First survey of twins in China Data were collected by Junsen Zhang in collaboration with the national bureau of Statistics in 5 cities. 2002 Adult twins between 18-65 Sampling rather representative: twins identified through various channels such as advertising neighborhood notices, public security bureau etc a Questionnaires filled by face-to-face interviews We got 810 pairs of twins with complete Information for our study Also surveyed non-twins for comparison
11 First survey of twins in China ◼ Data were collected by Junsen Zhang in collaboration with the National Bureau of Statistics in 5 cities, 2002 ◼ Adult twins between 18-65 ◼ Sampling rather representative: twins identified through various channels such as advertising, neighborhood notices, public security bureau etc. ◼ Questionnaires filled by face-to-face interviews ◼ We got 810 pairs of twins with complete information for our study ◼ Also surveyed non-twins for comparison
Data summary Generally twins sample comparable to other sample Tables-education pdf There is a large within-twin-pair variation of education 48% of the twins having within-twin difference in education 24%having a difference of 2 years or more
12 Data summary ◼ Generally, twins sample comparable to other sample Tables-education.pdf ◼ There is a large within-twin-pair variation of education ◼ 48% of the twins having within-twin difference in education ◼ 24% having a difference of 2 years or more
Results returns to education OLS estimate: 8.4% Tables-education pdf Within-twin-pair Fe/Gls estimates: 2.70 a Our results differ from twins studies of other countries--Tables-education pdf a True return in China is lower Selection effect in China is larger 13
13 Results: returns to education ◼ OLS estimate: 8.4% Tables-education.pdf ◼ Within-twin-pair FE/GLS estimates: 2.7% ◼ Our results differ from twins studies of other countries—Tables-education.pdf ◼ True return in China is lower ◼ Selection effect in China is larger
Potential biases of within-twin pair estimates(1) Within-twin-pair difference may not completely remove the omitted ability bias: Within-twin-pair difference in ability may still be correlated with the error term However, within-twin-pair estimates may be less biased than OLs estimates, and thus establish an upper bound We examine this by checking the correlations between education and other ability variables According to ashenfelter and rouse(1998), we have confidence if within-twin correlations are smaller than between-family correlations Tables-education pdf Indeed, within-twin-pair estimates establish an upper bound for the true returns to education, which is 2.7% in our case 14
14 Potential biases of within-twinpair estimates (1) ◼ Within-twin-pair difference may not completely remove the omitted ability bias: Within-twin-pair difference in ability may still be correlated with the error term ◼ However, within-twin-pair estimates may be less biased than OLS estimates, and thus establish an upper bound ◼ We examine this by checking the correlations between education and other ability variables ◼ According to Ashenfelter and Rouse (1998), we have confidence if within-twin correlations are smaller than between-family correlationsTables-education.pdf ◼ Indeed, within-twin-pair estimates establish an upper bound for the true returns to education, which is 2.7% in our case
Potential biases of within-twin pair estimates(2) Measurement error: people may miSs-report equcation Causes the return to be underestimated Correction: use the iv method designed bv Ashenfelter and Krueger(1994) Using cross-reported education as anⅠⅤ Tables-education.pdf 15
15 Potential biases of within-twinpair estimates (2) ◼ Measurement error: people may miss-report education ◼ Causes the return to be underestimated ◼ Correction: use the IV method designed by Ashenfelter and Krueger (1994) ◼ Using cross-reported education as an IV ◼ Tables-education.pdf