SEARCH

SEARCH BY CITATION

Keywords:

  • I22;
  • I28;
  • J24
  • Human capital;
  • education;
  • merit aid;
  • tuition assistance;
  • employment

Abstract

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References
Thumbnail image of graphical abstract

This paper examines how merit-based tuition assistance policies implemented by a growing number of states affect one important dimension of college graduate behaviour: the conditional probability of working in the state. Using a restricted-use administrative dataset for West Virginia, we find that, conditional on in-state enrolment and graduation, graduates that received merit aid are less likely than similar non-recipient graduates to work at establishments in the state. These results suggest that the positive influences of merit aid in West Virginia on human capital accumulation stemming from increased in-state college enrolment and graduation are dampened by the locational decisions of the recipients after graduation.

Resumen

Este artículo examina cómo las políticas de asistencia académica por méritos, implementadas por un número cada vez mayor de estados, afectan a una dimensión importante del comportamiento de graduados universitarios: la probabilidad condicional de trabajar en el estado. Mediante el uso de un conjunto de datos administrativos de uso restringido de Virginia Occidental encontramos que, condicionado a la inscripción y la graduación en el estado, los graduados que recibieron ayuda por mérito son menos propensos que los graduados no receptores similares a trabajar en establecimientos en ese estado. Estos resultados sugieren que las influencias positivas de la ayuda por mérito en Virginia Occidental en la acumulación de capital humano derivado del aumento de la matrícula y la graduación en la universidad del estado disminuyen por las decisiones de localización de los receptores después de la graduación.

Introduction

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

Federal and state governments place an enormous emphasis on the importance of providing higher education to their citizens as a means to improve the quality of the labour force, increase innovation, and boost economic growth. As a result, states have implemented policies that aim to increase the college-level educational attainment of their residents. For example, approximately one third of the states have invested in merit-based financial aid programmes for higher education. Arkansas and Georgia began the trend in the early 1990s and during the past 20 years 16 states have enacted some form of merit-based aid programme.1 Although the nature of each state's merit-aid programme varies considerably, they all aim to provide in-state tuition assistance to the best local students, regardless of income.

In 2001, West Virginia implemented the PROMISE (Providing Real Opportunities for Maximizing In-State Student Excellence) scholarship, a merit-based college tuition assistance programme, with the idea of promoting college education among West Virginia residents. In addition to the common objectives of merit-aid policies identified by Heller (2002), political leaders believed that keeping more students in West Virginia for college would increase the probability that they would stay in the state to work after graduation (DeFrank-Cole et al. 2007).2 Policy-makers value this because in-state work after graduation means increased economic growth and an expanded tax base.

Hickman (2009) presents evidence that merit-based financial aid can increase the share of college-educated residents living in a state. The ultimate impact of a merit-aid programme on state human capital accumulation can be disaggregated along three margins: (i) the impact of the programme on in-state college attendance rates (ii) the effect of the programme on college graduation rates of scholarship recipients (iii) the influence of the programme on post-graduation decisions of recipients, conditioned on in-state enrolment and graduation.

Several researchers have addressed the impact of merit-aid programmes on in-state college attendance (Dynarski 2000, 2008; Kane 2003; Cornwell et al. 2006; Goodman 2008; Pallais 2009), and for the most part the results suggest that the programmes have a positive impact.3 Less research has addressed the effect of merit-aid programmes on college graduation rates (including Dynarski 2008; Scott-Clayton 2011) and here again the available evidence suggests that the programmes increase college graduation rates. There has been no research published to date on the results at the third margin, although in a related study, Groen (2004) concludes that attending college in a state has a small positive effect on the probability of working in that state. Further, Bound et al. (2004) find that states which produce more graduates tend to have higher shares of college educated residents, although the impact is rather small.

This paper contributes to the existing literature by exploiting a new administrative dataset to estimate the likelihood that merit-aid recipients work in the state, conditional on in-state enrolment and graduation (which pertains to the third margin mentioned above). This adds to the literature by increasing our understanding of the process by which merit-aid affects local human capital accumulation. Using restricted-use, individual-level data on Bachelor's degree graduates from public higher education institutions in West Virginia and employment in the state, we estimate the conditional likelihood of PROMISE college graduates to be employed in the state. Specifically, we use quasi-experimental methods (regression discontinuity (RD) designs) to account for the impact of selection bias on the results.

We find that, conditional on in-state college enrolment and graduation, PROMISE recipients are less likely to work in West Virginia than similar in-state college graduates that did not receive the scholarship. It is important to remember, however, that the ultimate impact of the scholarship on human capital accumulation depends on all three dimensions. Even though scholarship recipients may be less likely to work in West Virginia (conditioned on enrolment at and graduation from state universities), it does appear to increase graduation rates from state public higher education institutions (Scott-Clayton 2011). Furthermore, based on the financial incentives, and the results obtained by Dynarski (2000) and Cornwell et al. (2006) for Georgia, it likely increases the probability of in-state enrolment. Thus, the ultimate impact of the scholarship on human capital accumulation in West Virginia is likely to be positive.

The paper is divided into 7 sections. The following section summarizes the structure of the West Virginia labour market and the PROMISE scholarship programme. Section 3 provides background and a review of the literature on the effects of merit-based aid on each of the three dimensions discussed above. Section 4 discusses the administrative data used for our analysis. Section 5 provides a description of the empirical approach and results of the RD analysis. Section 6 provides a back-of-the-envelope illustration of the magnitude of effects across all three margins. The paper concludes with section 7.

West Virginia and the PROMISE scholarship programme

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

Merit-based college tuition assistance programmes aim to provide financial aid to in-state college students that meet certain criteria for academic performance, but without regard to family or student financial resources. In 1999, the West Virginia Legislature authorized such a merit-based programme: the PROMISE scholarship, with the passage of Senate Bill 431. Funding, however, did not become available until 2001, when it was determined that revenues to fund the programme would come from taxation and regulation of the West Virginia video lottery machines. West Virginia created the programme with the objectives of improving the state's workforce, promoting college access, keeping the ‘best and brightest’ students in their home state for college, and rewarding students who work hard while encouraging younger students to achieve academic excellence during their high school years (DeFrank-Cole et al. 2007).

Improving the quality of the West Virginia's workforce is an important consideration for state policy-makers, since the state ranks low on many socio-economic indicators. Indeed, according to the 2011 American Community Survey from the Census Bureau, just 18.5 per cent of the state's residents age 25 and older had a bachelor's degree or better. That was well below the national rate of 28.5 per cent and ranked West Virginia last in the nation. Raising the level of educational attainment in the state may help to improve other key socio-economic metrics like labour force participation (West Virginia ranked last out of the 50 states and the District of Columbia in 2011), per capita personal income (West Virginia ranked 48th in 2011), and job growth (West Virginia ranked 46th in average annual employment growth during the 1969–2011 period).4

The structure of the PROMISE scholarship is slightly different than that in other states, although less complicated.5 According to the legislation's requirements, graduating high school students must earn a minimum 3.0 GPA in overall coursework, and a minimum score of 21 in the composite ACT in order to be eligible to receive the scholarship. However, more students than expected qualified for the scholarship and initial cost estimates turned out to underestimate true programme costs. Indeed, initial estimates suggested that funding for the first four classes of PROMISE scholars would be $27 million, but actual funding requirements turned out to be about $40 million. In 2007 the state decided to raise the minimum ACT composite score to 22 with a minimum sub-score of 22 for each of the subject areas. This would affect cohorts of high-school students starting college in 2007 or later, for which 2011 would be the earliest college graduation year.6 Alternatively, students may obtain a combined SAT score of 1,020 with a minimum score of 420 in critical reading and 480 in mathematics.7

In contrast to most state merit-aid programmes, PROMISE grantees during our period of interest were provided with a scholarship that covers full tuition in any of the state institutions, or an equivalent amount for in-state private institutions.8 In order to keep the scholarship, college students must earn a 2.75 grade point average on a 4.0 scale the first year and a 3.0 cumulative grade point average during all following years. Additionally, students must successfully complete 30 credit-hours per year (DeFrank-Cole et al. 2007).

Background and literature review

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

One goal of merit-based college tuition assistance is to increase state human capital accumulation. The role of human capital in local economic development has attracted a great deal of attention from researchers. In part, this attention derived from a desire to understand the relative importance of basic growth determinants at the local level, including human capital, public infrastructure investment, and private capital investment. One result of this research was evidence suggesting that human capital accumulation plays a large role in generating long-run local income growth (see for example, Glaeser et al. 1995; Moretti 2004; Higgins et al. 2006; Shapiro 2006; Hammond and Thompson 2008; Hammond and Tosun 2011).

The importance of human capital in long-run growth has stimulated the interest of policy-makers looking to improve local and regional economic performance. One avenue that has been proposed to increase local human capital accumulation has been the provision of merit-based financial aid to residents wishing to pursue higher education. Indeed, the legislation authorizing the PROMISE scholarship in West Virginia specifically cites (along with other issues) the fact that West Virginia ranks low among states in the share of the population with a Bachelor's degree or better.

The impact of merit-aid programmes on state college attainment rates (the share of the population with a college-level education) is concentrated on three dimensions: (i) the impact of the programme on in-state college attendance rates (ii) the effect of the programme on college graduation rates, conditioned on in-state enrolment (iii) and the influence of the programme on post-graduation decisions, conditioned on in-state enrolment and graduation.

In terms of the direct impact on in-state high-school graduates, we can summarize these dimensions and their relationship in the following way. Let HS represent the stock of high-achieving high school graduates in a state (graduates that would be eligible for merit-aid) and let pin-state College be their probability of attending college in state. Since the aid provides direct financial incentives to enroll in state, this probability needs to be conditioned on receiving the award. We can express the number of high-achieving high school graduates that are enrolled in state colleges as E = HS × pin-state College.

The number of high-achieving in-state high-school graduates that complete college in the state (G) is then the product of the number of high-achieving high school graduates attending college in state (E) multiplied by the probability of actually graduating from college (G = pgraduate × E). Thus, pgraduate is the probability of graduation, conditioned on in-state enrolment. Due to the scholarship's academic requirements and time limits, pgraduate also depends on, and is affected by, whether or not a student has been awarded the aid.

Finally, the number of high-achieving in-state high school graduates that graduate from college in the state and work in the state is given by W = pin-state residence × G, where pin-state work is the conditional probability that in-state college graduates work in the state after graduation. Note that work in the state is one of many post-graduation decisions that might be influenced by the merit-aid programme.

Merit-aid programmes have explicit financial incentives designed to raise in-state college attendance (and thus raise pin-state College), so it is natural to expect a positive impact along this dimension. The empirical results reported in Dynarski (2000) for Georgia suggest that merit-aid programmes can increase in-state college attendance of 18–19 year-olds by as much as 7.0 to 9.0 percentage points (or 25.0%). Cornwell et al. (2006), however, estimate that the scholarship in Georgia increased overall total enrolment by 5.9 per cent instead. They emphasize that the increase in college attendance arises from a reduction in students leaving the state for college. Kane (2003) reports similar results using individual data for a California higher education grant programme. In contrast, while Pallais (2009) finds that the merit-aid programme induced increases on ACT performance near the eligibility threshold in Tennessee, it did not induce more high-achieving students to apply for enrolment at in-state colleges. Goodman (2008) finds that merit-aid induced more students to enroll at public colleges in Massachusetts, but did not affect overall college enrolment.

Merit-aid programmes also have academic performance requirements that persist once the student is attending college (usually minimum GPA and course-load requirements, as well as a maximum duration of the award). Through the direct financial aid, as well as these requirements, a merit-aid programme may affect student academic performance while in college and the probability of graduation (pgraduate). While the impact of college costs on college degree completion has been studied (e.g., Angrist 1993; Bound and Turner 2002 address the impact of the GI bill on degree completion), there has been less research to date on the effect of merit-aid programmes on college completion. Dynarski (2008) estimates that the merit-aid programmes in Arkansas and Georgia reduced the college dropout rate by 3.0 to 5.0 percentage points, increasing the share of high-school graduates from these states with a college degree. Scott-Clayton (2011) presents detailed results on the effect of the West Virginia PROMISE scholarship on graduation rates and overall student achievement. The results suggest that PROMISE recipients are 9.4 percentage points more likely to graduate in four years than similar in-state college students who did not receive the award. Similarly, PROMISE recipients are 4.5 percentage points more likely to graduate in five years. Scott-Clayton (2011) also finds that the scholarship increased the overall four-year graduation rates by 7.0 percentage points and five-year graduation rates by 4.0 percentage points, for the initial waves of potential recipients.

The magnitude of the overall effect of the policy on the stock of college educated individuals in states that adopt a similar programme still depends on location decisions of college graduates. Dynarski (2008) finds a positive effect on the share of college graduates. However, Dynarski's analysis looks at college completion and does not discuss location after graduation. Groen (2011) discusses some of the existing evidence and shows that while merit-aid policies promote college attendance and completion among high-school graduates eligible, there is no clear and conclusive evidence that the programme increases the share of state residents with college degrees.

The relationship between college graduate production and the accumulation of human capital depends on the graduates' opportunities in and out of the state. Indeed, the connection between college location and work location is less than one-to-one. Groen (2004) provides indirect evidence on the effectiveness of these policies on keeping graduates in the state. Groen uses individual-level data from two separate longitudinal samples of students that attended college in the 1970s, which also tracks the students' state of residence 10–15 years after college. He finds a positive, yet modest link between college attendance in a state and remaining in that state to work after graduation. Indeed, the results suggest that the impact of attending college in-state raises the probability of working in state by about 9 or 10 percentage points. Groen (2004) argues that the causal effect of attending college in the home state is likely overstated due to the limitations on information about location preference driven college selection. He suggests that labour market impacts alone are unlikely to justify the adoption of merit-based scholarships.

The modest results support the evidence on the link between college production and the stock of college educated labour in the state. Bound et al. (2004) use state data on Bachelor's and Medical Doctor Degree production and Census data on the stock of degree holders residing in each state to measure the relationship between state degree production and state human capital accumulation. They find that the supply of college graduates in a state is modestly positively correlated with the stock of college graduates, with a positive and significant elasticity of 0.3 in the long run for Bachelor's degree holders. They find a much smaller positive correlation for medical degree recipients. Bound et al. (2004) suggest that states can only modestly influence local human capital accumulation through policies designed to affect higher-education financing.

Hickman (2009) directly analyses a merit-aid programme in Florida, using secondary data from the Census and American Community Survey. The results suggest the probability that Florida natives with college degrees between the ages of 23 and 27 will locate in Florida increased by 3.4 percentage points when the scholarship was introduced (from a baseline of 51%). While positive, this modest result suggests that the programme may indeed reduce the total effect of attending college in a home state on the local accumulation of human capital. As pointed out by Groen (2011), merit-aid programmes do not provide direct incentives for college graduates to either work or live in the state.

After graduation, the location of merit-aid recipients will be influenced, in part, by location-specific capital accumulation (i.e., the development of local networking relationships such as part-time work and/or internships). Graduates will likely shape these according to both predetermined location preferences, and future expected earnings. Merit-aid graduates with strong local family ties, for instance, will try to look for local internships even if out-of-state labour markets provide better returns. However, it is also possible that the increased human capital associated with college-level educational attainment will increase the geographical mobility of graduates, leading more of them to locate outside the state. High-achieving graduates, such as merit-aid recipients, may be offered internships in better labour markets, outside the state. In such cases, the (positive) marginal effect of attending college in the state may be reduced by the out-of-state capital accumulation.

It is important to note that none of the research on this dimension to date analyses the individual experiences of merit-aid recipients. Groen (2004) examines individual-level data, but does not have information on merit-based scholarships. Bound et al. (2004) use secondary data, which provides no direct information on merit scholarship programmes. Dynarski (2008) and Hickman (2009) also analyse secondary data for states that adopted merit-based scholarships.

This paper contributes to the literature by focusing attention on the third margin: the likelihood of scholarship recipients to work in the state. We examine this issue using a novel dataset, which contains information on the individual experience of recent graduates, including those receiving merit-aid. By looking at the influence of merit-aid on the location decisions of college graduates, our research complements previous work by closing the loop between merit-aid, college attendance and graduation, and human capital accumulation. Thus, this research is crucial to increasing our understanding of the process by which merit-aid affects local human capital accumulation, and hence the modest effects found in Hickman (2009). This information is critical to policy-makers who are investing huge sums of money in merit-based aid programmes, in the hope of increasing local human capital, economic growth, and the local tax base.

Data

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

The data used in this paper were provided by the West Virginia Higher Education Policy Commission (WVHEPC). The WVHEPC compiled demographic information on graduates from West Virginia public higher education institutions, including graduates' PROMISE scholarship status at graduation, gender, race, residency for fee purposes, age, ACT score, high-school and college GPA, graduation year, highest degree earned, current enrolment status, and area of concentration.9 Data on graduates during the 2006–2007 academic year were then matched with wage and salary employment records for 2008 maintained by Workforce West Virginia. A similar process matched graduates during 2007–2008 to employment records for 2009. Data for the match was maintained by Workforce West Virginia for one year after the end of the calendar year. After that time, individual wage records are no longer available.10

Employment is measured by place of work and covers jobs and wages reported by firms that participate in the West Virginia Unemployment Compensation system. Firms employing one or more workers for some part of a day in at least 20 different weeks of a calendar year are required to contribute to the state's unemployment insurance system. The advantage of this dataset is that it allows us to investigate the conditional likelihood of in-state employment for merit-aid recipients, which is new in the literature. However, the data do not allow us to account for graduates that may have gone on to further study out of state, remained in state but could not find work or chose not to work, or found work in another state.

Further, these data exclude certain classes of employers, namely railroad companies and the federal government, which contribute to different unemployment insurance systems.11 In addition, this dataset does not cover the self-employed, student workers, most church workers, and unpaid family workers. A graduate is counted as employed in the state if he/she is employed for at least one quarter during the year of measurement (2008 for 2006–2007 graduates, and 2009 for 2007–2008 graduates).

Our data on employment in the state may generate different conclusions than data on state residence. This is because employment is measured by place of work, not by place of residence. When local labour markets spill across state lines, these two concepts (work and residence) do not necessarily match very closely. This may be an issue for legislators in West Virginia concerned about residents rather than workers. There are 10 metropolitan statistical areas with component counties in West Virginia, and some of these do in fact cross state borders. Thus, while some college-educated individuals may reside in West Virginia, they may actually work in another state. The opposite, however, is also true for some college graduates who may reside in another state, but work in West Virginia.12

Since we are concerned with the influence of merit-based financial aid on the conditional probability of in-state employment, we limit our sample to include graduates that are potentially eligible for the award only. Hence, we exclude those classified as out-of-state residents for fee purposes, as well as those with high-school GPAs lower than 3.0.13 Students that meet these criteria are eligible to receive PROMISE so long as they score above the ACT threshold (or SAT equivalent).

Our dataset includes 10,060 in-state graduates from public higher education institutions during the 2006−2007 academic year, and 10,220 during the 2007−2008 year. Of those that graduated during 2006−2007, 4,633 met the GPA requirement. This number is similar to the 4,876 that met the requirement in the second cohort of graduates. Table 1 summarizes graduate characteristics and in-state employment rates of these potentially eligible graduates in West Virginia.

Table 1. West Virgina public higher education graduates and West Virginia employment in 2008 and 2009 graduates meeting the high-school GPA requirement for PROMISE scholarship
 Graduates During 2006–2007 Working in WV inGraduates During 2007–2008 Working in WV inGraduates During 2006–2007 and 2007–2008
Non-PROMISE GraduatesPROMISE GraduatesNon-PROMISE GraduatesPROMISE GraduatesNon-PROMISE GraduatesWorking in WVPercent of Non-PROMISE Working in WV after GraduationPROMISE GraduatesWorking In WVPercent of PROMISE Working in WV after Graduation
  1. Notes: These are the authors' calculations based on a sample of PROMISE eligible, in-state graduates during the academic years 2006–2007 and 2007–2008. Total number of In-State Graduates during the two years was 20,280. After restricting the sample to those that are eligible to receive PROMISE based on the high-school criterion (HS GPA >= 3.0) and dropping observations with missing data, the total number of graduates is 9,509. n/d: ‘Not Disclosed’ due to confidentiality requirements.

Total74.2%63.5%73.8%64.1%5,2753,90374.0%4,2342,70263.8%
Gender          
Female76.1%64.9%77.4%66.4%3,3022,53576.8%2,5791,69465.7%
Male71.1%61.2%67.4%60.7%1,9731,36869.3%1,6551,00860.9%
Race          
White74.1%63.8%74.2%64.4%4,9923,70274.2%4,0872,62064.1%
Non-White76.0%56.2%65.4%55.4%28320171.0%1478255.8%
Graduated and Enrolled73.2%62.8%73.5%56.2%61244973.4%89453059.3%
Mean Age    25.64  23.20  
Mean ACT Score    20.78  24.38  
Mean College GPA    3.15  3.43  
Degree          
Associate77.8%81.0%80.5%82.9%98678179.2%44836781.9%
Bachelor74.5%60.7%73.5%60.6%2,8372,09974.0%3,4492,09260.7%
First Professional63.5%n/d58.0%77.8%31619361.1%272177.8%
Masters73.2%59.3%72.2%69.3%99572472.8%24316367.1%
Masters Certificaten/dn/dn/dn/dn/d2n/dn/d0n/d
Doctoraln/dn/d62.5%n/d342058.8%n/d0n/d
Two-Year College77.7%82.3%0.0%83.2%85667278.5%33727982.8%
Area of Concentration          
Agriculture69.0%42.5%61.5%44.0%684464.7%903943.3%
Architecturen/dn/dn/dn/dn/d2n/dn/d0n/d
Bio Sciences48.1%41.2%58.7%47.9%1256552.0%35515944.8%
Business76.4%66.2%75.2%64.1%83563375.8%74448465.1%
Comm. Technologyn/dn/dn/d100.0%171270.6%181688.9%
Communication68.5%60.0%71.6%62.1%18713170.1%19712060.9%
Computer & Info Sci.55.2%69.2%75.9%66.7%583865.5%624267.7%
Culinaryn/dn/dn/dn/dn/d6n/dn/d1n/d
Education81.4%69.6%83.5%80.7%89874082.4%46835475.6%
Engineer Technology73.3%67.5%78.0%69.2%1259475.2%795468.4%
Engineering54.9%46.2%68.5%45.7%1257660.8%24211145.9%
English & Literature70.0%54.2%58.6%61.5%694565.2%1005858.0%
Family Sciences69.0%62.5%78.1%52.4%614573.8%372156.8%
Foreign Languages54.5%53.8%54.5%38.5%221254.5%261246.2%
General Studies71.9%68.7%68.4%68.9%43930870.2%1419768.8%
Health Professions77.2%76.8%71.8%75.2%1,04978274.5%66050175.9%
History72.2%78.9%80.0%46.7%433376.7%835161.4%
Legal Professions74.2%n/d74.4%n/d1057874.3%n/d2n/d
Math and Statisticsn/dn/dn/d35.0%n/d6n/d281242.9%
Mechanic & Repair Techn/dn/dn/dn/dn/d10n/dn/d3n/d
Multidisciplinary Studies81.6%62.5%64.3%68.2%916773.6%1006565.0%
Natural Resources82.8%66.7%62.9%67.9%644671.9%523567.3%
Philosophyn/dn/dn/dn/dn/d1n/dn/d2n/d
Physical Sciences73.9%38.1%55.6%55.3%412765.9%894247.2%
Precision Designn/dn/dn/dn/dn/d4n/dn/d1n/d
Psychology69.2%64.7%72.3%59.6%16111470.8%18411462.0%
Public Administration64.3%66.7%75.9%80.0%966870.8%403075.0%
Recreation64.1%61.9%56.7%n/d694260.9%311754.8%
Science Technology85.7%n/d86.4%n/d363186.1%n/d9n/d
Security79.2%70.6%81.4%78.3%19815980.3%947175.5%
Social Sciences81.3%62.2%72.0%55.3%14611176.0%18410858.7%
Visual & Perf. Arts69.2%78.0%66.1%58.2%1087367.6%1057167.6%

PROMISE and non- PROMISE graduates had similar general characteristics, with 60.9 per cent of PROMISE graduates female, 96.5 per cent white, 81.5 per cent earned a Bachelor's degree, and 47.5 per cent majored in Business, Education, General Studies, or Health Professions. PROMISE graduates were on average a bit younger than the average graduate (23.2 years versus 25.6), had higher mean ACT scores (24.4 versus 20.8), and had higher college GPAs (3.4 versus 3.2).

Table 1 also presents West Virginia employment rates for graduates, defined as the number of graduates working at establishments in the state divided by the total number of graduates. Of the non-PROMISE graduates, 3,903 (or 74.0%) worked at establishments located in West Virginia for at least one quarter in the year immediately after graduation. Employment rates tended to be higher for women and for white graduates. We also note that employment rates tended to be higher for those earning Associate's degrees.

For PROMISE graduates, the in-state employment rate one year after graduation was lower than that of non-PROMISE graduates at approximately 63.8 per cent. We again find that employment rates were higher for PROMISE graduates that are female, white, and those that earned Associate's degrees. Note, however, that the employment rate of PROMISE recipients with a Bachelor's degree (education level obtained by most of the award-winning graduates in our sample) was much lower than that for non-PROMISE (74.0% versus 60.7%).

In order to disentangle the influences on work participation in West Virginia, we now turn to an econometric analysis. This will allow us to estimate the likelihood of PROMISE-scholarship graduates of working in the state, conditional on in-state enrolment and other important characteristics, which help reduce the influence of selection bias as much as possible. Since the focus of PROMISE is on the transition from high school to college, we restrict our econometric analysis to those who earned a Bachelor's degree only.

Empirical method and results

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

Methodology

A major issue in estimating the influence of a merit-aid programme on residence or work in a state arises from the fact that the scholarship targets high-achieving students. This can create a selection bias because high-achieving students are not only likely to be the most mobile, but also more prone to continue their education without the need to work while in school. Thus, naïve estimates will tend to find a negative impact of the scholarship on residence and/or work participation. One response to this issue is to include the measures of individual ability that are available and to structure the sample in ways designed to compare individuals with similar mobility characteristics (so that the only difference is the receipt of the scholarship).

We employ a quasi-experimental approach to investigate the impact of the PROMISE scholarship on the probability of remaining in West Virginia for work, conditional on in-state college enrolment and graduation. Using a regression discontinuity (RD) design (Imbens and Lemiux 2008), we estimate the effect of crossing the ACT threshold needed for students to qualify for the scholarship. In theory, graduates' characteristics, employment opportunities, and other policy conditions should be the same for those just above and just below the threshold (i.e., one or two ACT points above and below).

Following Scott-Clayton (2011) we use a local linear regression (LLR) of the form:

  • display math(1)

where yi is an indicator of employment in West Virginia for individual i, Abovei (Belowi) is an indicator that the individual is above (below) the threshold (21 for these early cohorts of graduates), ACTdisti is the distance between the student's individual score and the underlying cutoff score, and X is a matrix including demographic variables such as gender, race, age and area of concentration. The parameter β estimates the difference in outcomes at the threshold.14 As Scott-Clayton argues, (after limiting the sample to those with GPA at or above 3.0) the ACT score alone is a strong predictor of PROMISE receipt. If being above the threshold and receiving the scholarship serves as a positive indication for employers, it may provide additional opportunities in more attractive labour markets outside the state borders. Through the RD design we isolate the effect of the programme, ruling out other explanatory factors (Scott-Clayton 2011). We present results separately for 2008 and 2009, in order to investigate whether the estimates are affected by the business cycle. West Virginia employment peaked in 2008 and declined during 2009. We also present pooled results to estimate an average impact.

Figures 1 and 2 show the raw means of employment for Bachelor's degree graduates in West Virginia for 2008 and 2009 by ACT score (including linear predictions). As the figures show, we find discontinuities around the ACT threshold, with lower work participation rates above the threshold than below it. This is particularly evident in 2009.

figure

Figure 1. West Virginia employment rates of graduates with Bachelor's degrees in 2008 by ACT score: size of circle indicates the number of students with the ACT score

Download figure to PowerPoint

figure

Figure 2. West Virginia employment rates of graduates with Bachelor's degrees in 2009 by ACT score: size of circle indicates the number of students with the ACT score

Download figure to PowerPoint

Results

Regression results based on Equation (1) above yield ‘intent-to-treat’ (ITT) effects, which are estimates generated assuming all qualified graduates apply for and accept the scholarship. Results for these regressions are presented in Table 2.

Table 2. Intent-to-treat regression discontinuity results estimate of the effect of West Virginia's PROMISE scholarship on work participation one year after graduationa
 Coefficient for ABOVE (2008)Coefficient for ABOVE (2009)Coefficient for ABOVE (Pooled)b
  1. Notes: This analysis uses administrative data from WVHEPC and worker's compensation record for Bachelor's degree graduates of West Virginia Public Higher Education institutions and their work participation in the state the year immediately after they graduated. It uses the cohort of graduates from 2007, receiving wages in 2008, and the cohort of graduates from 2008 receiving wages in 2009. Both datasets include the graduates that met the high school GPA requirements for PROMISE (3.0+) only, and for which information is available. a Robust Standard Errors are clustered by values of the running variable (ACT score), following Lee and Card (2008) for the case of discrete running variables. b The pooled model includes both measurement years (2008 and 2009) and includes a year dummy variable to account for the difference in year of graduation and/or employment. c The Simple Model represents the LLR of the dependent variable (Work Participation in this case) on the forcing variable only. d This model includes a quadratic function of the distance between the ACT score and the cut-off point. e Individual characteristics are additional variables that control for gender, race, age, and college GPA. f Adds controls for the area of concentration studied (e.g. Mathematics, Engineering, etc). Stars indicate the significance of findings at the 1% level (***), 5% level (**), and 10% level (*).

Small BandwidthSimple Modelc−0.0031−0.0345−0.0188
(18 <= ACT <= 23) (0.030)(0.019)(0.025)
 With a Quadratic Termd−0.1956***−0.1381***−0.1670***
  (0.000)(0.000)(0.002)
 With Individual Characteristicse−0.0104−0.0253*−0.0183
  (0.026)(0.010)(0.018)
 With a Quadratic Term−0.1801***−0.0955***−0.1362***
  (0.018)(0.008)(0.008)
 Full Modelf−0.0180−0.0365*−0.0207
  (0.032)(0.014)(0.019)
 With a Quadratic Term−0.1863***−0.1142***−0.1339***
 (0.020)(0.014)(0.014)
Number of Observations 1,6691,7033,372
Medium BandwidthSimple Model−0.0383−0.0518−0.0453
(16 <= ACT <= 25) (0.022)(0.031)(0.025)
 With a Quadratic Term0.01940.01420.0169
  (0.047)(0.043)(0.045)
 With Individual Characteristics−0.0468*−0.0433−0.0455
  (0.023)(0.033)(0.025)
 With a Quadratic Term0.01300.03040.0217
 Full Model−0.0489*−0.0448−0.0420
 With a Quadratic Term0.00810.02140.0226
 (0.048)(0.044)(0.042)
Number of Observations 2,3692,4544,823
Large BandwidthSimple Model−0.0456**−0.0456**−0.0396*
  (0.017)(0.017)(0.018)
(11 <= ACT <= 30)With a Quadratic Term−0.0611***−0.0611***−0.0553***
  (0.017)(0.017)(0.014)
 With Individual Characteristics−0.0561**−0.0340−0.0452**
  (0.018)(0.028)(0.017)
 With a Quadratic Term−0.0464−0.0004−0.0142**
  (0.038)(0.031)(0.004)
 Full Model−0.0575**−0.0246−0.0395**
  (0.018)(0.028)(0.016)
 With a Quadratic Term−0.0558−0.0068−0.0253
 (0.039)(0.032)(0.025)
Number of Observations 3,0383,1036,141

The estimates arise from a series of different model specifications, different bandwidths, and different data sets based on Equation (1) above. The first column shows the ITT effects of the programme on the in-state work experience in 2008 for Bachelor's degree graduates during the 2006–2007 academic year. The second column shows analogous results for the next cohort of graduates (graduates from 2007–2008 working in the state in 2009), while the third column shows the results for the two panels combined. The combined results can be interpreted as the average treatment effect of the programme on the work experience of potentially eligible graduates the year after graduation. We use a dummy variable to account for the graduation year. As suggested by Lee and Card (2008), we cluster the standard errors by values of the assignment variable. In addition, while many RD analyses use non-parametric estimation of LLR, this cannot be done for discrete running variables given the increased distance between the actual scores and the threshold.

To check the robustness of our results, we use three different bandwidths and three different models, where each model includes additional control variables. The simple model includes no control variables (X is excluded). The model with individual characteristics includes gender, race, age, and college GPA. The full model includes all individual characteristics, as well as area of concentration as defined in Table 1. In addition, we present results with and without a local quadratic term ACTdist to account for functional form issues.15

The results in Table 2 suggest that PROMISE eligible graduates are less likely to work in West Virginia (conditioned on in-state enrolment and graduation) than are similar graduates that could not receive the scholarship. Most of the estimated coefficients are negative and many are significant at conventional significance levels. Overall, the size of the estimates seems reasonable, with significant negative coefficients in the pooled results ranging from −0.167 to −0.014.

The ITT results assume that all qualified graduates apply for and accept the scholarship. However, in our data, being above the threshold does not necessarily translate into treatment. The reasons for this are not clear, but could include imperfect take up, as well as outright misclassification of graduates (see Scott-Clayton 2011, for more on this issue). Therefore, to estimate the effects of receiving the scholarship as opposed to just being above the threshold, we combine an instrumental variables approach with the RD design, which gives ‘treatment-on-the-treated’ (TOT) effects. In a local linear regression, these take the form:

  • display math(2)
  • display math(3)

where Pi represents actual PROMISE receipt, inline image represents predicted PROMISE receipt, and all other variables are defined as before.

These results are presented in Table 3 and are similar to those generated by the ITT estimates. We again find that the majority of estimates are negative and many are significantly negative. The TOT coefficient estimates overall are a bit larger that the ITT estimates, with the significant coefficients in the pooled results ranging from −0.292 to −0.064.

Table 3. Treatment-on-the-treated regression discontinuity results estimate of the influence of West Virginia's PROMISE scholarship on the work participation one year after graduationa
 Coefficient for PROMISE(2008)Coefficient for PROMIS E(2009)Coefficient for PROMISE (Pooled) b
  1. Notes: This analysis uses administrative data from WVHEPC and worker's compensation record for Bachelor's degree graduates of West Virginia Public Higher Education institutions and their work participation in the state the year immediately after they graduated. It uses the cohort of graduates from 2007, receiving wages in 2008, and the cohort of graduates from 2008 receiving wages in 2009. Both datasets include the graduates that met the high school GPA requirements for PROMISE (3.0+) only, and for which information is available. a Robust Standard Errors are clustered by values of the running variable (ACT score), following Lee and Card (2008) for the case of discrete running variables. b The pooled model includes both measurement years (2008 and 2009) and includes a year dummy variable to account for the difference in year of graduation and/or employment. c The Simple Model represents the LLR of the dependent variable (Work Participation in this case) on the forcing variable only. d This model includes a quadratic function of the distance between the ACT score and the cut-off point. e Individual characteristics are additional variables that control for gender, race, age, and college GPA. f Adds controls for the area of concentration studied (e.g. Mathematics, Engineering, etc). Stars indicate the significance of findings at the 1% level (***), 5% level (**), and 10% level (*).

Small BandwidthSimple Modelc−0.0051−0.0609−0.0320
(18 <= ACT <= 23) (0.049)(0.036)(0.042)
 With a Quadratic Termd−0.2885***−0.2923***−0.2915***
  (0.000)(0.000)(0.003)
 With Individual Characteristicse−0.0188−0.0472*−0.0335
  (0.047)(0.020)(0.034)
 With a Quadratic Term−0.2653***−0.2042***−0.2404***
  (0.025)(0.019)(0.013)
 Full Modelf−0.0328−0.0672*−0.0376
  (0.058)(0.028)(0.034)
 With a Quadratic Term−0.2827***−0.2504***−0.2376***
  (0.024)(0.036)(0.024)
Number of Observations 1,6691,7033,372
Medium BandwidthSimple Model−0.0618−0.0886−0.0752
(16 <= ACT <= 25) (0.036)(0.053)(0.042)
 With a Quadratic Term0.03110.02630.0290
  (0.076)(0.078)(0.077)
 With Individual Characteristics−0.0820*−0.0774−0.0802
  (0.040)(0.058)(0.043)
 With a Quadratic Term0.02270.06070.0406
  (0.076)(0.068)(0.073)
 Full Model−0.0859*−0.0794−0.0739
  (0.041)(0.057)(0.042)
 With a Quadratic Term0.01410.04260.0418
  (0.083)(0.086)(0.077)
Number of Observations 2,3692,4544,823
Large BandwidthSimple Model−0.0727**−0.0548−0.0636*
(11 <= ACT <= 30) (0.027)(0.043)(0.029)
 With a Quadratic Term−0.0683−0.0403−0.0561
  (0.057)(0.050)(0.045)
 With Individual Characteristics−0.0938**−0.0555−0.0744**
  (0.030)(0.046)(0.028)
 With a Quadratic Term−0.0862−0.0007−0.0456
  (0.071)(0.062)(0.047)
 Full Model−0.0969**−0.0400−0.0652**
  (0.031)(0.045)(0.025)
 With a Quadratic Term−0.1037−0.0136−0.0488
  (0.073)(0.064)(0.048)
Number of Observations 3,0383,1036,141

It is important to note at this point that since the RD approach uses local linear estimation, it does not estimate the influence of the programme for those well above the ACT threshold (e.g., ACT of 27 or above). In other words, in the RD we do not account for potential decreasing (or increasing) marginal effects in ACT score. While this is a limitation, as Lee and Lemieux (2010) point out, ‘[T]he RD estimate can be interpreted as a weighted average treatment effect, where the weights are the relative ex-ante probability that the value of an individual's assignment variable will be in the neighborhood of the threshold’ (p. 292). This is analogous to choosing a bandwidth which assigns very low weights (or none at all) to students with high ACT scores. This is justified on the grounds that it is likely that post-assignment outcomes such as graduation rates and in-state employment outcomes can be mainly driven by their individual abilities and characteristics, rather than by the receipt of the scholarship.

Ideally, however, in the RD analysis the individual characteristics of graduates just below and just above the threshold, as defined above, should not differ significantly. Table 4 presents the results of tests for differences across the ACT threshold for selected control variables.16 These regressions have the same general form as Equation (1), except the controls become dependent variables (and we exclude other control variables). As the table shows, we find several instances in which the controls are statistically significantly different across the ACT threshold for the medium bandwidth, which is not what we would hope. However, the differences tend to be very small in magnitude. For instance, graduates above the threshold tend to be younger than graduates below the threshold by only 0.4 years on average. We address this issue in the main regression by controlling for other characteristics such as gender, race, age, enrolment, degree, and area of concentration. Tables 2 and 3 show that the negative results tend to persist whether we add controls or simply use a specification with the assignment variable only.17

Table 4. Test for differences in basic demographic characteristics around ACT threshold (standard errors in parentheses)a
 Coefficient for ABOVE (2008)Coefficient for ABOVE (2009)Coefficient for ABOVE (Pooled)b
  1. Notes: This analysis uses an LLR specification similar to that in equation (1), but in this case the dependent variable is the corresponding covariate and no other controls added. The pooled model includes both measurement years (2008 and 2009) and includes a year dummy. a Robust Standard errors are in parenthesis and are clustered by the values of the running variable (ACT score), following Lee and Card (2008) for the case of discrete running variables. Stars indicate the statistical significance of findings at the 1% level (***), 5% level (**), and 10% level (*).

Medium BandwidthFemale−0.0394−0.0440*−0.0418**
(16 <= ACT <= 25) (0.022)(0.019)(0.014)
 Non White0.02410.00420.0142
  (0.026)(0.005)(0.015)
 Age−0.4860***−0.3151***−0.3933***
  (0.101)(0.064)(0.073)
 College GPA0.0650**0.0602*0.0623***
  (0.022)(0.092)(0.008)
Number of Observations 2,3692,4544,823
Table 5. Falsification exercise for graduates before 2002 using ITT estimates (standard errors in parentheses)a
 Coefficient for ABOVE (2008)Coefficient for ABOVE (2009)
  1. Notes: This analys is uses administrative data from WVHEPC and worker's compensation record for Bachelor's degree graduates of West Virginia Public Higher Education institutions before 2002 and their employment in the state in 2008 and 2009. Both datasets include the graduates that met the high school GPA requirements for PROMISE (3.0+) only, and for which information is available. a Robust Standard Errors are clustered by values of the running variable (ACT score), following Lee and Card (2008) for the case of discrete running variables. b The Simple Model represents the LLR of the dependent variable (Work Participation in this case) on the forcing variable only. c Individual characteristics are additional variables that control for gender, race, age, and college GPA. d Adds controls for the area of concentration studied (e.g. Mathematics, Engineering, etc). Stars indicate the significance of findings at the 1% level (***), 5% level (**), and 10% level (*).

Medium BandwidthSimple Linear Modelb0.00870.0220
(16 <= ACT <= 25) (0.016)(0.014)
 With Individual Characteristicsc0.01060.0245
  (0.017)(0.015)
 Full Modeld0.01170.0253
  (0.019)(0.017)
Number of Observations 4,9534,834

A remaining issue is that the cut-off point for both ACT and GPA is known to the high-school student before graduation. The student is not only able to work harder in class, but is also allowed to retake the ACT exams several times. Consequently, the approach may be sensitive to selection problems around the threshold. Indeed, those students who are induced to work extra hard and/or retake the exam several times in order to become eligible can be systematically different in unobserved ways from those who were eligible before the implementation of PROMISE. An examination of the distribution of graduates by ACT score does suggest that the number of graduates with scores at and beyond the threshold slightly increased after the programme was implemented (similar to results reported in Scott-Clayton 2011).

For the most part, individuals will hardly have precise or perfect control over the assignment variable and under imprecise control over the assignment variable the ‘variation in the treatment in a neighborhood of the threshold is “as good as randomized” ’ (Lee and Lemieux 2010, p. 292). While not a significant problem, the fact that these differences can provide some control over the treatment leads us to believe that the negative results could be partly driven by these unobservables.

Unobserved differences between high-school graduates below and above the threshold may include motivation, family ties, or locational preferences.18 These differences can in fact shape the incentives to qualify for the scholarship in the first place. For some students, for instance, the scholarship may represent more than just a reduction of college costs. At the margin, students with location preferences outside the state may view the award as a positive signal to employers in other labour markets. Other things equal, students around the threshold may seem better to out-of-state employers if they have a college scholarship in their resume. While this is also true for in-state employers, this signal can help students looking to move to more competitive markets outside of West Virginia.

College achievement may also yield a higher share of eligible graduates leaving the state. Scott-Clayton (2011) shows that, at the margin, the scholarship promoted college achievement by providing financial incentives through the renewal requirements. Her work showed that the scholarship not only improved graduation rates, but also college GPAs. Because higher ability/achieving students have a higher tendency to leave West Virginia in search for better labour markets, higher college achievement also supports the findings that once graduated, PROMISE recipients are less likely to work in the state.

A third factor, which is not completely independent of the two factors mentioned above, is the probability of enrolling in graduate school. Scholarship recipients may be more likely to apply and be accepted into graduate programmes. On one hand, many schools in West Virginia do not offer graduate degrees. Some offer a limited number of Master's degrees, and very few offer higher degrees. Thus, many of these students may enrol in programmes outside the state. On the other hand, as our data show, some do enrol in graduate programmes in West Virginia. Depending on assistantships and other sources of funding arrangements, full-time graduate students may appear in our data set as not being employed in the state, even if they live in West Virginia.

It is not clear which of the aforementioned mechanisms explains the negative estimates. In fact, our results may be driven by a combination of all, including the unobserved traits. Since we cannot control for locational preferences, family ties, or personal motivation, we test our results using a cohort of students without access to the scholarship. Presumably, these characteristics should not be affected by the scholarship. Thus, an examination of a different cohort allows us to assess whether the scholarship itself influences the location decisions of college graduates.

Robustness

We now consider a falsification exercise similar to that in Scott-Clayton (2011), by examining the work experience of graduates not affected by the PROMISE scholarship. In particular, we analyse the work experience in 2008 and 2009 of individuals that graduated from college before 2002 and thus were not subject to the scholarship programme. Thus, we expect to find no significant impact around the threshold. Using the ITT regression in Equation (1), we estimate each of the three models (all include a quadratic term). The results are summarized (using the medium bandwidth only) in Table 5 and they show no significant overall impact around the threshold in the conditional probability of working in state after graduation.

Overall, our estimates are robust to the choice of bandwidth and the inclusion of quadratic terms. The regression discontinuity results suggest that, conditional on enrolment and graduation, PROMISE-scholarship recipients are less likely to be employed in West Virginia after graduation. More specifically, the results from the pooled TOT analysis indicate that the negative estimate on PROMISE is around 7.0 to 8.0 percentage points, on average. This can be interpreted as follows: if the number of PROMISE recipients doubled at the threshold, the work participation rate for these marginal graduates will be likely to decrease by 7.0 to 8.0 percentage points from the baseline (75.2% at ACT = 20).

Policy analysis

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

The results above suggest that, conditioned on other outcomes, a merit-based scholarship recipient is less likely to work in the state after college graduation. It is important to keep in mind that the estimates found operate at the margin. Thus, the reduced likelihood at the third margin does not necessarily mean that the programme reduces the total share of local graduates working in the state. These results do suggest that states which are considering adopting merit-based scholarship programmes should be aware that the total impact of the programme on local human capital accumulation may be lower than expected.

The descriptive statistics show that the majority of PROMISE graduates so far have remained in the state to work. As noted, such merit-based aid tends to increase the share of high-achieving students enrolled in and graduated from in-state colleges and universities. Hence, after accounting for a lower likelihood of in-state employment on the third margin, such a policy may induce more high-achieving high-school graduates to enroll, graduate, and work in the state, and thus increase the overall quantity and quality of local college educated workers. This would be consistent with the overall positive results found in Hickman (2009).

We now present a back-of-the-envelope calculation to illustrate how our estimate of the third margin interacts with the first two margins. In Table 6, we analyse the potential impact of the scholarship on the in-state work participation of the first three cohorts of West Virginia high-school graduates subject to the implementation of PROMISE.19 As shown in panel A of Table 6, between the 2001–2002 and 2003–2004 academic years, West Virginia reported a total of 55,150 students in 12th grade (approximately 18,000 per cohort). The West Virginia Department of Education (WVDE) reported that high-school graduation rates in the state have oscillated around the 83.0 percent mark in the past decade. Applying this rate to the total number of high-school seniors, we estimate that the state graduated about 45,775 high-school seniors (or approximately 15,200 per year).

Table 6. Potential marginal impact of PROMISE on in-state work participation ratesThumbnail image of

Panel A of Table 6 then illustrates the first margin: the impact of the scholarship on in-state college enrolment. Scott-Clayton (2011) roughly estimated that 14.0 percent of high-school graduates either leave the state or enrol in a private college. Also, prior to the implementation of PROMISE, roughly 39.0 per cent enrolled in public higher education institutions, with approximately 24.0 per cent being non-eligible for the award and the remaining 15.0 per cent being above the threshold. Scott-Clayton (2011) then estimated that when PROMISE was introduced in West Virginia, overall enrolment rates went up by approximately 4.0 percentage points (to 43.0%).20 As line 6 in Table 6 shows, the enrolment rate for students above the threshold (eligible) increased by 5.0 percentage points. However, this marginal increase was slightly offset by a 1.0 percentage-point decrease in the enrolment rate of non-eligible students. Presumably, such offset may be due to some graduates crossing the threshold and becoming eligible.

The last columns of panel A use the rates in the first two columns to illustrate the marginal effect of PROMISE on enrolment levels. As shown, the increase of 5.0 percentage points for eligible high-school graduates resulted in approximately 2,289 more enrollees in in-state colleges and universities. Taking into account the small drop in non-eligible enrollees, overall enrolment at in-state universities rose by 1,831, which illustrates the positive impact of the scholarship at the first margin.

Not all enrollees graduate from college. Scott-Clayton (2011) estimated that in West Virginia, the probability of graduating in four years for those students just below the cut-off point for the scholarship was approximately 15.6 per cent, while the probability of doing it in five years was 36.7 per cent. She also estimated that, at the margin, the probability of PROMISE enrollees graduating in four years was 9.4 percentage points higher than the one for those below the cut-off point. Similarly, she estimated that PROMISE increased the probability of graduating in five years by 4.5 percentage points. These probabilities determine the second margin and are presented in panel B of Table 6 (see lines 8 and 9).21 We combine these estimates with the numbers on line 6 (last two columns of panel A) to calculate the approximate number of enrollees above and below the cut-off that would graduate within four and five years.

Overall, using the benchmark rates, we find that approximately 5,506 of high-school graduates non-eligible for PROMISE from the 2001–2002 to 2003–2004 cohorts graduated from West Virginia higher education public institutions between the 2005–2006 and 2008–2009 academic years. Accounting for the marginal effect found by Scott-Clayton (2011), this number is slightly higher for PROMISE-eligible students at 6,061, which illustrates the joint impact of the scholarship at both the first and second margins.

In panel C we estimate the third margin. In it, we calculate the number of PROMISE eligible and non-eligible college graduates employed in the state. Our empirical results suggest that of those graduates who were not eligible for the scholarship (those below the threshold), an average of 75.2 per cent worked in the state after graduation, and that at the margin this percentage is 7.0 to 8.0 percentage points lower for PROMISE recipients. Hence, the in-state employment of graduates above the threshold is approximately 67.2 to 68.2 per cent. Applying these rates, and our most conservative estimate (–7.0 percentage points) to the numbers found in line 10 (panel B),22 we find a total of 4,141 non-eligible and 4,133 PROMISE-eligible individuals employed in the state after graduation. This shows the joint impact of the scholarship across all three margins. In order to measure the overall impact of the scholarship on the share of workers with college degrees, one must look at the numbers as a percent of all high-school graduates below and above the threshold.

We estimate the work participation rates for West Virginia high-school graduates above and below the threshold in panel D of Table 6. Assuming that the ACT distribution of the entire high-school graduate population follows that of those that enrolled in public colleges and universities in line 7, we estimate that out of the 45,775 high-school graduates, approximately 24,484 were below the threshold, being ineligible for the scholarship. Similarly, out of those that graduated high-school between 2002 and 2004, approximately 21,290 were eligible to receive the award. These numbers suggest that 16.9 percent of high-school graduates below the threshold remained to work in the state after graduation. In contrast, the in-state employment rate of merit-aid recipients was approximately 2.5 percentage points higher (at 19.4%). Our illustration suggests that the scholarship may have had a positive impact on human capital accumulation in West Virginia, even though we estimate a negative impact at the third margin.

Although the overall impact of the scholarship was likely positive, these calculations and the results obtained by Hickman (2009) suggest that the policy had a modest impact. As mentioned above, our empirical results suggest that the scholarship may indeed serve as a tool for some graduates to seek employment opportunities elsewhere, and thus reducing the overall impact that policy-makers had expected. This may be a particular issue for West Virginia, which already has a very low educational attainment rate, tends to generate slow job growth, and has an economic base that is heavily influenced by mining activity (primarily coal, but increasingly natural gas extraction). This sector tends to employ far fewer college graduates in the state than do sectors such as education and health care (Bowen 2013).

Indeed, states planning to adopt policies aiming at increasing in-state educational attainment should contemplate additional alternatives. Groen (2011), for instance, suggests that basing financial aid on need rather than merit might be a more effective way to increase the supply of college graduates in a state's labour market, though the evidence on this is also not conclusive. Alternatively, degree production can be concentrated among certain groups. While increasing the share of residents with a Bachelor's degree seems important, labour demand in some states may be concentrated around those with technical degrees. Promoting access to technical careers and two-year colleges may be more effective at meeting some states' labour demands (see Rouse 1998, for a detailed discussion).

Alternatively, states may look for ways to reinforce local labour markets. Partnerships between local companies and universities that promote internships may broaden the students' local networks. These partnerships may also help retain out-of-state students that may otherwise return to their home state after graduation. Finally, programmes that increase the competitiveness of local labour markets help in attracting graduates that educated in other states. As Groen (2011) points out, college-educated workers are equally valuable whether they graduated in or out of the state.

Conclusion

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References

The results presented in this paper suggest that merit-based scholarship recipients are less likely to be employed in the state, conditional on in-state college enrolment and graduation. Thus, after controlling for a wide range of demographic characteristics, as well as attempts to reduce the impact of selection bias, we find that PROMISE scholarship recipients are on average less likely to be employed in West Virginia than graduates that did not receive the scholarship. These results also suggest that the positive influences of the PROMISE scholarship on human capital accumulation stemming from increased in-state college enrolment and graduation are dampened by a reduced conditional probability of working in state.

Merit-aid programmes like the one in West Virginia and Florida show evidence of small positive overall effects on the accumulation of local human capital. These effects seem modest and may not be very cost efficient. Note, however, that the implementation of these policies has yielded improved academic achievement in some states, including West Virginia and Tennessee (Scott-Clayton 2011; Pallais 2009). Thus, the PROMISE scholarship may have improved the overall quality of graduates that did remain in the state to work, which is an important dimension of regional human capital development.

Footnotes
  1. 1

    They are: Alaska, Arkansas, Florida, Georgia, Kentucky, Louisiana, Massachusetts, Michigan, Mississippi, Missouri, Nevada, New Mexico, South Carolina, South Dakota, Tennessee and West Virginia.

  2. 2

    Common objectives included: promote college access; keep the best and brightest students in their home state for college; and reward students who work hard (See also DeFrank-Cole et al. 2007).

  3. 3

    See Groen (2011) for a comprehensive review of previous studies.

  4. 4

    Full and part-time employment measured by US Bureau of Economic Analysis.

  5. 5

    The structure and academic requirements vary from state to state. Georgia's HOPE, for instance, has three different funding alternatives. Most students are only required to meet the high-school GPA requirement (3.0). However, students with high-school GPA of 3.7 and ACT scores of 26 (1,200 SAT) qualify for the Zell Miller scholarship, which is also part of HOPE and provides a higher award. Alternatively, some students are eligible for HOPE grants. Florida's Bright Futures Program provides a similar structure. The highest award goes to those with high-school GPA of 3.5 and ACT score of 28 or higher (1,270 SAT). A less restrictive part of the program requires students to earn a minimum ACT score of 21 or an equivalent 980 SAT, plus 75 hours of community service. These requirements are set to increase in 2014.

  6. 6

    Since the latest graduation year in our data set is 2007–2008, the change in the ACT requirement does not affect our analysis.

  7. 7

    Over 90 per cent of students take the ACT. Those who took the SAT in our sample had their scores converted. See Scott-Clayton (2011) for a complete explanation on the rule of conversion. Also, for a detailed description of the eligibility requirements see www.wvhepcnew.wvnet.edu.

  8. 8

    In 2009, the scholarship was capped at the lesser of tuition and mandatory fees or $4,750 at the West Virginia public or private college they choose to attend.

  9. 9

    Area of concentration and degree are identified by Classification of Instructional Programs (CIP) codes. These are eight-digit codes that identify the degree (Associate's, Bachelor's, etc), area of concentration (business, education, engineering, etc), and major (accounting, kindergarten teaching, chemical engineering, etc) of graduates. An area of concentration is an umbrella classification that includes one or more majors.

  10. 10

    This process means that a match of graduates to employment records for the years before the implementation of the PROMISE scholarship is no longer possible. In turn, this means that we cannot pursue the same sort of cohort analysis found in Scott-Clayton (2011), because we cannot observe the outcome variable (employment in the state) before the implementation of the scholarship. Additionally, the WVHEPC only matched the wage records of college graduates (not all college enrollees), which precludes us from using the same data used by Scott-Clayton (2011).

  11. 11

    Hammond and Leguizamon (2008) estimate that of the 107,455 West Virginia public higher education graduates during the 1997 to 2007 period, 1,996 worked in the federal government sector in the state in 2007 (the only year for which such an estimate exists).

  12. 12

    Note that our data do not allow us to observe whether the individual resides inside or outside the state, only whether he/she works in the state.

  13. 13

    As Scott-Clayton (2011) points out, one could choose the ACT score criterion to determine the eligibility of students. However, the high-school GPA requirement seems not to be a decisive factor for whether or not receiving the scholarship.

  14. 14

    See Scott-Clayton (2011) for details on the calculation of ACTdist.

  15. 15

    The number of graduates above and below the threshold is determined by the bandwidth size. As such, as we increase the bandwidth, the analysis includes more than those who are ‘just’ above and below the margins (one or two points above the cut-off point). Controlling for the possibility of non-linearity becomes more important as we increase the size of the bandwidth.

  16. 16

    Tables 2 and 3 use different bandwidths to show the robustness of the results. There is an ongoing debate in the econometric literature on whether there is an ‘optimal’ bandwidth for this estimator, but no consensus has been reached. For brevity, the results on Tables 4 and 5 use our selection of medium bandwidth, but the results also persist to the choice of bandwidth, and are available upon request.

  17. 17

    The model specification including areas of concentration as controls reveals that, in general, Education and Health majors are more likely to work in the state. Architecture and Biological Science majors are less likely to do so.

  18. 18

    Groen (2004, 2011) discusses the importance of controlling for these locational preferences.

  19. 19

    College graduates in our dataset include those from these three cohorts that enroll in West Virginia colleges and graduated in four to five years after enrollment.

  20. 20

    Scott-Clayton (2011, Figure 6 on page 634).

  21. 21

    Note that the baseline graduation rates and marginal impact of PROMISE are broadly and uniformly applied to all graduates below and above the threshold. As noted before, this does not account for potential increasing or decreasing marginal effects. As such the estimates may overestimate graduation rates for those that are well below the cut-off, and underestimate the graduation rates on the other end of the distribution. This is a simplification that serves for illustration purposes only.

  22. 22

    Similar to the calculation for graduation rates, our estimates are also broadly applied to all graduates below and above the cut-off point.

References

  1. Top of page
  2. AbstractResumen
  3. Introduction
  4. West Virginia and the PROMISE scholarship programme
  5. Background and literature review
  6. Data
  7. Empirical method and results
  8. Policy analysis
  9. Conclusion
  10. References
  • Angrist JD (1993) The effects of veterans benefits on education and earnings. Industrial and Labour Relations Review 46: 637652
  • Bound J, Groen J, Kezdi G, Turner S (2004) Trade in university training: Cross-state variation in the production and stock of college-educated labour. Journal of Econometrics 121: 143173
  • Bound J, Turner S (2002) Going to war and going to college: Did World War II and the GI bill increase education attainment for returning veterans? Journal of Labour Economics 20: 784815
  • Bowen E (2013) From higher education to work in West Virginia 2011. Bureau of Business and Economic Research. Manuscript. West Virginia University
  • Cornwell C, Mustard D, Sridhar D (2006) The enrollment effects of merit-based financial aid: Evidence from Georgia's HOPE scholarship. Journal of Labor Economics 24: 761786
  • DeFrank-Cole L, Noland B, Green A (2007) Merit-based aid in West Virginia: An overview of the PROMISE scholarship Program. Paper presented at the annual forum of the Association for Institutional Research
  • Dynarski S (2000) Hope for whom? Financial aid for the middle class and its impact on college attendance. National Tax Journal 53: 629661
  • Dynarski S (2008) Building the stock of college-educated labour. Journal of Human Resources, 43: 576610
  • Glaeser E, Scheinkman J, Shleifer A (1995) Economic growth in a cross-section of cities. Journal of Monetary Economics 36: 117143
  • Goodman J (2008) Who merits financial aid?: Massachusetts' Adams scholarship. Journal of Public Economics 92: 21212131
  • Groen J (2011) Building knowledge stocks locally: Consequences of geographic mobility for the effectiveness of state higher education policies. Economic Development Quarterly 25: 316329
  • Groen J (2004) The effect of college location on migration of college-educated labour. Journal of Econometrics 121: 125142
  • Hammond GW, Leguizamon JS (2008) From higher education to work in West Virginia 2007. Bureau of Business and Economic Research, Manuscript. West Virginia University
  • Hammond GW, Thompson E (2008) Determinants of income growth in US metropolitan and nonmetropolitan regions. American Journal of Agricultural Economics 90: 783793
    Direct Link:
  • Hammond GW, Tosun M (2011) The impact of local decentralization on economic growth: Evidence from US counties. Journal of Regional Science 51: 4764
  • Heller DE (2002) State merit scholarship programmes: An introduction. In: Heller DE, Marin P (Eds) Who should we help? The negative social consequences of merit scholarships. The Civil Rights Project at Harvard University, Cambridge, MA
  • Hickman DC (2009) The effects of higher education policy on the location decision of individuals: Evidence from Florida's bright futures scholarship program. Regional Science and Urban Economics 39: 553 Cambridge, MA: 562
  • Higgins MJ, Levy D, Young AT (2006) Growth and convergence across the United States: Evidence from county-level data. Review of Economics and Statistics 88: 671 Cambridge, MA: 681
  • Imbens GW, Lemieux T (2008) Regression discontinuity designs: A guide to practice. Journal of Econometrics 142: 615635
  • Kane TJ (2003) A quasi-experimental estimate of the impact of financial aid on college-going. NBER Working Paper No. 9703
  • Lee DS, Card D (2008) Regression discontinuity inference with specification error. Journal of Econometrics 142: 655674
  • Lee DS, Lemieux T (2010) Regression discontinuity designs in economics. Journal of Economic Literature 48: 281355
  • Moretti E (2004) Estimating the social return to higher education: Evidence from longitudinal and repeated cross-section data. Journal of Econometrics 121: 175212
  • Pallais A (2009) Taking a chance on college: Is the Tennessee education lottery scholarship program a winner? Journal of Human Resources 44: 199222
  • Rouse CE (1998) Do two-year colleges increase overall educational attainment? Evidence from the States. Journal of Policy Analysis and Management 17: 595620
  • Scott-Clayton J (2011) On money and motivation: A quasi-experimental analysis of financial incentives for college achievement. Journal of Human Resources 46: 614646
  • Shapiro JM (2006) Smart cities: Quality of life, productivity, and the growth effects of human capital. Review of Economics and Statistics 88: 324355