- Email: [email protected]
Private transfers and college students’ decision to work
Economics of Education Review 42 (2014) 34–42
Contents lists available at ScienceDirect
Economics of Education Review journal homepage: www.elsevier.com/locate/econedurev
Private transfers and college students’ decision to work Andreas Bachmann a,1, Stefan Boes b,* a b
University of Bern, Department of Economics, Schanzeneckstrasse 1, CH-3001 Bern, Switzerland University of Lucerne, Department of Health Sciences and Health Policy, Frohburgstrasse 3, PO Box 4466, CH-6002 Lucerne, Switzerland
A R T I C L E I N F O
A B S T R A C T
Article history: Received 14 June 2012 Received in revised form 6 May 2014 Accepted 25 May 2014 Available online 8 July 2014
We estimate the impact of external financial support on the labor supply of students during their tertiary education. Using a dynamic labor supply model and accounting for the endogeneity of income from private transfers, we find a significantly lower likelihood of being employed for transfer recipients. Our results suggest that private transfers lead to a shift in students’ time allocation, lowering their hours devoted to working and increasing their time devoted to studying. We find evidence for a psychological component of receiving transfers through an increase in the perceived risk of failure in academic studies. ß 2014 Elsevier Ltd. All rights reserved.
JEL classification: I21 I23 J22 Keywords: Financial support Student employment Time trade-off Simultaneity bias
1. Introduction When young adults enter tertiary education, they face a number of expenses, including tuition fees and costs for housing, transportation, and living in general.2 Students therefore often rely on the continuing support of their parents. Alternatively, they may decide to work during their academic studies. These two sources of financial support are not mutually exclusive and, in fact, are highly related, as
* Corresponding author. Tel.: +41 41 229 5949; fax: +41 41 229 5635. E-mail addresses: [email protected] (A. Bachmann), [email protected] (S. Boes). 1 Tel.: +41 31 631 4777; fax: +41 31 631 3783. 2 Actual expenses vary by location and life-style but can be as high as 5000 US dollars per month at top US universities (see, for example, http:// www.topuniversities.com/student-info/student-finance/how-muchdoes-it-cost-study-us [accessed 08.08.13]). In Switzerland, tuition fees are relatively low compared to those in the US, but living expenses can easily add up to 2500 US dollars per month, or more (see http:// www.crus.ch/information-programme/study-in-switzerland.html?L=2#8_Costs [accessed 08.08.13]). http://dx.doi.org/10.1016/j.econedurev.2014.05.005 0272-7757/ß 2014 Elsevier Ltd. All rights reserved.
standard labor supply models predict that a higher nonworking income is associated with a lower propensity to work (e.g., Blundell, MaCurdy, & Meghir, 2007). How does income from private transfers affect the student labor supply? Building on Becker’s work on intrafamily transfers (e.g., Becker, 1993), theoretical models with an endogenous determination of income through part-time employment predict an inverse relationship between transfers and a child’s labor supply (e.g., Juerges, 2000). The underlying mechanisms are highly interdependent, however, because non-working time for students may be not only devoted to leisure but also invested in human capital (through additional time devoted to studying). Parents can influence the time allocation of their children by providing transfers. For example, parents may encourage their child to cut back on work and study more if the child’s academic performance diminishes (Kalenkoski & Pabilonia, 2010). Whether the student works more in some periods and receives more financial support in others is likely a simultaneous decision (e.g., Dustmann, Micklewright, & van Soest, 2009; Lee & Orazem, 2010).
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
This simultaneity poses a severe problem for the estimation of a causal effect of private transfers on the student labor supply. In an early descriptive study, Pabilonia (2001) examines the employment behavior of American youths under the age of 16 years and finds a significantly negative association between labor supply and parental allowances. In more recent studies, Wolff (2006), Dustmann et al. (2009), Gong (2009), and Kalenkoski and Pabilonia (2010) use parental income as an instrument for parental transfers, arguing theoretically that the effect of parental income on a child’s labor supply can only go through a transfer. While the last three studies find a negative impact of transfers on youth employment, Wolff (2006) does not find a significant effect. In all these studies, the endogeneity of transfers plays a major role in the empirical argument, and failing to account for the endogeneity is shown to produce biased estimates. We add to the literature by (i) analyzing a large panel of college students in Switzerland, (ii) employing a different identification strategy from previous studies, and (iii) explicitly addressing students’ allocation of time. The use of longitudinal data serves two purposes. First, we can control for time-constant unobserved confounders. Gong (2009) argues that such confounders are particularly important in this context because they take into account idiosyncratic preferences for working and studying and personality traits that jointly determine the amount of transfers and the student’s labor supply. Second, we can use dynamic labor supply models, which crucially distinguishes our work from Gong (2009).3 As we condition on past employment and individual-specific effects, we can use lagged transfers as instruments for current transfers to identify a causal effect. Using over-identification and model specification tests, we find plausible evidence to support our instruments. This strategy also allows us to test the previously used exclusion restriction on parental income, which we confirm in our data. Our results provide new evidence on the trade-off in students’ allocation of time. Using the dynamic labor supply framework, we find that private transfers significantly reduce work hours and increase study hours. We conclude that transfers provide an incentive for students to shift work time toward study time. On the downside, we do not find an improvement in academic performance: the perceived risk of failure in academic studies is substantially higher for transfer recipients than for non-recipients, which might indicate an implicit increase in study-related stress levels and pressure to perform well associated with transfers. The remainder of the paper is organized as follows. Section 2 describes the data and the variables that are used in the analysis. Section 3 outlines the model and the estimation methods. Section 4 presents the results. We first analyze the labor supply decision, then look at the time trade-off between working and studying, and finally
3 Gong (2009) does not estimate dynamic labor supply models but rather addresses the simultaneity problem by using a fixed-effects twostage least squares procedure with variation in parental income as the instrument.
35
examine the risk of failure in academic studies. Section 5 discusses the limitations of our study and concludes the paper.
2. Data Our analysis of the impact of private transfers on the student labor supply is based on data from the Swiss Household Panel (SHP), a representative annual panel survey of the Swiss residential population; see Voorpostel et al. (2012).4 The SHP started in 1999 with a total of about 5100 interviewed households. It is a comprehensive survey that covers a wide range of topics, including household and family background, education, work, income, health, and socio-psychological information. In 2004, a refreshment sample of about 2500 households was added to overcome the initially high panel attrition. We employ all waves of the SHP until the most recent, W13 (year 2011). We confine our analysis to individuals enrolled at a higher education institution.5 The students can be fulltime or part-time students. The maximum sample consists of 857 students and 4419 person-year observations. The response rate among students per year is about 50%, which is slightly lower than the response rate for the total SHP (about 60%). We checked whether panel attrition could be a problem in the data we use, but we found little supporting evidence. First, attrition in the total SHP only affects a few variables, mainly political and leisure variables that are unrelated to our main variables (Voorpostel et al., 2012). Second, we estimated a simple logit model for the probability of missing information as a function of gender, age, and location of residence. None of these variables is statistically significant. Even if we include the mean transfers over the non-missing years as a predictor, this variable has no explanatory power (the detailed results are available upon request). We use information on labor supply for our main outcome employment: a dummy equal to one if the student is actively occupied and zero otherwise. Actively occupied individuals in the SHP comprise all respondents who are employed or self-employed (even if the amount of work is only one hour per week). In addition, we use work hours and study hours per week as outcome variables to investigate students’ allocation of time and the impact of private transfers on this allocation. As a final outcome, we take the perceived risk of failure in academic studies in the next 12 months, which also serves as an endogenous right-hand side variable in the labor supply models. This variable is recorded on an 11-point scale, where 0 means ‘‘no risk at all’’ and 10 means ‘‘sure a risk’’. Our key explanatory variable is yearly social informal transfers (in Swiss Francs, CHF), consisting of all payments received from people living in the same household and outside the household, henceforth referred to as amount of private transfers. While these transfers are directly paid to
4
Free access to the data can be acquired via http://www.swisspanel.ch. Most students study at an academic university (about 84 percent), but the sample also includes students of universities of teacher education (about 4 percent) and universities of applied sciences (about 12 percent). 5
36
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
the student, there may also be indirect transfers from other household members, for example, if they pay for accommodation or utilities. The SHP does not record such transfers, but the amount of such transfers is likely correlated with the income of other household members, which we use as a proxy control. For each individual i in year t, the income of all other household members is constructed as the sum of total net personal income in t over all members except i. Both monetary variables are inflation adjusted with 2011 as the base year. Information on the income of other household members is closely related to the exclusion restriction imposed in previous studies (Dustmann et al., 2009; Gong, 2009; Kalenkoski & Pabilonia, 2010). If the student still lives in the parents’ house (which is the case for more than 80% of the respondents in our sample), this information mainly captures parental income. It also contains the income of partners. Nevertheless, it may not be a priori clear whether such information can be excluded from the labor supply decision, as it may be associated with non-monetary benefits for the student (e.g., through contacts to possible employers). Hence, we do not impose this restriction but instead include the income of other household members as a separate control. We instrument direct transfers using panel internal information (lagged transfers, to be specific). In doing so, we can test the previously imposed exclusion restriction. Table 1 shows basic descriptive statistics for our sample. After deleting observations with missing information, we obtain a final (estimation) sample of 672 students (1443 student-year observations). Of these, 439 have been employed in at least one wave, and 554 received transfers at least once. The average employment rate is about 52 percent. If we compare students who received transfers with those who did not, the former have an employment rate that is about 13 percentage points lower (48% compared to 61%, respectively). The average transfer paid is about CHF 6300 per year. Apart from employment, there are substantial differences between transfer recipients and
non-recipients: on average, non-recipients work more and spend less time studying, and the amount of income of the other members in their households is lower. 3. Econometric model and estimation strategy While Table 1 suggests that a negative association exists between transfers and student employment, the raw difference is likely biased with respect to the causal effect. First, the raw difference does not adjust for confounding influences, such as gender, age, and individual preferences. Second, transfers themselves depend on the labor supply of students. Regarding the student labor supply, it can be conjectured that the mean comparison is downward biased, as working more reduces the amount of transfers (Dustmann et al., 2009). For the confounders, the bias is ambiguous. More motivated students might receive more transfers, but they might also be more likely to engage in extracurricular activities (such as part-time employment). Preferences might work in the opposite direction if students enjoy their independence and try to avoid receiving financial support from their parents. Following Eckstein and Wolpin (1990), Hyslop (1999), and Blundell et al. (2007), we specify a linear dynamic panel data model for the labor force participation of student i in time period t with individual-specific effects as yit ¼ ayi;t1 þ xit 0 b þ ci þ uit
(1)
where yit is the employment status in period t, yi;t1 is lagged employment, xit is the vector of time-varying regressors (including private transfers, the risk of failure in academic studies, and the log of the income of other household members), ci is an individual-specific unobserved effect that may be correlated with xit , and uit is a time-varying heteroscedastic error. Several aspects of the model should be highlighted. First, estimation by ordinary least squares will unlikely produce consistent results, as the lagged dependent
Table 1 Descriptive statistics by financial transfer status. No transfers Mean Employed (yes/no) Hours of work Hours spent studying Risk of failure in academic studies (0–10 scale) Amount of transfers (in CHF 1000) Income of others in household (in CHF 1000) Living with parents (yes/no) Female Age Married (yes/no)
61.5% 11.48 34.01 2.43 – 89.24 86.0% 59.3% 21.89 2.2%
Reference: no partner Partner, living together Partner, not living together
9.0% 45.2%
Kids (yes/no) Number of kids
27.8% 0.388
Number of observations Source: Swiss Household Panel 1999–2011, own calculations.
356
Transfers SD 14.20 13.86 2.03 82.29
1.65
Mean 48.4% 7.25 38.89 2.73 6.29 110.69 84.1% 52.8% 21.74 1.1%
SD 12.13 13.01 2.18 5.37 89.19
1.61
7.7% 44.9%
0.763
22.4% 0.281 1087
0.580
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
37
Table 2 Regression results for student employment. Random effects (1) Private transfers
(3)
Dynamic/IV (4)
0.115*** (0.044)
0.013** (0.006) 0.003 (0.009)
0.011 (0.008) 0.010 (0.014)
0.018** (0.008) 0.006** (0.003) 0.013 (0.008) 0.012 (0.014)
1443
1443
1443
1443
Lagged employment
Number of observations ar2p ar3p hansenp
(5)
(6)
0.192** (0.077)
0.017*** (0.006) 0.004 (0.003) 0.015** (0.006) 0.004 (0.009)
Amount of transfers2/10
Log (income others in hh)
(2)
0.105*** (0.031)
Amount of transfers
Risk of failure
Fixed effects
0.001 (0.010) 0.015 (0.012) 0.445*** (0.160)
0.032** (0.013) 0.010** (0.005) 0.003 (0.010) 0.015 (0.012) 0.452** (0.165)
760 0.418 0.325 0.710
760 0.424 0.328 0.646
Source: Swiss Household Panel 1999–2011, own calculations. Note. All models control for civil status, living in parents’ household, number of kids and year dummies. (1) and (2) add gender, a second order polynomial in age, Swiss nationality and prior education completed. (5) and (6) are estimated by the Blundell and Bond (1998) system GMM procedure with lagged dependent variable, transfers and risk of failure instrumented. ar2p, ar3p indicate p-values for the Arellano and Bond (1991) autocorrelation tests of order 2 and 3. hansenp indicates the p-value for the over-identification test as proposed by Hansen (1982). Cluster-adjusted standard errors in parentheses. * p < 0.1. ** p < 0.05. *** p < 0.01.
variable, by model assumption, is correlated with the individual effect. As outlined in Arellano and Bond (1991), estimation of (1) may proceed by first differencing the level equation to eliminate ci and to use yi;t2 (and earlier) to instrument the lagged difference on the right-hand side. Instruments are needed because the difference transformation generates a first-order serial correlation in the errors such that the lagged difference in employment status is an endogenous regressor. Second, Arellano and Bover (1995) raise the issue that lagged levels can be weak instruments, particularly with persistent data. For all our dependent variables (employment status, as well as work hours, study hours, and risk of failure in academic studies), we cannot rule out a priori the possibility that they are persistent variables. We therefore explore lagged differences in the dependent variable as additional instruments for the level equation, as in Blundell and Bond (1998). This procedure also improves the precision of the estimator. Third, estimation is performed by using the Stata command xtabond2 of Roodman (2009). The dynamic instrumental variable (IV) strategy requires at least three consecutive waves of data for an individual to be included in the estimation, which we have for 228 students (760 person-year observations). Fourth, the instruments are only valid if the level errors do not exhibit serial correlation, an assumption that can be tested and that we cannot reject in any of our models. Fifth, the simultaneity of transfers and employment yields an additional correlation between uit and current transfers that would bias the results if it is not accounted for. Under the assumption of serially uncorrelated errors, we can instrument current transfers (or the differences thereof due to the elimination of ci ) with two- and three-period lags of the transfer variable. These lags convincingly pass
the relevance condition (F-statistics of joint significance larger than 20), and we use over-identification tests to check the plausibility of the exogeneity condition. Again, for all our models, we find no indication of a violation of these tests. Sixth, as discussed in Kalenkoski and Pabilonia (2010), there might be additional simultaneity with study performance. We include the perceived risk of failure in academic studies as a control variable and instrument it analogously to transfers. We discuss the quality of this variable as a proxy for academic performance below. Seventh, with employment status as the dependent variable, we obtain a linear probability model. On the one hand, this type of model implies heteroscedastic errors, which we address by using heteroscedasticityrobust standard errors. On the other hand, estimated probabilities from the model are not restricted to the unit interval. Since we are mainly interested in estimating mean effects and since model predictions outside the unit interval occur for less than one percent of the observations in our sample, this issue should not substantially affect our results (see also Horrace & Oaxaca, 2006). Eighth, we use the same type of model to estimate the impact of private transfers on both work hours and study hours, again instrumenting the lagged dependent variable, transfers, and the risk of failure in academic studies. Finally, we add control variables for gender, age, nationality, highest level of education achieved before current studies,6 and family situation (living with parents,
6 This variable includes indicators for compulsory education, apprenticeship, (higher) vocational education, maturity, and a previous tertiary degree (the base category is no compulsory education).
38
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
living with a partner, marital status, and number of children in the household), and we allow for regional and time dummies to capture general labor market conditions. It should be noted that there are no restrictions on these control variables in a random effects model without dynamics (except for the usual exclusion of the reference group for categorical variables). In the dynamic models, the first differences eliminate not only the unobserved effect ci but also all time-constant controls. 4. Private transfers and student labor supply 4.1. The employment decision Our analysis starts with the effect of private transfers on the likelihood of being employed. We first ignore dynamics and simultaneity in the decision-making process by estimating a random effects (RE) model for student employment. Columns 1 and 2 in Table 2 present the estimated coefficients for a model with an indicator for transfers and a more flexible model with a quadratic form of the amount of transfers, respectively. Compared to the raw difference in Table 1, we regression-adjust for the controls listed above. The results suggest that the probability of being employed is about 10.5 percentage points lower for transfer recipients than for non-recipients. Increasing the amount of transfers by CHF 1000 is predicted to reduce the probability of working by about 0.9 to 1.7 percentage points.7 The effect of the income of the other household members is negligible8, and students who are at high risk of failure in academic studies are less likely to be employed (an increase in perceived risk by one point on the scale is predicted to reduce the likelihood of being employed by about 1.3 percentage points). In columns 3 and 4, we relax the zero-correlation assumption between the individual-specific effect and the right-hand side variables. In these fixed effects (FE) models, the effect of private transfers on employment is somewhat larger, but still in the range of the RE model. The difference in the likelihood of working between transfer recipients and non-recipients is now predicted to be 11.5 percentage points, and the continuous effect is about 0.6 to 1.8 percentage points per additional CHF 1000 of income from transfers. The latter reduction is largest (in absolute terms) for the initial transfer. The diminishing effect may be explained by a decreasing marginal utility of consumption and/or adaptation effects, i.e., students may become accustomed to the transfers that they receive. We reject the RE models against the FE models (Hausman test p-values less than 105), highlighting the importance of including individual-specific unobserved heterogeneity in the labor force participation equation.
7 Here and in the following analysis, we consider a range of CHF 0 to 10,000 to interpret an increase of CHF 1000, which covers about 80 percent of the transfer distribution. 8 Measurement error in the income variable, in particular as a proxy for indirect transfers, could be a possible explanation for the insignificant results. Under classical measurement error theory (e.g., Cameron & Trivedi, 2005, chap. 26), this would bias the results toward zero. We cannot rule out the possibility of such error in our analysis.
Private transfers might still be endogenous in the FE model if, for example, a student works more in one period and if parents, in anticipation of their child’s labor supply, decide to provide less financial support in that period (we discuss evidence in favor of such an argument below). Exploring the panel structure further, we can instrument private transfers (after first differencing) with the two- and three-period lagged levels of private transfers. We consider such an IV strategy to be credible because we rule out the possibility of an inter-temporal relationship by conditioning on past employment. The results of the dynamic IV model are presented in columns 5 and 6. Since the model is over-identified, we can use a Hansen test to evaluate the validity of the implied moment conditions, which is not rejected in any of the models. The estimated difference in the propensity to work between transfer recipients and non-recipients is predicted to be 19.2 percentage points, and the continuous effect is about 1.2 to 3.2 percentage points. Regarding the dynamics, we find positive persistence in the data; that is, students who decided to work in the last period are about 45 percentage points more likely to work in the current period. Thus, not all students who work in one year continue to work in the next year. As we condition on year fixed effects and thus control for the general business cycle, one possible explanation for this result may be the work-study cycle of students in Switzerland, who work in some years but not in others. The latter is reflected in the employment transition matrix, with almost 25 percent of the students switching their status from employed to not employed and about 40 percent switching their status in the other direction from one year to another. It should be noted that the estimates are robust to further lags in employment and any lags of private transfers, which are all insignificant (adjusting the instrument matrix accordingly). Our results are consistent with those in Dustmann et al. (2009), Gong (2009), and Kalenkoski and Pabilonia (2010), who allow for mutual dependence in the work-transfer relationship but use a different identification strategy. To the extent that income of the other household members resembles their parental income instrument, our small and insignificant estimates highlight the validity of the exclusion restriction imposed in these studies, at least in our data.9 4.2. Allocation of time In a next step, we examine the trade-off in the allocation of time between working and studying. We again estimate a simple RE model, a FE model, and a
9 When we use parental income as an instrument for private transfers, as in Dustmann et al. (2009), Gong (2009), and Kalenkoski and Pabilonia (2010), our first stage F-statistic is very low (1.4 in the model for the binary transfer indicator and 0.1 in the model for continuous transfers). Hence, the instruments do not pass the weak identification tests, and the IV estimation is likely biased in our case. We also estimated RE and FE logit models for employment status; the results confirm the results of our linear model specification in Table 2. Since non-linear models with both fixed effects and dynamics are still non-standard, we confine our analysis to the outlined identification strategy, specifically linear dynamic models.
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
39
Table 3 Regression results for work hours. Random effects (1) Private transfers
(3)
Dynamic/IV (4)
2.562** (1.038)
0.252* (0.147) 0.240 (0.236)
0.253 (0.176) 0.170 (0.337)
0.502** (0.201) 0.105 (0.080) 0.299* (0.176) 0.113 (0.342)
1443
1443
1443
1443
Lagged work hours
Number of observations ar2p ar3p hansenp
(5)
(6)
4.707** (2.193)
0.496*** (0.150) 0.116 (0.071) 0.299** (0.146) 0.230 (0.237)
Amount of transfers2/10
Log(income others in hh)
(2)
3.251*** (0.814)
Amount of transfers
Risk of failure
Fixed effects
0.124 (0.241) 0.145 (0.360) 0.380 (0.402)
0.887*** (0.321) 0.218* (0.127) 0.137 (0.235) 0.240 (0.379) 0.156 (0.325)
760 0.160 0.340 0.382
760 0.169 0.337 0.282
Source: Swiss Household Panel 1999–2011, own calculations. Note. See Table 2.
dynamic model with panel internal instruments for private transfers and the risk of failure in academic studies. Table 3 presents the results for work hours. We obtain a similar pattern of effects as that for the labor force participation equation. The RE model predicts that transfer recipients work, on average, 3.3 fewer hours per week than transfer non-recipients. The FE model predicts a decrease of about 2.6 h. The bias in the RE model suggests that students who tend to work more (owing to their timeconstant unobserved background) are those students who are less likely to receive transfers. As for the labor force participation decision, the differences in estimates for the continuous transfer variable are small between the RE and FE models. The dynamic model predicts a difference of about 4.7 h, i.e., transfer recipients work almost five fewer hours than non-recipients. Regarding continuous transfers,
we find a reduction in work hours per week of between 0.4 and 0.9 for each additional CHF 1000 transferred to the student. Evaluated at the average transfer and the average hours of work, this result implies an (external) income elasticity of labor supply of about 0.65, which is close to the elasticity of 0.6 reported in Dustmann et al. (2009) (using different data and a different identification strategy). The larger impact of transfers in the dynamic model compared to the random and fixed effects models suggests that students who worked more in the previous period are more likely to receive transfers in the current period, which indicates an upward bias in the estimated coefficients (RE/FE estimates are less negative). The estimated impact of private transfers on study hours confirms this result (Table 4), i.e., the estimates are opposite of those for work hours. While the RE model shows an increase of about 2.7 study hours per week for a
Table 4 Regression results for study hours. Random effects (1) Private transfers
2.694*** (0.810)
Amount of transfers
Log (income others in hh)
(3)
Dynamic/IV (4)
1.319 (0.940)
0.278* (0.164) 0.116 (0.232)
0.278 (0.209) 0.007 (0.326)
0.236 (0.162) 0.010 (0.068) 0.300 (0.206) 0.031 (0.329)
1443
1443
1443
1443
Lagged study hours Number of observations ar2p ar3p hansenp
Source: Swiss Household Panel 1999–2011, own calculations. Note. See Table 2.
(5)
(6)
3.282** (1.618)
0.400*** (0.147) 0.057 (0.067) 0.321** (0.162) 0.142 (0.233)
Amount of transfers2/10 Risk of failure
Fixed effects (2)
0.746** (0.330) 0.198 (0.398) 0.508** (0.241) 760 0.295 0.317 0.589
0.827*** (0.274) 0.203** (0.096) 0.694** (0.302) 0.052 (0.366) 0.506** (0.236) 760 0.573 0.318 0.503
40
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
Table 5 Regression results for risk of failure in academic studies. Random effects (1) Private transfers
(2)
0.221* (0.130)
Amount of transfers
(3)
Dynamic/IV (4)
0.345* (0.178)
0.012 (0.037)
0.106* (0.054)
0.019 (0.040) 0.009 (0.024) 0.105* (0.054)
1443
1443
1443
1443
Lagged risk of failure
Number of observations ar2p ar3p hansenp
(5)
(6)
0.651** (0.265)
0.012 (0.031) 0.000 (0.018) 0.018 (0.037)
Amount of transfers2/10 Log (income others in hh)
Fixed effects
0.040 (0.050) 0.383*** (0.132)
0.040 (0.052) 0.007 (0.022) 0.024 (0.051) 0.397*** (0.139)
760 0.194 0.307 0.789
760 0.245 0.368 0.775
Source: Swiss Household Panel 1999–2011, own calculations. Note. See Table 2.
transfer recipient, the effect is about 3.2 study hours per week in the dynamic IV models. The bias is negative in the models neglecting dynamics, indicating that students who studied less in the previous period receive more transfers in the current period. Thus, our results suggest that students who work more and study less in one year are more likely to receive transfers in the next year. A possible explanation for this result is that parents jointly decide with their children about their allocation of time and increase their financial support in order to allow their children to shift their work hours toward study hours. Similar to the results regarding labor force participation, the results for work hours do not provide evidence that the income of other household members and the risk of failure in academic studies significantly affect students’ work hours. However, a higher risk of failure in academic studies increases the time devoted to studying. For both, work hours and study hours, we find positive persistence in the data (although the lagged effects are insignificant for work hours). The lagged effects are about 0.5 for study hours and about 0.2–0.4 for work hours. This time pattern regarding work and study hours as well as the employment decision confirms that students combine their academic studies with part-time work and that they work more in some periods than in other periods, for example, depending on exam periods. 4.3. Perceived risk of failure in academic studies As a final aspect of our analysis, we consider the effect of private transfers on academic achievements. According to the traditional human capital perspective, transfers positively affect study performance through effort and time devoted to studying (e.g., Kalenkoski & Pabilonia, 2010). An opposing view is rooted in the moral hazard literature, which posits that income from transfers creates an educational disincentive (e.g., Bodvarsson & Walker, 2004; Hamilton, 2013). The main idea of this stream of literature is that transfers may offset the economic costs of performing poorly in school.
Unfortunately, the SHP does not contain direct performance-related measures such as grade point averages. Instead, it contains a record regarding the perceived risk of failure in academic studies in the next year, which is measured on an 11-point scale, where 0 means ‘‘no risk at all’’ and 10 means ‘‘sure a risk’’. Of course, the validity and quality of this variable as a proxy for study performance may be questioned. On the one hand, there is a related stream of literature that links failure expectations with academic performance (e.g., Perry, Hladkyj, Pekrun, Clifton, & Chipperfield, 2005; Stupnisky et al., 2007; Stupnisky, Perry, Hall, & Guay, 2012). On the other hand, we can use information on the highest level of education achieved to evaluate the link between failure expectations with academic performance in our dataset. We construct a dummy equal to one for individuals who passed their university studies and zero for individuals who left university without a degree.10 For each individual, we calculate the average perceived risk of failure in academic studies during the time enrolled in a university. Collapsing the data to a cross-section, we estimate a probit model to analyze the relation between the risk of failure variable and the binary outcome of receiving a university degree. We find that a one-point increase in the perceived risk of failure in academic studies is associated with a significant 4.4-percentage-point lower success probability (the result holds even after we control for gender, age, and region of residence). This result supports our view that the perceived risk of failure can be used as a proxy for study performance. We now turn to the effect of private transfers on this variable. As indicated above, any regression model for the risk of failure that includes transfers as the explanatory variable suffers from the same type of simultaneity/
10 We exclude individuals from this sub-analysis if they were enrolled in a university during the last year they were surveyed and if they did not report having received a degree. These individuals may or may not continue their studies in the subsequent year, and hence, we exclude them rather than code them as having failed.
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
endogeneity problems as the labor supply decision. Our preferred specification is therefore an individual unobserved effects model with dynamics in the risk of failure variable and lagged instruments for transfers. Table 5 shows the results. The dynamic IV model yields a significant 0.65-point higher risk of failure in academic studies for transfer recipients than for non-recipients. The impact of the continuous measure is small and insignificant. This result suggests a non-linear form for the transfer effect beyond the quadratic form that we included. We test several specifications, but the only evidence comes from the extensive margin, i.e., the comparison of transfers and no transfers. From traditional human capital theory, we would expect transfer recipients to worry less about financing their studies than non-recipients and, hence, to have more time to study and to be more successful. While the first part of this hypothesis is supported by our data, i.e., transfer recipients study more than non-recipients, this increased study time does not translate into better performance (as measured by our proxy). Instead, the results provide more support for the moral hazard view, i.e., the perceived risk of failure increases for transfer recipients. A complementary interpretation of the estimated pattern could be that students feel more pressure to meet parental expectations if they receive a financial transfer (irrespective of the amount). This interpretation would explain why only the effect of the 0/1 dummy for transfers on the risk of failure is statistically significant and large. 5. Concluding remarks Being employed during tertiary education significantly affects student outcomes, both at university and in the labor market. On the one hand, employment provides practical skills and early work experience (Stephenson, 1981; Ruhm, 1997; Hotz, Xu, Tienda, & Ahituv, 2002; Molitor & Leigh, 2005). On the other hand, it detracts from study-related activities, prolongs the duration of studies, and negatively affects academic performance (Ruhm, 1995; Stinebrickner & Stinebrickner, 2003; Oettinger, 2005; Dustmann and van Soest, 2007; Sabia, 2009). In this light, our study yields two major findings. First, providing financial support lowers a student’s likelihood of being employed and, to a large extent, shifts work time toward time devoted to studying. Private transfers do not necessarily increase study performance but instead may raise study-related stress levels through an implicit negative incentive. Second, the dynamics in students’ employment decisions and unobserved heterogeneity are important in the estimation of the impact of private transfers on students’ labor supply, highlighting the importance of using panel data and dynamic models. This empirical approach seems particularly promising, as one has to account for the simultaneous decision process regarding the amount of transfers and the allocation of time, for example, by using panel internal instruments. We also see two shortcomings of our paper. First, and related to the last remark, the use of panel internal instruments is certainly not the best choice, but
41
unfortunately, we do not have access to experimental data, ideally in a panel data setting. Such a design would allow for a rich set of inferences; e.g., one could study dynamic treatment effects while controlling for the intertemporal characteristics of the labor supply. We are not aware of any study that exploits this type of approach. Second, we do not have a direct measure of study performance, such as end-of-year grades or exam results, which might be more reliable than the risk of failure in academic studies. Having precise information on performance might render some of our results significant, particularly regarding the impact of the amount of transfers. While we cannot rule out such a possibility at this stage, it would be constructive to address this issue in future research. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.econedurev.2014.05.005. References Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. The Review of Economic Studies, 58, 277–297. Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of error-components models. Journal of Econometrics, 68, 29– 51. Becker, G. S. (1993). A treatise on the family: Enlarged edition. Harvard University Press. Blundell, R., & Bond, S. (1998). Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics, 87, 115–143. Blundell, R., MaCurdy, T., & Meghir, C. (2007). Labor supply models: Unobserved heterogeneity, nonparticipation and dynamics. In J. J. Heckman & E. E. Leamer (Eds.), Handbook of econometrics. Elsevier. Bodvarsson, O. B., & Walker, R. L. (2004). Do parental cash transfers weaken performance in college? Economics of Education Review, 23, 483495. Cameron, A. C., & Trivedi, P. K. (2005). Microeconometrics. Methods and Applications. Cambridge University Press. Dustmann, C., & Van Soest, A. (2007). Part-time work, school success and school leaving. Empirical Economics, 32, 277–299. Dustmann, C., Micklewright, J., & Van Soest, A. (2009). In-school labour supply, parental transfers, and wages. Empirical Economics, 37, 201–218. Eckstein, Z., & Wolpin, K. I. (1990). On the estimation of labor force participation, job search, and job matching models using panel data. In Y. Weiss & G. Fishelson (Eds.), Advances in the theory and measurement of unemployment. Macmillan. Gong, T. (2009). Do parental transfers reduce youths’ incentives to work? Labour, 23, 653–676. Hamilton, L. T. (2013). More is more or more is less? Parental financial investments during college. American Sociological Review, 78, 70–95. Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators. Econometrica, 50, 1029–1054. Horrace, W. C., & Oaxaca, R. L. (2006). Results on the bias and inconsistency of ordinary least squares for the linear probability model. Economics Letters, 90, 321–327. Hotz, V. J., Xu, L. C., Tienda, M., & Ahituv, A. (2002). Are there returns to the wages of young men from working while in school? The Review of Economics and Statistics, 84, 221–236. Hyslop, D. R. (1999). State dependence, serial correlation and heterogeneity in intertemporal labor force participation of married women. Econometrica, 67, 1255–1294. Juerges, H. (2000). Of Rotten kids and Rawlsian parents: The optimal timing of intergenerational transfers. Journal of Population Economics, 13, 147– 157. Kalenkoski, C., & Pabilonia, S. W. (2010). Parental transfers, student achievement, and the labor supply of college students. Journal of Population Economics, 23, 469–496. Lee, C., & Orazem, P. F. (2010). High school employment, school performance, and college entry. Economics of Education Review, 29, 29–39.
42
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
Molitor, C. J., & Leigh, D. E. (2005). In-school work experience and the returns to two-year and four-year colleges. Economics of Education Review, 24, 459–468. Oettinger, G. S. (2005). Parents financial support, students employment, and academic performance in college. Austin: University of Texas (unpublished manuscript). Pabilonia, S. W. (2001). Evidence on youth employment, earnings, and parental transfers in the national longitudinal survey of youth 1997. Journal of Human Resources, 36, 795–822. Perry, R. P., Hladkyj, S., Pekrun, R. H., Clifton, R. A., & Chipperfield, J. G. (2005). Perceived academic control and failure in college students: A three-year study of scholastic attainment. Research in Higher Education, 46, 535–569. Roodman, D. (2009). How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal, 9, 86–136. Ruhm, C. J. (1995). The extent and consequences of high school employment. Journal of Labor Research, 16, 293–303. Ruhm, C. J. (1997). Is high school employment consumption or investment? Journal of Labor Economics, 15, 735–776.
Sabia, J. J. (2009). School-year employment and academic performance of young adolescents. Economics of Education Review, 28, 268–276. Stephenson, S. P. (1981). In-school labor force status and post-school wage rates of young men. Applied Economics, 13, 279–302. Stinebrickner, R., & Stinebrickner, T. R. (2003). Working during school and academic performance. Journal of Labor Economics, 21, 473–491. Stupnisky, R. H., Renaud, R. D., Perry, R. P., Ruthig, J. C., Haynes, T. L., & Clifton, R. A. (2007). Comparing self-esteem and perceived control as predictors of first-year college students’ academic achievement. Social Psychology of Education, 10, 303–330. Stupnisky, R. H., Perry, R. P., Hall, N. C., & Guay, F. (2012). Examining perceived control level and instability as predictors of first-year college students’ academic achievement. Contemporary Educational Psychology, 37, 81–90. Voorpostel, M., Tillmann, R., Lebert, F., Kuhn, U., Lipps, O., Ryser, V.-A., et al. (2012). Swiss household panel user guide (1999–2011), wave 13, October 2012. Lausanne: FORS. Wolff, F.-C. (2006). Parental transfers and the labor supply of children. Journal of Population Economics, 19, 853–877.
Contents lists available at ScienceDirect
Economics of Education Review journal homepage: www.elsevier.com/locate/econedurev
Private transfers and college students’ decision to work Andreas Bachmann a,1, Stefan Boes b,* a b
University of Bern, Department of Economics, Schanzeneckstrasse 1, CH-3001 Bern, Switzerland University of Lucerne, Department of Health Sciences and Health Policy, Frohburgstrasse 3, PO Box 4466, CH-6002 Lucerne, Switzerland
A R T I C L E I N F O
A B S T R A C T
Article history: Received 14 June 2012 Received in revised form 6 May 2014 Accepted 25 May 2014 Available online 8 July 2014
We estimate the impact of external financial support on the labor supply of students during their tertiary education. Using a dynamic labor supply model and accounting for the endogeneity of income from private transfers, we find a significantly lower likelihood of being employed for transfer recipients. Our results suggest that private transfers lead to a shift in students’ time allocation, lowering their hours devoted to working and increasing their time devoted to studying. We find evidence for a psychological component of receiving transfers through an increase in the perceived risk of failure in academic studies. ß 2014 Elsevier Ltd. All rights reserved.
JEL classification: I21 I23 J22 Keywords: Financial support Student employment Time trade-off Simultaneity bias
1. Introduction When young adults enter tertiary education, they face a number of expenses, including tuition fees and costs for housing, transportation, and living in general.2 Students therefore often rely on the continuing support of their parents. Alternatively, they may decide to work during their academic studies. These two sources of financial support are not mutually exclusive and, in fact, are highly related, as
* Corresponding author. Tel.: +41 41 229 5949; fax: +41 41 229 5635. E-mail addresses: [email protected] (A. Bachmann), [email protected] (S. Boes). 1 Tel.: +41 31 631 4777; fax: +41 31 631 3783. 2 Actual expenses vary by location and life-style but can be as high as 5000 US dollars per month at top US universities (see, for example, http:// www.topuniversities.com/student-info/student-finance/how-muchdoes-it-cost-study-us [accessed 08.08.13]). In Switzerland, tuition fees are relatively low compared to those in the US, but living expenses can easily add up to 2500 US dollars per month, or more (see http:// www.crus.ch/information-programme/study-in-switzerland.html?L=2#8_Costs [accessed 08.08.13]). http://dx.doi.org/10.1016/j.econedurev.2014.05.005 0272-7757/ß 2014 Elsevier Ltd. All rights reserved.
standard labor supply models predict that a higher nonworking income is associated with a lower propensity to work (e.g., Blundell, MaCurdy, & Meghir, 2007). How does income from private transfers affect the student labor supply? Building on Becker’s work on intrafamily transfers (e.g., Becker, 1993), theoretical models with an endogenous determination of income through part-time employment predict an inverse relationship between transfers and a child’s labor supply (e.g., Juerges, 2000). The underlying mechanisms are highly interdependent, however, because non-working time for students may be not only devoted to leisure but also invested in human capital (through additional time devoted to studying). Parents can influence the time allocation of their children by providing transfers. For example, parents may encourage their child to cut back on work and study more if the child’s academic performance diminishes (Kalenkoski & Pabilonia, 2010). Whether the student works more in some periods and receives more financial support in others is likely a simultaneous decision (e.g., Dustmann, Micklewright, & van Soest, 2009; Lee & Orazem, 2010).
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
This simultaneity poses a severe problem for the estimation of a causal effect of private transfers on the student labor supply. In an early descriptive study, Pabilonia (2001) examines the employment behavior of American youths under the age of 16 years and finds a significantly negative association between labor supply and parental allowances. In more recent studies, Wolff (2006), Dustmann et al. (2009), Gong (2009), and Kalenkoski and Pabilonia (2010) use parental income as an instrument for parental transfers, arguing theoretically that the effect of parental income on a child’s labor supply can only go through a transfer. While the last three studies find a negative impact of transfers on youth employment, Wolff (2006) does not find a significant effect. In all these studies, the endogeneity of transfers plays a major role in the empirical argument, and failing to account for the endogeneity is shown to produce biased estimates. We add to the literature by (i) analyzing a large panel of college students in Switzerland, (ii) employing a different identification strategy from previous studies, and (iii) explicitly addressing students’ allocation of time. The use of longitudinal data serves two purposes. First, we can control for time-constant unobserved confounders. Gong (2009) argues that such confounders are particularly important in this context because they take into account idiosyncratic preferences for working and studying and personality traits that jointly determine the amount of transfers and the student’s labor supply. Second, we can use dynamic labor supply models, which crucially distinguishes our work from Gong (2009).3 As we condition on past employment and individual-specific effects, we can use lagged transfers as instruments for current transfers to identify a causal effect. Using over-identification and model specification tests, we find plausible evidence to support our instruments. This strategy also allows us to test the previously used exclusion restriction on parental income, which we confirm in our data. Our results provide new evidence on the trade-off in students’ allocation of time. Using the dynamic labor supply framework, we find that private transfers significantly reduce work hours and increase study hours. We conclude that transfers provide an incentive for students to shift work time toward study time. On the downside, we do not find an improvement in academic performance: the perceived risk of failure in academic studies is substantially higher for transfer recipients than for non-recipients, which might indicate an implicit increase in study-related stress levels and pressure to perform well associated with transfers. The remainder of the paper is organized as follows. Section 2 describes the data and the variables that are used in the analysis. Section 3 outlines the model and the estimation methods. Section 4 presents the results. We first analyze the labor supply decision, then look at the time trade-off between working and studying, and finally
3 Gong (2009) does not estimate dynamic labor supply models but rather addresses the simultaneity problem by using a fixed-effects twostage least squares procedure with variation in parental income as the instrument.
35
examine the risk of failure in academic studies. Section 5 discusses the limitations of our study and concludes the paper.
2. Data Our analysis of the impact of private transfers on the student labor supply is based on data from the Swiss Household Panel (SHP), a representative annual panel survey of the Swiss residential population; see Voorpostel et al. (2012).4 The SHP started in 1999 with a total of about 5100 interviewed households. It is a comprehensive survey that covers a wide range of topics, including household and family background, education, work, income, health, and socio-psychological information. In 2004, a refreshment sample of about 2500 households was added to overcome the initially high panel attrition. We employ all waves of the SHP until the most recent, W13 (year 2011). We confine our analysis to individuals enrolled at a higher education institution.5 The students can be fulltime or part-time students. The maximum sample consists of 857 students and 4419 person-year observations. The response rate among students per year is about 50%, which is slightly lower than the response rate for the total SHP (about 60%). We checked whether panel attrition could be a problem in the data we use, but we found little supporting evidence. First, attrition in the total SHP only affects a few variables, mainly political and leisure variables that are unrelated to our main variables (Voorpostel et al., 2012). Second, we estimated a simple logit model for the probability of missing information as a function of gender, age, and location of residence. None of these variables is statistically significant. Even if we include the mean transfers over the non-missing years as a predictor, this variable has no explanatory power (the detailed results are available upon request). We use information on labor supply for our main outcome employment: a dummy equal to one if the student is actively occupied and zero otherwise. Actively occupied individuals in the SHP comprise all respondents who are employed or self-employed (even if the amount of work is only one hour per week). In addition, we use work hours and study hours per week as outcome variables to investigate students’ allocation of time and the impact of private transfers on this allocation. As a final outcome, we take the perceived risk of failure in academic studies in the next 12 months, which also serves as an endogenous right-hand side variable in the labor supply models. This variable is recorded on an 11-point scale, where 0 means ‘‘no risk at all’’ and 10 means ‘‘sure a risk’’. Our key explanatory variable is yearly social informal transfers (in Swiss Francs, CHF), consisting of all payments received from people living in the same household and outside the household, henceforth referred to as amount of private transfers. While these transfers are directly paid to
4
Free access to the data can be acquired via http://www.swisspanel.ch. Most students study at an academic university (about 84 percent), but the sample also includes students of universities of teacher education (about 4 percent) and universities of applied sciences (about 12 percent). 5
36
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
the student, there may also be indirect transfers from other household members, for example, if they pay for accommodation or utilities. The SHP does not record such transfers, but the amount of such transfers is likely correlated with the income of other household members, which we use as a proxy control. For each individual i in year t, the income of all other household members is constructed as the sum of total net personal income in t over all members except i. Both monetary variables are inflation adjusted with 2011 as the base year. Information on the income of other household members is closely related to the exclusion restriction imposed in previous studies (Dustmann et al., 2009; Gong, 2009; Kalenkoski & Pabilonia, 2010). If the student still lives in the parents’ house (which is the case for more than 80% of the respondents in our sample), this information mainly captures parental income. It also contains the income of partners. Nevertheless, it may not be a priori clear whether such information can be excluded from the labor supply decision, as it may be associated with non-monetary benefits for the student (e.g., through contacts to possible employers). Hence, we do not impose this restriction but instead include the income of other household members as a separate control. We instrument direct transfers using panel internal information (lagged transfers, to be specific). In doing so, we can test the previously imposed exclusion restriction. Table 1 shows basic descriptive statistics for our sample. After deleting observations with missing information, we obtain a final (estimation) sample of 672 students (1443 student-year observations). Of these, 439 have been employed in at least one wave, and 554 received transfers at least once. The average employment rate is about 52 percent. If we compare students who received transfers with those who did not, the former have an employment rate that is about 13 percentage points lower (48% compared to 61%, respectively). The average transfer paid is about CHF 6300 per year. Apart from employment, there are substantial differences between transfer recipients and
non-recipients: on average, non-recipients work more and spend less time studying, and the amount of income of the other members in their households is lower. 3. Econometric model and estimation strategy While Table 1 suggests that a negative association exists between transfers and student employment, the raw difference is likely biased with respect to the causal effect. First, the raw difference does not adjust for confounding influences, such as gender, age, and individual preferences. Second, transfers themselves depend on the labor supply of students. Regarding the student labor supply, it can be conjectured that the mean comparison is downward biased, as working more reduces the amount of transfers (Dustmann et al., 2009). For the confounders, the bias is ambiguous. More motivated students might receive more transfers, but they might also be more likely to engage in extracurricular activities (such as part-time employment). Preferences might work in the opposite direction if students enjoy their independence and try to avoid receiving financial support from their parents. Following Eckstein and Wolpin (1990), Hyslop (1999), and Blundell et al. (2007), we specify a linear dynamic panel data model for the labor force participation of student i in time period t with individual-specific effects as yit ¼ ayi;t1 þ xit 0 b þ ci þ uit
(1)
where yit is the employment status in period t, yi;t1 is lagged employment, xit is the vector of time-varying regressors (including private transfers, the risk of failure in academic studies, and the log of the income of other household members), ci is an individual-specific unobserved effect that may be correlated with xit , and uit is a time-varying heteroscedastic error. Several aspects of the model should be highlighted. First, estimation by ordinary least squares will unlikely produce consistent results, as the lagged dependent
Table 1 Descriptive statistics by financial transfer status. No transfers Mean Employed (yes/no) Hours of work Hours spent studying Risk of failure in academic studies (0–10 scale) Amount of transfers (in CHF 1000) Income of others in household (in CHF 1000) Living with parents (yes/no) Female Age Married (yes/no)
61.5% 11.48 34.01 2.43 – 89.24 86.0% 59.3% 21.89 2.2%
Reference: no partner Partner, living together Partner, not living together
9.0% 45.2%
Kids (yes/no) Number of kids
27.8% 0.388
Number of observations Source: Swiss Household Panel 1999–2011, own calculations.
356
Transfers SD 14.20 13.86 2.03 82.29
1.65
Mean 48.4% 7.25 38.89 2.73 6.29 110.69 84.1% 52.8% 21.74 1.1%
SD 12.13 13.01 2.18 5.37 89.19
1.61
7.7% 44.9%
0.763
22.4% 0.281 1087
0.580
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
37
Table 2 Regression results for student employment. Random effects (1) Private transfers
(3)
Dynamic/IV (4)
0.115*** (0.044)
0.013** (0.006) 0.003 (0.009)
0.011 (0.008) 0.010 (0.014)
0.018** (0.008) 0.006** (0.003) 0.013 (0.008) 0.012 (0.014)
1443
1443
1443
1443
Lagged employment
Number of observations ar2p ar3p hansenp
(5)
(6)
0.192** (0.077)
0.017*** (0.006) 0.004 (0.003) 0.015** (0.006) 0.004 (0.009)
Amount of transfers2/10
Log (income others in hh)
(2)
0.105*** (0.031)
Amount of transfers
Risk of failure
Fixed effects
0.001 (0.010) 0.015 (0.012) 0.445*** (0.160)
0.032** (0.013) 0.010** (0.005) 0.003 (0.010) 0.015 (0.012) 0.452** (0.165)
760 0.418 0.325 0.710
760 0.424 0.328 0.646
Source: Swiss Household Panel 1999–2011, own calculations. Note. All models control for civil status, living in parents’ household, number of kids and year dummies. (1) and (2) add gender, a second order polynomial in age, Swiss nationality and prior education completed. (5) and (6) are estimated by the Blundell and Bond (1998) system GMM procedure with lagged dependent variable, transfers and risk of failure instrumented. ar2p, ar3p indicate p-values for the Arellano and Bond (1991) autocorrelation tests of order 2 and 3. hansenp indicates the p-value for the over-identification test as proposed by Hansen (1982). Cluster-adjusted standard errors in parentheses. * p < 0.1. ** p < 0.05. *** p < 0.01.
variable, by model assumption, is correlated with the individual effect. As outlined in Arellano and Bond (1991), estimation of (1) may proceed by first differencing the level equation to eliminate ci and to use yi;t2 (and earlier) to instrument the lagged difference on the right-hand side. Instruments are needed because the difference transformation generates a first-order serial correlation in the errors such that the lagged difference in employment status is an endogenous regressor. Second, Arellano and Bover (1995) raise the issue that lagged levels can be weak instruments, particularly with persistent data. For all our dependent variables (employment status, as well as work hours, study hours, and risk of failure in academic studies), we cannot rule out a priori the possibility that they are persistent variables. We therefore explore lagged differences in the dependent variable as additional instruments for the level equation, as in Blundell and Bond (1998). This procedure also improves the precision of the estimator. Third, estimation is performed by using the Stata command xtabond2 of Roodman (2009). The dynamic instrumental variable (IV) strategy requires at least three consecutive waves of data for an individual to be included in the estimation, which we have for 228 students (760 person-year observations). Fourth, the instruments are only valid if the level errors do not exhibit serial correlation, an assumption that can be tested and that we cannot reject in any of our models. Fifth, the simultaneity of transfers and employment yields an additional correlation between uit and current transfers that would bias the results if it is not accounted for. Under the assumption of serially uncorrelated errors, we can instrument current transfers (or the differences thereof due to the elimination of ci ) with two- and three-period lags of the transfer variable. These lags convincingly pass
the relevance condition (F-statistics of joint significance larger than 20), and we use over-identification tests to check the plausibility of the exogeneity condition. Again, for all our models, we find no indication of a violation of these tests. Sixth, as discussed in Kalenkoski and Pabilonia (2010), there might be additional simultaneity with study performance. We include the perceived risk of failure in academic studies as a control variable and instrument it analogously to transfers. We discuss the quality of this variable as a proxy for academic performance below. Seventh, with employment status as the dependent variable, we obtain a linear probability model. On the one hand, this type of model implies heteroscedastic errors, which we address by using heteroscedasticityrobust standard errors. On the other hand, estimated probabilities from the model are not restricted to the unit interval. Since we are mainly interested in estimating mean effects and since model predictions outside the unit interval occur for less than one percent of the observations in our sample, this issue should not substantially affect our results (see also Horrace & Oaxaca, 2006). Eighth, we use the same type of model to estimate the impact of private transfers on both work hours and study hours, again instrumenting the lagged dependent variable, transfers, and the risk of failure in academic studies. Finally, we add control variables for gender, age, nationality, highest level of education achieved before current studies,6 and family situation (living with parents,
6 This variable includes indicators for compulsory education, apprenticeship, (higher) vocational education, maturity, and a previous tertiary degree (the base category is no compulsory education).
38
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
living with a partner, marital status, and number of children in the household), and we allow for regional and time dummies to capture general labor market conditions. It should be noted that there are no restrictions on these control variables in a random effects model without dynamics (except for the usual exclusion of the reference group for categorical variables). In the dynamic models, the first differences eliminate not only the unobserved effect ci but also all time-constant controls. 4. Private transfers and student labor supply 4.1. The employment decision Our analysis starts with the effect of private transfers on the likelihood of being employed. We first ignore dynamics and simultaneity in the decision-making process by estimating a random effects (RE) model for student employment. Columns 1 and 2 in Table 2 present the estimated coefficients for a model with an indicator for transfers and a more flexible model with a quadratic form of the amount of transfers, respectively. Compared to the raw difference in Table 1, we regression-adjust for the controls listed above. The results suggest that the probability of being employed is about 10.5 percentage points lower for transfer recipients than for non-recipients. Increasing the amount of transfers by CHF 1000 is predicted to reduce the probability of working by about 0.9 to 1.7 percentage points.7 The effect of the income of the other household members is negligible8, and students who are at high risk of failure in academic studies are less likely to be employed (an increase in perceived risk by one point on the scale is predicted to reduce the likelihood of being employed by about 1.3 percentage points). In columns 3 and 4, we relax the zero-correlation assumption between the individual-specific effect and the right-hand side variables. In these fixed effects (FE) models, the effect of private transfers on employment is somewhat larger, but still in the range of the RE model. The difference in the likelihood of working between transfer recipients and non-recipients is now predicted to be 11.5 percentage points, and the continuous effect is about 0.6 to 1.8 percentage points per additional CHF 1000 of income from transfers. The latter reduction is largest (in absolute terms) for the initial transfer. The diminishing effect may be explained by a decreasing marginal utility of consumption and/or adaptation effects, i.e., students may become accustomed to the transfers that they receive. We reject the RE models against the FE models (Hausman test p-values less than 105), highlighting the importance of including individual-specific unobserved heterogeneity in the labor force participation equation.
7 Here and in the following analysis, we consider a range of CHF 0 to 10,000 to interpret an increase of CHF 1000, which covers about 80 percent of the transfer distribution. 8 Measurement error in the income variable, in particular as a proxy for indirect transfers, could be a possible explanation for the insignificant results. Under classical measurement error theory (e.g., Cameron & Trivedi, 2005, chap. 26), this would bias the results toward zero. We cannot rule out the possibility of such error in our analysis.
Private transfers might still be endogenous in the FE model if, for example, a student works more in one period and if parents, in anticipation of their child’s labor supply, decide to provide less financial support in that period (we discuss evidence in favor of such an argument below). Exploring the panel structure further, we can instrument private transfers (after first differencing) with the two- and three-period lagged levels of private transfers. We consider such an IV strategy to be credible because we rule out the possibility of an inter-temporal relationship by conditioning on past employment. The results of the dynamic IV model are presented in columns 5 and 6. Since the model is over-identified, we can use a Hansen test to evaluate the validity of the implied moment conditions, which is not rejected in any of the models. The estimated difference in the propensity to work between transfer recipients and non-recipients is predicted to be 19.2 percentage points, and the continuous effect is about 1.2 to 3.2 percentage points. Regarding the dynamics, we find positive persistence in the data; that is, students who decided to work in the last period are about 45 percentage points more likely to work in the current period. Thus, not all students who work in one year continue to work in the next year. As we condition on year fixed effects and thus control for the general business cycle, one possible explanation for this result may be the work-study cycle of students in Switzerland, who work in some years but not in others. The latter is reflected in the employment transition matrix, with almost 25 percent of the students switching their status from employed to not employed and about 40 percent switching their status in the other direction from one year to another. It should be noted that the estimates are robust to further lags in employment and any lags of private transfers, which are all insignificant (adjusting the instrument matrix accordingly). Our results are consistent with those in Dustmann et al. (2009), Gong (2009), and Kalenkoski and Pabilonia (2010), who allow for mutual dependence in the work-transfer relationship but use a different identification strategy. To the extent that income of the other household members resembles their parental income instrument, our small and insignificant estimates highlight the validity of the exclusion restriction imposed in these studies, at least in our data.9 4.2. Allocation of time In a next step, we examine the trade-off in the allocation of time between working and studying. We again estimate a simple RE model, a FE model, and a
9 When we use parental income as an instrument for private transfers, as in Dustmann et al. (2009), Gong (2009), and Kalenkoski and Pabilonia (2010), our first stage F-statistic is very low (1.4 in the model for the binary transfer indicator and 0.1 in the model for continuous transfers). Hence, the instruments do not pass the weak identification tests, and the IV estimation is likely biased in our case. We also estimated RE and FE logit models for employment status; the results confirm the results of our linear model specification in Table 2. Since non-linear models with both fixed effects and dynamics are still non-standard, we confine our analysis to the outlined identification strategy, specifically linear dynamic models.
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
39
Table 3 Regression results for work hours. Random effects (1) Private transfers
(3)
Dynamic/IV (4)
2.562** (1.038)
0.252* (0.147) 0.240 (0.236)
0.253 (0.176) 0.170 (0.337)
0.502** (0.201) 0.105 (0.080) 0.299* (0.176) 0.113 (0.342)
1443
1443
1443
1443
Lagged work hours
Number of observations ar2p ar3p hansenp
(5)
(6)
4.707** (2.193)
0.496*** (0.150) 0.116 (0.071) 0.299** (0.146) 0.230 (0.237)
Amount of transfers2/10
Log(income others in hh)
(2)
3.251*** (0.814)
Amount of transfers
Risk of failure
Fixed effects
0.124 (0.241) 0.145 (0.360) 0.380 (0.402)
0.887*** (0.321) 0.218* (0.127) 0.137 (0.235) 0.240 (0.379) 0.156 (0.325)
760 0.160 0.340 0.382
760 0.169 0.337 0.282
Source: Swiss Household Panel 1999–2011, own calculations. Note. See Table 2.
dynamic model with panel internal instruments for private transfers and the risk of failure in academic studies. Table 3 presents the results for work hours. We obtain a similar pattern of effects as that for the labor force participation equation. The RE model predicts that transfer recipients work, on average, 3.3 fewer hours per week than transfer non-recipients. The FE model predicts a decrease of about 2.6 h. The bias in the RE model suggests that students who tend to work more (owing to their timeconstant unobserved background) are those students who are less likely to receive transfers. As for the labor force participation decision, the differences in estimates for the continuous transfer variable are small between the RE and FE models. The dynamic model predicts a difference of about 4.7 h, i.e., transfer recipients work almost five fewer hours than non-recipients. Regarding continuous transfers,
we find a reduction in work hours per week of between 0.4 and 0.9 for each additional CHF 1000 transferred to the student. Evaluated at the average transfer and the average hours of work, this result implies an (external) income elasticity of labor supply of about 0.65, which is close to the elasticity of 0.6 reported in Dustmann et al. (2009) (using different data and a different identification strategy). The larger impact of transfers in the dynamic model compared to the random and fixed effects models suggests that students who worked more in the previous period are more likely to receive transfers in the current period, which indicates an upward bias in the estimated coefficients (RE/FE estimates are less negative). The estimated impact of private transfers on study hours confirms this result (Table 4), i.e., the estimates are opposite of those for work hours. While the RE model shows an increase of about 2.7 study hours per week for a
Table 4 Regression results for study hours. Random effects (1) Private transfers
2.694*** (0.810)
Amount of transfers
Log (income others in hh)
(3)
Dynamic/IV (4)
1.319 (0.940)
0.278* (0.164) 0.116 (0.232)
0.278 (0.209) 0.007 (0.326)
0.236 (0.162) 0.010 (0.068) 0.300 (0.206) 0.031 (0.329)
1443
1443
1443
1443
Lagged study hours Number of observations ar2p ar3p hansenp
Source: Swiss Household Panel 1999–2011, own calculations. Note. See Table 2.
(5)
(6)
3.282** (1.618)
0.400*** (0.147) 0.057 (0.067) 0.321** (0.162) 0.142 (0.233)
Amount of transfers2/10 Risk of failure
Fixed effects (2)
0.746** (0.330) 0.198 (0.398) 0.508** (0.241) 760 0.295 0.317 0.589
0.827*** (0.274) 0.203** (0.096) 0.694** (0.302) 0.052 (0.366) 0.506** (0.236) 760 0.573 0.318 0.503
40
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
Table 5 Regression results for risk of failure in academic studies. Random effects (1) Private transfers
(2)
0.221* (0.130)
Amount of transfers
(3)
Dynamic/IV (4)
0.345* (0.178)
0.012 (0.037)
0.106* (0.054)
0.019 (0.040) 0.009 (0.024) 0.105* (0.054)
1443
1443
1443
1443
Lagged risk of failure
Number of observations ar2p ar3p hansenp
(5)
(6)
0.651** (0.265)
0.012 (0.031) 0.000 (0.018) 0.018 (0.037)
Amount of transfers2/10 Log (income others in hh)
Fixed effects
0.040 (0.050) 0.383*** (0.132)
0.040 (0.052) 0.007 (0.022) 0.024 (0.051) 0.397*** (0.139)
760 0.194 0.307 0.789
760 0.245 0.368 0.775
Source: Swiss Household Panel 1999–2011, own calculations. Note. See Table 2.
transfer recipient, the effect is about 3.2 study hours per week in the dynamic IV models. The bias is negative in the models neglecting dynamics, indicating that students who studied less in the previous period receive more transfers in the current period. Thus, our results suggest that students who work more and study less in one year are more likely to receive transfers in the next year. A possible explanation for this result is that parents jointly decide with their children about their allocation of time and increase their financial support in order to allow their children to shift their work hours toward study hours. Similar to the results regarding labor force participation, the results for work hours do not provide evidence that the income of other household members and the risk of failure in academic studies significantly affect students’ work hours. However, a higher risk of failure in academic studies increases the time devoted to studying. For both, work hours and study hours, we find positive persistence in the data (although the lagged effects are insignificant for work hours). The lagged effects are about 0.5 for study hours and about 0.2–0.4 for work hours. This time pattern regarding work and study hours as well as the employment decision confirms that students combine their academic studies with part-time work and that they work more in some periods than in other periods, for example, depending on exam periods. 4.3. Perceived risk of failure in academic studies As a final aspect of our analysis, we consider the effect of private transfers on academic achievements. According to the traditional human capital perspective, transfers positively affect study performance through effort and time devoted to studying (e.g., Kalenkoski & Pabilonia, 2010). An opposing view is rooted in the moral hazard literature, which posits that income from transfers creates an educational disincentive (e.g., Bodvarsson & Walker, 2004; Hamilton, 2013). The main idea of this stream of literature is that transfers may offset the economic costs of performing poorly in school.
Unfortunately, the SHP does not contain direct performance-related measures such as grade point averages. Instead, it contains a record regarding the perceived risk of failure in academic studies in the next year, which is measured on an 11-point scale, where 0 means ‘‘no risk at all’’ and 10 means ‘‘sure a risk’’. Of course, the validity and quality of this variable as a proxy for study performance may be questioned. On the one hand, there is a related stream of literature that links failure expectations with academic performance (e.g., Perry, Hladkyj, Pekrun, Clifton, & Chipperfield, 2005; Stupnisky et al., 2007; Stupnisky, Perry, Hall, & Guay, 2012). On the other hand, we can use information on the highest level of education achieved to evaluate the link between failure expectations with academic performance in our dataset. We construct a dummy equal to one for individuals who passed their university studies and zero for individuals who left university without a degree.10 For each individual, we calculate the average perceived risk of failure in academic studies during the time enrolled in a university. Collapsing the data to a cross-section, we estimate a probit model to analyze the relation between the risk of failure variable and the binary outcome of receiving a university degree. We find that a one-point increase in the perceived risk of failure in academic studies is associated with a significant 4.4-percentage-point lower success probability (the result holds even after we control for gender, age, and region of residence). This result supports our view that the perceived risk of failure can be used as a proxy for study performance. We now turn to the effect of private transfers on this variable. As indicated above, any regression model for the risk of failure that includes transfers as the explanatory variable suffers from the same type of simultaneity/
10 We exclude individuals from this sub-analysis if they were enrolled in a university during the last year they were surveyed and if they did not report having received a degree. These individuals may or may not continue their studies in the subsequent year, and hence, we exclude them rather than code them as having failed.
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
endogeneity problems as the labor supply decision. Our preferred specification is therefore an individual unobserved effects model with dynamics in the risk of failure variable and lagged instruments for transfers. Table 5 shows the results. The dynamic IV model yields a significant 0.65-point higher risk of failure in academic studies for transfer recipients than for non-recipients. The impact of the continuous measure is small and insignificant. This result suggests a non-linear form for the transfer effect beyond the quadratic form that we included. We test several specifications, but the only evidence comes from the extensive margin, i.e., the comparison of transfers and no transfers. From traditional human capital theory, we would expect transfer recipients to worry less about financing their studies than non-recipients and, hence, to have more time to study and to be more successful. While the first part of this hypothesis is supported by our data, i.e., transfer recipients study more than non-recipients, this increased study time does not translate into better performance (as measured by our proxy). Instead, the results provide more support for the moral hazard view, i.e., the perceived risk of failure increases for transfer recipients. A complementary interpretation of the estimated pattern could be that students feel more pressure to meet parental expectations if they receive a financial transfer (irrespective of the amount). This interpretation would explain why only the effect of the 0/1 dummy for transfers on the risk of failure is statistically significant and large. 5. Concluding remarks Being employed during tertiary education significantly affects student outcomes, both at university and in the labor market. On the one hand, employment provides practical skills and early work experience (Stephenson, 1981; Ruhm, 1997; Hotz, Xu, Tienda, & Ahituv, 2002; Molitor & Leigh, 2005). On the other hand, it detracts from study-related activities, prolongs the duration of studies, and negatively affects academic performance (Ruhm, 1995; Stinebrickner & Stinebrickner, 2003; Oettinger, 2005; Dustmann and van Soest, 2007; Sabia, 2009). In this light, our study yields two major findings. First, providing financial support lowers a student’s likelihood of being employed and, to a large extent, shifts work time toward time devoted to studying. Private transfers do not necessarily increase study performance but instead may raise study-related stress levels through an implicit negative incentive. Second, the dynamics in students’ employment decisions and unobserved heterogeneity are important in the estimation of the impact of private transfers on students’ labor supply, highlighting the importance of using panel data and dynamic models. This empirical approach seems particularly promising, as one has to account for the simultaneous decision process regarding the amount of transfers and the allocation of time, for example, by using panel internal instruments. We also see two shortcomings of our paper. First, and related to the last remark, the use of panel internal instruments is certainly not the best choice, but
41
unfortunately, we do not have access to experimental data, ideally in a panel data setting. Such a design would allow for a rich set of inferences; e.g., one could study dynamic treatment effects while controlling for the intertemporal characteristics of the labor supply. We are not aware of any study that exploits this type of approach. Second, we do not have a direct measure of study performance, such as end-of-year grades or exam results, which might be more reliable than the risk of failure in academic studies. Having precise information on performance might render some of our results significant, particularly regarding the impact of the amount of transfers. While we cannot rule out such a possibility at this stage, it would be constructive to address this issue in future research. Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.econedurev.2014.05.005. References Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. The Review of Economic Studies, 58, 277–297. Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of error-components models. Journal of Econometrics, 68, 29– 51. Becker, G. S. (1993). A treatise on the family: Enlarged edition. Harvard University Press. Blundell, R., & Bond, S. (1998). Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics, 87, 115–143. Blundell, R., MaCurdy, T., & Meghir, C. (2007). Labor supply models: Unobserved heterogeneity, nonparticipation and dynamics. In J. J. Heckman & E. E. Leamer (Eds.), Handbook of econometrics. Elsevier. Bodvarsson, O. B., & Walker, R. L. (2004). Do parental cash transfers weaken performance in college? Economics of Education Review, 23, 483495. Cameron, A. C., & Trivedi, P. K. (2005). Microeconometrics. Methods and Applications. Cambridge University Press. Dustmann, C., & Van Soest, A. (2007). Part-time work, school success and school leaving. Empirical Economics, 32, 277–299. Dustmann, C., Micklewright, J., & Van Soest, A. (2009). In-school labour supply, parental transfers, and wages. Empirical Economics, 37, 201–218. Eckstein, Z., & Wolpin, K. I. (1990). On the estimation of labor force participation, job search, and job matching models using panel data. In Y. Weiss & G. Fishelson (Eds.), Advances in the theory and measurement of unemployment. Macmillan. Gong, T. (2009). Do parental transfers reduce youths’ incentives to work? Labour, 23, 653–676. Hamilton, L. T. (2013). More is more or more is less? Parental financial investments during college. American Sociological Review, 78, 70–95. Hansen, L. P. (1982). Large sample properties of generalized method of moments estimators. Econometrica, 50, 1029–1054. Horrace, W. C., & Oaxaca, R. L. (2006). Results on the bias and inconsistency of ordinary least squares for the linear probability model. Economics Letters, 90, 321–327. Hotz, V. J., Xu, L. C., Tienda, M., & Ahituv, A. (2002). Are there returns to the wages of young men from working while in school? The Review of Economics and Statistics, 84, 221–236. Hyslop, D. R. (1999). State dependence, serial correlation and heterogeneity in intertemporal labor force participation of married women. Econometrica, 67, 1255–1294. Juerges, H. (2000). Of Rotten kids and Rawlsian parents: The optimal timing of intergenerational transfers. Journal of Population Economics, 13, 147– 157. Kalenkoski, C., & Pabilonia, S. W. (2010). Parental transfers, student achievement, and the labor supply of college students. Journal of Population Economics, 23, 469–496. Lee, C., & Orazem, P. F. (2010). High school employment, school performance, and college entry. Economics of Education Review, 29, 29–39.
42
A. Bachmann, S. Boes / Economics of Education Review 42 (2014) 34–42
Molitor, C. J., & Leigh, D. E. (2005). In-school work experience and the returns to two-year and four-year colleges. Economics of Education Review, 24, 459–468. Oettinger, G. S. (2005). Parents financial support, students employment, and academic performance in college. Austin: University of Texas (unpublished manuscript). Pabilonia, S. W. (2001). Evidence on youth employment, earnings, and parental transfers in the national longitudinal survey of youth 1997. Journal of Human Resources, 36, 795–822. Perry, R. P., Hladkyj, S., Pekrun, R. H., Clifton, R. A., & Chipperfield, J. G. (2005). Perceived academic control and failure in college students: A three-year study of scholastic attainment. Research in Higher Education, 46, 535–569. Roodman, D. (2009). How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal, 9, 86–136. Ruhm, C. J. (1995). The extent and consequences of high school employment. Journal of Labor Research, 16, 293–303. Ruhm, C. J. (1997). Is high school employment consumption or investment? Journal of Labor Economics, 15, 735–776.
Sabia, J. J. (2009). School-year employment and academic performance of young adolescents. Economics of Education Review, 28, 268–276. Stephenson, S. P. (1981). In-school labor force status and post-school wage rates of young men. Applied Economics, 13, 279–302. Stinebrickner, R., & Stinebrickner, T. R. (2003). Working during school and academic performance. Journal of Labor Economics, 21, 473–491. Stupnisky, R. H., Renaud, R. D., Perry, R. P., Ruthig, J. C., Haynes, T. L., & Clifton, R. A. (2007). Comparing self-esteem and perceived control as predictors of first-year college students’ academic achievement. Social Psychology of Education, 10, 303–330. Stupnisky, R. H., Perry, R. P., Hall, N. C., & Guay, F. (2012). Examining perceived control level and instability as predictors of first-year college students’ academic achievement. Contemporary Educational Psychology, 37, 81–90. Voorpostel, M., Tillmann, R., Lebert, F., Kuhn, U., Lipps, O., Ryser, V.-A., et al. (2012). Swiss household panel user guide (1999–2011), wave 13, October 2012. Lausanne: FORS. Wolff, F.-C. (2006). Parental transfers and the labor supply of children. Journal of Population Economics, 19, 853–877.