Working during school and academic performance

Todd R. Stinebrickner and Ralph Stinebrickner1

please direct correspondence to Todd R. Stinebrickner Dept. of Economics The Social Science Centre The University of Western Ontario London Ontario Canada n6a 5c2 trstineb@julian.uwo.ca phone 519 679-2111 ext. 5293 fax 519 661-3666

Unique new data from a college with a mandatory work-study program are used to examine the relationship between working during school and academic performance. Particular attention is paid to the importance of biases that are potentially present because the number of hours that are worked is endogenously chosen by the individual. A “naive” OLS regression, which indicates that a positive and statistically significant relationship exists between hours-worked and grade performance, highlights the potential importance of endogeneity bias in this context. Although a fixed effects estimator suggests that working an additional hour has an effect on grades which is quantitatively very close to zero, we suggest that there are likely to exist causes of endogeneity which are not addressed by the fixed effects estimator. Indeed, an instrumental variables approach, which takes advantage of unique institutional details of the work-study program at this school, indicates that working an additional hour has a negative and quantitatively large effect on grade performance at this school. The results suggest that, even if results appear “reasonable,” a researcher should be cautious when drawing policy conclusions about the relationship between hours-worked and a particular outcome of interest unless he/she is confident that potential problems associated with the endogeneity of hours have been adequately addressed.

The University of Western Ontario and Berea College. We would like to thank John Bound and Dan Black for useful comments.

1

I. Introduction Important policy decisions have been based on beliefs about the relationship between working during school and a student’s current and future academic performance. However, although some previous research examines these matters, currently no consensus exists on the effect that youth employment has on these outcomes. If working during high school has a harmful effect on academic performance, it might be reasonable to strengthen laws that regulate the number of hours that youth can work. Similarly, if working during college is detrimental, individuals who need to work during college in order to pay tuition costs may be at a disadvantage when compared to students from wealthier backgrounds, and work-study based financial aid programs may have certain undesirable side effects. Difficulty in determining the true impact that work has on academic performance arises largely because the number of hours that an individual works is endogenously chosen. For example, it is

sometimes posited that individuals who fare well academically in school tend to be blessed with high levels of “motivation” that may also make them more likely than other students to become involved with nonacademic activities such as work.2 In empirical work, if “motivation” is not fully observed, some of the variation in academic performance that should be attributed to differences in motivation may mistakenly be attributed to differences in work status. Thus, in this scenario, simple econometric models may understate the negative impact that working has on school performance. Indeed, as discussed in the next section, previous studies have sometimes found that academic performance is highest among individuals who are working a moderate number of hours. Credibly dealing with the endogeneity of work hours in empirical work is typically very difficult.

This may be a reasonable scenario for students working low to moderate number of hours. However, students who work large numbers of hours may be individuals who are not particularly interested in academics.

2

1

It is true that certain individual characteristics may be able to capture some portion of factors such as motivation. However, given the large amount of variation in academic performance that typically remains unexplained after controlling for available individual characteristics, it is likely that including individual characteristics provides only a relatively low level of protection against this type of endogeneity. Instrumental variable (IV) estimators represent a theoretically appealing way to deal with the endogeneity problem. However, in practice it is often difficult to find instruments that both explain a reasonable portion of the variation in work hours and are unrelated to an individual’s academic performance except through their effect on hours worked.3 Unfortunately, when methods for controlling for endogeneity are not entirely satisfactory, the researcher can be left without any manner in which to determine the extent to which bias may be present in his/her estimates. Even when estimates seem “reasonable,” a large amount of bias may be present. In this paper, we utilize unique new data in an attempt to examine the extent to which the endogeneity of hours may bias estimates of the effect of employment on academic performance. The data are obtained directly from the administrative records of Berea College. Located in central Kentucky, this liberal arts institution operates under a mission of providing an education to those who “have great promise, but limited economic resources.” As part of this mission, all students who attend Berea receive full tuition scholarships. Part of the cost of schooling is defrayed through a mandatory work-study program. Although all students must work at least a minimum of ten hours a week, variation in hours worked arises because students can often choose to earn extra income by working additional hours. We wish to note in advance that, given the unique nature of Berea College, it is our belief that our results

See Bound et al. (1995) for a discussion of the potential problems that can arise in instrumental variables estimation when the correlation between instruments and the endogenous explanatory variable is weak.

3

2

should be viewed cautiously. Nonetheless, as will be described throughout the paper, the institutional details of the Berea College labor program and the detailed nature of our administrative data provide us with a unique opportunity to obtain some information about the potential importance of the endogeneity problem described above. In particular, we are able to compare estimates of the effect of working on academic performance obtained using an Ordinary Least Squares (OLS) estimator, a fixed effects (FE) estimator, and an instrumental variables (IV) estimator. The FE estimator takes advantage of semester-by-semester

information on hours worked that is available in our administrative data. The IV approach takes advantage of two features of the labor program at Berea that imply that the job to which a person is assigned in the first semester can serve as an instrument for hours worked. First, as will be discussed in detail, it is reasonable to believe that the first-year job assignment process (which takes place before students arrive at Berea) does not create a situation in which the “motivation” level of students is higher in some jobs than in others. Second, the amount of hours that are “available” for a person to work depends on the job to which he/she is assigned. There are intuitive reasons to believe that both the OLS and FE estimators may tend to understate the negative effect that working has on academic performance. Our results suggest that the amount of bias may be substantial. Whereas the OLS and FE estimators both suggests a positive relationship between working and academic performance, the IV estimator indicates that working additional hours has a harmful effect on academic performance. The remainder of the paper proceeds as follows. In Section II, we briefly examine past research that considers the relationship between employment while in school and academic or labor market outcomes. In Section III, we describe the data from Berea College and we highlight some of the advantages of these data In Section IV, we provide the details of the results summarized above. In Section V we conclude.

3

II. Previous Work Ruhm (1997) provides a very thorough overview of the research from several academic disciplines which examines the effect of working during school. He concludes that for employment in high school “there is currently no consensus whether student employment improves or worsens school performance.” His summary suggests that previous studies have varied substantially in terms of the samples and the methodologies that have been employed. In many cases, the endogeneity of the work decision and the hours decision was not addressed in these studies. Further, the very few cases where attempts were made to deal with the endogeneity problem met with limited success because of difficulty in finding credible instrumental variables.4 For example, OLS estimates in Turner (1994) indicate that, for students in the High School and Beyond survey, working a moderate number of hours is positively related to high school performance. The effect of hours is found to be statistically insignificant when fixed effects and IV estimators are employed. However, because some of his instruments are potentially related to academic performance as well as the decision to work in high school, it becomes difficult to tell whether working during high school is beneficial, harmful, or unimportant from the standpoint of academic performance.5 Ruhm reports that previous studies of the effect of work in high school on future wage outcomes are more conclusive and typically conclude that work in high school is associated with increased future earnings. Although most previous studies had concentrated on the period immediately following high school completion, Ruhm reaches similar conclusions when examining labor market outcomes six to nine years after the scheduled date of high school graduation. However, these studies are subject to the same
Eckstein and Wolpin (1997) take a very different approach to dealing with the endogeneity problem by formulating and estimating an explicit sequential decision model of high school attendance and work. They find that working while attending high school does reduce academic performance, but the quantitative effects are small. For example, a student’s “access to money from parents” would seem to be correlated with family background/income which has a well-known effect on academic achievement.
5 4

4

difficulties in adequately controlling for the potential selectivity in hours worked. This problem is recognized by Ruhm who uses geographic characteristics to identify a model which controls for the endogeneity of work. Although his instruments pass the Newey (1985) test for exogeneity based on overidentification restrictions, they provide only weak predictive power for hours. As a result, the effect of hours is measured very imprecisely and the correction for selection bias meets with “limited success.”6 Less research examines the effect of working during college on academic performance. However, a lack of consensus also appears in this context. For example, Paul (1982) finds that working is detrimental to academic performance in college, Hood et al. (1992) find that grade point averages are highest among students with moderate amounts of work, and Ehrenberg and Smith (1987) find positive effects of working in on-campus jobs but negative effects of working in off-campus jobs. III. Berea College and the Berea College Labor Program III.1 Overview of Financial Aid and the Labor Program Berea College is located in Berea, Kentucky where the “Bluegrass meets the foothills of the Appalachian Mountains.” Consistent with the mission of the school discussed in the introduction, students who are admitted to Berea are typically quite poor; Stinebrickner and Stinebrickner (1998) find that the mean and standard deviation of family income for dependent students who entered Berea between 1989 and 1997 is 22,450 and 13,590. However, each student at Berea receives a full tuition subsidy, and the average student receives more than $3,000 of additional financial aid from private, state, and federal

For other recent work which examines the relationship between employment during school and earnings see Light (forthcoming) who uses a sample of male, terminal high school graduates from the NLSY and finds that high school employment has a “small and relatively short-lived effect on post-school wages” and Hotz et al. (1998) who estimate a dynamic discrete choice model using the NLSY. Light (1998a, 1998b) examines the relationship between in-school work experience and returns to schooling.

6

5

sources.7 As part of this financial aid package, each student is required to work a minimum of ten hours a week in the school’s labor program. In addition to receiving a $2800 labor grant which is part of the financial aid package described above, the student/worker receives a small hourly wage. Before their arrival to Berea, students are assigned to specific “service” type jobs for their first year which pay approximately two dollars an hour.8 Hourly wages in a particular job increase after the first year with “length of service.” In addition, after the first year, students may choose to compete for higher-paying jobs that they find desirable. Students at the upper-end of the pay scale receive approximately four dollars an hour.9 III.2 Measurement Benefits of Our Administrative Data From their administrative records, Berea College made available records for all of the full!time students that matriculated between the fall semester of 1989 and the fall semester of 1997. There are several reasons that these administrative data are useful from the standpoint of measuring the variables of interest for this study. First, because hours will not tend to be constant across the entire school year, it is desirable to link academic performance and hours within time units that are as short as possible. Our data allow us to measure hours and our measure of academic performance, grades, at a semester level rather than at the yearly level which is typically the smallest time unit possible in other studies.10 Second, the use of administrative data ensures that our semester hours measure is not subject to the types of measurement
7

This leaves the entering student with an average total room, board, and college fee bill of only approximately $1000. These numbers were taken from the Berea College admissions brochure, 1998. Students graduate from Berea with an average of approximately $1,000 in student loans.
8

In 1997, freshmen received $1.95 per hour (in 1997 dollars). The highest paying job paid $3.95 in 1997.

9

Although it cannot be examined directly using our data, students at Berea tend to work roughly the same amount each week during a particular semester. Thus, it also seems likely that less inter-week variation in hours will be present for a person at Berea than a person in other data.

10

6

error that are potentially problematic in longitudinal data which rely on self-reports of hours worked.11 In other studies, where the outcomes of interest (e.g., grades or test score performance) are measured at yearly intervals, the appropriate hours measure should reflect the total number of hours worked over the entire academic year. However, past evidence that employment hours tend to be misreported in retrospective data suggests that precisely measuring total yearly hours may be difficult. For example, in a validation study of the Panel Study of Income Dynamics, Duncan and Hill (1985) find that respondents overstate the number of hours that they work by between 10% and 12%, and evidence of this problem is also found for the youth in the NLSY by Ruhm (1997).12 In part because of this problem and in part because some surveys do not contain measures of yearly hours, researchers have often sensibly chosen to characterize hours using responses to questions about the number of hours which were worked by the respondent in a particular reference week. Unfortunately, even if respondents answer this question accurately, using reference week hours to construct yearly work hours will lead to measurement error in the latter if, as indicated by Ruhm (1997), variation exists across weeks in the number of hours that are worked.13 Finally, because we study a single school, our outcome measure is directly comparable across the different individuals in the sample.14 Of course, with respect to the latter, it is important that our

11

Students are not permitted to hold off-campus jobs.

He finds irreconcilable differences between measures of average weekly work hours for the sample based on retrospective questions about yearly work histories and measures of average weekly work hours for the sample based on a survey question about hours worked in the week prior to the interview. In the NLSY, it is reasonable to believe that these averages should be reasonably similar because students are interviewed at different times in the year.

12

Ruhm (1997) finds evidence that work hours increase as the academic year progresses. This implies that the measure of a particular person’s hours worked in the NLSY will be determined in part by the respondent’s interview date.
This is particularly true in the early semesters when individuals are satisfying general, liberal arts requirements. As mentioned earlier, one also might worry when using data on students from different schools that
14

13

7

conclusions take into account the unique nature of our sample. III.3 A Description of Employment Hours During the first semester, individuals are not officially “encouraged” to work more than the ten hour a week minimum. In practice, this implies that students are not allowed to add “secondary” jobs to the “primary” jobs that they are assigned for their freshman year. However, variation does appear in work hours during the first semester because some degree of hours flexibility exists within a person’s primary job; for our sample, the mean and standard deviation of hours in the first semester is 10.97 and 1.40 respectively. As we will discuss in more detail, some of the variation in first semester hours is explained by the fact that individuals are assigned to different labor jobs. What is not entirely clear is why students often work more than the minimum number of hours given the low marginal wage rate. One possibility is that, even with full tuition subsidies, the marginal utility of income may be very high for students at Berea who are typically from low income families.15 A second possibility is that, when particular labor departments face worker “shortages” (in the sense that all necessary hours would not be covered if all assigned workers worked the ten hour minimum), students offer to work additional hours or decide to satisfy requests to work additional hours in order to remain in

the number of hours worked is correlated with school quality, which has a direct influence on academic outputs. In this case, the effect of hours on performance would tend to be biased if school quality is not fully observed. In results not shown here, we did not find evidence that lower income individuals within our sample choose to work more hours than higher income individuals. However, even the highest family income earned at Berea is not particularly high. The fact that many students choose to add secondary jobs of a service nature in the second semester may represent some evidence that this explanation is relevant. Average hours are 11.53, 12.34, 12.76, 13.22, 13.55, 14.04, and 14.02 in the second through eighth semesters respectively. It is likely that this increase is due to a combination of 1) an increase in the wage that is paid as individuals increase their job rank and experience standing, 2) an increase in the benefits of human capital accumulation as students are able to switch from service-type jobs to jobs which are more relevant to their future careers, 3) an increase in the non-pecuniary benefits of working as students become able to choose jobs which are more interesting to them, and 4) an increase in the ease of adding hours as students become familiar with the labor system and learn about potential job opportunities.
15

8

good standing with their supervisors. Remaining in good standing with a supervisor may make working generally more enjoyable or may lead to positive recommendations from the supervisor in the future. A final possibility is that students find that time spent working is somewhat substitutable with other available leisure activities. IV. Estimation and Results In order to estimate the relationship between academic performance as measured by semester grade point averages and semester hours worked, we examine the equation (1) Git ="Xi + $Hit +,it

where Git is the person’s semester grade point average at time t, Hit is the number of hours worked by person i at time t, and Xi is the vector of person i’s individual characteristics described in the first column of Table 1.16 Our primary interest is in $, the effect of an additional hour of work on the person’s semester grade point average at time t. If Xi and Hit are uncorrelated with ,it, a consistent estimate of $ can be obtained using the OLS estimator. The second column of Table 1 shows the results when equation (1) is estimated using data from the first semester in college (t=1). The third column of Table 1 shows the results when equation (1) is estimated using data from the second semester in college (t=2). The two columns indicate that, on average, individuals who work more hours receive higher grades, with an additional hour of work being associated with increased grade point averages of .043 in the first semester and .021 in the second semester. The estimates are statistically significant with t statistics of 3.929 and 3.138 respectively.

In order to keep the sample as large as possible, family income, which is missing for a subset of our students (e.g., those who report being independent), is not included in X. However, the results reported in the remainder of the paper were found to be very similar when family income was included (and the models were estimated with the subset of individuals with non-missing family income information). For example, as noted in footnote 15, family income does not explain the variation in hours worked.

16

9

Given the types of service jobs in which individuals work in the first year, it seems unlikely that a given person’s academic performance would improve if he decided to add hours above the weekly minimum of ten hours. Instead, it seems likely that the OLS estimator of $ is biased by the fact that the number of hours is endogenously chosen. The benefit of our data is that they allow us to take further steps to analyze the extent to which bias might be present. Endogeneity bias will be present to some extent if there exist unobserved determinants of grades (i.e., elements of ,it) which also influence the number of hours that a person works. Perhaps the most commonly discussed scenario is one in which the “problematic” unobserved determinants are personspecific, permanent, individual attributes, 0i , such as “motivation” or “energy” which might drive certain people to both work more and to study more. In this scenario, equation (1) can be rewritten as: (2) Git ="Xi + $Hit + 0i+