Add Journal to My Library
Journal of Economic Geography
, Volume Advance Article – Mar 13, 2018

30 pages

/lp/ou_press/commute-costs-and-labor-supply-evidence-from-a-satellite-campus-mjfwLTrOtc

- Publisher
- Oxford University Press
- Copyright
- © The Author(s) (2018). Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com
- ISSN
- 1468-2702
- eISSN
- 1468-2710
- D.O.I.
- 10.1093/jeg/lby006
- Publisher site
- See Article on Publisher Site

Abstract Using the transition of undergraduate teaching from an urban to suburban campus and an exogenous increase in faculty’s commute time, we estimate the causal effect of commute costs on labor supply. Difference-in-differences estimates using individual commute time changes imply that the 1.0–1.5-h round-trip increase in commute time reduces annual teaching hours by 22 (8.4%). Substitution to alternative work activities is minor but class sizes increase and research output decreases. This suggests that work time is highly responsive to commute time for workers with flexibility and has important ramifications for transport policy, city growth and business strategies. 1. Introduction Because commute costs are variable with respect to work days but fixed within a work day, theoretically a longer commute can either increase or decrease total work time depending on the relative changes in days worked and daily hours (Cogan, 1981; Parry and Bento, 2001).1 In what direction, and how much, commute costs affect work time is therefore an empirical question. Because controlling for endogeneity and selection bias in estimation are difficult, extant answers are limited and biased downward. Using a unique empirical setting, we estimate the causal effect without selection bias and find a large work time reduction. A significant labor supply response to commute costs has important ramifications for government policy, city growth and business strategies.2 Cost–benefit analyses of transportation infrastructure investments and traffic congestion policies should consider not only the opportunity cost of commute time changes but also the accompanying change in labor supply and, therefore, output. The negative relationship between congestion and long-run employment growth (Hymel, 2009; Duranton and Turner, 2012), the presence of coordination and knowledge spillovers in cities (Moretti, 2004), and the longer commute times and distances caused by urban sprawl (Glaeser and Kahn, 2001) imply that commute time’s influence on labor supply plays a role in city growth. For firms, understanding the causal effect of commute costs on labor supply helps them design policies to attract talent and influence their work time and productivity. Quantifying the effect of commute costs on labor supply is difficult because they are endogenous and suitable instruments are scarce.3 Workers consider commute costs when choosing residences, job locations, and commute modes and firms consider them when choosing wages and locations.4 Workers with high commute-cost sensitivity are likely to choose residence–job combinations with short commutes, whereas workers with low sensitivity are likely to tolerate those with longer commutes. Failing to correct for endogeneity will understate commute costs’ effect on work time. Measuring commute costs is also difficult. Commute costs can include time, monetary costs and disutility and even time and distance are usually measured imprecisely.5 Most extant evidence on this question is either indirect or subject to endogeneity. Gibbons and Machin (2006) state that there is no direct empirical evidence of commute time’s causal effect on labor supply. The only subsequent papers we know of that deal with the endogeneity issue are Gutiérrez-i-Puigarnau and van Ommeren (2010a), Mulalic et al. (2014) and Gershenson (2013). The first two use workplace relocations and exclude workers who change residences to maintain exogeneity. While this solves the in-sample endogeneity problem, it understates the out-of-sample effects because workers who change residences have high commute-cost sensitivity. The commute costs measures may also involve error because transport mode is unobserved in the first paper and commute time is self-reported in the second. The effects for the in-sample groups are small: 15 fewer work minutes per week from an extra 40 km in daily round-trip commute distance in the first paper and an insignificant effect of commute time in the second. Gershenson (2013) uses random daily assignments of substitute school teachers to overcome the endogeneity problem and estimates commute time’s effect on teachers’ daily job acceptance probabilities rather than their work time. Our causal estimates are based on the addition of a suburban satellite campus to a main urban campus at a typical, well-established Chinese university. For classes taught at the satellite campus, commute time increases exogenously since faculty has a strong financial incentive not to quit their jobs6 or move their residences to the satellite campus and virtually none do. Moreover, the increased time and distance are known and homogeneous across teachers7 since virtually all faculty live at or near the main campus and ride a free university shuttle bus. Faculty members choose their teaching time freely within an internal labor market and are paid on a piece-rate system subject to a linear wage8 allowing us to measure the market response of work time. The transition sequence of undergraduate teaching to the satellite campus allows us to use teacher-level variation to identify the causal effect of commute costs on teaching time. The satellite campus opens in academic year 20049 (a ‘year’ will refer to an ‘academic year’ unless otherwise noted) but undergraduate students transition one class level per year until all four levels are taught there in 2007. This incremental transition imposes different levels of commute costs on different teachers in different years depending on their course schedule. Teaching a freshmen course10 imposes commute costs beginning in the transition’s first year, a sophomore course beginning in the second year, a junior course beginning in the third year and a senior course beginning in the fourth year. This allows a difference-in-differences (DD) analysis comparing work time effects for teachers with differential changes in commute time. This is immune to confounding factors not correlated with individual-level commute time. In particular, university-wide changes such as wages, student body size, faculty size, class size and teaching load are differenced out. We provide evidence that possible individual-level confounders do not materially affect our estimates—in particular class size. Because the transition sequence was known only shortly before it began faculty could only avoid these commute costs through very costly efforts of developing new courses rapidly. The 1.0–1.5 h increase in commute time per teaching day reduces annual undergraduate teaching time by 21.9 ‘class hours’ (a ‘class hour’ is the amount of time a faculty member spends in the classroom to receive one ‘hour’ of pay) per teacher after the full transition or 8.4% of the pre-transition average of 260.0 ‘class hours’. The elasticity of work time with respect to commute time is −0.044 and teachers value commute time at 18–27% of their hourly wage. Since wages are linear, the elasticity of income with respect to commute time is also −0.044. Commute costs vary with days worked but are fixed with respect to daily hours conditional on working that day. Consistent with this, we find that the full transition reduces a teacher’s annual undergraduate teaching days by 18 and increases daily undergraduate ‘class hours’ by 0.50. This means confounding factors are even less likely. They must decrease days worked but increase daily hours and operate at the individual faculty level—affecting those with longer commute times by more than those with shorter. Faculty may substitute to other work activities to offset their reduced undergraduate teaching time. Possibilities include graduate teaching, research and consulting. Since the location of these is unaffected, they become relatively more attractive once undergraduate teaching transitions. However, time spent on these activities might decrease if increased time and fatigue from commuting crowds them out. The effect on graduate teaching is insignificant. Research output falls with the number of academic research papers published in top journals decreasing significantly. We are unable to observe consulting time but the drop in undergraduate teaching time for assistant professors, who rarely consult, is similar to the effect for more senior faculty. Unless research productivity fell dramatically, time diverted to other work activities is minimal and the total decrease in work time approximately equals the decreased undergraduate teaching time. Since we do not observe preparation time outside the classroom, we cannot quantify the effect on work time inclusive of it.11 If preparation time per ‘class hour’ remained unchanged after the transition then our estimates understate the decrease in total work time by the ratio of preparation time to in-class plus preparation time. If preparation time changed because of the transition then we need to worry whether this confounds our estimates. In particular, since the university accommodated the decreased teaching time primarily by increasing class sizes, preparation time outside class may have increased causing faculty to further reduce their ‘class hours’. However, if increased preparation time is uncorrelated with individual changes in commute time then the DD estimates are unbiased. Class size increases are correlated with commute time changes at the individual level but the magnitude is small. As we examine a specific type of worker, university professors, how well might our results extend to other types of workers? Equilibrium commute times and distances generally increase with education (Groot et al., 2012) suggesting that our estimates may understate the effects more generally. To the contrary, for workers who have less flexibility on the intensive margin of labor supply, our estimates may overstate the effects. However, a nontrivial fraction of workers have significant discretion over their work time and this fraction is expected to increase.12 The proportion of ‘knowledge workers’, who generally have flexible schedules, is projected to increase over time (Moretti, 2012) and smartphone applications are allowing increasing numbers of people to work flexibly.13 At least 19% of employed urban workers in China have some amount of flexibility over their work time.14 Russo et al. (2012) offer some evidence to determine which of these opposing effects, education or flexibility, dominates. They estimate a long-run marginal willingness to pay (WTP) to avoid commuting for faculty at a Dutch university and find it equals approximately 35% of their wage.15 This is somewhat lower than the 50% midpoint estimate for workers more broadly (Gibbons and Machin, 2006) which suggests that our estimates underestimate the effects for other types of workers. The next two sections provide institutional and theoretical background for our estimates. We then provide the econometric model, data and results before concluding. 2. Institutional details We examine commute costs created from transitioning the location of undergraduate teaching at a well-established Chinese university.16 The university is a research institution serving both undergraduate and graduate students and is a highly ranked specialized (in particular academic disciplines) rather than general university. Capacity-constrained and faced with acquiring very expensive land adjacent to the original campus in the city center, the university decided to add a satellite campus 20 km away in a suburban area. Planning began in calendar year 2000 with a search for land and the university signed a contract with the city government to buy a parcel the next calendar year. Bidding for the campus design was held in calendar year 2002 and later that year a national newspaper announced that incoming freshman would live and be taught at the satellite campus beginning in 2003, later postponed to 2004. The timing of this announcement is critical because it means that prior to late 2002 faculty was aware that a campus was being built but unaware of the transition timing. This is after academic year 2002 had begun and teaching schedules had been finalized for that year. Therefore, any faculty efforts to change their teaching schedule away from teaching freshman classes (to delay commuting) began in the academic year 2003 at the earliest. The school held a groundbreaking ceremony in early calendar year 2003 and, in academic year 2004, the entering freshmen lived and took courses at the satellite campus while higher class levels remained at the main campus. In 2005, the entering class again lived and took courses at the satellite campus so that freshmen and sophomores took courses at the satellite campus while juniors and seniors remained at the main campus. In 2006, only seniors remained at the main campus while the other class levels lived and took courses at the satellite campus. From 2007 onward, all four class levels lived and took courses at the satellite campus. Our data spans the years 2000–2009. This provides 4 years before the transition and 3 years in which all four undergraduate class levels took courses at the satellite campus. Graduate courses remained at the main campus during the undergraduate transition. Entering Master’s students began taking courses at the satellite campus in 2008. Since they generally study for 2 years, one-half of them were at the satellite campus in 2008 and all of them were at the satellite campus in 2009.17 Entering Ph.D. students began taking courses at the satellite campus in 2009. Since most Ph.D. students study for 3 years, approximately one-third took courses at the satellite campus in 2009. After the end of our sample period, Ph.D. students finished transitioning to the satellite campus and executive MBA, professional education, continuing education, some business classes, some research institutes and some administrative offices remained at the main campus. Almost all teachers resided at or near the main campus during the sample period because the university continued to provide subsidized housing in residential neighborhoods adjacent to the main campus and did not complete construction of similar subsidized housing at the satellite campus until after 2010.18 The university provided a convenient shuttle bus between the two campuses by which virtually all faculty commuted. The shuttle was free so we estimate the effect of increased commute time but not monetary costs. The shuttle trip takes about 30 min one way plus up to 15 min of walking and waiting on each end. Since the time required depends on random variation in weather, traffic and wait times, we assume that commute time increased 1.0–1.5-h round-trip per commute day (a ‘teaching day’ refers to a day on which a faculty member teaches at least one class regardless of location while a ‘commute day’ refers to a ‘teaching day’ at the satellite campus). Our primary data consists of the university’s complete undergraduate class schedule from 2000 to 2009. For each class, we know its course title, academic semester, teacher, class level (freshman, sophomore, junior, senior and other), number of students (class size), day and time of meeting, weekly ‘class hours’ and number of weeks. We can identify class level because in China most undergraduate courses are taught to a single class level. This is important since it allows us to determine which classes were taught at which campus during the transition.19 Our primary measure of labor supply is a ‘class hour’ (50 min at the main campus and 45 min at the satellite campus). We do not observe time spent outside of class preparing, grading and responding to students. If outside time remains the same per ‘class hour’ after versus before the transition, our estimates can be scaled up by the appropriate multiplier to obtain total hours from ‘class hours’ and our estimates are not biased. If preparation time increased because of the transition (e.g. due to the accompanying increase in class sizes) our estimates would be biased. We present evidence when we discuss our results that changes in class size are largely orthogonal to changes in individual commute time so that bias in our DD estimates is minimal. Teachers allocate their time among five major activities: undergraduate teaching, graduate teaching, research, consulting and leisure. A teacher’s total annual compensation can be represented as F+BTR+wUTU+wGTG. F is a fixed payment based on seniority, position and administrative duties and is primarily based on a nationwide standard. B is an annual bonus paid for research publications where TR is time spent on research and we assume that there are diminishing or constant returns to research ( B0=0, B'>0 and B''≤0). Research also provides nonpecuniary benefits such as prestige, personal satisfaction and future career advancement and we can think of B as including these effects as well.20 The last two components are the linear payments for teaching where wU and wG are ‘hourly’ wages for undergraduate and graduate teaching and TU and TG are annual ‘class hours’ taught for each. The ‘hourly’ undergraduate wage increased over time: RMB 20 in 2001 and 2002, RMB 40 in 2003 and 2004, RMB 60 from 2005 to 2007 and RMB 90 from 2008 onward.21 Domestic faculty were compensated for a graduate ‘class hour’ at 1.5 times the rate for undergraduate ‘class hours’ and those with a foreign Ph.D. (from a non-mainland China university) were compensated at the same rate for both. F, B, wU and wG are otherwise common to all faculty and do not change within academic years. The minimum annual teaching load was 240 ‘class hours’ from 2001 to 2004 and 225 from 2005 onward.22 Teachers are paid for ‘class hours’ both used to satisfy their teaching load and those above. Teachers may fulfill this minimum through other activities (and get paid for them) that we do not observe. These include supervising graduate theses, administrative tasks, and supervising student internships and study trips.23 Therefore, many faculty have fewer ‘class hours’ than the minimum. Although we do not observe these activities, we show that our results are robust to them. Unlike at many American universities in which department heads or deans have more control, faculty at most Chinese universities, including the one we study, have great discretion in choosing their teaching time. Subject to the wage, a teacher chooses teaching hours to maximize their utility. The process is as follows: each faculty member submits their chosen courses to the department staff which figures out course scheduling. The schedule is submitted to a university-wide administrative office that assigns classrooms. At this university, faculty could also choose their teaching time in fine increments for three reasons. First, course credits range from one to six. A course credit corresponds to one ‘class hour’ per week over a 16-week semester. Thus, courses allow annual teaching time to be chosen in increments as small as 16 ‘class hours’. Second, co-teaching arrangements further reduce this increment. Co-teaching with one person allows annual teaching time to be chosen in increments of eight ‘class hours’ and co-teaching with more than one faculty member is allowed and common. Third, seminar-style classes, which meet once, allow faculty members to receive teaching credit in increments as small as two ‘class hours’. The figure in Online Appendix A illustrates this flexibility. It shows the distribution of annual undergraduate ‘class hours’ across all faculty members from 2000 to 2009. Although the distribution exhibits spikes at some multiples of course credits, it exhibits significant dispersion. Given faculty’s flexibility to adjust their ‘class hours’, how could the university accommodate a decrease in teacher ‘class hours’ while still providing sufficient student ‘class hours’24 for graduation and distribution requirements? The university could adjust along two dimensions: class sizes and faculty size. We provide rough numbers on the adjustment after we discuss our results. The figure in Online Appendix A also shows many faculty with fewer ‘class hours’ than the minimum teaching load, consistent with it not binding due to availability of substitute activities. We return to this in our results. Faculty members also had significant flexibility in allocating their ‘class hours’ across days of the week. Classes of three credits or less met once per week and a faculty member could choose the day of the week. Longer classes met twice per week (e.g. a four-credit course meets twice per week for two ‘class hours’ each day) and the faculty member could choose any two non-contiguous days. The overall constraint of not scheduling too many courses for a particular class level on the same day might require ‘horse-trading’ of days among colleagues.25 Faculty size, student enrollment, graduation requirements and class sizes could affect university-level teaching demand. However, contractually the university could only require teachers to fulfill their minimum teaching load. Above this, the university could influence demand in the internal labor market only through the wage it offered. Department heads may attempt to pressure faculty to teach more or less. Pressure applied to all teachers will be differenced out in our DD estimates. Individualized pressure would bias our results away from zero only if department heads systematically exerted more pressure on faculty facing longer commute costs to teach less. The opposite seems more likely. If a faculty member facing a commute to teach freshman at the satellite campus were convinced to teach less the department head would have to then convince another teacher unfamiliar with the course to commute and teach it. 3. Theoretical background We model the effect of increased commute time on daily ‘class hours’, annual teaching days and annual ‘class hours’. For manageability, we first consider a model with no graduate teaching or research TG=TR=0 so that all work time is subject to commute costs and all teaching days are commute days. We reintroduce these in an alternative model in Online Appendix C. Because additional commute time increases fixed costs per teaching day, faculty will concentrate more daily ‘class hours’ in fewer annual teaching days. Total teaching time could increase or decrease. We show this using a modified version of the model in Gutiérrez-i-Puigarnau and van Ommeren (2009) which generalizes a labor supply model with commute costs to allow for the choice of days worked and daily hours. We adapt their model to our setting in two main ways. Their model allows for a concave wage function due to declining marginal productivity. We instead use a linear wage function and assume that a convex effort cost diminishes the value of leisure. We also exclude monetary commute costs consistent with the university’s free shuttle service. The two models’ implications are qualitatively similar. A teacher’s annual utility is v=VC,L where C is annual consumption, L is annual leisure time and V is differentiable with VL>0,VC>0,VLL<0,VCC<0, and VCL>0. Without graduate teaching and research, annual compensation is F+wUDH where annual undergraduate ‘class hours’ TU is decomposed into annual days D and daily ‘class hours’ H. A teacher’s annual budget constraint is C=Y+F+wUDH, where Y is annual nonlabor income. Annual time is divided between undergraduate teaching and leisure and each teaching day requires round-trip commute time of t.26 Daily ‘class hours’ require effort that decreases utility from daily leisure by eH with e'H>0 and e”’H>0 denominated in leisure hours. The disutility can be interpreted as diminishing the quality of each leisure hour or additional time spent resting to recover from the fatigue of commuting. Although we have stated annual teaching days, daily ‘class hours’ and annual ‘class hours’ in terms of in-class time, it is simple to allow for exogenous fixed preparation time outside class. If total work hours per ‘class hour’ is H'=ρH with ρ>1, redefine the effective hourly wage as wU'=wU/ρ and all the results go through. A teacher’s annual time constraint is T¯=L+DH+t+eH where T¯ is total annual hours. Substituting the budget and time constraints:27 v=V(Y+F+wUDH,T¯−D(H+t+e(H))). (1) The two first-order conditions are: ∂v/∂H=VCwUD−VLD(1+e'(H))=0, (2) and ∂v/∂D=VCwUH−VL(H+t+e(H))=0. (3) Equation (2) says that the marginal utility of consumption from an extra daily ‘class hour’ equals the foregone marginal utility of daily leisure including the effect of fatigue. Equation (3) says the same from working an extra teaching day during the year. Combining these two, the optimally chosen daily ‘class hours’ fulfills e'(H)=t+e(H)H. (4) The teacher equates the marginal disutility of effort to the average daily disutility of working (including commute time and effort). The teacher smoothes daily ‘class hours’ across days to avoid escalating the costs from working very long days (e.g. it is better to have two 5-h days than one 10-h day).28 If we totally differentiate Equation (4) letting daily ‘class hours’ adjust to a change in commute time, it follows that an increase in daily commute time increases daily ‘class hours’ dHdt=1e”(H)H>0. (5) Given a longer daily commute, teachers work more ‘class hours’ once at the satellite campus so as to avoid additional trips on other days. In Online Appendix B, we show that increased commute time decreases annual teaching days. Teachers concentrate their teaching in fewer days to avoid the extra commute time incurred each teaching day. Thus, increased commute time increases daily ‘class hours’ but decreases annual teaching days. In Online Appendix B, we also show that increased commute time could increase or decrease annual ‘class hours’ DH, which happens depends in particular on the curvature of the effort costs. If effort costs do not increase too rapidly with daily ‘class hours’ then increased commute time may increase annual ‘class hours’. In Online Appendix C, we modify the model to consider two work activities—one affected by commute time (undergraduate teaching) and the other not. The other activity could either be paid according to a wage linear in hours worked (as with graduate teaching) or increase a teacher’s annual bonus according to a weakly concave function of hours worked (as with research). To simplify the analysis, we collapse the separate choices of annual teaching days and daily ‘class hours’ into a single choice of annual ‘class hours’ for each activity. The model shows that time spent on the other activity could increase or decrease with commute time when undergraduate teaching time decreases. Faculty may substitute toward these activities since they do not require commuting; however, increased commute time may crowd them out. 4. Data Our primary sample is the university’s complete undergraduate course schedule provided by the university’s Undergraduate Education Administrative Office. We supplement this data with rank, gender and PhD source for each teacher from the university’s website. A teacher is included in this sample as long as they taught at least one undergraduate course. If a teacher taught only graduate level courses or no classes at all they are not included.29 For each class, we use weekly ‘class hours’ and number of weeks taught to compute total ‘class hours’. For co-taught classes, we divide total ‘class hours’ by the number of co-teachers to obtain total ‘class hours’ for each teacher. We then aggregate across all classes for a teacher in a year to obtain annual ‘class hours’ for each teacher-year observation. To determine the number of teaching days for each teacher, we use the days of week for each class they teach to identify all the dates on which their classes are taught during the semester.30 We then identify any overlap in these dates to obtain unique teaching dates for each semester. Aggregating across the two semesters, we obtain annual teaching days for each teacher-year observation. Finally, we compute daily ‘class hours’ (conditional on teaching that day) for each teacher-year observation by dividing annual ‘class hours’ by annual teaching days. A second sample consists of graduate course information. Since we were unable to obtain complete graduate course data from the university, we downloaded it from its graduate school website. As a result, we do not observe day and time of meeting or class size but we do observe course title, academic semester, teacher, weekly ‘class hours’ and number of weeks. A teacher is included in this sample if they taught at least one graduate course. Teachers who taught only undergraduate courses or no classes at all are not in this sample. A third sample consists of data on faculty research output from the university’s Research Support Office website. We observe author’s name, journal name and publication date. Because it is important in both determining faculty salaries and promotions and establishing the university’s reputation, we believe the data are accurate and comprehensive. During our sample period, China’s Ministry of Education attributes research output only to the first author’s affiliation. The university applied this same criterion in evaluating faculty so we count a paper only toward the first author. We designate papers as appearing in either ‘top’ or ‘non-top’ journals31 and compute annual publications per teacher. The table in Online Appendix D summarizes how various undergraduate teaching variables evolve over time. Annual teacher ‘class hours’ summed across all teachers in Column 7 hint at the effect that our formal tests reveal. Prior to the transition, teacher ‘class hours’ increase each year rising from 96.5 in 2000 to 144.0 in 2003 before dropping to 142.9 in 2004 when the transition begins. Columns 1 and 2 show that class-level-specific courses comprise a large and stable fraction of all classes from 2000 to 2008 (ranging from 82.0% to 88.9%).32 The table in Online Appendix D also hints at how the university responded to the decreased teaching time: by increasing the average number of students per class post-transition (Column 10)—average class size is 54.6 in 2003 and jumps to 57.0 in 2004 when the transition begins and rises to 67.0 by the time the transition ends in 2007. 5. Identification strategy Our identification strategy is similar to Duflo (2001) who implements DD estimation in which treatment intensity varies by geographic region. In our case, treatment intensity varies across faculty and increases differentially as the campus transition progresses. We utilize individual commute-cost variation by using the fact that class levels transition one at a time each year to the satellite campus. Work time should be disproportionately affected for those who teach class levels that have transitioned relative to those who teach levels that have not. We assume that a teacher’s exposure to commute costs during the transition period is proportional to the commute days a teacher would incur based on their teaching schedule in 2003—just prior to the transition. We call this the expected commute days for teacher i in year tCDit and calculate it as the annual number of days a teacher would commute during and after the transition if their 2003 schedule had persisted. For our DD estimation, we exclude ‘other’ courses in our calculation of expected commute days because we cannot infer their location but include them in calculating work time. This will understate expected commute days and bias against finding an effect.33 To illustrate, consider a teacher who taught 16 weeks in 2003 and taught a freshman and a junior class on Monday, a freshman and a sophomore class on Tuesday, no classes on Wednesday, two senior classes on Thursday, and a sophomore and a junior class on Friday. Prior to 2004, expected commute days is zero. In 2004, when freshmen transition to the satellite campus, expected commute days equals the number of unique dates that a teacher taught freshman classes in 2003: 32 in this example (two commute days weekly for 16 weeks). In 2005, when freshmen and sophomores are taught at the satellite campus, it equals the number of unique dates that a teacher taught freshman or sophomore classes or both in 2003. In this example, it would equal 48 (three commute days weekly for 16 weeks). Expected commute days in 2006 equals the number of unique dates that a teacher taught freshman, sophomore, or junior classes or some combination. In the example, it would remain 48 because there are no dates on which junior classes are taught that freshman or sophomore classes are not. In 2007 and after, expected commute days equals unique dates on which any class level or some combination is taught: 64 in this example (4 days weekly for 16 weeks). We believe that a teacher’s 2003 schedule provides the best measure of commute costs. It is necessary to use a measure from the pretreatment period because once the transition begins teachers alter their schedules to avoid teaching classes that impose more commuting.34 We also want as current a measure as possible because teachers’ schedules may change over time for random reasons such as changes in student or teacher interests. This will not bias the estimates but it will lower their precision. This suggests using the last year before the transition. For the same reason, our main results use only the pre-transition and transition periods (2000–2007); however, we show that our results are robust to also including the post-transition period (2008 and 2009).35 Since the transition sequence was announced in late 2002, there was a brief window in which faculty could attempt to shift away from teaching freshman and sophomore courses which would transition first. If this occurs, this would be reflected in their 2003 teaching schedule and the causal effect would be biased toward zero. This behavior is endogenous because it affects the treatment intensity. It is separate from teachers attempting to consolidate their teaching into fewer teaching days once the transition begins—the causal effect we estimate. Such avoidance behavior is costly as it requires incurring the fixed costs of developing a new junior or senior course. If such avoidance behavior is significant, a teacher’s freshman and sophomore ‘class hours’ should decline in 2003 relative to 2002. On average, across 452 faculty teaching undergraduates in both 2002 and 2003, the fraction of freshman ‘class hours’ dropped by only 2.2% in 2003 and is insignificant (standard error of 1.9%).36 The fraction of sophomore ‘class hours’ actually increased by 2.0% although not significantly (standard error of 1.9%). This is consistent with little avoidance behavior occurring in 2003. Table 1 contains descriptive statistics for our three samples. We focus on 2000–2007 data since this is used in most of our estimates. Panel A shows data for the 513 faculty teaching undergraduates in 2003 (necessary to calculate expected commute days) and at least 1 year after 2003 (necessary to aid in identification). An observation is a teacher-year.37 Expected commute days increase from an average of 24.6 in 2004 to 75.4 in 2007 with the smallest increase in 2007 due to fewer senior courses being taught. Panel B summarizes the data for graduate teaching for the 275 faculty who taught undergraduate students in 2003 (necessary to compute expected commute days) and taught undergraduate (necessary to identify the teacher fixed effects) and graduate students in at least 1 year from 2004 to 2007. Expected commute days are lower for this group given their graduate teaching requirements. Panel C summarizes the research output data which spans 2001–2008 because we lag teaching measures in our research estimation. This includes the 521 teachers who taught undergraduates in 2002 and in at least 1 year from 2003 to 2007 (necessary to identify the teacher fixed effects). Table 1 Descriptive statistics Variable N Mean Std. Dev. Min Max Panel A: Undergraduate teaching 2000–2007; 513 teachers Annual undergraduate ‘class hours’ 3402 260.25 171.16 2.00 1121.00 Annual undergraduate teaching daysa 3399 75.36 39.07 3.00 196.00 Daily undergraduate ‘class hours’a 3399 3.34 1.41 0.21 10.00 Expected commute days—2004 490 24.60 36.91 0.00 162.00 Expected commute days—2005 438 52.62 44.45 0.00 172.00 Expected commute days—2006 414 72.15 42.12 0.00 176.00 Expected commute days—2007 387 75.43 41.27 0.00 195.00 Male 3402 0.58 0.49 0.00 1.00 Position—assistant professor 3402 0.50 0.50 0.00 1.00 Position—associate professor 3402 0.29 0.46 0.00 1.00 Position—full professor 3402 0.17 0.37 0.00 1.00 Panel B: Graduate teaching 2000—2007; 275 teachers Expected commute days—2004 185 15.91 31.48 0.00 126.00 Expected commute days—2005 209 39.18 38.41 0.00 156.00 Expected commute days—2006 223 58.98 39.41 0.00 176.00 Expected commute days—2007 239 66.84 40.40 0.00 195.00 Annual graduate ‘class hours’ 1319 111.34 78.28 6.00 696.00 Panel C: Research output 2001—2008: 521 teachers Lagged expected commute days—2005 456 25.07 36.93 0.00 162.00 Lagged expected commute days—2006 426 52.38 45.00 0.00 172.00 Lagged expected commute days—2007 399 70.03 42.51 0.00 176.00 Lagged expected commute days—2008 386 73.34 42.65 0.00 195.00 Annual publications per teacher 3367 0.96 1.58 0.00 15.00 Annual top publications per teacher 3367 0.01 0.10 0.00 2.00 Annual non-top publications per teacher 3367 0.95 1.56 0.00 15.00 Variable N Mean Std. Dev. Min Max Panel A: Undergraduate teaching 2000–2007; 513 teachers Annual undergraduate ‘class hours’ 3402 260.25 171.16 2.00 1121.00 Annual undergraduate teaching daysa 3399 75.36 39.07 3.00 196.00 Daily undergraduate ‘class hours’a 3399 3.34 1.41 0.21 10.00 Expected commute days—2004 490 24.60 36.91 0.00 162.00 Expected commute days—2005 438 52.62 44.45 0.00 172.00 Expected commute days—2006 414 72.15 42.12 0.00 176.00 Expected commute days—2007 387 75.43 41.27 0.00 195.00 Male 3402 0.58 0.49 0.00 1.00 Position—assistant professor 3402 0.50 0.50 0.00 1.00 Position—associate professor 3402 0.29 0.46 0.00 1.00 Position—full professor 3402 0.17 0.37 0.00 1.00 Panel B: Graduate teaching 2000—2007; 275 teachers Expected commute days—2004 185 15.91 31.48 0.00 126.00 Expected commute days—2005 209 39.18 38.41 0.00 156.00 Expected commute days—2006 223 58.98 39.41 0.00 176.00 Expected commute days—2007 239 66.84 40.40 0.00 195.00 Annual graduate ‘class hours’ 1319 111.34 78.28 6.00 696.00 Panel C: Research output 2001—2008: 521 teachers Lagged expected commute days—2005 456 25.07 36.93 0.00 162.00 Lagged expected commute days—2006 426 52.38 45.00 0.00 172.00 Lagged expected commute days—2007 399 70.03 42.51 0.00 176.00 Lagged expected commute days—2008 386 73.34 42.65 0.00 195.00 Annual publications per teacher 3367 0.96 1.58 0.00 15.00 Annual top publications per teacher 3367 0.01 0.10 0.00 2.00 Annual non-top publications per teacher 3367 0.95 1.56 0.00 15.00 Note: Panel A includes data for faculty who teach undergraduates in 2003 and in at least 1 year after. Panel B includes data for faculty who teach undergraduate students in 2003 and teach undergraduate and graduate students in at least 1 year after. Panel C includes data for any faculty who teachs undergraduate students in 2002 and in at least 1 year after. aNumber of observations for annual teaching days and daily ‘class hours’ is <3402 because some class-year observations are missing day-of-week information. These are included for annual ‘class hours’ because hours are available even if day of week is not. Table 1 Descriptive statistics Variable N Mean Std. Dev. Min Max Panel A: Undergraduate teaching 2000–2007; 513 teachers Annual undergraduate ‘class hours’ 3402 260.25 171.16 2.00 1121.00 Annual undergraduate teaching daysa 3399 75.36 39.07 3.00 196.00 Daily undergraduate ‘class hours’a 3399 3.34 1.41 0.21 10.00 Expected commute days—2004 490 24.60 36.91 0.00 162.00 Expected commute days—2005 438 52.62 44.45 0.00 172.00 Expected commute days—2006 414 72.15 42.12 0.00 176.00 Expected commute days—2007 387 75.43 41.27 0.00 195.00 Male 3402 0.58 0.49 0.00 1.00 Position—assistant professor 3402 0.50 0.50 0.00 1.00 Position—associate professor 3402 0.29 0.46 0.00 1.00 Position—full professor 3402 0.17 0.37 0.00 1.00 Panel B: Graduate teaching 2000—2007; 275 teachers Expected commute days—2004 185 15.91 31.48 0.00 126.00 Expected commute days—2005 209 39.18 38.41 0.00 156.00 Expected commute days—2006 223 58.98 39.41 0.00 176.00 Expected commute days—2007 239 66.84 40.40 0.00 195.00 Annual graduate ‘class hours’ 1319 111.34 78.28 6.00 696.00 Panel C: Research output 2001—2008: 521 teachers Lagged expected commute days—2005 456 25.07 36.93 0.00 162.00 Lagged expected commute days—2006 426 52.38 45.00 0.00 172.00 Lagged expected commute days—2007 399 70.03 42.51 0.00 176.00 Lagged expected commute days—2008 386 73.34 42.65 0.00 195.00 Annual publications per teacher 3367 0.96 1.58 0.00 15.00 Annual top publications per teacher 3367 0.01 0.10 0.00 2.00 Annual non-top publications per teacher 3367 0.95 1.56 0.00 15.00 Variable N Mean Std. Dev. Min Max Panel A: Undergraduate teaching 2000–2007; 513 teachers Annual undergraduate ‘class hours’ 3402 260.25 171.16 2.00 1121.00 Annual undergraduate teaching daysa 3399 75.36 39.07 3.00 196.00 Daily undergraduate ‘class hours’a 3399 3.34 1.41 0.21 10.00 Expected commute days—2004 490 24.60 36.91 0.00 162.00 Expected commute days—2005 438 52.62 44.45 0.00 172.00 Expected commute days—2006 414 72.15 42.12 0.00 176.00 Expected commute days—2007 387 75.43 41.27 0.00 195.00 Male 3402 0.58 0.49 0.00 1.00 Position—assistant professor 3402 0.50 0.50 0.00 1.00 Position—associate professor 3402 0.29 0.46 0.00 1.00 Position—full professor 3402 0.17 0.37 0.00 1.00 Panel B: Graduate teaching 2000—2007; 275 teachers Expected commute days—2004 185 15.91 31.48 0.00 126.00 Expected commute days—2005 209 39.18 38.41 0.00 156.00 Expected commute days—2006 223 58.98 39.41 0.00 176.00 Expected commute days—2007 239 66.84 40.40 0.00 195.00 Annual graduate ‘class hours’ 1319 111.34 78.28 6.00 696.00 Panel C: Research output 2001—2008: 521 teachers Lagged expected commute days—2005 456 25.07 36.93 0.00 162.00 Lagged expected commute days—2006 426 52.38 45.00 0.00 172.00 Lagged expected commute days—2007 399 70.03 42.51 0.00 176.00 Lagged expected commute days—2008 386 73.34 42.65 0.00 195.00 Annual publications per teacher 3367 0.96 1.58 0.00 15.00 Annual top publications per teacher 3367 0.01 0.10 0.00 2.00 Annual non-top publications per teacher 3367 0.95 1.56 0.00 15.00 Note: Panel A includes data for faculty who teach undergraduates in 2003 and in at least 1 year after. Panel B includes data for faculty who teach undergraduate students in 2003 and teach undergraduate and graduate students in at least 1 year after. Panel C includes data for any faculty who teachs undergraduate students in 2002 and in at least 1 year after. aNumber of observations for annual teaching days and daily ‘class hours’ is <3402 because some class-year observations are missing day-of-week information. These are included for annual ‘class hours’ because hours are available even if day of week is not. The idea of our identification strategy is illustrated in Table 2. The top panel shows means of annual ‘class hours’ before versus during the campus transition for high versus low commute cost groups based on their 2003 teaching schedule. The high-cost group—those in the top decile of annual freshman teaching days—faces greater commute costs than those in the low-cost group—those in the top decile of annual senior teaching days. The high-cost group teaches more overall, both before and during the transition, than the low-cost group. Both groups decrease their teaching during the transition although only significantly so for the high-cost group. The difference of these differences is the causal effect assuming that in the absence of the transition the groups’ work times would not significantly differ. The transition decreased work time by 16.7 ‘class hours’ annually and the effect is very significant. The middle panel shows the same estimate for annual teaching days and implies a significant decrease of 18.2 days, while the bottom panel shows a significant increase in daily ‘class hours’ of 0.72. Table 2 Mean annual work time measures for high and low commute costs cohorts before (2000–2003) versus during (2004–2007) campus transition Pre-transition (2000–2003) During transition (2004–2007) Difference Annual ‘class hours’ Top decile—freshman annual 403.6 372.4 −31.2* teaching days (158.8) (186.0) (17.1) 204 205 Top decile—senior annual 265.7 251.2 −14.5 teaching days (136.0) (167.4) (13.9) 226 251 Difference 137.9*** 121.2*** −16.7*** (14.3) (16.7) (1.5) Annual teaching days Top decile—freshman annual 113.3 79.9 −33.4*** teaching days (34.7) (32.3) (3.3) 204 205 Top decile—senior annual 85.8 70.6 −15.2*** teaching days (37.0) (35.9) (3.4) 224 251 Difference 27.5*** 9.3*** −18.2*** (3.5) (3.2) (0.3) Daily ‘class hours’ Top decile—freshman annual 3.60 4.64 1.04*** teaching days (1.24) (1.77) (0.15) 204 205 Top decile—senior annual 3.13 3.45 0.32*** teaching days (0.99) (1.31) (0.11) 224 251 Difference 0.47*** 1.18*** 0.72*** (0.11) (0.15) (0.01) Pre-transition (2000–2003) During transition (2004–2007) Difference Annual ‘class hours’ Top decile—freshman annual 403.6 372.4 −31.2* teaching days (158.8) (186.0) (17.1) 204 205 Top decile—senior annual 265.7 251.2 −14.5 teaching days (136.0) (167.4) (13.9) 226 251 Difference 137.9*** 121.2*** −16.7*** (14.3) (16.7) (1.5) Annual teaching days Top decile—freshman annual 113.3 79.9 −33.4*** teaching days (34.7) (32.3) (3.3) 204 205 Top decile—senior annual 85.8 70.6 −15.2*** teaching days (37.0) (35.9) (3.4) 224 251 Difference 27.5*** 9.3*** −18.2*** (3.5) (3.2) (0.3) Daily ‘class hours’ Top decile—freshman annual 3.60 4.64 1.04*** teaching days (1.24) (1.77) (0.15) 204 205 Top decile—senior annual 3.13 3.45 0.32*** teaching days (0.99) (1.31) (0.11) 224 251 Difference 0.47*** 1.18*** 0.72*** (0.11) (0.15) (0.01) Note: The three panels compare work time before (2000–2003) versus during (2004–2007) the transition for teachers in the top decile of freshman annual teaching days with those in the top decile of senior annual teaching days based on their 2003 teaching schedule. The top panel compares the change in annual ‘class hours’ for the two groups, the middle panel annual teaching days, and the bottom panel daily ‘class hours’ (conditional on working that day). Each cell displays the mean, standard deviation, and number of observations. * = 10% significance, ** = 5% significance, *** = 1% significance for differences in means. Table 2 Mean annual work time measures for high and low commute costs cohorts before (2000–2003) versus during (2004–2007) campus transition Pre-transition (2000–2003) During transition (2004–2007) Difference Annual ‘class hours’ Top decile—freshman annual 403.6 372.4 −31.2* teaching days (158.8) (186.0) (17.1) 204 205 Top decile—senior annual 265.7 251.2 −14.5 teaching days (136.0) (167.4) (13.9) 226 251 Difference 137.9*** 121.2*** −16.7*** (14.3) (16.7) (1.5) Annual teaching days Top decile—freshman annual 113.3 79.9 −33.4*** teaching days (34.7) (32.3) (3.3) 204 205 Top decile—senior annual 85.8 70.6 −15.2*** teaching days (37.0) (35.9) (3.4) 224 251 Difference 27.5*** 9.3*** −18.2*** (3.5) (3.2) (0.3) Daily ‘class hours’ Top decile—freshman annual 3.60 4.64 1.04*** teaching days (1.24) (1.77) (0.15) 204 205 Top decile—senior annual 3.13 3.45 0.32*** teaching days (0.99) (1.31) (0.11) 224 251 Difference 0.47*** 1.18*** 0.72*** (0.11) (0.15) (0.01) Pre-transition (2000–2003) During transition (2004–2007) Difference Annual ‘class hours’ Top decile—freshman annual 403.6 372.4 −31.2* teaching days (158.8) (186.0) (17.1) 204 205 Top decile—senior annual 265.7 251.2 −14.5 teaching days (136.0) (167.4) (13.9) 226 251 Difference 137.9*** 121.2*** −16.7*** (14.3) (16.7) (1.5) Annual teaching days Top decile—freshman annual 113.3 79.9 −33.4*** teaching days (34.7) (32.3) (3.3) 204 205 Top decile—senior annual 85.8 70.6 −15.2*** teaching days (37.0) (35.9) (3.4) 224 251 Difference 27.5*** 9.3*** −18.2*** (3.5) (3.2) (0.3) Daily ‘class hours’ Top decile—freshman annual 3.60 4.64 1.04*** teaching days (1.24) (1.77) (0.15) 204 205 Top decile—senior annual 3.13 3.45 0.32*** teaching days (0.99) (1.31) (0.11) 224 251 Difference 0.47*** 1.18*** 0.72*** (0.11) (0.15) (0.01) Note: The three panels compare work time before (2000–2003) versus during (2004–2007) the transition for teachers in the top decile of freshman annual teaching days with those in the top decile of senior annual teaching days based on their 2003 teaching schedule. The top panel compares the change in annual ‘class hours’ for the two groups, the middle panel annual teaching days, and the bottom panel daily ‘class hours’ (conditional on working that day). Each cell displays the mean, standard deviation, and number of observations. * = 10% significance, ** = 5% significance, *** = 1% significance for differences in means. 6. Econometric model We generalize this identification approach in a regression model of teacher i’s work time in academic year t: Yit=αi+δt+βTrt*CDit+γXit+εit, (6) where Yit is one of three measures of work time (annual ‘class hours’, annual teaching days and daily ‘class hours’), αi is a teacher fixed effect, δt is a year fixed effect and Trt is a dummy variable set to one after the transition begins in 2004 (the treatment effect) and zero before. Expected commute days CDit captures the treatment intensity for faculty member i in year t and equals the expected number of days teacher i would have to commute to the satellite campus based on their 2003 academic-year schedule. We control for teacher-year characteristics Xit such as rank. β captures the effect of an additional commute day on work time. Online Appendix E shows that Equation (6) is consistent with a model allowing for heterogeneous commute costs across teachers as long as these costs are independent of the individual’s treatment intensity and the error structure allows for general heteroskedasticity and clustering within teacher-transition cell (pre- versus post-transition). We therefore allow for this. Our DD estimates compare the responses during the transition of teachers with varying levels of expected commute days—their treatment intensity. Identification of the treatment effect requires that conditional on the control variables in Equation (6) the treatment effect is uncorrelated with the error; otherwise, bias will be introduced. The academic year fixed effects δt in Equation (6) capture year-specific unobserved factors affecting work time. These include university-level changes in wages, student enrollment, faculty size, class size, curriculum, graduation course requirements and national education policies since these are constant within an academic year as well as any aggregate trends in work time common to all faculty.38 The teacher fixed effects αi in Equation (6) absorb unobserved teacher-specific work-time preferences that are invariant across years such as the value of leisure time. Time-varying, teacher-specific changes could bias the estimates if correlated with individual expected commute days. For example, individual-level changes in returns to outside activities such as consulting would bias the estimates if they were correlated with individual expected commute days. In particular, a key identifying assumption of Equation (6) is that work time trends would be identical for the control and treatment groups absent the treatment (Angrist and Pischke, 2009). Table 3 provides evidence that this ‘common trends’ assumption holds. It shows estimates using placebo ‘transitions’ in the pretreatment period. Since we only have data from 2000 to 2003 available for these tests, we can only implement this in two different ways. The first assumes a ‘transition’ beginning in Year 2001 which uses Year 2000 data to define expected commute days. That is, expected commute days in 2000 equal zero, in 2001 equal freshman teaching days in 2000, in 2002 equal freshman and sophomore teaching days in 2000, and in 2003 equal freshman, sophomore and junior teaching days in 2000. The second way assumes a ‘transition’ beginning in Year 2002 which uses Year 2001 data to define expected commute days. In this case, expected commute days in 2000 and 2001 equal zero, in 2002 equal freshman teaching days in 2001, and in 2003 equal freshman and sophomore teaching days in 2001. As Table 3 shows, neither of these artificial policies has a significant effect on any of the three work time measures consistent with an absence of preexisting trends that would bias the estimates. Table 3 Placebo tests for pretreatment (2000–2003) trends in work time measures 1 2 3 4 5 6 Annual ‘class hours’ Annual teaching days Daily ‘class hours’ 2000–2003 2001–2003 2000–2003 2001–2003 2000–2003 2001–2003 (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) Placebo expected −0.157 −0.218 −0.062 −0.091 0.0002 0.0013 commute days (0.274) (0.385) (0.086) (0.097) (0.0024) (0.0028) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes R2 0.734 0.660 0.660 0.750 0.633 0.729 Number of teachers 225 221 225 221 225 221 N 659 431 658 430 658 430 1 2 3 4 5 6 Annual ‘class hours’ Annual teaching days Daily ‘class hours’ 2000–2003 2001–2003 2000–2003 2001–2003 2000–2003 2001–2003 (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) Placebo expected −0.157 −0.218 −0.062 −0.091 0.0002 0.0013 commute days (0.274) (0.385) (0.086) (0.097) (0.0024) (0.0028) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes R2 0.734 0.660 0.660 0.750 0.633 0.729 Number of teachers 225 221 225 221 225 221 N 659 431 658 430 658 430 Note: Columns 1, 3 and 5 include data from 2000 to 2003 and include all teachers from the full sample present in 2000 and in at least 1 year from 2001 to 2003. Columns 2, 4 and 6 include data from 2001 to 2003 and include all teachers from the full sample present in 2001 and at least 1 year from 2002 to 2003. Dependent variable is: annual undergraduate ‘class hours’ in Columns 1 and 2, annual undergraduate teaching days in Columns 3 and 4, and undergraduate daily ‘class hours’ (conditional on teaching that day) in Columns 5 and 6. Placebo variables are constructed based on Year 2000 academic schedule in Columns 1, 3 and 5 and based on Year 2001 academic schedule in Columns 2, 4 and 6 as described in the text. Standard errors in parentheses. Standard errors allow for clustering by teacher and general heteroskedasticity in all regressions. * = 10% significance, ** = 5% significance, *** = 1% significance. Table 3 Placebo tests for pretreatment (2000–2003) trends in work time measures 1 2 3 4 5 6 Annual ‘class hours’ Annual teaching days Daily ‘class hours’ 2000–2003 2001–2003 2000–2003 2001–2003 2000–2003 2001–2003 (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) Placebo expected −0.157 −0.218 −0.062 −0.091 0.0002 0.0013 commute days (0.274) (0.385) (0.086) (0.097) (0.0024) (0.0028) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes R2 0.734 0.660 0.660 0.750 0.633 0.729 Number of teachers 225 221 225 221 225 221 N 659 431 658 430 658 430 1 2 3 4 5 6 Annual ‘class hours’ Annual teaching days Daily ‘class hours’ 2000–2003 2001–2003 2000–2003 2001–2003 2000–2003 2001–2003 (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) (‘2001 Transition’) (‘2002 Transition’) Placebo expected −0.157 −0.218 −0.062 −0.091 0.0002 0.0013 commute days (0.274) (0.385) (0.086) (0.097) (0.0024) (0.0028) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes R2 0.734 0.660 0.660 0.750 0.633 0.729 Number of teachers 225 221 225 221 225 221 N 659 431 658 430 658 430 Note: Columns 1, 3 and 5 include data from 2000 to 2003 and include all teachers from the full sample present in 2000 and in at least 1 year from 2001 to 2003. Columns 2, 4 and 6 include data from 2001 to 2003 and include all teachers from the full sample present in 2001 and at least 1 year from 2002 to 2003. Dependent variable is: annual undergraduate ‘class hours’ in Columns 1 and 2, annual undergraduate teaching days in Columns 3 and 4, and undergraduate daily ‘class hours’ (conditional on teaching that day) in Columns 5 and 6. Placebo variables are constructed based on Year 2000 academic schedule in Columns 1, 3 and 5 and based on Year 2001 academic schedule in Columns 2, 4 and 6 as described in the text. Standard errors in parentheses. Standard errors allow for clustering by teacher and general heteroskedasticity in all regressions. * = 10% significance, ** = 5% significance, *** = 1% significance. 7. Results 7.1. Annual undergraduate ‘class hours’ The left panel of Table 4 shows the results of DD estimates using Equation (6) with annual undergraduate ‘class hours’ as the dependent variable. These estimates are based on an unbalanced panel that includes any faculty present in 2003 and in at least 1 year of the transition. This relates individual-level changes in teaching time to individual-level changes in commute time during the transition years. Column 1 includes teacher and year fixed effects. This is our preferred specification and it shows that an increase of one additional expected commute day decreases annual ‘class hours’ by 0.27 at a very high level of significance. These teachers on average taught 82.3 days annually from 2000 to 2003 (pre-transition). Using this as an estimate of the total number of commute days the full transition would impose implies a decrease of 21.9 annual ‘class hours’ from the full transition (8.4% of the average 260.0 annual ‘class hours’ that these faculty taught pre-transition). This implies an elasticity of −0.044 for work time with respect to commute time or distance.39 Since wages are linear the elasticity of income with respect to commute time is the same. Gutiérrez-i-Puigarnau and van Ommeren (2010a) estimate an elasticity of −0.009 for work time with respect to commute distance, consistent with their excluding workers with high commute-cost sensitivity, the workers in our sample having more flexibility over work time or both. Table 4 Effect of expected commute days on work time measures (2000–2007) 1 2 3 4 5 6 7 8 9 10 11 12 Annual undergraduate ‘class hours’ Annual undergraduate teaching days Daily undergraduate ‘class hours’ Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Expected commute days −0.266*** 0.108 −0.414*** −0.213*** −0.129** −0.203*** 0.0061*** 0.0078*** 0.0026** (0.100) (0.268) (0.152) (0.024) (0.062) (0.041) (0.0008) (0.0020) (0.0012) (Expected commute days)2 −0.0029 −0.00064 −0.000013 (0.0021) (0.00045) (0.00001) Quintiles—Average expected commute days 1st quintile −3.459 −2.734 0.0626 (10.115) (2.683) (0.0976) 2nd quintile −10.579 −10.893*** 0.3426*** (11.729) (2.915) (0.1010) 3rd quintile −0.992 −10.015*** 0.4612*** (11.318) (2.991) (0.1040) 4th quintile −16.442 −18.532*** 0.5948*** (12.528) (3.031) (0.1159) 5th quintile −52.360*** −29.861*** 0.6680*** (15.250) (3.678) (0.1177) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Teacher-specific linear trend No No No Yes No No No Yes No No No Yes Number of teachers 513 513 513 513 513 513 513 513 513 513 513 513 Prob > F (linear + quadratic) 0.027 0.000 0.000 Prob > F (quintiles) 0.020 0.000 0.000 R2 0.635 0.636 0.635 0.744 0.522 0.523 0.523 0.655 0.550 0.549 0.551 0.678 N 3402 3402 3402 3402 3399 3399 3399 3399 3399 3399 3399 3399 1 2 3 4 5 6 7 8 9 10 11 12 Annual undergraduate ‘class hours’ Annual undergraduate teaching days Daily undergraduate ‘class hours’ Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Expected commute days −0.266*** 0.108 −0.414*** −0.213*** −0.129** −0.203*** 0.0061*** 0.0078*** 0.0026** (0.100) (0.268) (0.152) (0.024) (0.062) (0.041) (0.0008) (0.0020) (0.0012) (Expected commute days)2 −0.0029 −0.00064 −0.000013 (0.0021) (0.00045) (0.00001) Quintiles—Average expected commute days 1st quintile −3.459 −2.734 0.0626 (10.115) (2.683) (0.0976) 2nd quintile −10.579 −10.893*** 0.3426*** (11.729) (2.915) (0.1010) 3rd quintile −0.992 −10.015*** 0.4612*** (11.318) (2.991) (0.1040) 4th quintile −16.442 −18.532*** 0.5948*** (12.528) (3.031) (0.1159) 5th quintile −52.360*** −29.861*** 0.6680*** (15.250) (3.678) (0.1177) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Teacher-specific linear trend No No No Yes No No No Yes No No No Yes Number of teachers 513 513 513 513 513 513 513 513 513 513 513 513 Prob > F (linear + quadratic) 0.027 0.000 0.000 Prob > F (quintiles) 0.020 0.000 0.000 R2 0.635 0.636 0.635 0.744 0.522 0.523 0.523 0.655 0.550 0.549 0.551 0.678 N 3402 3402 3402 3402 3399 3399 3399 3399 3399 3399 3399 3399 Dependent variable is: annual undergraduate ‘class hours’ in left panel, annual undergraduate teaching days in middle panel, and undergraduate daily ‘class hours’ (conditional on teaching that day) in right panel. Standard errors in parentheses. Standard errors allow for clustering within teacher-transition cell and general heteroskedasticity in all regressions. * = 10% significance, ** = 5% significance, *** = 1% significance. The F-tests are the p-value for the joint significance level of the coefficients on the five quintile expected commute day terms in Columns 2, 6 and 10 and the joint significance of the coefficients on the linear and quadratic expected commute day terms in Columns 3, 7 and 11. Table 4 Effect of expected commute days on work time measures (2000–2007) 1 2 3 4 5 6 7 8 9 10 11 12 Annual undergraduate ‘class hours’ Annual undergraduate teaching days Daily undergraduate ‘class hours’ Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Expected commute days −0.266*** 0.108 −0.414*** −0.213*** −0.129** −0.203*** 0.0061*** 0.0078*** 0.0026** (0.100) (0.268) (0.152) (0.024) (0.062) (0.041) (0.0008) (0.0020) (0.0012) (Expected commute days)2 −0.0029 −0.00064 −0.000013 (0.0021) (0.00045) (0.00001) Quintiles—Average expected commute days 1st quintile −3.459 −2.734 0.0626 (10.115) (2.683) (0.0976) 2nd quintile −10.579 −10.893*** 0.3426*** (11.729) (2.915) (0.1010) 3rd quintile −0.992 −10.015*** 0.4612*** (11.318) (2.991) (0.1040) 4th quintile −16.442 −18.532*** 0.5948*** (12.528) (3.031) (0.1159) 5th quintile −52.360*** −29.861*** 0.6680*** (15.250) (3.678) (0.1177) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Teacher-specific linear trend No No No Yes No No No Yes No No No Yes Number of teachers 513 513 513 513 513 513 513 513 513 513 513 513 Prob > F (linear + quadratic) 0.027 0.000 0.000 Prob > F (quintiles) 0.020 0.000 0.000 R2 0.635 0.636 0.635 0.744 0.522 0.523 0.523 0.655 0.550 0.549 0.551 0.678 N 3402 3402 3402 3402 3399 3399 3399 3399 3399 3399 3399 3399 1 2 3 4 5 6 7 8 9 10 11 12 Annual undergraduate ‘class hours’ Annual undergraduate teaching days Daily undergraduate ‘class hours’ Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Baseline model Quintiles Quadratic effect Teacher- specific trend Expected commute days −0.266*** 0.108 −0.414*** −0.213*** −0.129** −0.203*** 0.0061*** 0.0078*** 0.0026** (0.100) (0.268) (0.152) (0.024) (0.062) (0.041) (0.0008) (0.0020) (0.0012) (Expected commute days)2 −0.0029 −0.00064 −0.000013 (0.0021) (0.00045) (0.00001) Quintiles—Average expected commute days 1st quintile −3.459 −2.734 0.0626 (10.115) (2.683) (0.0976) 2nd quintile −10.579 −10.893*** 0.3426*** (11.729) (2.915) (0.1010) 3rd quintile −0.992 −10.015*** 0.4612*** (11.318) (2.991) (0.1040) 4th quintile −16.442 −18.532*** 0.5948*** (12.528) (3.031) (0.1159) 5th quintile −52.360*** −29.861*** 0.6680*** (15.250) (3.678) (0.1177) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Teacher-specific linear trend No No No Yes No No No Yes No No No Yes Number of teachers 513 513 513 513 513 513 513 513 513 513 513 513 Prob > F (linear + quadratic) 0.027 0.000 0.000 Prob > F (quintiles) 0.020 0.000 0.000 R2 0.635 0.636 0.635 0.744 0.522 0.523 0.523 0.655 0.550 0.549 0.551 0.678 N 3402 3402 3402 3402 3399 3399 3399 3399 3399 3399 3399 3399 Dependent variable is: annual undergraduate ‘class hours’ in left panel, annual undergraduate teaching days in middle panel, and undergraduate daily ‘class hours’ (conditional on teaching that day) in right panel. Standard errors in parentheses. Standard errors allow for clustering within teacher-transition cell and general heteroskedasticity in all regressions. * = 10% significance, ** = 5% significance, *** = 1% significance. The F-tests are the p-value for the joint significance level of the coefficients on the five quintile expected commute day terms in Columns 2, 6 and 10 and the joint significance of the coefficients on the linear and quadratic expected commute day terms in Columns 3, 7 and 11. In unreported results, we estimate the baseline model but include the post-transition years of 2008 and 2009. This potentially adds more noise because it is further in time from the 2003 data used to proxy for commute costs; nonetheless, the estimates are very similar to the baseline results (a significant decrease of 21.3 ‘class hours’). Column 2 of Table 4 allows for a piecewise linear effect of commute costs. It uses dummy variables to indicate each quintile of expected commute days. The coefficients are jointly significant at the 2% level. Dividing each coefficient by the average expected commute days within the quintile yields marginal effects of an additional expected commute day of −0.16, −0.23, −0.01, −0.17 and −0.40. With the exception of the third quintile, these bracket our baseline estimate (−0.27). Column 3 allows for a quadratic function of expected commute days. The coefficients are collectively significant at the 2.7% level. The average partial effects are −0.202 (close to the baseline estimates) with a standard error of 0.098. Evaluated at the 25th, 50th and 75th percentiles of the pre-transition teaching days, the estimates imply a decrease of 1.7, 9.6 and 24.2 annual ‘class hours’. Column 4 introduces teacher-specific time trends αit that relax the parallel slopes assumption. This also allows for the possibility that an individual teacher’s desire to work may change over time due to promotions, changes in research productivity, changing financial conditions or changes in the attractiveness of outside options. The point estimate is greater in magnitude than the baseline estimate. The next two subsections provide additional evidence consistent with commute time as the cause of decreased teaching time by testing theoretical predictions for annual teaching days and daily ‘class hours’. 7.2. Annual undergraduate teaching days The middle panel of Table 4 shows estimates of Equation (6) for annual undergraduate teaching days. Our preferred specification with academic-year and teacher fixed effects in Column 5 shows a decrease of 0.21 annual teaching days for each additional expected commute day. Grossing this up in the same way as for annual ‘class hours’ implies a decrease of 17.5 annual teaching days from the full transition or 21.3% of the pre-transition 82.3 annual teaching days. In unreported results, the estimate is robust to extending the data through the post-transition period (significant decrease of 17.2 days). For quintile estimates (Column 6), all but one coefficient is significant on its own and jointly they are significant at better than the 0.1% level. Divided by the mean expected commute days within each quintile yield marginal effects of an additional expected commute day (−0.13, −0.23, −0.14, −0.19 and −0.23) that bracket the baseline estimate (−0.21). The coefficients for a quadratic function of expected commute days (Column 7) are highly jointly significant and yield average partial effects of −0.20 with a standard error of 0.025. Evaluated at the 25th, 50th and 75th percentile of the pre-transition teaching days, the estimates imply decreases of 8.0, 18.5 and 22.7 annual teaching days. A teacher-specific time trend (Column 8) yields results almost identical to the baseline. 7.3. Daily undergraduate ‘class hours’ The right panel of Table 4 shows the DD estimates of Equation (6) for daily undergraduate ‘class hours’. The preferred specification in Column 9 shows an increase of 0.0061 daily ‘class hours’ for each additional expected commute day. Grossing up these changes over the full transition yields an increase of 0.50 daily ‘class hours’ or 16.2% of the pre-transition average of 3.1 daily ‘class hours’. In unreported results, the estimate is almost identical if the post-transition years are included (significant increase of 0.51 ‘class hours’). For quintile estimates (Column 10), all but one coefficient is significant on its own and jointly they are significant at better than the 0.1% level. Dividing the coefficients by the mean expected commute days within each quintile yield marginal effects of an additional expected commute day (0.0030, 0.0073, 0.0065, 0.0061 and 0.0051) that bracket the baseline estimate (0.0061). The coefficients on a quadratic function of expected commute days (Column 11) are jointly significant at better than the 0.1% level and yield average partial effects of 0.0063 with a standard error of 0.00084. Evaluated at the 25th, 50th and 75th percentiles of the pre-transition teaching days implies decreases of 0.36, 0.54 and 0.71 daily ‘class hours’. Allowing for teacher-specific time trends reduces the estimate by about one-half (Column 12) to an increase of 0.22 daily ‘class hours’ from the full transition. 7.4. Summary Together, these results confirm the theoretical predictions of Section 3. Annual teaching days decrease and daily ‘class hours’ increase consistent with a longer commute imposing higher daily fixed costs. Teachers averaged 82.3 annual undergraduate teaching days before the transition. Given a daily round-trip commute time of 1−1.5 h, this would require 82.3−123.4 commute hours per year after the transition. We estimate teachers decreased annual ‘class hours’ by 21.9 in response to this implying that they value commute time at 18−27% of their ‘hourly’ wage.40 Multiplying by the wage, the average teacher would pay RMB 16–24 (USD 2.5–3.8)41 to avoid one commute hour given their Year 2011 ‘hourly’ wage of RMB 90 (USD 14.2). Put differently, faculty on average dislike undergraduate teaching more than commuting and would prefer 3.8−5.6 h commuting to one ‘class hour’ teaching. This is consistent with greater disamenity from undergraduate teaching than commuting (Becker, 1965) or it may partially reflect preparation time spent outside the classroom. Shirking could increase as a substitute for leisure time lost to longer commutes.42 Such effects are likely small in our setting given that teachers work in front of a class. Longer commute times can also increase absenteeism (Gutiérrez-i-Puigarnau and van Ommeren, 2010b) but in this case teachers must make up any missed classes. The estimates for annual teaching days and daily ‘class hours’ imply roughly the same impact on overall work time as the direct estimates of annual ‘class hours’. Teachers taught 3.22 daily ‘class hours’ and 81.2 annual teaching days in 2003. Using our preferred estimates and calculating (daily ‘class hours’ in 2003 + estimated change in daily ‘class hours’) × (annual teaching days in 2003 + estimated change in annual teaching days) – (daily ‘class hours’ in 2003) × (annual teaching days in 2003) yields a decrease of 24.7 ‘class hours’ which is close to our estimate of a decrease of 21.9 annual ‘class hours’. 7.5. Robustness checks Table 5 contains robustness checks for our baseline estimates of annual ‘class hours’ reproduced in Column 1. Column 2 tests whether teachers with a high degree of flexibility in course scheduling were differentially affected relative to those with little. This serves as a test of how sensitive our results are to our assumption that faculty freely choose their working time subject to the university’s offered wage. If they are not freely able to choose then we might expect the commute time response to be lower for the group that displays less flexibility. Table 5 Effect of expected commute days on undergraduate annual ‘class hours’ (2000–2007)—robustness checks 1 2 3 4 5 6 Close to minimum teaching load Baseline model ‘High’ versus ‘low’ flexibility 240 ≤ 2003 ‘class hours’ ≤ 250 240 ≤ 2003 ‘class hours’ ≤ 260 Above minimum threshold Class size Expected commute days −0.266*** −0.392* −0.266*** −0.267*** −0.251* 0.065*** (0.100) (0.205) (0.100) (0.101) (0.135) (0.014) (Expected commute days)* 0.002 (‘Low’ flexibility) (0.253) (Expected commute days)* 0.154 0.119 (Close to threshold) (0.777) (0.400) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Number of teachers 513 322 513 513 288 513 N 3402 928 3402 3402 2020 3400 R2 0.635 0.752 0.635 0.635 0.533 0.410 1 2 3 4 5 6 Close to minimum teaching load Baseline model ‘High’ versus ‘low’ flexibility 240 ≤ 2003 ‘class hours’ ≤ 250 240 ≤ 2003 ‘class hours’ ≤ 260 Above minimum threshold Class size Expected commute days −0.266*** −0.392* −0.266*** −0.267*** −0.251* 0.065*** (0.100) (0.205) (0.100) (0.101) (0.135) (0.014) (Expected commute days)* 0.002 (‘Low’ flexibility) (0.253) (Expected commute days)* 0.154 0.119 (Close to threshold) (0.777) (0.400) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Number of teachers 513 322 513 513 288 513 N 3402 928 3402 3402 2020 3400 R2 0.635 0.752 0.635 0.635 0.533 0.410 Note: Dependent variable in Columns 1 through 5 is annual undergraduate ‘class hours’ and in Column 6 is class size. Columns 1, 3, 4 and 6 include teachers present in 2003 and in at least 1 year after 2003. The number of observations in Column 6 differ due to missing class size information for some teachers. Column 2 includes teachers in the top and bottom deciles of flexibility as defined in the text. Column 5 includes all teachers present in 2003 and in at least 1 year after 2003 and exceed the minimum teaching load in the previous year. The variable ‘Close to Threshold’ is set equal to 1 in Column 3 (4) if the teacher had between 240 and 250 (240 and 260, respectively) ‘class hours’ inclusive in Year 2004 and zero otherwise. Standard errors in parentheses. Standard errors allow for clustering within teacher-transition cell and general heteroskedasticity in all regressions. * = 10% significance, ** = 5% significance, *** = 1% significance. Table 5 Effect of expected commute days on undergraduate annual ‘class hours’ (2000–2007)—robustness checks 1 2 3 4 5 6 Close to minimum teaching load Baseline model ‘High’ versus ‘low’ flexibility 240 ≤ 2003 ‘class hours’ ≤ 250 240 ≤ 2003 ‘class hours’ ≤ 260 Above minimum threshold Class size Expected commute days −0.266*** −0.392* −0.266*** −0.267*** −0.251* 0.065*** (0.100) (0.205) (0.100) (0.101) (0.135) (0.014) (Expected commute days)* 0.002 (‘Low’ flexibility) (0.253) (Expected commute days)* 0.154 0.119 (Close to threshold) (0.777) (0.400) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Number of teachers 513 322 513 513 288 513 N 3402 928 3402 3402 2020 3400 R2 0.635 0.752 0.635 0.635 0.533 0.410 1 2 3 4 5 6 Close to minimum teaching load Baseline model ‘High’ versus ‘low’ flexibility 240 ≤ 2003 ‘class hours’ ≤ 250 240 ≤ 2003 ‘class hours’ ≤ 260 Above minimum threshold Class size Expected commute days −0.266*** −0.392* −0.266*** −0.267*** −0.251* 0.065*** (0.100) (0.205) (0.100) (0.101) (0.135) (0.014) (Expected commute days)* 0.002 (‘Low’ flexibility) (0.253) (Expected commute days)* 0.154 0.119 (Close to threshold) (0.777) (0.400) Teacher fixed effects Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Number of teachers 513 322 513 513 288 513 N 3402 928 3402 3402 2020 3400 R2 0.635 0.752 0.635 0.635 0.533 0.410 Note: Dependent variable in Columns 1 through 5 is annual undergraduate ‘class hours’ and in Column 6 is class size. Columns 1, 3, 4 and 6 include teachers present in 2003 and in at least 1 year after 2003. The number of observations in Column 6 differ due to missing class size information for some teachers. Column 2 includes teachers in the top and bottom deciles of flexibility as defined in the text. Column 5 includes all teachers present in 2003 and in at least 1 year after 2003 and exceed the minimum teaching load in the previous year. The variable ‘Close to Threshold’ is set equal to 1 in Column 3 (4) if the teacher had between 240 and 250 (240 and 260, respectively) ‘class hours’ inclusive in Year 2004 and zero otherwise. Standard errors in parentheses. Standard errors allow for clustering within teacher-transition cell and general heteroskedasticity in all regressions. * = 10% significance, ** = 5% significance, *** = 1% significance. We quantify a teacher’s flexibility based on changes across class levels in their schedule from 2002 to 2003.43 Specifically, we define flexibility as the absolute value of the change in ‘class hours’ between class levels across all levels: ∑j∈{Fr,So,Ju,Se}|Hi2003j-Hi2002j| where Hitj is annual ‘class hours’ teacher i taught class level j in year t. We then identify those teachers in the top and bottom deciles of this measure and estimate Equation (6) distinguishing the effect of commute costs for these two groups. This also serves as a test of whether avoidance behavior in the window between 2002 and 2003, when the transition sequence became known, is of concern. Comparing teachers who change schedules greatly (and likely have greater discretion to engage in avoidance behavior) with those who do not serves as a test of this avoidance behavior. Although the significance is lowered by a smaller sample, Column 2 shows that commute time has a negative effect on annual ‘class hours’ and that the effects on high- and low-flexibility teachers are not significantly different. The university reduced the minimum teaching load during the transition (in 2005) from 240 to 225 annual ‘class hours’; however, this does not appear to confound our estimates. The proportion of faculty at the minimum of 240 ‘class hours’ is 2 out of 225 (proportion of 0.0089 with a standard error of 0.0063) in 2003 just before the transition and 7 out of 254 (proportion of 0.0276 with a standard error of 0.0103) in 2004 right after the transition begins. Although the fraction is somewhat higher right after the transition it is not statistically significantly different. These small fractions are also consistent with no significant bunching at the minimum thresholds which is confirmed by the absence of any significant discontinuities in the figure in Online Appendix A. Because of the availability of substitutes for teaching (supervising graduate theses, administrative tasks, and supervising student internships and study trips), there is also a significant proportion of faculty at or below the minimum: 0.439 (standard error of 0.022) in 2003, 0.518 (0.023) in 2004 and 0.477 (0.024) in 2005. The fraction is slightly higher once the transition begins but not statistically so and the fraction falls in the second year of the transition. Columns 3 and 4 of Table 5 use regression analysis to test whether the reduction in the minimum teaching load in 2005 caused some of the reduction in teaching time. In Column 3 we create a dummy variable set to one across all years if the teacher’s annual ‘class hours’ were close to the threshold (between 240 and 250 annual ‘class hours’ inclusive) in 2003 (pre-transition) and zero otherwise. These are the teachers for whom the minimum constraint was most binding in 2004 and, therefore, should adjust their teaching time the most in 2005. We then interact this dummy variable with expected commute days in each year to distinguish the effect of commute costs on this group. The interaction term is positive and insignificant while the baseline effect of expected commute days is unchanged. This is consistent with reduced teaching time being attributable to increased commute costs rather than the minimum teaching load change. Column 4 widens the definition of ‘close to the threshold’ to between 240 and 260 ‘class hours’. The results are very similar. Column 5 tests whether there is a differential effect on teachers who are above the minimum teaching load in the previous year (240 annual ‘class hours’ in 2003 and 2004 and 225 in 2005 through 2007). It is useful to check this subsample because they are less likely to use other activities besides teaching to fulfill their minimum teaching load and we do not observe the change in these. ‘Class hours’ decrease by 0.25 per expected commute day. Average annual teaching days for this subsample pre-transition is 108.3 implying a decrease of 27.1 annual ‘class hours’ from the full transition. This is above the estimate from the full sample; however, the effects are actually somewhat lower in percentage terms—6.9% (the subsample averages 391 annual ‘class hours’ pre-transition) versus 8.4% for the full sample. Any confounding factors that would bias our DD results must be correlated with teacher-specific changes in commute days. Unilateral actions by the university to increase class sizes for all faculty members are captured by the year fixed effects. However, administrators may put greater pressure to teach larger classes on teachers who face large commute time increases because few faculty will tolerate commuting. Larger classes may require more outside preparation time and could lead teachers with larger commute time increases to further reduce their in-class teaching time leading to a bias away from zero. We show this more formally in Online Appendix F by extending our theoretical model to allow effort to depend on class size in addition to ‘class hours’. At the aggregate level, average class size at the satellite campus does exceed that at the main campus during the transition years (71.0 versus 63.0 significant at the 1% level) consistent with faculty wishing to avoid commuting. This is a relatively small difference and, even if the larger class sizes at the satellite campus required more outside preparation time, this would bias our DD estimates only if class sizes are correlated with individual-level commute costs. To examine whether this is the case we estimated Equation (6) using class size as the dependent variable to relate class size with individual-level commute costs. The results are shown in Column 6 of Table 5. Expected commute days is significantly correlated with class size; however, the effect is small. Each additional expected commute day is associated with an additional 0.065 students. Grossing up by the average pre-transition teaching days implies a class size increase of 5.3 students from the full transition. This is only 8.6% of the average pre-transition class size of 61.8 students. 7.6. Sensitivity analysis An alternative way to estimate the effect of commute costs is an interrupted time series comparing work time before (2000–2003) versus after (2007–2009) the transition of all four undergraduate class levels to the satellite campus. Online Appendix G describes a model for these estimates along with a discussion of identification conditions, summary statistics and results. Annual ‘class hours’ decline by 20.3 due to the transition which is similar to our DD estimate of 21.9. Estimates for annual teaching days are larger (a drop of 27 days) than the DD estimates (a drop of 18 days) while those for daily ‘class hours’ are higher (an increase of 0.91) than the DD estimate of 0.50. Online Appendix H compares teachers with high commute costs to those with low to examine whether the timing of work time adjustments is consistent across the different groups. This also serves as another test of the parallel trends assumption across the groups. We construct a ‘highly-treated’ group that includes teachers for whom freshman teaching hours comprised 80% or more of their total ‘class hours’ and who were in the 80th percentile of total freshman ‘class hours’; and a ‘minimally-treated’ group that includes teachers for whom junior plus senior ‘class hours’ comprise 60% or more of their total ‘class hours’ and who were in the 60th percentile or above of total junior plus senior ‘class hours’. Consistent with expectations, the effects are larger for the ‘highly treated’ group than for the full sample—each additional expected commute day lowers annual ‘class hours’ by 0.69. For the ‘minimally treated’ group, as expected, there is little reduction in the first 2 years of the transition because junior and senior class levels have not yet transitioned to the new campus but significant effects in the third and fourth years of the transition when juniors and seniors move. 7.7. Role of demographics The table in Online Appendix I examines the role of faculty demographics in the response of annual undergraduate ‘class hours’ to commute time. In contrast to previous evidence that female work time is more sensitive to commute costs, there is no significant difference by gender in the effect of the transition on work time.44 The lack of any significant difference is not because a disproportionate fraction of senior faculty is male and the increased bargaining power that conveys offsets higher commute cost sensitivity among female faculty: interacting rank and gender reveals no significant differences. The transition also has similar effects across faculty ranks consistent with robustness to differences in outside options to undergraduate teaching across ranks (graduate courses are usually taught by associate or full professors, full professors are the only faculty rank legally allowed to supervise Ph.D. theses and senior faculty has greater consulting opportunities).45 7.8. Possible university responses How did the university accommodate the decrease in per-teacher undergraduate teaching time? We cannot precisely answer this question but we offer some evidence based on the annual demand for and supply of undergraduate student ‘class hours’: Demand=(Number of Students)*(Student ‘Class Hours’/Student), (7a) Supply=(Number of Teachers)*(Teacher ‘Class Hours’/Teacher)*(Class Size). (7b) There are four possible margins of adjustment which are not mutually exclusive. The university could reduce demand for teaching time by: (i) admitting fewer students or (ii) reducing the number of student ‘class hours’ required per student; or it could increase supply by: (iii) hiring more teachers or (iv) increasing class sizes.46 In Online Appendix J, we approximate these margins of adjustment by taking differentials of Equations (7a) and (7b) and evaluate the changes using the average of 2003 (just before the transition) and 2007 (just after the full transition) data. Demand for undergraduate teaching time increased by 2.83 million student ‘class hours’ between 2003 and 2007. A large increase in the student body increased demand by 3.32 million student ‘class hours’ which was offset by 0.49 million due to an aggregate decline in student ‘class hours’ per student. Given such a large increase in demand, the university was likely encouraging faculty to teach more rather than applying pressure for them to teach less or effectively demoting them by reducing their teaching time and therefore wages. A larger faculty size increased the supply of teaching time by 2.73 million student ‘class hours’ between 2003 and 2007.47 However, decreased teaching time per faculty member decreased aggregate supply by 1.78 million student ‘class hours’. Therefore, without accounting for class size changes, supply increased by 0.95 million student ‘class hours’ annually and demand exceeded supply by 1.88 million. This excess was met by a dramatic increase in class size. The university increased the average number of students per class by 12.4 students between 2003 and 2007. Thus, preparation time outside of class likely increased on average for all faculty. However, as the estimates in Column 6 of Table 5 show this was largely uncorrelated with individual-level commute costs and therefore differenced out in our DD estimates. 7.9. Graduate teaching and research output Column 1 of Table 6 presents estimates from a DD specification using Equation (6) with annual graduate ‘class hours’ as the dependent variable and expected commute days based on undergraduate teaching as the explanatory variable. A teacher is included in this sample if they taught undergraduates in 2003 and in at least 1 year after that and taught graduate students in at least 1 year of the transition. The results show no evidence of substitution toward graduate teaching due to the increased commute costs. Table 6 Effect of expected commute days on annual graduate ‘class hours’ (2000–2007) and annual research output (2001–2008) 1 2 3 4 5 6 7 Annual graduate ‘class hours’ (2000–2007) Total publications (2001–2008) Non-top publications (2001–2008) Top publications (2001–2008) All teachers Active researchers All teachers Active researchers All teachers Active researchers Expected commute days 0.0165 (0.1059) Lagged expected commute days 0.0003 0.0003 0.0004 0.0005 −0.00015* −0.00018* (0.0010) (0.0011) (0.0010) (0.0011) (0.00008) (0.00009) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Number of teachers 448 521 441 521 441 521 441 N 1319 3367 2909 3367 2909 3367 2909 R2 0.598 0.527 0.498 0.520 0.491 0.255 0.254 1 2 3 4 5 6 7 Annual graduate ‘class hours’ (2000–2007) Total publications (2001–2008) Non-top publications (2001–2008) Top publications (2001–2008) All teachers Active researchers All teachers Active researchers All teachers Active researchers Expected commute days 0.0165 (0.1059) Lagged expected commute days 0.0003 0.0003 0.0004 0.0005 −0.00015* −0.00018* (0.0010) (0.0011) (0.0010) (0.0011) (0.00008) (0.00009) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Number of teachers 448 521 441 521 441 521 441 N 1319 3367 2909 3367 2909 3367 2909 R2 0.598 0.527 0.498 0.520 0.491 0.255 0.254 *Note: Dependent variable is annual graduate ‘class hours’ in Column 1, annual output of total research publications in Columns 2 and 3, annual output of non-top research publications in Columns 4 and 5, annual output of top research publications in Columns 6 and 7. Standard errors in parentheses. Standard errors allow for clustering within teacher-transition cell and general heteroskedasticity in all regressions. = 10% significance, **= 5% significance, *** = 1% significance. Column 1 includes any faculty who teach undergraduate students in 2003 and at least 1 year after 2003 and graduate students in at least 1 year of the transition. Columns 2, 4 and 6 include faculty who teach undergraduate students in 2002 and at least 1 year after. Columns 3, 5 and 7 include faculty who teach undergraduate students in 2002 and at least 1 year after and produced at least one research paper during the transition. Table 6 Effect of expected commute days on annual graduate ‘class hours’ (2000–2007) and annual research output (2001–2008) 1 2 3 4 5 6 7 Annual graduate ‘class hours’ (2000–2007) Total publications (2001–2008) Non-top publications (2001–2008) Top publications (2001–2008) All teachers Active researchers All teachers Active researchers All teachers Active researchers Expected commute days 0.0165 (0.1059) Lagged expected commute days 0.0003 0.0003 0.0004 0.0005 −0.00015* −0.00018* (0.0010) (0.0011) (0.0010) (0.0011) (0.00008) (0.00009) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Number of teachers 448 521 441 521 441 521 441 N 1319 3367 2909 3367 2909 3367 2909 R2 0.598 0.527 0.498 0.520 0.491 0.255 0.254 1 2 3 4 5 6 7 Annual graduate ‘class hours’ (2000–2007) Total publications (2001–2008) Non-top publications (2001–2008) Top publications (2001–2008) All teachers Active researchers All teachers Active researchers All teachers Active researchers Expected commute days 0.0165 (0.1059) Lagged expected commute days 0.0003 0.0003 0.0004 0.0005 −0.00015* −0.00018* (0.0010) (0.0011) (0.0010) (0.0011) (0.00008) (0.00009) Teacher fixed effects Yes Yes Yes Yes Yes Yes Yes Academic-year fixed effects Yes Yes Yes Yes Yes Yes Yes Number of teachers 448 521 441 521 441 521 441 N 1319 3367 2909 3367 2909 3367 2909 R2 0.598 0.527 0.498 0.520 0.491 0.255 0.254 *Note: Dependent variable is annual graduate ‘class hours’ in Column 1, annual output of total research publications in Columns 2 and 3, annual output of non-top research publications in Columns 4 and 5, annual output of top research publications in Columns 6 and 7. Standard errors in parentheses. Standard errors allow for clustering within teacher-transition cell and general heteroskedasticity in all regressions. = 10% significance, **= 5% significance, *** = 1% significance. Column 1 includes any faculty who teach undergraduate students in 2003 and at least 1 year after 2003 and graduate students in at least 1 year of the transition. Columns 2, 4 and 6 include faculty who teach undergraduate students in 2002 and at least 1 year after. Columns 3, 5 and 7 include faculty who teach undergraduate students in 2002 and at least 1 year after and produced at least one research paper during the transition. Columns 2 through 7 of Table 6 present estimates of the effect of commute costs on research productivity. These are estimated using the DD specification in Equation (6) with various measures of annual research output (total publications, top publications and non-top publications) per teacher as the dependent variable. We lag expected commute days by 1 year since we estimate it takes about 1 year to write and publish a paper in a Chinese journal and 96.5% of the publications in our sample appear in such journals. A teacher is included in this sample if they taught undergraduates in 2002 and in at least 1 year after that. We also estimate using an ‘active researcher’ subsample of faculty who produced at least one research paper during the transition. Commuting has no effect on total publications (Columns 2 and 3). If we separate effects, commuting has no effect on non-top publications (Columns 4 and 5) but appears to ‘crowd out’ top publications using either all faculty or active researchers (Columns 6 and 7). This could be caused by increased fatigue or more juggling of tasks due to interruptions from commuting (Coviello et al., 2015). Using the results for all teachers, each additional expected commute day reduces publications produced a year later by 0.00015. Grossing up using the average annual days taught pre-transition implies 0.012 fewer annual per capita publications. 8. Conclusion and discussion There is little evidence about the causal effect of commute costs on labor supply. The few available results are subject to endogeneity, imprecise measures of commute costs or lack of comparability in predicting out of sample. Subject to these caveats, previous results indicate a small or no change in labor supply from commute costs changes. In contrast, we find that teaching time drops significantly in response to an exogenous increase in commute time, estimating a commute distance elasticity of work time of −0.044 which is almost five times larger than previous estimates. Vis-à-vis the previous literature, our results suggest caution in concluding that work time responds little to commute costs. Our results suggest that cost–benefit analyses of transportation design should consider labor supply responses to changes in commute costs. Similarly, evaluations of policies alleviating traffic congestion such as driving restrictions, staggered work hours and reversible lanes should incorporate the resulting work time changes. It has been suggested that congestion taxes replace income taxes because the former reduces the negative externalities from driving even though both distort labor supply.48 Our results imply that faster commutes under a congestion tax will offset some of the labor market distortion arising from monetary commute costs. Our results also have ramifications for theoretical labor supply models. Some assume that work days are fixed and daily hours chosen (Cogan, 1981) while others assume the opposite (Parry and Bento, 2001). We find that both margins adjust. Our results imply a role for commute time in the long-run level and rate of city growth. Longer commutes will directly negatively impact a city’s productivity especially as it concerns attracting high human capital or knowledge workers. The presence of knowledge spillovers in the workplace (Fu, 2007; Rosenthal and Strange, 2008) implies that productivity growth would also suffer from less workplace time. This is particularly relevant given the longer commute times caused by urban sprawl (Brueckner, 2001). The competition between ‘edge’ and core cities will also be affected by the decreased work time of those who commute between the two (Henderson and Mitra, 1996). Those with flexible work time (to which our estimates are most relevant) exert an outsized effect on the economy. The self-employed, especially entrepreneurs, create positive employment spillovers (van Praag and Versloot, 2008). Their work time and these spillovers are affected by longer commutes (Viard and Fu, 2015). Florida (2004) argues that a ‘creative class’, about 30% of the U.S. workforce, sets their own hours and is critical to development of postindustrial U.S. cities. High human-capital and high-technology workers often have flexible schedules and exert a multiplier effect on local employment due to increased demand for local goods and services (Moretti, 2010; Moretti and Thulin, 2013). City growth is particularly sensitive to the presence of high human-capital workers due to spillovers from knowledge sharing (Jovanovic, 1992; Glaeser, 2003). For businesses, our results suggest that locating close to employees or easing their commutes can yield more time at work and likely higher productivity (Ross and Zenou, 2008; Gutiérrez-i-Puigarnau and van Ommeren, 2010b). Firms must compensate workers who have longer commutes with higher wages (Timothy and Wheaton, 2001; Fu and Ross, 2013) suggesting an added benefit for a firm in shortening commutes. Our results have important implications for the expansion of higher education in China. Total undergraduate enrollment in China increased from 2.0 million in 1998 to 8.7 million in 2010.49 The number of universities has not kept pace leading to higher enrollments: about 14,000 students per university in 2006 compared with 4000 in 1997. Universities have accommodated this expansion by increasing campus sizes—often by adding satellite campuses. As of 2009, more than 60 universities had established satellite campuses.50 Use of satellite campuses will lead to reduced teaching supply which, unless compensated for with a larger faculty, will reduce faculty–student interaction and diminish educational quality (Angrist and Lavy, 1999; Arias and Walker, 2004; DeGiorgi et al., 2012). Supplementary material Supplementary data for this paper are available at Journal of Economic Geography online. Acknowledgements We thank the university administrators who answered our questions and helped us with the data and thank the editor, two anonymous referees, Yong Suk Lee, Gan Li, Yi Lu, Hongliang Zhang and seminar participants at University of International Business and Economics, Renmin University of China, National University of Singapore, Chinese University of Hong Kong, 2012 China Economist Society Annual Conference in Kaifeng, Huazhong University of Science and Technology, Shandong University, Tsinghua University, Xiamen University, the 2012 International Conference on Industrial Economics at Zhejiang University, South China Normal University and the 2015 AREUEA-ASSA meetings for comments. All errors are our own. Footnotes 1 Daily work hours could also change as workers adjust their start and end times to avoid congested periods of the day as in ‘bottleneck’ theories examined by Vickrey (1969), Arnott et al. (1990, 1993) and Arnott et al. (2005). 2 Commute time may also influence labor supply through the labor participation rate. We are able to measure only the change in work time of already-employed workers. 3 Burchfield et al. (2006) emphasize quantifying the consequences of urban sprawl but note the necessity of using good instruments. 4 Many papers examine these equilibrium outcomes. Manning (2003) provides empirical evidence on the positive relationship between commute costs and wages. Gin and Sonstelie (1992) examine residential location changes due to commute cost changes. Zax (1991) and Zax and Kain (1996) empirically examine residence and job changes in response to commute cost changes. van Ommeren and Rietveld (2005) provide a theoretical relationship between commute time and wages in a job-matching model. White (1988) provides a theoretical model of location choice with endogenous residence and work locations. 5 Examples of monetary commute costs are gasoline, depreciation and tolls. Disutility includes discomfort from noise, pollution and effort. 6 We found no evidence of faculty quitting due to the longer commute. Leaving is costly—tenure-track faculty would have to break a 3-year contract with huge financial penalties if untenured or give up secure employment if tenured—and faculty could relocate to subsidized housing at the satellite campus once it opened in 2010. 7 For brevity, we will use the terms ‘teacher’ and ‘teachers’ interchangeably with ‘faculty member’ and ‘faculty’ even though our sample includes faculty who both teach and research. 8 We discuss later, the possibility of administrators intervening in the market and applying non-wage pressures. 9 As in the U.S. universities, academic year t spans fall semester of calendar year t to spring semester of calendar year t+1. 10 When we use the term ‘course’, we allow for the possibility of multiple sections. We, therefore, use the term ‘class’ to refer to a course with a single section or one section of a course with multiple sections. 11 This problem is faced by any study of work time since unofficial work is unobserved as is sharing of household chores and paid work within the household (Knowles, 2013). 12 ‘Millennials at Work: Reshaping the Workplace’, (PwC, 2011) estimates that 32% of millennials expect to have mainly flexible working hours in the future. 13 ‘There’s an App for That’, The Economist, 3 January 2015. In a survey of 5000 American workers, 34% engaged in some form of freelancing (‘Freelancing in America: A National Survey of the New Workforce’, Elance-oDesk, 2014). 14 Based on the 2005 China Inter-Census Population Survey, the survey classifies urban workers into one of four categories (hired, employer, self-employed and family business). We classify the first category as lacking flexibility and the last three as having flexibility. 15 The paper estimates a WTP based on job quits of €9.8 for the whole sample which includes all university workers and a WTP of €11.7 for non-faculty. Non-faculty workers represent 74% of the sample implying a WTP of €4.5 (35% of the average wage) for the faculty subsample. The paper cannot directly estimate a WTP for the faculty subsample because the effect of wage on job quits is not precisely estimated for the subsample. 16 For confidentiality reasons, we cannot identify the university nor can we provide references for the background information on the campus opening all of which were obtained from local newspapers. 17 Most Master’s programs in China take 3 years but some universities have 2-year programs. 18 The provision of faculty housing on or near the university campus is a common practice among Chinese public universities and most Chinese universities are public. Limited faculty offices and overnight apartments were available at the satellite campus and might limit the teaching time decline resulting from the longer commute. 19 Three categories of courses are offered to all class levels—‘sports’, ‘university’ and ‘double degree’ courses—which we call ‘other.’ ‘Sports’ courses teach athletics and ‘university’ courses concern culture or personal development. Courses are usually taught only to students within a major (corresponding to a university department) and only to a single class level. The exceptions to this, ‘double-degree’ courses, are offered to students outside the major and can be taken by any class level. 20 Faculty with a foreign Ph.D. and domestic faculty hired since 2006 have 3-year contracts. All other domestic faculty has permanent contracts. Regardless of contract length, research output affects promotion from an assistant to an associate or from an associate to full professor. 21 We do not believe that faculty anticipates wage changes because they are determined by human resources or a university-level committee and only then announced to faculty members. Therefore, they will not change their teaching schedules dynamically in anticipation of wage changes. 22 This is for department-specific courses. For ‘university-wide’ and ‘sports’ courses the minimum was 320 ‘class hours’ per year from 2001 to 2004 and 300 from 2005 onward. These courses are taught primarily by faculty in the English, sports and math departments. For the few teachers with a foreign Ph.D., the minimum was 160 ‘class hours’ per year. The university did not allow faculty to carry-forward or carry-back teaching credits and examined faculty workload year-by-year. The financial penalties for not meeting the teaching load were severe. 23 The activities available for meeting the minimum teaching load vary by rank. For example, only associate and full professors can supervise Master’s theses and only full professors can supervise Ph.D. theses. We check the robustness of our results to this by including faculty rank controls in some specifications. 24 One teacher ‘class hour’ yields x student ‘class hours’ where x is the class size being taught. Throughout the paper a ‘class hour’ refers to a teacher ‘class hour’ unless otherwise qualified. 25 To the extent that this constraint was binding during the transition years it was due to avoiding scheduling conflicts for students—the satellite campus was well below capacity without all class levels present. 26 As discussed earlier, teachers may have other work obligations besides teaching. Time spent on these is subsumed into leisure and income from these is subsumed in Y. Our model assumes equal ‘class hours’ across teaching days. In our data, they are unevenly distributed but this does not qualitatively change the model’s implications. 27 The problem should also include constraints on the maximum number of daily ‘class hours’ and for the minimum teaching load. For simplicity, we assume an interior solution. 28 Consistent with this, Connolly (2008) finds that male workers increase their work time on rainy days and decrease it the following day to equalize the marginal utility of leisure across days. 29 We drop class-year observations taught by faculty appearing in only 1 year that would be dropped with teacher fixed effects and those missing a teacher name. We also drop those taught by teachers under short-term contracts who are not permanent staff including visiting, retired, rehired (after retirement) and adjunct faculty. The number of observations for faculty rank information is slightly lower because we were unable to collect this information for some faculty. 30 For co-taught courses, we assign each teaching day to all teachers of the course. Although we would ideally allocate them proportional to the number of co-teachers this is impossible because these courses often meet multiple times per week and we do not observe which teacher teaches on which day. This makes it impossible to determine the overlap with each co-teacher’s other courses. Since we are unable to allocate them we make the conservative assumption to overstate teaching days and therefore commute days. 31 The Research Support Office ranks Chinese journals as ‘A1’, ‘A2’, ‘B1’, ‘B2’ or ‘C’ and English journals as ‘A’, ‘B’ or ‘C’. ‘A1’ and ‘A2’ Chinese journals are the top general interest and field journals in China. English ‘A’ journals are top general interest journals and ‘B’ are top field journals. Since publishing papers in English is difficult, we designate Chinese ‘A1’ and ‘A2’ and English ‘A’ and ‘B’ journals as ‘top’. All other journals we designate as ‘non-top’. 32 This drops to 69.9% in 2009 because the university re-classified some courses that were department-specific and offered separately to the four class levels as university-wide courses taught to the four class levels collectively. 33 This could also bias our results if ‘other’ courses were primarily taught by teachers with low or high commute-cost sensitivity and also primarily located at either the main or satellite campus. We have no means to check for this possibility. 34 This includes both not teaching courses that require commuting and consolidating classes in fewer teaching days to reduce the number of commute days. A similar issue arises in the environmental literature. In estimating the causal effect of pollution on health outcomes, it is important to control for the fact that people may avoid the impact of pollution by, for example, spending less time outside or wearing protective masks (Graff Zivin and Neidell, 2013). 35 Post-transition estimates are also affected by the transition of graduate and Ph.D. students to the satellite campus as well as the re-classification in 2009 of other courses shown in Columns 1 and 2 of the table in Online Appendix D. 36 This is not a tautology: annual ‘class hours’ per teacher (‘class hours’ summed across all courses taught by a teacher) can change on average across all teachers even as total student ‘class hours’ across all students stays the same as the university changes class sizes or number of faculty. 37 Day and time is available for classes meeting on weekends only beginning with the second semester of 2005. Before this, we have no way of determining whether a missing value is due to the class being taught on a weekend or some other reason. To be conservative, we include weekend days taught as a teaching day after second semester 2005 but drop missing values both prior to and after this. This will understate teaching days prior to 2006 and bias us against finding a decrease in annual teaching days due to the transition. 38 Online courses were not used by the university and teaching support and research fund guidelines did not change over time. 39 This is calculated using the midpoint method, a decrease of 21.9 ‘class hours’ per year and an average of 260.0 annual ‘class hours’ prior to the transition. The percentage change in commute time or distance is 200% since there was no commute time before the transition. 40 Many studies estimate the value of commute relative to work time. These only estimate the equilibrium trade-off and do not provide structural parameters for evaluating transport policy or labor market outcomes (Gibbons and Machin, 2006, 7). This literature has yielded a large range for the tradeoff: from 0.2 to 3 times the wage rate (Small, 1992; Calfee and Winston, 1998; Timothy and Wheaton, 2001; Small and Verhoef, 2007). Gibbons and Machin (2006) place the center of these estimates at 0.5. 41 Throughout the article we use an exchange rate as of August 2012: 6.35 USD:RMB. 42 Ross and Zenou (2008) find evidence for this among highly supervised blue-collar workers. 43 Because of changes in course names and numbers over time it is impossible to trace the evolution of specific courses over time to see how frequently the faculty who teach them change. 44 Blau and Kahn (2007) provide evidence of significant female labor supply changes but also conclude that female labor supply characteristics converge toward those of males over time. Black et al. (2014) find that female work time is more sensitive to commute costs as an equilibrium outcome. White (1986) finds evidence that male and female commute times respond differently to income, home ownership and presence of children. 45 We also tested for demographic differences in annual teaching days and daily ‘class hours’ and found no significant differences in either across gender or rank. 46 The number of teachers, students and student ‘class hours’ per student are clearly determined by the university. Teachers indirectly influence class size through their teaching quality and class requirements; however, the equilibrium effects are determined university-wide. 47 Temporary and adjunct faculty played a tiny role in the university’s response—the teacher-course observations we dropped for them numbered only 109 over the 10 years. 48 Parry and Bento 2001; Mayeres and Proost, 2001; and De Borger and van Dender, 2003 discuss endogenizing work time in theoretical models analyzing the welfare implications. 49 According to Ministry of Education data available at http://www.moe.gov.cn/. 50 ‘Development Patterns of College Towns in China’, Wei Zhou (2009), M.A. Thesis (in Chinese), Zhongshan University. References Angrist J. D. , Lavy V. ( 1999 ) Using Maimonides’ rule to estimate the effect of class size on scholastic achievement . The Quarterly Journal of Economics , 114 : 533 – 575 . Google Scholar CrossRef Search ADS Angrist J. D. , Pischke J. ( 2009 ) Mostly Harmless Econometrics . Princeton, NJ : Princeton University Press ,. Arias J. J. , Walker D. M. ( 2004 ) Additional evidence on the relationship between class size and student performance . The Journal of Economic Education , 35 : 311 – 329 . Google Scholar CrossRef Search ADS Arnott R. , de Palma A. , Lindsey R. ( 1990 ) Economics of a bottleneck . Journal of Urban Economics , 27 : 111 – 130 . Google Scholar CrossRef Search ADS Arnott R. , de Palma A. , Lindsey R. ( 1993 ) A structural model of peak-period congestion: a traffic bottleneck with elastic demand . The American Economic Review , 83 : 161 – 179 . Arnott R. , Tilman R. , Schöb R. ( 2005 ) Alleviating Urban Traffic Congestion . Cambridge, MA : The MIT Press . Becker G. S. ( 1965 ) A theory of the allocation of time . Economic Journal , 75 : 493 – 517 . Google Scholar CrossRef Search ADS Black D. A. , Kolesnikova N. , Taylor L. J. ( 2014 ) Why do so few women work in New York (and so many in Minneapolis)? Labor supply of married women across U.S. cities . Journal of Urban Economics , 79 : 59 – 71 . Google Scholar CrossRef Search ADS Blau F. D. , Kahn L. M. ( 2007 ) Changes in the labor supply behaviour of married women: 1980–2000 . Journal of Labor Economics , 25 : 393 – 438 . Google Scholar CrossRef Search ADS Brueckner J. ( 2001 ) Urban Sprawl: Lessons from Urban Economics . Washington, DC: Brookings Institute Press. Burchfield M. , Overman H. G. , Puga D. , Turner M. A. ( 2006 ) Causes of sprawl: a portrait from space . Quarterly Journal of Economics , 121 : 587 – 633 . Google Scholar CrossRef Search ADS Calfee J. , Winston C. ( 1998 ) The value of automobile travel time: implications for congestion policy . Journal of Public Economics , 69 : 83 – 102 . Google Scholar CrossRef Search ADS Cogan J. F. ( 1981 ) Fixed costs and labor supply . Econometrica , 49 : 945 – 963 . Google Scholar CrossRef Search ADS Connolly M. ( 2008 ) Here comes the rain again: weather and the intertemporal substitution of leisure . Journal of Labor Economics , 26 : 73 – 100 . Google Scholar CrossRef Search ADS Coviello D. , Ichino A. , Persico N. ( 2015 ) The inefficiency of worker time use . Journal of the European Economic Association , 13 : 906 – 947 . Google Scholar CrossRef Search ADS De Borger B. , van Dender K. ( 2003 ) Transport tax reform, commuting, and endogenous values of time . Journal of Urban Economics , 53 : 510 – 530 . Google Scholar CrossRef Search ADS DeGiorgi G. , Pellizzari M. , Woolston W. G. ( 2012 ) Class size and class heterogeneity . Journal of the European Economic Association , 10 : 795 – 830 . Google Scholar CrossRef Search ADS Duflo E. ( 2001 ) Schooling and labor market consequences of school construction in Indonesia: evidence from an unusual policy experiment . The American Economic Review , 91 : 795 – 813 . Google Scholar CrossRef Search ADS Duranton G. , Turner M. A. ( 2012 ) Urban growth and transportation . Review of Economic Studies , 79 : 1407 – 1440 . Google Scholar CrossRef Search ADS Florida R. ( 2004 ) Cities and the Creative Class . New York : Basic Books . Fu S. ( 2007 ) Smart café cities: testing human capital externalities in the Boston metropolitan area . Journal of Urban Economics , 61 : 86 – 111 . Google Scholar CrossRef Search ADS Fu S. , Ross S. L. ( 2013 ) Wage premia in employment clusters: how important is worker heterogeneity? Journal of Labor Economics , 31 : 271 – 304 . Google Scholar CrossRef Search ADS Gershenson S. ( 2013 ) The causal effect of commute time on labor supply: evidence from a natural experiment involving substitute teachers . Transportation Research Part A , 54 : 127 – 140 . Gibbons S. , Machin S. ( 2006 ) Transport and labour market linkages: empirical evidence, implications for policy and scope for further UK research, Background paper for the Eddington report to the Department of transport, UK. Gin A. , Sonstelie J. ( 1992 ) The streetcar and residential location in 19th century Philadelphia . Journal of Urban Economics , 32 : 92 – 107 . Google Scholar CrossRef Search ADS Glaeser E. ( 2003 ) The new economics of urban and regional growth. In Clark G. , Feldman M. , Gertler M. (eds) The Oxford Handbook of Economic Geography, pp. 83–98 . Oxford : Oxford University Press . Glaeser E. L. , Kahn M. E. ( 2001 ) Decentralized Employment and the Transformation of the American City . Washington, DC: Brookings Institute Press. Graff Zivin J. , Neidell M. ( 2013 ) Environment, health, and human capital . Journal of Economic Literature , 51 : 689 – 730 . Google Scholar CrossRef Search ADS Groot S. P. T. , de Groot H. L. F. , Veneri P. ( 2012 ) The educational bias in commuting patterns: micro-evidence for the Netherlands, Tinbergen Institute Discussion Papers 12–080/3: Tinbergen Institute, Available online at SSRN: https://ssrn.com/abstract=2119929. Gutiérrez-i-Puigarnau E. , van Ommeren J. N. ( 2009 ) Labour supply and commuting: implications for optimal road taxes, Proceedings of the BIVE-GIBET transport Research Day 2009 (Part II), edited by Macharis, Cathy and Turcksin, Laurence, pp. 621–641, Brussels: VUBPRESS Brussels University Press. Gutiérrez-i-Puigarnau E. , van Ommeren J. N. ( 2010a ) Labour supply and commuting . Journal of Urban Economics , 68 : 82 – 89 . Google Scholar CrossRef Search ADS Gutiérrez-i-Puigarnau E. , van Ommeren J. N. ( 2010b ) Are workers with a longer commute less productive? An empirical analysis of absenteeism . Regional Science and Urban Economics , 41 : 1 – 8 . Henderson J. V. , Mitra A. ( 1996 ) The new urban landscape developers and edge cities . Regional Science and Urban Economics , 26 : 613 – 643 . Google Scholar CrossRef Search ADS Hymel K. ( 2009 ) Does traffic congestion reduce employment growth? Journal of Urban Economics , 65 : 127 – 135 . Google Scholar CrossRef Search ADS Jovanovic B. ( 1992 ) Coordination and Spillovers, Working Papers 91–65. C.V. Starr Center for Applied Economics, New York University. Knowles J. A. ( 2013 ) Why are married men working so much? An aggregate analysis of intra-household bargaining and labour supply . Review of Economic Studies , 80 : 1055 – 1085 . Google Scholar CrossRef Search ADS Manning A. ( 2003 ) The real thin theory: monopsony in modern labour markets . Labour Economics , 10 : 105 – 131 . Google Scholar CrossRef Search ADS Mayeres I. , Proost S. ( 2001 ) Marginal tax reform, externalities and income distribution . Journal of Public Economics , 79 : 343 – 363 . Google Scholar CrossRef Search ADS Moretti E. ( 2004 ) Human capital externalities in cities. In Henderson J. V. , Thisse J. F. (eds) Handbook of Regional and Urban Economics , Vol. 4 , pp. 2243 – 2291 . North Holland : Elsevier . Moretti E. ( 2010 ) Local multipliers . American Economic Review , 100 : 1 – 7 . Google Scholar CrossRef Search ADS Moretti E. ( 2012 ) The New Geography of Jobs . Boston and New York : Houghton Mifflin Harcourt . Moretti E. , Thulin P. ( 2013 ) Local multipliers and human capital in the United States and Sweden . Industrial and Corporate Change , 22 : 339 – 362 . Google Scholar CrossRef Search ADS Mulalic I. , van Ommeren J. N. , Pilegaard N. ( 2014 ) Wages and commuting: quasi-natural experiments’ evidence from firms that relocate . The Economic Journal , 124 : 1086 – 1105 . Google Scholar CrossRef Search ADS Parry I. W. H. , Bento A. ( 2001 ) Revenue recycling and the welfare effects of road pricing . Scandinavian Journal of Economics , 103 : 645 – 671 . Google Scholar CrossRef Search ADS Rosenthal S. S. , Strange W. C. ( 2008 ) The attenuation of human capital spillovers . Journal of Urban Economics , 64 : 373 – 389 . Google Scholar CrossRef Search ADS Ross S. L. , Zenou Y. ( 2008 ) Are shirking and leisure substitutable? An empirical test of efficiency wages based on urban economic theory . Regional Science and Urban Economics , 38 : 498 – 517 . Google Scholar CrossRef Search ADS Russo G. , van Ommeren J. N. , Rietveld P. ( 2012 ) The university workers’ willingness to pay for commuting . Transportation , 39 : 1121 – 1132 . Google Scholar CrossRef Search ADS Small K. ( 1992 ) Urban Transportation Economics: Fundamentals of Pure and Applied Economics , Vol. 51 . New York : Harwood Academic Publishers . Small K. , Verhoef E. ( 2007 ) The Economics of Urban Transportation . New York : Routledge . Timothy D. , Wheaton W . ( 2001 ) Intra-urban wage variation, employment location, and commuting times . Journal of Urban Economics , 50 : 338 – 366 . Google Scholar CrossRef Search ADS van Ommeren J. N. , Rietveld P. ( 2005 ) The commuting time paradox . Journal of Urban Economics , 58 : 437 – 454 . Google Scholar CrossRef Search ADS van Praag C. M. , Versloot P. H. ( 2008 ) The economic benefits and costs of entrepreneurship: a review of the research . Foundations and Trends in Entrepreneurship , 4 : 65 – 154 . Google Scholar CrossRef Search ADS Viard V. B. , Fu S. ( 2015 ) The effect of Beijing’s driving restrictions on pollution and economic activity . Journal of Public Economics , 125 : 98 – 115 . Google Scholar CrossRef Search ADS Vickrey W. S. ( 1969 ) Congestion theory and transport investment . American Economic Review , 59 : 251 – 260 . White M. J. ( 1986 ) Sex differences in urban commuting patterns . The American Economic Review , 76 : 368 – 372 . White M. J. ( 1988 ) Location choice and commuting behaviour in cities with decentralized employment . Journal of Urban Economics , 24 : 129 – 152 . Google Scholar CrossRef Search ADS Zax J. S. ( 1991 ) The substitution between moves and quits . The Economic Journal , 101 : 1510 – 1521 . Google Scholar CrossRef Search ADS Zax J. S. , Kain J. F. ( 1996 ) Moving to the suburbs: do relocating companies leave their black employees behind? Journal of Labor Economics , 14 : 472 – 504 . Google Scholar CrossRef Search ADS © The Author(s) (2018). Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oup.com This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal of Economic Geography – Oxford University Press

**Published: ** Mar 13, 2018

Loading...

personal research library

It’s your single place to instantly

**discover** and **read** the research

that matters to you.

Enjoy **affordable access** to

over 18 million articles from more than

**15,000 peer-reviewed journals**.

All for just $49/month

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Read from thousands of the leading scholarly journals from *SpringerNature*, *Elsevier*, *Wiley-Blackwell*, *Oxford University Press* and more.

All the latest content is available, no embargo periods.

## “Hi guys, I cannot tell you how much I love this resource. Incredible. I really believe you've hit the nail on the head with this site in regards to solving the research-purchase issue.”

Daniel C.

## “Whoa! It’s like Spotify but for academic articles.”

@Phil_Robichaud

## “I must say, @deepdyve is a fabulous solution to the independent researcher's problem of #access to #information.”

@deepthiw

## “My last article couldn't be possible without the platform @deepdyve that makes journal papers cheaper.”

@JoseServera

DeepDyve ## Freelancer | DeepDyve ## Pro | |
---|---|---|

Price | FREE | $49/month |

Save searches from | ||

Create lists to | ||

Export lists, citations | ||

Read DeepDyve articles | Abstract access only | Unlimited access to over |

20 pages / month | ||

PDF Discount | 20% off | |

Read and print from thousands of top scholarly journals.

System error. Please try again!

or

By signing up, you agree to DeepDyve’s Terms of Service and Privacy Policy.

Already have an account? Log in

Bookmark this article. You can see your Bookmarks on your DeepDyve Library.

To save an article, **log in** first, or **sign up** for a DeepDyve account if you don’t already have one.

All DeepDyve websites use cookies to improve your online experience. They were placed on your computer when you launched this website. You can change your cookie settings through your browser.

ok to continue