Ability, Schooling Inputs and Earnings: Evidence from the NELS Ozkan Eren University of Nevada, Las Vegas June 2008
Introduction I The earnings dispersion among individuals for a given age, education level, gender and race has increased substantially in U.S. over the past couple of decades. I Growing importance of productive skills in the labor market (Juhn et al. 1993). I Traditionally, the primary example of productive skills is cognitive ability (measured by knowledge and aptitude tests). I Viewing cognitive skills as the sole/primary aspect of productive skills, however, may be misleading.
Introduction I Green et al. (1998)-British Employers Manpower and Skills Practices. I 1998 survey by U.S. Census Bureau in collaboration with the Department of Education. I Sociology and psychology literature (i.e., Barrick and Mount 1991, Hogan and Holland 2003).
Introduction I Goldsmith (1997) using the NLSY data (self-esteem). I Bowles et al. (2001) with di erent data sets (self-esteem, optimism and aggression). I Coleman and DeLeire (2003) using the NELS data (locus of control). I Cuhna et al. (2006) and Heckman et al. (2006) using the NLSY data (self-esteem and locus of control).
Introduction I Primary focus is on a single measure of central tendency, the conditional mean. I Uninformative about ability (cognitive or noncognitive) when the e ects are heterogeneous (construction worker vs manager). I Self-rated structure of noncognitive ability-measurement error (i.e., Borghans et al. 2008).
Introduction I The analysis in this paper is based on the NELS data. I The NELS includes subject test scores as well as the self-esteem and locus of control scales, which constitutes our measure of noncognitive ability. I In addition to conditional mean, we estimate the e ects of pre-market cognitive and noncognitive ability over the earnings distribution controlling for the measurement error in the latter. I Our distributional approach is based on the (instrumental) quantile regression. I Apart from the ability focus of the paper, the availability of the schooling inputs in the NELS data.
Empirical Findings I The eighth grade noncognitive ability is an important determinant of earnings at age 26-27. However, the e ects are not homogeneous throughout the distribution. I Substantial measurement error inherent in noncognitive ability. I The impact of eighth grade cognitive ability works mainly through the educational channels (no direct e ect on earnings). I Unidimensional vision of ability is a faulty one. I The eighth grade pupil-teacher ratio has a negative and signi cant impact on earnings. However, similar to noncognitive ability, the e ects are not uniform.
Empirical Methodology w = αnc + δc + Γ 0 β + ε w - (log weekly) earnings NC - noncognitive ability (measured with error) C - cognitive ability Γ - individual, family and schooling characteristics ε -mean zero, possibly heteroskedastic, normally distributed error term. I OLS and IV for the conditional mean. I QR and IQR (Chernozhukov and Hansen 2006, 2008) for the distributional e ects.
Data I The data comes from the National Educational Longitudinal Study (NELS) of 1988, a large longitudinal study of eighth grade students. I Follow-up surveys were administered in 1990 (10th grade), 1992 (12th grade), 1994 and 2000. I The respondents were administered cognitive tests in reading, social science, mathematics and science during the spring of the base year, rst and second follow-up. I We use the NELS constructed eight grade standardized (mean of zero and standard deviation of one) composite mathematics and reading test score as the measure of cognitive ability.
Data I For noncognitive ability, we utilize Rosenberg Self-Esteem and Rotter Locus of Control scales. I The Rosenberg scale refers to the perceptions of self-esteem (Rosenberg 1966). I The Rotter scale refers to the extent to which individuals believe that they can control outcomes that a ect them (Rotter 1965).
Data Rosenberg Self-Esteem Scale I I feel good about myself; I I feel I am a person of worth, the equal of other; I I am able to do things as well as most other people; I On the whole, I am satis ed with myself; I I feel useless at times; I At times I think I am no good at all; I I feel I do not have much to be proud of.
Data Rotter Locus of Control Scale I I do not have enough control over the direction my life is taking; I In my life, good luck is more important than hard work for success; I Every time I try to go ahead, something or somebody stops me; I On the whole, I am satis ed with myself; I I feel useless at times; I At times I think I am no good at all; I I feel I do not have much to be proud of.
Data I Self-esteem and locus of control scales were measured on a four point Likert scale ranging from strongly agree to strongly disagree. I Similar to cognitive ability, the NELS constructed composite measures, which constitutes the Rosenberg and Rotter scales. I We use the eighth grade standardized (mean of zero and standard deviation of one) average of these two scales as our measure of noncognitive ability. I With respect to schooling inputs, we focus on the pupil-teacher ratio of the school and school type variables (i.e., public, Catholic...).
Data I The dependent variable is log weekly earnings. I We restrict our analysis solely to young men who were not enrolled in school at the time of the interview, who reported working at least 25 weeks in 1999 and were not self-employed.
Data I The set of controls include the following variables: Individual: race, region, educational attainment; Family: father s education, mother s education, parents marital status, socioeconomic status of the family, family size, family income, indicators for home reading materials (books and daily newspaper), indicator for a home computer; School: indicators for school type, pupil-teacher ratio, percentage of students from single parent homes, percentage of minority students, percentage of students receiving free lunch, urban/rural status and region. I Information on family and schooling variables come from the base year survey questionnaires and data pertaining to individual characteristics are obtained from the fourth-follow up survey.
Estimation Problems I The use of pre-labor market measures of cognitive and noncognitive ability allows us to avoid the reverse causality problem (i.e., the possibility that earnings develop self-esteem). I The cohort e ects contamination is ruled out by the very nature of the NELS data (i.e., NLSY). I The endogeneity issues may arise due to omission of input variables that a ect both the outcome variable and the respective schooling inputs. To overcome (or at least to substantially reduce) the potential biases of schooling inputs, we follow Dearden et al. (2002) and Dustmann et al. (2003) and utilized a lengthy vector of family background and schooling characteristics. I The self rated structure of noncognitive ability-measurement error problem. IV estimation.
Evaluation of the Instrument I The most common solution to measurement error is the use of instrumental variable estimation. I Suitable instruments or repeated observation of the variable measured with error. I The panel structure of the NELS data allows us to observe the noncognitive ability at multiple points of time. I We use the standardized tenth grade Rosenberg and Rotter scales as instruments for eighth grade noncognitive ability.
Evaluation of the Instrument I A regression, controlling for all the other covariates, of eighth grade noncognitive ability on tenth grade Rosenberg and Rotter scales yields a partial R 2 and F -statistics as 0.199 and 266.22, respectively. I The overidenti cation test with Hansen s J-statistics yields a p-value=0.25.
Table 1: OLS Estimations Specification (1) (2) (3) (4) (5) (6) 8th Grade ncognitive Ability 0.060*** (0.007) 0.045*** (0.008) 0.047*** (0.008) 0.042*** (0.008) 0.042*** (0.008) 8th Grade Cognitive Ability.. 0.060*** (0.008) 0.054*** (0.008) 0.023*** (0.008) 0.021*** (0.008) Pupil Teacher Ratio........ 0.0043** (0.0021) 0.038*** (0.008) 0.006 (0.009) 0.0037* (0.0020) Other Controls: Individual Family Schooling Education
Table 2: IV Estimations Specification (1) (2) 8th Grade ncognitive Ability 0.095*** (0.017) 8th Grade Cognitive Ability 0.008 (0.009) Pupil Teacher Ratio 0.0046** (0.0021) 0.085*** (0.017) 0.002 (0.009) 0.0041** (0.0020) Other Controls: Individual Family Schooling Education
Conditional Mean Findings I The eighth grade noncognitive ability is an important determinant of earnings for young men. I Taken at the face value, the labor market seems to put more weight on noncognitive ability than it does for cognitive ability. I Substantial measurement error inherent in noncognitive ability. I The relation between cognitive ability and earnings seems to mainly work through the educational channels (no direct e ect on earnings). I A unidimensional vision of ability may be a faulty one. I On average, we nd a negative and signi cant impact of pupil-teacher ratio and no evidence on the e ectiveness of the school type variables on young men s earnings.
Table 3: Traditional Quantile Regression Estimations Panel A θ=0.20 θ=0.40 θ=0.50 θ=0.60 θ=0.80 8th Grade ncognitive Ability 0.059*** 0.042*** 0.036*** 0.035*** 0.029*** (0.011) (0.010) (0.010) (0.010) (0.011) 8th Grade Cognitive Ability 0.011 0.018 0.030*** 0.029*** 0.029*** (0.015) (0.011) (0.011) (0.010) (0.011) Pupil Teacher Ratio 0.0034 0.0034 0.0042* 0.0046* 0.0037 (0.032) (0.0027) (0.0025) (0.0025) (0.0028) Panel B θ=0.20 θ=0.40 θ=0.50 θ=0.60 θ=0.80 8th Grade ncognitive Ability 0.050*** 0.037*** 0.032*** 0.035*** 0.027** (0.012) (0.010) (0.009) (0.010) (0.011) 8th Grade Cognitive Ability 0.005 0.001 0.004 0.009 0.017 (0.016) (0.011) (0.011) (0.010) (0.012) Pupil Teacher Ratio 0.0018 0.0032 0.0041* 0.0048* 0.0027 (0.0033) (0.0026) (0.0025) (0.0026) (0.0028)
Table 4: Instrumental Quantile Regression Estimations Panel A θ=0.20 θ=0.40 θ=0.50 θ=0.60 θ=0.80 8th Grade ncognitive Ability 0.100*** 0.110*** 0.090*** 0.080*** 0.080*** (0.024) (0.024) (0.026) (0.021) (0.023) 8th Grade Cognitive Ability 0.016 0.001 0.013 0.015 0.025** (0.015) (0.013) (0.013) (0.011) (0.013) Pupil Teacher Ratio 0.0044 0.0035 0.0058** 0.0040 0.0039 (0.0030) (0.0029) (0.0027) (0.0025) (0.0027) Panel B θ=0.20 θ=0.40 θ=0.50 θ=0.60 θ=0.80 8th Grade ncognitive Ability 0.090*** 0.090*** 0.080*** 0.070*** 0.070*** (0.027) (0.023) (0.022) (0.020) (0.024) 8th Grade Cognitive Ability 0.017 0.007 0.002 0.001 0.014 (0.016) (0.012) (0.012) (0.011) (0.013) Pupil Teacher Ratio 0.0026 0.0030 0.0049* 0.0046* 0.0027 (0.0033) (0.002) (0.0026) (0.0025) (0.0028)
Conclusions I The noncognitive ability is an important determinant of earnings even after controlling for educational attainment. However, the e ects are not homogeneous throughout the distribution. We nd noncognitive ability to be most e ective for low earners. I Cognitive ability, on the other hand, does not yield any e ect either at the mean or at the distributional level once we augment the educational controls. The association between cognitive ability and earnings mainly work through the educational channels. I The labor market seems to put more weight on noncognitive ability and a unidimensional vision of ability may be misleading.
Conclusions I On average, our ndings indicate a negative and signi cant impact of pupil-teacher ratio on earnings. However, similar to noncognitive ability, we observe heterogeneity in the coe cient estimates of the pupil-teacher ratio. I The noncognitive ability accompanied with the pupil-teacher ratio estimates suggest that the conditional mean obscures some important information.
Policy Implications I The educational/social policy interventions during early childhood or adolescence aiming to alter noncognitive rather than the cognitive ability may be more e ective to combat adverse labor market and educational controls. I For economic success, focusing on the ways to boost noncognitive ability may be a more e ective tool than simply pouring more funds into schools to lower the pupil-teacher ratio (premature in the absence of a detailed cost-bene t analysis).