Leadership Skills and Wages Revisited: Is There a Causal Relation?

Leadership Skills and Wages Revisited: Is There a Causal Relation?

Ozkan Eren *

I. Serkan Ozbeklik †

University of Nevada, Las Vegas

Claremont McKenna College

Abstract This paper examines the role of holding a leadership position in high school on adult earnings and assesses the sensitivity of the previously found positive association to nonrandom selection bias. Even though the estimate is still relatively large in magnitude, controlling for pre-high school activities produces insignificant coefficient estimates. This may indicate the leadership activities students engage in during high school to merely act as observable indicators for leadership skills. Our sensitivity checks using a recently developed procedure raise further concerns regarding the causality of high school leadership activities on earnings; the selection on unobservables has to be only around 10 percent as strong as selection on observables to explain the entire leadership effect. Finally, in the presence of fixed effects, we show that the use of selection on observables to address non-random selection bias may yield misleading results if the fixed effects are not purged out properly.

JEL: J24, J30, I20 Keywords: Implied Ratio, Noncognitive Skills and Selection on Unobservables.

*

Ozkan Eren, Department of Economics, College of Business, University of Nevada, Las Vegas, 4505 Maryland Parkway, Las Vegas,NV 89154-6005, U.S.A. Tel: 1-702-895-3653. Fax: 1-702-895-1354 ] E-mail: [email protected].

†

Corresponding Author: I. Serkan Ozbeklik, the Robert Day School of Economics and Finance, Claremont McKenna College, Claremont, CA 91711. Tel: 909-607-0721. Fax: 909-621-8249. E-mail: [email protected]. The data sets used in this paper are available upon request.

1

Introduction

Researchers have traditionally focused on cognitive skills, measured by knowledge and aptitude tests, as the primary source of productive skills. Viewing cognitive traits as the sole aspect of productive skills, however, may be misleading because there is prominent competing evidence. For instance, Green et al. (1998) report the survey results from the British Employers Manpower and Skills Practices in which roughly one-third of the establishments respond positively to the skill shortage inquiry and identify the recruitment problem arising mainly from poor personality, motivation and attitude rather than the lack of cognitive skills. Similarly, in a 1998 survey conducted by the U.S. Census Bureau in collaboration with the Department of Education, employers, when considering the hiring process, rank non-cognitive skills as far more important than the years of schooling or academic performance. Moreover, the sociology and psychology literature have given the non-cognitive skills an equally predictive power for many labor market and social outcomes as they do to cognitive skills (see, for example, Barrick and Mount 1991, Hogan and Holland 2003). In the last decade or so, this growing evidence on the role of non-cognitive skills for economic outcomes has attracted the interest of researchers. Using the National Longitudinal Survey of Youth (NLSY), Goldsmith et al. (1997) …nd positive and signi…cant e¤ects of self-esteem on earnings. Bowles et al. (2001) with di¤erent data sets discuss the e¤ects of personal traits such as self-esteem, optimism and aggression on earnings and schooling. Coleman and DeLeire (2003), using the National Educational Longitudinal Study (NELS) data, obtain a signi…cant impact of locus of control on the expected earnings at age 30. Kuhn and Weinberger (2005) examine the association between leadership activities during high school and labor market outcomes of roughly a decade after graduation. Heckman et al. (2006) with the NLSY data demonstrate that non-cognitive ability as measured by locus of control and self-esteem scales are important in explaining various aspects of social and economic life. In a special volume in 2008, several papers investigate the formation of noncognitive skills and their role on social and labor market outcomes. For instance,

1

Cuhna and Heckman (2008) introduce a new model with an emphasis on the complementarity between cognitive and noncognitive traits and Borghans et al. (2008) and Krueger and Schkade (2008) present evidence about the role of noncognitive skills on the occupational decision and Fortin (2008) using the NELS data examines the role of several noncognitive traits (for example, self-esteem and locus of control) in explaining the gender wage gap and …nds a modest but signi…cant role of these traits in accounting for the gender wage gap. Segal (2011) with the same data …nds a negative and signi…cant association between early adolescence misbehaving and schooling/earnings. Finally, Lindqvist and Vestman (2011) show that personal traits, based on interviews conducted by psychologists during the military enlistment, are strong predictors of labor market outcomes for Swedish men. Of particular importance among the aforementioned studies in the context of our paper is Kuhn and Weinberger (2005). In their innovative paper, Kuhn and Weinberger have meticulously examined the impact of leadership activities on adult earnings and have found statistically signi…cant correlations. The authors’ further exploration of the issue indicate that high school leadership does not act as a proxy for physical or psychological attributes and that the signi…cant correlation between high school activities and earnings persist even after controlling for these attributes. With that said, it is important to note that Kuhn and Weinberger (2005) never argue the e¤ect of high school leadership activities on earnings (or other labor market outcomes) to be causal. The purpose of this paper is two-folds. First, we revisit the association between the leadership activities students engage in during high school and their adult earnings using the more recent NELS data. The richness of the data set and the longitudinal nature of the leadership variables (i.e., prehigh school leadership activities) allow us to investigate this association in more details. Second, we examine whether the correlation between the leadership activities and their adult earnings is a causal one. Using the NELS data, as well as the two other data sets (National Longitudinal Survey 1972 (NLS72) and High School and Beyond (HSB)), from Kuhn and Weinberger (2005), we analyze the sensitivity of the high school leadership e¤ect to non-random selection by employing the econometric 2

strategy developed in Altonji et al (2005). Speci…cally, this technique allows us to calculate how large the extent of selection on unobservables relative to selection on observables must be in order to fully explain the treatment e¤ect (leadership in our case) estimated under the assumption of exogeneity. Our …ndings using the NELS are consonant with Kuhn and Weinberger (2005); the estimated high school leadership e¤ect on earnings is in the range of 10 to 14 percent. Controlling for physical and psychological attributes do not largely alter the results. Adding the pre-high school leadership activities produces relatively large leadership coe¢ cient estimates, but these estimates are not distinguishable from zero. This may indicate the leadership activities students engage in during high school to merely act as observable indicators for leadership skills. Our sensitivity checks further raise concerns regarding the causality of high school leadership activities on earnings; the selection on unobservables has to be only around 10 percent as strong as selection on observables to explain the entire leadership e¤ect. Finally, in the presence of …xed e¤ects, we show that the use of selection on observables to address non-random selection bias may yield misleading results if the …xed e¤ects are not purged out properly.

2

Data and Estimation

The primary data set for this study is drawn from National Education Longitudinal Survey (NELS). The NELS is a large longitudinal study of eighth grade students conducted by the National Center for Education Statistics in 1988. The strati…ed sample of NELS was chosen in two stages. In the …rst stage, a total of 1032 schools on the basis of school size were selected from a universe of approximately 40,000 schools. In the second stage, up to 26 students were selected from each of the sample schools based on race and gender. For subsample of respondents, follow-up surveys were administered in 1990 (…rst-follow up, tenth grade), 1992 (second-follow up, twelfth grade), 1994 (third-follow up) and 2000 (fourth-follow up). To be consonant with Kuhn and Weinberger (2005) (KW, hereafter), we use the same sample

3

restrictions and focus on only white men who graduated from high school and were employed in 2000. Focusing on only white males mitigates the possible contamination of the parameter of interest due to racial and gender discrimination in the labor market. The students of the NELS were asked a series of questions pertaining to their sport/club participation and the responses to these questions from the second-follow up constitute our measure of leadership activity. The NELS also allows us to condition on the same set of control variables (math test scores, parents’ educational attainment and school …xed e¤ects) used in KW. The dependent variable is the log hourly earnings obtained by dividing annual earnings by weeks worked times hours worked in a typical week. We exclude those whose hourly earnings are below $3.75 and above $48.8. This corresponds to the 1st and 99th percentile of the earning distribution. The e¤ective sample excludes observations with missing data on hourly earnings, on the leadership activity measures, as well as on the math test score.1 Dummy variables are used to control for missing values of parents’educational attainment. The …nal sample contains 1,779 white males.2 In addition to NELS, we supplement our analysis with two other data sets. KW, in their examination of the leadership premium, have relied on three di¤erent data sets covering a time span from 1960s to 1990s. The …rst one is Project TALENT where the students in this study were surveyed during high school in 1960 and were followed longitudinally for 11 years after high school. The second data set is National Longitudinal Study of the Class of 1972 (NLS72) that collected detailed personal, academic and behavioral information from 1972 high school seniors and followed them up until 1986. The …nal data set comes from the High School and Beyond (HSB), which surveyed 1980 high school sophomores and continued longitudinally up until 1992. Among these three data sets, we employ 1

We imputed the tenth grade math test score for those with missing twelfth grade math test score. The imputation increased our sample size by 9.3%. 2 The fourth follow-up includes a raw sample of 12,051 observations. Focusing on only white males leads to a roughly 69% reduction in the sample leaving us with 3,735 young men. Excluding those who were enrolled in school at the time of interview, who were high school dropouts or unemployed restricts the sample to 2,166 individuals. Finally, dropping observations with missing values on ability measures, weekly income, as well as leadership measures and trimming the weekly income distribution by 1% from the lower and upper ends leads to our …nal sample of 1,779 individuals.

4

the NLS72 and HSB to have some benchmark estimates and more importantly, to test whether the leadership e¤ects found in KW are causal. Applying the same data restrictions, we end up with 3,109 and 2,440 observations for NLS72 and HSB, respectively. We present the selected sample means for the variables of interest for all three data sets in Table 1. The …rst and third columns display the means for the KW’s NLS72 and HSB samples while the second and fourth columns show the means for our respective samples. There are very slight di¤erences between the KW’s and our samples. However, as discussed in the subsequent section, this does not generate any signi…cant discrepancy in our replication of the KW’s results. Finally, the …fth column of Table 1 presents our sample means for the NELS. Slightly more than 11% of the students in the NELS occupied both a team captain and club president position in the senior year of high school. The similarity among the means of the covariates provides strong assurance that the three data sets are comparable. In order to measure the leadership e¤ect, similar to KW, we estimate the following outcome equation:

0

Wi = Leaderi + Xi +

(1)

i

where W is the hourly earnings for NLS72 and NELS (and annual earnings for HSB) of student i, X is the vector of control variables including the senior level math scores, highest level of parental educational attainment, indicators for occupying only a team captain or a club president position, indicators for team or club membership (and both), as well as the school …xed-e¤ects and

is the

usual error term. Leader is a dummy indicator that is equal to 1 if the student was both a team captain and a club president in high school and zero, otherwise. The OLS estimate of true leadership e¤ect on adult earnings as long as Cov(Leader; ) = 0 is satis…ed.

5

yields the

3

Results

3.1

Baseline Results

Table 2 presents the school …xed e¤ect estimates. The KW’s original estimates using NLS72 and HSB are given in columns 1 and 3, our replication of their results using the same data sets are given in columns 2 and 4 and …nally, the results from NELS are presented in Column 5 of Table 2.3 As it is visible from the table, we were able to replicate the KW results to a great extent. With respect to the leadership premium, the estimates indicate an e¤ect within the range of 0.126 to 0.241. That is, those holding a team captain and a club president position in high school earn 12.6% to 24.1% higher hourly or annual income relative to men who joined both teams and club but held no leadership position. For HSB and NELS data sets, we also observe that men who were either captain or president signi…cantly earn more than compared to those who were only members of teams or clubs.

3.2

Do Leadership Skills Act as a Proxy for Some Other Attributes?

In addition to …nding a strong association between leadership activities and adult earnings, KW examined the source of this positive correlation. Speci…cally, the authors looked at whether leadership skills are acting as proxies for physical or psychological attributes. Several di¤erent speci…cations such as including measures for physical attractiveness (for example, body mass index) and for psychological characteristics (for example, self-esteem index) leaves the leadership coe¢ cient almost identical. Similar to KW, in this section, we augment several di¤erent variables to examine the robustness of the leader coe¢ cient. To account for physical attractiveness, we utilize the self-reported questionnaires from the second-follow up for whether the student identi…es him or herself as being athletic and popular. The responses for each question are divided into three categories: very, somewhat and not at all. The …rst column of Table 3 reports the results from this speci…cation and adding these attributes 3

The coe¢ cients on other control variables are suppressed for the ease of presentation.

6

do not a¤ect the leadership coe¢ cient estimate. To investigate if the leadership e¤ect proxies for some observable psychological attributes, we add the following measures from the second-follow up to our baseline speci…cation: self reported questionnaire for whether the student identi…es him or herself as being social (very, somewhat and not at all) and an average of the composite Rosenberg self-esteem and Rotter locus of control scales. The Rosenberg Scale refers to the perceptions of self-esteem based on 7 statements such as “I feel I am a person of worth, the equal of other”and “On the whole, I am satis…ed with myself”(Rosenberg 1965). The Rotter Scale, on the other hand, refers to the extent to which individuals believe that they can control outcomes that a¤ect them based on 6 statements such as “I do not have enough control over the direction my life is taking”and “Chance and luck are very important for what happens in my life” (Rotter 1966). A larger value of the average of the composite index implies higher noncognitive skills. The results shown in the second column of Table 3 indicate that self esteem and locus of control a¤ect the adult earnings, but the leadership e¤ect is not simply capturing these traits. In the last column, we add both the physical attractiveness and psychological measures at the same time and doing so does not a¤ect the coe¢ cient estimate of leadership either.

3.3

Are Leadership Skills Acquired During High School?

A crucial question for policy implication would be whether the leadership skills are acquired in high school or are individuals endowed with them even before reaching to high school. KW use school-level leadership opportunities available to students during high school to shed some light on this issue. Speci…cally, the authors show that students prone to leadership who attend high schools with more leadership opportunities earn signi…cantly more compared to otherwise similar students attending schools with less leadership opportunities. This …nding potentially implies a kind of complementarity between acquired and endowed components of leadership skills. In this section, using a unique feature of the NELS, we take this analysis one step further. Our

7

contribution over KW is two-folds. First, we have measures of activities and the positions held before the high school entry, which enables us to condition and examine the impact of pre-high school skills. We would expect a signi…cant impact of eighth grade variables on earnings under the endowed skill hypothesis. Similar to senior year in high school, the students were asked a series of questions pertaining to their sport/club participation in eighth grade. The responses are divided into three categories: did not participate, participate as a member and participated as an o¢ cer. Depending on the nature of participation activity, we de…ne the eighth grade o¢ cers as team o¢ cer or club o¢ cer (or both) and eighth grade membership is de…ned similarly. Second, unlike KW, we can still control for the school …xed e¤ects since our examination of the endowed versus acquired skills is based on individual level data. To get a grasp of team/club participation across the grades, we …rst present the distribution of the o¢ cer and leadership positions held in eighth and twelfth grades separately (Panel A of Table 4). In our e¤ective sample (1,560 observations once we drop the missing observations for eighth grade team/club participation), 14% of all eighth graders were either a club o¢ cer or a team o¢ cer (or both), while the corresponding fraction for leadership activity in twelfth grade was around 40%. Next, in Panel B of Table 4, we display a matrix of twelfth grade leadership and eighth grade o¢ cer/membership involvement. More than 8% of the students held some type of position (leadership or o¢ cer) in both eighth and twelfth grades. Among the twelfth graders who were team captains and club president, 26% were also either a team or a club o¢ cer (or both) in eighth grade. More interestingly, we observe a substantial overlap between the twelfth grade leadership indicator and holding both eighth grade team and club membership position. Around 70% of those who acted as leaders in twelfth grade were also team and club members in eighth grade. Given this large overlap, team and club membership position in eighth grade may better proxy the skill set of those twelfth grade leaders under the endowed skill hypothesis. In the …rst column of Table 5, we …rst present our baseline speci…cation by replacing the twelfth 8

grade leader and membership indicators with those from eighth grade to distinguish between acquired versus endowed skills. Without any school …xed e¤ects, holding both club and team o¢ cer position in eighth grade signi…cantly a¤ects hourly earnings with an estimated impact of 11.3%, while holding membership in teams and clubs has an e¤ect of around 8%. Adding the eighth grade school …xed e¤ects in the second column reduces the coe¢ cient estimate of the o¢ cer position to 4.7% and the e¤ect is no longer statistically signi…cant. The coe¢ cient estimate of holding both team and club membership, on the other hand, is still statistically signi…cant with a magnitude just above 15%. Next, we incorporate the eighth grade variables to our twelfth grade speci…cation. Given the reduction in our sample size, we …rst present the equivalent estimates from Table 2 for 1,560 observations in the third column. Although the leadership coe¢ cient falls around 4 percentage points (from 0.142 to 0.106), it is still statistically signi…cant. In the next column, we add the eighth grade o¢ cer and membership variables. Doing so reduces the magnitude of the leadership coe¢ cient to a small extent, but the impact falls short of signi…cance at the conventional levels. The estimate for holding a team and club membership in eighth grade continues to be statistically signi…cant with a magnitude of 0.130 (0.053). In the last two columns, we replicate the same exercise from columns 3 and 4, but this time by adding the physical attractiveness and psychological attributes to our model for a sample of 1,529 observations. The leadership coe¢ cient estimate with the baseline controls (column 5) is 0.104 (0.050), while it falls to 0.084 (0.055) in the last column of Table 5 and is not di¤erent than zero. In this most extensive speci…cation, it is only the coe¢ cient estimate of holding both a team and a club membership position in eighth grade that is statistically signi…cant. Overall, we observe the leadership impact to be somewhat sensitive to the inclusion of pre-high school leadership activities. The estimates from the fourth and the last column of Table 5 may lend support to endowed skill hypothesis. Alternatively, the insigni…cance of the leadership estimates may have been driven because of a smaller sample. In the subsequent section, we further examine whether the leadership activities students engage in during high school merely re‡ects observable indicators for 9

leadership skills or whether it is the actual high school experience that leads to higher earnings.

3.4 3.4.1

Nonrandom Selection and The Extent of Unobservables Selection on Observed and Unobserved Variables with Fixed E¤ects

As noted, the reliability of the coe¢ cient estimates of holding a leadership position hinge upon the selection on observables assumption. Failure to assess the validity of this assumption and problems with coming up with a proper instrument prevents one from making a …rm inference about causality. Fortunately, we can evaluate the potential role of non-random (positive) selection by applying the procedure developed in Altonji et al. (2005). Intuitively, the basic idea is to assess how much selection on unobservables there must be, relative to selection on observables, to fully account for the positive association between leadership e¤ect and earnings. Given that the leadership variable takes only the values of 0 and 1, one can compute the (normalized) amount of selection on unobservables by the ratio

E[ jLeader = 1] E[ jLeader = 0] V ar( )

where

(2)

is the error term from equation (1) and denotes the unobservables in the outcome equation.

Similarly, the (normalized) amount of selection on observables can be written as the ratio

0 0 E[X ~ jLeader = 1] E[X ~ jLeader = 0] V ar(X 0 ~ )

(3)

where X is the set of control variables from equation (1) and ~ is the corresponding parameter vector. Using equations (2) and (3), one can, then, ask how large the selection on unobservables relative to selection on observables must be in order to explain the entire leadership e¤ect :

10

To begin with, we express the participation equation as

0

Leaderi = Xi + vi

(4)

where Leader is the leadership indicator as described in equation (1), X is the identical set of control variables and v is the usual error term of the participation equation. Substituting equation (4) into (1) gives 0

Wi = Xi ( +

The probability limit of the OLS estimator of

p lim ~ = =

) + vi +

(5)

i

is then given by

Cov(v; ) V ar(v) V ar(Leader) + E[ jLeader = 1] V ar(v)

(6)

+

E[ jLeader = 0]

Under the assumption that the degree of selection on observables is equal to the degree of selection on unobservables, the bias term in (6) turns out to be

V ar(Leader) Cov(v; ) = V ar(v) V ar(v)

Setting

(

) 0 0 E[X ~ jLeader = 1] E[X ~ jLeader = 0] V ar( ) V ar(X 0 ~ )

(7)

in equation (5) equal to zero, one can consistently estimate ~ via OLS. The estimated value

of ~ and variance of the residual, along with sample values of V ar(Leader) and V ar(v) would provide an estimate of the asymptotic bias under equal degrees of selection on observables and unobservables. The key assumption leading to equal selection on observables and unobservables is that the set of covariates are a random draw from all factors a¤ecting the outcome of interest (i.e., earnings). This assumption may not hold exactly. Large data sets such as the NELS, however, are designed to serve multiple purposes rather than to address a speci…c question. Moreover, as indicated in Altonji et al.

11

(2005), as a result of the limits on the number of the factors that we know matter and that we know to collect and can a¤ord to collect, the variables that are available may be more or less a random subset of the ones relevant for a particular question. It would be hard to imagine that the sole purpose of NELS is to examine the e¤ects of high school leadership activities.4 One other necessary assumption is to have a large set of observables and no observable or unobservable factor plays too large role in the determination of the outcome variable. Under these assumptions, dividing the unconstrained estimate of

from equation (1) by (7) yields

a measure, which Altonji et al. (2005) describe as the implied ratio. The implied ratio indicates how much larger the extent of selection on unobservables needs to be, relative to selection on observables, to entirely explain the treatment (leadership) e¤ect. Small values of the implied ratio point out that the treatment e¤ect is highly sensitive to selection on unobservables. Even though there is not a threshold value in order to make a distinction between a small or a large implied ratio, it may be reasonable to consider a ratio of at least in the neighborhood of one to conclude for a causal relation.5 With that said, a potential complication, however, may arise in applying the procedure of Altonji et al. (2005) in this speci…c case. Our speci…cation includes the school …xed e¤ects and as it is well known, the …xed e¤ects inherently captures both observables and unobservable characteristics. Keeping the …xed e¤ects in the speci…cations thus contradicts with the whole idea of using the degree of selection on observables as a guide to the degree of selection on unobservables. To overcome this, we purge out the school …xed e¤ects by demeaning our data sets at the school level

Wij

Wi = ::

Wi =

0

(Leaderij ::

Leaderi ) + (Xij ::

Leaderi + X i +

:: i

0

Xi ) + (

ij

i)

(8) (9)

4 The assumption of equal selection on observables and unobservables should also not necessarily be taken literally. Researchers often do not choose their righthand side variables at random and there will be less selection on unobservables than selection on observables. This assumption represents one extreme. At the other extreme, there stands no selection on unobservables (i.e., OLS). 5 A large value of the implied ratio only provides insights about a causal relation. It does not imply that the magnitude obtained, say, from OLS estimation is unbiased or consistent.

12

where bar on each variable represents the corresponding school level average and note that the binary indicator turns out to be a continuous variable after demeaning. In the case of a continuous endogenous variable, the asymptotic bias from OLS in equation (6) can be generalized to

Cov(Leader; ) . Setting V ar(v)

the degree of selection on observables to be equal to the degree of selection on unobservables, the bias term turns out to be 0

Cov(X ~ ; Leader) V ar( ) Cov(Leader; ) = V ar(v) V ar(v) V ar(X 0 ~ ) where this equality follows from the fact that equation (2) and (3) can be written as

(10) Cov(Leader; ) V ar( )

0

Cov(X ~ ; Leader) ; respectively in the case of a continuous endogenous variable. Similar to the and V ar(X 0 ~ ) binary case, dividing the unconstrained estimate of

3.4.2

from equation (1) by (10) gives the implied ratio.

Assessing the E¤ectiveness of High School Leadership Activities

The implied ratios are provided in Table 6. The …rst row of Table 6 presents the demeaned (school …xede¤ect) estimates of

for di¤erent data sets and for di¤erent speci…cations, the estimated asymptotic

bias is given in row 2 and the last row displays the estimated implied ratios. For NLS72, the implied ratio is 0.08; that is if the (normalized) amount of selection on unobservables is only 8 percent as strong as the amount of selection on observables, the entire e¤ect of holding a leadership position in high school on adult earnings is fully explained. The evidence for leadership premium from the NLS72 is strong enough to completely rule out the possibility that the positive e¤ect is causal. Similarly, the estimated ratio for HSB (0.29) is well below the threshold value of one. For our baseline speci…cation from NELS (column 3), the implied ratio is once again very small (0.13) indicating the fact that if the selection on unobservables is 13 percent of selection on observables, the e¤ect of leadership would be negligible. In the remaining columns, we calculate the implied ratios for several other speci…cations. The implied ratios range between 0.12 to 0.08. Overall, the leadership e¤ect seems to be very sensitive to non-random selection.

13

Finally, as discussed above, Altonji et al. (2005) caution against the misuse of the idea of using observables to draw inferences about selection bias if the observables are small in number and explanatory power. Following this caveat, we have extended the speci…cation from the last column of Table 6 to include the following variables: family income, socioeconomic status of the family, family size, indicators for skipping and cutting school and for school suspension and for whether the student got into trouble. All the additional control variables come from the second follow-up. The regression results from this most extensive speci…cation yields a statistically insigni…cant leadership coe¢ cient of 0.081 (0.053) and an implied ratio of 0.11.

3.4.3

Using Demeaning versus Fixed E¤ect Dummies

There has been an upsurge in interest among researchers in using selection on observables to address non-random selection bias. Several …xed e¤ect speci…cation studies have employed the econometric technique developed in Altonji et al. (2005) to support their …ndings in favor of a causal (or noncausal) relation (e.g., Emran and Shilpi 2008, Cavalcanti and Guimaraes 2010 and Kingdon and Teal 2010). In applying the Altonji et al. (2005), these studies simply add the …xed e¤ect dummies into the speci…cations. To our knowledge, since we are the …rst to deal with …xed e¤ects correctly in this setting, we want to discuss the potential threat to inference. As noted above, the …xed e¤ects capture both observable and unobservable characteristics. Keeping them in the regression thus contradicts with the whole idea of using selection on observables as a guide to the degree of selection on unobservables. To see this, suppose we rerun the regressions of Table 6 with school …xed e¤ects rather than demeaning and purging them out. Then, our leadership measure would still be an indicator and we can calculate the implied ratios using equations (2), (3) and (7). The results from this exercise are given in Table 7. As it is visible from the table, the implied ratios are signi…cantly overstated. Taking the ratio of one as our benchmark, it is hard to rule out a causal relation between leadership activities and earnings for the second, third and fourth columns. Therefore, prior to applying the procedure

14

in Altonji et al. (2005), dealing with the …xed e¤ects properly seems crucial to avoid any misleading interpretation.

4

Conclusion

In recent years, noncognitive skills in explaining social and labor outcomes have received considerable attention. Among many di¤erent aspects, this paper examines one dimension of noncognitive skills, namely the leadership skills. Using the NELS and following Kuhn and Weinberger (2005), we revisit the association between the leadership activities students engage in during high school and their adult earnings. Viewing the complete set of results, we have the following …ndings. There is a strong positive association between high school leadership activities and adult earnings. This holds for three data sets covering the period from early 1980s to early 2000. Conditioning on several physical or psychological attributes do not change the coe¢ cient estimates of leadership indicator. The positive association turns out to be statistically insigni…cant, but still large in magnitude, once we control for pre-high school leadership activities. Our further …ndings, however, indicate that only a small amount of selection on unobservables is su¢ cient to explain the entire positive e¤ect of high school leadership activities on earnings. In this respect, the leadership activities in high school may seem to merely act as observable indicators for leadership skills. Finally, we show that caution is warranted in using selection on observables to address non-random selection in the presence of …xed e¤ects. Unless …xed e¤ects are dealt properly, the results may lead to incorrect inference.

15

References [1] Altonji, Joseph G., Todd E. Elder, and Christopher R. Taber. 2005. “Selection on Observed and Unobserved Variables: Assessing the E¤ectiveness of Catholic Schools.” Journal of Political Economy, Vol. 113, pp. 151-184. [2] Barrick, Murray R., and Michael K. Mount. 1991. “The Big Five Personality Dimensions and Job Performance: A Meta-Analysis.” Personnel Psychology, Vol. 44, No. 1, pp. 1-26. [3] Borghans, Lex, Bas ter Weel, and Bruce A. Weinberg. 2008. “Interpersonal Styles and Labor Market Outcomes” Journal of Human Resources, Vol. 43, pp. 815-858. [4] Bowles, Samuel, Herbert Gintis, and Melissa Osborne. 2001. “The Determinants of Earnings: A Behavioral Approach.” Journal of Economics Literature, Vol. 39, No. 4, pp. 1137-1176. [5] Card, David. 1999. “The Causal E¤ect Education on Earnings.” Handbook of Labor Economics ed. Orley Ashenfelter and David Card, Vol. 3A, pp. 1801-1863, Amsterdam, Elsevier. [6] Carneiro, Pedro, and James J. Heckman. 2003. “Human Capital Policy”in Inequality in America: What Role for Human Capital Policies (ed. James J. Heckman and Alan Krueger), MIT Press. [7] Cavalcanti, Tiago, Juliana Guimaraes, and Breno Sampaio. 2010. “Barriers to Skill Acquisition in Brazil: Public and Private School Students Performance in a Public University Entrance Exam.” The Quarterly Review of Economics and Finance, Vol. 50, No. 4, pp. 395-407. [8] Coleman, Margo, and Thomas DeLeire. 2003. “An Economic Model of Locus of Control and the Human Capital Investment Decision.” Journal of Human Resources, Vol. 38, No. 3, pp.701-721. [9] Cunha, Flavio, James J. Heckman, Lance Lochner, and Dimitriy V. Masterov. 2006. “Interpreting the Evidence on Life Cycle Skill Formation” in Handbook of the Economics of Education (ed. Eric Hanushek and Finis Welch), North Holland. 16

[10] Emran, Shahe M., and Forhad Shilpi. 2008. “Cultural Inheritance, Gender, and Intergenerational Occupational Mobility: Evidence from a Developing Economy.” Unpublished Manuscript. [11] Fortin, Nicole M. 2008. “The Gender Wage Gap among Young Adults in the United States: The Importance of Money versus People.” Journal of Human Resources, Vol. 43, No. 4, pp.884-918. [12] Goldsmith, Arthur H., Jonathan R. Veum, and William Jr. Darity. 1997. “The Impact of Psychological and Human Capital on Wages.” Economic Inquiry, Vol. 35, No. 4, pp. 815-829. [13] Heckman, James J., and Flavio Cunha. 2008. “Formulating, Identifying and Estimating the Technology of Cognitive and Noncognitive Skill Formation”Journal of Human Resources, Vol. 43, pp. 738-782. [14] Heckman, James J., Jora Stixrud, and Sergio Urzua. 2006. “The E¤ects of Cognitive and Noncognitive Abilities on Labor Market Outcomes and Social Behavior.” Journal of Labor Economics, Vol. 24, No. 3, pp. 411-482. [15] Hogan, Joyce, and Brent Holland. 2003. “Using Theory to Evaluate Personality and Job Performance Relations: A Socioanalytic Perspective.” Journal of Applied Psychology, Vol. 88, No. 1, pp. 100-112. [16] Kingdon, Geeta, and Francis Teal. 2010. “Teacher Unions, Teacher Pay and Student Performance in India: A Pupil Fixed E¤ects Approach.” Journal of Development Economics, Vol.91, No. 2, pp.278-288. [17] Krueger, Alan B., and David Schkade. 2008. “Sorting in the Labor Market: Do Gregarious Workers Flock to Interactive Jobs? Journal of Human Resources, Vol. 43, pp. 859-883.

17

[18] Lindqvist, Erik, and Roine Vestman. 2011. “The Labor Market Returns to Cognitive and Noncognitive Ability: Evidence from the Swedish Enlistment.” American Economic Journal: Applied Economics, Vol. 3 (January), pp. 101-128. [19] Kuhn, Peter, and Catherine Weinberger. 2005. “Leadership Skills and Wages.” Journal of Labor Economics, Vol. 23, pp. 395-436. [20] Rosenberg, Morris. 1965. Society and the Adolescent Self-Image. Princeton, NJ., Princeton University Press. [21] Rotter, Julian B. 1966. Generalized Expectancies for Internal versus External Control for Reinforcement. Washington, DC., American Psychological Association. [22] Segal, Carmit. 2011. “Misbehavior, Education and Labor Market Outcomes.”The Journal of the European Economic Association, forthcoming.

18

Table1: Sample Means of Various Data Sources NLS72-KW

NLS72

12.600

12.954

HSB-KW

HSB

26,100

26,884

NELS

Earnings: Hourly Earnings Annual Earnings Leadership: Both Captain and President Captain Only President Only Membership: Both on Team Member and in Club On Team Only In Club Only Math Score (percentile/100) Parents' Education: High School College or Higher Number of Schools Sample Size

14.798

0.065 0.102 0.156

0.060 0.099 0.174

0.121 0.133 0.222

0.114 0.145 0.230

0.112 0.101 0.190

0.427 0.163 0.237 0.465

0.451 0.160 0.234 0.604

0.490 0.111 0.267 0.538

0.508 0.125 0.252 0.463

0.463 0.058 0.301 0.570

0.554 0.312

0.553 0.275

0.601 0.326

0.628 0.290

0.585 0.331

824 3,083

833 3,109

699 2,383

710 2,440

683 1,779

19

Table 2: The Effects of High School Leadership Activities, Various Data Sources NLS72-KW

NLS72

HSB-KW

HSB

NELS

Coefficient (Standard Error) Leadership: Both Captain and President Captain Only President Only Membership: Both on Team and in Club On Team Only In Club Only

Controls: Senior Level Math Test Score Parental Education School Fixed Effects R²

0.122* (0.054) 0.042 (0.040) -0.005 (0.032)

0.126* (0.052) 0.044 (0.039) 0.018 (0.032)

0.231** (0.056) 0.111* (0.048) 0.167** (0.038)

0.241** (0.046) 0.135** (0.038) 0.138** (0.035)

0.142** (0.046) 0.060 (0.047) 0.101** (0.037)

0.094** (0.036) 0.056 (0.039) 0.069 (0.036)

0.092** (0.035) 0.052 (0.038) 0.066 (0.035)

0.017 (0.051) 0.065 (0.064) -0.073 (0.053)

0.026 (0.043) 0.089 (0.052) -0.063 (0.046)

-0.018 (0.040) -0.050 (0.063) -0.055 (0.041)

Yes Yes Yes

Yes Yes Yes

Yes Yes Yes

Yes Yes Yes

Yes Yes Yes

…

0.38

…

0.36

0.42

** Significant at 1%, *Significant at 5%

20

Table 3: Introducing Physical and Psychological Traits into Leadership Earnings Regression (1)

(2)

(3)

0.137** (0.047) 0.056 (0.048) 0.096** (0.037)

0.130** (0.048) 0.037 (0.048) 0.087* (0.038)

0.133** (0.049) 0.036 (0.049) 0.089* (0.038)

-0.036 (0.042) -0.073 (0.064) -0.054 (0.041)

-0.040 (0.041) -0.058 (0.064) -0.063 (0.041)

-0.050 (0.042) -0.073 (0.065) -0.062 (0.042)

0.038 (0.036) -0.018 (0.047) -0.077 (0.045) -0.103 (0.058)

….

0.030 (0.038) -0.011 (0.050) -0.080 (0.049) -0.093 (0.064)

NonCognitive Skills: Somewhat Social

….

Not Social At All

….

Self-Esteem and Locus of Control

….

-0.004 (0.031) -0.077 (0.045) 0.021* (0.010)

0.028 (0.038) -0.027 (0.053) 0.019 (0.013)

0.42 1,779

0.42 1,733

0.43 1,733

Yes Yes Yes

Yes Yes Yes

Yes Yes Yes

Leadership: Both Captain and President Captain Only President Only Membership: Both on Team and in Club On Team Only In Club Only Physical Attractiveness: Somewhat Athletic Not Athletic At All Somewhat Popular Not Popular At All

R² Sample Size

Controls: Senior Level Math Test Score Parental Education School Fixed Effects ** Significant at 1%, *Significant at 5%

21

…. …. ….

Table 4:Descriptive Statistics for the 8th and 12th Grade Leadership Activities Panel A:

8th Grade

12th Grade

Any Officer/Leadership Position

217 (13.9%)

620 (39.7%)

No Officer/Leadership Position

1343 (86%)

940 (60.2%)

Panel B:

12th Grade 8th Grade Both Captain and President

Captain Only

President Only

Both Team and Club Officer

9 (0.5%)

2 (0.1%)

4 (0.2%)

Team Officer Only or Club Officer Only

34 (2%)

32 (2%)

48 (3%)

116 (7.4%)

103 (6.6%)

169 (10.8%)

Both Team and Club Member

NOTES: The eighth and twelfth grade leadership activities are available for 1,560 white males and there are 166 males, who held both a team captain and a club president position in twelfth grade. Percentages are with respect to total sample size.

22

Table 5: Introducing the 8th Grade Activities into Leadership Earnings Regression (1)

(2)

(3)

(4)

(5)

(6)

0.113* (0.050) 0.057 (0.036) 0.024 (0.050)

0.047 (0.086) -0.034 (0.083) 0.083 (0.053)

….

0.038 (0.099) -0.123 (0.076) 0.036 (0.054)

….

0.019 (0.101) -0.112 (0.077) 0.011 (0.056)

0.077* (0.034) 0.027 (0.038) 0.069 (0.038)

0.158** (0.053) 0.108** (0.054) 0.104** (0.055)

….

0.130* (0.053) 0.099 (0.055) 0.090 (0.053)

….

….

….

Captain Only

….

….

President Only

….

….

0.106* (0.052) 0.047 (0.052) 0.103* (0.041)

0.089 (0.053) 0.044 (0.052) 0.089* (0.041)

0.104* (0.051) 0.033 (0.053) 0.101* (0.042)

0.084 (0.055) 0.023 (0.054) 0.077 (0.042)

….

….

On Team Only

….

….

In Club Only

….

….

-0.001 (0.045) -0.062 (0.070) -0.055 (0.045)

-0.014 (0.046) -0.080 (0.070) -0.061 (0.045)

-0.010 (0.046) -0.066 (0.070) -0.058 (0.046)

-0.038 (0.049) -0.091 (0.072) -0.063 (0.046)

R² Sample Size

0.04 1,560

0.44 1,560

0.46 1,560

0.46 1,560

0.45 1,529

0.47 1,529

Yes Yes No No No

Yes Yes Yes No No

Yes Yes Yes No No

Yes Yes Yes No No

Yes Yes Yes Yes Yes

Yes Yes Yes Yes Yes

8th Grade Leadership: Both Team and Club Officer Team Officer Only Club Officer Only 8th Grade Membership: Both on Team and in Club On Team Only In Club Only 12th Grade Leadership: Both Captain and President

12th Grade Membership: Both on Team and in Club

Controls: Senior Level Math Test Score Parental Education School Fixed Effects (8th or 12th Grade) Physical Attributes Psychological Attributes ** Significant at 1%, *Significant at 5%

23

…. ….

…. ….

…. ….

…. ….

0.107* (0.053) 0.087 (0.056) 0.066 (0.055)

Table 6: Implied Ratios for Demeaned Regressions NLS72

HSB

α

0.126

0.241

0.142

0.137

0.130

0.084

(Cov(Leader,ε))/(Var(v))

1.548

0.831

1.048

1.114

1.112

1.105

Implied Ratio

0.081

0.290

0.135

0.122

0.116

0.076

Controls: Senior Level Math Test Score Parental Education Physical Attributes Psychological Attributes Pre-High School Leadership Activities

Yes Yes No No No

Yes Yes No No No

Yes Yes No No No

Yes Yes Yes No No

Yes Yes Yes Yes No

Yes Yes Yes Yes Yes

NOTES: Implied Ratio is the first row divided by the second row.

24

NELS-(1) NELS-(2) NELS-(3) NELS-(4)

Table 7: Implied Ratios with School Fixed Effects NLS72

HSB

α

0.126

0.241

0.142

0.137

0.130

0.084

(Cov(Leader,ε))/(Var(v))

0.188

0.220

0.103

0.130

0.191

0.114

Implied Ratio

0.670

1.100

1.378

1.053

0.680

0.736

Yes Yes No No No

Yes Yes No No No

Yes Yes No No No

Yes Yes Yes No No

Yes Yes Yes Yes No

Yes Yes Yes Yes Yes

Controls: Senior Level Math Test Score Parental Education Physical Attributes Psychological Attributes Pre-High School Leadership Activities

NOTES: Implied Ratio is the first row divided by the second row.

25

NELS-(1) NELS-(2) NELS-(3) NELS-(4)