Nature of Attrition Manajemen | Fakultas Ekonomi Universitas Maritim Raja Ali Haji 536.full

is presented for all known children and three subsamples: 1 children continuously in the sample, 2 children for whom supplemental information is sometimes col- lected, and 3 children for whom supplemental information is never collected. The comparisons are done separately for the full and cross-sectional samples. Although the mothers of children in the continuous sample score higher on the AFQT than the mothers of children who are either sometimes or never in the NLSY-C, along most other dimensions they look similar to the mothers of the children for whom supple- mental information is sometimes collected. The mothers of the children for whom supplemental information is never collected look different in that they are the most likely to have had a child and the least likely to have never been married at the 1979 interview. Overall, children who are surveyed in some of the child supplement years in which they are eligible are similar to those who are surveyed in all eligible years. The mothers of these two groups of children also appear similar. However, the children who are not interviewed in any of the mother-child supplement years and their moth- ers are different. These mothers more often experienced an early marriage or birth. Attrition from the NLSY79 prior to the start of the child supplement appears nonran- dom with respect to women’s fertility through 1979 and marital status at that time. Evaluating the effects that this attrition has on the sample of children is made difŽcult because there is no information on children whose mothers attrite before their births, but the fact that there are not many of these children minimizes the chance that their absence will have large consequences.

III. Nature of Attrition

This section further examines how attritors are different from nonat- tritors by estimating two types of equations. The samples for these equations are restricted to children who have been born in 1994 or earlier and have not attrited prior to having been assessed for the Žrst time on the PPVT eligible at age three or the BPI eligible at age four depending on the assessment of interest. The Žrst of these equations is a probit that estimates the probability that an individual is an attritor at a future NLSY79-C interview. 1 P [A i 2000 5 1] 5 FX i b l 1X it b 2 where A i2000 indicates that the individual attrites by the year 2000 interview, X is the general notation for the covariates included in the analysis, some of which are perma- nent characteristics of the child or her mother and others are measured at age t in the survey where the child is Žrst age-eligible for the assessment, and b are the parameters estimated to gauge the extent to which attritors come from certain seg- ments of the population. The equation is estimated for the full and cross-sectional samples of the children in the NLSY79-C. Four speciŽcations are estimated; they focus on family income early in the child’s life measured by average income over the child’s Žrst three years in 1984 or maternal employment in the child’s Žrst year measured by hours worked during the child’s Žrst year and a dummy variable indicating the mother worked at all in the child’s Žrst year and contain either BPI or PPVT standard scores at the Žrst survey in which the child was age-eligible for the assessment. By including individual characteristics, one can test whether children who attrite come from particular backgrounds. 7 Table 6 presents coefŽcient estimates and marginal effects for Equation 1. The full sample results indicate that Black and Hispanic children are more likely than other children to attrite. The children whose mothers have never married are less likely to attrite compared to those whose mothers are married, separated, or divorced. The probability of attrition increases in average family income during the child’s Žrst three years. In the speciŽcations that include family income, the probability of future attrition is decreasing in the educational attainment of one’s maternal grand- father by 0.6 of a percentage point for each additional grade completed by one’s grandfather. Moreover, in the speciŽcation that includes early maternal employment and BPI scores, attrition appears nonrandom with respect to BPI scores. As was the case in the full sample, in the cross-section a child from higher-income family or whose mother is divorced or separated or married is more likely to attrite, and the effects remain small. However, attrition appears less connected to character- istics of the children when the cross-section is examined separately, implying that the process of attrition for the cross-section may be different than that in the Black and Hispanic supplements. Although attrition is related to some of the characteristics included in the probit equations on attrition, the psuedo-R 2 for these equations is quite low, ranging from 0.012 to 0.034 which indicates that much of the attrition is not associated with the variables controlled for in these speciŽcations. Attrition appears random with respect to maternal employment in the child’s Žrst year of life, but is related to family income early in the child’s life. The second type of equation examines the effect of attrition on the estimates of the effect of family characteristics measured early in the child’s life on his BPI and PPVT standard scores. 2 AS it 5 a 1 rFC iE 1 b 1 Z i 1 b 2 Z it 1 e it where AS it is child i’s assessment standard score on either the PPVT or BPI and t is the Žrst interview at which the child is age-eligible for that assessment, FC iE is a characteristic of the child’s parent or family measured early in the child’s life at time E, Z are the additional covariates controlled for, e it is a random error, and r is the coefŽcient on the family characteristics of interest. These equations are esti- mated using the same samples that are used in the attrition probits above, as well as, the subsample of nonattritors. The differences in the estimates of r for the sample with and without the attritors are then compared to see how attrition impacts the estimated relationship. Four models are estimated for each sample and subsample: 1 Model 1 contains no additional covariates, 2 Model 2 includes a dummy vari- able for race, a dummy variable for ethnicity, child’s gender, child’s age in months, and highest grade completed by each of the child’s maternal grandparents in the Vector Z, 3 Model 3 adds a set of dummy variables indicating child’s year of birth 7. This analysis is similar to Mott 1998 where the probability of having a valid PIAT-math or BPI score at the next child-supplement interview is examined conditional on having a valid score at the current interview. A u ghinba ugh 551 Table 6a Attrition Probit, Child Ever Out: Full Sample CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal Estimate Effect Estimate Effect Estimate Effect Estimate Effect Black 0.348 0.119 0.379 0.132 0.312 0.107 0.338 0.118 0.061 0.062 0.071 0.073 Hispanic 0.497 0.175 0.520 0.186 0.468 0.167 0.520 0.188 0.072 0.074 0.080 0.082 Mother’s age 1979 20.013 20.004 20.012 20.004 20.023 20.008 20.015 20.005 0.018 0.019 0.021 0.021 Mother’s highest grade 1979 0.001 0.000 0.006 0.002 0.013 0.004 0.010 0.003 0.019 0.020 0.022 0.023 Number of children 1979 20.093 20.030 20.106 20.035 20.053 20.018 20.038 20.013 0.060 0.062 0.066 0.066 Mother’s marital status Married 0.177 0.061 0.213 0.074 0.194 0.067 0.192 0.068 0.081 0.083 0.089 0.093 Divorcedseparated 0.417 0.151 0.493 0.183 0.404 0.148 0.432 0.160 0.194 0.210 0.213 0.218 Male child 0.083 0.027 0.070 0.023 0.079 0.026 0.056 0.019 0.038 0.040 0.044 0.046 The Journa l of Hum an Re so ur ce s Table 6a continued CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal Estimate Effect Estimate Effect Estimate Effect Estimate Effect Grandmother’s highest grade 0.005 0.002 0.004 0.001 0.002 0.001 0.005 0.002 0.011 0.011 0.012 0.012 Grandfather’s highest grade 20.018 20.006 20.017 20.006 20.019 20.006 20.018 20.006 0.009 0.009 0.009 0.010 PPVT score 20.000 20.000 0.000 0.000 0.001 0.001 BPI score 20.003 20.001 20.003 20.001 0.002 0.002 Average income in Žrst three years 0.012 0.004 0.016 0.005 10,000 0.006 0.006 Maternal hours worked 0.053 0.018 0.069 0.023 In Žrst year 1,000 hours 0.042 0.045 Work Žrst year 20.084 20.028 20.044 20.015 0.073 0.080 Number of observations 5,080 4,582 3,872 3,493 Psuedo R-squared 0.030 0.034 0.031 0.034 Notes: Samples are composed of children born in 1994 or earlier and who do not attrite prior to Žrst assessment on either the BPI or PPVT. PPVT and BPI scores are standardized for age of the child. indicates signiŽcance at the 1 percent level and at the 5 percent level. Standard errors are clustered by mother id and are presented in parentheses. Regressions also include two dummy variables indicating grandfather’s education and grandmother’s education are missing. A u ghinba ugh 553 Table 6b Attrition Probit, Child Ever Out: Cross-Section CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal Estimate Effect Estimate Effect Estimate Effect Estimate Effect Black 0.008 0.002 0.068 0.020 0.018 0.005 0.036 0.011 0.103 0.104 0.116 0.119 Hispanic 20.002 20.001 0.059 0.017 0.073 0.022 0.145 0.045 0.129 0.134 0.136 0.139 Mother’s age 1979 0.007 0.002 0.005 0.001 0.009 0.003 0.014 0.004 0.027 0.028 0.029 0.030 Mother’s highest grade 1979 20.029 20.008 20.016 20.005 20.032 20.009 20.023 20.007 0.030 0.031 0.032 0.034 Number of children 1979 20.074 20.021 20.086 20.025 20.081 20.024 20.037 20.011 0.086 0.091 0.092 0.092 Mother’s marital status Married 0.157 0.046 0.213 0.065 0.209 0.065 0.207 0.065 0.104 0.107 0.113 0.117 Divorcedseparated 0.635 0.218 0.642 0.222 0.685 0.243 0.667 0.237 0.224 0.244 0.243 0.249 Male child 0.059 0.017 0.069 0.020 0.014 0.004 20.000 20.000 0.053 0.055 0.057 0.061 The Journa l of Hum an Re so ur ce s Table 6b continued CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal CoefŽcient Marginal Estimate Effect Estimate Effect Estimate Effect Estimate Effect Grandmother’s highest grade 20.010 20.003 20.010 20.003 20.022 20.007 20.017 20.005 0.017 0.017 0.018 0.019 Grandfather’s highest grade 20.008 20.002 20.005 20.001 20.003 20.001 20.004 20.001 0.012 0.013 0.013 0.013 PPVT 20.002 20.001 20.001 20.000 0.002 0.002 BPI 20.001 20.000 20.002 20.001 0.002 0.002 Average income in Žrst three years 0.014 0.004 0.018 0.005 10,000 0.007 0.008 Maternal hours worked 20.004 20.001 20.024 20.007 In Žrst year 1,000 hours 0.055 0.061 Work Žrst year 20.067 20.019 20.015 20.005 0.095 0.103 Number of observations 3,139 2,820 2,417 2,155 Pseudo R-squared 0.012 0.014 0.019 0.020 Notes: Samples are composed of children born in 1994 or earlier and who do not attrite prior to Žrst assessment on either the BPI or PPVT. PPVT and BPI scores are standardized for age of the child. indicates signiŽcance at the 1 percent level and at the 5 percent level. Standard errors are clustered by mother id and are presented in parentheses. Regressions also include two dummy variables indicating grandfather’s education and grandmother’s education are missing. to the variables included in Model 2, and 4 Model 4 adds mother’s AFQT score to those variables in Model 3. As Fitzgerald, Gottschalk, and MofŽtt 1998 note, some attrition will have oc- curred before the child reaches age t and consequently the estimates may already be biased. To test whether further attrition is biasing, Fitzgerald, Gottschalk, and MofŽtt assume that attrition pre- and post- age t biases the coefŽcient estimates in the same direction. This assumption is sufŽcient to permit inference that there is attrition bias in the sample if attrition bias is found in the post-period. I also make this assumption. Table 7 examines how attrition impacts estimation of the effect of average income during the child’s Žrst three years measured in units of 10,000 1984 on PPVT and BPI standard scores. The top half presents coefŽcient estimates for the full sam- ple while the bottom half present the estimates for the cross-section. As in past literature that has examined the effect of family income on child devel- opment Blau 1999, Mayer 1997, income tends to improve test scores by a small, but statistically signiŽcant amount. In general, the magnitude of the family income effect is slightly larger among the nonattritors; however, in no case is the difference between the coefŽcient estimates using all observations and using the nonattritors signiŽcantly different from zero. 8 Table 8 examines the impact of attrition on estimates of the effect of maternal employment on children’s PPVT and BPI scores. Hours of work in units of 1,000 hours during the child’s Žrst year and a dummy variable indicating whether the mother worked during the child’s Žrst year are used to measure maternal employ- ment. 9 In Panel A, the coefŽcient estimates for the maternal employment variables are presented for the full sample for both all children and for the nonattritors. Panel B presents this same information for the cross-sectional sample. The estimates pre- sented here are in accord with a large number of studies that use the NLSY79-C and Žnd small effects of maternal employment on early PPVT and BPI scores for example, James-Burdumy 1999 and Harvey 1999. 10 Although in the full sample hours worked in the child’s Žrst year do not have a signiŽcant impact on PPVT scores, the dummy variable indicating that the mother worked implies that it is beneŽcial for the child to have a mother who works during his Žrst year of life. These beneŽts decline as controls are added, which seems to imply positive selection of women into the labor force during their children’s Žrst year. In Models 1 and 2, additional hours of work by the mother appear to reduce behavioral problems and the dummy variable for working in the child’s Žrst year has 8. The standard errors are calculated using the fact that the variance of the difference in the coefŽcients for the total sample and the nonattriting subsample is equal to the difference in the variances. See Fitzgerald, Gottschalk, and MofŽtt 1998 and reference therein for more detail. 9. In much of the literature on the impact of maternal employment on child development, maternal employ- ment is measured by hours worked in the child’s Žrst three years measured with three variables, one for each year. Owing to the high correlation between hours worked in these years, changes in either the sample used for estimation or the speciŽcation can have large effects on all three of the coefŽcients estimates on hours worked that may balance each other out. Consequently, to study the impact of attrition, I measure early maternal employment using only employment in the child’s Žrst year. 10. A larger detrimental effect from early maternal employment is found when either the outcomes of white children Han and Waldfogel 2001 or achievement test scores for 5- and 6- year-olds Ruhm 2001 are examined. The Journa l of Hum an Re so ur ce s Table 7 Estimates of Effect of Average Family Income in First 3 Years on Child Assessments PPVT BPI Full Sample Nonattritors All Nonattritors All n 5 3,872 n 5 2,782 Difference n 5 5,080 n 5 3,695 Difference Model 1 0.957 1.012 0.055 20.440 20.487 0.047 0.160 0.218 0.148 0.064 0.075 0.039 [0.030] [0.030] [0.014] [0.015] Model 2 0.247 0.236 0.011 20.320 20.345 0.025 0.095 0.120 0.073 0.058 0.067 0.034 [0.321] [0.342] [0.033] [0.037] Model 3 0.228 0.235 0.007 20.221 20.237 0.016 0.097 0.124 0.077 0.054 0.062 0.030 [0.325] [0.345] [0.060] [0.068] Model 4 0.061 0.027 0.034 20.175 20.182 0.007 0.089 0.114 0.071 0.052 0.062 0.034 [0.357] [0.379] [0.064] [0.072] A u ghinba ugh 557 Cross-Section All n 5 2,417 n 5 1,883 Difference n 5 3139 n 5 2,496 Difference Model 1 0.663 0.762 0.099 20.400 20.440 0.040 0.152 0.203 0.135 0.073 0.084 0.042 [0.020] [0.021] [0.014] [0.014] Model 2 0.144 0.202 0.058 20.274 20.302 0.028 0.093 0.114 0.066 0.064 0.074 0.037 [0.291] [0.295] [0.041] [0.038] Model 3 0.135 0.193 0.058 20.194 20.203 0.009 0.097 0.121 0.072 0.061 0.069 0.032 [0.296] [0.301] [0.071] [0.074] Model 4 0.007 0.022 0.015 20.165 20.165 0.000 0.091 0.114 0.069 0.060 0.070 0.036 [0.327] [0.333] [0.074] [0.077] Notes: Samples are composed of children born in 1994 or earlier and who do not attrite prior to Žrst assessment on either the BPI or the PPVT. PPVT and BPI scores are standardized for age of the child. Income is measured in units of 10,000 1984 . Model 1 includes no additional controls, Model 2 includes a dummy variable for race, a dummy variable for ethnicity, child’s gender, child’s age in months, highest grade completed by each of the child’s maternal grandparents, and dummy variables indicating grandparents’ highest grade is missing in the Vector X, Model 3 includes a set of dummy variables indicating child’s year of birth and the variables included in Model 2, and Model 4 includes mother’s AFQT score and those variables in Model 3. indicates signiŽcance at the 1 percent level and at the 5 percent level. Standard errors are clustered by mother id and are presented in parentheses. R-squared’s are presented in square brackets. The Journa l of Hum an Re so ur ce s Table 8a Estimates of Effect of Maternal Employment in First Year on Child Assessments: Full Sample PPVT BPI Nonattritors All Nonattritors All n 5 3,493 n 5 2,478 Difference n 5 4,582 n 5 3,289 Difference Model 1 Hours 21.055 21.270 0.215 21.009 21.402 0.393 0.649 0.764 0.403 0.416 0.471 0.221 Any work 8.909 9.007 0.098 20.651 20.179 0.472 1.195 1.359 0.647 0.725 0.828 0.400 [0.029] [0.028] [0.006] [0.008] Model 2 Hours 0.128 20.497 0.625 21.139 21.489 0.350 0.543 0.622 0.303 0.411 0.466 0.220 Any work 3.540 3.632 0.092 0.441 0.933 0.492 0.976 1.086 0.476 0.726 0.830 0.402 [0.331] [0.347] [0.033] [0.038] A u ghinba ugh 559 Nonattritors All n 5 3,493 n 5 2,478 Difference n 5 4,582 n 5 3,289 Difference Model 3 Hours 0.130 20.401 0.531 20.467 20.785 0.318 0.577 0.641 0.279 0.412 0.469 0.224 Any work 3.536 3.702 0.166 0.151 0.591 0.440 0.986 1.098 0.483 0.718 0.825 0.406 [0.335] [0.352] [0.061] [0.069] Model 4 Hours 20.169 20.630 0.461 20.384 20.725 0.341 0.543 0.620 0.299 0.413 0.471 0.226 Any work 2.294 2.225 0.069 0.458 0.959 0.501 0.986 1.078 0.436 0.717 0.827 0.412 [0.363] [0.383] [0.066] [0.075] Notes: Samples are composed of children born in 1994 or earlier and who do not attrite prior to Žrst assessment on either the BPI or the PPVT. PPVT and BPI scores are standardized for age of the child. Hours are measured in units of 1,000. Model 1 includes no additional controls, Model 2 includes a dummy variable for race, a dummy variable for ethnicity, child’s gender, child’s age in months, highest grade completed by each of the child’s maternal grandparents, and dummy variables indicating grandparents’ highest grade is missing in the Vector X, Model 3 includes a set of dummy variables indicating child’s year of birth and the variables included in Model 2, and Model 4 includes mother’s AFQT score and those variables in Model 3. indicates signiŽcance at the 1 percent level and at the 5 percent level. Standard errors are clustered by mother id and are presented in parentheses. R-squared’s are presented in square brackets. The Journa l of Hum an Re so ur ce s Table 8b Estimates of Effect of Maternal Employment in First Year on Child Assessments: Cross-Section PPVT BPI All Nonattritors All Nonattritors n 5 2,155 n 5 1,670 Difference n 5 2,820 n 5 2,225 Difference Model 1 Hours 20.773 20.965 0.192 20.483 20.991 0.508 0.762 0.849 0.374 0.521 0.554 0.188 Any work 6.407 6.110 0.297 20.850 20.227 0.623 1.366 1.538 0.707 0.890 0.967 0.378 [0.018] [0.015] [0.003] [0.004] Model 2 Hours 0.002 0.041 0.039 20.551 21.057 0.506 0.648 0.712 0.295 0.511 0.548 0.198 Any work 2.332 1.913 0.419 0.033 0.540 0.507 1.126 1.240 0.519 0.892 0.972 0.386 [0.288] [0.282] [0.039] [0.038] A u ghinba ugh 561 All n 5 2,155 n 5 1,670 Difference n 5 2,820 n 5 2,225 Difference Model 3 Hours 0.135 0.200 0.065 0.166 20.313 0.479 0.666 0.739 0.320 0.516 0.554 0.202 Any work 2.198 1.859 0.339 20.226 0.266 0.492 1.130 1.246 0.525 0.884 0.967 0.392 [0.294] [0.291] [0.071] [0.072] Model 4 Hours 20.156 20.024 0.132 0.221 20.267 0.488 0.647 0.716 0.307 0.516 0.555 0.204 Any work 1.170 0.766 0.404 20.008 0.509 0.517 1.106 1.227 0.531 0.882 0.967 0.396 [0.322] [0.322] [0.075] [0.076] Notes: Samples are composed of children born in 1994 or earlier and who do not attrite prior to Žrst assessment on either the BPI or the PPVT. PPVT and BPI scores are standardized for age of the child. Hours are measured in units of 1,000. Model 1 includes no additional controls, Model 2 includes a dummy variable for race, a dummy variable for ethnicity, child’s gender, child’s age in months, highest grade completed by each of the child’s maternal grandparents, and dummy variables indicating grandparents’ highest grade is missing in the Vector X, Model 3 includes a set of dummy variables indicating child’s year of birth and the variables included in Model 2, and Model 4 includes mother’s AFQT score and those variables in Model 3. indicates signiŽcance at the 1 percent level and at the 5 percent level. Standard errors are clustered by mother id and are presented in parentheses. R-squared’s are presented in square brackets. no effect. In both the PPVT and BPI equations, the differences between the coefŽcient estimates on the maternal employment variables when the attritors are included in the estimation and when they are not included are small and never signiŽcant. When these equations are estimated using the cross-sectional sample, the coefŽ- cient estimates on the maternal employment variables are rarely signiŽcant. For all four speciŽcations of BPI standard scores, the differences in the coefŽcient estimates on hours worked in the child’s Žrst year are signiŽcant. Though the coefŽcient esti- mates themselves are not signiŽcantly different from zero in three of the four speci- Žcations, exclusion of the attritors makes it appear that hours worked by the mother in the child’s Žrst year are more effective in decreasing behavioral problems com- pared to those estimates that include the attritors. This may imply that the intensity of work during a child’s Žrst year varies systematically by future attrition with char- acteristics that are not controlled for here. I also estimate Equations 1 and 2 using an older sample that employs the assess- ment scores from two years later and requires children not to have attrited up to that point. The sample sizes are much smaller and hence it is not surprising that almost none of the regressors in the attrition probits are signiŽcant. However, family income remains signiŽcantly negative in the probit that includes PPVT. In the estimates of Equation 2 that use the older children, the differences in the estimates of r between samples that include and exclude the attritors are never signiŽcant. Compared with the results presented in Tables 7 and 8, the computed differences and their standard errors are approximately an order of magnitude larger.

IV. Conclusion