Someone proposed that most of us are naturally short equities, in the same sense we are all short housing. Though a little unbelievable at first – aren’t we all better off as the market booms? – I’m starting to think that disdain confuses correlation and causation.

We can explain this by drawing an analogy to housing, something we all need. When the Mayflower reached the continent, all the Pilgrims were short housing. Except some were natural builders who could net that short with a claim on future demand. That net long now resides with landlords and families with vacation homes. Someone is on the other side of every short.

So too it must be with equities, or more simply the residual claim on production after more senior creditors have been paid off. Most individuals have a comparative advantage at being senior, since most goods and services are commodified they demand a market rate and offer security and certainty. Most individuals may also be unqualified to generate enough residual income through entrepreneurship to offset risk aversion.

Suppose the risk aversion of those around you falls, reducing the equity premium and increasing the subordinated capitalization. There are two effects. More people are engaging in risky but positive expected value activities, increasing production. However, as equity values rise your senior claim as a laborer commands an even smaller claim.

The asymptotic effect is that the boom in subordinated value crowds out everyone else, resulting in unaffordable inflation for other goods we are commonly short – housing, food, apparel, etc. Those naturally inclined to be entrepreneurs generate the initial long position on subordinated claims. The rest must hedge this short by investing in the stock market. The story doesn’t really work the other way (where entrepreneurs are short the senior claims) because the more correct analogy might be a call which has limited downside. This creates the confusion between correlation and causation. We are all better off as the market booms since that reflects better technology. But only the subordinated claims have this upside. Another way of reaching the same conclusion is that as the economy shifts towards new technology, you need to be able to afford that new technology; something only exposure to the subordinated claims can offer you.

In The Second Machine Age, Erik Brynjolfsson and Andrew McAfee capture the GDP effect of some technological progress as the price of a good falling from infinity to zero. This might get the supply-side tension right, but there are a few other problems in measuring long-term changes in standard of living. This post is a question more than answer and one that touches on interesting aspects of the debate on unmeasured consumer surplus, magnitude of economic wellbeing, and secular stagnation.

As a baseline, consider the BLS method of hedonic adjustment. To account for changes in product quality, they generate a regression of log price on various attributes (for example pixelation and screen size for television). They then consider improvements in these factors over time to make sure estimated inflation doesn’t overstate the truth.

This gets the year-on-year figures right, but makes long-term adjustment really hard. Specifically, consider your Internet service that’s probably too expensive. Hedonic adjustments will adjust for increases in speed, at least at a first order. Now of course, the consumer Internet didn’t exist in 1980 and therefore it’s theoretical price was infinite.

But suppose you could move the Internet back to 1980. Everything. Routers, telecom wires, transpacific cables, etc. The market price of this service would be next to nothing, since no one has computing technology to browse the web, since no engineers exist to create worthwhile content, and since building the necessary infrastructure from the state of science at that point in time is extremely hard.

In that sense, the shadow market price of the Internet that doesn’t exist yet is 0. This isn’t true for all technology. Shoes today are way better than shoes in 1980, and even if the technology didn’t exist to create them as we do now, the market price would still be far greater than that for the 1980 equivalent.

This framework as easily suggests there was massive deflation in the price of shoes (which is true) and a massive inflation in the price of Internet (which is false).

Our examples don’t need to be so extreme. An iPhone would be considered vaguely useful as a portable camera, but without technology to listen to music it wouldn’t be nearly as popular as it is today. Still it’s an incredible product and well-above its time in technology and should command a premium (this is the entire argument of a hedonic adjustment). And this of course assumes a stability in tastes and preferences.

This is intimately connected to the question of consumer surplus. It’s frequently said that productivity improvements are underestimated given the “explosion” of consumer surplus from web services like Netflix, Google, and Facebook – “how much would you pay for Facebook”, the question goes. Still it’s unlikely that this surplus actually goes unmeasured. I almost certainly wouldn’t pay $100/mo for LTE data on my iPhone if I couldn’t access Facebook, Google, or iMessage. Without considering the cross elasticities between free products, and the new industries they tempt, it’s hard to argue that GDP is underestimated relative to the true economic benefit.

Consider clean air. One day China will have a technology that cleans its air despite extremely intense energy consumption, and it will be extremely cheap relative to the life and economic savings it generates. Its creator will be praised for creating all sorts of consumer surplus unmeasured in China’s GDP. Unmeasured except for the boom in foreign investment, outdoor playgrounds, and botanical gardens that is.

Arguing that we didn’t have Netflix or Google in 1980 isn’t enough. We pay for Comcast and Apple, which are both prominent in GDP numbers. Of course this doesn’t say much about the distribution of that surplus – it might be that this benefits certain percentiles more than others, but this isn’t an easy claim to make without reference to the relevant cross elasticities of demand.

Does this mean inflation isn’t transitive? How would we model that? An error term that grows unreasonably when considering changes over decades or more?

It might be that the right answer is extremely large standard errors in estimates of long-term inflation. As certain markets grow large at the expense of others – cars versus cabs, for example – the statistical basis for making hedonic adjustments against an increasingly non-representative base challenges statistical conviction.

This is relevant in finance as well. 30 year Treasuries price in some expected degree of inflation. But the annualized inflation rate is likely very different from the true inflation, even though it’s correct on a year-on-year basis. Moreover, to the extent inflation is low from hedonic adjustment, it is because of a dramatic increase in the real growth rate.

Is there another premium in long-term bonds? A term premium yes, but also a “hedonic error premium” – which would include the expected inflation as measured (i.e. the breakeven rate) along with the error term that is larger in 30 years than it is in 10 years?

We could have a market for how much better 2015 is than 2014. We don’t have a market for how much better 2015 is than 1800. A diligent team of investors could answer the question “how much would you have to pay me to use 2010 healthcare instead of 2015 healthcare”. It’s not clear the same team could answer the question “how much would you have to pay me to use 1980 healthcare instead of 2015 healthcare” (a question posed by Larry Summers to suggest there is unmeasured productivity improvement since 1980).

Or it could just be that the existence of old age dating apps and the Internet makes being 85 much more tolerant – and the consumer surplus of the needless to say cheap apps is masked in elderly healthcare costs.

Now this isn’t to say technological progress is underrated in its contribution to humanity. If anything we probably owe earnings growth in legacy industries to technological surplus.


Most people agree GDP is probably a flawed tool. Results from a few papers I’ve recently read – which I’ll get to shortly – increase my curiosity. Ignoring its many other shortcomings, I want to focus on the difficulty of measuring realized GDP. My entire premise might be flawed, and definitely contact me if you can explain why. I might be very confused about the whole thing.

Consider an economy with households, firms, entrepreneurs, and banks. Households keep all their income as deposits in a bank, which invests in firms in the form of long-term debt, with residual business income accruing to entrepreneurs who hold equity.

Suppose tastes and preferences change and households no longer purchase widgets. Unfortunately at the beginning of this year firms had invested in large factories to scale their widget factory, financed largely by debt. Widget factories, having lots of unsecured debt, decide to default on the principal. Continuing operations is profitable for another year or two, however, and therefore they remain current on interest payments for another 2 years.

If widget factories decided to go bankrupt today, banks would have to writedown all their debt – taking a huge expense – and resulting in a sharp fall in production. Instead, widget factories will slowly default over the next decade, overstating ex ante GDP.

Now this effect might be moderated with a sensible requirement that banks immediately expense income for a loss reserve as they lend – with symmetric errors between reported and true earnings as borrowers default and recover. However my understanding is that the BEA uses current earnings excluding such accounting practices; and even if such were included the effect still remains large, especially when historical delinquency rates fail to capture the prevailing economic environment.

A working paper from Amir Sufi and Atif Mian prompted this question. In particular, they find that a one standard deviation increase in the household debt to GDP ratio over the last 3 years (6.2 percentage points) is associated with a 2.1 percentage point decline in GDP over the next three years, and outcome that seemed extremely large to me, especially given the econometric robustness they detailed.

Could it be that periods of substantive increase in household debt, and the rising delinquency rate that comes with it, make it harder to match economic revenues with economic expenses – perverting growth estimates? Even if the effect on GDP itself is small – which, if true, it cannot be given the meaningful role banks play in advanced economies – the effect on growth is meaningful. Consider valuation adjustments in financial inventories:

Screenshot 2016-02-02 00.48.50.png

Economic losses are unlikely to have been so concentrated. Non-recourse borrowers were economically delevered as soon as housing prices started to fall; and the ex post GDP in these years was much lower than that reported ex ante (not to be confused with revisions).

In this particular case, it is also likely the regression coefficient on HHD is also overstated. The increase in (Reported HHD / Reported GDP) is about 6% lower than that in (Economic HHD / Economic GDP) under standard calculations – where banks reserve losses at historical rates, growth is about 3% in the period of increasing debt, and the debt cycle gathered momentum over 10 years. In practice the effect is likely a little sharper as GAAP loss reserves are not included in the BEA estimate. In other words, the increase in the HHD/GDP ratio needed to achieve the same decrease in GDP is actually larger than estimated.

The particular problem I think this poses for economic analysis in general is from measurement error. Even if revised reported GDP is negligibly erroneous against perfectly reported GDP, both are meaningfully erroneous against perfectly reported economic GDP. There’s no good fix to this problem – it’s really hard to find out who has economically defaulted and who hasn’t. Doing a better job adjusting for loan reserves might help, but also open a host of other problems regarding standardized measures.

Of course the reported GDP eventually converges to the economic GDP, as losses must eventually be recognized. However, this process is subject both to institutional flaws, such as slow bankruptcy proceedings, as well as macroeconomic trends, such as low interest rates that allow borrowers to remain current on bad principal for a longer period of time.

Maybe more importantly, the measurement error attenuates regression coefficients and economic GDP turns out to be a lot more important than we gave it credit for.

I’ve worked out a back-of-the-envelope model to see the discrepancy that I’ll write about soon. In any case it generates an annualized 0.3% to 0.75% error in annualized growth rates under reasonable economic environments.

Some remedies might include a push towards using market, rather than face, value of debt. For example, sometime in 2007 the market value of subprime started collapsing even as no writedowns were realized – because investors realized the owners of these houses either had, or were going to, default and hence effectively unlevered.


The harder question is estimating the current, capitalized value of GDP. If there is any interesting writing on this topic, I’m interested.

By now you’ve heard that YC wants to learn all about it. I’m completely unqualified as an economic experimenter, but I have a few thoughts about the curious premise.

Very obviously any affordable study can only observe the individual rather than structural benefits of a basic income. There are genuine nonlinearities in the progression towards a world where (1) everyone has meaningful purchasing power, (2) where inefficiently-administered government welfare systems don’t create needless problems, and (3) where (1) and (2) are taken for granted.

But it’s also very likely that the move towards basic income will probably start with annual awards on the order of $1,000 to $5,000 in concert with a host of other legacy welfare technology rather than a complete revolution. Especially in the United States.

Therefore the reach of Sam’s fundamental question – “do people, without the fear of not being able to eat, accomplish far more and benefit society far more” – probably exceeds its grasp. But we can study the latter, something I’ll get to soon.

It’s also worthwhile being mindful of creating an observer effect. It certainly seems like a bad idea telling a young family with a newborn that it has a good fortune of a basic income, and don’t worry economists and sociologists will only study its developments every month. This effect would be particularly acute if the basic income is anything meaningful. Call this the “be careful not to create a lottery” bias.

If compelled I’d propose the following.

  1. Scrap the 5 year limit. It dramatically constrains the economic scope of a basic income and is unlikely to yield an interesting result. Individuals smooth income over a large period of time, and this is likely to suffer from a “lottery bias” as a result.
  2. Opportunity is one of the most important themes in American socioeconomic discourse, and opportunity is most elastic during childhood. Find a way to have a substantive answer to the question “how does basic income level the playing field for poor children”.
  3. Avoid the observer effect by finding ways to give people a basic income without their knowledge: though this is tricky for reasons I’ll outline below.
  4. Attempt to maintain general best practices for experiments (randomization, large sample for high statistical power, etc.)

The best way to satisfy (1), (2), and (4) might be something like an annual award on the order of $3,000 to $10,000 for couples that have just had a child for randomly selected people in randomly selected cities (ensuring some diversity in both type of people and type of city) for 20 years.

500 families in 10 cities at $5,000 would cost $25 million a year. That sounds doable for the Silicon Valley elite. Though there are plenty of ways to reduce this and still learn a lot. We can start with a single city. We can cut the study period to 10 years (and divide the group into early and late childhood – to see where the effect is strongly felt). I imagine the project should start in a single city – the administrative cost of studying 10 over 20 years isn’t negligible.

Though $5,000 isn’t a basic income, it is a little under what the average family spends on food. $10,000 to $15,000 is affordable, especially if we have only a single city. But at this order we sacrifice statistical power and also tempt strong observer effects.

For one, $10,000 to $15,000 awards will create envy. In neighborhoods where this program is most useful, this figure eclipses 100% of disposable income. The point of a basic income isn’t to create a table of special kids eating avocados amidst an inner city cafeteria serving mass frozen pizza. It would also be impossible to provide such awards without alerting the recipients.

Indeed doing so in general is hard. Working with employers is one option, though as economists know telling an employer you’re giving employee X an extra $5,000 means it can spend less to keep that employee than it would have to otherwise.

  1. It would be good to work with corporate echelons of a distributed employer that would randomly increase the salary without the knowledge of more immediately located hiring managers.
  2. Working with the IRS to randomly increase tax refunds to selected families could also work though this would need someone willing to work with the IRS.

The ability to precisely study beneficiaries is diminished if one is vigilant about a non-observable experiment. That’s a tradeoff YC will have to make. But the guilt, envy, awe, shock, and puzzle of receiving a lot of money will have large effects.

As a guiding principle, it may be cheapest to start an experiment that researchers can most quickly learn from and modify as needed (sequential hypothesis testing is valuable here). Building an apparatus where making modifications and adjusting the experiment doesn’t void the statistical validity of results already obtained is valuable.

Maybe the simplest advice is spend a lot of money by giving reasonable chunks of cash to a lot of people over a long period of time. Once collected data is made public the academic value of this data will be unbeatable. This is a landmark study if done correctly. It would directly challenge the government as both a provider of welfare, but philosophically demonstrate the capacity of Valley billionaires to produce meaningful research answering the hardest questions in social science.

I don’t know. Alex Tabarrok has an answer, though I’m not sure its sufficiently skeptical.

Of course, housing is overrated as a financial investment, largely because for most people it isn’t a financial investment, as much as a hedge against an increase in price of rent. You don’t get “a great investment and a place to live” primarily because you only get a place to live and very little investment. When the S&P500 shoots up, most of us can sell and live richer lives. When the housing market goes up most of us can sell… and buy another house a market that has already realized appreciation.

There are instances in which you might notice a genuine increase in wealth associated with a tighter housing market. If the price of large homes outpaces that of small homes, and you’re about to retire at that moment, you might be in luck. If a new job takes you away from San Francisco where prices have skyrocketed, you might be in luck.

But in general it’s not exactly easy to predict these trends – and it certainly shouldn’t be the province of homebuyers and the real estate agents that inform them. That doesn’t mean home ownership is a bad idea. In general, we don’t like cash flow volatility. A young couple working at Facebook, with enough cash for a downpayment and enough earning power for a good mortgage, might prefer to avoid the ups and downs of a tumultuous Bay Area rental market by purchasing a house. We’re all short the market to some extent.

In parts of the country without stratospheric home prices, a rental market is pretty nonexistent for affluent buyers. To the extent you want a nice family home with a pool in a safe, Iowa suburb, you may not have the option to rent. Indeed, “should you buy or rent a house” is question relevant only for affluent urbanites who have both the option to rent a decent place and credibility to borrow on good terms. The best argument to buy is that you probably live in a city where only kids at the local college rent.

The tax benefits of ownership are also questionable. While a mortgage interest deduction certainly exists, this is passed through to renters. In general it subsidizes building property since land is inelastic, and therefore is likely most beneficial for affluent midwesterners who like a lot of property on worthless land. This is not to say ignore the benefit in your calculus, as much as not to tell yourself that this benefit militates in your favor without actually looking at the numbers – more precisely, learning that there is such a deduction without any further information shouldn’t change your decision.

A thought experiment is useful. Imagine a town with some borrowers, owners, and a few landlords. If the government subsidizes interest on houses, market competition only increases the value of land relative to everything else but doesn’t change the relative price. Marginal investment is poured into building more property on top of that land. The supply of property-intense housing increases, and the price falls. Renters and owners both benefit equally, and because of competition not at all. The only obvious winners are landlords. The mortgage deduction doesn’t benefit the affluent, it benefits large landholders and their investors. Yet another tax debacle that benefits the superrich.

The pass through effect is muted somewhat by the asymmetry of potential buyers and renters in that some individuals might be “captive renters” as they can’t access affordable, long-duration credit on the principal at hand. Though this is almost tautologically irrelevant for markets where the question “should I buy or rent” persists; for any given house the individual who can pay the most, either capitalized or amortized, can afford to borrow.

Of course, accepting that housing is rarely an investment also challenges the associated wealth effects, either positive or negative, that many economists have observed. To focus the skepticism a particular theory, the Sufi-Mian leveraged loss multiplier, consider liquidity instead of solvency. The prevalence of non-recourse borrowing in the United States meant borrowers could pull residual equity into cash and leave if that value was above the collateral value.

Between 2003 and 2007 we had a bunch of people with high propensities to spend with access to non-recourse financing collateralized on a volatile asset. This was money that would be paid back only if prices kept rising. They weren’t leveraged precisely in the state of the world where that leverage would have been binding. It turns out that this was the state of the world that came to be, but the decline in spending wasn’t a question of solvency but loosing access to extremely cheap liquidity. Of course consumption and residential investment fell most in the leveraged areas where it increased most because that’s where closing the tap of free cash most dramatically affected ability to spend.

The best reason to rent might be that many educated people think housing is a good investment increasing the price of housing credit beyond what is justified – and an unlimited supply of conforming loans doesn’t really matter in this case since the people for whom this question is relevant are almost certainly buying a house above the conforming limit. People may also tell themselves that they get a tax credit one way and not the other (even though this is priced in both ways to begin with) which would further distort the market in favor of renters.

And even if housing was an investment, buying a house would infrequently be the best way to get there. The home you own faces plenty of idiosyncratic risk against the market as a whole – by region, by local policy, by construction style, etc. It would be foolish to expose yourself to this risk without relevant expertise – something both homebuyers and their real estate agents certainly lack.

There are other ways to express your beliefs about the housing market. If you already own a home, get a second mortgage against it and use that to buy an ICF index, your financial position might be closer to what many financial writers ascribed to middle-class Americans in 2005. Though, as you can see, that is probably far from where you actually were at the time.


I’m not saying this is a good thing or that it is necessarily true, but it appears some people are confusing “this sucks for poor people” (which it does) with “this increases income inequality” (which it may not).

It’s important to distinguish gentrification from zoning inequality in this context. While the price rises with gentrification so does quality – and while that may not be at parity with the dollar for poorer people it is still important. Let’s stipulate that ridiculous rents in San Francisco are primarily a result of extreme zoning laws and red tape.Then there are two cases we must consider, of inequality in and around SF and inequality nationwide. Dexterous definition becomes important.

On the one hand, by raw income, inequality almost certainly decreases as people are priced out of the bottom but not the top. The second-order effects of gentrification might modify this somewhat but in any case would be a redistribution within the upper middle class. And whatever metric you use – top n%/bottom n%, gini, mean/median, etc. – removing the bottom 10% will reduce the inequality except perhaps in pathological cases.

And yet it is the case that most residents get some sort of consumer surplus from living in SF and adjust by paying higher rents, diminishing consumption of other items. There are also many moving parts here. Of first-order importance is the relative adjustment of the rich vs. poor. But there is also the question of which goods and services realize lower demand, and who that affects from both a demand and supply side.

There is the undeniable cost of dislocation but so long as housing regulations don’t become progressively harsher this is one-off and regardless not something economists should be concerned about. If having artists is important, the city can easily include an artist tax credit in its next budget. One would need to do a lot more work than simply acknowledge extreme rents to conclude that rents increase inequality within San Francisco.

That leaves the question of nationwide inequality. Surely the absence of affordable housing should increase inequality versus the counterfactual, right? Maybe. In effect pretty much everyone, including the rich tech family that’s paying $10,000 for a 2 b/r, suffers from high prices.

The only beneficiaries are those who have a net long position in property. The simplest way to explain this would be that if I am contractually sworn to be a taxi driver, I observe no increase in wealth if the price of my medallion increases since I am always net short – but I would if I lent an extra medallion to a friend.

It’s clear that poor people are definitely short, middle class people maybe flat, and a sliver of rich people long the housing market. Home ownership rate in SF is around 35%. That means about 65% of people are somewhere short, and of the 35% a whopping majority will be flat.

There is also the point that the “original short” of a tech employee is much higher than a hair stylist. To work at Facebook you need to live around the extremely expensive region. To work as a stylist you can work anywhere. A stylist obviously prefers not to move, but can do so with less penalty to earnings than a top software engineer.

There are other effects too. San Francisco is obviously a high-paying region for relatively unskilled, service jobs. More affordable housing would let more people indulge in this market. But it’s not clear which way the net effect runs, in particular home prices which command over 50% of the income of such individuals has a strong pass through into wage rate demanded.

Important considerations motivated by research Raj Chetty etc. are access to public transportation and commutability of a city. That is within-city inequality might be preferred to across-city inequality, and ceteris paribus more zoning laws encourage the latter.

It might well be the case that income inequality is increased by zoning regulations but the argument needs more evidence than dislocation and expensive rents. However without a doubt this sucks for poor people and zoning regulations should be abolished immediately on simply that account.



(Edit: I would be remiss not to reference other work critiquing Unz’s point from other angles – particularly from Nurit Baytch or Janet Mertz, both of whom have more thoroughly considered documents which similarly challenge Unz’s claims).

Several years ago political activist Ron Unz wrote a lengthy essay suggesting that elite universities have held Asian applicants to much higher academic standards than other groups, particularly Jews. He cites the disproportionate number of Jews at top universities relative to the population in general, and uses some questionable analysis of PSAT winners to conclude that elite universities are part of a vast Jewish conspiracy (not that you’ve ever heard this story before…):

Taken in combination, these trends all provide powerful evidence that over the last decade or more there has been a dramatic collapse in Jewish academic achievement, at least at the high end.

Unz reaches this conclusion based on looking for Jewish last names on various lists of high performers, especially National Merit Scholars, who seem to be overly Asian, at least by last name. Without disputing that this may be true – though there is significant bias given that one can with near certainty identify a name as Asian, but would have a relatively harder time achieving such power with Jewish names – it certainly isn’t a good way to measure bias.

There have been a number of criticisms of this article, based both on new evidence and identifying flaws in Unz’s article. The point of this post is to conclusively add to the repository of new evidence that Unz had no basis beyond racially-charged intuition to write his essay. In particular the point here is not to make any new claim based on data I’ve gathered as much as dispute everything that has been written on the subject so far (edit: I specifically mean in relation to writing in favor of this thesis – there has been plenty of writing that deals with other and similar flaws in Unz’s article much more throughly than mine.)

If the claim is that an admissions committee is biased towards Jews and therefore holds Asians to a relatively higher standard, it would follow that the academic profile of admitted Jews falls below other ethnic categories. The honest way to test this hypothesis would be to look at the academic performance of graduating students and test whether ethnicity is at all associated with academic performance.

Maybe Unz and others were too lazy to compile this data but, given what it reveals, they probably just didn’t like what it reveals. While universities don’t explicitly make available academic performance of accepted students by race, let alone ethnicity, this information can be inferred from public data.

Specifically, a number of universities publish a PDF of commencement programs, which usually contain the names of all graduating seniors, their department of study, and important academic honors they’ve received (Latin honors, prestigious scholarships, and the like).

The somewhat similar structure of these documents (for example, see the University of Pennsylvania or Princeton) makes it possible to parse the PDF files for the underlying information (it’s a somewhat messy task – send me an email if you want the code). Since these files are available from 2007 or so, with various omitted years for various universities, it’s possible to generate matched data of name, graduation year, major, school, and any academic awards received.

It is also not remarkably difficult to infer gender from the first name, and not catastrophically harder inferring ethnicity from the last. The method for identifying someone as various subcategories of Asian consisted of a simple cross reference to a large list of common last names (from the Census and other sources). The false positive rate here is extremely low (that is Varun Agarwal is extremely unlikely not to be Indian).

Identifying Jewish, specifically Ashkenazic, names is harder, as there is commonality with European, especially German, names. Further, plenty of Jewish-Americans have generic names that makes powerful identification challenging. However, for a given name, using an extremely large list of common Ashkenazic last names, and adding the score of the top 3 matches based on a Levenshtein distance fuzzy matching algorithm, it is possible to get some sense of probability that a name is Ashkenazic.

Since I don’t have verified data on the ethnicity of graduates by name, it’s difficult to test or train this strategy. That said this doesn’t really disturb the results. For one, this error is probably random – i.e. associations between ethnicity and achievement, if any, are unlikely to depend on the conviction with which an algorithm can determine ethnicity. To the extent there are attenuated coefficients, the bias will be in favor of Unz’s claim – i.e. odds appear to be lower than expected by a factor proportional to the ratio between classification error and total error. Furthermore, even if there is a strong association one way or the other between Jewish sounding names and odds of success, we propose that as an interesting finding in and of itself without making any further judgement.

For a flavor of the classification strategy, the table below charts 10 randomly sampled names at various thresholds. Clearly there is a link between strictness of threshold and what one might consider to be typical Jewish-American last names.

Screenshot 2015-12-18 19.26.10

Then to estimate the order of attenuated variables, you can compare the estimated coefficient on Jewish odds by threshold of an “Ashkenazi score” determined as outlined above:

Screenshot 2015-12-18 19.24.48

When including scores above 100 (which includes pretty much every non-Asian name, and some Asian names as well) the variable has little explanatory power if any – the measurement error dwarfs the residual error because this factor is true for pretty much every name. Clearly as we increase the threshold at which we admit a name as Jewish, the associated variable increases. This doesn’t imply anything about those who have more obviously Jewish surnames as much as give an order of magnitude estimate of attenuated coefficient bias, which is obviously significant.


It’s worth discussing what exactly we mean by academic achievement. To maintain comparability over time by school, Phi Beta Kappa and equivalent honor societies are used as proxies for underlying academic achievement, exactly the dimension along which Ron Unz seems to think Jews underperform. Phi Beta Kappa (and equivalents for engineering and business majors) is composed to students roughly in the top 10% of their graduating classes by GPA, with additional input from faculty recommendations.

There could easily be bias ingrained in the election process. That said, it is unlikely that this bias is somehow systematically focused against Asians in favor of Jews and, moreover, is moderated by the fact that the dominant theme is the GPA requirement. This does not settle the question, but if Ron Unz wants to argue that Jewish students with a 3.8 GPA are somehow academically non-performant relative to Asians with the same GPA, I would be interested in hearing that story indeed.

And here are the results based on regression on almost every Penn graduate over the past decade:

Screenshot 2015-12-17 23.28.20

The important point is that, controlling for graduation year, gender, school, and dual-degree status, the odds of being in an academic honor society increases by about 1.5x for those very likely to have Jewish last names.

I have been able to compile similar data for Brown and Princeton, where I notice similar trends (though no variables of significance for Princeton).

The point of this data – to be clear – is not to make a judgement on the academic achievement of various ethnic groups and therefore a claim on how admissions offices should operate. Other factors are probably much more important – however this data is a useful antidote to anti-scientific crusaders bent on ascribing one story of bias or the other that maligns large groups of people, without realizing that the underlying statistics on these questions are inherently complicated and do not yield easy interpretation.

Frankly, the set of vapid claims Ron Unz makes (and the New York Times editorial page approvingly cites) is predicated on bogus data. Not only does it fail to make any effort to identify Jewishness beyond perfect matching on to 100 lists, it fails to use data on actual achievement at school in favor of questionable data from PSAT scores.

Further, without even looking at a broad scope of data on graduate achievements – instead looking at meaninglessly small lists on Putnam exam winners or IMO olympiad results – Unz and everyone citing him favorably has decided to decree that an entire ethnic group has realized a “dramatic collapse in achievement”.

This is of course sloppy data work in general, but also evidence of complete disregard to any scholastic standards whatsoever before committing to a negative story about an entire category of students, which would be a cautionary tale if it was not so laughable in face of actual results.

Be careful when citing this post. You might notice that while Jews perform extremely well relative to baseline in every school where there is a significant variable at all, the coefficient for engineers is low (even though math and physics would be included in the first category where performance is fine). Ron Unz has an explanation for that:

We should also remember that Jewish intellectual performance tends to be quite skewed, being exceptionally strong in the verbal subcomponent, much lower in math, and completely mediocre in visuospatial ability; thus, a completely verbal-oriented test such as Wordsum would actually tend to exaggerate Jewish IQ.

He reaches this conclusion, of course, based on some colorless tale of ultra-religious orthodox Jewish reproduction patterns in urbanized environments (or something like that I didn’t actually bother to figure out what he meant). Unz might be banking on the hope that Jews are too mathematically illiterate to see through his lies.







Get every new post delivered to your Inbox.

Join 2,287 other followers