Multivariate results Directory UMM :Data Elmu:jurnal:E:Economics of Education Review:Vol20.Issue2.2001:

156 B.J. Surette Economics of Education Review 20 2001 151–163 A final plausible explanation for women’s higher two- year attendance rates and lower transfer rates is that women simply prefer two-year colleges to four-year col- leges for reasons that are not directly observable. For example, it is well known that women and men tend to end up in different occupations. If the female-dominated occupations require training at two-year colleges, and the male-dominated occupations require training at four-year colleges, one might expect women and men to make dif- ferent schooling decisions. We would like to control for such preferences to test this explanation, but we cannot directly observe them. However, one may be able to infer them based on post-schooling occupational outcomes. We estimate the attendance model both with and without proxies for these preferences. Specifically, we model schooling decisions using indicator variables for whether or not an individual ever works in several specific female-dominated occupations. 13 3.2. The transfer model The attendance model provides partial information about whether and why women and men differ in their college attendance patterns and their rates of transfer to four-year college. This section models the decision to transfer directly. We define “transfers” as individuals who have attended two-year college at some point in their lives and have subsequently attended a four-year college. Only the last observation of each individual is used in this model: for an individual observed in all 12 waves of the NLSY, she is a “transfer” if she attended a two-year college and subsequently attended a four-year college at some point prior to 1990. The transfer model is based on the same human capital theory that motivates the attendance model. The main distinction is that the transfer model describes a decision about subsequent schooling, conditional on having pre- viously chosen to attend two-year college. High school graduates who never attend college, and students who attend only four-year college, are not included in this part of the analysis. The non-transfers consist of three groups: those who continue to attend two-year college, those who complete an associate’s degree, and those who leave school. Because all these individuals could subsequently enroll in a four-year college at some point, it would be inappro- priate to exclude any of them from the analysis. The model treats them as potential transfers, but includes a 13 As the inclusion of post-schooling outcomes in schooling decisions poses problems for interpreting causality, we present models that exclude occupational indicators. The discussion in Section 4.1 outlines the main differences between the attend- ance models with and without these variables. variable to identify individuals who have a gap in their schooling history of more than one year. 14 The explanatory variables used to describe whether an individual transfers are very similar to those used in the attendance model. Exceptions are that broad field of study is incorporated and the explanatory variables that can vary with time are set to their values as of age 20. The inclusion of field of study addresses one of the limi- tations of the attendance model. Expectations for vari- ables included in both the attendance and transfer mod- els, where different from the expectations outlined above, are noted in the discussion of the results.

4. Multivariate results

This study uses univariate probits to describe college attendance decisions and the decision among two-year college students to transfer to a four-year college. 15 The results are expressed as Probit derivatives for ease of exposition and interpretation. 16 Estimation results from the attendance models are reported in Table 4. 17 Table 5 presents the results of the transfer models. 4.1. Two- and four-year college attendance models We initially argued that women choose two-year over four-year attendance as a result of practical consider- ations—such as family responsibilities—and that such considerations might also explain why women transfer at lower rates than men. The results from the attendance models partially bear out this expectation. The gender- 14 Leigh Gill 1997 report that individuals who leave school and then return still earn a wage premium on their initial college credits. This indicates that the value of college credits does not depreciate fully over even several years. 15 Other researchers have used multinomial logit to examine the decision to attend different types of post-secondary insti- tutions Manski Wise, 1983; Ordovensky, 1995; Hilmer, 1997. It seems likely that for the questions addressed in this paper, the “independence of irrelevant alternatives” IIA con- dition required for that method would not be satisfied. 16 Probit derivatives are simply the change in the probability of an outcome caused by a change in one of the explanatory variables. For y = FxB, the Probit derivative is calculated y 9 = fxbb, where b is the estimate of B. The probabilities are calculated at the x-variable means. 17 Several of the variables included on the right-hand side of the decision equations are the product of past decisions and may therefore cause endogeneity bias. Methods for reducing such bias such as the Instrumental Variables method generally require a variable that affects past decisions but does not other- wise directly affect the current decision. Identification based solely on functional form is generally viewed as an unsatisfac- tory solution to this problem. Given the difficulties such correc- tions pose, the model does not address endogeneity bias. 157 B.J. Surette Economics of Education Review 20 2001 151–163 Table 4 Probit models of two-year and four-year college attendance probability derivatives listed, Huber–White t-scores in parentheses a Two-year attendance b Four-year attendance b 1 Pooled 2 Women 3 Men 4 Pooled 5 Women 6 Men Demographics Female 1.19 4.03 0.64 1.89 Married 22.11 5.33 22.65 5.15 21.34 2.10 24.45 9.11 25.92 9.30 22.62 3.47 Young kids 22.50 5.87 22.97 5.46 22.03 2.63 -2.51 4.07 24.08 5.33 0.25 0.22 AFQT score 0.83 8.36 0.85 5.66 0.75 5.86 2.48 20.46 2.59 14.09 2.17 14.83 Black cross 1.51 2.33 1.37 1.51 1.71 1.89 4.30 4.72 4.87 3.95 3.77 2.94 Hispanic cross 1.89 2.16 1.41 1.13 2.43 2.05 4.15 3.46 6.06 3.37 2.16 1.52 Black over-sample 0.94 1.98 1.63 2.33 0.16 0.26 4.00 7.17 4.45 5.19 3.52 5.25 Hispanic over-sample 2.32 4.36 1.94 2.62 2.63 3.60 1.73 2.68 0.89 0.93 2.29 2.73 Parents education included included included included included included Urban location 1.98 6.05 2.02 4.14 1.95 4.55 0.37 0.99 0.60 1.08 0.30 0.66 Age 13.7010.27 15.99 8.49 11.33 6.10 11.25 7.36 9.77 4.79 12.98 6.12 Age squared 215.6310.84 217.86 8.68 213.26 6.67 215.53 9.04 213.70 5.98 217.28 7.24 Age cubed 5.34 10.91 5.96 8.51 4.66 6.96 5.88 9.70 5.25 6.53 6.41 7.60 Time trend 0.00 0.32 0.05 0.33 20.02 0.12 20.17 1.34 0.27 1.40 20.11 0.67 Costs and benefits College wage premium 3.60 2.64 3.46 1.79 3.78 2.05 24.33 2.87 21.08 0.48 27.35 3.83 HS graduate wage 4.23 2.41 1.71 0.64 6.07 2.68 24.27 2.24 24.96 1.59 23.46 1.64 Two-year tuition 23.02 6.39 23.45 5.22 22.68 4.09 2.06 3.54 2.26 2.65 1.60 2.19 Four-year tuition 20.93 2.45 21.00 1.87 20.76 1.47 21.27 2.85 21.68 2.56 0.78 1.42 Unemployment rate 0.22 5.12 0.19 3.04 0.25 4.43 0.16 3.20 0.21 2.87 0.13 2.04 Human capital 2-yr credits 4.38 21.49 4.22 14.46 4.44 18.96 1.44 3.94 0.52 1.07 2.12 4.25 4-yr credits 25.37 10.85 26.46 9.93 24.06 5.62 15.36 29.82 14.95 19.09 14.55 23.79 4-yr credits 2 0.59 6.15 0.81 7.30 0.28 1.74 22.43 18.14 22.37 11.32 22.30 14.92 AA degree n.a. n.a. n.a. 0.43 0.48 20.18 0.15 1.80 1.43 Experience 20.94 7.22 20.97 5.21 20.98 5.60 22.14 9.06 21.78 5.29 22.46 8.44 Vocational training 20.39 0.73 20.88 1.23 0.03 0.04 21.33 1.67 21.72 1.61 21.07 1.13 Mean S.E. of dep. var. 0.086 0.280 0.092 0.288 0.080 0.270 0.184 0.387 0.179 0.382 0.185 0.388 Sample size 36,223 18,798 17,425 36,223 18,798 17,425 Pseudo-R 2 0.1719 0.1809 0.1682 0.4520 0.4434 0.4700 a Standard errors are calculated using the Huber 1967 and White 1980 method to account for multiple observations of individuals. b Includes all high school graduates in each year they are observed in the data. Once individuals complete a bachelor’s degree, they are excluded from the estimation. 158 B.J. Surette Economics of Education Review 20 2001 151–163 Table 5 Probit transfer models Probability derivatives listed, t-scores in parentheses Transfer from two-year college a 1 Pooled sample 2 Women 3 Men Demographic factors Female 26.28 2.05 Married 217.39 3.51 216.51 2.90 222.15 2.16 Young children 215.17 2.59 216.07 2.60 27.36 0.41 AFQT score 9.48 8.57 9.36 6.09 10.11 6.09 Black random 4.00 0.60 21.30 0.15 11.87 1.17 Hispanic random 11.47 1.36 22.76 1.89 2.82 0.22 Black over-sample 9.84 2.15 4.30 0.70 14.97 2.16 Hispanic over-sample 4.07 0.78 23.84 0.54 12.15 1.60 Parents education included included included Urban location 4.27 1.10 10.65 1.97 23.78 0.67 Age 21.73 1.43 20.67 0.40 23.14 1.75 Time trend 2.50 1.54 0.11 0.50 4.20 1.75 Costs and benefits College premium 210.94 0.97 26.90 0.43 217.35 1.07 High school wage 235.99 2.92 232.85 1.99 249.30 2.65 Tuition ratio 4-yr2-yr 21.00 2.22 21.31 2.07 20.89 1.35 Unemployment rate 20.22 0.49 20.29 0.46 0.19 0.27 Human capital Two-year credits 27.37 4.91 210.91 5.35 23.15 1.38 AA degree 27.62 8.09 27.63 5.88 26.58 5.30 Experience 225.16 8.51 231.96 7.37 219.65 4.56 Ever train 27.58 2.50 26.70 1.64 210.63 2.30 Gap in schooling 8.94 1.81 10.34 1.56 6.60 0.88 Field of study Vocational 23.79 5.23 18.19 2.98 29.38 4.25 Academic 36.58 4.25 44.33 3.87 27.28 2.05 Mean of dep. var. 0.415 0.493 0.385 0.487 0.452 0.498 Sample size 1390 764 626 Pseudo-R 2 0.2477 0.3022 0.2183 a Includes only individuals who attended a two-year college. High school graduates who never attended either type of college and four-year college students who never attended a two-year college are excluded. specific regressions show that having young children or being married affects women’s college attendance decisions more than men’s. In the two-year and four-year attendance equations columns 2, 3, 5, and 6 of Table 4 the coefficients on “married” and “young kids” are larger in absolute value for women than for men. 18 Moreover, consistent with notion that two-year colleges are more accommodating than four-year colleges, having young children and being married generally reduce four-year attendance more than two-year attendance. We also suggested that living in an urban area might proxy for proximity to college, and that gender differ- ences in the effects of “proximity” thus defined might 18 Gender differences in the effects of marital status and hav- ing young children are statistically significant at the 5 level in only the four-year attendance equation. explain observed attendance patterns. The urban indi- cator is positive and significant in all three two-year attendance equations, consistent with the hypothesis that proximity is an important determinant of two-year attendance. However, this effect does not differ by gen- der, nor does it affect four-year attendance, which sug- gests that proximity does not explain the gender differ- ence in attendance patterns. To test whether differences in educational ability can explain the gender difference in transfer rates, the model includes the AFQT score in the attendance models. All the columns in Table 4 show that the AFQT is an important determinant of attendance and that it has a larger effect on four-year than two-year attendance. However, there is no gender difference in the effects of AFQT on either two-year or four-year attendance. The AFQT effects on attendance indicate that marginal stu- dents tend to attend two-year college, but there is no evidence that the AFQT matters less for women. 159 B.J. Surette Economics of Education Review 20 2001 151–163 The attendance models include a set of economic fac- tors to shed further light on how students decide between and use two- and four-year colleges. We focus first on the direct costs of attendance measured by two-year and four-year tuition. The signs of the own-tuition effects are generally negative, as expected, and do not differ sig- nificantly by gender. The signs of the cross-tuition effects are a priori ambiguous; their estimated values tell us something useful about how students utilize the two types of college. The positive effect of two-year tuition in the four-year attendance equations columns 4, 5, and 6 indicates that when two-year college is expensive, stu- dents are more likely to attend four-year college and are less likely to attend two-year college. These cross- effects are not symmetric, however. The negative effect of four-year tuition in the two-year attendance equations columns 1, 2, and 3 suggests that when four-year col- lege is expensive, students are less likely to attend two- year college. This latter effect is consistent with two- year colleges’ transfer role: when four-year college is expensive and, therefore, more difficult to justify on economic grounds, the transfer option is less attractive and attendance at two-year colleges is reduced. 19 The opportunity cost of college attendance is charac- terized by the wage earned by high school graduates in each individual’s state of residence. 20 As expected, this cost reduces four-year college attendance. Contrary to expectations, this opportunity cost generally raises two- year attendance. Turning next to the benefits of attend- ance, the four-year college graduate wage premium raises the probability of two-year college attendance for both women and men, but reduces the probability each particularly men attends four-year college. The latter effect also contradicts expectations. These contradictory results may stem from a number of causes; state level aggregate wages capture any num- ber of state-specific effects making interpretation of such variables problematic. I retained them because human capital theory argues strongly for their inclusion in col- 19 Intuitively, this asymmetry implies that the cross-price sub- stitution effect is small compared to the income effect in the two-year attendance equation and vice versa in the four-year attendance equation. This asymmetry may stem from intertem- poral considerations playing a larger role in the two-year decision equation—in effect reducing the cross-substitution effect—than in the four-year decision equation. 20 Median wages by education level are obtained from the Current Population Survey CPS. They are calculated by state and year for 18- to 40-year-olds reporting highest grade com- pleted equal to 12, and for 22- to 40-year-olds for individuals reporting highest grade completed equal to 16 or more. There are not enough observations to calculate these variables by gen- der for each state and year. It is straightforward, but uninforma- tive, to include the wage premium earned by individuals with 1 to 3 years of college. lege attendance models. Note that none of the results are altered by excluding either the high school graduate wage or the return to college or both. Despite controlling for the many factors that human capital theory indicates should affect attendance, the female coefficient in the pooled two-year decision equ- ation remains positive and statistically significant: women remain more likely than men to attend two-year college. This suggests that the explanations advanced thus far for why women are less likely than men to attend two-year colleges only tell part of the story. Family responsibilities, proximity, monetary costs, and ability all play an important role, but they do not completely explain the significant gender difference in two-year attendance rates. As noted earlier, attendance decisions are almost cer- tainly driven in part by occupational preferences. We cannot observe occupational preferences, but we can proxy for them using ex-post-occupational outcomes. In models not reported, we include indicators for whether each individual ever worked in one of three female- dominated occupations: the secretarial profession, the allied health profession, and primary or secondary school teaching. Training for the first two professions is gener- ally provided by two-year colleges. 21 Training for the latter is generally provided by four-year colleges. The inclusion of occupational indicators causes the coefficient on female in the pooled two-year college regression to become very small and statistically insig- nificant. Women who subsequently work in either the allied health or secretarial professions are much more likely to attend two-year colleges than those who sub- sequently work in other fields. Working as a primary or secondary school teacher or in an allied health profession significantly raises the probability of four-year attend- ance. Controlling for these variables “explains away” the unexplained gender difference in two-year college. The inclusion of such post-schooling outcomes is hard to justify on logical and econometric grounds—and for that reason we present in Table 4 the models that exclude them. However, such models do tell us something important about the attendance decision and are therefore worthy of note. They strongly suggest that women attend two-year colleges at higher rates than men as a result of occupational preferences. The question remains as to whether the entire set of explanatory variables explains the gender difference in transfer rates as opposed to attendance rates. The four- year attendance model allows us to examine this ques- tion, albeit indirectly. Controlling for other factors, one expects women and men with equivalent numbers of two-year credits to be equally likely to attend four-year 21 Some of the allied health professions, most notably nurs- ing, may require a bachelor’s degree. 160 B.J. Surette Economics of Education Review 20 2001 151–163 college. 22 While this is not, strictly speaking, the same as transferring, it is very similar. The gender-specific four-year attendance regressions columns 5 and 6 show that the effect of accumulated two-year credits on four- year attendance is smaller for women than for men; one year of two-year credits raises the probability men attend four-year college by 2 or 3 percentage points but has no effect on women’s probability of attendance. This differ- ence is statistically significant at the 1 percent level. Put another way, women with one year of two-year credits are significantly less likely than similar men to attend four-year college and have no higher a probability of attending a four-year college than an otherwise similar high school graduate with no two-year credits. Note that it is not the case that credits simply matter less in women’s schooling decisions. The effects of accumulated two-year credits on further two-year attend- ance columns 2 and 3, and the effects of accumulated four-year credits on further four-year attendance columns 5 and 6, do not differ by gender. Women sim- ply appear less likely than similar men to use two-year colleges as stepping stones to four-year colleges. This is so despite the inclusion of many variables that could theoretically explain the gender difference. Moreover, the inclusion of the occupational indicators discussed above but not included in the models in Table 4 does not alter this conclusion. 4.2. The transfer model The attendance models tell us much about what might and might not explain the lower transfer rates observed among women. However, the effects of two-year credits on four-year attendance described above does not meas- ure gender differences in transfer rates per se. Rather, we infer lower rates of transfer for women from the fact that men and women with equivalent numbers of two- year credits and other similar characteristics make sub- sequent educational decisions differently. An alternative way to test whether women are less likely than men to transfer is to directly estimate a model with an indicator for whether or not a student transferred at some point during the 12-year panel as the dependent variable. High school graduates who never attend college and students who enrolled only in a four-year college are not considered in the transfer model. Estimation results for three groups—one pooled and one each for women and men—are reported in Table 5. 22 Four-year colleges vary in whether or not two-year credits can be applied by transfer students toward a bachelor’s degree. Usually, such requirements involve grade received and course content. I have no way of identifying which credits would be acceptable in transfer so all credits earned are included in these models. The explanatory variables are similar to those used in the college attendance equations. However, because the dependent variable is retrospective over up to 12 years, variables that can change with time are set to their values at age 20. 23 Thus, marital status, the presence of children under the age of 5, tuition, the unemployment rate, urban status, and others are fixed at their values when each individual was 20, even though the transfer could have occurred after that age. The pooled estimates in Table 5 column 1 show clearly that women are much less likely than men to transfer. Even after controlling for a wide range of other factors that could explain this trend, being female reduces the probability of transferring by about 6 per- centage points. 24 We postulated earlier that two-year col- leges’ more flexible class schedules better accommodate domestic responsibilities, which may be borne dispro- portionately by women. Consistent with this hypothesis, the gender-specific effects of having young children are much larger and negative for women column 2 than for men column 3. Contrary to expectations, being mar- ried reduces men’s transfer rates more than women’s, though this difference is not statistically significant. Finally, the effects of living in an urban area on the transfer decision are substantially higher for women than for men, which is consistent with proximity mattering more for women than for men. 25 None of these variables, however, fully explains the gender difference in trans- fer rates. Because the transfer model is defined only for individ- uals who attended college, it is possible to address the possibility that women attend two-year colleges at higher rates than men because the fields of study offered there prepare them satisfactorily for their chosen occupations, which may differ from the fields of study men choose. We include indicators for whether the field of study was vocational or academic, as defined by a taxonomy pro- vided by the National assessment of vocational edu- cation United States Department of Education, 1989. The omitted category is “no major field specified”. Despite the inclusion of these indicators, the coefficient on “female” in the pooled regression is negative, indicat- ing that women remain less likely than men to transfer. Women and men with similar numbers of two-year cred- its, the same broad field of study, who are unmarried and have no small children, and are otherwise identical have different rates of transfer to four-year college. 23 Setting these variables to their values at age 21 or 19 does not alter the qualitative results. 24 The coefficient on female is fairly robust, remaining close to 6 percent for a wide range of models. Changes in the model rarely causes the t-statistic on this parameter to drop below 1.95. 25 The gender difference in the urban coefficient is statisti- cally significant at the 5 level. 161 B.J. Surette Economics of Education Review 20 2001 151–163 We also hypothesized that women’s lower transfer rates might be explained by differences in educational ability; that more marginal women than men attend two- year college. The AFQT score is included in the transfer model to address this possibility, but the effects of the AFQT score do not differ much across gender. The AFQT is probably an imperfect proxy for edu- cational ability. Unobserved components of ability that drive schooling decisions will be partially captured by accumulated college credits. The effects of two-year cre- dits are much lower for women than for men. While hav- ing an associate’s degree raises the probability of trans- ferring equally for men and women, completing an associate’s degree requires approximately 2 year’s worth of credits. The combined effects of credits and associ- ate’s degree completion are very small in the female equ- ation, while in the male equation the combined effects of credits and degree completion are quite large. 26 These results confirm the findings from the attendance models; women who attend two-year college are less likely than similar men to transfer. A third explanation advanced for why women transfer at lower rates than men is that they simply prefer two- year colleges to four-year colleges due to occupational ambitions. To test this hypothesis, we estimated the transfer model with indicators for whether an individual ever worked as a primary or secondary school teacher, in the allied health profession, or as a secretary estimates not reported. Women who subsequently work in the allied health profession or as teachers are more likely to transfer. However, the key results of this paper are not affected by the inclusion of occupational dumm- ies; the coefficient on female remains negative and stat- istically significant, and in fact, becomes larger and sig- nificant at the 1 level. We conclude from these regressions that while occupation indicators have a lot of explanatory power, they do not alter the conclusion that women are less likely than men to transfer. 4.3. Credits and labor market earnings Human capital theory predicts that credits affect col- lege attendance by changing the tradeoff between current and future earnings. If the relative returns to two-year and four-year credits differ for women and men, it might be economically rational for women to transfer at lower rates. We tried to test this hypothesis by incorporating the rate of return to college in the attendance and transfer models. Those results were somewhat unsatisfactory, due 26 For women, the effect of completed two-year credits, mul- tiplied by 2 years worth of credits, substantially offsets the effect of having an associate’s degree 27.6 2210.9 + 6. For men the combined effect is about + 20. in part to limitations in using state-level aggregate wages to proxy for current and future earnings. To further evaluate this possible explanation, Table 6 reports rate of return estimates from four Mincer, 1974 human capital earnings functions based on the NLSY. Several previous authors have examined the returns to two-year college credits and associate’s degree com- pletion Grubb, 1993; Kane Rouse, 1995a,b; Surette, 1997. The Grubb and Kane and Rouse analyses are based primarily on the National Longitudinal Survey of the Class of 1972, and the college credit data is now over 20 years old. The NLSY data used here and in Surette is more recent though still over 10 years old and is more relevant for explaining attendance behavior during the 1980s. The estimates reported in Table 6 are broadly similar to those reported in previous studies. The regressions presented in Table 6 explain the log of hourly earnings and the log of annual earnings, by gender, as functions of two-year credits, four-year cred- its, degrees, and a large set of other explanatory variables see footnote in Table 6 typically included in human capital models. The estimates show that women and men do not differ appreciably in their returns to two-year cre- dits, and that for women, the returns to two-year and four-year credits do not differ much. Men’s return to four-year credits are lower than their return to two-year credits, and are much lower than women’s returns to four-year credits. This is the opposite of what one would expect if differences in returns to college explain the transfer rate difference. These estimates suggest that on economic grounds women should, if anything, be more likely than men to attend or transfer to four-year college, not less.

5. Summary and conclusion