Does raising the school leaving age reduce teacher effort? Evidence from a policy experiment.

Green, Colin ; Paniagua, Maria Navarro

I. INTRODUCTION

Raising the compulsory school leaving age (henceforth RoSLA) is a key policy instrument used to increase minimum educational attainment levels. Moreover, there is an ongoing debate in a number of countries, such as the United Kingdom and Spain, regarding further increases in the minimum high school leaving age. In addition, RoSLA has been widely used in the literature on returns to education as a source of exogenous variation in years of schooling/educational levels (see for instance, Harmon and Walker 1995 for the United Kingdom; Pischke and von Wachter 2008 for Germany; Pons and Gonzalo 2002 and Arrazola et al. 2003 who use the 1970 RoSLA in Spain).

However, teachers who take classes in the "affected" years of schooling are unlikely to be indifferent to this policy change. (1) Increasing the compulsory schooling age increases the number of students in those years, but also changes the distribution of ability and motivation of students that teachers have to instruct. For instance, teachers at the latter part of compulsory secondary school will now have lower ability students and/or those with less interest in formal schooling in their class, along with those students who would have voluntarily chosen post-compulsory schooling in the absence of the legislative change. Teaching (and managing) these students is likely to be more difficult. In the absence of compensating differentials, it is difficult to imagine that this will not affect teacher motivation and effort. (2)

This paper is the first to our knowledge that investigates this motivational effect of compulsory schooling laws on teachers. Specifically, we examine the impact of the increase in the school leaving age that occurred in Spain in the academic year 1998-1999 on one element of high school teacher behavior, absenteeism. Employing Spanish Labor Force Survey (SLFS) data that covers the relevant policy reform period, we estimate Difference-in-Difference models of absenteeism using count data approaches. We demonstrate that raising the compulsory schooling age led to an increase in teacher absenteeism. This is a matter of concern as previous research has demonstrated a negative causal effect of teacher absence on student achievement (Duflo, Hanna, and Ryan 2005; Miller, Murnane, and Willett 2007, 2008), and one that is more severe for students from lower socioeconomic backgrounds (Clotfelter, Ladd, and Vigdor 2009). This may be the result of absent teachers being replaced by less qualified substitutes and/or the disruption inherent in the use of replacement teachers.

This leads to a concern that increasing the compulsory school leaving age may decrease the quality of educational provision in the affected years. Furthermore, this has implications for estimation of the returns to education that rely upon RoSLA reforms as an instrumental variable. Namely, that if the education treatment because of RoSLA is of lower quality, then the local average treatment effect of education on wages will be biased downward.

II. DATA

The policy reform examined consisted of an extension of free, compulsory, and comprehensive education from 14 to 16 years. The reform was the General Regulation of the Education System passed in 1990. (3) Because of the economic crisis of the 1990s in Spain, the change in compulsory schooling was delayed until the last quarter of 1998. Primary and high school curriculum standardization across Spain was another part of this reform. Primary and early secondary school standardization occurred before the changing of the compulsory school leaving age. While curriculum standardization for 15- to 16-year-old students occurred at the same time, the compulsory leaving age was raised. This could influence high school teacher behavior if, for instance, this adjustment was onerous and/or disruptive. Hence, any RoSLA policy effect could include an adjustment effect as a result of curriculum change. We examine the potential for the policy effect to be confounded by this implementation period in later robustness checks. From the last quarter of 1998 on students that otherwise would have dropped out (in the previous academic year were in the last year of compulsory schooling) were obliged to stay two more years at school. This leads to compulsory education comprising a total of 10 years, divided into two educational levels: primary education (6-12 years old) and lower secondary education that it is ordinarily completed from the ages of 12-16 years old. (4)

Increasing the school leaving age may, because of the need to increase the teaching labor force, lead to a distributional change in teacher characteristics. This is similar to the point raised by Jepsen and Rivkin (2009) regarding the consequence of decreasing class sizes. Increasing the teaching workforce to cope with an increased student body may lead to, for instance, less qualified or less experienced teachers being hired who may be more frequently absent. Alternatively, increased compulsory schooling could lead to larger class sizes which could increase teacher disutility and/or exposure to contagious illness and thus increase absence. It does not appear, however, that high school class sizes increased at the time of RoSLA in Spain. In fact there was a small decrease, from 26.4 students per class in 1997-1998 to 26.0 and 25.5 in 1998-1999 and 1999-2000, respectively (MEC 2010b). At the same time, there is no evidence of increases in high school teacher numbers, either in official data MEC (2010a) or in the representative sample of high school teachers we use. While we cannot be definitive about the source of this apparent discrepancy, this may have occurred because of the reduction in the age cohort from which high school students were drawn over the 1990s. For instance, there were 3.3 million people in Spain aged 12-16 in 1990 (8.8% of the population) and this fell to 2.6 million (6.3% of the population) by 1998, representing a 31% decrease in this age group. This, coupled with the well-documented strong employment protection in Spain (most high school teachers are civil servants), meant that teachers were not fired when student numbers fell, and this we believe led to an increase in teaching capacity up to the time of the reform.

The data we use is drawn from the SLFS, a quarterly repeated cross-sectional survey that is representative at both the national and regional level in Spain. We select a sample of full-time employees in the period spanning 1st quarter of 1996 to 4th quarter of 2004. Self-employed workers are excluded. The full sample consists of about 988,329 workers, 2.57% of them are high school teachers. We are able to identify high school teachers (and distinguish them from primary school teachers) in our data as the SLFS reports three digit disaggregated occupation codes (ISCO) for all workers. In Spain, secondary education teachers must have a university degree, only teach subjects of their field of specialization, and are primarily civil servants who attained their post through state or regionally competitive exams. In all estimates, we control for both public sector and temporary contracts, 85% of the teachers work in the public sector and 64% are on permanent contracts.

To test the robustness of our results, we use a number of subsamples. This includes dropping the vacation period (third quarter of the year surveys) as teachers have more summer holidays than many other workers. Furthermore, to ensure that the timing of other holidays is not generating our results, we estimate our models on two successively more restrictive samples. The first is workers in the education industry only and in the second we include only primary and secondary school teachers. These latter two groups have essentially identical holiday schedules and provisions. The second sample contains 63,811 workers in the education sector, and the third sample is comprised of 49,711 primary and secondary school teachers (Table 1). Importantly our key results are robust to the choice of these samples.

We use information on the hours of absence per week reported as being due to sickness to generate our dependent variable. (5) We calculate this variable as the difference between usual hours and actual hours for those that report the reason of any difference between them as due to sickness. (6) Generally, paid sick leave is very generous in Spain. While coverage may vary slightly by sector or firm level agreement, the norm is 1 month of absence leave fully paid per year worked in Spain up to a maximum of 18 months leave. If the worker reaches their maximum leave entitlement they have to attend a special panel which assesses their sickness.

We appreciate that using sickness absence only may be quite restrictive. In unreported estimates our main results are robust to using more broad definitions where we include differences in usual and actual hours because of other forms of absence including personal/family responsibilities, bad weather, summer schedule/flexible hours, and "other reasons."

Figure 1 reports minutes of sickness absence before and after RoSLA for three groups, high school teachers, primary school teachers, and all other workers. Prior to the policy introduction, high school teachers had lower absence levels than primary school teachers. This is similar to the pattern found in other jurisdictions such as in Australia (Bradley, Green, and Leeves 2008) and in the United States (Clotfelter, Ladd, and Vigdor 2009). More detail is provided in Appendix Figures A 1 and A2 which show yearly absence rates before and after RoSLA, again for these three groups. These figures suggest that RoSLA increased high school teacher absence but not that of other workers. Moreover, this increase does not seem to dissipate over our sample period.

A range of control variables are available in the SLFS. We use gender, age, age squared, marital status, education, public sector, type of contract, industry dummies, occupation dummies, and size of the firm/establishment. We also control for year, quarter, and regional (CCAA) fixed effects so as to take regional differences into account such as wage differences and unemployment rates because of different industrial structures within regions and patterns of morbidity. (7)

III. METHODOLOGY

In our baseline model, workers' minutes of absence per week can be specified as follows:

(1) [Absm.sub.it] = [phi] + [delta][RoSLA.sub.it] + [gamma][beta][HST.sub.i] + [beta][RoSLA.sub.it] x [HST.sub.i] + [alpha][X.sub.i] + [[epsilon].sub.i]

i = 1, ..., 988,329 and

t = 1996Q1, ..., 2004Q4

where [Absm.sub.it] corresponds to the minutes of absence of worker i in the period t. [RoSLA.sub.it] is an indicator that takes value of unity if the worker is observed during the reform period. [HST.sub.i] is a dummy variable that equals 1 if the worker is a high school teacher and 0 otherwise. And the interaction term [RoSLA.sub.it] x [HST.sub.i] equals 1 for treated individuals (HSTeachers) in the post-treatment period (after the RoSLA was implemented). The OLS estimate of [beta] is equivalent to the Differences-in-Differences (DID) estimator and thus provides the absence caused by the reform for the treated group (i.e., the absence caused by the RoSLA for secondary school teachers) (Cameron and Trivedi 2005, 890-91).

Our dependent variable, minutes of absence, is a count variable. Moreover, there is an excess of zero outcomes. This could occur during the reference week both if (a) the worker/teacher never gets sick and does not skip work during the reference week but could have been absent in the case of illness (sampling zeros); (b) the worker/teacher always goes to work because of commitment and motivation despite illness (structural zeros). As a result, we estimate zero-inflated models that allow for these excessive number of zeroes in addition to overdispersion in the zero-inflated negative binomial (ZINB) case. Importantly, zero-inflated models permit the zero absence of teachers to be explained in a different manner than that of those workers that are absent for more than zero hours. It combines a count density with a binary process in such a way that a binary model is estimated to predict, with probability [[psi].sub.i], whether a worker will always have a zero count (i.e., type b). Then, a count model (Poisson or negative binomial) chosen with probability 1 - [[psi].sub.i] is generated to predict the counts for those who will not always have a zero count (type a). Then [Absm.sub.i] has a zero-inflated distribution given by Long (1997, 242-50). The appropriateness of the zero-inflated models over their non zero-inflated counterparts (i.e., Poisson vs. zero-inflated Poisson [ZIP], negative binomial vs. zero-inflated negative binomial) can be tested using the approach set out by Vuong (1989).

We estimate both ZIP and ZINB models:

(2) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where [f.sub.1] corresponds to a binary process (i.e., logit) and [f.sub.2] is a Poisson or negative binomial density function. An assumption of Poisson models is that the variance equals the sample mean:

(3) E([[mu].sub.i]|[x.sub.i]) = exp ([x.sub.i][beta]) = var ([[mu].sub.i]|[x.sub.i]).

If this is not the case and the variance exceeds the mean then the data are said to be overdispersed and the Poisson model will be inefficient. This can be overcome by using a negative binomial model that adds a random term that is assumed to be uncorrelated with the model's covariates. The Poisson is nested in the negative binomial model, as it is the case where [alpha] = 0 in Equation (5). The null hypothesis of no overdispersion, [alpha] : [H.sub.o] ([alpha] = 0), can be determined via a Likelihood-ratio test.

Zero-inflated Poisson:

(4)[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

Zero-inflated negative binomial:

(5) [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII].

IV. RESULTS

Table 2 presents estimates of the effect of RoSLA on high school teacher absence behavior. Two sets of estimates are reported for the ZIP and ZINB models, respectively. The Vuong test (reported in each of the tables) suggests that our zero-inflated models are a significant improvement over standard Poisson or negative binomial models. The likelihood-ratio test in all cases rejects the null hypothesis of no overdispersion (p value = .000). It indicates that because of overdispersion the ZINB can improve goodness-of-fit over the ZIP. Nonetheless, we report both sets of estimates to demonstrate that our results are not being driven by choice of model. For both models the policy effect dummy (RoSLA x HSTeacher) demonstrates a statistically significant increase in high school teacher absenteeism as a result of the RoSLA. Coefficient estimates from count data models can be difficult to interpret so we also present the incident rate ratio (IRR) or exponentiated coefficient [e.sup.[beta]]. Thus, an IRR greater than one indicates that the expected count in the exposed group is greater than the expected count in the unexposed group. These demonstrate that the effect of the policy was to increase the high school teacher's hours lost through sickness absence by 15% (IRR = 1.15).

Comparing high school teacher absence behavior and that of all other workers may be too broad. For instance, there may be unobserved changes to the incentives for worker absence occurring for other occupations/industries during the same period. Alternatively, there may be some unobserved shock to teachers' absence that coincides with the policy change. These may serve to bias our estimates of the policy effect in some unknown way. To mitigate this problem we examine our two successively more restrictive subsamples of workers, education workers and teachers (high school and primary school). As previously mentioned, focusing on the education sector aims to eliminate the possibility of bias originating from unobserved changes in absence incentives in other industries. While comparing primary and high school teachers only has the added advantage that these workers face a very similar holiday structure and leave timing. The cost of these robustness checks are decreased sample size and potentially less precise estimates.

Table 3 reports the estimates for these two subsamples, where for brevity we only report the key policy variable estimates. Again these estimates reveal that the change in compulsory schooling laws lead to an increase in teachers' hours of sickness absence. Moreover, restricting our sample leads to an increase in the magnitude of this effect to 18%-20%, suggesting that our earlier estimates were biased downward.

An additional issue is that we do not possess a nonpolicy control group (i.e., schools or regions where there is no RoSLA). Hence, there is the potential that our policy effect is in actuality because of some contemporaneous exogenous shock. To examine this, we conduct robustness checks exploiting the regional variation in pre-RoSLA high school participation rates. Specifically, we split our sample according to the post-compulsory high school participation rates of the region prior to the reform. The intuition is that schools in regions where participation rates beyond compulsory education were previously lower should be more affected by RoSLA, that is, a larger proportion of students in these schools are treated by the policy reform. Hence, if it is this policy that is actually causing the change in teacher absence, teachers in areas that previously had lower post-compulsory participation rates in secondary education should respond more. We re-estimate our absence models split according to whether the school was in a region with greater than or less than 20% post-compulsory participation rates prior to RoSLA, (8) and report these estimates in Table 4. (9) The results demonstrate that the policy caused teacher absence to increase by almost 50% in those regions with lower post-compulsory high school education participation before the reform was implemented. This result increases our confidence that the rise in absence we observe is caused by the policy reform.

We conduct four final sets of robustness checks. First, we investigate whether our results are generated by the year of policy implementation. For instance, this may have caused disruption to classrooms and teachers leading to an increase in absence. We do this by excluding the year of the reform from our sample. These estimates are reported in the first row of Table 5 and reveal that omitting the year of the reform does not alter the key results; the magnitude remains at approximately 20%. We use this approach as the basis to investigate the longevity of the policy effect by further omitting the year after the policy reform. The results in the last row of Table 5 demonstrates again the robustness of the main result and suggests that the policy continued to have an effect on high school teacher absence until, at least, 2 years after the reform year. This effect remained of the same magnitude. In addition, no policy effect was evident when we investigated a placebo reform for the year prior to the actual RoSLA.

Second, we re-estimate our models excluding the summer quarter of the SLFS. This is performed for two reasons, first the bulk of school holidays occur in the summer quarter hence the opportunity (or need) for teachers to take sickness absence in this quarter are diminished. Second, it has been suggested that increases in temperature are generally associated with increases in absence (Connolly 2008). Estimates of these further restricted models are reported in Table 6. The pattern of these estimates largely follows those reported in Tables 2 and 3. Together the results in Tables 3, 4, and 6 suggest that our estimated impact of RoSLA on teacher absence is not being driven by unobserved variations in holiday availability/timing or unobserved shocks to the absenteeism of teachers.

As a further robustness check that our results are not being driven by an unobserved exogenous shock to absence at the time of the reform, we examine RoSLA's impact on a group of workers who should be unaffected. Specifically, we exclude all the education sector workers from the sample and treat the remaining public sector workers as the treatment group. In Table 7, we show that the coefficient for this "incorrect" treatment group is not significantly different from 0. This further suggests that the change in teacher absence is because of the change in the school leaving age and not some contemporaneous shock. One might also be concerned that the March 11, 2004 attacks in Madrid may also have influenced Spanish work behavior regarding absenteeism. For instance, Hotchkiss and Pavlova (2009) have demonstrated that the September 11 attacks changed the working hours of New York residents. We investigated this by deleting 2004 from our sample and re-estimating our main models; our key policy estimates were unaffected by this.

Bertrand, Duflo, and Mullainathan (2004) demonstrate that standard errors in DID estimators are inconsistent when panels with longer time periods are used and the dependent variable is serially correlated. We examine this by collapsing the time dimension of our data into two periods, pre- and post-RoSLA. We then re-estimated (1) across these two periods and these are reported in Table 8. Our resulting estimate of the RoSLA effect is statistically significant at the 1% level, and suggests a larger policy effect of at least 50%. This suggests that the statistical significance of our RoSLA estimates is not the result of serial correlation in absenteeism.

Having established a robust policy effect, we now seek to provide some approximate quantification of the magnitude of the policy effect on teacher absence. In addition, while the estimates have been concerned with intensive margins of absence, we also provide some information on the incidence of sickness absence which can be considered a measure of the extensive margin. Panel A of Table 9 reports the RoSLA marginal effects from a probit of the probability of a worker taking sickness absence in the reference week. For our education worker sample this reveals that the increase of high school teachers' absence because of the reform was 1.6 and 2.4 percentage points excluding summer holidays, respectively. Slightly lower magnitudes are reported for the high school versus primary school sector although there is a loss of precision in the estimates and these miss statistical significance at standard levels.

Panel B of Table 9 reports the coefficients of an OLS regression for those workers that took absence due to sickness. Because of the reform, high school teachers take 308 more minutes of absence than their other education sector counterparts. They take 337 more minutes than primary education teachers because of the reform, and the effect is that of 280 and 352 minutes for the two samples, respectively when we exclude the summer quarter from the data. These are broadly in line with the 15%-20% magnitude of policy effect reported in our earlier count data models.

V. CONCLUSION

RoSLA is seen as a key instrument for increasing basic education levels within society. At the same time, it has been relied upon by many researchers as a source of exogenous variation in educational attainment in econometric studies of the returns to education. In this paper, we asked the question, how do teachers react to the changes in teaching environment implicitly created by retaining students at school who would have otherwise left? Specifically, we focus on one potential response, changes in teacher absence behavior.

We examined changes in high school teacher absence behavior because of an increase in the school leaving age in Spain in the academic year 1998-1999. Using representative labor force data we demonstrated that teachers affected by this reform increased their absenteeism by 15%, rising to almost 50% in regions that traditionally had lower post-compulsory school participation. Our interpretation of this result is that more onerous teacher environments lead to decreases in effort by high school teachers. This result is of importance for two related reasons. Given previous research that establishes a link between teacher absence and lower student performance, our results demonstrate that increasing the compulsory school leaving age has the potential to reduce educational quality. Second, our results suggest that researchers using RoSLA or other policy changes that may affect teaching conditions as an instrumental variable should consider their possible effects on educational "quality."

This paper has considered the effect of RoSLA on absence; future research should consider the range of other potential reactions of teachers (i.e., turnover, job satisfaction) and the subsequent effect these have on childrens' outcomes.

Finally, a policy recommendation that is derived from our work is that education authorities should consider the need for increased compensation or improved working conditions for teachers adversely affected by increasing the compulsory school leaving age.

doi: 10.1111/j.1465-7295.2011.00386.x

ABBREVIATIONS

DID: Differences-in-Differences

IRR: Incident Rate Ratio

RoSLA: Raising of the School Leaving Age

SLFS: Spanish Labor Force Survey

ZINB: Zero-Inflated Negative Binomial

ZIP: Zero-Inflated Poisson

APPENDIX

[FIGURE A1 OMITTED]

[FIGURE A2 OMITTED]

REFERENCES

Arrazola, M., J. De Hevia, M. Risueno, and J. F. Sanz. "Returns to Education in Spain: Some Evidence on the Endogeneity of Schooling." Education Economics, 11(3), 2003, 293-304.

Barmby, T. A., M. G. Ercolani, and J. G. Treble. "Sickness Absence: An International Comparison." Economic Journal, 112, 2002, F315-31.

Bertrand, M., E. Duflo, and S. Mullainathan. "How Much Should We Trust Differences-in-Differences Estimates?" The Quarterly Journal of Economics, 119(1), 2004, 249-75.

Boyd-Barret, O., and P. O'Malley. Education Reform in Democratic Spain.

London: Routledge, 1995.

Bradley, S., C. Green, and G. D. Leeves. "Worker Absence and Shirking: Evidence from Matched Teacher-School Data." Labour Economics, 14(3), 2008, 319-34.

Cameron, A.C., and P. K. Trivedi. Microeconometrics: Methods and Applications. New York: Cambridge University Press, 2005.

Clotfelter, C. T., H. F. Ladd, and J. L. Vigdor. "Are Teacher Absences Worth Worrying About in the United States?" Education Finance and Policy, 4(2), 2009, 115-49.

Connolly, M. "Here Comes the Rain Again: Weather and the Intertemporal Substitution of Leisure." Journal of Labor Economics, 26(1), 2008, 73-100.

Duflo, E., R. Hanna, and S. Ryan. "Monitoring Works: Getting Teachers to Come to School." Tech. Rep. 11880, NBER Working Paper, 2005.

Hamermesh, D. C., C. K. Myers, and M. L. Pocock. "Cues for Timing and Coordination: Latitude, Letterman, and Longitude." Journal of Labor Economics, 26(2), 2008, 223-46.

Harmon, C., and I. Walker. "Estimates of the Economic Return to Schooling for the United Kingdom." American Economic Review, 85(5), 1995, 1278-86.

Hotchkiss, J. L., and O. Pavlova. "The Impact of 9/11 on Hours of Work and Labour Force Participation in the US." Applied Economics Letters, 16(10), 2009, 999-1003.

Jepsen, C., and S. Rivkin. "The Potential Tradeoff between Teacher Quality and Class Size." The Journal of Human Resources, 44(1), 2009, 223-50.

Long, J. S. Regression Models for Categorical and Limited Dependent Variables. Thousand Oaks, CA: Sage Publications, 1997.

Lozano, F. A. "The Flexibility of the Workweek in the United States: Evidence from the FIFA World Cup." Economic Inquiry, 49(2), 2011, 512-29.

MEC. "Ministerio de Educacion. Estadistica de las Ensenanzas no universitarias. Profesorado," 2010a.

--. "Ministerio de Educacion. Estadistica de las Ensenanzas no universitarias. Unidades escolares por ensenanza," 2010b.

Miller, R. T., R. J. Murnane, and J. B. Willett. "Do Teacher Absences Impact Student Achievement? Longitudinal Evidence from Urban School District." Tech. Rep. 13356, NBER Working Paper, 2007.

--. "Do Worker Absences Affect Productivity?" International Labour Review, 147(1), 2008, 71-89.

Pischke, J.-S., and T. von Wachter. "Zero Returns to Compulsory Schooling in Germany: Evidence and Interpretation." The Review of Economics and Statistics, 90(3), 2008, 592-98.

Pons, E., and M. T. Gonzalo. "Returns to Schooling in Spain: How Reliable Are Instrumental Variable Estimates?" Labour, 16(4), 2002, 747-70.

Voung, Q. H. "Likelihood Ratio Tests for Model Selection and Non-nested Hypotheses." Econometrica, 57, 1989, 307-33.

COLIN GREEN and MARIA NAVARRO PANIAGUA *

* The authors would like to thank Ian Walker, Alfredo Paloyo, Fernando Lozano, and seminar participants at Lancaster University, the University of Aberdeen, and the University of Cyprus for their helpful comments. The authors are also grateful to the editor Peter Arcidiacono and two anonymous referees. M.N.P. gratefully acknowledges financial support from the Spanish Ministry of Science and Innovation Postdoctoral Grant 2008-0583, FEDEA, and the CYCIT project ECO2008-06395-C0503/ECON.

Green: Economics Department, Lancaster University, Lancaster LAI 4YX, UK. Phone 0044 1524 92950, Fax 0044 1524 594244, E-mail c.p.green@lancaster.ac.uk

Navarro: Economics Department, Lancaster University, Lancaster LA1 4YX, UK; FEDEA, Madrid, Spain. Phone 0044 1524 594667, Fax 0044 1524 594244, E-mail m.navarropaniagua @ lancs.ac.uk

(1.) That is, teaching previously noncompulsory years of schooling that became mandatory.

(2.) To our knowledge generally, and in the particular case we examine, compulsory schooling changes were not introduced with an increase in either the supply of high school teachers or improvements in teacher salaries and/or conditions (Boyd-Barret and O'Malley 1995).

(3.) In Spanish, Ley de Ordenacion General del Sistema Educativo. 1990 (LOGSE).

(4.) Although students can stay in school until they are 18 (or 21 in the case of pupils with special education needs).

(5.) The SLFS has been demonstrated to have an internationally consistent definition of absence (Barmby, Ercolani, and Treble 2002).

(6.) We consider usual hours as synonymous with contractual hours. This is similar in spirit to the approach used in previous research such as Hamermesh, Myers, and Pocock (2008) and Lozano (2011).

(7.) CCAA stands for Autonomous Communities in Spanish. Spain is administratively divided into 17 regions and two autonomous cities. This division corresponds to the NUTS level 2 established by Eurostat for statistical purposes.

(8.) We estimate this for the education sector sample as there are insufficient observations if we use the teacher's only sample. These results are robust to alternative splits of the regions such as greater or less than 25% of students attending post-compulsory secondary education prior to the reform.

(9.) Specifically they are biased downwards causing overestimation of significance levels.

TABLE 1
Descriptive Statistics 1996-2004

 High/
 Primary
 All Education School
 Workers Workers Teachers

Total observations 988,329 63,811 49,711
Absence due to 18,349 927 789
 illness
 Excluding summer quarter

Total observations 742,458 48,433 37,771
Absence due to 14,012 805 694
 illness

Source: SLFS, authors' own calculations.

TABLE 2
Changes in Compulsory Schooling Laws and Teacher Absenteeism, All
Workers, SLFS 1996-2004

 ZIP (a) ZINB (b)

RoSLA x HSTeacher 0.139 ** (0.069) 0.140** (0.071)
 [1.150] ** [1.150] **
HSTeacher -0.202 *** (0.063) -0.203 *** (0.065)
 [0.8171 *** [0.817] ***
RoSLA -0.028 * (0.014) -0.029 ** (0.015)
 [0.972] * [0.972] **
Age 0.001 (0.001) 0.001 (0.001)
 [1.0011 [1.001]
[Age.sup.2] 0.000 (0.001) 0.000 (0.001)
 [1.000] [1.000]
Female 0.003 (0.003) 0.003 (0.003)
 [1.0031 [1.003]
Married 0.004 (0.003) 0.004 (0.003)
 [1.004] [1.0041
Secondary education -0.009 (0.007) -0.009 (0.007)
 [0.991] [0.9911
Higher education -0.004 (0.008) -0.004 (0.008)
 [0.9961 [0.9961
Public sector -0.027 *** (0.006) -0.027 *** (0.006)
 [0.974] *** [0.974] ***
Temporary contract -0.006 (0.005) -0.006 (0.006)
 [0.994] [0.994]
Establishment size
 Workers 0-5 -0.022 *** (0.005) -0.022 *** (0.005)
 [0.978] *** [0.978] ***
 Workers 6-10 -0.002 (0.005) -0.002 (0.005)
 [0.998] [0.9981
 Workers 11-19 0.004 (0.004) 0.004 (0.004)
 [1.004] [1.0041
 Workers 20-49 -0.003 (0.005) -0.004 (0.005)
 [0.997] [0.9961
Observations 988,329 988,329
Vuong test 1,222.14 -2.60
p value 0 0.9953
Likelihood-ratio test 1.6 x [10.sup.6]
p value .0000

Notes: Controls for industry, workers' occupation, region, year, and
quarter are included but not reported. Robust standard errors
clustered at the regional level are in parentheses. IRR are in
brackets.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1% levels, respectively.

TABLE 3
Changes in Compulsory Schooling Laws and Teacher Absenteeism,
Alternative Subsamples

 ZIP (a)

 High/Primary
 Education Workers School Teachers

RoSLA x HSTeacher 0.162 *** (0.053) 0.180 ** (0.077)
 [1.176] *** [1.198] **
HSTeacher -0.163 *** (0.055) -0.162 ** (0.072)
 [0.849] *** [0.850]
RoSLA -0.029 (0.036) -0.145 (0.134)
 [0.9711 [0.865]
Observations 63,811 49,711
Vuong test 331.68 288.96
p value .0000 .0000
Likelihood-ratio test
p value

 ZINB (b)

 High/Primary
 Education Workers School Teachers

RoSLA x HSTeacher 0.172 *** (0.058) 0.189 ** (0.081)
 [1.187] *** [1.208] **
HSTeacher -0.170 *** (0.059) -0.168 ** (0.074)
 [0.843] *** [0.846] **
RoSLA -0.032 (0.039) -0.155 (0.136)
 [0.968] [0.857]
Observations 63,811 49,711
Vuong test 3.04 2.89
p value .0012 .0019
Likelihood-ratio test 8.0 x [10.sup.4] 6.5 x [10.sup.4]
p value .0000 .0000

Notes: All other controls are as in Table 2. Robust standard errors
clustered at the regional level are in parentheses. IRR are in
brackets.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1% levels, respectively.

TABLE 4
Changes in Compulsory Schooling Laws and Teacher Absenteeism, by
Previous RoSLA Participation Rates in Post-Compulsory Secondary
Education

 ZIP (a)

 <20% [greater than or
 equal to] 20%

RoSLA x HSTeacher 0.359 *** (0.078) 0.073 (0.070)
 [1.432] *** [1.076]
HSTeacher -0.342 *** (0.086) -0.084 (0.074)
 [0.711] *** [0.919]
RoSLA -0.094 (0.109) -0.015 (0.048)
 [0.910] [0.985]
Observations 20,925 42,886
Vuong test 161.35 266.94
p value .0000 .0000
Likelihood-ratio test
p value

 ZINB (b)

 <20% [greater than or
 equal to] 20%

RoSLA x HSTeacher 0.397 *** (0.087) 0.080 (0.073)
 [1.487] *** [1.084]
HSTeacher -0.374 *** (0.095) -0.089 (0.078)
 [0.688]*** [0.915]
RoSLA -0.125 (0.111) -0.018 (0.054)
 [0.883] [0.9821
Observations 20,925 42,886
Vuong test 1.84 2.46
p value .0329 .0070
Likelihood-ratio test 2.4 x [10.sup.4] 5.4 x [10.sup.4]
p value .0000 .0000

Notes: All other controls are as in Table 2. Robust standard errors
clustered at the regional level are in parentheses. IRR are in
brackets. <20% and [greater than or equal to] 20% columns comprise 7
and 11 regions, respectively.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1% levels, respectively.

TABLE 5
Changes in Compulsory Schooling Laws and Teacher Absenteeism,
Implementation Effects and Policy Longevity

 Zip (a)

 High/Primary
 Education Workers School Teachers

 Excluding the reform period (1 year)

RoSLA x HSTeacher 0.182 *** (0.050) 0.182 ** (0.075)
 [1.199] *** [1.200] **
Observations 61,268 47,783

 Excluding the reform period (2 years)

RoSLA x HSTeacher 0.184 *** (0.049) 0.181 ** (0.076)
 [1.202] *** [1.199] **

Observations 52,748 41,064

 ZINB (b)

 High/Primary
 Education Workers School Teachers

 Excluding the reform period (1 year)

RoSLA x HSTeacher 0.191 *** (0.055) 0.192 ** (0.079)
 [1.210] *** [1.211] **
Observations 61,268 47,783

 Excluding the reform period (2 years)

RoSLA x HSTeacher 0.194 *** (0.055) 0.193 ** (0.080)
 [1.214] *** [1.212] **

Observations 52,748 41,064

Notes: All other controls as in Table 2. Robust standard errors
clustered at the regional level are in parentheses. IRR are in
brackets.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1% levels, respectively.

TABLE 6
Changes in Compulsory Schooling Laws and Teacher Absenteeism,
Excluding Summer Quarter

 ZIP (a)

 High/Primary
 Education Workers School Teachers

RoSLA x HSTeacher 0.145 *** (0.044) 0.182 *** (0.070)
 [1.156] *** [1.200] ***
HSTeacher -0.140 *** (0.047) -0.163 ** (0.065)
 [0.870] *** [0.850]**
RoSLA -0.019 (0.046) -0.144 (0.132)
 [0.982] [0.8661
Observations 48,433 37.771
Vuong test 315.20 289.98
p value .0000 .0000
Likelihood-ratio test
p value

 ZINB (b)

 High/Primary
 Education Workers School Teachers

RoSLA x HSTeacher 0.151 *** (0.048) 0.188 ** (0.076)
 [1.163] *** [1.207] **
HSTeacher -0.143 *** (0.050) -0.166 ** (0.070
 [0.867]*** [0.847] **
RoSLA -0.020 (0.047) -0.153 (0.138)
 [0.980] [0.858]
Observations 48,433 37,771
Vuong test 3.10 3.05
p value .0010 .0011
Likelihood-ratio test 7.5 x [10.sup.4] 6.1 x [10.sup.4]
p value .0000 .0000

Notes: All other controls as in Table 2. Robust standard errors
clustered at the regional level are in parentheses. IRR are in
brackets.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1% levels, respectively.

TABLE 7
Changes in Compulsory Schooling Laws and Public Sector (Non-Education
Worker) Absenteeism

 All Periods

 ZIP (a) ZINB (b)

RoSLA x Treatment 0.056 (0.046) 0.053 (0.048)
 [1.0571 [1.0551
Treatment -0.082 * (0.046) -0.080* (0.048)
 [0.921] * [0.923] *
RoSLA -0.052 ** (0.026) -0.051 * (0.027)
 [0.949] ** [0.951] *
Observations 923,586 923,586
Vuong test 1,186.29 -2.66
p value 0.0000 0.9961
Likelihood-ratio test 1.5 x [10.sup.6]
p value .0000

 Excluding Summer Quarter

 ZIP (a) ZINB (b)

RoSLA x Treatment 0.045 (0.061) 0.043 (0.064)
 [1.0461 [1.0441
Treatment -0.069 (0.061) -0.066 (0.064)
 [0.934] [0.936]
RoSLA -0.037 (0.031) -0.035 (0.033)
 [0.964] [0.966]
Observations 693,317 693,317
Vuong test 1,022.04 -2.03
p value 0.0000 0.9787
Likelihood-ratio test 1.3 x [10.sup.6]
p value .0000

Notes: All other controls as in Table 2. Robust standard errors
clustered at the regional level are in parentheses. IRR are in
brackets.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1% levels, respectively.

TABLE 8
Changes in Compulsory Schooling Laws and
Teacher Absenteeism Collapsed, Excluding
Summer Quarter

 Zip (a) ZINB (b)
 Education Education
 Workers Workers

RoSLAx 0.496 *** (0.173) 0.653 *** (0.171)
 HSTeacher [1.6431 *** [1.9211 ***
HSTeacher -0.183 (0.197) -0.314 (0.224)
 [0.8331 [0.7301
RoSLA -0.399 ** (0.200) -0.332 (0.213)
 [0.6711 ** [0.7181
Observations 12,347 12,347
Vuong test 117.24 6.09
p value .0000 .0000
Likelihood-ratio test 2.0 x [10.sup.5]
p value .0000

Notes: All other controls as in Table 2. Robust standard
errors clustered at the regional level are in parentheses. IRR
are in brackets.

(a) Zero inflated Poisson.

(b) Zero inflated negative binomial.

*, **, and *** indicate statistical significance at the 10%,
the %, and the 1% levels, respectively.

TABLE 9
Quantifying the Effect of RoSLA on Extensive and Intensive Margins of
Absence

 All Periods

 Education High/Primary
 Workers School Teachers

Panel A. Sickness absence (incidence), probit marginal effects

RoSLA x HSTeacher 0.016 0.011
 (0.009) * (0.012)
HSTeacher -0.010 -0.009
 (0.008) (0.012)
RoSLA 0.005 0.011
 (0.012) (0.012)
Observations 21,782 17,420

Panel B. Minutes of absence (OLS)

RoSLA x HSTeacher 307.570 336.969
 (93.672) *** (140.689) **
HSTeacher -302.293 -299.363
 (94.654) *** (129.098) **
RoSLA -53.577 -258.431
 (87.580) (265.185)
Observations 927 789

 Excluding Summer Quarter

 Education High/Primary
 Workers School Teachers

Panel A. Sickness absence (incidence), probit marginal effects

RoSLA x HSTeacher 0.024 0.013
 (0.011) ** (0.019)
HSTeacher -0.012 -0.012
 (0.011) (0.020)
RoSLA 0.010 0.025
 (0.016) (0.020)
Observations 12,172 9,561

Panel B. Minutes of absence (OLS)

RoSLA x HSTeacher 279.665 351.523
 (80.259) *** (127.227) **
HSTeacher -259.205 -308.671
 (83.303) *** (117.256) **
RoSLA -31.578 -266.458
 (105.083) (264.981)
Observations 805 694

Notes: All other controls as in Table 2. Robust standard errors
clustered at the regional level are in parentheses.

*, **, and *** indicate statistical significance at the 10%, the 5%,
and the 1 % levels, respectively.

FIGURE 1 Minutes of Absence for Full-Time Workers Before and After
the Reform

All workers

Minutes Absence per week

 Worker High School Teacher Primary School Teacher

Pre-RosLA 43 20 34
RoSLA 42 30 35

Workers with non-zero absence

Minutes Absence per week

 Worker High School Teacher Primary School Teacher

Pre-RosLA 2258 1790 1961
RoSLA 2249 1959 1969

Note: Table made from bar graph.