Robots for the retired?
Do aging populations and a scarcity of workers drive the adoption of robots?
The source recording for this episode has no speaker labels, so this transcript runs as continuous dialogue without attributing individual lines.
>> Welcome to the Justified Posteriors Podcast, the podcast that updates beliefs about the economics of AI and technology. This is Seth Benzell, apparently wasting my youth in comparative advantage by not specializing in production tasks directly coming to you from Chapman University in sunny Southern California. This is Andrey Fradkin, ready with a cross-country regression whenever I have a theory I need to justify. Coming to you from sunny San Francisco today. >> A cross-country regression. Yeah, we're going to see a lot of these in the paper today. I guess we'll talk about some. What do you think when you see a paper with a cross-country regression?
I know that's not something that usually gets you excited, Andrey. >> It gets me excited to make fun of it. >> No, no, I mean, look, this is the difference being academic economics and practical questions, right? In the end, there are certain issues in which we have nothing better than comparing countries, and so we have to look at this evidence even though how much we update based on it, may not be as much as the authors of the paper would help. >> Yeah, that's a fair point. >> But I guess we'll introduce the hypothesis that's under consideration today very shortly. But I think the hypothesis isn't necessarily only going to be at the country level.
So they'll try to do some things at the sub-country level, but also going to find limited data issues. So the question that we're talking about today is from a very concretely named paper by our two, I would say heroes or people looked up to by this podcast, Derona Simoglu and Pasquale Restrepo. This is a conversation based on their paper, demographics and automation, which asks the question to what extent does the decision to automate is that a function of the demographics of your country, and they're going to take a very particular take on why demographics are important. Why do they think demographics are important, Andrey? >> Well, if you need to get some stuff done, and you no longer have able-bodied humans to do those things, so carrying stuff, screwing stuff and walking around, etc.
Then it might be worthwhile for you to invest in a technology that can do some of those same things. So that's the automation of hypothesis. So if you have less labor, you have a labor scarcity, you're going to automate more. >> Right. We should talk about that this paper is looking backward at the period like approximately 1990 to the late 2000s, where they're specifically the automation that is relevant is physical task, industrial robot automation. That's not to say that there aren't other kinds of automation happening at the time. But when the title of this paper says, automation in demographics, you should think industrial robots, because that's what this paper wants to talk about. >> Yeah.
So if you're, let's say, manufacturing a car, we're talking about the sorts of robots that might be in the factory that are combining the different pieces together and moving them around. >> Exactly. So we're thinking about these big robot arms coming in, and the hypothesis is as you employ that regions that are older, that have too many old workers relative, excuse me, to middle-aged workers. If older workers are specialized in not being really small and lifting heavy things, but rather other skills, it's hard for me to rack my brains to think of what you could possibly be good at other than being small, Andrey. Maybe you'll give us some ideas that then you're going to want to differentially adopt these technologies that substitute for being able to lift heavy thing, fit heavy thing, sit in uncomfortable positions, which a lot of these jobs, the paper talks about them as production tasks.
These production tasks tend to entail things that middle-aged people are better at or younger people are better at than older people. So that's the direction of the prediction, more old people, more automation. I think we should think about this Andrey as we start to move into our priors. I want to think about this in a backward-looking way and a forward-looking way, right? So I think first, let's ask ourselves about what's our prior going in terms of looking backward. Do we think in different countries and different US regions that were differentially older, were they causally more likely to adopt these industrial robots? That's the looking backwards question.
Now we can ask the looking forwards question. We know the world is getting way older. Specifically in the developed world, there's still a big demographic boom going on in Africa. But in the Western and developed countries, we're really seeing a baby crunch or a baby bust, however you want to talk about it. So then crisis, if you will. A fertility crisis. Once you have a crisis that allows the, if only there were a way to use emergency tariff powers to solve that crisis. Although this paper, we'll talk about endogenous friendability later on, I'm sure. So like, yeah, the question looking forward is, do we think that this will be an important effect moving forward?
Do we think that a world with a baby crunch is going to automate faster and more deeply than a counterfactual world without a baby crisis fertility crunch? So, Andrey, you want to take on one or both of those questions? Yeah, so let's start with the first. I think there's a question of whether the theory is directionally right and how much it explains. I think those are maybe two different questions. I think it's hard to even imagine an econ-based model where labor scarcity doesn't result in some amount of investment or things that are substitutable to it, right? So if we have some factories in our country and now we have no one to work in those factories, you'd imagine that the owners of those factories face two choices.
One of those choices is to invest in technologies that allow the factory to keep running. They could raise their wages, but obviously in equilibrium, they will do other things. And then the other thing is to close down the factory and open it up somewhere else. So those are kind of the two choices that they have. So surely there is some adjustment on this margin. I think this paper makes the argument that this is like a huge causal factor. It's not just directionally right, it's a big causal factor here. I think that's with a pretty high probability. With some probability, you've got to put a number on it, that's the name of the podcast, dude.
Let's say a lot, let's say, does it explain, I don't know, 20% of the variation in industrial robot adoption across countries, put that at, I don't know, 40%. OK, 40%. So more likely than not, that the effect is less than explaining 20% of the variation. Yes, but if we're talking about like 1% of the variation, I'd be like a 99%. So I just-- Oh, OK, I see. So as we move up the strength of the causal thing. OK, so definitely some directional effect. Well, I'll tell you, I came at this kind of thinking at it already from like what data was going to be available, right?
And I know that you're going to find a strong correlation, because I know at the global level, the world's been getting older and we've been adopting robots, right? So that correlation's going to be there. And then at the regional level, I know in the United States, we kind of have this Rust Belt area that is both getting older and automating, right? So I feel like before I even looked at the evidence in this paper, I was coming in kind of understanding that the correlation is going to be there, that older places, older time periods are going to have more robots. So I was not surprised to see that correlation later on when we look at the evidence.
But as I knew that correlation was there, I knew in the back of my head that it isn't necessarily a causal relationship. And I'll tell you why, Andrey, well, maybe I was a little bit more skeptical about the hypothesis that the direction has to be in this one direction, right? Okay. So we'll talk about this a little bit later. But if the very narrow point is when the wages of factory workers are relatively high, you want to automate, setter as parabas, everything, that effect's got to be real, right? However, the question is about demographics and actually kind of like a lot of shit. There's a lot of moving parts in demographics other than one particular price difference, right?
You can imagine that young people bring energy and dynamism and they're going to, lead to more technology. I guess that's going to be my main kind of forward-looking point. Without going too far down that road, what I would say about, doubting any sort of observed relationship we see in the data is that given how universal ish this phenomenon is of the Western world getting more automation and getting older, I'm worried that there's like some sort of like overdetermination here and maybe like the right way of thinking about this is there's an emitted variable, which is like a modernity, which is causing oldness and robots.
So that's kind of my, so that's my backward-looking, I guess we'll go at each other. What do you think forward-looking, Andrey? Yeah, so forward-looking, I mean, once again, given that, I understand you can come up with, models where it's not so obvious, but I think on some margins, it must be true that older people mean like that we're going to need some robots to do the jobs that, they used to be doing. But I do think, and I think this is exactly where I was thinking as well, said, is that, as automation goes from something that's relatively nichely situated in a manufacturing sector to something that happens throughout the economy, it becomes just innovation, right?
It's just the table stakes for competition. And at that point, I think the demographics as an explanatory variable might stop losing, it might stop being as best powerful. So then there are, of course, two issues that are involved here. One is like, who is inventing the robots, or who's inventing the automation technology, right? And, certainly if we look at LMs and related technologies, here the key inventors seem to be happening in the US and in China, right? And, you can, there are a lot of differences between the US and China and the rest of the world, right? And then in terms of adoption, it may be the case that adoption is still determined more by demographics, right?
You do have more of an incentive to adopt certain technologies if you're in a world of labor scarcity. But once again, there is a little bit of a countervailing force in that older countries, or you might think that they are old people, old habits die hard, as they say. Science, what, the TFP advances one death at a time. Yeah, so you might be worried that if you have a bunch of old people in an organization, and, a country that's pretty aging, they might not adopt those technologies. And, but once again, I think when we're thinking about future-looking automation technologies, we have to, distinguish green, robotics, physical, automation, and non-robotics automations.
So I'd still imagine it to be the case that for manufacturing related robotics, or for, let's say medical robotics, the incentives are going to be so strong that you're going to see older societies adopt them more. But automation technology in general, I think, is going to be more driven by just overall R&D, capacity of a country. Right. So that's not something that's contemplated in Pascual and Inassimo-Glu's model, right? That the idea that you need young people to make the automation. In fact, as we'll see, the automation kind of comes out of final production. There's no sense in which you, need young people's ideas.
And, I don't know if you kind of kill the one thing that young people can do. Maybe it's not surprising that you want to, that they can just automate away from them. Okay. My forward-looking thoughts are pretty similar to yours. So again, we both agree on this main sort of, youth to innovation angle seems ignored here. Another angle that I think actually kind of pushes in the direction of this hypothesis, but is not contemplated in the model that we're about to read is, and this is kind of course I'm thinking about this because the only thing I ever think about, which is older people are wealthier than younger people.
If you have an older society, you have a wealthier society. And we think that the automation versus augmentation decision, you should automate more when interest rates are low and wages are high, right? Which should be the period of time when you have old wealthy people. And that's, this is the, I mean, I don't know, DeRone is just not interested in that mechanism. Yeah. Yeah. All right. So now that we've stated our... Wait, I didn't give him a percent. I didn't give him a percent. Okay. So my percent is, I would say 60% chance. So more likely than not, the net effect of aging is to increase the rate of automation.
But I, that's with wide uncertainty principles given that I feel like we need a second paper measuring the usefulness of the youth in innovating to balance against this other mechanism. Okay. Now I'm ready for the evidence. Well, being a podcast about, hey, we should mention that if you invent AGI, presumably you'll figure out how to automate everything else, right? So the epic AI guys say... Yeah. So, whatever conditions lead to AGI, we need to lead down the road to automation. Right. That's the correct rationalist take. The rationalist take is how dare you even contemplate partial automations that our viewers are listening to.
Yeah. It's all the transition, you know. Wait, so I got, did we get a percent, can I get a numerical prediction out of you for our forward looking question? So I wanted to frame it as the percentage chance that all else equal the aging society advances increases the rate of automation. 70%. 70%. All right. Okay. Well, let's see if we can move you at all a little bit with this data. So there's a theory and there's some data they bring to it. Maybe, Andrey, why don't you start telling us a little bit about the theoretical model? It's a pretty straight forward extension of asymo blue wrist repo raised between the end of the machine, right?
Yeah. So we have a bunch of industries. They're going to use a mix of inputs and they're going to have some manufacturing production required and then there's going to have some services required. And then they're also going to have a choice of what share of the tasks involved in the production process are automated or not. And then they're going to have, for each industry, I believe, a monopolist innovator, who's going to be, that's going to be the mechanism by which they're going to determine how much to invest in increasing the automation level for a given industry. Right. The way there's this, there's a basic model where we're just going to exogenously say here's the maximum level of automation and endogenous people will use the maximum level of automation.
And then they have an endogenizing the automation decision. Yeah. And then the production and the services are going to use a different, mix of old and young workers and there's going to be wages determining equilibrium. And they, show that under certain conditions, there exists an equilibrium or sometimes there are multiple equilibrium and they prove some comparative statics about it, such as the aforementioned statements about older societies resulting in more automation. Right. And I think, and there's a couple of other things that, automation is good for productivity. Is ambiguous for wages. But I think we can focus in on kind of just this main hypothesis who's automating, who's developing the automation, who's improving.
Yes. All right. I think that was a really kind of fair description. It's coming, this is all inspired by the task based model where we're going to explain different differential impacts on workers from technology as well, they were doing different tasks. And that's why they're going to be affected differently. So that's the task based framework that's been so influential. So in terms of data that they bring to this is where, so maybe let's talk about the theory section for a little bit. I think for what is in the theory, it is, straightforwardly makes an argument that, makes sense to us the relative, when the relative price of factory workers goes up, you want less factory workers.
It's the stuff that's kind of not in here. That is the interesting thing about the model. Yeah. So I think that's sometimes true with macroeconomic models. But by necessity, they can't have everything. Right. So in some ways, it's their rhetorical arguments pointing us towards one causal mechanism. That's very often how these models work. Right. Yeah. So this is not a bad way. Yeah, not in a bad way. I will say that like, if I'm teaching this, especially to people who are not super obsessed with the task-based model, I think you can get like 90% of the points of the model by just, drawing a supply and demand curve, but you don't need a lot.
There's a lot of apparatus here to make a pretty simple point. But it works. It does what it says. What's not in here that we both would both think are important. The first thing is you buy automation out of final production, which is in a world where we think young people have some special role in innovation that's probably not kosher. The second thing that's not missing here is the essentially, in the simplified model, the exogenous price of the automation technology. Basically, there's no there's no modeling here of the price of capital side. Right. So if you think that aging is going to have an effect on the price of capital, then that's also missing for you here.
Yes. There's also no trade in this model. Certainly something really. There's an extension of trade. There's an extension of trade. There's trade in the extension. It's simplified. It's fine. It does what they need it to do. When you want robots, you also import robots. Yes. And then on the shore that that's there. I guess like to the extent that you're shifting your industry composition. That's right. There's like there's short, there's trading goods, but there's no trade in industry and there's no migration, right? So you might imagine like in the long run, just like the industries go where the young people are, right?
Rather than vice versa. Yeah. This is a key, question, in the current policy debate, right? Why does China have so much manufacturing? Why did you, why did the US lose so much manufacturing? The price of labor seems really, really important here, right? And obviously this paper's not trying to address that, but it's getting at some underlying mechanisms. Yeah, but like you're totally right. So exactly. So what do I try to say here that's missing? It seems like the story about the US automating is mostly about wages being expensive and capital being cheap and like much secondarily about middle-aged people being relatively expensive compared to old people, right?
And that, and that's the mechanism from the other paper, not this paper. Yes. Okay, so let's talk about the data. We're going to look across countries for some of the analysis and we're going to look within the United States for some of the analysis within the United States. Whoa, within the United States, we got two measures. We got one measure by commuting zone, which is our robot integrator's presence that's like the presence from this other paper of people who specialize in helping you install robots. That's a binary indicator at the county level. I think it's not even panel. I think it's just a cross-section.
So you get a sense of how thin some of this data is. Okay, so they got a cross section. And then they've got a second cross section, which is actually a result from one of their previous papers, Asamogul and Restrepo. We've talked about them a lot. They're leaders in thinking about these questions in a paper that we have not yet to discuss, but perhaps someday we will. They look across the United States to see whether quote unquote exposure to robots is bad for wages and employment in routine factory jobs. Exposure to robots is a combination of the region's industrial specification in like the 80s combined with how much Europeans in those industries start to use robots over time.
Okay, so this is like kind of a real kind of like going all the way around the block to try to get how many robots are there in a US industry times location, right? But they do it. So that's their country level measures, their US level measures, I should say. The US level, yes. And then across countries, they've got a couple of different sources. Also of dubious reliability. One is the IFR, the International Federation of Robots. Andrey, do you remember that brief time in like the late 2010s or it felt like every paper used this data set? It was not without its controversies further.
Well, one of the controversies is so I, to the listener, we're economists. And so to us, the natural unit of measurement of the universe is the almighty dollar, right? You know, how many chairs do you have versus sofas? I don't need to fucking know how many. I just know the value of my chairs and the value of my sofas in a bizarre perversion of basic accounting practices. The way that the data on robots is accumulated is not by the value of the robot or the sale price of the robot or the tax rate of the robot or any economically interpretable measure of amount of robots.
What instead do they measure? Andrey, can you tell us a number of robots? The number of robots. I don't know what a robot is to be clear. Is it an arm? Is it a is it a Roomba? I don't know. Completely opaque, heterogeneous counting variable is going to be our main treatment or I guess mean outcome for our main outcome. I've been told to curse less on this podcast. Hey, viewers at home, let us be known. Let us know on, the comments. How much do you want to listen? The kids are listening, Seth, and they might be offended. They want to listen in their cars.
Not all children are from New Jersey. They're not used to our rough ways. Okay, so I'm going to we're going to limit our cursing to moments of really intense emotion, I think. Okay, so IFR data. So that gives you a measure of the quote unquote number of robots in a country times industry year. Then we have com trade, which is some international trade data, and that gives us some information on I believe the value of robots being imported and exported in different countries. I believe that those are the two, and then there's like some variations on those two data sets. Did I miss a big data set?
Yes, and actually there's a third one. There's also patenting. So we also can see across different regions of the world, the share of the patents that are in automation versus not automation technologies. So before we go into those results, so I've already talked to you about some problems with the international federations of robots data. Let me continue to list additional challenges with this data. Did you notice the footnote? We're darron in past while I'll say we're just going to we're going to emit mena countries because they don't fit our hypothesis. I thought you were going to say we're going to the footnote where they say that you're panins.
Oh, no, they don't but yeah, they just they have it in Korea, they have it in Korea. So and all they say it's robust, including or drop exactly. But no, the mena countries are the ones that blew my mind because it's just like they I think the footnote literally says we omit them because they do not fit our hypothesis, right? Let me see if I can find that footnote. What jumped out at you about some of these data selection decisions? Look, I don't I just have very little bar for cross-country regressions. I assume that they're that is just going to be like this because so many cross-country regression papers are like that.
I actually just like looking at the scatter plots. Like I just feel like rather than focusing on these regression results, which I feel are not super, they're credible to the extent that they point us in the right direction. We'll talk about the instrumenting for fertility because I think that that helps. That only helps. Okay, so let me I just want to read this footnote because I feel like I should always have an aside to read the best footnote in the paper. All right, footnote seven. Although international federation of robots data reports numbers for Japan and Russia, the danger for these countries underwent major reclassifications, etc, making the numbers reported for Japan not comparable over time.
Therefore, we exclude Russia and Japan. The IFR also reports data for Belarus, Bosnia, Herzegovina, North Korea, Puerto Rico, and Uzbekistan, which are omitted because we lack data on key covariates. Finally, we emit data from Iran, Kuwait, Oman, Saudi Arabia, and UAE, which are excluded both because they have few robots and also because their demographics are highly influenced by immigration. If you emit a third of your data, because it doesn't fit your hypothesis, I'd be surprised if you couldn't get a regression. Yeah, I mean, look, like I said, I look at the plots with the points and I look at each point, I'm like, oh, that's interesting.
Tell me what you see in the South Korea is. What's on the X axis? What's on the Y axis? The X axis aging, the Y axis number of robots, the increase in robots, let's be very precise, because it's really trying to measure the change in the in the aging and the change in the robots. Yes, well, no, not a day. Yeah, so we have South Korea has a ton of robots. And it's also experienced a lot of aging. So I think that's it's kind of the clear outlier one side. On the other side, a very old society without a lot of robots is Hong Kong.
And it's very easy for us to post hoc rationalize, maybe why Hong Kong doesn't have that many robots, right? It's a financial capital. Robots are a complement to space. Yeah, it hasn't been, all the manufacturing happens in Shenzhen. It's right across the border that used to happen in Hong Kong, right? But I think like just by doing enough of these case examples, we can learn something. I'm not sure we learned something from the regression itself. We also have, countries like Pakistan and Egypt, which are pretty young and have no robots. I think an interesting one. You can't you think of an admitted variable that in different H Pakistan and Egypt from South Korea and Taiwan?
Yeah, so I mean, look, sorry, who was you saying the middle? I interrupted your view of the discussion of the figure. No, the US is in the middle. The US is in the middle. Pretty much on the line. We're pretty close to the line. They're in the middle in terms of aging. They're in the middle in terms of robots. So yeah, that's kind of the basic outline of the story. They have another plot where they kind of focus on that OECD countries. And they're, interestingly, we see kind of below trend is Italy, which is a very old society, which doesn't have that many robots.
And then we have Sweden, which is a pretty young society that has a lot of robots. Yes. Yeah. I guess you would, I guess, if Pasquale was here, he'd say, well, that's the immigration story, right? Yeah. But, yeah. Look, the immigration and demographics are like, join at the hip. What are we talking about here? Right? So certainly it's an endogenous to aging, whether you have an immigration. Okay, so yeah. So all right. So to lock that in for in the minor or listeners, kind of the headline result is if you just put up a scatter block across countries, the older ones have the ones who are aging more or getting more robots, the planet's face, the relationship holds.
Is it the same way within countries within the United States? Yeah. So it seems to be the same way within the United States. And, within even industries within the United States, right? So kind of, I mean, one of the strengths of this paper is that they really do bring a lot of different data sets that are all kind of pointing to the same trend, even if each piece of the evidence is not, the best piece of evidence that we can imagine. Okay. Well, each piece of the evidence. All right. So I'm looking at their main regression table here, table four, which you told me not to look at Andrey, but you know how I love at this respecting of, reports to include.
All right. So we've got imports and exports. I think there is a similar table three, which does the same for just the amount of robots. Maybe we'll talk through table three. We have these OLS regressions of amount of robots on aging, pretty strong, significant, positive relationship of about one more robot per thousand workers per year, which is a tongue twister. Then what they do, I think this is probably like the most compelling thing in the entire paper, more compelling than the elaboration of the model, more compelling than all of this different data gathering. It's just instrumenting for olderness age of the population using historical fertility, right?
So like you point out, like population is endogenous. People do migrate, but a lot of the population of any given place is going to be its native born population. And the hypothesis that, economic factors today lead to migration today, obviously that's pretty strong, but it's a lot harder to argue that economic conditions today are creating fertility 30 years ago, right? So if we look at these instrumental variable results, let's do a little reader listener service and explain what instrumental variables are. All right. So they're like, they tell you like how many oboes you have, how many violins you have. I prefer the drums, you know.
What's an instrumental variable, Andrey? Yeah. Well, all right. So this is kind of the perfect case where we might be worried that we can't control for all things that are happening across countries. And we might be worried that countries who have different ages might have other factors that affect automation, their religion or their war causes, work causes, and stuff on automation, war, various economic conditions, like the financial crises and so on, just you can list so many things. And obviously, with very few observations, you can't control for all that. And so one idea was, is that what if you could find a variable that affects just one of the causal pathways and nothing else, then econometrics gives you the ability to use that to isolate a causal effect.
So actually, the best case scenario for an instrumental variable is actually in an experiment. Let's say that I want to understand what the effect of Seth buying a new mic is on his audio quality, but I can't force him to use using a new mic. Sorry, using that. I can't force him to use it, but I can buy it for him. And so the fact, let's say we have a bunch of Seth's and I've randomly bought some microphones for some of them. And then I also measure where they end up using the microphone. And then I can kind of use that experiment to say, what is the causal effect of using this microphone that I bought on Seth's audio quality?
That's kind of the best case scenario. The way the reason it's the best case scenario is if I've randomized, then the people who I've bought the microphones for and the people who I didn't buy the microphones for are on average very similar to each other. Now, with instrumental variables, you're look, without an experiment, you're looking for something that's similar in nature to that. Right. So what's the natural experiment here? So what's the natural experiment? So the authors are trying to say that, how many the fertility rate a long time ago, it obviously affects how old the society is today, but they're kind of two key assumptions.
So the fertility rate long ago isn't correlated with other persistent country level things that might affect robot adoption today. And also the fertility rate a long time ago doesn't affect any other causal pathways that have nothing to do with the relative number of workers. You might be thinking, dear listener, that there might be perhaps other things going on. Now, I for one thing, one of the key problems in our society is the boomer generation continuously electing their very old candidates. Four positions of power. Back to your medicine. This is not a political podcast, but, we'll let it put as an ageist, but it is ageist.
Old people have a lot of time to go vote, but they certainly vote differently than young people. And they might vote in specific ways that affect the economy. But for example, the instrument would also affect the political outcomes in a country, right? And the political outcomes are not in any way modeled in a theory model here. Right, so the answer to the question that was positive for the instrument is fertility rate. So you're imagining that more fertile people might have different politics than less fertile people. Well, that's certainly true. But the instrument here is historical fertility, right? Yeah. So that affects some.
It's not that it's just contemporaneous. Okay, so you're saying it's not just, but I guess that's what the question is, right? So when I read this paper, I took it to make the narrower claim that the causal impact of age on automation decisions, right? What you're pointing out is there could be a mechanism for that, which is not the contemplated mechanism of the paper, which is, but that's not necessarily saying the claims wrong. Yeah, so that's right. There are kind of two critiques. I think one critique is, if historical fertility is correlated with other things that still matter today, religion is certainly one that you can think about, that's one thing.
And then the other thing is just that there are many reasons that age composition can matter. For example, it might affect immigration and immigration might affect the adoption of automation, right? Or it might affect politics and that might affect automation. And I think, yeah, the authors are a little playing loose with the theory is very narrow. The theory is very narrow. It's really, yeah, it's really just about this, the factor supply, but the cross country regret fins with instrument are potentially, even if we kind of believe them on its face are including many different mechanisms. Right. Okay, but so let's actually interpret these results.
Do you have a table three and table four in front of you? I have table three in front of me. Yes. Okay. Well, table four, let's start with table four and let's work back to table three. So look at table four, this is imports and exports. On the top row, we've got the OLS estimates of the effect of aging on imports and exports. And we see maybe three more robots per thousand people as a function of agent, depending on the specification. And when you move to the instrumented version, again, it's a way since that says let's be precise. It's imports of robots to intermediate. Oh, yes, it's the share.
It's the percentage share of robots. That's a good point. Because obviously, out of intermediate imports. Right. So it is, it's pretty specific measure. And when we do the instrumenting, I would say that that is cutting the point estimates slightly, maybe not even consistently. It's not even consistently. Wait estimates don't change that much confidence intervals. Whiting a little bit. This is kind of what you expect to happen if you've just kind of got a shitty answer. Sorry. So I actually think that's not how I put it all. Okay. I would just put it as that the it's super predictive, right? Like it's pretty, demographics is one of those things that's very easy to forecast in a way.
Why is there F stat so bad? I don't know if it's so bad. Isn't cut off for being good? Yeah. And I see a 13. All right. Fair enough. Yeah, I don't know. I don't know. I just like now we're getting into the like the when you look at one of these tables, what do you think about right? Or I mean, I guess my thought process with you, they're kind of isolating a piece of, variation that it's demographics are highly correlated, hardly predictable over time. And so yes, you can predict demographics, but it's still not really telling you why demographics may matter or whether there's a minute variable bias.
Right. And we'll come back to the minute variable in a little bit. All right. Now, if we may move up to table three, are you seeing what I'm seeing in table three? What are you seeing set up? Compare the OLS estimates to the IV estimates, panel A to panel B in table three. They are identical. Right. That's got this is an error, right? I think they reported the IV twice because why is there an F stat for the OLS? That can't be. I'm sure I read this like three times trying to figure out what was going on here. No, yeah, it has it has to be a double paced issue, I guess.
Yeah. All right. So yeah, I didn't look at it this like too closely, but yes. I noted it and I was like, Oh, that's probably something. It's not it's not quite log of any ICS, but it's a little bit embarrassing that the restud editors let this through. Yeah. Yeah. All right, dirt cream. We love you. We love you senior editor, Dirk Kruger, you're a great economist. Sometimes the stuff actually gets introduced in the very last bit like when you're editing. Oh, okay. Yeah, I there's no way like the referees would have noticed this. I actually think there must have been some at some point some sort of caught like past the referees or the copy mistake.
You've heard it here first. Andrey was more sympathetic to an error in a paper than I was. I was not so not anticipating that in today's episode. Well, I guess like my presumption is that this table was correctly written in the submitted version. That's kind of what I, but if it definitely if it got through, if I got through the referees in this form, that would be pretty much now I have to look up the preprint. Yeah. Yeah. All right. All right. So we checked the tape Andrey and you were totally right, of course, past wall and to run are perfect. And it was an error introduced after the preprint.
What can you say? Now you're going to get me talking about how, pointless, the modern publication processes where you have a working paper five years before the damn thing is published. And by the time it's published, you don't even want to look at it ever again. By the time it's published, nobody even cares that no one at Courtney research has ever done at his experiment like this. So Andrey, the last result I wanted to highlight and I think maybe you wanted to talk about patents for a minute is that they've got this like we talked about earlier to run in past wall have this measure of kind of differential exposure to robots in United States regions and industries.
And so what if the hypothesis, one of the predictions of their model is when you bring in these industrial robots, said or is parabas, if we think they're substitutes for the middle aged workers and compliments to the old people, we should see the wages and employment of the middle aged people go down when the robots come in. And, maybe a boost for the old people. And they do seem to have that effect. So to the extent that you trust this exposure measure as being a good measure and being fully exogenous, it is consistent with this idea that it is the middle aged guys who are getting automated by the robot.
So that's another complimentary piece of evidence. Yeah, I mean, what do you, what do you think of that? I think that's compelling on the mechanism side, right? Remember, I was earlier arguing that I think an important mechanism is the wealth to population distribution, right? And that's a story where the relative age isn't so much important, so much as just the quantity of old people who have assets. And this, I mean, this differential effect on its face says, Hey, look, it looks like this is more, this technology is more substitutable for the middle aged people. So that I think that speaks to the mechanism. Okay, so then let's talk a little bit about the patents.
So I think an interesting question, is this, I think we're very late at the forward lookingness of this is, are we going to see more innovation and automation in older societies, right? And so what they're going to do is they're going to measure patents related to various automation technologies. Yes, so table five. So table five estimates the impact of aging on patents related to robotics. And they're going to find once again, statistically significant results. So that's true. And both of their all less and the instrumental variable strategy. So the argument is that countries that are older are patenting more in automation. I think what do you think about that?
I think this is kind of dependent on the quality of the quote or quote, automation measure. And specifically about whether this is a kind of automation that is relevant to middle school people, right? So I do not dive deep enough into this appendix to figure out if I like, I like their definition of automatable and automation patents. Another running feeling that I make clear in this show is that it's often not obvious from the technology level, whether the technology is an automation technology or a labor complimenting technology. So I'm just kind of suspicious of this at the, at the like, what they're trying to do phase is to get that right from the patenting.
But I'm sure they did a solid job. This, it's a nice complimentary piece of evidence, but I don't know how much I really leave it. Yeah, I mean, I think they do try to isolate it to robotics. So they are robotics related patents. So that might comfort us a little bit. They call it robotics patents, not automation patents. Sorry, they do, they do. It's a lot of robotics related patents. They do call it. They do. Oh, man. I'm maybe in the pre they called it automation patents. Okay. Yeah, I guess I'd want to know whether we think of these patents as being good patents or not, because that I think is where the predictive value comes in, because I think both of our concern is that the true innovations are happening due to young people or young, middle-aged people around talking young people talking about middle-aged people in their 30s and so on.
And if this is really just measuring like, oh, we have like a factory and we have to do some minor adjustments, we're going to patent everything. That's quite different than like, we've invented a new way to do robots. Right. Right. I feel like this should just be like, it's weird that it's share of patents that are automation, right? It seems like the argument should be it's still like a level effect, or like level as a function of the populate. Maybe this should be like a amount of automation patents scaled by the number of people, right? It's not because the decision isn't necessarily between automation and other technology, the decision should be between automation and like nothing, right?
Yeah, I agree with that as well. All right. Any other evidence you want to cover said? No, I think we kind of, I mean, listen, I think we talked as much as you can of how these, in some of the most specified specifications, we're down to 30 observations. Honestly, how much can you say about 30 observations, dude? You know, the older place. Well, we have the industry level results. The industry level, okay. But they split by industry. Yeah, they point in the same direction. You might be, that was, when I think about like, in the intro, I, thought like the different countries have different amounts of manufacturing, and that's going to be something you really want to think hard about when thinking about robots used in manufacturing.
And so they kind of throw that to the extent that industries are now allowing you to control for how much scope for robots there is, still within country industry pairs that this is there. Okay. So maybe let's like keep on talking a little bit about our limitations before we start getting into our past years. I think that like the kind of the macro idea here is that either age times industry or fertility times industry is somehow exogenous by region, right? That's kind of like the framework that makes this work. But as we've talked about, like industries move, people move, it seems like a theory of demographics and automation would at least need to think about those things as not being like measurement error, therefore, delete observation.
Second one, I don't know. I, did you understand why the scientists in the international model care about their domestic oversupplier undersupply of factory workers? Didn't they care about the international mix of factory workers when they're making? If I'm selling my thing abroad, it seems like it's the international prices that should matter, right? It depends on which types of science. This goes back to the small versus big innovation. So like, look, I'm like the fact, the fact there is scientists, engineer, and I make this adjustment to this invariable that makes it more autumn, they're a part of less oversight.
It's patentable, but, it's obvious too. And, the other kind of people in the other country will do it as well. We think about Chinese manufacturing sector, my understanding is that a lot of the innovations there in terms of productivity are not related. We're not patented. Right. Okay. Yeah. That's a huge, that's why we're not focused on patents, guys. Obviously patents don't cover that much of innovation. Okay. So you tell me if you are worried about this kind of emitted variable, right, which is my fertility decisions today are influenced by my anticipated regional economic expectations 30 years from now.
Is it really that implausible that a 30 year old making fertility decisions in 1980 is more, is what is differently excited if the region is making computer chips instead of buggy whips? It's not obvious that that can't be a factor. Sure. Yeah. Certainly, if we think that GDP per capita is persistent at the very least, and you would think that that's going to affect everything, including fertility, everything about communism affecting fertility. Well, you came up. You came up. That's a big one. I'm anticipating political changes. You brought up one very potential important emidifactor here, which is religion, which we think is going to have impacts on fertility decisions and like adoption of new technology.
Maybe as a bonus episode someday, we'll release more on economics of religion. But yeah, and the emitted variable I'm kind of thinking about is just like post modernity, right? It just seems to be this vibe internationally that, having kids isn't cool and robots are cool, right? So with only 30 data points and not a good way, not a good natural experiment of controlling for the vibe effect. I kind of, we're back to this is sort of like over-determined, right? This is the culture. Culture is different across countries. Yes. Right. Right. It's just that there's a big like thing called Western modernity, and inside of that package are things like an aging population and robotics, but also things like, texting a lot and social media and, having a liberal democracy.
Obviously that's going out of fashion, but. Well, but Seth, but what about them? What about the industry by country results? Right. So now we're going to have, so in those specifications, we distinguish by industry. The, so what are you're saying that that controls for race, for religion, because you think that the religious mix in one industry should be the same as the religious mix in the adjacent industry? Well, no, to the extent that you are then able to add in country fixed effects, that's getting at some of your concerns, no? Well, it could be, well, country trend in one specification that you can have the other country trend.
That helps. There's no reason to think that this relationship is linear. I'm not, at this point, I'm not appealing to there is a statistical test that would satisfy me. This is more just kind of skepticism about you have 30 observations and a lot of things are going together. Right. I guess I was responding to 30 observations, right? In the sense that like, there are 12,000 observations. Well, okay. So you tell me, which specification do you really love? You tell me, which one should I look at, the one that really convinces you? Well, it's not, I don't want to be in position at this, but this really convinces me.
I'm just saying that, it's a bit unfair to say that there are only 30 observations. When you think about table seven, we have an outcome as the replaceability index, and we have, this is country industry pairs. So when they're able to go to the country industry level, they multiply up from 58 countries up to 11,000 country industry years, which is obviously that's a lot more compelling. Yeah, right. That's kind of awesome. No, that's fair. Okay. Fair point. Fair point. Okay. Any other limitations you want to add or limitations that I said that you wanted to take down? No, I think, I think I've mentioned a bunch of them.
Okay. All right. So it's time to move into our posterior years. Andrey, are you ready to justify your posterior? Yes. All right. Our backward looking question asked, did differential aging of US regions and countries have an economically significant causal role on which regions adopted industrial robots? Where did you start? And where did you end up? Yeah, so I think I had this statement about, 40% that the effect explained 20% of the variation. Yeah. Now, I think now I'm probably like 50% or something like that. I moved up. And I think just generally, I do think that aging should affect automation adoption. And I'm probably that on the margin that has to, not that it has to be true, but that it's very likely to be true.
Yeah. And I still believe that. Yeah. Okay. So then I guess I started more confident and got more confident. So I was at about 75% that this is an economically important effect in contributing to automation that went, I went up from 75% to 85% on that question. Excuse me. They're really squeezing this stone. There's, in some of these specifications, they're really squeezing it. But every time they show it to us, every reasonable specification, every reasonable additional control, after throwing out the observations they don't like, which fair enough, it does seem to hold up. What, why am I not at 100%? I think 10% of my Bayesian prior is you should really, like you should not split off this as a separate effect that the correct model is that modernity is causing this, right?
And like there wouldn't be any statistical test that we necessarily convince me, but it's just, it's over determined maybe. And then I would say 5% is on, these are not super duper high quality measures, right? So we talked about like, what is a robot? You know, we talked about, they throw out Nigeria, because they don't trust their data. They throw out Russian, Japan, because they don't trust the data. I think like, yeah, so I think there's at least a 5% chance that like this is just garbage and garbage out and like, we shouldn't learn anything from this favor, but like non-zero, but also not high.
Okay. What about the future? What about the future? That's what we're here for. That's what we're here for, dude. That and justifying our posterior's. I would say that the way I ask the question is, do I think ceteris parabas compared to a world without a baby bust, does automation happen faster in our world, right? Because we got this baby bust coming. I see. I move from 60% to 65%, that automation happens faster in that counterfact. Automation moves faster in our world versus the counterfactual. Why is that? Well, the model in this paper is very narrowly about kind of like, the relative prices of old versus young labor, right?
Coming into this paper, I thought that that mechanism was going to be relatively unimportant, but we would still see this big correlation. And then like the paper would be about convincing me that that was the important mechanism, right? I come away feeling like, yeah, actually that probably is an important mechanism. We see that, especially, I think in the US data, that kind of convinces me is that, here's our AR approved measure of exogenous robot shock, and it does differentially hurt the middle aged people compared to the old people. That's strong evidence to me for this specific mechanism. I still think that like, at a meta level, probably the more important reason that aging is going to lead to automation is due to this old people are wealthier.
And there's less labor overall, rather than it being a, there's less physical labor and more service labor because that like old people provide service labor argument. It's like a part that's in here, but it's not like a long term compelling argument, right? And in fact, we think new waves of automation might differentially come for those kinds of jobs. So yeah, I move slightly in the direction that, yeah, you do want to automate harder when people are old, but we've talked about the youth energy, we've talked about other reasons that the effect might go in the other direction. So, more likely than not, but not overwhelmingly confident where you at.
Yeah, I mean, I think like I said, 70% that there's at least some increasing and adoption at the very least of automation technologies from having an older population. And I think, I'm would say updated to 75%. You're both looking a little bit in that direction. We both still have this concern that young people really are essential for the innovation process. You were right to point out that there's like kind of a differentiation here between kind of marginal or less exciting innovations where old people might be fine. Maybe Boopers can handle that. And kind of the more extreme innovations that we think are going to make the difference between AGI and not AGI.
And then, obviously, if you believe the Epic AI model, AGI is the only thing that matters. Yeah. It's one way to like, since now we're talking about the future posterior, right, is think about like autonomous vehicles, right? They are a form of robot. Although I'm not sure that that would be counted in this paper. Not mess. I'm not sure. I'm not sure like robotics patents. Maybe they do include some of that. Maybe they don't. Yeah, I don't know. But you know, when we think about the innovators of autonomous vehicle industry, they were all pretty young at the time they were doing their innovating.
And my understanding is that the reason they were doing it isn't that there was a shortage of Uber drivers, right? It was really that this is a pretty useful fundamental technology for a lot of things. Some of it might end up, surely will end up substituting for Uber drivers. But a lot of it will end up substituting for regular drivers and doing all sorts of rides that wouldn't have happened otherwise. And I think that's kind of the more exciting form of automation that we should be thinking about in the future. What's the form of automation that we should be thinking about in the future?
The exciting kind? We're like, we're creating new products based on the automation. New products. Yeah, like we've talked about, and this goes back to our a breast and a hand paper, reading about, I mean, I think both of us agree that automation going from, 50%, whatever, what's going from a certain percentage of jobs automated to, that plus X percent of jobs automated. At most, you're only getting GDP of the amount of jobs you automated, right? Until you get up to like 95, 99% of jobs automated and then sort of different dynamics take place. The place that you can have like unlimited gains without getting hurt by scale is innovation in these intellectual property products, right?
If they can invent, if the AI, the real gains come from the AI, inventing new medicines and inventing new games and inventing new art, that's the stuff that we aren't held down by the sclerotic bureaucracy that prevents us from building things in the physical world. And I know, I know that there are some who have told us it's a time to build. But in the short term, if you want TFP, it's better to keep it in cyberspace. And that's a great way to end this episode, Seth. Keep it in the cyberspace of listening to our podcast on all fine purveyors of podcasts. And make sure to like and comment and subscribe.
And we'll see you next time. Keep your pasta, your yours justified. It's good.