How To Find Out If A Woman Can Complete An Mba

Transcription

1 The Quarterly Review of Economics and Finance 47 (2007) Best laid plans: Gender and the MBA completion rates of GMAT registrants Mark Montgomery a,, Katharine Anderson b a Department of Economics, Grinnell College, Grinnell, IA 50112, United States b Department of Economics, University of Michigan, United States Received 5 October 2004; received in revised form 8 April 2005; accepted 15 August 2005 Available online 10 January 2007 Abstract Evidence suggests that while women are more likely to go to college than men, they are less likely to go to graduate school. Moreover, in fields like science and engineering, women who do pursue advanced degrees are less likely than men to complete them. This paper compares MBA completion rates for women and men who registered to take the Graduate Management Admission Test (GMAT). We find that female test registrants are about 30% less likely than men to complete the MBA. This difference in completion rates seems only partly due to gender disparity in family responsibilities Board of Trustees of the University of Illinois. All rights reserved. JEL classification: I21; J16; J44 Keywords: MBA; GMAT; Gender disparity 1. Introduction The political, scientific and business leaders of the 21st century are likely to have more than 4 years of college education. Yet empirical evidence suggests that while women are more likely to go to college than men, they are less likely to go to graduate school. Moreover, those women who do pursue advanced degrees appear less likely than men to complete them. This conclusion, however, rests on a literature that is very thin. Only a handful of studies have looked at completion rates beyond undergraduate degrees, and those that do, cover only a narrow range of fields. Moreover, Corresponding author. Tel.: ; fax: address: montgome@grinnell.edu (M. Montgomery) /$ see front matter 2006 Board of Trustees of the University of Illinois. All rights reserved. doi: /j.qref

2 176 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) these studies tended to focus only on top-flight students, those coming from, and/or going to, elite colleges and universities. This paper examines male and female graduate school completion rates in an area where they have yet to be studied: business. We use a recent longitudinal survey of people who registered for the Graduate Management Admissions Test (GMAT) between June 1990 and March This is the first study of graduate school completion to model the school enrollment and degree completion decisions simultaneously. We estimate a nested logit model of two sequential choices: (1) which, if any, of four types of MBA programs to enroll in and (2) whether to complete the degree once enrolled. 1 Program type was defined by whether it was full-time or part-time and whether it was more-selective or less-selective. As a test of their robustness, the nested logit model results are compared with those from an ordinary probit, and with those from a bivariate probit with sample selection. We find that after controlling for a wide range of academic credentials and personal characteristics, female GMAT registrants were about 0.08 (30%) less likely to complete an MBA (within 7 years) than their male counterparts. This is partly because women test takers are less likely to pursue the degree, but mostly because those who do are less likely to complete it. Simulations from a nested logit model show that had the female school-goers in our sample been otherwise-identical males, their completion rates would have been 24 55% higher, depending upon the type of MBA program chosen. Among the young, single and childless enrollees the gender completion gap was somewhat smaller but did not disappear. We interpret this combination of results to suggest that while family responsibilities are a greater impediment to MBA completion for women than men, they explain only a portion of the gender completion gap. 2. Studies of completion rates in higher education There is a substantial body of literature on the determinants of undergraduate educational choice: whether to go to college and what college to attend. Most studies of undergraduate attendance agree that after controlling for individual characteristics like family background and financial status, women are more likely to attend college than men (Ganderton & Santos, 1995; Light & Strayer, 2000; McPherson & Schapiro, 1991). There is disagreement, however, about the effects of gender on college completion rates. Ganderton and Santos (1995) find that for white and black students (though not hispanics) women are significantly less likely than men to obtain the bachelor s degree. 2 Light and Strayer (2000), however, using better controls for student ability, find that female college students are significantly more likely to graduate. 3 Compared to those on college attendance, the studies of post-graduate educational choice are few. This is a significant omission in the literature, because post-graduate education should be an area of increasing interest. Marriage and child rearing, which are educational impediments that differentially affect women, seem to be shifting down the female life cycle. For example, between 1970 and 2000 the median age of a woman at first marriage rose from 20.8 to 23.2 years (Fields & Casper, 2001). And the trend in postponed fertility has been quite dramatic: the rate of first live births among women aged years increased 280% between 1960 and 1 Throughout this paper the term MBA is used as a familiar shorthand for the several masters level management degrees. 2 Interestingly, they found no relationship between gender and completion among hispanic students. 3 The discrepancy between Ganderton and Santos (1995) and Light and Strayer (2000) may partly reflect the fact that the observed male female completion gap appears to be narrowing over time.

3 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) (Centers for Disease Control, 1997). Thus, it seems clear that increasingly young women face the conflict between family and education after college that their grandmothers faced after high school. In this paper we hypothesize that the constraints imposed by marriage and family are likely to differentially affect women s ability to complete a graduate management degree. That women are more burdened by actual pregnancy and child birth is obvious, but there are other likely additional impediments as well. For example, South and Spitze (1994) find that whereas women tend to do more housework than men in all living arrangements, the absolute burden is highest for married women. Moreover, most MBA s are pursued part-time on nights and weekends. Presser (1995) shows that children make it harder to work (and by extension, to go to school) during such nonstandard hours, and that this constraint binds hardest for women. 4 It is reasonable to suppose, therefore, that women MBA aspiriants such as our GMAT registrants, will likely encounter more barriers to degree completion than will their male counterparts. The comparatively few studies that do focus on post-graduate educational choice have tended to focus on the impact of financial aid (Ehrenberg & Mavros, 1995; Fox, 1992; Jantzen, 2000; Weiler, 1994) or on the quality/reputation of the undergraduate institution (Eide, Brewer, & Ehrenberg, 1998). There are, however, two studies of graduate level schooling that examine gender specifically. Schapiro, O Malley, and Litten (1991), using data from the Consortium on Financing Higher Education (COFHE), find that men are significantly more likely to attend graduate school than women. Baker (1998), looking at completion rates for degrees in science and engineering using applications for National Science Foundation fellowships, finds that women s attendance and completion rates lag far behind those of men. Our study extends the small literature described above in three ways. First, we examine the gender gap in graduate degree completion in an area in which it has not been studied: business. This is useful because business is a less classically male-dominated field than science and engineering, as studied by Baker (1998). Second, unlike previous studies of graduate education, we follow the approach that Light and Strayer (2000, 2002) took for undergraduate education, modeling the decision to enroll simultaneously with the decision to complete the degree. This provides a much better defense against any unobserved heterogeneity that could bias the apparent impact of gender on school completion. Third, the graduate school studies of Schapiro et al. (1991) and Baker (1998) were generally restricted to the best students from the best colleges. Our GMAT registrants, on the other hand, range from third year undergraduates to executives in their 1950s The GMAT registrant survey The data for this analysis come from the GMAT Registrant Survey conducted by Battelle Memorial Institute (of Seattle) on behalf of the Graduate Management Admissions Council, which is the organization that administers the GMAT. The survey was conducted in four waves from 1990 through 1998, as summarized in Table 1. Wave 1 contacted 7006 individuals who registered to take the GMAT on test dates between June 1990 and March 1991; 5885 registrants responded. The final follow-up, Wave 4, was completed in 1998, roughly 7 years after test registration. Our sample is drawn from 3771 of the Wave 4 respondents who answered the necessary questions. 6 Table 1 reports the wave structure of the GMAT survey. 4 For a recent review of the differential effects of housework and childrearing on women, see Stratton (2003). 5 We did, however, eliminate a few registrants whose ages at GMAT registration exceeded This number includes 332 people who responded to Wave 4 but not Wave 3.

4 178 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Table 1 Wave structure of longitudinal survey of GMAT registrants Wave number Survey dates Time since registration Number responding Wave 1 6/90 3/ Wave 2 9/91 6/92 15 months 4883 Wave 3 1/94 6/ years 4533 Wave 4 1/98 12/98 7 years 4188 Table 2 Business school attendance and completion rates by sex All registrants (%) Female (%) Male (%) Completed MBA Enrolled in business school Enrollees who went FT Enrollees who completed Sample The GMAC survey tells us which school, if any, a registrant attended and whether she went part-time or full-time. An important feature of this data set is that survey responses could be matched with GMAC s own registration and test records. This gives us exceptionally good information about applicants academic credentials, including GMAT score, grade point average, and undergraduate major. Moreover, because we know the applicant s undergraduate school, information about its quality can be drawn from various published sources. Ability to control for these credentials is useful since they help determine which schools an applicant can attend and some are correlated with gender Sample attendance and completion rates Table 2 reports MBA enrollment and completion rates for our sample, by sex. As the first row shows, only 26% of GMAT registrants end up with an MBA by Wave 4 of the survey. 7 As rows 3 and 5 show, only 61% of the registrants in our sample eventually enrolled in business school; of these 39% earned the degree. Some gender differences are apparent. The completion rate for women is substantially lower than for men: 21% versus 29%. Women registrants were somewhat less likely to enroll than men 57% versus 64% and those who did enroll were less likely to finish: 34% of women enrollees as opposed to 42% of men. 8 Thus, two factors seem to make women GMAT registrants less likely to obtain an MBA: they are less likely to pursue the degree and less likely to complete it if they do. 7 About 16% of our sample of GMAT registrants never took the test. Nevertheless, many went on to pursue, and some to complete, MBAs. 8 About 6% of the sample reported still being enrolled in business school at Wave 4. For estimation purposes we treated them the same as dropouts. Leaving them out of the sample affected the results slightly, mainly by making the negative effect of being female on completion somewhat stronger.

5 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Modeling MBA completion We should expect that the two forms of non-completion of the MBA distinguished in Table 2 never enrolling, and dropping out would be influenced by personal characteristics in different ways and would therefore require separate econometric models. (The decision to enroll and the decision to dropout are, after all, often made years apart.) But modeling enrollment and completion as entirely separate phenomena involves some well-recognized econometric dangers. It is easy to see that measuring the effect of gender on MBA completion must confront the fact that the sample of those who are business school-enrolled is self-selecting. To illustrate, we borrow an example from Light and Strayer (2002) and adapt it to our specific case. Let us assume there is an unobserved personal attribute called career ambition that critically influences both the pursuit and completion of an MBA. 9 On average, we assume, men and women have equal amounts of career ambition. But a young woman who undertakes graduate business study which happens during her peak years of fertility requires more of this attribute than a similarly situated man. Hence, the typical woman studying in any given MBA program has more career ambition than her typical male classmate. Because more ambitious people are more likely to complete the MBA, the observed graduation rate among females will be higher, ceteris paribus, than among males. Their gender will appear to give women students an advantage in graduating that they do not in fact possess they are simply more ambitious, on average. Or, by extension, any gender disadvantage they may actually suffer will be understated if we ignore sample selection. The problem of self-selection can be avoided by using a model that estimates the parameters of the business school enrollment and MBA completion decisions simultaneously (as Light & Strayer 2002, 2004 did for undergraduate study). A number of sequential discrete choice models are available to fill this need. There is a difficulty with such models, however: the complexity of estimation severely restricts the number of estimable parameters. Often, as in our case, some interesting and useful covariates must be jettisoned from the model before estimation becomes feasible. In light of the above difficulties, and in order to obtain the most robust picture possible of the influence of gender on MBA completion rates, this paper employs three alternative discrete choice models: (1) Probit models. These can be used to estimate the unconditional probability of completing the degree. (2) Bivariate probit models with sample selection. These models simultaneously estimate the probability of choosing to pursue the MBA and the conditional probability of completing the MBA having enrolled. (3) A nested logit model. This approach simultaneously estimates the probabilities of choice among four alternative types of MBA program part-time or full-time, and more-selective or less-selective and the (four) conditional probabilities of having completed the MBA, given the type of program chosen. Models of types (1) and (2) make use of the full range of available covariates, while model (3) most effectively addresses the question of sample selection. We provide details of the models below. 9 Light and Strayer (2002) use drive.

6 180 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) As usual in discrete choice analysis, our models define the utility to individual i of completing an MBA, Ui C, and of not earning one, UNC i,as Ui C = β C x i + ε ic,ε C N(0,σ C ) and Ui NC = β NC x i + ε inc,ε NC N(0,σ NC ) (1) The individual completes the MBA if Ui C >Ui NC, that is if ε ic < ε inc. Defining the error for the net utility of completing as ε i = ε ic ε inc, the probit models estimates the probability that y i =1 as Prob(y i = 1) = Prob(ε i βx i ) = Φ(βx i ) (2) where Φ is the cumulative distribution function of the standard normal. The simple probit model does not acknowledge that there are several ways to not complete an MBA. It is important to distinguish the role of gender in abandoning plans for business school from its role in the very different phenomenon of enrolling but not finishing. One way to examine the effect of gender on dropping out is simply to run an ordinary probit on a sample of registrants who enrolled in some program. But this approach courts the problem of sample selection bias, as described above. Therefore, to consider the effect of gender on the probability of completing conditional upon some attempt to earn one (i.e. enrollment), we use a bivariate probit with sample selection. The model was developed by Van de Ven and Van Praag (1981). We posit a net utility, U E, from enrolling in an MBA program and a net utility, U C, from completing the degree once enrolled, and assume that their respective error terms share a bivariate normal distribution: Ui C Ui E = β C x Ci + ε Ci = β E x Ei + ε Ei ε Ci,ε Ei Φ 2 (0, 0, 1, 1,ρ) (3) Φ 2 is the cumulative density function for the bivariate normal and ρ is the correlation between ε C and ε E. We define dummy variable y Ei to equal one if a registrant enrolled in an MBA, and zero if she did not; we define y Ci to equal 1 if she completed the MBA program, and zero if she did not. Then, Prob(y Ci = 1 y Ei = 1) = Φ 2 (β C x Ci,β E x Ei,ρ) (4) While the bivariate probit model has the advantage over ordinary probit of confronting sample selection, it is not sophisticated enough to differentiate among types of program. This is a significant limitation in that part-time and full-time attendance represent two distinct regimes of business school study. Part-time students, who are about twice as numerate as full-timers, usually have full-time day jobs and take courses on nights and weekends. They are much more constrained in their geographic choice of school than are full-time students. On the other hand, a prospective part-time student has many more programs to choose from most MBA programs do not even offer day classes. The part-time/full-time distinction is especially meaningful for this study because a married MBA student with children will likely have access to child care in the evening (via the spouse) than during the day. We might expect women to more often choose the part-time option. Moreover, gender could influence completion rates differently among the various available types. For example, a woman pursuing an MBA in the daytime may have more hours with her family than one who is in class during nights and on weekends. To account for the possibility that gender might affect enrollment and completion in different regimes of study differently, we use a nested logit model. This approach posits two sequential

7 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) decisions on the part of a GMAT registrant: (1) what type of MBA program (if any) to attend and (2) whether to complete the degree once enrolled. In our nested logit model, decision (1) comprises the first nest and decision (2) the second. The parameters that determine the probability of making a particular choice within a nest are estimated simultaneously for both nests. For our first nest we divide all 573 US MBA programs into four types depending upon whether they are part-time or full-time and whether they are more-selective or less-selective in admissions. 10 A program s selectivity was determined by the average GMAT score of its students: 550 or higher was deemed more-selective. By the laws of probability, the unconditional probability of completing an MBA program of type k (call it P C k ) can be expressed as the product of two other probabilities: the conditional probability of completing an MBA program given that it is type k (which we will call P C k ) and the probability of choosing a program of type k (P k ). We can write this as P Ck = P C k P k (5) The nested logit model assumes the right-hand probabilities are related to the vectors of individual and MBA program characteristics that determine enrollment and completion call these vectors x C and x E in the following way (suppressing i for the individual): P C k = eβ Ckx Ck e β Ekx Ek +λ k I k 1 + e β, P k = (6) Ckx k k K eβ Ekx Ek +λ k I k As above, C refers to completion and E to enrollment. Note that the expression for P k looks like a standard multinomial logit probability except for the appearance of the variable I k. The variable I k is the so-called inclusive value for choice k, and is defined as I k = ln(1 + e γβ Ckx Ck ). It represents the utility value (to within scale factor γ) of having MBA programs of type k available. The inclusive value provides a handy measure of the usefulness of nesting: if the coefficients on the inclusive values are all zero, then the completion decision is independent of the type of school chosen. In this case there would be no statistical advantage to nesting the decisions. Fig. 1 shows the basic structure of the nested logit and the proportions of the sample falling into each category of outcome. The nested logit does a better job than the probit models of avoiding selection bias when estimating the probabilities of MBA completion conditional upon enrollment. It also has some significant limitations which is why we do not rely upon it exclusively. The complexities of the nested logit substantially restrict the number of estimable parameters. Achieving convergence in this model compelled us to make several structural concessions including paring down the set of independent variables and constraining some parameters to equality across branches of the two nests. We also had to assume that the inclusive values (the I k, above) were equal across the four types of MBA programs. In spite of its limitations, however, the nested logit is a critical supplement to the probit and bivariate probit models in providing a robust and consistent picture of the impact of gender on MBA completion. 5. Independent variables Table 3 reports descriptive statistics for our independent variables by gender. A first set of variables comprises characteristics of the individual GMAT registrants, a second represents character- 10 The choice of four types of program, as opposed to six or eight, was governed by what the nested logit model would tolerate. Four was also the number of school options used by Light and Strayer (2002, 2004) in their multinomial probit models of undergraduate enrollment and completion.

8 182 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Fig. 1. Structure of the nested logit model. istics of the MBA programs they attended (if any). To avoid simultaneity between enrollment and demographic characteristics like marriage and fertility, most demographic variables are measured as of Wave 1 of the survey. About 44% of the registrants were female. About a third were married, with men slightly more likely to be married than women. The average age for men was 27 years, Table 3 Means and standard deviations of variables used in estimation (an omitted standard error indicates a dummy variable) All registrants Female Male Mean Standard deviation Mean Standard deviation Mean Standard deviation Personal characteristics Age (Wave 1) Female Married (Wave 1) Number of kids (Wave 1) Predicted hourly earnings w/o MBA Asian Black or hispanic Had FT job Wave Undergraduate GPA GMAT score Did not take test College selectivity Social science major Science major Humanities major Characteristics of MBA program Average GMAT Enrollment Private Tuition cost of MBA (US$ 000) Top tier (US News and WP) Sample Went to business school

9 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) and for women 26 years. Children were relatively few: 0.38 children on average for men and 0.28 for women. Men had a somewhat higher average GMAT score: 513 compared with 469 for women. Two demographic variables require further explanation. To avoid the expense of including several dummies in a nested logit, the college selectivity variable converts the various admissions categories in Barron s Profiles of American Colleges into a single value representing the proportion of test registrants who attended a college less-selective than that of the individual in question. 11 Values of our variable ranged from 95 (registrant s college was highly selective) to 5 (registrant s college was not selective). The variable representing predicted hourly earnings (without an MBA) is a measure of the potential opportunity cost of attending business school. It was generated for each sample member using a selection-corrected regression of hourly compensation at the primary job in Wave 4. The specification of the regression equation is based on the wage models that were estimated by Montgomery and Powell (2003) using Waves 1 3 of the GMAT registrant survey. This variable is especially useful in the nested logit, where slots for independent variables are scarce. The predicted earnings variable compresses a large number of demographic characteristics into a single index. The second set of independent variables includes characteristics of the MBA program attended, if any. These variables are drawn mainly from Barron s Guide to Graduate Business Schools (Miller, 1994) and Peterson s MBA Programs 2000 (Peterson, 2000). The tuition variable is the cost of acquiring the minimum credits to obtain an MBA part-time, or full-time, whichever is relevant. 12 The average sample member would have to pay about US$ 12,000 to obtain an MBA. The average GMAT score at schools attended by sample members was A final program characteristic is whether it was ranked among the top 20 American business schools by US News & World Report in About 11% of women and 15% of men enrolled in top-tier programs. 6. Estimation results Table 4 presents the results from our probit models of MBA completion. The first column reports the single-equation probit of the probability of completing an MBA. Here the coefficient of the female dummy is negative and five times its standard error; the effect is highly statistically significant. However, because interpretation of probit coefficients is not straightforward, in Table 5 we convert them into changes in probability (evaluated at the sample mean of the independent variables). By the calculations in Table 5, being female reduces the unconditional probability of finishing the MBA by 0.078, or about 30% of the mean completion rate of The other models in Table 4 all use bivariate probit to dissect not completing into the component parts of not enrolling and dropping out (failing to complete within 7 years) once enrolled. The coefficients in the top of the table are for the completion equation, and those in the bottom are for the enrollment equation. In the second column of Table 4 we observe that female 11 The categories were most competitive, highly competitive, very competitive, competitive and less competitive. As a test of robustness we included dummies for the actual Barrons categories in the bivariate probit models. There was virtually no effect on our key results. 12 To distinguish between in-state and out-of-state tuition we had to infer state of residence based on: (1) reported address over the first three waves of the survey and on (2) the undergraduate school attended. For about 15% of the sample, no state of residence could be established. These people were assigned out-of-state tuition. 13 For about 30% of the schools we had to infer average GMAT based on a school-selectivity variable created by Battelle from GMAC records (see Montgomery, 2002 for details). The great majority of non-reporting schools were small, regional institutions that almost certainly had average scores near the bottom of the distribution.

10 Table 4 Probit models of likelihood of GMAT registrant completing an MBA Probit Bivariate probit with sample selection All registrants All enrollees All enrollees Part-time enrollees Full-time enrollees Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Completion equation Female Married (Wave 1) Number of kids (Wave 1) Age (Wave 1) Predicted earnings Asian Black or hispanic Undergraduate GPA GMAT score Did not take GMAT College selectivity Average GMAT at B-school Enrollment at B-school B-school private B-school top tier log B-school tuition Female married Female kids Constant Enrollment equation Female Married (Wave 1) Number of kids (Wave 1) Age (Wave 1) Had FT job Wave Asian Black or hispanic M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007)

11 Table 4 (Continued ) Probit Bivariate probit with sample selection All registrants All enrollees All enrollees Part-time enrollees Full-time enrollees Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Humanities major Social science major Science major Undergraduate GPA GMAT score Did not take GMAT College selectivity Female married Female kids Constant σε 1 ε Sample in completion equation M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007)

12 186 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Table 5 How being female changes the probability of completing an MBA Probit Bivariate probit with sample selection All registrant All enrollees Part-time enrollees Full-time enrollees Sample completion rate in unconditional probability of completion in probability of completion enrollment T-statistics for in unconditional probability Completion sample GMAT registrants are both less likely to enroll in business school and less likely to complete once enrolled. As the corresponding column of Table 5 shows, being female reduces the unconditional probability of completing the MBA by Note that most of that effect, 0.063, comes from increasing the dropout rate, not discouraging enrollment. A central hypothesis of this paper is that family responsibilities prove a greater obstacle to degree completion for women than men. One approach to testing that hypothesis is to interact the female dummy with our two measures of family commitment: married (in Wave 1) and number of kids (Wave 1). The third column includes these interaction terms. The coefficient suggests that for a woman, being married decreases the probability of enrolling but increases the chances of finishing once enrolled. Children have just the opposite effect: they increase the chances of enrolling but decrease the chance of finishing. There are plausible stories that could be told to explain these seemingly contradictory results, but in no case is the coefficient on the interaction terms statistically significant at conventional levels. The last two columns in Table 4 treat part-time and full-time attendance as separate regimes. The contrast in the effect of being female is instructive. Once enrolled, women are less likely to complete either type of MBA program; the negative effect of gender is highly significant for parttimers and somewhat significant for full-timers. But the attendance models shows that women are more likely to choose part-time enrollment over the alternative options (going full-time or not at all), but less likely to choose full-time enrollment. Note, from Table 5, however, that for both part-timers and full-timers, the effect of being female on the unconditional probability of completing the MBA comes (as in the simple probit model) almost entirely through conditional probability of completing once enrolled. In other words, being female mainly makes people more likely to dropout, not less likely to enroll. The bivariate probit models in Tables 4 and 5 are limited in their ability to separate part-time and full-time enrollment. In the part-time model, for example, the only alternative to enrolling parttime is not enrolling part-time. The model cannot distinguish between the quite distinct options of: (a) choosing a full-time program instead of part-time and (b) abandoning all plans for an MBA. We address this limitation by estimating the nested logit model reported in Table 6. As shown in Fig. 1 above, we create five alternative enrollment options, which include no enrollment and enrollment in one of four types of MBA program: (i) part-time and more-selective, (ii) part-time and less-selective, (iii) full-time and more-selective and (iv) full-time and less-selective. The results of the nested logit are reported in Table 6. The top half of the table reports the enrollment nest: the coefficients in each column relate to impact of the independent variable upon the likelihood of choosing the type of MBA program for that column instead of not enrolling at

13 Table 6 Nested logit model of the likelihood of completing an MBA (n = 3762) Independent variables Enrolled part-time Enrolled full-time More-selective program Less-selective program More-selective program Less-selective program Coefficient Standard error Coefficient Standard error Coefficient T-statistics Coefficient Standard error Nest 1: enrolled in this type of program vs. no enrollment Female a Number of schools in state b Married (Wave 1) a Age Number of kids (Wave 1) a Black or hispanic b College quality b GMAT score b Had FT job in Wave 1 a Constant Nest 2: completed MBA vs. did not complete Female Married (Wave 1) Number of kids (Wave 1) Age a Earnings w/o MBA (predicted) a Constant a Coefficient varies by part-time/full-time but not by more-selective/less-selective. b Coefficient varies by more-selective/less-selective but not by part-time/full-time. M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007)

14 188 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Fig. 2. Simulated MBA completion rates for female enrollees if they were male (n = 1673). all. The bottom half is the completion nest: the coefficients relate to the likelihood of completing the relevant program versus not completing it. Some aspects of the nested logit specification bear discussion. Note that the set of independent variables is much trimmed compared to the sets in Table 4 this was necessary in order to make the model converge. The variable number of schools in state is designed to address the issue of access to different types of school. For example, many registrants live in places where there simply is no more-selective full-time program. Our model assumes that the more programs of the given type available, the more probable a registrant will find one compatible with her needs and therefore choose a program of that type. We see in Table 6 that the coefficient on this variable is positive and significant for more-selective programs but not less-selective. This suggests that the supply of programs is constraining only for those seeking the more-selective type. The bottom half of Table 6 reports Nest 2 which shows the effects of the independent variables upon the probability of completing relative to not completing. The nested logit results are consistent with those of the probit models in Table 4: the female coefficients are negative for all four types of program. The effect is highly statistically significant for the most common type of program, part-time and less-selective, and moderately significant for the other three types of program. Interpretation of nested logit coefficients is complicated. To judge the magnitude of the effect of being female on completion rates, therefore, we simulate the change in the predicted probability of completion. The simulation results are as shown in Fig. 2. The figure uses the coefficients in Table 6 to predict how the MBA completion rates of women enrollees in our sample would change if they were male instead. The simulations suggest that if these women were otherwise-identical males their completion rates would be higher for all categories of MBA program. They would be especially higher for those choosing a part-time, less-selective program, which is about half the sample. Note that the effects predicted by the nested logit are somewhat larger than reported in Table 5 for the probit models. Our central hypothesis is that women are less likely to complete MBA programs because of the greater pressure of family responsibilities. The parsimony of the nested logit model, however, does not allow us to include interaction terms between being female and being married. As an alternative we try re-estimating the model on those in our sample who were single, childless and 30 years or younger at GMAT registration. We would expect that women in this situation had no greater family responsibilities than their male counterparts (at least initially). The coefficients are not reported here, but the simulated probabilities are shown in Fig. 3. The figure shows that the

15 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Fig. 3. Simulated MBA completion rates for single, childless 30 or younger female enrollees if they were male (n = 1009). predicted completion gap is substantially reduced for all types of MBA program, but still persists. We conclude from this that family responsibilities are part, but only part, of why women who pursue MBAs are less likely to finish than their male counterparts. 7. Earnings effects How do the lower completion and attendance rates for women translate into dollars of salary lost? We simulated the hourly earnings that the women in our sample would have achieved had their MBA attendance and completion rates matched those of otherwise-identical men. This required predicting post-mba wages for each individual using a selection-corrected hourly wage regression like that used to predict non-mba wages for the nested logit (see Table 3). To make these predictions specific to the type of school attended, the regressions included dummies for whether the MBA was earned part-time or full-time and whether at a more-selective or lessselective school. Thus, we predict the hourly wage of individual i with an MBA from school of type k, and call it W ik. For convenience we can simply define k = 0 as the no-attend option and W i0 as the no-mba wage. So prior to enrolling in school the individual has five possible eventual wage outcomes: W ik (k = 1 5). Suppose, as above, we let P ik be the probability that person i chooses a type k school and P ic k be the probability she completes the MBA having done so. The expected wage associated with attending this type of school is E(W ik ) = P ik W ik + (1 P ic k )W i0 (7) The expected wage prior to any enrollment, therefore, is E(W i ) = P i0 W i0 + 4 P ik [P ic k W ik + (1 P ic k )W i ] (8) k=1 By simulating the above probabilities with the female dummy turned off, we can calculate how the expected wage for the women in our sample would change if they were male. Conducting this experiment we predict that if women had the attendance and completion rates of men, their average wages would be about 7% higher; for childless women under 30, the effect would be almost 8% These estimates should be viewed as upper bounds only. The simple Heckit regressions used to predict wages show surprisingly large wage effects of obtaining an MBA, leading us to suspect heterogeneity bias. Moreover, for computational reasons, our simulations allowed us only to change the female dummy in the completion nest, not the attendance nest.

16 190 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Summary and conclusions This paper examined the impact of gender on MBA completion rates. The effect of gender on the attainment of post-graduate degrees like the MBA should be of increasing interest to economists as the conflict between education and fertility arises later in the female life cycle than it did a generation or two ago. This paper adds to the literature on gender and school completion in several ways. This was the first study of post-graduate education to model the enrollment and completion decisions simultaneously. Also, our area of study, business, is a less traditionally male-dominated field than those examined by some of the previous studies. Finally, our sample used a broader base of degree aspirants: everyone with sufficient interest in an MBA to register to take the GMAT test. We used a number of discrete choice models to analyze enrollment and completion behavior: probit, bivariate probit with sample selection and nested logit. The various models consistently suggested that a woman who enrolls in an MBA program is about 30% less likely to complete her degree than an otherwise-identical man. The negative effect of being female was quite robust across the different models employed. This result implies that similar findings by Baker (1998) for NSF fellowship applicants are not solely due to the relative male domination of science and engineering. Why are women less likely to finish their degrees? Our hypothesis that the lower completion rates would be attributable to the differential burden of family responsibility found some support. In the probit models, interaction terms between the female dummy and the number of children and being married were not statistically significant. But simulation of the nested logit model showed that among those who were young, single and childless at the time of GMAT registration, the completion gap was narrower. It did not disappear, however. We think that the effect of gender on graduate school completion rates is a worthy one for further study. Some form of graduate study is increasingly a prerequisite for a professional career. As the study toward a terminal degree extends into the late twenties and early thirties, women will face more distractions from education, for biological if not social reasons. This could slow down the gains in relative education and earnings that women have made in the last few decades. Acknowledgements This research was supported in part by grants from Grinnell College. The authors would like to thank participants in the workshop series at the Federal Reserve Bank of Chicago for helpful comments. References Baker, J. G. (1998). Gender, race and Ph.D. completion in natural science and engineering. Economics of Education Review, 17(2), Centers for Disease Control. (1997). Natality, Tables 1 and 2. Vital statistics of the United States Centers for Disease Control. Ehrenberg, R., & Mavros, P. (1995). Do doctoral students financial support patterns affect their times-to-degree and completion probability? Journal of Human Resources, 30(3), Eide, E., Brewer, D. J., & Ehrenberg, R. G. (1998). Does it pay to attend an elite private college? Evidence on the effects of undergraduate college quality on graduate school attendance. Economics of Education Review, 17(4), Fields, J., & Casper, L. M. (2001). America s families and living arrangements: March 2000 current population reports, P Washington, DC: U.S. Census Bureau.

17 M. Montgomery, K. Anderson / The Quarterly Review of Economics and Finance 47 (2007) Fox, M. (1992). Student debt and enrollment in graduate and professional school. Applied Economics, 24(7), Ganderton, P., & Santos, R. (1995). Hispanic college attendance and completion: Evidence for the high school and beyond surveys. Economics of Education Review, 14(1), Jantzen, R. H. (2000). Price and quality effects on the demand for US business programs. International Advances in Economic Research, 6(4), Light, A., & Strayer, W. (2000). Determinants of college completion: School quality or student ability? Journal of Human Resources, 35(2), Light, A., & Strayer, W. (2002). From Bakke to Hopwood: Does race affect college attendance? Review of Economics and Statistics, 84, Light, A., & Strayer, W. (2004). Who receives the college wage premium? Assessing the labor market returns to degrees and college transfer patterns. Journal of Human Resources, 39(3), McPherson, M. S., & Schapiro, M. O. (1991). Does student aid affect college enrollment? New evidence on a persistent controversy. American Economic Review, 81(1), Miller, E. (1994). Barron s Guide to Graduate Business Schools (2nd ed.). New York, NY: Barron s Educational Series. Montgomery, M. (2002). A nested logit model of choice of a graduate business school. Economics of Education Review, 25(4), Montgomery, M., & Powell, I. (2003). Does an advanced degree reduce the gender wage gap? Evidence from MBAs. Industrial Relations, 42(3), Peterson. (2000). Peterson s MBA Programs 2000: Us, Canadian, and International Business Schools. Princeton, NJ: Petersons. Presser, H. B. (1995). Job, family and gender: Determinants of non-standard work schedules among American workers in Demography, 33(4), Schapiro, M. O., O Malley, M. P., & Litten, L. H. (1991). Progression to graduate school from Elite colleges and universities. Economics of Education Review, 10(3), South, S. J., & Spitze, G. (1994). Housework in martial and nonmartial households. American Sociological Review, 59(3), Stratton, L. (2003). Gains from trade and specialization: Division of housework among married couples. In K. S. Moe (Ed.), Women, family, and work: Writings on the economics of gender (p. 2003). Oxford: Blackwell Publishing. Van de Ven, W. P. M. M., & Van Praag, B. M. S. (1981). The demand for deductibles in private health insurance: A probit model with sample selection. Journal of Econometrics, 17(2), Weiler, W. C. (1994). Expectations, undergraduate debt and the decision to attend graduate school: A simultaneous model of student choice. Economics of Education Review, 13(1),