Worker Absenteeism: Peer Influences, Monitoring, and Job Flexibility∗ Per Johansson†,‡

Arizo Karimi§

J. Peter Nilsson¶

May 3, 2018

Abstract We study the presence of other-regarding preferences in the workplace by exploiting a randomized experiment that changed the monitoring of workers’ health during sick leave. We show that workers’ response to an increase in co-worker shirking, induced by the experiment, is much stronger than the response to a decrease in co-worker shirking. The asymmetric spillover effects are consistent with evidence of fairness concerns documented in laboratory experiments. Moreover, we find that the spillover effect is driven by workers with highly flexible and autonomous jobs, suggesting that co-worker monitoring may be at least as important as formal monitoring in alleviating shirking. Keywords: Productivity; social preferences; fairness; sick leave. JEL-codes: J22, J16, C93, I13.

∗ We are grateful for valuable comments on earlier versions of this paper from Anna Dreber Almenberg, Hans Grönqvist, Lena Hensvik, V. Joseph Hotz, Peter Skogman Thoursie, Olof Åslund, and two anonymous referees. Comments from seminar participants at the Swedish National Conference 2014, UCLS Spring Meeting 2014, and The University of Duisburg-Essen are also gratefully acknowledged. Arizo Karimi gratefully acknowledges financial support from the Jan Wallander and Tom Hedelius foundation. † The Department of Statistics, Uppsala University; Institute for Evaluation of Labor Market and Education Policy (IFAU); Uppsala Center for Labor Studies (UCLS), and IZA. ‡ Corresponding Author: Department of Statistics, Uppsala University, Box 513, SE-751 20, Uppsala. Phone: +46 (0)18 471 51 46. E-mail: [email protected]. § Department of Economics, Uppsala University; Uppsala Center for Labor Studies (UCLS), and Institute for Evaluation of Labor Market and Education Policy (IFAU). ¶ Institute for International Economic Studies (IIES), Stockholm University, and Uppsala Center for Labor Studies (UCLS).

1

Introduction

A substantial amount of theoretical and empirical work has established that workers not only care about maximizing their material payoffs, but are willing to forego compensation for e.g., fairness and reciprocity concerns. Moreover, social interactions between co-workers is a salient feature of many workplaces, potentially amplifying the importance of social preferences for workers’ behavior through peer effects. Thus, the existence of peer effects in the workplace has potential implications for whether an efficient workplace design is one characterized by teamwork or one in which employees work alone. We empirically assess the presence of other-regarding preferences in the workplace and how such preferences vary with job characteristics by exploiting a large-scale randomized field experiment implemented in Sweden’s second largest city, Göteborg.1 The experiment increased work absence incentives by relaxing formal monitoring of workers’ health status during sick leave spells for around half of the 60,000 workers employed in more than 3,000 workplaces in our sample. Workers assigned to the treatment group were allowed to be absent from work due to illness for 14 days before needing to hand in a certificate from a doctor verifying reduced work capacity due to illness. The control group was as usual required to hand in the certificate after 7 days.2 The direct effect of the experiment was studied in Hartman et al. (2013) and potential spillover effects in the workplace was studied in Hesselius et al. (2009). We add to the previous spillover paper in several ways. In order to firmly provide evidence of the mechanisms of the spillover effects found in Hesselius et al. (2009) we add data from a similar but reversed experiment, implemented at the same time in a different part of the country. Moreover, we add survey data on job characteristics which provides another source for pinpointing the mechanisms of the spillover effects. Our paper adds to the growing literature on productivity spillovers within the workplace using matched worker-firm data, many times from single firms (see e.g. Bandiera et al., 2005, 2010; Mas and Moretti, 2009; Falk and Ichino, 2006). Our empirical analysis proceeds in two steps. First, we show that the experiment leads to a sharp increase in absence among the treated workers. In particular, as shown by Hartman et al. (2013), exits from sickness spells are conspicuously bunched on the day before health screening was required for 1 Following the literature, we define other-regarding preferences as deviations from pure self-interest motives, e.g. due to altruism, fairness concerns, reciprocity, and inequity aversion (see e.g. Fehr and Schmidt, 1999; Rabin, 1993). 2 One major advantage of the experiment is that it addresses many of the severe identification problems associated with the identification and estimation of peer effects (see e.g. Manski, 1993; Angrist, 2014, for a description of the difficulties in the estimation of social interaction effects). In particular, because relaxed monitoring of work absence was randomly assigned to a subset of employees within a workplace, we can estimate how the non-treated respond to having a large share of treated co-workers, which allows us to draw inference on how the reference group affects individual behavior and not the other way around.

2

continued sick leave. In the treatment group, the spike in the exit rate shifts from day 7 to day 14 (see Figure 1). This stark shift clearly indicates that postponing formal monitoring increases shirking among treated workers.3 Moreover, we show that the share of treated workers varies substantially across workplaces and that during the experiment, the share of treated co-workers is positively related to individuals’ work absence. Before the experiment, neither future assignment to treatment, nor the share of treated co-workers is related to pre-experiment absence levels. That work absence, conditional on own treatment status, is increasing in the share of co-workers with additional unmonitored sick leave provides strong evidence for spillover effects in shirking among co-workers. An explanation for such spillover effects could be that workers have otherregarding preferences in the form of fairness concerns. If the workload is affected by the increase in unmotivated absence among the treated workers, it is conceivable that the non-treated workers feel unfairly treated. There are numerous ways in which shirking co-workers could be punished, but a natural way to reciprocate may be to increase one’s own absence. The key assumptions required to exploit the experiment to empirically assess the strength of other-regarding preferences in the workplace is that the spillover effect indeed reflects preferences for fairness and reciprocity. However, alternative mechanisms, including knowledge transfers about the sick-leave system and/or leisure complementarities, could also lead to positive effect estimates. We present two pieces of evidence in support of fairness concerns being the key underlying mechanism. First, we find that the share of treated peers affects control group workers, but not treated workers. Since there is no obvious reason why increased information or joint leisure should only affect the non-treated, we follow Hesselius et al. (2009) and interpret the heterogeneity in the spillover effects across individual treatment status as suggesting that leisure complementarities or information sharing are not main underlying mechanisms. Instead, fairness concerns manifested as negative reciprocity or conditional cooperation seems like more plausible explanations. This hypothesis is consistent with the findings in which we exploit data from the reversed experiment implemented at the same time in another region, Jämtland. The experiment in Göteborg implied a relaxation of monitoring during an absence spell, whereas in Jämtland it implied stricter monitoring by shifting the health screening from day 14 to day 7 for half of the population. A wellestablished result from laboratory experiments is that the response to unfair treatment (the increase in co-worker shirking) is much stronger than the response to kind behavior (same-sized decrease in co-worker shirking) (see e.g. Offerman, 2002; Fehr et al., 2009). Outside the lab, Mas (2006) documents large reductions in police performance following unexpected wage cuts, but that performance 3 See also Nagin et al. (2002) for field experimental evidence on the impacts of monitoring in workers’ shirking behavior.

3

is hardly affected by wage-increases. These asymmetric effects provide us with the prior that if preferences for fairness and reciprocity are a key mechanism behind the spillover effect, the decrease in shirking among treated co-workers (Jämtland) should not generate nearly as large spillovers as those from an increase in shirking (Göteborg). Consistent with this prediction, the estimated spillover effect is much smaller and not statistically significant among those whose colleagues experienced stricter monitoring. In the second part of the paper, we present new evidence on how the spillover effect in our experiment varies with job characteristics. In particular, employees working in groups or in designs with extensive communication in the workplace may be more exposed to co-worker monitoring, potentially allowing peer pressure to reduce shirking behavior. We test whether this conjecture is valid in a general setting with varying job tasks, by estimating heterogeneous treatment and spillover effects by job flexibility. Workers with highly flexible jobs are arguably less likely to work in teams, and thus less exposed to co-worker monitoring. Such workers are also less likely to impose externalities on co-workers when shirking. Thus, this analysis is also an indirect test of whether the other-regarding preferences detected in our experiment is consistent with conditional cooperation, i.e., increased shirking because one’s co-workers are not contributing to their fair share, or negative reciprocity, i.e., increased shirking to punish shirking co-workers. To this end, we exploit a supplementary data set that provides survey responses to questions about job flexibility. We use multiple variables measuring e.g., temporal job flexibility, whether the job allows working from home, the extent of freedom to decide how to structure and perform the tasks, etc., in a principal components analysis to create an index of job flexibility. We find that the spillover effect is fully accounted for by workers with highly flexible jobs, suggesting that co-worker monitoring may be at least as important as formal monitoring in alleviating shirking in the workplace. These results also suggest that the fairness concerns driving the estimated spillover effects in this experiment are due to conditional cooperation, rather than (negative) reciprocal behavior towards shirking co-workers. In summary, we confirm that one well-documented result from the laboratory is also present in the labor market: the positive response to fair treatment is small, while the negative response to unfair treatment is large, in particular when the extent of co-worker monitoring is low. The remainder of the paper is organized as follows. Section 2 provides information on the institutional setting and the experiment. In Section 3, we describe the data and the empirical strategy. In Sections 4 and 5 we present the main results and the findings on heterogeneous effects by job characteristics. Section 6 concludes the paper.

4

2

Background

In Sweden, foregone income due to temporary illness is covered by public sickness insurance. During the experimental period, the replacement rate was 90 percent of previous earnings up to a cap, and workers were covered from the first absence day. In addition to the public insurance, most Swedish workers are covered by top-up sickness insurance regulated in agreements between the unions and employers’ confederations. Total compensation could be as high as 100 percent. The sickness insurance covers all workers, unemployed, and students. At the time of the experiment, the sick-pay was paid by the Swedish government. This implies that for the employer, the costs were only associated with finding and hiring replacement workers and/or foregone productivity. There was no limit to how long or how often sickness benefits could be claimed. For the worker, there is a high degree of discretion when reporting sick. A sick spell starts when the worker calls the employer and the public insurance office to report sick. During the first 7 days of an absence spell there is no formal monitoring. From the 8th day, however, the worker must present a medical doctor’s certificate confirming reduced work capacity due to temporary illness for eligibility for continued sick leave. The certificate is reviewed by the public insurance office, which approves or, in very few cases, declines further sick leave.

2.1

The Experiment

In the fall of 1988, the regional social insurance boards in Göteborg – Sweden’s second-largest city, and Jämtland – a large, sparsely populated region in the North of Sweden, implemented a randomized social experiment in the timing of the requirement of the medical certificate. For those born on even dates, a medical certificate was not required until the 15th day of an absence spell during the experiment. Those born on an uneven date were required to confirm their eligibility status on the 8th day. The idea was that, by postponing the timing of the certificate requirement, public health care spending would be reduced. This belief was based on a presumption that doctors prescribed unnecessarily long absences, and that more workers would have time to recover before having to visit a doctor. The monitoring rules differed between Göteborg and Jämtland before the experiment was implemented. In Jämtland, the two-week rule for non-monitored absence was already in place for all residents since 1987. In Göteborg, however, the usual (national) rule of one-week non-monitored absence applied. Thus, in Göteborg, the experiment implied less strict eligibility screening for half of the population, while in Jämtland the experiment implied stricter eligibility screening for those born on an uneven date. 5

The experiment was preceded by a large information campaign via mass media, and pamphlets and posters were distributed to workplaces. In addition, information was printed on the form submitted to the insurance office to receive sickness benefits. Thus, all employers, employees, and doctors were informed about the experiment. Hartman et al. (2013) show that, in contrast to what was expected by the Social Insurance Board, there was a substantial increase in the work absence spell duration due to the experiment. The upper and lower panels of Figure 1 shows the hazard rate out of sickness absence for the treatment and the control group in Göteborg and Jämtland, respectively, during the experiment. In the control group, the spike is shifted to day 14. Hence, people typically returned to work the day before their health status was to be formally monitored.

3

Data and Empirical Strategy

We use a matched employer-employee data set constructed by combining population-wide administrative registers, maintained by Statistics Sweden. The data contains individual background characteristics and information on sickness absence spells during 1987 and 1988, as well as workplace characteristics - such as the composition of the workforce in terms of demographic characteristics, industry affiliation, and number of employees. We restrict attention to individuals residing in Göteborg municipality and employed at workplaces with 10-100 employees.4 Similarly, for the reversed experiment in Jämtland, we focus on individuals residing in the region and employed at workplaces with 10-100 employees. Among employees residing in Göteborg and Jämtland, the share of treated co-workers varies substantially, as shown in Figure 2. The average proportion treated co-workers is, however, higher in Jämtland (0.45) than in Göteborg (0.3). The difference in the share treated co-workers for the typical employee in the respective samples is due to that the variation in the share treated stems not only from the randomization, but also from co-workers who live outside the experiment region (henceforth “commuters”) and who are therefore not assigned to either the treatment or the control group. As seen from the summary statistics presented in Table A.1, the workplace of the typical individual in the Göteborg sample has a larger share of commuters than does the workplace of the typical individual in the Jämtland sample. In addition to share commuting co-workers, the labor markets in Göteborg and Jämtland differs in that the educational and earnings levels are higher, on 4 The restriction on workplace size is to ensure that co-workers constitute a reasonably well-defined peer group. Workplaces are defined based on the physical address and the firm to which it belongs (one firm can have several establishments). A few individuals have multiple workplaces and in those cases, for simplicity, we assign individuals to the workplace from which they earn their main income.

6

average, in Göteborg. However, as established in Hartman et al. (2013), despite these differences, the estimated effects of the timing of the doctor’s certificate on sickness absence in the two experiments was very similar, which suggests that the impacts of the main experiment can be generalized. To establish the presence of social spillovers, we test whether the share of treated co-workers affects individual work absence using the following empirical specification, estimated on individuals in the treatment area:

yig = β 0 + β 1 Tig + β 2 π(−i) g + xig β 3 + z(−i) g β 4 + eig

(1)

where yig is the number of sick leave days during the fall of 1988 for individual i in workplace g, Tig equals one if the individual is treated, and π(−i) g is the share of treated co-workers, excluding individual i. xig is a vector of individual characteristics; age, gender, earnings, and education. z(−i) g is a vector of workplace characteristics including dummy variables for the share of commuters, industry dummies, number of employees, share of females, average education, average age, and lagged workplace average characteristics. e is assumed to be an idiosyncratic error term. The reported standard errors are clustered at the workplace level. β 1 measures the direct effect of treatment for individual i in workplace g. The parameter of interest is β 2 , which captures the effect of the share of treated co-workers. The intuition is simple. In the absence of social spillovers, individuals’ work absence decisions should not be affected by the share of treated co-workers. The randomization balances unobserved covariates between the treated and the control group (see Table A.2 for evidence that observed covariates are balanced across treatment status). However, the variation in the share of treated co-workers also stems from the share of commuters. If workplaces with a large share of commuters differ in factors that also correlate with work absence, then β 2 will not only capture the social spillover effects induced by the experiment. Thus, we control for lagged workplace average sickness absence as well as the share of commuting co-workers in all specifications. Still, there might be some concern that controlling for these important workplace characteristics is not sufficient. Unobserved factors could potentially still bias the estimates. In order to address this concern, we also present estimates using data from the year preceding the experiment. If our main results simply reflect unobserved differences between individuals in workplaces with different shares of treated co-workers, we would expect to see similar effects before and during the experiment. In the next section, we discuss possible interpretations for the coefficients on β 1 and β 2 .

7

4

Fairness Concerns, Joint Leisure, Information Sharing?

This section provides evidence on the important question of whether we can really distinguish other regarding preferences from alternative explanations that could also lead to positive spillover effects across co-workers. For example, knowledge about the sick-leave system might have been incomplete before the experiment, and having many treated co-workers increases the probability of acquiring such information. Alternatively, absence decisions may change in response to co-workers’ absence because of complementarities in the utility of leisure (Lindbeck et al., 1999). However, we believe that fairness concerns constitute the key mechanism behind the spillover effect in this setting. This conclusion builds on two separate results, presented in Table 1: (a) Differential Spillover Effects by Treatment Status. Panel A shows the estimates from model (1) for all workers, treated and control group, respectively. Columns (1) and (2) show that for treated workers, the average number of days in absence spells shorter than 8 days decreases, while the number of days in spells shorter than 15 days increases for the treated workers. Moreover, the share of treated co-workers is positively related to co-worker absence. Columns (3) to (6) show the results separately for the treated and control groups. Note that the share of treated co-workers does not significantly affect individuals in the treatment group. However, the share of treated co-workers is significantly related to work absence in the control group, and the point estimates are also substantially larger than in the treatment group. Panel B shows that in the year prior to the experiment, there is no significant effect of treatment or the share of treated co-workers on workers’ sickness absence, except for a weakly significant effect on the (wrong-signed) coefficient in column (3). Following Hesselius et al. (2009), we interpret the difference in the estimated effects across individual treatment status as suggesting that joint leisure or information sharing are not the main mechanisms. There is no obvious reason why the treatment or control workers should react so differently to increased information, or the possibility to enjoy leisure time together. Instead, other-regarding preferences in the form of reciprocity and/or fairness concerns seem like a more plausible explanation. For example, in cooperative settings where workers have to fill in for each other, negative reciprocity could materialize as increased absence as a response to co-worker shirking. Thus, if workers care about fairness, control group workers could, as a response to an increase in shirking among their treated peers, increase their own absence to punish the shirking co-worker. Alternatively, one could interpret the fairness concerns in terms of conditional cooperation, i.e., decreasing one’s own effort as a response to co-workers not contributing to their fair share. (b) Exploiting the Reversed Experiment. We also provide evidence on the underlying mechanism by exploiting data from the reversed experiment that was implemented at the same time in 8

another part of the country, Jämtland. Before the experiment, workers in the Jämtland region needed to provide a doctor’s certificate before the 15th day of an absence spell. During the experiment, those born on an uneven date needed to provide a certificate already on the 8th day, which led to shorter absence spells in the treatment group (Hartman et al., 2013). Thus, the experiment in Göteborg implied a relaxation of health screening for half of the population during the absence spell, whereas in Jämtland it implied stricter health screening for half the population. Lab experiments often find that negative reciprocity is a much stronger motivator than positive reciprocity/altruism (Fehr et al., 2009). Such asymmetries have been suggested to lead to differential behavioral effects among workers experiencing e.g. wage cuts and wage increases (Offerman, 2002). Moreover, in the field, Card et al. (2012) find that workers with salaries below the median in their co-worker peer group report lower job satisfaction after their peers’ salaries were disclosed, whereas those above the median report no higher satisfaction. Also, in a field experiment on wages and worker effort, Cohn et al. (2015) find that wage increases affects effort mainly through the removal of perceived unfairness, or the elimination of negative reciprocity, whereas no effects were found of wage increases among workers who perceived being adequately paid at the base wage. The analogue in our setting is that responses to unfairness (an increase in shirking among co-workers) are stronger than the response to kind behavior (a decrease in shirking among co-workers). Based on these findings, our prior is that if fairness concerns constitute the key underlying mechanism, the share of treated co-workers in Jämtland (a decrease in shirking) should have a differential effect from the share of treated in Göteborg (an increase in shirking). Instead, if information sharing or joint leisure were the key mechanisms, we would expect a symmetric response from the experiments in Göteborg and Jämtland. Panel C presents the results for the Jämtland experiment. The direct effects of monitoring on treated workers are of similar size but, as expected, in the opposite direction than for Göteborg (see columns (1) and (2)). The average number of days in spells shorter than 8 days increases, while the number of days in spells shorter than 15 days decreases. More interestingly, unlike in Göteborg, columns (3)-(6) also show that the share of treated co-workers in Jämtland is: (i) not systematically negatively related to work absence; (ii) the same for treated and controls and (iii) small and never statistically significant. In Table A.3, we also provide the results from a placebo test for Jämtland, where we estimate equation (1) on sickness absence in the year preceding the experiment. As seen, just as in Göteborg, there are no effects either of treatment or of the share treated on pre-experiment absence. In summary, these results are consistent with the interpretation of the results under (a), and with

9

our prior that fairness concerns should generate asymmetric spillover effects in Göteborg and Jämtland. Together the results provide support for the fact that other regarding preferences in the form of fairness concerns constitute a key underlying mechanism behind the spillover effects.5 Finally, we note that while our findings suggest that the spillover effects are driven by fairness concerns and not by non-social peer effects such as information sharing, we are unable to explicitly differentiate between different types of fairness concerns, e.g., negative reciprocity or conditional cooperation. Moreover, the asymmetry in the spillover effect across treatment status suggests that co-workers increase their shirking only if both my co-workers shirk and if they do so because they have better opportunities to shirk than they themselves have. While less general a conclusion than one in which all workers respond similarly to an increased co-worker shirking, we believe that our findings are still relevant as it has implications for how – real or perceived – unmotivated differences at the workplace, for example in working conditions or remuneration, across co-workers can affect worker morale and effort. We return to this discussion in more detail in the next section, where we analyze the spillover effect by job characteristics.

5

Heterogeneous Effects by Job Types - Co-worker Monitoring

As noted by Falk and Ichino (2006), the existence of peer effects in productivity raises questions of whether an efficient design of the workplace is one in which employees work in groups or alone. Their results from a controlled field experiment show positive spillover effects in worker productivity, thus suggesting that employees should work in groups rather than alone. Moreover, employees working in groups or in designs with otherwise extensive communication are more exposed to coworker monitoring, potentially allowing peer pressure to reduce shirking behavior. We test whether this conjecture is valid in a general setting with varying job tasks, by exploiting a supplementary data set to estimate heterogeneous treatment and spillover effects by job characteristics, in particular job flexibility in terms of both working time and the freedom to decide how to structure, plan, and perform one’s tasks. Workers with highly flexible jobs are arguably less likely to work in teams, and are thus less exposed to co-worker monitoring. Such workers are also less likely to impose externalities on coworkers when shirking. Thus, this analysis also allows for an indirect test of whether the otherregarding preferences detected in our experiment is consistent with conditional cooperation, i.e., 5 An additional channel through which an increased co-worker absence might affect individual absence is through direct negative effects on health, e.g. from the stress of facing an increased workload. Although we cannot rule such an channel out, we argue that the short duration of the experiment, and our focus on short-term sickness absence most likely diminishes the risk of such effects to be a main driver of the spillover estimates. Also, we note that negative health externalities would likely yield a similar effect on both treated and non-treated.

10

increased shirking because one’s co-workers are not contributing to their fair share, or negative reciprocity, i.e., increased shirking to punish shirking co-workers. The latter channel requires that one is indeed able to punish co-workers through one’s own absenteeism behavior, which is only relevant if one’s own absenteeism causes an externality on the co-workers’ workload. We investigate potential heterogenous effects across job characteristics by adding information from a supplementary data set; the Swedish Living Conditions Surveys (ULF/SILC). The survey is conducted annually and covers 11,000–13,000 nationally representative individuals per year. Respondents are asked about their health, financial situation, housing arrangements, and the characteristics of their jobs. We extract information on job flexibility using five variables, which measure the extent to which workers are free to decide (a) when to start and end their workday, (b) how to structure their work, (c) how to plan their tasks, (d) how to allocate their time across tasks, and (e) physical location of work (e.g., work from home).6 The ULF survey does not contain workplace identifiers, so we are unable to match the job characteristics to individual workers or workplaces. Instead, we collapse these job characteristics by industry affiliation, which are then matched to our analysis data. The variation in job characteristics derived from ULF is thus at the industry level. In Table A.4, we show that the industry composition differs somewhat between Göteborg and Jämtland, but that all industry groups are represented in both regions. To capture the aspects of job flexibility in a parsimonious way, we construct a job flexibility index using principal component analysis, separately for Göteborg and Jämtland. Table A.5 in the Appendix reports the results from our principal component analysis to generate the flexibility index for Göteborg. The first factor explains 68 percent of the variation in the 5 flexibility variables, with each variable having loadings that range from 0.39 to 0.49. Similarly, Table A.6 shows that the for Jämtland sample, the first factor explains 66 percent of the variation in the same variables. We use the first factor for our index of job flexibility, and classify jobs as being flexible if they have an index score above the median score, and non-flexible if they have an index score below the median. We then estimate our main specification separately for high- and low-flexibility jobs. The results are presented in Table 2 and show that the effect of being assigned to treatment is higher in low flexibility jobs, suggesting that workers who have less freedom to decide how and when to allocate their work increase shirking more when formal monitoring is lower. However, the spillover effect is fully accounted for by workers with high flexibility jobs (column 2), and the effect is only driven by non-treated workers with highly flexible jobs (column 6). In the reversed experiment (Panel B of Table 2), there is no significantly different effect of monitoring on absence between workers with high 6 This

data is used in a similar way by Hotz et al. (2018) to create an index of job flexibility in Sweden.

11

and low flexibility jobs. Consistent with the conjectures stated above, we thus find that the spillover effects are fully accounted for by workers with highly flexible jobs, suggesting that co-worker monitoring and peer pressure may be at least as important as formal monitoring in alleviating shirking in the workplace. These results also suggest that the fairness concerns driving the estimated spillover effects in this experiment are due to conditional cooperation, rather (negative) reciprocal behavior towards shirking co-workers.7

6

Conclusion

In this paper, we investigate the impact of other-regarding preferences on workers’ shirking behavior. We exploit a large-scale randomized field experiment that increased work absence incentives by relaxing formal monitoring of workers’ health status, and a reversed experiment that increased formal monitoring. In the first part of the paper, we show that workers’ response to an increase in co-worker shirking, induced by the experiment, is much stronger than the response to a decrease in co-workers shirking. These asymmetric responses are consistent with the evidence of vertical fairness concerns - e.g. workers visavis managers - documented in the lab and in the field; reciprocal behavior to fair treatment is typically small, while the negative effects of unfair treatment on behavior are large (Offerman, 2002; Mas, 2006; Fehr et al., 2009). A plausible interpretation is therefore that horizontal fairness concerns, i.e., between co-workers, also play a role for worker effort and productivity in the labor market. Our findings also point to the difficulties in performing field experiments, and possibly also in policy implementation. First, the possibility of unintended spillover effects might complicate the analysis and interpretation of policy (experiments). Moreover, if policy interventions are perceived as unfair, by either those subject to the changes or by individuals without access, the behavioral responses may lead to inefficient outcomes. In the second part of the paper, we present evidence on how the spillover effect varies with job characteristics. In particular, we create an index of job flexibility, capturing the extent to which workers are free to decide whether to work from home, when to start and end their work-day, and how to structure; plan, and execute their tasks. Workers with highly flexible jobs are arguably less likely to work in teams, and are thus less exposed to co-worker monitoring and impose less externalities on co-workers when shirking. Consistent with this conjecture, we find that the spillover effects are 7 Another potential reason for differences in spillover effects by job flexibility is that workers in flexible jobs might not observe their co-workers behavior to the same extent, if they are less likely to work in teams, e.g., than those with nonflexible jobs. However, such a channel would work against the results that we find, in which case our estimates in the group of workers with flexible jobs would be attenuated towards zero.

12

driven by workers with highly flexible jobs, suggesting that co-worker monitoring and peer pressure may be at least as important as formal monitoring in alleviating shirking in the workplace. This, in turn, implies that a more efficient workplace design - in terms of reducing negative productivity spillovers - is one characterized by teamwork, visibility and communication within the workplace. In summary, we confirm that one well-documented result from the laboratory is also present in the labor market: the positive response to fair treatment is small, while the negative response to unfair treatment is large, in particular when the extent of co-worker monitoring is low. Note that our real-world setting confirms this behavior even when the cost of punishing unfair treatment, i.e., when the price of fairness (Eckel and Grossman, 2008), may be high; increasing shirking to punish one’s co-workers may well be costly and damaging to the employer-employee relationship.

13

References A NGRIST, J. D. (2014). The perils of peer effects. Labour Economics, 30, 98–108. B ANDIERA , O., B ARANKAY, I. and R ASUL , I. (2005). Social preferences and the response to incentives: Evidence from personnel data. The Quarterly Journal of Economics, 120 (3), 917–962. —, — and — (2010). Social incentives in the workplace. The Review of Economic Studies, 77 (2), 417–458. C ARD , D., M AS , A., M ORETTI , E. and S AEZ , E. (2012). Inequality at work: The effect of peer salaries on job satisfaction. The American Economic Review, 102 (6), 2981 – 3003. C OHN , A., F EHR , E. and G OETTE , L. (2015). Fair wages and effort provision: Combining evidence from a choice experiment and a field experiment. Management Science, 61 (8), 1777–1794. E CKEL , C. C. and G ROSSMAN , P. J. (2008). Chapter 113 men, women and risk aversion: Experimental evidence. Handbook of Experimental Economics Results, vol. 1, Elsevier, pp. 1061 – 1073. FALK , A. and I CHINO , A. (2006). Clean evidence on peer effects. Journal of Labor Economics, 24 (1), 39–57. F EHR , E., G OETTE , L. and Z EHNDER , C. (2009). A behavioral account of the labor market: The role of fairness concerns. Annual Review of Economics, 1 (1), 355–384. — and S CHMIDT, K. M. (1999). A theory of fairness, competition, and cooperation. The quarterly journal of economics, 114 (3), 817–868. H ARTMAN , L., H ESSELIUS , P. and J OHANSSON , P. (2013). Effects of eligibility screening in the sickness insurance: Evidence from a field experiment. Labour Economics, 20 (0), 48–56. H ESSELIUS , P., J OHANSSON , P. and N ILSSON , J. P. (2009). Sick of your colleagues’ absence? Journal of the European Economic Association, Papers and Proceedings, 7 (2-3), 583–594. H OTZ , V. J., J OHANSSON , P. and K ARIMI , A. (2018). Parenthood, Family Friendly Firms, and the Gender Gaps in Early Work Careers. Tech. rep., National Bureau of Economic Research. L INDBECK , A., N YBERG , S. and W EIBULL , J. W. (1999). Social norms and economic incentives in the welfare state. The Quarterly Journal of Economics, 114 (1), pp. 1–35. M ANSKI , C. F. (1993). Identification of endogenous social effects: The reflection problem. The review of economic studies, 60 (3), 531–542. M AS , A. (2006). Pay, reference points, and police performance. The Quarterly Journal of Economics, 121 (3), pp. 783–821. — and M ORETTI , E. (2009). Peers at work. The American Economic Review, 99 (1), 112–145. N AGIN , D. S., R EBITZER , J. B. and S ETH S ANDERS , L. J. T. (2002). Monitoring, motivation, and management: The determinants of opportunistic behavior in a field experiment. The American Economic Review, 92 (4), 850–873. O FFERMAN , T. (2002). Hurting hurts more than helping helps. European Economic Review, 46 (8), 1423– 1437. R ABIN , M. (1993). Incorporating fairness into game theory and economics. The American economic review, pp. 1281–1302.

14

F IGURE 1. Hazard rate during the experimental period

(A) Göteborg

(B) Jämtland N OTE .— The figure shows the hazard rates of sick leave days during the experiment for the treated and control groups in Göteborg. Source: Hartman et al. (2013).

15

0

1

Density 2

3

4

F IGURE 2. Share treated co-workers of workers residing in Göteborg and Jämtland

0

.2

.4 .6 Share treated

.8

1

Mean: 0.297

0

1

2

Density 3

4

5

(A) Gothenburg

0

.2

.4 .6 Share treated

Mean: 0.452

(B) Jämtland

16

.8

1

17 22,402

2.46

All

22,402

3.42

-0.23∗∗∗ (0.07) -0.18 (0.36) 2.25∗∗∗ (0.38)

61,715

2.62

0.03 (0.04) -0.09 (0.22) 1.77∗∗∗ (0.38)

61,715

4.17

0.36∗∗∗ (0.05) 0.82∗∗ (0.33) 3.14∗∗∗ (1.18)

<15 days

(2)

11,300

2.59

-0.12 (0.34) 1.64∗∗∗ (0.37)

30,339

1.80

-0.35∗ (0.20) 1.58∗∗∗ (0.45)

30,339

2.82

0.10 (0.29) 1.34∗∗∗ (0.46)

<8 days

(3) Treated

11,300

3.30

-0.52 (0.46) 2.52∗∗∗ (0.52)

30,339

2.63

-0.40 (0.31) 2.18∗∗∗ (0.54)

30,339

4.38

0.53 (0.47) 3.41∗∗∗ (0.89)

<15 days

(4)

11,102

2.33

0.29 (0.38) 0.87∗∗ (0.35)

31,376

1.79

-0.06 (0.22) 1.17∗∗∗ (0.42)

31,376

3.13

0.92∗∗∗ (0.32) 2.04∗∗∗ (0.67)

11,102

3.54

0.24 (0.57) 1.73∗∗∗ (0.52)

31,376

2.61

0.25 (0.33) 1.62∗∗∗ (0.54)

31,376

3.97

1.19∗∗∗ (0.42) 2.90∗∗ (1.13)

(6) Non-Treated <8 days <15 days

(5)

N OTE .— The outcome variables are the number of days in sick leave spells that are shorter than 8 and 15 days, corresponding to non-monitored absence for the treatment group before and during the experiment. Included covariates are gender, age, earnings, dummies for schooling level, dummies for the share of commuters at the workplace (divided into 10 percent bins), share of female employees, average age at the workplace, average earnings at the workplace, share of employees with compulsory-, high school- and college education, dummies for industry affiliation, workplace average sickness absence days in the fall and spring of 1987 and the spring of 1988. Panel C presents results from the reversed experiment in Jämtland, shifting monitoring from the 14th day to the 7th day. Standard errors are clustered at the workplace level. * p < 0.1., ** p < 0.05, *** p < 0.01.

Observations

Mean absence days

Constant

Share of treated co-workers

0.25∗∗∗ (0.05) 0.07 (0.26) 1.13∗∗∗ (0.27)

61,715

Observations

C. Work absence days in Fall 1988, reversed experiment Treatment

1.79

0.00 (0.02) -0.22 (0.14) 1.15∗∗∗ (0.27)

61,715

2.97

-0.28∗∗∗ (0.03) 0.48∗∗ (0.22) 1.73∗∗∗ (0.53)

<8 days

(1)

Mean absence days

Constant

Share of treated co-workers

B. Work absence days in Fall 1987 (placebo) Treatment

Observations

Mean absence days

Constant

Share of treated co-workers

A. Work absence days in Fall 1988 Treatment

TABLE 1. The effect of treatment and of the share treated co-workers on work absence

18

7939

-0.2927∗∗∗ (0.1120) 0.2812 (0.5502) 1.7855∗∗∗ (0.6470)

23,155

0.4335∗∗∗ (0.0803) 0.0042 (0.5580) 2.4006∗∗∗ (0.4867)

Low Job flexibility

(1) All

7439

-0.1960∗ (0.1182) -0.3312 (0.6471) 2.0392∗∗∗ (0.5831)

23,484

0.3623∗∗∗ (0.0759) 0.9572∗ (0.5006) 2.4640∗∗∗ (0.4707)

High Job flexibility

(2)

4032

-0.2522 (0.5969) 2.8241∗∗∗ (0.6906)

11,411

-0.6298 (0.8390) 3.5360∗∗∗ (0.7997)

(4)

3716

-0.3487 (0.8359) 1.7240∗∗ (0.8034)

11,525

0.6457 (0.7061) 3.2427∗∗∗ (0.6552)

High Job flexibility

Treated Low Job flexibility

(3)

3907

0.8637 (0.9804) 0.4185 (1.0055)

11,744

0.5849 (0.6988) 1.7825∗∗∗ (0.5382)

(6)

3723

-0.2658 (1.0273) 2.0646∗∗ (0.8594)

11,959

1.3600∗∗ (0.6744) 2.0459∗∗∗ (0.6127)

High Job flexibility

Non-treated Low Job flexibility

(5)

N OTE .— The outcome variable is the number of days in sick leave spells that are shorter than 15 days in the fall of 1988. Included covariates are age, earnings, dummies for schooling level, dummies for the share of commuters at the workplace (divided into 10 percent bins), share of female employees, average age at workplace, average earnings at workplace, share of employees with compulsory-, high school- and college education, workplace average sickness absence days in the fall and spring of 1987 and the spring of 1988. Standard errors are clustered at the workplace level. * p < 0.1., ** p < 0.05, *** p < 0.01

Observations

Constant

Share treated

B. Work absence days in Fall 1988, reversed experiment Treatment

Observations

Constant

Share treated

A. Work absence days in Fall 1988 Treatment

Sample Job type

TABLE 2.

The effect of treatment and of the share treated co-workers on the work absence of workers with low and high job flexibility

Appendix A

Additional Tables and Figures TABLE A.1.

Summary statistics: individual and co-worker attributes for the samples in Göteborg and Jämtland Göteborg Mean (sd)

Jämtland Mean (sd)

Individual attributes Treated Female Age Annual earnings (1000s SEK) At most high school College Absence < 15 days, Spring 1987 Absence < 15 days, Fall 1987 Absence < 8 days, Spring 1987 Absence < 8 days, Fall 1987 Absence < 15 days, Spring 1988 Absence < 15 days, Fall 1988 Absence < 8 days, Spring 1988 Absence < 8 days, Fall 1988

0.491 0.509 36.298 98.852 0.445 0.255 2.712 2.619 1.821 1.794 3.456 4.173 2.699 2.975

(0.500) (0.500) (12.680) (68.924) (0.497) (0.436) (4.916) (4.829) (3.087) (3.134) (5.390) (6.252) (4.079) (4.290)

0.505 0.523 36.975 85.134 0.471 0.183 2.696 2.621 1.367 1.258 3.025 3.419 2.106 2.460

(0.500) (0.499) (13.041) (54.294) (0.499) (0.386) (5.390) (5.385) (2.546) (2.545) (5.123) (5.323) (3.380) (3.706)

Co-worker & workplace attributes Avg. age Avg. earnings (1000s SEK) Share with high school education Share with college education Share females Worklplace size (count) Share treated Share commuters Avg. absence < 15 days, Spring 1987 Avg. absence < 15 days, Fall 1987 Avg. absence < 15 days, Spring 1988 Avg. absence < 15 days, Fall 1988

36.545 99.812 0.427 0.234 0.506 39.453 0.297 0.380 2.655 2.549 3.351 3.975

(5.901) (37.502) (0.176) (0.250) (0.312) (25.411) (0.137) (0.239) (1.350) (1.383) (1.647) (1.931)

37.005 85.310 0.457 0.181 0.525 37.799 0.452 0.089 2.644 2.570 3.019 3.426

(5.482) (25.783) (0.173) (0.227) (0.323) (23.988) (0.149) (0.223) (1.414) (1.470) (1.474) (1.576)

N OTE .— The table reports the means and standard deviations of characteristics of workers (and their co-workers) residing in Göteborg and Jämtland in 1988, and who were employed at workplaces with 10-100 employees.

19

TABLE A.2. Covariate balance tests for the experiments in Göteborg and Jämtland Göteborg

Jämtland

Point Est.

SE

t-stat

Point Est.

SE

t-stat

Individual attributes Female Annual earnings (1000s SEK) Age At most high school College Absence < 15 days, Spring 1987 Absence < 15 days, Fall 1987 Absence < 15 days, Spring 1988 Absence < 8 days, Spring 1987 Absence < 8 days, Fall 1987 Absence < 8 days, Spring 1988

-0.002 0.634 -0.102 0.003 -0.002 -0.048 0.024 0.024 -0.018 0.004 -0.003

0.004 0.538 0.102 0.004 0.003 0.037 0.037 0.041 0.024 0.024 0.032

-0.571 1.177 -1.004 0.742 -0.671 -1.293 0.653 0.569 -0.779 0.184 -0.103

0.000 0.389 0.384 -0.007 0.002 0.103 -0.010 0.001 0.020 0.028 0.039

0.007 0.725 0.184 0.007 0.005 0.068 0.069 0.069 0.034 0.033 0.045

0.010 0.537 2.082 -1.064 0.328 1.524 -0.141 0.009 0.583 0.853 0.855

Co-worker & workplace attributes Avg. age Avg. earnings (1000s SEK) Share high school education Share college education Share females Workplace size (count) Share commuters Avg. absence < 15 days, Spring 1987 Avg. absence < 15 days, Fall 1987 Avg. absence < 15 days, Spring 1988

-0.072 0.544 -0.001 0.002 -0.003 0.124 0.005 -0.018 -0.016 -0.012

0.045 0.289 0.001 0.002 0.002 0.193 0.002 0.010 0.011 0.013

-1.624 1.883 -0.892 0.913 -1.180 0.643 2.747 -1.691 -1.513 -0.923

0.108 0.225 -0.001 0.005 0.002 0.429 -0.002 -0.012 -0.017 -0.002

0.072 0.354 0.002 0.003 0.004 0.338 0.003 0.020 0.021 0.020

1.490 0.636 -0.318 1.614 0.522 1.268 -0.557 -0.595 -0.828 -0.115

Joint χ2 p-value

28.699 0.121

18.563 0.613

N OTE .— The table reports the estimated difference in individuals’ attributes between the treated and non-treated in Göteborg and Jämtland. The χ2 − test statistic of joint significance of all coefficients was obtained by estimating a seemingly unrelated regression equation system for the included covariates, allowing for correlated error terms across equations.

20

21

22,233

0.02 (0.03) -0.19 (0.17) 1.27∗∗∗ (0.30)

All

22,233

-0.02 (0.07) -0.08 (0.33) 2.00∗∗∗ (0.74)

<15 days

(2) Treated

<15 days

(4)

11,206

-0.22 (0.23) 1.05∗∗∗ (0.30) 11,206

-0.03 (0.47) 2.08∗∗∗ (0.63)

Work absence days in Fall 1987, Reversed experiment

<8 days

(3)

11,027

-0.13 (0.26) 0.69∗ (0.38)

<8 days

(5) Non-Treated

11,027

-0.03 (0.51) 2.14∗ (1.15)

<15 days

(6)

N OTE .— The table presents results from the reversed experiment in Jämtland, shifting monitoring from the 14th day to the 7th day, in the year preceding the experiment. The outcome variables are the number of days in sick leave spells that are shorter than 8 and 15 days, corresponding to non-monitored absence for the treatment group during and before the experiment. Included covariates are gender, age, earnings, dummies for schooling level, dummies for the share of commuters at the workplace (divided into 10 percent bins), share of female employees, average age at the workplace, average earnings at the workplace, share of employees with compulsory-, high school- and college education, dummies for industry affiliation, workplace average sickness absence days in the fall and spring of 1987 and the spring of 1988. Standard errors are clustered at the workplace level. * p < 0.1., ** p < 0.05, *** p < 0.01.

Observations

Constant

Share of treated co-workers

Treatment

<8 days

(1)

TABLE A.3. Placebo analysis for Jämtland

TABLE A.4. Summary statistics: industry composition for the samples in Göteborg and Jämtland Göteborg

Jämtland

Frequency

Percent

Frequency

Percent

Agriculture, hunting, forestry & fishing Mining and quarrying Manufacturing Electricity, gas-, and water supply Construction Wholesale and retail trade Transport and communications Financial intermediation Real estate, renting & business activities Data management operations Other business activities Hotels and restaurants Education R&D Health and social work Lobbying, and religious activities Recreation-, culture-, and sports, and other activities Public administration

170 71 7,778 665 4,525 12,987 4,325 1,688 2,068 952 5,344 3,326 4,235 233 11,608 1,318 2,490 1,914

0.259 0.108 11.839 1.012 6.888 19.768 6.583 2.569 3.148 1.449 8.134 5.063 6.446 0.355 17.669 2.006 3.790 2.913

566 46 3,321 503 1,374 3,168 882 465 559 74 525 1,401 2,556 7 5,495 371 497 1,029

2.478 0.201 14.541 2.202 6.015 13.871 3.862 2.036 2.448 0.324 2.299 6.134 11.191 0.031 24.060 1.624 2.176 4.505

Total

65,697

100

22,839

100

N OTE .— The table reports the industry composition of workers residing in Göteborg and Jämtland in 1988, and who were employed at workplaces with 10-100 employees.

TABLE A.5. Principal component analysis for the job flexibility index: Göteborg Factor Loadings (1)

(2)

(3)

(4)

(5)

Job Flexibility Work location Start & end of work-day Planning of work Allocate time across tasks Structure of work

0.4318 0.3866 0.4758 0.4499 0.4851

-0.5199 0.7002 -0.2022 -0.2706 0.3541

-0.0258 0.3089 -0.4983 0.7086 -0.3917

0.7026 0.1354 -0.5810 -0.3520 0.1630

0.2213 0.4966 0.3832 -0.3136 -0.6777

Eigenvalue Percent of variance

3.4033 0.6806

0.8936 0.1787

0.4215 0.0843

0.2470 0.0494

0.0347 0.0069

N OTE .— The table presents the results from a factor analysis used to construct the job flexibility index.

TABLE A.6. Principal component analysis for the job flexibility index: Jämtland Factor Loadings (1)

(2)

(3)

(4)

(5)

Job Flexibility Work location Start & end of work-day Planning of work Allocate time across tasks Structure of work

0.3662 0.4132 0.4742 0.4644 0.5046

0.6676 -0.6639 0.2609 0.0276 -0.2114

0.5583 0.2770 -0.4171 -0.5887 -0.3917

0.2996 0.2245 -0.5841 0.5974 -0.4022

0.1370 0.5112 0.4381 -0.2831 - 0.6692

Eigenvalue Percent of variance

3.2902 0.6580

0.8303 0.1661

0.5167 0.1033

0.2950 0.0590

0.0679 0.0136

N OTE .— The table presents the results from a factor analysis used to construct the job flexibility index.

22