The Effect of Height on Earnings: Is Stature Just a Proxy for Cognitive and Non-Cognitive Skills?

This study investigates the degree to which the association of height and earnings in Pakistan is independent of other cognitive and socioemotional skills. While taller workers are regularly observed to earn more, they commonly have higher cognitive ability. Thus, there is debate concerning the independent contribution of stature. The study explores the relationship between height and earnings when a measure of cognitive ability -- performance on Raven's matrices -- and an index of socioemotional capacity are included. The study finds that there is only modest attenuation of the coefficient of height?treated as endogenous or exogenous?when these additional indicators of human capital are included.


Policy Research Working Paper 8254
This paper is a joint product of the Social Protection and Labor Global Practice Group and the Human Development Team, Development Research Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at lbossavie@worldbank.org.
This study investigates the degree to which the association of height and earnings in Pakistan is independent of other cognitive and socioemotional skills. While taller workers are regularly observed to earn more, they commonly have higher cognitive ability. Thus, there is debate concerning the independent contribution of stature. The study explores the relationship between height and earnings when a measure of cognitive ability-performance on Raven's matrices-and an index of socioemotional capacity are included. The study finds that there is only modest attenuation of the coefficient of height-treated as endogenous or exogenous-when these additional indicators of human capital are included.

Introduction
Earnings and wages are regularly found to be associated with height in both developed economies (Case and Paxson, 2008;Lundborg, Nystyedt and Rooth, 2014) and low-income settings (LaFave and Thomas, 2017;Schultz, 2003;Thomas and Strauss, 1997;Haddad and Bouis, 1991). In some occupations this may be due to a direct impact of height on physical capacity for work; in others it may reflect the indirect effect of height on schooling or on status or a combination of these (Pitt, Rosenzweig, and Hassan, 2013). Plausibly, however, height may have relatively little direct impact on earnings but may be a proxy for other dimensions of human capital that are less often measured and -for employers -less easily observed at the time of hiring.
Differences in the measured impact of nutrition on earnings or wages may reflect -as is often the case -context. However, distinctions across studies may also reflect whether the results are net of schooling or learning (Behrman et al. 2103) or include the pathway through schooling.
Differences in results may also depend on whether health has been assumed to be exogenous or not (Alderman et al. 1996;Schultz 2003) and whether cognitive skills are included in the analysis (Vogl 2012;LaFave and Thomas 2017). The current study looks at these issues testing whether the impact of height on wages remains robust to the inclusion of cognitive skills as well as an additional measure of socio-emotional skills. These latter skills -which are also referred to as non-cognitive skills, particularly in economics literature -have been shown to be an important determinant of labor market outcomes in the United States (Heckman, Stixrud and Urzua 2006) but have only recently been included in studies of earnings in a wider context. 3 Our results are generally consistent with the findings of Vogl (2014) and Lafave and Thomas (2017) as well as a similar paper by Bargain and Zeidan (2017) in that height remains a determinant of earnings even when cognitive and non-cognitive measures are included. The current study, however, adds to the small pool from which generalizations can be made. Moreover, unlike the previous studies, the current investigation takes the endogeneity of height into consideration.

Basic Conceptual Framework
In order to view human capital over a lifetime, we consider three periods. In the first period, the foundations for an individual's health and nutrition (H i1) are established as a function of investments (Ii1) in that period as well as the individual's own genetic makeup (Xi), his or her family's characteristics (F1) and community infrastructure (V1). These latter two categories can be time varying.
In the following period the child accumulates other forms of human capital (Si2), which can be considered as schooling or learning (Hanushek and Woessmann, 2008) and which reflect health accumulated earlier as well as current inputs and individual, family and community characteristics.
2) Si2 = s(h(Hi1), I i2, Xi, F2, V2) When the individual enters employment his or her wages or earnings reflect both health and learning along with other individual characteristics as well as local market conditions.
More detailed models can illustrate how inputs in one period influence the returns to inputs in subsequent periods or can fine tune different periods of sensitive investments (Cunha and Heckman, 2007). In addition, the number and types of investment in each period included in models of inter-4 period accumulation of human capital depend on the nature of the analysis. However, the model is general and a parsimonious illustration suffices for the study at hand.

Data
This paper uses data from the second wave of the Labor Skills Survey (LSS) conducted in Pakistan in the last quarter of 2013. The survey was designed to be nationally representative and covers all regions of Pakistan except Balochistan and the Federal Administered Tribal Areas, which jointly represent less than 7% of the total population. The sample was drawn using a stratified three-stage design. Twenty districts (7 in urban and 13 in rural areas) were first selected through a random sampling in each of the urban and rural strata. In the second stage, 100 primary sample units at the union council level were selected within each stratum systematically with probability proportional to size. Finally, in the third stage, a random systematic sampling was used to select 25 target households. A total of 2,500 households in 20 districts and 100 union councils were finally selected.
Interviews were completed for 2,354 households in 94 union councils, due to security issues in six union councils of Khyber Pakhtunkhwa (KPK).
The LSS household survey consists of a questionnaire for the household head as well as a separate questionnaire for a subset of one male and one female randomly selected within the household, among all mentally able household members aged between 15 and 64. The household head questionnaire collected general information on all household members including age, gender, height and general education. The male and female questionnaires collected detailed information on individual employment, income, and individual skills reported by the individual himself.
Additionally, cognitive abilities were assessed for all males and females aged 15 to 64 in the household. The height variable used in the paper was obtained from actual measurement of all household members aged 2 and above.
Our measure of cognitive ability is derived from the Raven's test of progressive matrices, administered to all adults aged 15 to 64 in the LSS households. The Raven's test of progressive matrices aims at measuring logical reasoning ability. The adult instrument consists of 60 questions 6 of increasing difficulty in which the respondent is asked to identify the missing figure in a logical sequence of figures. We construct our measure of cognitive development from the answers given to each of the 60 items using Item Response Theory (IRT). 2 The IRT method estimates the relationship between the latent trait of interest, in our case cognitive ability, and the Raven's question items intended to measure the trait using maximum likelihood methods. In the context of this paper, we use a two-parameter logistical model to estimate IRT scores. 3 The main advantage of this approach, compared to using raw scores, is to consider differences in the difficulty of the 60 test questions in the calculation of the cognitive score. The correlation between the Raven's raw scores and IRT scores is, however, very large in our sample at 0.97.
The measure of non-cognitive abilities used in the study is based on the Big 5 Personality Test that was included in the male and female questionnaires. The Big 5 personality test consists of a set of 24 questions aimed at measuring 5 different dimensions of non-cognitive abilities. Each question is a statement about a given behavior of the respondent in his daily life which can be answered to be always true, true most of the time, rarely true or never true. The Big 5 personality traits measured are openness, conscientiousness, extraversion, agreeableness, and neuroticism. The methodology underlying the Big Five taxonomy is the Five Factor Model (FFM) and has been widely used in the psychology and economics literature (Heckman and Kautz, 2012;Heckman and Mosso, 2014).
The Big Five indicators have been proven to show consistency in interviews, selfdescriptions and observations (Costa and McCrae, 1987). We construct our indicator of noncognitive abilities from the score obtained in the 5 components of the Big 5 personality traits, using 7 a principal component analysis. The first principal component of the Big 5 indicators obtained from this procedure is used as our index of non-cognitive skills.
The Pakistan LSS surveys a total of 11,533 individuals in 2,239 households. Among those, 2,045 males age 15-64 were administered the detailed labor questionnaire. Since income data are non-zero only for individuals who work and who receive payment or for those respondents who report earning a profit, our final sample consists of 1,419 working male individuals that were identified as either self-employed (including agriculture), or paid employees. Given the small number of paid working females in the sample, the paper restricts the analysis to a sample of male workers. We classify individuals who report earnings from their own activity as self-employed, and individuals who receive payment from their employer as wage earners. Self-employed workers were asked the amount of net profits generated over the last period of work, defined as sales minus expenses to the individual. We use this amount converted into monthly profits as our measure of earnings for the self-employed. Table 1 reports summary statistics on this sample. There is a legitimate concern that land owners could conflate rental earnings with crop production minus expenses. However, this is not an obvious issue with the data. Land owners -20% of the entire category of self-employed and 44% of all individuals who reported cultivation as their main employment -reported agricultural profits that were not significantly different than those reported by renters. The median monthly profit for land-owners in the sample is PK Rps 7,000, against PK Rps 7,500 for tenants. Table 2 presents OLS regressions of the logarithm of earnings from both wage and own employment.

Results
The first three columns focus on the impacts of stature, cognitive ability and non-cognitive skills 8 entered separately. Regressions also include categorical dummies for the highest level of schooling completed by the individual as well as potential experience and its quadratic term. 4 An additional centimeter of height contributes nearly 0.8% to earnings. The regression in the fourth column includes all three of these measures jointly. As indicated, there is a modest attenuation of the magnitude of height (0.7 percentage points or 11% of the initial estimate) and cognition compared to that observed in the first three columns, but they remain individually significant.
Moreover, as shown in column 5 when we add the education of the father, that variable proves significant. As the labor market does not directly reward the ability of the father -many of whom are deceased -this may indicate that there are aspects of the ability of the current generation of workers that are not directly measured by stature or the skills included here yet are recognized in the labor market. Possibly the father's education is a proxy for unmeasured skills that are genetically transmitted. Alternatively, or additionally to this interpretation, the coefficient of father's education may indicate learning that is imparted by parental guidance. Furthermore, the coefficient can also reflect access to networks that an educated father can facilitate. In this specification, the magnitude of the height coefficient decreases slightly (approximately 10 percent) compared to column 4, but remains statistically significant.
The regression in column 5 may be considered as the full reduced form impact of these categories of skills. That is, the coefficients capture the indirect impact of these aspects of human capital on wages via schooling and the impact of stature and skills on labor market choices regarding sector and labor supply. In addition, they measure any direct impact on earnings conditional on these choices. Cawley, Heckman, and Vytlacil (2001) show that the estimated impact of ability on wages is substantially smaller when it is conditional on levels of schooling. This is also observed in our data. Column 6 reports the impact of stature and skills conditional on schooling. The coefficient on 9 height remains significant but the magnitude declines relative to column 5. The impact of height conditional on school might either be because it influences learning per year of school or because it conveys abilities that are rewarded in earnings beyond the returns to schooling per se or both.
In contrast, the measure of cognitive skills is no longer significant in the regression in column 6. The standard errors for the coefficient of Raven's declines relative to the previous estimate, so the loss of significance is driven by the reduction of the point estimate. The impact implied by the coefficient of Raven's in column 5 likely works primarily through the indirect impact of skills on schooling. That is, the coefficients in column 5 can be viewed largely as δW/δS*δS/δH. At the same time, it appears that schooling is not merely a signal for these skills since the inclusion of schooling increases the portion of earnings explained in the regression.
However, as mentioned, wage earnings and earnings from own employment are pooled in the first 6 columns. Columns 7 and 8 indicate how these skills influence earnings in these two sectors respectively. Height is far more important in wage employment than in own employment. This is in partial contrast to Pitt, Rosenzweig, and Hassan (2013) who argue that employment in agriculture may reflect the relative importance of physical capacity in that sector. As 41.1 percent of the individuals in our sample who are self-employed are in agriculture the results in column 7 should reflect the role of stature in agriculture. Our measure of socio-emotional skills is not important in wage employment although it is in self-employment. Conditional on schooling, Raven's scores are not significant in either sector, although non-cognitive skills have a role in self-employment earnings while the coefficient of father's education is significant in both sectors.
The results in table 2, however, tacitly assume that equations 1 and 2 can be considered lagged endogenous and also that they have no common unobservables with equation 3 and, thus, no correlation of errors with estimations based on equation 3. We relax this assumption in table 3, which 10 reports the same regressions as in table 2 with stature instrumented. 5 The land holding of the father of the current employed individual (as this individual recollects it) is used as one of the instruments.
The first stage regression is reported in table A1 of the appendix. As the survey obtained height information from all adults but earnings for only a subset, the full sample was used for instrumenting height and standard errors for the second stage IV were obtained by bootstrapping rather than running the IV estimates as a simultaneous set of equations. This approach was also motivated by the fact that the first step of the IV corresponds to equation 1 presented in the methodology while the wage equation is an estimate for equation 3. Time variant information that is observed at the time of employment does not pertain to the production of skills in equation 1 and thus is not appropriate in the estimation of height. The F statistic for the first stage regressions reported in columns 1-4 exceeds the rule of thumb for plausible instruments from Stock and Yogo (2005).
Columns 5 and 6 offer results using an alternative instrument, the residual of a regression for the child's height. This residual is assumed to pick up common genetic factors as well as other elements of the child height regression which are orthogonal to other determinants of child height.
Carslake et al. (2013) use a similar approach as an instrument for father's height; however, in their study, son's height rather than a residual is employed as an instrument. As only a subset of the sample had children who were measured, the regressions in columns 5 and 6 have fewer observations. This has a moderate effect on the significance of the height coefficient, but the point estimate is similar to that in columns 1 and 2. The point estimates are again similar when both potential instruments are included in columns 7 and 8 but the standard errors increase. The regressions in those columns, however, allow an overidentification test which provides some reassurance that the area of land owned by the father is a valid instrument. 6 While the magnitude of the contribution of height to explaining wages in table 2 using OLS regressions is somewhat lower than reported elsewhere in the literature (such as in Vogl 2014), the magnitude of the coefficient of instrumented height in table 3 increases substantially. 7 This may reflect a combination of errors in measurement for height as well as endogenous choices. In contrast to the OLS results the IV coefficient of height is significant for self-employed (including those engaged in agriculture) while it is not for wage workers.
In principle, schooling is also an endogenous choice and thus it would be desirable to also instrument the education variables in table 3. Duflo (2001) conversely, the probability of being self-employed including working in agriculture increases with father's landholding as is logical. The IRT cognitive score is significant in the choice of wage employment even when years of completed schooling are included but neither height nor the index of socio-emotional skills influence selection into wage employment. The absence of a role of height in sectoral choice is in contrast to the results in both Vogl (2014) and LaFave and Thomas (2013).
However, the significance of the IRT cognitive score is consistent with other evidence in the literature including the studies cited in the introduction.
The OLS results for wages conditional of selection reported in table 5 do not differ appreciably from the corresponding tables which do not account for selection into wage employment. The IV results, however, reflect the challenge of accounting for both selection as well as endogeneity with the limited instruments for lagged decisions on health. In effect, we have to rely on selection by functional form which is not preferred; the results can be considered indicative albeit weakly identified. Statistical power is lost, for example, on the coefficient of height in selfemployment although the point estimate is similar to that in table 3.

Conclusion
The results reported in this paper support the view that height provides independent information on labor productivity rather than only serving as a proxy for other common measures of human capital. This general point is consistent with other studies that include one or more measures of skills in addition to height. The IV results reported here also provide no indication that the association of height and wages in OLS results is biased upwards due to unmeasured aspects of early home environment. Moreover, the IV coefficient of height in the current study is similar in magnitude to 13 those in Schultz (2003) as well as the OLS results conditional on the inclusion of cognitive ability in Vogl (2014). Thus, while conceptually an IV approach is preferred to OLS for measuring human capital impacts, the results here reinforce rather than challenge the core evidence from OLS studies in the literature.
However, drilling down a bit into the manner by which height explains labor choices indicates that there also are some differences compared to previous studies. For example, height seems to have no role in selection into wage work in Pakistan and, of course, given the bivariate choice in the choice of occupations this also implies it has no role in selection into other selfemployment activities including those in which brawn is assumed to be more central.
The differences with other results in the literature likely reflect context; there are too few studies that include both height and a range of other measures of human capital from which to generalize. In addition, the data required for precise estimation of selection and endogenous choice over the life-cycle are daunting. Still, given that the paper also shows that socio-emotional skills have additional explanatory power, a finding that is more commonly noted in studies from high income countries, the three relatively accessible measures of skills studied here confirm not only that ability is multi-dimensional but that insights into these dimensions are available with judicious modifications of standard labor force surveys.       (1) and (2), the area of land owned by the worker's father, available for the full sample, is the excluded instrument used as an additional predictor of the worker's height. In columns (3) and (4) The residual of that equation should reflect the child's genetic makeup in addition to the influence of unobserved community elements and conventional measurement error. While genes flow from parent to child, the association is two ways and including the average residual of all measured descendants as a regressor in the equation for the height of the adult provides information not generally available and the coefficient is plausible. Unfortunately, not all individuals in the labor sample had children young enough to be included in the measurement of height, which reduces the sample size in the specifications where the child's height residual is used as instrument for the worker's height. In columns (5) and (6), father's landholding and children's height residuals are both used to predict height.
22  Adjusted R-squared 0.096 Number of observations 6,165 Note: *: statistically significant at the 10% level; **: statistically significant at the 5% level; ***: statistically significant at the 1% percent level. Robust standard errors clustered at the village level are reported in parenthesis. The estimation ample excludes individuals with height-for-age z-scores lower than -6 or greater than 6, following the WHO guidelines. The residuals of the estimation reported in column (2) are used as instrument for the worker's height in the IV estimation of the earnings equation.