This is a lightly edited repost of a twitter thread. So, short paragraphs ahoy:
I’ve finished Bryan Caplan’s The Case Against Education and it’s definitely very interesting, although I remain unconvinced.
PART 1: I’M AN EDUCATION ECONOMIST AND I REALLY LOVE SIGNALING
On a personal level I really love the signaling model. If you are not familiar, it’s basically the idea that if you want to prove you have quality X, then you do wasteful and useless thing Y that someone WITHOUT X would have a harder time doing than you.
That’s it. Then, if someone sees you do Y they’re convinced you’re X. Sure, someone without X could copy you by doing Y too, but if it’s harder enough for them than it is for you, they don’t bother.
I think this is just straight up *cool*. I learned it in some economics class, immediately glommed onto it, and looked for it wherever I could. Once you start looking, you see plenty of examples.
I think it’s one of the most important ideas economists have popularized, and I’ve included it in every course I’ve ever taught (although, alas, it has fallen off the behavioral econ syllabus).
I’ve certainly spent my fair share of time looking into it in the educational context — I still have a strange fascination with Bedard 2001, which finds evidence of it. Signaling in ed ~ ed *certifies* your ability, rather than *improving you*.
To be clear, the book is trying to address: (1) we see that getting education raises your earnings. (2) So, how much of that raise is human capital, and how much is signaling? (3) The book suggests it’s mostly signaling. (4) So what?
Nowadays, I’m an economist who mostly works on education. So I read a whole lot of papers on the economics of education. Granted, not usually on human capital (the “ed improves you” story) vs signaling directly, but I see the papers for sure.
And what do I see? Not a massive amount of work that supports signaling as the *primary* explanation for the impact of education. Signaling effects? Yes. MOSTLY signaling effect? It’s just not there. Now, why is it not there? There are a few reasons.
A biggie is that it’s just super hard to be able to point to data and be able to say with real confidence “that right there? It’s signaling. It’s not anything else, it’s signaling.” The predictions of human capital (HC) and signaling (S) overlap A LOT.
Most of the time, when talking about what you observe in the real world, nearly all stories can be twisted to be signaling or human capital [this is, by the way, something that I’ll come back to later when discussing the particular claims that The Case Against Education makes]
There are papers that do it — Bedard 2001 is an example. Tyler, Murnane, & Willett 2000 is another. But these rely on super-specific circumstances to be able to tease out something that is undeniably signaling.
The reason I’m focusing on this is to address where I — and much of the ed economics literature — is coming from. Caplan takes the position throughout the book that, as a field, we’re basically ignoring the possibility of signaling, treating it as too unimportant to bother with.
Maybe this is rhetorically necessary for the book to be framed as him Standing Up to the Man. Or to explain why so many people interested in the question have come to a different conclusion.
But that’s not what I really see. I Googled around for Econ of Ed syllabi — I’ve yet to find one with human capital but without signaling. Ed signaling even pops up in Principles textbooks — heck, Mankiw has it.
Papers finding plausible evidence of signaling publish quite well, and, without having actually gone out to count, it seems more common to see studies of the earnings boost from college that omit any mention of HC or S than to see ones that just mention HC without S.
I think Caplan paints the opposition incorrectly. Caplan claims education is mostly signaling (his estimates vary from 50–80%) and says that econ opponents are ignoring common sense or are following Social Desirability Bias. I just think they don’t see evidence for that claim.
Now, one point here that Caplan hints at but does not actually make is that people tend to assume that human capital is the *default* explanation. Lots of stuff could be human capital or signaling. So if you can’t prove it’s definitely signaling, it’s HC.
That’s wrong too. My personal take (and I had a tweetstorm on this long ago) is that there’s so much overlap between the theories that using evidence to actually pick a percentage, *any* percentage, of HC vs. S is basically a dead end.
Working on a paper about this, actually. TM TM TM Don’t Steal My Ideas.
[Caveat: wouldn’t be a dead end if the truth were 0% or 100%. But it’s not. On that note, if you respond to these tweets under the assumption that I’m gung-ho HC 100%, you haven’t read this far.]
PART 2: TALK MORE ABOUT THE BOOK ALREADY
The book is separated into a few parts. The first explains the basic idea behind educational signaling and clearly delineates what it’s not (i.e. “haha, people with BAs are working at Starbucks, education is just signaling” is incorrect).
This part is great. It’s a crash course in the basic theories of human capital and signaling, and lays out several basic facts about the education system. The *fact* is that education raises the average person’s wages, causally. The question is why.
The second part covers his case for signaling being the *why* explanation that covers most of that raise. The third is a lot of back-of-the-envelope calculations of what that means for educational investment.
The fourth takes those back-of-the-envelope calculations and makes policy recommendations (cut education by a lot, cut non-job-focused classes like music, and do more vocational ed). The fifth spends more time clarifying the argument against common rebuttals.
The first and fifth parts I don’t really have much to add on, but I’ll continue by focusing on two things: the argument for signaling, and the policy implications that follow from it.
In the true fashion of a reviewer who really wants to talk more about himself, as I write this I’m pretty sure it’s already more than half over. [edit: it’s not]
PART 3: THE CASE FOR SIGNALING
I’m going to argue here that the case for signaling isn’t nearly as strong as Caplan makes it out to be. I have two main points. The first is that a whole lot of evidence fits multiple theories, and the second is that Caplan surprisingly gets signaling wrong several times.
As I go through this, if you happen to be a HC-hater (tend to find a lot of those in smartypants autodidactic internet circles, wonder why), keep in mind that I’m NOT assuming that arguing against his claims means I’m claiming HC is the correct explanation instead.
That would be the “default-HC” approach that I was talking about earlier. I’m instead making the case that for a lot of evidence that we see, there’s no clear reason to pin it on one theory, and we end up just going with what feels right.
There’s a lot of overlap between the two theories. So if one theory can claim overlaps by default, it’s automatically going to be the biggest explanation.
Caplan does this. In effect, he’s taking a “default-S” approach, where many people take a “default-HC” approach. No clear evidence that it’s HC? Must be S! Fair enough to him, I guess. Neither is right, but at least doing both is fair!
[btw, both “default” approaches fail not just because of overlap, but because HC and S aren’t even the only explanations for the premium! For example, students may use school to DISCOVER their speciality, or there may be NETWORKING. Those are neither HC nor S]
So let’s talk about some overlaps that are claimed in the book as signaling!
Let’s start with one of his claims that I found intriguing — much of the signal that education sends is that you’re willing to be a conformist. If ed is just a signal, why not just replace it with an IQ test? Because that wouldn’t tell you about social skills & conformity.
However, this overlaps neatly with a HC explanation that school *teaches you* to be a conformist, and by forcing you to be around other children raises social skills.
I don’t know which explanation is right. But you’d need some pretty wild data to be able to tease them apart. Absent that (and we are absent that), interpreting the effect is down to common sense. Unfortunately, common sense from different people gives different answers here.
Another is similar — Caplan, throughout the book, targets job-irrelevant skills taught in classes like history, high school English, advanced math, and music. There ARE returns to taking these classes. But why?
It could be to just signal that you’re capable of completing these classes (S). However, it could also be that, even though it’s teaching you skills that don’t transfer to your job, it is also teaching you skills that do (HC).
Caplan cuts off this particular argument by pointing out that ed psych shows that we tend to learn what we actually do, not something else.
So what do we actually do in these courses? Push through hard problems, endure boredom, write, follow instructions, coordinate and communicate. Those are all job-relevant.
That’s not exactly a glowing endorsement of ed, but it IS a HC story. A common mistake I see in discussions of HC vs S is that any negative aspect of ed must point towards signaling. Nope! Caplan largely avoids this, but not always; this is a good example.
There are lots of small bits of overlap like this, but let’s turn to the big one — sheepskin effects. The sheepskin effect refers to the fact that much of the wage return to education comes in the year you graduate.
Under basic human capital, you earn more because you learn more. But you probably learn similar amounts each year you’re in school, so the return should be similar each year you’re in school. But it’s not. One-year-from-grad dropouts earn way less than graduates.
This is generally taken to be pretty strong evidence for the presence of signaling, and I agree. But is it *all* signaling? Again the overlaps emerge. Let me make a very Caplan-style argument leading to a very non-Caplan conclusion:
“What schools actually *do* when they allow you to continue in your education is, effectively, measure what you’ve learned and see if it passes some minimum standard. If you don’t, you drop out. We end up with those failing the (lax) minimum populating the dropout years. They’ve learned little, so they earn little. In the final year, you see everyone who passes the minimum, whether they learned just enough or WAY MORE than enough. The final year contains a wide range of big learners, so on average there’s a big jump in earnings that year.”
Intuitive, based on literal actions people perform, and a totally-HC explanation of sheepskin effects. Do I believe sheepskin effects are all HC? Certainly not (I’m not Noah Smith). But it’s probably not 100% signaling either.
And that’s a VERY IMPORTANT distinction. Pretty much every calculation that Caplan makes rides heavily on the exact size of the sheepskin effect, and attributing it 100% to signaling. I’ll come back to this later when I talk policy.
My point is that this is mushy. It’s all real mushy and it’s nearly impossible to use data and evidence to make much sense of it. HC and S aren’t the most easily falsifiable of theories.
Caplan would criticize that, saying that, sure, academics are going to get real picky about this sort of thing, but we need action now. You have to choose SOME education policy.
This is true, but my point is that the justified action necessarily depends on which theory just sorta feels right, not data. Not a great basis. Or, rather, *since it’s so indeterminate, maybe this isn’t the right question to ask when considering policy.* I’ll come back to this.
There’s a related issue here on Caplan’s handling of the evidence. The book is very well-researched, but omits a rather important vein: pretty much all the studies that really attempt to isolate HC or S effects. Largely omitted or glazed over.
Caplan’s says that these often rely on advanced statistics or natural experiment and so are untrustworthy in their own way. This is true! But the reason they do this is so they can actually make the difficult step of being able to confidently say “This is S.” or “This is HC.”
They HAVE to do the advanced stuff so as to avoid being mushy, and combating alternative explanations. Because the differences between the theories really do only exist in the cracks. Caplan prefers more intuitive evidence. But this evidence is mushy in its interpretation.
Convenient, at least, that form of evidence least subject to being bent into shape to fit a favored argument is disallowed and ignored. To Caplan’s credit, he sticks to his rule and doesn’t make exceptions for these papers that DO find signaling.
The other big omission in Caplan’s reading is Goldin & Katz’s The Race Between Education and Technology. Unlike the other papers, it is a book-length, more-intuitive, less-fancy piece of work.
G & K is widely read (for an academic work), well-regarded, and basically fits in the style of evidence that Caplan says he favors. And it comes down heavily for education-as-productive. But it is omitted except for one technical footnote. This is a big omission.
PART 3-B: CAPLAN VERSUS SIGNALING
The Case Against Education is not Caplan’s first foray into the topic, and he has been blogging about it for years. I’ve been following them loosely as they come out. One big problem I keep coming back to, and which is still present in the book, is this:
Caplan gets the details of signaling wrong with some regularity. More charitably, his model of signaling is different from the classical one in some important ways. But if it’s the latter, that means that broadly-accepted implications of signaling don’t necessarily apply to him.
In particular, as a lover of signaling (see Part 1) this is the thing that has always bugged me the most. I can truck a disagreement on policy, but God help your mildly inaccurate application of some nerdy, nerdy rules.
To dig a bit more into the ed signaling model: (1) It must be the case that there are differences between people before education, (2) getting education must be *more costly* for undesirable employees than desirable ones or undesirables will get ed too, and it falls apart, (3) as a result, there is a difference in *average* ability between people who complete ed and those who don’t, even if ed didn’t improve them, (4) this lets employers find the high-skill workers, who they are willing to pay more, so you see people earn more with ed.
My first quibble on this is minor, and in regards to (3). Caplan spends a fair amount of the early part of the book establishing that college students don’t actually know that much. Not just that they haven’t *learned* much, but that they don’t *know* too much.
Which, if true, is not a great look for HC! But it’s not a great look for S either. If there’s little difference between educated and non people post-hoc, that’s not a signaling explanation. These tests must not be covering skills employers care about (learned or innate!)
The second issue has to do with costs, in regards to (2). Basically, for signaling to work, education has to be very difficult or expensive for the low-skilled. One strain of research Caplan brings up is the Academically Adrift line showing that students spend less time
learning in college than they used to. Study time is down, socializing up, passing easier. Not a great HC look! This should make college a LESS effective signal, but Caplan strikes it as a win for signaling, even though the return is NOT dropping in line with these changes.
Additionally, we actually see a broadening base of college grads over time, coming from lower down in the ability distribution, and yet the return maintains its size or grows.
Further on the costs issue, Caplan advocates for trying to push down overall attainment; if the return is due to your relative education rather than absolute, then we’d all be fine if we just all got one less degree.
But this ignores how costs work in signaling. In a signaling equilibrium, you have to make sure that sending a particular signal is costly enough to keep imitators away. If you put everyone on one less degree, a whole lot more “imitators” would spring up.
This would quickly bring us back to where we are. “Degree creep” is likely a signaling phenomenon (although, mushy mushy, also explainable via HC if labor market demand for high skill is increasing, which it appears to be). However, AS a signaling phenomenon,
degree creep is an equilibrium outcome that is likely to reconstruct itself unless the cost structure changes. Granted, that is part of what Caplan is advocating.
The biggest issue I have with Caplan’s interpretation of the signaling model has to do with its policy implications, which I will address in the next section!
PART 4: SO WHAT?
Another big part of the book is Caplan’s calculations on how much value education brings to individuals, and how much it brings to society. His conclusion: The individual (“selfish”) return suggests most people should get HS degrees, and college if they’re smart.
But socially we should discourage college, which has negative returns for everyone but the real smarties, and also cut way back on high school too.
The calculations are rather impressive, and he manages to fit in a whole lot of small effects that are not typically done in returns calculations (I could certainly point to details he missed, but mostly not that would change the results too much).
However, when it comes to social returns, an absolutely massively important part of the calculation is “the signaling share” — how much of the return to education is S, and how much is HC? He uses 50% (“reasonable”) and 80% signaling (“preferred”).
I have a big problem with this for two reasons! One I’ve already covered — I think it’s basically impossible to come up with a realistic estimate of the signaling share using data. That 50% comes from the size of sheepskin effects, which I don’t think are 100% signaling.
In fact, I think most effects of education you see, whether it seems to make sense that you’d assign them to S or HC, are not fully either. That makes Solomon’s sword a whole lot wobblier in terms of where it decides the share is.
The second problem, and it’s a major one too, is related to Caplan’s understanding of signaling, as I mentioned in 3-B. And it’s this: Caplan assumes that the signaling share is the *nonproductive* share. And that’s not accurate!
What I mean is this: society’s *economic* interest in education works to the extent that education actually makes people more productive, rather than just letting them get ahead at the expense of others.
And so when Caplan (and some others) make a HC vs S argument, they’re letting HC proxy for “productive education” and S proxy for “nonproductive” — but this is wrong. Because signaling can be productive!
In fact, signaling almost certainly *is* productive. Sure, in the simplest signaling model where there’s only one job, education signaling is zero-sum (if you get ahead it’s at the expense of others). But as soon as you introduce multiple, different, occupations, that changes!
For a basic example of this, say you have two occupations, each with room to employ half the population, “Simple” and “Complex.” In Simple, everyone produces 1. In Complex, smarties (half the pop) produce 4 and everyone else produces 0.
In extreme-land where there’s no signaling, people are basically assigned to jobs randomly, everyone gets paid a wage of 1, and average production is .5*1 + .5*(.5*4+.5*0) = 1.5.
Now introduce a completely wasteful, no-HC education signal. Ed is free for smarties, but it costs 3 for everyone else because they’d have to hire tutors. Now, employers know who the smarties are, and they all get Complex jobs. Others get Simple jobs.
Production is .5*1 + .5*(4) = 3. Smarties get a wage of 4 and others get a wage of 1. Others don’t bother going to school — they’d get a wage bump of 3 but they’d spend it all on school anyway, so why bother?
Here, signaling alone doubled production (this approach is generally called “matching”). There’s a reason we call signaling outcomes the “second-best” outcome in game theory, and not “the really awful outcome” — because it can increase efficiency over NOT having the signal.
Am I saying that *all* signaling is productive? Of course not. There’s a zero-sum element to it too. But *some* signaling certainly is. How much? No idea. But Caplan doesn’t even consider the question. What if it’s 80%? half? Or even 20%? His numbers would all be WAY off.
Finally, as one last aside here — Caplan goes into detail on his support for vocational education. I think this is great! Go voc ed. There are undeniably lots of students who just really hate being in a classroom.
If you push those students to HAVE to go to college, they’re unlikely to graduate. Vocational ed is a great alternative. More, please!
The one sticking point I have on this is in how Caplan illustrates his opponents. Because not only do *I* like vocational ed, but as someone who hangs around in a lot of ed policy circles, so do a *lot* of other people.
In fact, I hear a fair amount about it, all positive, but Caplan frames it as this taboo idea. I don’t think that’s accurate. Either in academia or policy.
I dunno, maybe it’s a new thing — I haven’t been around *that* long. Maybe fifteen years ago voc ed was the policy of which no-one must speak its name, and that’s what Caplan is thinking of. But now, this is a potential area of alignment with a lot of people.
PART 5: CONCLUSION
So what can I make of this book? There are certainly some praises I can make — it raises in prominence an important economic theory concerning education that is not well-known enough. It is very thoroughly researched, although it has some blind spots.
There are some ways in which it plays fast and loose with the theory, both in interpreting signaling itself and in using data to understand S vs. HC. But I’d be lying if I said that at least some of these sins weren’t also in play with the pro-HC crowd.
Ultimately, I think that the HC vs. S question is the wrong one to ask. They’re both interesting theories (although S is way cooler) but the extent to which they empirically and policy-wise lead to the same dead ends makes a concern about which one is “more right” inert.
So the answer is, I think, find a better question. That said, I do think the book is worth picking up. Especially if you are a blog-reading person (the style of the book is much more like a blog post than an academic study, with the pros and cons that entails).
If you are completely unfamiliar with the HC and S theories, this is a great intro, although do be aware that he is coming in with a very one-sided argument to prove. If you are a HC-person, it’s good to be fully aware of the places where that theory fails.
And if you’re an S person, well, you’ve probably already read his blog posts.
So do pick it up and read it, keeping in mind that, to whatever extent there are cracks and mushiness in the HC side, they exist here too. And then, come join me in the comfy, cozy confines of indeterminacy. /fin