To: H-1B/L-1/offshoring e-newsletter Sun May 12 22:08:28 PDT 2013 H-1Bs are paid 20% less than comparable Americans. No, they make 6.7% MORE than Americans. No, scratch that, it's only 2.6% more. Wait a minute, no, no, a new Brookings study (my focus in this posting) finds that the H-1Bs actually are paid TWENTY percent (20%) more than comparable Americans. Confused? Allow me to help, with a tip from a former statistics professor (me): The key word aobve is COMPARABLE. Most of the above studies don't make things comparable at all. And not only does this statistics professor have strong views regarding H-1B, I have even stronger views about proper use of statistics. So, I want to use this posting as a case study on misuse of statistics. I apologize in advance if I come across as preachy, especially when I bring up bias issues at the end. Moreover, I want to stress here that journalists, policy makers and so on should NOT, upon seeing such a wide range of study results like those above, simply throw up their hands and say that the overall result is "inconclusive." On the contrary, as I'll explain below, we ALREADY KNOW conclusively that the H-1Bs are paid less than comparable Americans. Let me repeat: The question of underpayment of the H-1Bs HAS ALREADY BEEN SETTLED, with the answer being Yes. That first figure above, -20%, comes from my study, published a couple of months ago in the academic journal Migration Letters. That last figure above, +20%, is from a study released on Friday by the Brookings Institution. (One of the Brookings authors is a reader of this e-newsletter, and I may post an update after I receive responses to a query I sent him yesterday.) I will explain all this in common-sense terms. You won't need to be a statistician or economist, or even be good at math, to understand the flaws in the study. And after you read this, I hope you use it as a model from which to view future studies you see on H-1B. As a dentist once pointed out to me, an X-ray is "only a shadow," not as accurate and informative as physically looking inside a tooth. A dentist can't do the latter, of course, so he/she resorts to an X-ray. X-rays can indeed be highly useful--providing they are properly read. Or better yet, providing they are properly aimed; when I broke my leg in two places a few years ago (fall from a pullup bar, long story), they only found one of the two breaks, because they only aimed the camera in one place. Regression analysis is used (among other things) as a technique for making things COMPARABLE, the famous "ceteris paribus" term from economics--"all other factors being equal." I use it in my own research on H-1B (though not for my -20% figure above), and, I think more importantly, it is my main research specialization in the statistical methodology work I publish in statistics and computer science journals. So, I don't have a problem with regression analysis per se. As with X-ray shadows, it can be highly useful, IF properly read and properly aimed. Unfortunately, regression is the "X-ray shadow" of the statistics world, often poorly aimed, and even more poorly read. In this posting, I will show why the Brookings authors did not properly "aim and read" its regression analysis. But before doing so, it is extremely important to state that THE QUESTION HAS ALREADY BEEN ANSWERED. We have ALREADY "looked inside the tooth, rather than relying on a shadow" in terms of H-1B wages, and we ALREADY know they are underpaid. Here's why. Two employer surveys commissioned by Congress (which sadly doesn't know what its own commissioned reports say) actually ASKED employers whether they paid H-1Bs less than comparable Americans. The NRC report, in describing the employer survey the NRC asked Hal Salzman to do, said "...based on interviews with some H-1B employers, Salzman reported that H-1B workers in jobs requiring lower levels of IT skill received lower wages, less senior job titles, smaller signing bonuses, and smaller pay and compensation increases than would be typical for the work they actually did." (Note by the way that the NRC survey covered a wide range of employers, not just the Indian bodyshops.) The GAO report had similar findings, though with less smoking-gun-ish language. You can't get any better than that folks, straight from the horse's mouth! No shadows, no superstition. No trying to just indirectly INFER underpayment by drawing lines through data, which is what linear regression analysis does (hyperplanes instead of lines if we have multiple variables). Just plain, extremely straightforward querying employers, "Do you underpay your H-1Bs?", with the answer being Yes. In addition, as I've pointed out before, you can see the underpayment of H-1Bs even without data or statistics, as follows. The H-1Bs tend to be immobile (definitely the case for those being sponsored for a green card, and true to various extents for many others). If you can't move around freely in the labor market, then you can't negotiate the best salary deal among the employers. Thus on average, the H-1Bs will be paid less than what they would command in the market if they had their freedom. This is basic economic theory, not to mention common sense. So, again, we ALREADY KNOW THE H-1BS ARE UNDERPAID, relative to their market value. Knowing that, i.e. knowing that Brookings is trying to answer a question that has already been definitively answered, let's take a look at the Brookings study. I will primarily discuss Table 2 in their full report (http://papers.ssrn.com/abstract=2262872), a regression analysis. Let's begin by discussing Brookings' H-1B data itself. They claim that this data is representative, but it is certainly NOT representative. Why not? The Brookings data are a snapshot of what H-1Bs are paid AT THE TIME OF HIRE. But workers in general tend to get periodic raises in the years AFTER they are hired, right? Well, not if you are an H-1B. The NRC quote above spoke of the H-1Bs getting smaller raises than Americans, and H-1B advocacy groups claim it's common to receive no raises at all. Indeed, even industry lobbyists have complained that, in the case of H-1Bs being sponsored for green cards, the law actually forbids giving raises, because that would force the employer to open up recruitment again to Americans, who might take the job at the higher price. (This action is basically forced if, for instance, the H-1B is given a promotion.) How does that affect the Brookings analysis? The analysis says that salaries in general grow at the rate of 10% per year, and since the visa is good for 3 years, and can be renewed for 3 years, the average H-1B has held his visa for 1.5 years. Thus his salary should have increased 15% since the time his visa petition was filed. Suppose, as indicated above, that he has gotten no raises. He is then underpaid by 15%--but the Brookings data wouldn't show it, because that data only shows salary at the time of hire. Now let's look at Brookings' education variable, reporting highest degree earned. Frankly, any professional economist or statistician would be shocked by the Brookings analysis regarding education. Their data set's education variable codes a bachelor's degree with the value 5, 6 for a master's and 7 for a PhD. Brookings used these codes as a variable, and performed statistical operations on it, e.g. finding the mean. But these are simply arbitrary codes! Finding their mean is not much better than averaging a bunch of ZIP codes. The standard way to handle this situation is to make three variables from the original one: One variable for bachelor's degree, another for master's and a third for PhD. Each variable would have the values 0 and 1, the latter meaning yes. So, if a worker in the data has a master's, he would have a value of 1 for the master's variable and 0s for the other two. (A 0 value for all would mean the worker has less than a bachelor's degree.) True, Brookings' variable was at least ordinal, but it assumes that the wage difference between a bachelor's and a master's is the same as the difference between a master's and a PhD. That is obviously false. For example, in my recent EPI study, the data showed that for the CS field, master's was worth $6,703 more than a bachelor's, while in turn a PhD was worth $21,543 more than a master's. The Lofstrom and Hayes paper, cited prominently by the Brookings authors, has similar values. Since, as the industry lobbyists love to point out, an H-1B is much more likely to have a PhD than Americans, this produces a major bias in Brookings' findings. They are NOT comparing H-1Bs to comparable Americans, in this case, NOT comparing PhDs to PhDs. Next, consider region. The problem is that Brookings DIDN'T consider region--their analysis had no region variable. And that factor matters a LOT. Look for example at wages for the job category Software Developers, Applications in the OES data, compiled by the federal Bureau of Labor Statistics. The mean wage in the Chicago area is $89,990, compared to $116,610 in the San Jose region. Quite a spread! Brookings did have location information in its data, but did not use it, causing yet another major error, since the H-1Bs are disproportionately in the high-cost-of-living regions. There are numerous other serious flaws in the Brookings study. I may discuss some in later postings, but the ones I presented above are already quite enough to show that the authors' findings should not be taken at face value. I will close with the topic of Brookings' relation to Microsoft. I bring this up only with considerable reluctance, because some who read this will overinterpret it. I know one of the authors a bit, and I judge him to be a decent, ethical person. Yet the Microsoft relation must be stated and discussed. Brookings, like any not-for-profit organization, survives through contributions. And it is a fact of life that these organizations cannot afford to do many, if any, studies that are counter to the goals of the donors. But unlike many other organizations that have clear agendas, Brookings presents itself as ideologically neutral, with a motto of "Quality, Independence, Impact." Its Web address even sports a .edu domain. Independence. Nice concept, but how independent can Brookings be on tech issues when its annual report lists a contribution from the Bill and Melinda Gates Foundation in the $2,500,000-4,999,999 category? (This is the highest category.) There are also contributions from Microsoft Corporation, Google and so on. Is it any wonder that Microsoft people are usually invited as featured panelists in conferences that Brookings holds on the H-1B issue? I'll say, with considerable understatement, that it's impossible for a researchers to be impartial with Microsoft/Gates' millions hanging over their heads. Bias doesn't have to be as flagrant as fudging data; it can be more subtle, such as telling only part of the story. The full paper for the Brookings study has NO papers in its bibliography that are directly negative about H-1B. They do have the recent EPI paper by Salzman et al showing a lack of a STEM labor shortage but nothing negative about H-1B. They have NO citations to Ron Hira's work, they don't cite the NRC and GAO employer surveys, they have nothing about the internal NSF memo arguing the U.S. should bring in foreign students to hold down STEM salaries (which then did occur), etc. Granted, Brookings' Web summary of the Brookings study, at http://www.brookings.edu/research/papers/2013/05/10-h1b-visas-stem-rothwell-ruiiiz links to my EPI paper. But there they describe that paper as being about cheap labor, which it is definitely not (it addresses the issue of whether the H-1Bs tend to be "the best and the brightest"). I have to assume the Brookings authors have never bothered to read it. This lack of citing countering views is actually a hallmark of the research literature that is positive about H-1Bs. This is not what research ought to be like, especially by those with a .edu in their Internet addresses. Norm Archived at http://heather.cs.ucdavis.edu/Archive/BrookingsFlawedStudy.txt