Printer Friendly

IQ in the production function: evidence from immigrant earnings.

The cross-country growth literature (especially Sala-i-Martin 1997; Sala-i-Martin, Doppelhofer, and Miller 2004) has found that traditional education measures rarely have a robust relationship with growth and productivity--elementary education being a rare exception. By contrast, a new empirical growth literature (Jones and Schneider 2006; Lynn and Vanhanen 2002; Ram 2007; Weede 2004; Weede and Kampf 2002) has shown that a nation's average IQ has a remarkably robust relationship with its productivity (Figure 1). That a test designed by psychologists should have such a robust relationship with economic variables is a puzzle that demands explanation.

For instance, Jones and Schneider showed that national average IQ was more robust than other human capital variables and was statistically significant at the 1% level in all 455 growth regressions that controlled for all 18 factors that passed Sala-i-Martin, Doppelhofer, and Miller's (2004) robustness tests. Of course, as with all the growth regression literature, a key difficulty is disentangling cause and effect. Thus, in this article, we run no growth regressions whatsoever. Instead, we perform a simple calibration of the IQ-productivity relationship based on widely agreed upon microeconomic parameters. That means we can directly estimate one causal channel running from cognitive ability to productivity. In the process, we learn the following:

(1) If one knows the average IQ of a nation's citizens as estimated by Lynn and Vanhanen (2002, 2006), one can predict the average wages that immigrants from that country will earn upon their arrival in the United States--whether or not one controls for immigrant education and even if the test is completely visual rather than verbal. In other words, national average IQ predicts part of what Hendricks (2002) calls "unmeasured worker skill."

(2) We find that a 1-point increase in national average IQ predicts 1% higher immigrant wages--precisely the value found repeatedly in microeconometric studies (note that by construction, 1 IQ point [approximately equal to] 1/15th of a standard deviation within any large national population). Together, Points 1 and 2 provide further evidence that cross-country IQ tests are valid predictors of worker productivity.

(3) When IQ is added to the production function in the form implied by traditional, externality-free human capital theory, differences in national average IQ are quantitatively significant in explaining cross-country income differences. That said, our productivity accounting exercise does not resolve the puzzle of why high-IQ countries are 15 times richer than low-IQ countries.

In a related vein, Hanushek and Kimko (2000) use national math and science test scores to verify that cognitive skills appear to matter more for groups than for individuals. Like us, they use immigrants to the United States as a way to test whether immigrants bring their home country productivity levels along with them when they immigrate to the United States. When they interpret their results within a Solow-type framework, they conclude that "the [cross-country] growth equation results are much larger than the corresponding results for individual earnings."

In sum, our article rigorously explores the quantitative magnitude of the puzzle uncovered by Hanushek and Kimko (2000) and Hanushek and Woessman (2007). But by using IQ tests rather than other widely used math and science test scores, we can often double our sample size while simultaneously using the most widely analyzed, best understood form of cognitive test.

We begin with an overview of the recent psychological literature on the validity of IQ tests and then proceed to our discussion of the link between IQ and immigrant wages. The discussion of IQ and immigrant wages yields a key parameter, [gamma], the IQ semi-elasticity of wages, which we use in our development accounting exercise. We then discuss the questions of reverse causality and trends in the IQ-productivity relationship over the past 40 yr and conclude by discussing how our results fit into the growth literature.


It is not possible to have confidence in these or any other IQ-related findings without an adequate understanding of how IQ is measured and why psychologists believe that well-constructed IQ tests are legitimate tools for the study of cognitive abilities. Unfortunately, space does not permit a comprehensive review of the large research literature that adequately addresses the many reasonable doubts and concerns a properly skeptical reader might have about the validity of IQ tests. Readers wishing for scholarly, balanced, and accessible introductions to intelligence research are advised to consult Bartholomew (2004), Cianciolo and Sternberg (2004), Deary (2001), or Seligman (1992).

Considerable effort has gone into producing nonverbal IQ tests that can be used in any culture. These "culture-fair" or "culture-reduced" IQ tests have been shown to predict important life outcomes with validity coefficients comparable to traditional IQ tests designed for specific populations (Court 1991; Kendall, Verster, and Von Mollendorf 1988; Rushton, Skuy, and Bons 2004). As we note below, the correlation between national average IQ and gross domestic product (GDP) per worker is essentially unchanged if we only use data from such culture-reduced tests. Unlike traditional IQ tests that measure a very diverse set of cognitive abilities, culture-reduced IQ tests necessarily measure a much smaller number of abilities, focusing on nonverbal reasoning and novel problem solving. Fortunately, the types of tests that lend themselves to cross-cultural research correlate very highly with the overall scores from traditional IQ tests (Jensen 1998). For our purposes, it does not matter if one believes that IQ tests are valid measures of whatever "real intelligence" is (if there is such a thing as "intelligence"). The tests measure a set of skills that appear to be very advantageous in societies with modern economies. Unlike other measures of human capital such as reading comprehension and mathematical reasoning tests, culture-fair IQ tests have no literacy prerequisites. Because the tests are nonverbal, the test items are the same for everyone, and thus, results are more comparable across language groups and cultures.

We do not conceptualize culture-fair IQ tests as measures of some immutable quantity that is solely determined by genes. Although it is quite clear that genes play an important role in the development of cognitive abilities, it is equally clear that cognitive abilities are quite sensitive to environmental inputs and can change considerably over the lifespan (Shaie 2005). It is relatively easy to disrupt the delicate processes of the brain with disease, malnutrition, parental abuse and neglect, environmental toxins, and brain injury. With considerable effort, it is also possible to raise IQ somewhat with high-quality personal health care, sound public health policies, adequate nutrition, reasonable parental involvement, and excellent education (Armor 2003). The fact that IQ scores have been rising 0.2 standard deviations per decade in most countries ever since mass IQ testing started in the 1920s (Dickens and Flynn 2001; Flynn 1987; Neisser 1988) suggests that in many societies, people have increased access to some of these things.


In this section, we investigate whether the average IQ in the immigrant's home country is a useful predictor of the wages of immigrants from that country. Our estimates of immigrant wages come from Hendricks (2002), who used data on earnings, education, and age from 106,263 immigrants from the 1990 Census of Population and Housing. (1) These immigrants were between the ages of 20 and 69 and worked full time in the United States and had immigrated as adults. For further information on the immigrant data, see Section II and especially table B1 of Hendricks (2002).

Hendricks extracted systematic wage differences due to education and age by comparing weighted averages of the earnings of native-born and immigrant workers. He did this by creating ten age categories and six education categories for each country's immigrants as well as for U.S. natives. The average immigrant wage per source country was weighted according to the U.S. distribution of education and age levels. Thus, countries whose emigrants have a low (high) average education level would have the wages of their highly educated emigrants overweighted (underweighted). For example, immigrants from Taiwan have an average of 15.9 yr of education (above the average of native U.S. workers), so the adjustment process would downweight the earnings of Taiwan's highest educated immigrants, putting more weight on the earnings of those with less than a high school education.

After thus controlling for age and education, Hendricks concludes that the only remaining explanation for wage differences between workers from different countries is what he calls differences in unmeasured worker skill. (2) Hendricks created estimates of unmeasured worker skill for 76 countries.

Perhaps surprisingly, this unmeasured worker skill estimate varies widely for immigrants from different countries. The standard deviation of log unadjusted immigrant wages is 0.29 across Hendricks's sample of 76 countries, while the standard deviation of log unmeasured worker skill across these countries is still a sizable 0.19. Henceforth, we refer to [uws.sub.i], the log of unmeasured worker skill in country i.

Our goal in this section was to show that national average IQ is a useful predictor of Hendricks's unmeasured worker skill. We use Lynn and Vanhanen's (2006) database of national average IQ estimates. Appendices 1 and 2 provide the raw country-level IQ estimates and some tests of the reliability of these IQ estimates, respectively. We should briefly review how Lynn and Vanhanen (henceforth LV) (2006) created their data set: they used hundreds of IQ tests from 113 countries across the 20th and 21st centuries in LV (2006). They aggregated these results using best practice methods to create estimates of "national average IQ" for these countries. (3) LV show that the IQ gaps between regions of the world have not appreciably changed during the 20th century.

LV's (2006) data set overlaps with 59 of Hendricks's observations. The mean and median IQ across these 59 countries are both 91 and the standard deviation of IQ across these countries is 9. This is a slightly more intelligent, less varied sample than the full 113 countries: LV's full-sample mean and median are both 87 and the standard deviation is 12. For comparison, we note that within the United Kingdom, mean IQ is defined as equal to 100, and the standard deviation of IQ within the UK population is defined as equal to 15.

Data in hand, we regress log [uws.sub.i] onto the level of national average IQ and a constant. The goal was to see whether the estimated relationship between immigrant wages and national average IQ is close to conventional microeconometric estimates of the IQ-wage relationship. In a variety of previous studies, the semi-elasticity of wages (denoted [gamma]) has been close to 1%: thus, 1 IQ point is associated with 1% higher wages, and a one standard deviation rise in IQ is associated with about a 1% rise in wages. (4) The semi-elasticity [gamma] has a similar magnitude whether one measures in developing countries or in the United States. Perhaps surprisingly, Zax and Rees (2002) find that [gamma] appears to rise later in life--so childhood IQ predicts one's wage better as one gets older--while the coefficient rises only by about one-third when one controls for education in a typical Mincer wage regression.

Now, let us return to our main question. Do our 59 observations roughly replicate these intracountry estimates of the IQ-wage relationship, where 1 IQ point predicts about 1% higher wages? Yes, they do, as seen in Figure 2 and Table 1. When we run a simple bivariate correlation between [uws.sub.i] and national average IQ, we find a correlation of +.47, and ordinary least squares (OLS) yields a regression coefficient of [gamma] = 0.95 (White standard error = 0.31). This is remarkably close to the coefficient estimates seen elsewhere.

Our estimate, which we round to unity, provides a number of insights. First, it shows that LV's national average IQ measures are useful for predicting more than just cross-country productivity differences, cross-country growth rates (both positive correlations), cross-country suicide rates (also a positive correlation [sic]: Voracek 2004, 2005), and other cross-country factors. We have now shown that they are also useful for predicting the age- and education-adjusted wages of the average immigrant coming from her home country to the United States. (5) This is surely evidence that national average IQ is an important predictor of what Hanushek and Kimko (2000) call "labor quality."

Further, we have shown that the estimate is quite close to conventional microeconometric estimates of the IQ-wage relationship. (6) Whatever an IQ test can tell us about worker wages, it appears to be measuring the same thing across countries as within countries. This is confirmatory evidence that cross-country IQ comparisons are indeed possible, despite the claims of many (e.g., Diamond 1999; Ehrlich 2000) to the contrary.




As mentioned above, Zax and Rees (2002) note that controlling for education may bias the [gamma] coefficient downward. After all, IQ is quite likely to have an impact on the quantity of future education a student acquires, so some of the estimated effect of education on earnings is likely to represent IQ's indirect impact on earnings. As a practical solution, they recommend a simple regression of earnings on IQ alone.

In our case, the equivalent regression would involve regressing Hendricks's "log unadjusted earnings" on IQ. This regression is then the average wage of all Mexicans or all Canadians or all Italians working in the United States, regressed on the average IQ in that country. This will provide us with an upper bound for IQ's impact on immigrant earnings. In such a regression (Table 2), the correlation coefficient is +.42, with an OLS regression coefficient of [gamma] = 1.3 (White standard error, 0.44). This is quite close to the upper bound of current estimates found in microlevel panel and cross-sectional studies and is only 30% larger than our baseline estimate of [gamma] = 1.

Further, our original [uws.sub.i] results do not appear to be sensitive to obvious outliers. There are three obvious outliers and all three tend to push [gamma] downward: high-wage South African immigrants (IQ = 72) and low-wage Chinese (IQ = 105) and South Korean immigrants (IQ = 106); they are the only three with regression residuals more then 2.5 standard deviations away from zero, and all three are in fact over 4 standard deviations away from zero. Thus, they are not small outliers. But are they driving our results from the previous section? It would appear not. One at a time omission of these outliers has a negligible impact on the [gamma] estimate, and eliminating all three raises the coefficient to just 1.4, at the high end of microeconometric estimates.

Another way to check for outliers would be to include dummies for regions of the world that appear to be econometrically "special." Sala-i-Martin (1997) and Sala-i-Martin, Doppelhofer, and Miller (2004) found that geographic dummies for East Asia, Latin America, and Sub-Saharan Africa were robust across millions of growth regressions. At the present state of knowledge, it is difficult to know just what these dummies are proxying for; it could be geography, culture, genetics, natural resource availability, persistent political institutions, or many other factors. When we include dummies for these regions (Table 2), we find that our result actually becomes slightly more robust. Thus, these data provide little evidence that a few special regions of the world are driving this result.

Overall, our results appear to be robust to endogenous education and to outliers. In our development accounting exercises below, we investigate the implications of imposing various [gamma] values. We tentatively conclude that cross-country IQ measures, as aggregated by LV, are a useful indicator of the private marginal productivity of workers. Cross-country IQ scores pass this "market test" with little difficulty, a result that strengthens our confidence in the validity of cross-country IQ tests as indices of one form of labor quality.


We now turn to the question of whether IQ's impact on the private marginal product of labor can explain the massive differences in living standards we see across countries. We begin by assuming an IQ-augmented Cobb-Douglas production function,


The subscript i is the country subscript; Y, K, A, and L are output, the capital stock, disembodied technology, and the labor supply, respectively; and [gamma] is the semi-elasticity of wages with respect to IQ. In other words, [gamma] is the impact of IQ on human capital. Since our concern is with cross-country comparisons, we suppress time subscripts. We reorganize the production function to make it amenable to development accounting:


This is the equation we use (sometimes in log form) to evaluate the impact of IQ differences on steady-state living standards. Writing in terms of a capital-output ratio is useful since in a Solow or Ramsey growth model, the economy heads to a steady-state capital-output ratio that is independent of the level of technology (or by extension, the level of IQ). IQ appears in the production function just as any other form of human capital would. As such, we can estimate IQ's impact on output in the same way that economists estimate education's impact on output: by looking at microeconometric estimates of the link between wages and this form of human capital. Thus, we will repeatedly reuse our [gamma] = 1 estimate from Section III, but will also consider [gamma] = 1.25 as an upper bound and [gamma] = 0.5 as a lower bound. Bowles, Gintis, and Osborne (2001), in a metastudy of the labor literature, find a median estimate of [gamma] = 0.5; their meta-study includes all possible studies, without regard to econometric technique.

Before we do so, let us briefly review the power of national average IQ to predict national productivity. LV (2002) found a correlation of .7 between national average IQ and the level of GDP per worker in 81 countries. Jones and Schneider (2006) found a correlation of .82 between national average IQ and log GDP per worker and also found that national average IQ was statistically significant at the 1% level in 455 cross-country growth regressions that used all the Sala-i-Martin, Doppelhofer, and Miller's (2004) robust growth variables as controls. (7)

In the results below, GDP per worker estimates are from the Penn World Tables. In total, we have complete data for 87 countries that are broadly representative of the world's economies. Data and software are available upon request, and the raw data underlying LV's IQ estimates are readily available in table form on the Web (Sailer 2004). The Sailer Web site's charts are especially useful for demonstrating that these IQ differences have been persistent and do not turn on the type of IQ test employed.

A. IQ Differences: Magnitude

In this section, we combine the IQ-augmented production function (2) with conventional parameter values for [gamma] to illustrate how IQ differences can impact steady-state living standards. Consider two countries that differ only in average IQ--that is, their levels of technology and their capital-output ratios are equal across countries. The ratio of living standards in these two countries would then be:

(3) [(Y/L).SUB.hi]/(Y/L).sub.lo] = [e.sup.[gamma][DELTA]IQ],

where [DELTA]IQ is the difference in IQ between the two countries. LV (2006) show that if countries are ranked according to IQ then the country in the 5th percentile has an estimated average IQ of 66, while the country in the 95th percentile has a median IQ of 104. This yields an IQ gap of 38 points--a bit more than two standard deviations if one were looking within the U.S. population. As noted above, we take [gamma] = 1 as our preferred estimate; under this assumption, a rise of 1 IQ point raises wages (and hence the marginal product of labor) by a modest 1%.

Therefore, as Figure 3 illustrates, if a country moved from the middle of the bottom IQ decile to the middle of the top IQ decile (a rise of 38 points), steady-state living standards would be about 1.5 times greater in the higher IQ country ([e.sup.0.33] [approximately equal to] 1.46). This compares to the factor of 2 commonly cited for the impact of cross-country differences in education on productivity--some of which may in fact reflect differences in IQ endogenously driving education choices. If the true [gamma] were equal to 1.25, toward the high end of current estimates, a 38 IQ point gap would raise living standards by a multiple of 1.64. And if [gamma] were half of our preferred estimate, as denoted in the lowest of the three lines, a 38-point IQ gap would cause living standards to diverge by a factor of 1.23.

But perhaps the 5th and 95th percentiles are outliers, driven by test error or idiosyncratic environmental factors. Therefore, we look at the 90:10 and 80:20 ratios. The gap between the 90th and 10th percentiles is 31 IQ points (102 and 71 points), while the gap between the 80th and the 20th percentiles is 21 IQ points (99 and 78 points). In these cases, productivity levels between these countries in the [gamma] = 1 case would differ by a bit more than 30% and a bit more than 20%, respectively.


Since living standards across countries differ by perhaps a factor of 30, and since the natural log of 30 is about 3.4, then if [gamma] = 1, the channel running from national average IQ to the private marginal product of labor explains perhaps 0.46/3.4 or a bit less than one-sixth of the log difference in living standards across countries.

We should note that these development accounting results do not depend on IQ being exogenous. We suggest below that simple reverse causality (running from productivity to IQ) is unlikely to be the main explanation for the strong empirical IQ-productivity relationship. However, even if reverse causality were important, the development accounting results would still hold since microeconomic studies demonstrate convincingly that IQ has an independent impact on the marginal product of labor.

B. Calibration Results

The calibration exercise is quite straightforward and is similar to that of Dhont and Heylen (2008). For each country in the data set, we predict the level of GDP per worker using Equation (2), assuming that technology and capital per worker are identical across countries. Thus, the only source of cross-country productivity differences is IQ working through the narrow channel of the private marginal product of labor, [gamma]. This gives us predicted values for 87 countries, which we then compare to the actual values of GDP per worker. We compare this prediction against the actual level of GDP per worker for each country.

As a goodness-of-fit measure, we use [R.sup.2] This [R.sup.2] is the percentage of the global income distribution that can be explained through a single channel: the steady-state impact of differences in national average IQ on labor productivity by way of the private marginal product of labor, [gamma]. For reference, note that the [R.sup.2] between log GDP per worker in 2000 and LV's (2006) national average IQ estimate is 58%, and in a cross-country OLS regression, 1 IQ point is associated with 6.7% higher GDP per worker (Figure 1).

Results are reported in Table 2. (8,9) For the preferred parameter value of [gamma] = 1.0, IQ can explain 16% of log cross-country income variation. Therefore, IQ's impact on wages would explain 29% (i.e., 16%/58%) of the relationship between IQ and log productivity.

If, instead, IQ has a 25% larger impact on wages ([gamma] = 1.25) then IQ's effect on wages can explain 20% of the variance in log productivity and 34% (=20%/58%) of the IQ/log productivity relationship. And even if [gamma] = 0.5--half of our preferred estimate--IQ's impact on wages explains 8% of the log global income distribution. So even under unusually conservative assumptions, IQ's impact on the private marginal product of labor appears to belong on any top 20 list of explanations for cross-country income differences.


The quantitative results of the last two sections imply that differences in national average IQ are substantial drivers of global income inequality. Can simple reverse causality explain this relationship? In other words, does a dramatic rise in GDP per worker cause a dramatic rise in national average IQ?

The region of the world that has witnessed the most rapid increases in living standards the world has ever known is unambiguously East Asia. Surely, this region would be an ideal testing ground for the productivity causes IQ hypothesis. If most of the IQ-productivity relationship was reverse causality then we would expect to see the East Asian economies starting off with low IQs in the middle of the 20th century, IQs that would rapidly rise in later decades, perhaps even converging to European IQ levels. In short, one would expect to see Solow-type convergence in national average IQ.

However, this is not the case. LV's (2006) country-level IQ data show that average East Asian IQs were never estimated below 100 before the 1980s (Figure 4). These IQ scores come from South Korea, Japan, Hong Kong, China, and an East Asian offshoot, Singapore. In all cases, IQ scores are above 100--even in a poor country like China. Thus, East Asians both started and ended the period with highIQ scores.

Another place to look for massive IQ increases would be in a region of the world that experienced a dramatic increase in the price of its exports: the oil-rich countries of the Middle East. But a glance at that data, likewise, shows little evidence that being richer, per se, increases IQ within 10 or 20 yr:
Year IQ Country

1957 77 Egypt
1957 82 Lebanon
1959 84 Iran *
1972 81 Egypt
1972 83 Iran *
1972 87 Iraq *
1972 87 Iraq *
1987 80 Iran *
1987 84 Jordan
1987 78 Qatar *
1989 83 Egypt
1992 89 Iran *
1997 85 Yemen
2005 86 Kuwait *

Note: Asterisk indicates Organization of Petroleum
Exporting Countries member.

If one uses 1973 as a breakpoint--since real oil prices increased fourfold between 1973 and 1986 before declining--then one would expect IQ scores to be higher in oil-rich countries if simple reverse causality drove IQ scores. Casual inspection of the evidence does not show such a relationship--indeed, Qatar and Kuwait, two low-population, high-GDP per capita countries, fail to stand out along the IQ dimension.

Further, after 1973, there is no clear difference between OPEC and non-OPEC countries, contrary to what one would expect if income caused IQ in an important way. Finally, a simple difference-in-difference test shows that OPEC countries have a median IQ score falling 5.5 points lower compared to non-OPEC countries after 1973 (given the small sample size, we will refrain from calculating standard errors--consider these results as suggestive). All told, if one wants to use a reverse causation argument to explain the IQ-productivity relationship, it will have to be more subtle than the simple tests of East and Southwest Asian IQs presented here.



Hendricks (2002) showed convincingly that workers from different countries have different average levels of what he calls unmeasured worker skill. We have provided evidence that conventional, out-of-the-box IQ tests can measure an important part of that heretofore unmeasured skill. This supports the claims of LV (2002, 2006) that national average IQ is an important determinant of economic outcomes across countries.

We have further shown that the between-country coefficient of the IQ semi-elasticity of wages, [gamma], is essentially identical to the within country coefficient, and we have used that fact to conduct a conventional, externality-free development accounting exercise. In such an exercise, we found IQ's impact on productivity to be quantitatively modest: it explains about one-sixth of the variance in log productivity between countries and about one-sixth of the predicted steady-state relationship between IQ and log productivity.

To put this in perspective, note that if a nation moved from the 5th to the 95th percentile of national average IQ, our development accounting exercise predicts that its output per worker would rise by perhaps 50%. But in reality, these countries have living standards that differ by a factor of 15, not 1.5. We hope that future research investigates why these relatively modest IQ differences between countries are correlated with such massive differences in national living standards.

We also hope that economists can bring their powerful econometric and theoretical tools to bear on the question of why IQ gaps across poor countries are so large. If economists can find ways to narrow these persistent IQ gaps, the world's poorest citizens may be able to make full use of their productive potential.

IQ and Earnings Data

 National Log Adjusted
Country Average IQ Earnings

Argentina 93 4.63
Australia 98 4.88
Austria 100 4.84
Barbados 80 4.56
Belgium 99 4.84
Bolivia 87 4.36
Brazil 87 4.54
Canada 99 4.83
Chile 90 4.51
China 105 4.35
Colombia 84 4.43
Denmark 98 4.88
Dominican Republic 82 4.37
Ecuador 88 4.41
Egypt 81 4.54
Fiji 85 4.40
France 98 4.84
Germany 99 4.76
Ghana 71 4.25
Greece 92 4.63
Guatemala 79 4.33
Honduras 81 4.29
Hong Kong 108 4.59
Hungary 98 4.61
India 82 4.58
Indonesia 87 4.57
Iran 84 4.51
Iraq 87 4.48
Ireland 92 4.78
Israel 95 4.70
Italy 102 4.78
Jamaica 71 4.50
Japan 105 4.92
Jordan 84 4.51
Kenya 72 4.60
Malaysia 92 4.54
Mexico 88 4.34
Netherlands 100 4.70
New Zealand 99 4.84
Norway 100 4.88
Pakistan 84 4.41
Peru 85 4.35
Philippines 86 4.34
Poland 99 4.53
Portugal 95 4.70
South Africa 72 4.91
South Korea 106 4.35
Spain 98 4.66
Sri Lanka 79 4.61
Sweden 99 4.86
Switzerland 101 4.88
Syria 83 4.67
Taiwan 105 4.60
Thailand 91 4.42
Turkey 90 4.67
United Kingdom 100 4.87
Uruguay 96 4.57
Venezuela 84 4.49
Yugoslavia 89 4.71

Notes: National Average IQ data are from LV (2006).
Adjusted earnings data are from table B1 of Hendricks
(2002) and draw on the 1990 U.S. census.


Since the LV (2002, 2006) IQ scores have been used in only a few papers in the economics literature, some effort to measure the reliability of their database is warranted. In LV (2006), more than 300 IQ tests from 113 nations are used. Their database combines many types of IQ tests from the purely visual, multiple-choice Raven's Progressive Matrices to the 3-hr long Wechsler, which is always given one-on-one by a professional psychometrician. When LV have multiple IQ estimates for a country, they choose the median score. LV's data come from a variety of sources, but the two most important categories are "standardization samples" and individual published studies. Standardization samples are typically created by publishers of IQ tests to learn about the first, second, and higher moments of the distribution of IQ scores within a particular national population. By doing so, they can convert a raw test score into a percentile ranking within a national population.

For example, Angelini et al. (1988) gave more than 3,000 Brazilian children the Raven's Colored Progressive Matrices test (a simple multiple-choice visual IQ test, but a powerful predictor of overall IQ). In creating a standardization sample, psychometricians attempt to create a genuinely representative sample, so that their product--purchased by school districts around the world--will build a good reputation and find many customers. In LV (2006), 25 countries have at least one score from a standardization sample. In LV (2006), most standardization samples are from Raven's-type tests.

The other individual published studies tend to come from "opportunity samples," perhaps a classroom or a school district near the researchers. Some such studies are simply an attempt to document how typical children perform on one type of IQ test, while other studies look into how nutrition, level of schooling, environmental lead, or other forces impact an individual child's IQ. An important question is how the "best" studies--the standardization samples--compare to the "rest." Within a country, are the standardization scores similar to the individual study scores? And are the standardization scores similar to LV's country-level estimated national average IQ?

The answer to both questions appears to be "yes." We assembled data from countries that had at least one standardization sample IQ estimate plus at least one individual published study IQ estimate. In the "standardization" category, we also include five Latin American estimates from a UNESCO (1998) IQ test of verbal reasoning; each country had a sample of 4,000 students. Omitting these observations had no noticeable impact on the results below. When a country had more than one standardization sample (common in rich countries plus India), we took the median standardization sample IQ score and compared it pairwise against the other lower quality IQ scores.

We arrive at a total sample of 23 countries and 63 comparisons. The mean absolute deviation between a country's median standardization score and that country's other lower quality scores is 3.2 IQ points, one-fifth of a standard deviation within the United States (the standard deviation is 4.4 IQ points). Therefore, it will be rare for a lower quality IQ score to be off by 15 points, a full U.S. standard deviation. The mean deviation is -0.2 IQ points, so the lower quality scores display negligible bias, with standardization scores ever so slightly lower than other scores. The correlation between high-quality and lower quality IQ scores is .9, but since this sample is weighted toward the higher IQ countries where there is little variation in IQ scores, a correction for restriction of range would raise this correlation even higher.

Since both high-quality and lower quality samples appear to tell roughly the same story about a country's IQ, there is little to be gained from painstakingly creating a standardization sample for every country: "the best" differ little from "the rest." This finding shows up in LV's estimated national average IQ: the mean absolute deviation between the median standardization score and LV's national average IQ is a negligible 1.1 IQ points (standard deviation 1.9 IQ points), while the mean deviation is 0.1 IQ points. So for countries where we have standardization scores for comparison, LV's method of choosing the median IQ score is quite similar to choosing the highest quality score.

Another important question is how the Flynn effect impacts these scores: might the Flynn correction introduce some bias? All scores used thus far in this article employ only Flynn-adjusted IQ scores. LV's Flynn adjustment uses 1979 as a base year and following the best practice in the literature assumes that scores on the Raven's increase by 3 IQ points per decade and increase by 2 points per decade on all other IQ tests. LV (2002) report both Flynn-adjusted IQ scores and raw IQ scores: using that data, we find that the correlation between LV's national average IQ and year 1998 log GDP per worker (Heston et al. 2002) is .83 with unadjusted scores and .85 with Flynn-adjusted scores, a minor difference.

Indeed, part of the reason Flynn adjustments cannot matter much is because both poor and rich countries have IQ scores going back many decades so on average, the Flynn adjustment impacts all types of countries about equally. Therefore, even if the Flynn adjustments are incorrect, they combine an irrelevant shift in intercept with a shift in slope. A mere glance at the data sets in LV (2002) will be enough to convince many readers that Flynn corrections are unlikely to be relevant; Sailer (2004) has put the LV (2002) database into a convenient online tabular format.

Finally, there is the question of whether the quality of the IQ scores impacts the immigrant wage results reported in this article. Apparently, the answer is no: when we use only standardization sample and UNESCO (1998) scores, we have observations for a mere 21 countries, but an OLS regression finds that 1 IQ point predicts 1.2% higher immigrant wages (p = .02, corr = .5), similar to the results reported using the full data sample. Similarly, at the cross-country level, one "high-quality" IQ point predicts 8% higher national GDP per worker (p = .0001, corr = .7, n = 27). So even using high-quality national average IQ estimates, IQ predicts small within-country productivity differences but large cross-country productivity differences. This replicates the central finding of this article.


GDP: Gross Domestic Product

OLS: Ordinary Least Squares

doi: 10.1111/j.1465-7295.2008.00206.x


Angelini, A. L., I. C. Alves, E. M. Custodio, and W. F. Duarte. Manual Matrizes Progressivas Coloridas. Sao Paolo, Brazil: Casa do Psicologo, 1988.

Armor, D. J. Maximizing Intelligence. New Brunswick, N J: Transaction Publishers, 2003.

Bartholomew, D. J. Measuring Intelligence: Facts and Fallacies. Cambridge: Cambridge University Press, 2004.

Behrman, J., H. Alderman, and J. Hoddinott. "Copenhagen Consensus--Challenges and Opportunities: Hunger and Malnutrition." 2004. Accessed November 6, 2008. asp?ID=223

Bishop, J. H. "Is the Test Score Decline Responsible for the Productivity Growth Decline?" American Economic Review, 79, 1989, 178-97.

Bowles, S., H. Gintis, and M. Osborne. "The Determinants of Earnings: Skills, Preferences, and Schooling." Journal of Economic Literature, 39, 2001, 1137-76.

Cawley, J., K. Conneely, J. Heckman, and E. Vytlacil. "Cognitive Ability, Wages, and Meritocracy," in Intelligence, Genes, and Success: Scientists Respond to The Bell Curve, edited by B. Devlin, S. E. Fienber, D. P. Resnick, and K. Roeder. New York: SpringerVerlag, 1997, 179-92.

Cianciolo, A. T., and R. J. Sternberg. Intelligence: A Brief History. Malden, MA: Blackwell Publishing, 2004.

Court, J. H. "Asian Applications of Raven's Progressive Matrices." Psychologia, 34, 1991, 75-85.

Deary, I. J. Intelligence: A Very Short Introduction. Oxford: Oxford University Press, 2001.

Dhont, T., and F. Heylen. "Why Do Europeans Work (Much) Less? It Is Taxes and Government Spending." Economic Inquiry, 46, 2008, 197-207.

Diamond, J. Guns, Germs, and Steel: The Fates of Human Societies. New York: W.W. Norton, 1999.

Dickens, W. T., and J. R. Flynn. "Heritability Estimates versus Large Environmental Effects: The IQ Paradox Resolved." Psychological Review, 108, 2001, 346-69.

Ehrlich, P. Human Natures: Genes, Culture, and the Human Prospect. Washington, DC: Island Press, 2000.

Flynn, J. R. "Massive IQ Gains in 14 Nations." Psychological Bulletin, 101, 1987, 171-91.

Gould, S. J. The Mismeasure of Man. New York: W.W. Norton, 1981.

Hanushek, E., and D. Kimko. "Schooling, Labor Force Quality, and the Growth of Nations." American Economic Review, 90, 2000, 1184-208.

Hanushek, E., and L. Woessman. "The Role of School Improvement in Economic Development." NBER Working Paper No. 12832, 2007.

Hendricks, L. "How Important is Human Capital for Economic Development? Evidence from Immigrant Earnings." American Economic Review, 92, 2002, 198-219.

Heston, A., R. Summers, and B. Aten. Penn Worm Table Version 6.1. Center for International Comparisons at the University of Pennsylvania, 2002.

Jensen, A. R. The g Factor: The Science of Mental Ability. Westport, CT: Praeger, 1998.

Jones, G., and W. J. Schneider. "Intelligence, Human Capital, and a Bayesian Averaging of Classical Estimates (BACE) Approach." Journal of Economic Growth, 11, 2006, 71-93.

Kendall, I. M., M. A. Verster, and J. W. Von Mollendorf. "Test Performance of Blacks in Southern Africa," in Human Abilities in Cultural Context, edited by S. H. Irvine and J. W. Berry. Cambridge: Cambridge University Press, 1988, 299-339.

Lynn, R., and T. Vanhanen. IQ and the Wealth of Nations. Westport, CT: Praeger Publishers, 2002.

--. IQ and Global Inequality. Augusta, GA: Washington Summit Publishers, 2006.

Whetzel, D. L., and M. A. McDaniel. "Prediction of National Wealth." Intelligence, 34, 2006, 449-58.

Neal, D. A., and W. R. Johnson. "The Role of Premarket Factors in Black-White Wage Differences." Journal of Political Economy, 104, 1996, 869-95.

Neisser, U. The Rising Curve. Washington, DC: American Psychological Association, 1998.

Ram, R. "IQ and Economic Growth: Further Augmentation of Mankiw-Romer-Weil Model." Economics Letters, 94, 2007, 7-11.

Rushton, J. P., M. Skuy, and T. A. Bons. "Construct Validity of Raven's Advanced Progressive Matrices for African and Non-African Engineering Students in South Africa." International Journal of Selection and Assessment, 12, 2004, 220-29.

Sailer, S. "IQ and the Wealth of Nations, Lynn and Vanhanen: Data Table of National Mean IQ Studies." 2004. Accessed November 6, 2008. http://www.isteve. com/IQ_Table.htm

Sala-i-Martin, X. "I Just Ran Two Million Regressions." American Economic Review, 87, 1997, 178-83. Accessed November 6, 2008.

Sala-i-Martin, X., G. Doppelhofer, and R. Miller. "Determinants of Long-Run Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach." American Economic Review, 94, 2004, 813-35.

Seligman, D. A Question of Intelligence: The IQ Debate in America. New York: Birch Lane Press, 1992.

UNESCO. Statistical Yearbook 1998. Paris, France: UNESCO Publishing and Bernan Press, 1998.

Vinogradov, E., and L. Kolvereid. "Cultural Background, Home Country National Intelligence, and Self-Employment Rates among Immigrants in Norway."

Working paper, Bodo Graduate School of Business, 2006.

Voracek, M. "National Intelligence and Suicide Rate: An Ecological Study of 85 Countries." Personality and Individual Differences, 37, 2004, 543-53.

--. "National Intelligence, Suicide Rate in the Elderly, and a Threshold Intelligence for Suicidality: An Ecological Study of 48 Eurasian Countries." Journal of Biosocial Science, 37, 2005, 721-40.

Weede, E. "Does Human Capital Strongly Affect Economic Growth Rates? Yes, But Only If Assessed Properly." Comparative Sociology, 3, 2004, 115-34.

Weede, E., and S. Kampf. "The Impact of Intelligence and Institutional Improvements on Economic Growth." Kyklos, 55, 2002, 361-80.

Zax, J. S., and D. I. Rees. "IQ, Academic Performance, Environment, and Earnings." Review of Economics and Statistics, 84, 2002, 600-16.


* We would like to thank participants at the Missouri Economics Conference, the Southern Economic Association meetings, the Society for Economic Dynamics, DEGIT XI, McGill University, and George Mason University for helpful comments. We especially thank Francesco Caselli, Michael Davis, William Smith, Petia Stoytcheva, Bryan Caplan, editor Vincenzo Quadrini, and an anonymous referee for particularly helpful recommendations, and the Graduate School of Southern Illinois University Edwardsville for financial support. An earlier version of this article circulated under the title "IQ in the Ramsey Model." The usual disclaimer applies with particular force.

Jones. Assistant Professor, Department of Economics and Center for Study of Public Choice, George Mason University, Fairfax, VA. Phone 314-973-7243, E-mail

Schneider: Assistant Professor, Department of Psychology, Illinois State University, Normal, IL. Phone 309-438-8410, E-mail

(1.) Hendricks's census data on "earnings" combine all forms of income, but we will follow Hendricks's practice and treat them as useful proxies for wages.

(2.) Hendricks addresses the question of immigrant self-selection in detail and finds little evidence that this is quantitatively important. We refer interested readers to his valuable analysis.

(3.) LV made one noteworthy change between their 2002 and 2006 IQ estimates: in cases where they had more than two IQ estimates for a country, they chose the median as their national average IQ estimate rather than their mean.

(4.) For instance, the widely cited work of Zax and Rees (2002) uses data from Wisconsin to estimate the impact of teenage IQ on lifetime earnings. They find that for men in their 50s, [gamma] = 0.7% higher earnings when they control for education and [gamma] = 1.4% when they do not. Since some education is surely caused by prior IQ, and since that education causes higher wages, Zax and Rees note that we should place some weight on the estimates that do not control for education when trying to determine the impact of IQ on wages. They find that IQ--which was measured when these men were teenagers--does a better job predicting wages in a worker's 50s than in his 20s. Neal and Johnson (1996) find that one IQ point is associated with [gamma] = 1.3%, while Bishop (1989) finds [gamma] = 1.1%. Cawley et al. (1997) find U.S. estimates in a similar range, even when they break the estimates down by ethnic group and gender--and their estimates drop by about one-third when they control for education. Behrman, Alderman, and Hoddinott (2004) survey some developing country studies and find that the mean and median estimates both imply [gamma] = 0.8%. We take [gamma] = 1% as reasonable estimate of best practice labor econometric work; U.S. estimates often run a bit larger, while developing country estimates and estimates that control for education often run a hit smaller.

(5.) Vinogradov and Kolvereid (2006) show that LV's national average IQ estimates are good predictors of the self-employment rates of immigrants coming to Norway.

(6.) At first glance, this is surprising: if one thought that workers from low-IQ countries faced enormous hardships, hardships that would impact their level of human capital in ways that would not show up on a so-called pencil and paper IQ test, then one would expect immigrants from those countries to have much lower earnings upon their arrival in the United States than an IQ test would predict. In other words, an IQ of 81 for an American citizen would mean something much less serious than an IQ of 81 for a person from Ecuador. The Ecuadorean 81 would likely come bundled with a history of poor nutrition and education, weak public health services, and other adverse factors. Can a mere pencil and paper IQ test capture the impact of all these various insults on a person's wage-earning ability? The answer appears to be yes, on average. So while one might have expected [gamma] >> 1 in this cross-country regression, that was not the case. At the same time, one might have expected the OLS estimate of [gamma] to be smaller than 1 : if IQ tests in general were a Mis-measure of Man (Gould 1981) then one would expect cross-country IQ tests that were aggregated to the national level and then imputed to the average immigrant to have multiple levels of errors-in-variables problems. This would likely bias the IQ coefficient downward, yielding [gamma] << 1. But neither turned out to be the case: our estimated coefficient is quite close to conventional microeconometric estimates.

(7.) Does this strong IQ-productivity correlation depend on the type of IQ test used? Apparently not, if we look at the IQ tests underlying LV's (2002) estimates. For example, looking only at the 25 scores (out of the 163 total in their 2002 book) that used Cattell's culture-fair test, the correlation with 1998 purchasing power parity-adjusted log GDP per capita was .74, slightly below the .82 in the aggregated sample. For one form of Raven's progressive matrices (a nonverbal, visual pattern-finding IQ test), the correlations were .92 (35 tests), and for the other form of the Raven, the correlation was .69 (53 tests). These were the only three tests appearing more than 25 times in the LV (2002) database. Clearly, regardless of the type of test used, national average IQ can still predict about half or more of a nation's productivity.

(8.) Results were substantially unchanged if 2000 log GDP per person was used instead of log GDP per worker. They were also substantially unchanged if national average IQ was windsorized at values of 70, 80, or even 90 IQ points (first recommended by McDaniel and Whetzel 2004). For example, IQ scores less than 70 were set equal to 70, and the estimates were substantially unchanged when rerun. This windsorizing addresses the concern that IQ scores from the poorest countries are "too low to believe": even if we bump the lowest scores up a few (dozen) points, the results still hold.

(9.) Results were likewise substantially unchanged if we omitted the eight observations that Jones and Schneider (2006) also omitted. They omitted observations from LV's data set that were based on fewer than 100 test subjects per country or that relied exclusively on immigrant data. They also omitted two observations (Peru and Colombia) that partially relied on imputing IQ scores based on the average IQs of residents of nearby countries. Omitting these possibly weaker data points had no substantial effect on the results.
National IQ as a Predictor of Immigrant Earnings

 Coefficient Standard Error p value [R.sup.2]

Dependent variable: log unmeasured worker skill
(i.e., education-and age-adjusted earnings)

IQ 0.95% 0.31% 0.3% 23%
Controls: None
IQ 1.16% 0.35% 0.2% 41%
Controls East Asia/Sub-Saharan Africa/Latin America
 dummies (from Sala-i-Martin 1997;
 Sala-i-Martin, Doppelhofer, and Miller 2004)

Dependent variable: log unadjusted earnings

IQ 1.30% 0.44% 0.5% 18%
Controls: None

Note: 59 observations; White standard errors.
COPYRIGHT 2010 Western Economic Association International
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2010 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Jones, Garett; Schneider, W. Joel
Publication:Economic Inquiry
Geographic Code:1USA
Date:Jul 1, 2010
Previous Article:Male marital wage differentials: training, personal characteristics, and fixed effects.
Next Article:Dead on arrival: zero tolerance laws don't work.

Terms of use | Privacy policy | Copyright © 2019 Farlex, Inc. | Feedback | For webmasters