Economist Fatih Guvenen joins Allison Schrager to discuss his research into some of the most pressing issues facing the American economy—from wage stagnation to income risk—and how the use of big data can demystify the nation’s economic challenges.
Audio Transcript
Allison Schrager: Welcome to Risk Talking, a podcast about economics. I'm your host, Alison Schrager. And I'm delighted to be joined this week by Fatih Guvenen. He is the Curtis Carlson professor of economics at the University of Minnesota, the director of the Minnesota Economics Big Data Institute, and a research associate at the National Bureau of Economic Research.
When I think of what Big Data has done for economics, I immediately think of his research, really because his research more than anyone else I think of really has completely turned on its head so many things we thought were wrong with the economy, or maybe are wrong, but just in a different way. And this is not only really critical for the way we think about the economic challenges we face, but really should be foremost of our minds when we're shaping policy and thinking about what the economy's going to look like in the future. So I'm so excited that he's here this week to share his research, because it's really going to change how you think about a lot of very pressing economic issues, particularly this week when we talk about wage stagnation and income risk. So thank you so much for joining us.
Fatih Guvenen: Thanks, Allison. Thanks for having me and for the very kind introduction.
Allison Schrager: So let's just jump in with wage stagnation, because this has been the big economic issue, which is blamed pretty much everything from stagnant living standards, allegedly to the rise of Donald Trump, everything. I think, when we talk about big data, you got your hands with a big team on social security records, which really has shed a lot of light on the true nature of wage stagnation. Because I think a lot of people don't quite understand when we look at here are statistics about stagnating wages. We are just getting a snapshot of the whole population. That actually means, an individual is often less important because what we really want to focus on is whether an individual is getting raises each year and look at his lifetime income. And if you just take the whole population, you might be capturing things like an aging population, which doesn't tell you as much.
So when you got social security records, you could really look at what happened to people's incomes at an individual level over their lifetime and how that's been changing. So I guess to sum up, what did your research tell you about wage stagnation in America?
Fatih Guvenen: That's a very good way to describe actually what we are doing. So what we did in this project with some co-authors, we examined about 60 years of social security records on earning histories. So the data goes back to late 1950s, all the way to mid 2010s. And we track individuals over their lifetime, as you described and we are able to follow about 27 cohorts, and by cohorts I mean, year of birth cohort, people who enter the labor market pretty much in the same year and we follow as their labor market careers evolve.
So what we have done for each of these cohorts, for each individual, we calculated the lifetime income. So not one year, two year, three years. It is actually we add up the entire income they earn over their lifetime, and then we look at the distribution of these values. So you get really not a snapshot in time, but you get a summary of the entire lifetime earnings of a generation, a cohort that enters the economy in a given year. And then what we do, we follow each cohort that enters the US economy, and we look at how that has evolved. Have subsequent cohorts gotten richer and richer? Have they made higher incomes over time?
And what you see is a pretty striking picture. Early on from 1957 to late 1960s, for about the first 10, 12 cohorts, there are very solid gains. And now I'm going to focus first on males because men and women look very different when it comes to these facts. For men, you observe very solid gains that was widespread. One interesting fact is, the top 1% and the bottom 1% income growth in lifetime income was equal. The entire actual economy grew at the same rate in lifetime incomes for these cohorts.
But something happened. There's a turning point around 1968. And if you thought that was an important year, now we have one more reason to think it's an important year because whatever happened that turns out to be the peak year for lifetime income. The cohort that entered that year was a turning point. And subsequently, every cohort that came later earned a lower and lower lifetime income. To give you some numbers, the 1983 cohort, that's the cohort basically where we have a full lifetime, 30 years of data. So we observed them from age 25 to 55. The 1983 cohort entered only 17 years later after the 1968 cohort. But their lifetime income, depending on how you exactly do the calculation was between 10 to 20% lower than their, you can say older brother or young parents generation.
The reason this is striking is, during this time, the US economy kept growing. So GDP was growing, incomes overall were growing, but when you looked at the medium man, the average man in that cohort, their incomes were going down and that is really striking.
Allison Schrager: Is this true for every education level?
Fatih Guvenen: So the data we have, doesn't allow us to look at education. I will say more from other data sets maybe later, but what we look at is we have looked at where does this stop? So the average worker has seen this decline. How about the worker at the 70th percentile, two thirds of the way up. It goes up all the way to three quarters of the way. So the volume is 75% of the US male distribution. So almost no gains in lifetime income during this period.
Allison Schrager: So when you say gains, I guess there's different ways you could look at this. You could look at their starting income and how much it grew over time, or just the present discounted value of their entire lifetime income. When you say income, you mean that present discounted value or the changes they had over time?
Fatih Guvenen: Yes. So I mean the present discounted value. So that's the lifetime measure we have, but you can unpack this. You can ask, okay, the later cohorts had lower lifetime income, but at the risk of oversimplifying, that can happen through two different ways. One, it can be that the later cohorts entered at the same, entered meaning, when they first started working their entry wage was similar to the earlier cohorts, but their wages did not grow as fast as they got older. That's one possibility, okay? And then they will have lower income. The second one is that they actually entered with lower wages, but maybe their wage growth over their life cycle was similar.
It turns out that it is almost entirely the second one that the entry wage, yes, tracks pretty well what happened to lifetime income. And there is a very important message there because a lot of the topics we talk about with wage stagnation and I don't mean to say that they're unimportant, but things like trade things like in the labor market, how can we change certain policies to help with trade? They apply to workers after they enter the labor market, but this finding suggest that these cohorts were different already. By the time they entered the labor market, for whatever reason, they had lower income early on, but the evolution of their income over the life cycle was there were some differences, but they were second order compared to the lower entry wages. So this changes your perspective about what factors we should be focusing on to understand wage stagnation.
Allison Schrager: And also, as I said, changes how you think about it? Because we talk about wage stagnation that suggests people aren't getting raises. We often hear people say, "Americans, haven't had a raise in so many years." But it sounds like while people are certain, maybe worse off compared to earlier generations, they are actually seeing raises.
Fatih Guvenen: So it depends on what part of the distribution you are looking at. So, like I said, if you enter with lower wages, you will see wage gains over your lifetime, but compared to somebody who entered, let me give you one statistic that to me, when I teach my labor classes, to my students, I say, this is one of the most important statistics that I know in economics. If you look at a male worker aged 25, if you look at this worker, the medium worker of this kind in 1970, the entry wage was $35,000. If you work at the same worker in 2010, the entry wage was $25,000 adjusted for inflation.
This sounds pretty crazy because this is almost 50 years apart. 50 years later, a worker with the same demographic characteristics is making 35% less when they enter the labor market in an economy where the GDP is three times higher. So you do get raises over your lifetime. But the level you start at is much lower than before. And because people who are born now or who are entering the labor market now, they don't know how the world was 50 years ago. They may not be realizing this, but actually when you do this long time comparison, the picture becomes very, very clear.
Allison Schrager: It's even more striking that educational attainment has increased in that time, too.
Fatih Guvenen: That is true. One aspect of this discussion is about what can be done about this. We can talk about this now we can talk about this later. What do we make of this? What prescriptions can we think about to deal with this? And the immediate reaction of many economists is education, right? We jump and say, "We need more education." But if you look at statistics on education, some of them are pretty sobering because one statistic that I think is pretty eyeopening is, if you ask the following question, we hear a lot about the college premium and the college premium is the difference between the average wage of a college graduate and the average wage of a high school graduate. This is one definition that we can think about. So how much premium is there in getting a college degree?
This number is pretty big. It's almost double, okay? A college graduate makes almost double, a high school graduate. And this is very highly quoted. This number was much smaller in the 1970s. It has gone up a lot. And the implication that we draw is that, well, we should have more people going to college. And as you mentioned, they are going to college. College attainment is much higher than it was in the 1960s, '70s.
The problem is the following. If I ask another question, which I think somebody who wants to go to college will want to know, if I say, "What fraction of college graduates end up not making the same amount of money of a high school graduate in the top 25% of the high school distribution, okay?" If I go to college for four years, I pay all the tuition. I forgo four years of income, I better be at the top of the high school distribution, right? And it turns out, about 31% of college graduates are not at the top 25% of the high school distribution. And that number to me is pretty big, but that is not really the most striking fact. The striking fact is, this number was 31% in 1970, and it was 31% in 2015.
So even though the average college premium has gone up, because of the distributions overlap, meaning that outcomes are so dispersed that there are a lot of college graduates actually, who make a lot less than a lot of high school graduates. So when we go back to college attainment, it's the opposite. I'm not in any way suggesting college education is not important, but we also need to think more carefully about the content of the college education, about whether we are teaching the right skills that help workers do the tasks that are not directly competing with machines, robots, especially with software, especially with jobs that can be easily offshoot or can betray their way and so on. So I think the education discussion needs to be a lot more nuanced than what it is in some circles.
Allison Schrager: Have you also noticed that, as I said, you did mention that most of the fall and lifetime income for men has been about the starting wage, but have you also noticed ... When I studied income patterns, what we saw is we saw higher earners started low, but had very, very steep trajectories. Are you still seeing that's the case?
Fatih Guvenen: You see that a lot more for women. Yes. So the adventure, so to speak, of women has been very different in the sense that the start, because they were starting from a much lower baseline, right, the average wages in the 1960s, '70s were so low that it could not go anywhere but up, so to speak. So it has gone up quite significantly by about 60%. If you look at again, the same comparison, roughly speaking the late 1950s versus 2010s.
However, it has stagnated since the 1990s, but what has changed is what you just said, which is that they have entered, especially in services, the sectors, the industries that they have entered in were industries where wages start low, but then you see very fast wage growth. So when you plot the figures, you see that you actually see very fast wage growth for women, especially in the 1990s. It's not like throughout the period in the last, I would say 20 years since the turn of the century, women look more and more like men. Like across the board, if you look at labor market statistics, there's a level gap between men and women in terms of their income, but the overall pattern of how their career evolution looks like is very similar.
Allison Schrager: What do you think did happen in 1968? Do you have any ideas of what did change?
Fatih Guvenen: We don't have a smoking gun. So what we did, we looked at state level data to get a better idea. And the way the idea goes is, you can calculate the median lifetime income for each state in every year. And when you plot this over time, you see that they had very different trajectories, actually in some states, the median lifetime income declined a lot more than others. Whereas in some others, they were flat. They did not decline in some others. They went up. So the overall decline still masks quite a bit of heterogeneity across states and other outcomes also were different.
So we try to link these to some usual suspects. We look at the decline. For example, in the unionization rate, we looked at the rise in trade share, we looked at the share of the gender ratio in the labor market, in the demographics and a host of other variables. And to our disappointment, nothing really came back as a smoking gun. Some of these variables do change in the '70s, but there is some evidence that the demographics played some role. Some evidence that manufacturing share played some role in stagnation. But so far I view our paper, our research as more a call to action that there's something here that we need to better understand. And it's still ongoing research that needs to pin this down. Exactly what happened.
Allison Schrager: So what states should we live in if we want to have income growth?
Fatih Guvenen: Some of them are places you can guess, California, New York, New Jersey, those states actually have done quite well.
Allison Schrager: So let's run into another paper of yours. So something that really shook me up when I saw it, mainly because it disproved my dissertation. But as you say, things change. This is why we have Big Data, right? Better data. So things we thought were true, weren't true. And that makes us better economists and makes the world better. So I'm letting it go. So you have five myths, big myths, I think really have been taken as conventional wisdom that you completely blow apart. So I'd like to go through all of them if you don't mind.
So first one, myth one, the rise in income inequality is partly or largely driven by rising within firm inequality. Before the CEO made certain amount and the secretary made certain amount and now he makes so much more than she does. You found that that is actually not what's driving inequality. So what is?
Fatih Guvenen: Yes. So what we did in that project, we actually put together a data set that encompasses the entire United States, all the workers and all the firms in the United States. And we were able to do this. It was the first time ever that this dataset was put together. So it allowed us some unprecedented look at these questions because the statistics that you have always heard about like the CEO to medium worker wage that was cobbled together from different datasets, because we don't really have a dataset out there that tracks all the CEOs. There's about 6 million firms in the United States. If you look at firms with more than 20 employees, there is about a million firms and we did not have a data set that tracks all the CEOs.
So what we had was some data sets that track the top 500 CEOs or the top 200, 300 CEOs. And those statistics that we hear about, usually look at those very, very, very top CEOs and look at the medium pay in the economy and take the average. So instead, what we did, for the first time, we actually had a data set where you could track all the CEOs because we have all the firms and we have the top earner in these firms. And what we found was the following. If you put aside for a moment, the largest 1000 firms in the US economy, and these are firms roughly with more than 10,000 employees, if you put them aside for a moment, the remaining 99.9% of the firms within firm inequality has almost not changed to a first order approximation.
What did really change was the gap in the average pay across firms. In other words, it wasn't that the CEO, say, of Google was making a lot more than the medium worker at Google, but it was that the average pay, say, at a company like Google, has diverged from an average, say, grocery store in your neighborhood. And that is really driving the big rise in the wage inequality. And if you want, I can tell you about the top 1000 firms.
And then if you turn to the top 1000 firms, yes, you do see a rise in CEO pay, however, the magnitudes that you see are quite smaller. So there was a doubling of the CEO to medium worker wage gap between 1980 to today. And almost all of that happened in the early 1980s. So this 200 fold and so on gap, is really not the CEO in that firm and the medium worker in the same firm. It was comparing the CEO of one firm to the medium pay in the economy. But firms where the CEO is paid a lot, actually, the average worker is also often paid quite a bit. So that gap is smaller. When you put the two together, the very large firms and the 99% of the firms and ask, "If I were to remove all the CEOs from the economy and look at the rise in wage inequality in the US, okay, would I have seen a much lower rise in inequality because now there are no CEOs in the economy?"
It's pretty much indistinguishable from what you would have. And the reason for this is, remember, for the vast majority of them, the wage didn't rise much faster and for the 1000 or so, they are a very, very small part of the economy. There is 6 million firms and only for the top 1000 firms, the CEO pay has gone up. So those statements basically were based, I would say on more anecdotal data than really data that covered the entire us economy.
Allison Schrager: So when you say a thousand firms, do you mean just in terms of pay or do you mean they're just all big or they're fast growing?
Fatih Guvenen: No, this is size. This is employment based. About 10,000 employees or more. We call the mega firms. It's a name that I just made up and it's stuck, now I think now people are calling these mega firms as well. So I'm happy about that.
Allison Schrager: So do you think these firms are more productive? Before, working at a big firm tended to play rewards, but now it seems like if you want to make a lot of money, you really should be at a very big firm. Does that mean these big firms are more productive? Does that mean they just are getting rents? What do you think is going on?
Fatih Guvenen: Well, there's a reshuffling that's going on in the sense that there are two distinct trends that has been going on since the 1980s. One is what we call segregation. And by that I mean segregation by skill level. So the underlying change in the economy is really coming from the workers becoming more and more different in terms of their skills. So that is driving the wage inequality. Because they have different skill levels, the wages that they get is also more and more different, but by segregation, what we mean is, similar workers by skill are now more likely to work together in the same firms.
So you used to have basically firms employing high skilled workers, middle skill and low skilled workers. Whereas now you have some firms that are really focusing on employing a lot of high skilled workers, others that hire more middle skilled workers, other that focus more on low skilled workers.
And this is part of the story. So when you look at a cross firm wage inequality, that's what you are seeing because one firm is hiring high skill workers. The wages there are much higher. Whereas the other one the wages are lower because the average skill level is lower. And the second one is what is called sorting. So this high skilled workers are not working randomly at some firm. They actually happen to also work at firms that have higher ... We call it quality, but it's a very vague term. These are firms actually that pay workers more than their market wages. So that can be due to rents that these firms can be earning rents and sharing with their employees, or it could be that they have better technology and these workers are a better compliment to their technology, the high skilled workers, and they're able to be more productive and they're sharing some of that with their workers. So these two factors combined together to explain about more than two thirds of the rising firm inequality, the segregation and the sorting.
Allison Schrager: Do we see this in certain industries? Are all these a thousand firms in high skill services or do they have manufacturing too?
Fatih Guvenen: So these high paying firms or these large firms?
Allison Schrager: Are they the same?
Fatih Guvenen: No, they're not. So the large firms actually, that's another big change. So manufacturing used to have about 400 out of the 1000 largest firms in the US in 1980. And today the number is down to less than 200. And the opposite happened in retail, which had about 150 of the largest 1000 and they overtook manufacturing. And right now about 350 of the largest are in retail. This is the biggest change that happened and the others went up and down, but not to the same extent. In terms of the most productive at the top. Some of these are, I would say, a bit usual suspects firms that you can expect high tech companies are there. Pharmaceutical companies are there, chemicals, they are in there. And these are companies with very high margins, but at the same time, they tend to be very innovative companies like the Apples and the Googles and the Teslas and all these companies that you can think of will be in that high productivity category.
Allison Schrager: So getting into the next myth and as I said, this is something I said that is again, become a very pervasive thing. We hear all the time that the economy is now so precarious and workers have no wage stability and income risk is out of control. There's been books written about this. And to be honest, I said, this actually featured heavily in my dissertation. I very, like a whole generation of economists, very dutifully, decomposed lifetime wages using a PSID, which is a huge survey from the University of Michigan, or actually it wasn't that big. That might have been the problem. And we saw, wow, wages were getting so much more variable. Workers just do not have the income security that we used to have. But it turns out, once you got your hands on the social security data, you saw something very different. Do you want to explain?
Fatih Guvenen: Sure. I myself have taught the first 10 years of my career. Every year I went to my class and beautifully repeated what I learned, which was that income volatility has gone almost through the roof. It has doubled or more than doubled since the 1970s. And so some of the facts that we discovered, or we found in Big Data, we could smell it in advance. We had reasons to be suspicious. And we looked in that direction, some fall in your lap. And this one was more like that. We did not really expect it. We had a little suspicion. So when we looked at the source security data some of the basic statistics, some of the work we discussed so far requires some more advanced methods and you need to be careful about how you apply the methodology correctly.
And the fundamental fact is so simple. You just look at the right place and then once you see it, you cannot unsee it. So it turned out that when you plot the basic data, the volatility of income, we did not see the pattern that others have documented in the PSID. And there's nothing wrong with the work that they did, the methodology they applied, it was solid. It was that there were some issues with the survey data. And you alluded to survey data is very useful for many purposes. And I still use it. Even today I use it in a lot of my research. But when you look at long term trends, sometimes panel datasets especially, their composition will change and they may look a bit less and less like the US economy. So then you need to be pretty cautious about how to interpret a trend that you see in a long run survey versus the data.
And so that was pretty puzzling because there is about 30 papers that have documented that. And then we kept digging into it. We said, "Is this true for men and women separately? Is this true for different age groups? Is this true for workers working for big firms, small firms, by industry?" No matter how we cut the data, it always came the same. We looked by percentile by percentile of income distribution. Different people in different parts of the distribution, have they all experienced a decline in volatility? And it turned out that it was really pervasive and the volatility ... The amount by which it declined? We can debate that. Okay, is it 10%, 20%? But because the baseline belief was that it went up dramatically. And like you mentioned, I haven't counted it. But I think if we did, there would be hundreds of papers published and a lot of research. And I have had papers not published, but I worked on them where we took this as a fact. And we tried to explain other facts about the world.
Now we need to rethink some of those and say, well, if we cannot explain it with more uncertainty, then how do we explain the same phenomenon now? So that's still work in progress. We are still learning and we are happy, when you do something with a data set like this. You always want others to be able to replicate your work. And recently, there's another team that used another dataset, the LEHD by the Census Bureau. And they basically looked at the last, they are able to look at the last 20 years of data and thankfully, their data and ours aligned very well for the 15 years that the two data sets overlap. So we feel vindicated at least for that period, the two looks very much in line.
Allison Schrager: Yeah, I recall, I think it was Gottschalk and Moffitt were the wage people. And I believe one of them, or both of them did have a paper confirming your results, which for me was the end of it, like, this is a fact. But it's important for a lot of reasons. One I think like a financial economist and in some ways this has got to be good for Americans, right? In finance, an asset that gives you more stable returns is worth more than an asset that's more volatile. So in some ways, could this be good for workers, they have more predictability?
Fatih Guvenen: That's an interesting question. And for another reason as well, because Moffitt and Gottschalk, when they wrote their famous, the seminal paper in 1994, and they documented that the volatility was going up, I think without exception, every economist I know, they took that as very bad news. More volatile means more uncertainty. That's more risk. We don't like risk. So that must be bad. But now that we document, it's actually not the case, the volatile is declining. I don't feel ready to jump to the conclusion that this is good. There are reasons why it may be good and you are actually giving one reason, but at the same time, a less volatile economy could also be an indication of again, stagnation or more rigidity.
Let me give you one reason. A lot of the wage changes if you take the variance of wages, about 80% of wage variation in the data comes from job to job changes. It happens when a worker changes a job. So if there's a decline in job changes, you are going to actually also get a decline in the wage volatility. We did actually look at that. There's a small decline in job to job changes. And after you account for that, you still see this decline. So I am not immediately very worried that that might be driving it. But I think that one is also still, I would say the jury is out, whether it is really the way you're describing that there's less risk. That should be good. I hope that is really true. But is there something more sinister lying in the background that maybe it says more slowing the economy? I don't think I'm ready to rule out either one just yet.
Allison Schrager: Yeah. As I said, thinking like a finance person is this two sides to the same coin. Like what we spoke earlier about wage stagnation and also less risk than you would expect more wage stagnation, if you have less risk in the economy.
Fatih Guvenen: True. Exactly. But just to say the following, let me add one more thing. We did look at the following. The overall volatility in wages can decline because the size of positive income shocks the positive weight changes become smaller, or because the size of negative income changes become smaller. If the size of negative income changes become smaller on average, that's good news. It means that people don't have this big declines in income. If the size of big positive changes decline, that's bad news. It means people don't get big raise anymore.
And when we looked at that, we said, which one is it? Which one is driving the decline? It was equal. Pretty much both of them were driven by the same amount. And there's some indication even that it might be the negative shocks are becoming less severe for certain demographics. So the evidence I have, really, doesn't give me any reason to be very pessimistic, but maybe I don't know the word economy in generally, we are in a pessimistic place. So I'm trying to be cautious when I speak.
Allison Schrager: Okay. So myth three is, income risk over the business cycle has been mostly cyclical variance of shocks. So first of all, what does that mean?
Fatih Guvenen: So, yes, that's a bit like a technical terminology. So we have been trying. This goes back to 1990s, since economies have started thinking about heterogeneity, differences among people, we started taking this more seriously in the context of business cycles. We started asking, "As a profession, how does wage risk, income risk change from an expansion to a recession?" Now there's one obvious thing in a recession, right? On average, incomes go down. On average, wages go down. But beyond that, do we also perceive more uncertainty of wages looking forward?
That is not so obvious actually it might just be that everybody's wages fall by 5%, but looking at tomorrow, we don't have more uncertainty. I think if you do some introspection, if you look around, we also feel like there's a bit more uncertainty. There's a lot of people who have not been laid off, but their colleagues, their coworkers have been laid off. So you're worried that you might be laid off. You're applying for a job, but nobody's hiring. So there's a bit more uncertainty about if you will find a job in the next, say, three months, six months and so on.
So when economists have first started looking at this, they analyzed data against survey data. But for this particular question, it doesn't actually matter. It turns out that the methodology was a bit more important and they concluded that the way they concluded one, that uncertainty really goes up in recessions, but the way they modeled uncertainty, this is the curse of normality or the Gaussian assumption that we hear in many different contexts, right? Gaussian modeling or Gaussian distribution, the bell shaped curve is very convenient in statistics. So we use it or maybe overuse it.
So the way this was modeled was through a bell shaped curve, which as we know is symmetric. So the implication was that when the variance of a bell curve goes up, both tails open up by the same degree. What that means is, uncertainty goes up, but the likelihood of a big negative shock and the likelihood of a big positive shock, both of them go up in a recession. So that was the accepted result. That was the conventional wisdom that in recessions, the variance of a Gaussian goes up means both tails expand. That makes you scratch your head a little bit. The left tail, we understand you. It's more likely we'll get big negative shocks. We see how that can happen, but the right tail, how does that really happen? Who are these people who are getting bigger, more likely positive shocks?
So when we looked at this, we said, that's another thing in research. I feel very strongly about. We should not be making any parametric assumption when we absolutely don't need to, because we should let the data speak and not put words into the mouth of the data.
When you don't have very good data, that's unavoidable. Because if you have a small sample, a lot of noise to get to learn anything from it, you have to impose some restriction like the Gaussian distribution. So with better data, we basically said, let's make no assumption on the data. And let's look at the actual distribution of wage changes and see how that changes in a recession. And it turns out that the picture was somewhat different with important implications. So the left tail really expands, meaning that yes, really there are bigger negative shocks in recessions. Actually, the duration of them is also longer. So when there's a painful thing happening, it's worse in recessions. And I think everybody can agree with that, but the right tail, these mythical people who are getting these big, positive shocks and celebrating in the streets, they don't exist.
So instead, if you are going to get a promotion and maybe a 30% raise, you maybe get the promotion, but you only get a 20% raise or a 10% raise. So the right tail was actually compressing rather than expanding, which is skewness. So what was changing was the skewness and you cannot capture that with a Gaussian because Gaussian is restricted to zero skewness. So this is a bit more technical I know, but this is a different kind of uncertainty. And actually, if you put it in your models, which we are doing in an ongoing paper, you find that the models work better with this kind of risk, because it in a way makes the recession risk a bit bigger than you would have under a countercyclical variance risk that we have been assuming in previous research.
Allison Schrager: So has it gotten worse? So what you're saying is, one thing that you've learned from looking at this data is when it comes to a recession, people face very negative shocks. Have they gotten more negative over time?
Fatih Guvenen: There is some indication of that. Yes. So we have looked at data, again, for this one, we have looked at data going back to 1950, which there's a particular social security dataset that allows us to do that. And when you look at the biggest shifts to negative skewness, the last three recessions before COVID, so the early '90s recession, the early 2000 recession and the Great Recession were by far the largest negative skewness that you see, which is interesting because the early '90s and early 2000 recessions were very mild actually by many measures. But by this skewness measure, meaning for heterogeneity, for risk, it wasn't.
So there's some indication ... Actually, we also looked at for stock returns, we looked at COVID and you also see that's so obvious. I think that can already guess it. But when you plot all of them together, this is such a big spike that it wipes out the other spikes that you see in the data. So now I can add that too. The last 40 session, I would say looks more severe in terms of negative skewness.
Allison Schrager: So when we think about job loss and recession, there's always a human cost and it's awful, but on the upside, we always like to think, "Well, recessions are good for cleaning out unproductive companies that their day was done anyway." So I'm curious how persistent these income shocks are. You have a negative income shock during a recession, you lose your job. Do you end up then eventually finding yourself in maybe a more productive firm and then end up recovering from that income shock?
Fatih Guvenen: That's a very good point. So we looked at a couple of different things and each one actually surprised us more and more. So the first one is, we also looked at long term skewness. Meaning that not the change one year change in your income, but the five year change in income. So you are here to the recession today. And this applies to, even if you don't lose your job, but we looked at where your wage is five years down the road. And the facts I described apply equally well, actually to that one. So the effects seem to be very persistent.
The second one is, we have done this also for Germany, France, Sweden, and now we have an international project with 13 countries. This is a joint project with more than 50 economists and these different teams, they have also done it for all 13 countries. And this is pretty striking. Sometimes you see the data and they all look the same, which is beautiful because in economics, we never have laws, right? Everything's all over the place usually. But when it comes to this skewness behavior, it seems to be very much synchronized with the business cycle. And the point I want to make is the following, in those countries, we have good data on employment. So you can ask the following question. Suppose I only look at workers who are continuously employed at the same firm in the same establishment and they're employed full time. So now these workers could not have been unemployed.
Despite this, you still see the skewness fluctuating the way I described. The magnitude is a bit smaller, but the basic facts that I described are quite similar to the entire population of workers that I described. So there's something going on that applies to workers, even who don't lose their job.
And when we looked at this, we saw in France, we have very good data on this. We see for this full-time workers that about two thirds of the skews change happens because they adjust their wage. So somehow, in recessions, the firm somehow comes to an agreement with the worker that they take some wage cut. And one third happens through hours change. So the firm says, "You can continue to work, but we'll have to cut your hours by this much." And the worker, rather than losing the job, they agree to that arrangement.
So it goes a bit deeper than our initial instinct, which is also what you mentioned, that this may be all just job losses and it'll revert. It seems to be very persistent and comes not only from job losses, but is more pervasive across the workers employed than unemployed.
Allison Schrager: Do you think it might be different going forward now that inflation is back? Because before if you had say five, 6% inflation and you were in a recession, you could just not increase people's wages and you had this very natural wage cut or a real wage cut, that you didn't have to change nominal wages. But with low inflation, you really did have to cut nominal wages too. Do you think maybe that could have been driving some of this?
Fatih Guvenen: That's true. I guess the best evidence we have is the 1970s and the long time serious data I have, we have that. And in 1973 and the 1980 double-dip recession. So 1973, the inflation was rising, but it wasn't as high as today, and the 1980 early recession, the inflation was very high, but then came down. So they're not perfect examples, but they were surrounded by inflation and the skewness changes similar to what we see today. The magnitudes are a bit smaller, but in terms of the way they look, they look similar. So I'm not sure, I don't like to predict the future, but inflation may be affecting the averages, but I'm not sure how much it'll affect the shape of the distribution. But that's a good question.
Allison Schrager: On a more optimistic note, do you see the skewness to the other side during booms?
Fatih Guvenen: You do. So coming out of the recession, it's like roaring back, actually. That's why I love looking at these plots. I have two sons and sometimes they have interest in what I do. So they ask me what I do. I say, "I solve puzzles." And they say, "What puzzles?" And I show them the figures. And I try to tell the stories. They don't get too much out of it, but actually, some economics, there are stories in these figures. And when you look at them, you do see the skewness going from a very big negative number, in one year, two years out of a recession, they just zoom to a very big, positive number. And now we have another line of work with other co-authors where we look at firms and you see the exact same picture there. The skewness is negative in recessions and then it comes roaring back in expansions.
Allison Schrager: So taking this back to how we value wages, I said earlier, variance has gone down, so in some ways wages have become more valuable, but also thinking about it in terms of finance, it sounds like you're saying the wage beta has also gotten more up, which in a sense does make them riskier. Variance has gone down, but it sounds like the income's beta risk. So how much income's vary with the business cycle? It sounds like it's become more risky.
Fatih Guvenen: That's about the first moment. So this one I was describing is more about the shape of the distribution. So it's the third moment varying more with the business cycle. So it's like the skewness risk is covering more with the business cycle. Yes, that is true. But there's another concept that I also talk about, which is worker betas. And that is how much your income correlate with the business cycle. And that one we haven't, because it requires data that we don't have yet, so we haven't looked at whether that has changed over time, whether that correlation has become stronger over time. But that's a good question. I think in Europe there's data to do that. So somebody should look at that if it has gone up.
Allison Schrager: So myth four, I see we're running out of time. So I might have to speed through these next two is, the top 1% or largely immune to business cycle risk. Living in New York, I don't know why that was ever a myth. I know it is, but because a whole city, it goes to pieces every time there's a recession and everyone loses their job and finance.
Fatih Guvenen: True. I agree with that. I think what I was referring to and usually what people have in mind is if you go back to 2010, '11, if you remember in the media, there were a lot of articles that were looking at the top 1% share. And that's a statistic that we need to be very careful about. So that is you look at the total wage income in a given year and you look at how much of that is earned by the top 1% of workers in that year. And some economists have plotted this over time and you can infer a certain amount of limited amount of things from that. The reason is because the people who are in the top 1% this year and the next year can be completely different people. And there's a lot of turnover actually, at the top.
So in 2011 and '12, there were articles in the media saying that plotting this top 1% share and saying that, oh, the top 1% has completely recovered. So there's almost no loss to income. But the problem is you are comparing apples to oranges. People who entered the recession in the top 1%, they may have lost a lot of income and somebody else replaced them in the top 1%. Now you're comparing these two different people. So from a welfare perspective, as an economist, we cannot never do that. You cannot compare how it comes off to different people. So what we do in our work, we actually, that's why having a panel data, meaning data where you can follow people over time is critical. The structure we have is very different. We follow people over time and then we put somebody in the top 1% of the five year income distribution.
So if your average income was in top 1% for five years on average, then I say, buy 2006, you are in the top 1%. Then what I do, I look at the people in this bin, in this box, I look at where each of them is in 2011, five years ahead. And then when you compare this as a group for the top 0.1%, their income was about 50% lower. So this is not one or two people. This is the entire group's income was about 50% lower. And this was also true in the 1991 recession and the 2001 recession. So in the last three recessions, there was a very big income loss that was persistent. Yes, somebody now overtook that, but that's not a big consolation for the people, just like you said, who lost their job on Wall Street or some other high paying job and they never recovered from that. I personally met people, actually. I met an Uber driver who took me for a long distance trip and he was the president of a company that he completely lost. And since then he has been driving Uber.
Allison Schrager: I guess a lot of people do end up at the 1% at one point in their life. So what does the 1% even really mean?
Fatih Guvenen: Yeah. Yes. I think that's why it's makes more sense to look at more persistent measures like five year average. Were you there just by luck for one year or are you really a member of that community and that's different.
Allison Schrager: So last one idiosyncratic income shocks could be modeled fairly well with the log normal distribution. So again, what does that mean?
Fatih Guvenen: So that's fairly technical, but I think these discussions came up before the great recession, right? There were all this pricing of the CDSs with the AIG and other companies. And there was a lot of Gaussian modeling that went into those. And there was some criticisms of that. Yes. So economists, statisticians, we love Gaussian modeling traditionally because just like in math, we have linearity in statistics. We have Gaussianity. It simplifies a lot of analysis and it's everywhere. So for the longest time, until very recently, the way we modeled income shocks to individuals was assuming a lot of Gaussian shocks. We have fixed effect, meaning that's the permanent differences across people in their income. Let's make it Gaussian. We have annual income shocks that we get let's make that Gaussian. Let's also make the process linear, so every year it keeps being Gaussian.
And the Gaussian distribution has certain properties. We talked about one just a few minutes ago. It has zero skewness, it's symmetric. So you cannot have a shock that is more likely to be negative or likely to have larger negative shocks than positive shocks. You cannot have those. But another thing you cannot have is, in the data. When you look at fraction of workers, whose wage change by a small amount from year to year, this is what we sometimes call sticky wages whose wages change by say less than 5% from one year to another or 10% from one year to another, the Gaussian distribution will underestimate this by threefold or fourfold. So what this means is, in the data, there is a huge spike near zero of people whose wages change very little from year to year. The flip side of this is in the data. There are also very, very large shocks that we shouldn't be observing, if the data was really Gaussian.
To give you an idea, if you have a shock of that increases your income by eightfold, okay? The likelihood of that is in the data about 13 times higher than under a Gaussian. If you look at a shock that is 20 times, raise your income. I know we think these are very unlikely events, but actually in real life, how you model this does make a difference or think about your income, going down by 95%, that's something we can relate to more easily. You lose your job and you stay unemployed for most of the year. In a Gaussian world, this happens very, very rarely. And in the real data the likelihood of this is 20 times, 25 times higher than in data. So what we show is, if you just stare at the data, okay, do non parametrically unplotted, you'll see that the features of the data are very, very different than Gaussian.
So it was like a call to action to say that, yes, we understand it is convenient, but we have been doing this for the longest time. Gaussian lived more than 200 years ago. It's time to now ... And we offer actually, it's not our invention. This has also been studied for centuries now. There are modeling tools, for example, the simplest one is Gaussian mixing. Okay, you can mix Gaussian distributions and you can model the income data. You can fit it almost exactly with those mixtures.
So this is a more technical point, but at the end of the day, we put these in models. Then we do a policy analysis, I come up with a number and then some journalists will quote it as a result from some researcher. But at the bottom of this whole chain, actually it starts with an assumption I made. And if I made an assumption that really is so far detached from the data, the result will be really affected by that. So that's why I think this is important for the end result as well.
Allison Schrager: So I guess when we're thinking about what income risk looks like in America, you said, there's a lot, you're figuring out having access to this just tremendous data. And that I think the risk picture looks just, you said, a lot more complicated. It's not just a very that Americans or even Europeans have no sense of what their income's going to be in the year to year on average. They do, but they do some very big shocks and can move around the income distribution significantly.
Fatih Guvenen: It is. I think that's exactly the right word. It's more nuanced. It's more complex than the way we have been modeling. We have some calculations like we do in finance. We have a risk premium calculation. And what we do is, suppose I give you next year, you are going to consume, you're going to have to spend whatever income you make or alternatively, I'm going to take your average expected income, and I'm going to pay you that next year, but you have to pay me a premium. And we do that premium under the standard Gaussian assumption. And then under the data, the way it looks with its skewness and keratosis and all the empirical features we talked about, and we ask, how does the premium you would pay compare if you take the data seriously compared to, if you model it as a Gaussian with the same variance? The variance is the same in both cases and it's at least four times larger.
So that is to say can say that the risk is actually much higher than we have been modeling in that sense, but it's also more nuance because the Gaussian also implies that every year, the standard deviation of an income shock is about 50% in the data. So under Gaussian, a typical shock is 50% sorry. The standard deviation is a typical shock. So every year you can expect a shock roughly of 50%. And we don't really see that in the population. And the reason is, most years that doesn't happen. But once in a while you get these really, really big shocks that either raise your income or really reduce your income a lot. So the nature of it is also more complex than how we have been modeling traditionally.
Allison Schrager: Well, there's again, a lot like finance, where I think a lot of naive practitioners of finance tend to get very hung up on volatility as a mess risk measure and don't really think about the higher moments. Of course, I think like most economists, we tend to think of income as a lifetime asset too. So it shouldn't be such a stretch. And I think it's important because just like in finance, when people get hung up on volatility and they don't think about tail risk, that's when we tend to have bad things happen or we don't appropriately ensure. And I think one thing I've really enjoyed about your research is, if we get too hung up on the wrong thing, we don't come up with the right policies or this is that insurance, if you will, to help people manage risk better.
Fatih Guvenen: Absolutely true. Very true. Yeah. There are many similarities. Actually, next week I'm talking at a finance conference and I'm going to talk about all this research and somehow the topics I talk about, I feel like that audience is more ready in a way to understand them, because a lot of this research has already been going on under different names and different language already in finance. And so there are a lot of similarities. And my training when I first start my career actually was more heavily in finance. So maybe some of it is coming back in my research and helping me now.
Allison Schrager: Yeah, you can see it. I think being in macrofinance also, I have a life cyclist, I think that maybe that's why I've always appreciated your work. So-
Fatih Guvenen: Thank you. Thank you very much.
Allison Schrager: ... And as I said, I think it's a real triumph for Big Data. I wish all these results were on the cover of the New York Times because it really grows on the head, things we hear about all the time in the media. And not only is it misleading and makes people think there's risk where there isn't none or maybe not seeing risk where there is some, but I think it really skews policy, which I said could be much better at helping people maybe figure out different ways they could be having more wage growth and better ensuring people from negative wage shock.
Fatih Guvenen: Absolutely. Yeah. Very true.
Allison Schrager: So as I said, we've run out of time, so thank you so much for joining me. That's all we have. I'll link to his research in the show description and you can find him at Twitter @fatihguvenen, and you can also find City Journal on Twitter @Cityjournal and on Instagram @cityjournal_MI. And as always, if you like what you heard on this podcast, please give us a five star rating in iTunes. Fatih, thanks again for joining us.
Fatih Guvenen: Thanks for having me.
Photo: gremlin/iStock