could see the time.  
[02:44:11] >> I think it's it's a little bit too imprecise to do a lot on the timing. We might be able to explore a little bit more on the timing and see if if we push does it turn to mush. It's a question I would give to more than thank you.