The economic and social determinants of employee behavior: evidence from insider econometric studies / von Konstantin Böddeker, M.Sc.

Fakultät für Wirtschaftswissenschaften Lehrstuhl für Organisations-, Medien- und Sportökonomie ___________________________________________________________________________________ THE ECONOMIC AND SOCIAL DETERMINANTS OF EMPLOYEE BEHAVIOR: EVIDENCE FROM INSIDER ECONOMETRIC STUDIES ___________________________________________________________________________________ Der Fakultät für Wirtschaftswissenschaften der Universität Paderborn zur Erlangung des akademischen Grades Doktor der Wirtschaftswissenschaften - Doctor rerum politicarum - vorgelegte Dissertation von Konstantin Böddeker, M.Sc. geboren am 10.12.1985 in Paderborn (2016) The data used in this thesis are proprietary and can, therefore, not be made available to other researchers. The results, opinions and conclusions of this dissertation are those of the author and not necessarily those of the Volkswagen AG. Table of Content III TABLE OF CONTENT LIST OF FIGURES ............................................................................................................ V LIST OF TABLES ............................................................................................................. VI LIST OF ABBREVIATIONS ........................................................................................ VIII 1 INTRODUCTION ........................................................................................................... 1 2 WORK TEAM DIVERSITY AND THE OPTIMAL COMPOSITION OF TEAMS: A META-ANALYTIC APPROACH .................................................... 11 2.1 Introduction ........................................................................................................... 11 2.2 Diversity ................................................................................................................ 13 2.2.1 Definitions of Diversity ................................................................................. 14 2.2.2 Dimensions of Diversity ................................................................................ 16 2.2.3 Measures of Diversity .................................................................................... 16 2.3 Theoretical Background of Diversity Research .................................................... 18 2.3.1 Social Categorization and Similarity-Attraction Paradigm ........................... 18 2.3.2 Information Processing Perspective and Economic Theory .......................... 21 2.3.3 Moderators of the Diversity-Performance Relationship ................................ 23 2.4 Meta-Analysis Approach ...................................................................................... 25 2.5 Results ................................................................................................................... 30 2.6 Conclusion and Implications ................................................................................ 36 3 EMPLOYEE ABSENTEEISM: DETERMINANTS IN THE INTERNATIONAL CONTEXT ................................................................................. 39 3.1 Introduction ........................................................................................................... 39 3.2 Theoretical Framework ......................................................................................... 42 3.2.1 Social Peer Influences in Work Teams.......................................................... 45 3.2.2 Economic Influences / Incentives .................................................................. 48 3.2.3 Worker Characteristics .................................................................................. 49 3.2.4 Working Condition ........................................................................................ 52 3.3 Data Set and Descriptive Statistics ....................................................................... 52 3.4 Estimation Strategy ............................................................................................... 57 3.5 Empirical Results .................................................................................................. 61 3.6 Conclusion and Implications ................................................................................ 74 Table of Content IV 4 ON THE ROAD AGAIN: CROWDING-OUT EFFECTS OF EXTRINSIC MOTIVATION IN COMMERCIAL TRUCKING.................................................... 79 4.1 Introduction ........................................................................................................... 79 4.2 Theoretical Framework ......................................................................................... 82 4.3 Data Set and Descriptive Statistics ....................................................................... 88 4.4 Empirical Models and Estimation Results ............................................................ 93 4.5 Conclusion .......................................................................................................... 106 5 BEHAVIORAL CONSEQUENCES OF THE TRANSITION FROM TEMPORARY TO PERMANENT EMPLOYMENT ............................................. 110 5.1 Introduction ......................................................................................................... 110 5.2 Literature Review and Theoretical Framework .................................................. 112 5.3 Data Set and Descriptive Statistics ..................................................................... 117 5.4 Empirical Models and Estimation Results .......................................................... 121 5.5 Conclusion .......................................................................................................... 128 6 SUMMARY AND FUTURE OUTLOOK ................................................................. 131 APPENDIX ........................................................................................................................ IX REFERENCES ............................................................................................................ XXIX EIDESSTATTLICHE ERKLÄRUNG ......................................................................... LXI List of Figures V LIST OF FIGURES Figure 1.1: Diversity as Separation, Variety and Disparity ................................................. 15 Figure 3.1: Steers and Rhodes (1978) Process Model of Employee Absence .................... 43 APPENDIX Figure A.1 : Forest Plot – Age Diversity and Overall Performance.................................... IX Figure A.2: Forest Plot – Gender Diversity and Overall Performance ................................ X Figure A.3: Forest Plot – Culture Diversity and Overall Performance ............................... XI Figure A.4: Forest Plot – Tenure Diversity and Overall Performance .............................. XII Figure A.5: Forest Plot – Functional Background Diversity and Overall Performance... XIII Figure A.6: Forest Plot – Educational Background Diversity and Overall Performance . XIV Figure A.7: Forest Plot – Education Level Diversity and Overall Performance ............... XV Figure A.8: Fuel Consumption with / without Incentives ................................................ XXI Figure A.9: Kernel Density Plot of Fuel Consumption by Trip Evaluation ............... XXVIII List of Tables VI LIST OF TABLES Table 2.1: Formulas for the Calculation of Heterogeneity Measures ................................. 17 Table 2.2: Relationship between Diversity and Overall Team Performance ...................... 31 Table 2.3: Team Size as Moderator of the Performance-Diversity Relationship ................ 33 Table 2.4: Team Type as a Moderator of the Performance-Diversity Relationship............ 34 Table 3.1: Summary Statistics – Data Set Composition ...................................................... 54 Table 3.2: Summary Statistics – Means and standard deviations........................................ 55 Table 3.3: Fixed-Effects Interaction Terms on Absence Rates (I) ...................................... 63 Table 3.4: Fixed-Effects Interaction Terms on Absence Spell Duration (II) ...................... 64 Table 3.5: Fixed-Effects Interaction Terms on Absence Frequency (III) .......................... 65 Table 4.1: Summary Statistics - Driver Demographics ....................................................... 90 Table 4.2: Summary Statistics - Driving Style .................................................................... 90 Table 4.3: Summary Statistics - Traffic Light Performance Evaluation Tool..................... 92 Table 4.4: Baseline Results of the Fixed-Effects Estimations (Models 1 to 6) ................... 98 Table 4.5: Seemingly Unrelated Regression (SUR) Results on Driving Parameters ........ 102 Table 4.6: Seemingly Unrelated Regression (SUR) Results on Evaluation Parameters ... 103 Table 4.7: Estimated Probabilities for Trip Evaluation in both Incentive Environments . 104 Table 5.1: Summary Statistics – Driver Demographics .................................................... 118 Table 5.2: Summary Statistics – Dependent Variables ..................................................... 119 Table 5.3: Impact of Contract Status on Short-Term Fuel Consumption .......................... 124 Table 5.4: Impact of Contract Status on Long-Term Fuel Consumption .......................... 125 Table 5.5: Estimated Probabilities for Trip Evaluation ..................................................... 127 List of Tables VII APPENDIX Table A.1: Heterogeneity Measures Listed by Frequency of Use ....................................... IX Table A.2: Diversity-Performance Relation by Performance Measure ............................ XVI Table A.3: Fixed-Effects Estimations on Absence Rate (IV) ....................................... XVIII Table A.4: Fixed-Effects Estimations on Mean Absence Spell Duration (V).............. 19IXX Table A.5: Fixed-Effects Estimations on Absence Frequency (VI) .................................. XX Table A.6: Descriptive Statistics with and without Incentives in Practice ....................... XIX Table A.7: Two-Way Anova of Incentives and Driver on Fuel Consumption.............. XXIII Table A.8: Findings on Fuel Consumption after Installation of Incentives .................. XXIII Table A.9: Findings on Fuel Consumption after Abolition of Incentive ...................... XXIV Table A.10: Ordered Logit Estimation for Trip Evaluation .......................................... XXIV Table A.11: Baseline Results of Fixed-Effects Estimation (incl. Bonus Regime) ......... XXV Table A.12: Baseline Results of Fixed-Effects Estimation (incl. Trend Variable) ....... XXVI Table A.13: Ordered Logit Estimation of Driver Performance .................................... XXVII List of Abbreviations VIII LIST OF ABBREVIATIONS CI Confidence Interval CV Coefficient of Variation DWD Deutscher Wetterdienst (German Meteorological Service) ES Effect Size ESP Spain EU European Union FE Fixed-Effects (Regression) GDP Gross Domestic Product GER Germany GPS Global Positioning System HR Human Resources HRM Human Resource Management ILI Influenza-like Illnesses KPI Key Performance Indicator OECD Organization for Economic Co-Operation and Development OHC Occupational Health Care OLS Ordinary Least Squares (Regression) R&D Research and Development RE Random-Effects (Regression) ROA Return on Assets ROE Return on Equity ROS Return on Sales s.d. Standard Deviation SUR Seemingly Unrelated Regression TMT Top Management Team UK United Kingdom US United States of America WHO World Health Organization Introduction 1 1 INTRODUCTION Since the 1980s the strand of personnel economics has emerged from labor economics in order to address both the lack of economic perspectives in early human resource management research as well as the growing need for hands-on, business-oriented insights on human resource management (HRM) (Lazear 2000a). Initially influenced by informational economics and further extended by ideas originating from social psychology and organizational sociology, personnel economics is nowadays a well-established and substantial part of labor economics research. Reviews by Prendergast (1999), Lazear (1999, 2000a) and Lazear and Oyer (2013) provide a comprehensive overview of the historical development as well as the contemporary debate within the field of personnel economics. Personnel economics applies (labor) economic methods such as econometrics and game theory to provide a detailed understanding of the functioning of a firm and its human resource practices. However, due to limited data availability in the early years, personnel economics started off with mostly theoretical considerations. Later contributions combine theory with empirical analyses as they are built on rich firm-based data sets. In general, personnel economic models and theories play by the traditional rules of economics, in particular the assumptions of rational maximizing agents, specific equilibrium analysis of labor and product markets and efficiency as a result of market equilibria (Lazear 2000a, Lazear, Shaw 2007). Nevertheless, the main research fields of personnel economics are not limited to the classic HRM issues such as compensation, turnover and incentives. Instead, its scope of research is extended beyond traditional economic fields, including topics such as group norms, teamwork settings, or worker empowerment. In brief, one can summarize the main research interests of personnel economists to comprise incentives, worker-firm matching, compensation, skill development and organization of work (Lazear, Oyer 2013). These topics share the common thread in that personnel economics is all about enhancing worker productivity (Lazear, Gibbs 2009). In contrast to labor economics, personnel economics does not appear policy-oriented: It focuses exclusively on the welfare within an individual employment relationship and not on the overall social welfare (Lazear, Oyer 2013). In other words, personnel economics aims at offering practical implications for managers on how to improve operations and, thus, enhance the overall productivity of a given firm. As a consequence, personnel economics research mainly focuses on those variables that managers have an actual bearing in. Introduction 8 hand, similarity-attraction paradigm and social categorization theory suggest group heterogeneity to induce performance and productivity to be inferior due to sub-group formation (in-group vs. out-group) and work group conflict. As a consequence, homogeneous teams are assumed to be preferable (Tajfel 1982, Tajfel, Turner 1986, Turner 1987). On the other hand, the information processing perspective assumes heterogeneous teams to increase performance and productivity due to complementary cognitive resources, experiences and network ties. Hence, heterogeneous teams are expected to be advantageous in terms of performance and productivity (Cox, Blake 1991). There is extensive empirical evidence for both approaches. Applying meta-analytic research to a total of 66 individual samples derived from 63 primary studies on work group diversity, this paper seeks to analyze the broad literature on diversity in search of common statistical patterns on the advantageousness of either homogeneous or heterogeneous teams. Diversity is considered along less task-related (age, gender, culture) and highly task-related (tenure, function, educational background and educational level) attributes. Furthermore, empirical tests for potential moderating effects of team type (top management teams, work teams, research and development teams and mixed teams) and team size are performed. Overall, findings confirm the ambivalent nature of work group heterogeneity. In support of social categorization theory and similarity-attraction paradigm, the results suggest negative population correlations of gender and age heterogeneity on team performance. Likewise, educational background diversity is observed to be positively related to team performance, thus, confirming arguments stated by information processing perspective. Two moderating effects are observed. First, increasing team size moderates the influence of team diversity on performance negatively for less task-related diversity attributes and positively for highly task-related attributes. Second, top management team type and research and development team type positively moderate the diversity-performance relationship regardless of the task-relatedness of diversity. Moreover, work team type negatively moderates the diversity-performance relationship for work teams regardless of the task-relatedness of diversity attributes. Chapter three depicts joint work with Bernd Frick in which a hitherto unavailable data set covering 160 blue-collar work units from four international manufacturing plants of a large European automobile manufacturer is used to analyze social and economic determinants of employee absenteeism. In this chapter, the understanding of absenteeism is based on the Steers and Rhodes (1978) process model as well as on economic theories as proposed by neoclassical labor supply models (e.g. Allen 1981, Dunn, Youngblood 1986) and efficien- Introduction 9 cy wage theory (Shapiro, Stiglitz 1984). By means of a comprehensive review of existing empirical findings on absence research, a set of thirteen potential determinants of absence is identified that stem from social peer influences (work unit size, turnover, share of temporary co-worker, share of health-impaired co-workers), economic influences / incentives (sick pay regulations, employment protection laws, unemployment, prosperity level), worker characteristics (age, gender, tenure, acute health) and working conditions (shift system). Absence is found to be a multifaceted phenomenon since only few statistical patterns hold true for all four plants. Instead, core findings suggest results to vary significantly by production site. Based on empirical analyses as well as interviews with on-site experts (i.e. HR managers, worker representatives, managers and shop floor staff), results suggest absence to increase with unit size and overall unit turnover. Likewise, a high share of temporary co-workers increases group-level absence among permanent employees. In contrast, the hypothesized link between absence and the share of co-workers that suffer any kind of permanent or temporary health-impairment could not be supported. Additionally, positive relationships between absence and strict employment protection laws as well as favorable prosperity levels are identified. Findings with respect to the national (un)employment situation are inconclusive and results for national sickness benefits legislation fail to reach statistical significance. Furthermore, group-level absence is observed to increase with the share of female employees in the team. Moreover, both involuntary and voluntary absence is observed to increase with age. However, results for tenure as well as national influenza outbreaks are inconclusive. With regard to working conditions, absence is observed to be higher in units that operate in two- and three-shift systems compared to those following the standard one-shift system. Employees’ behavioral responses to changes in monetary incentive design are discussed in chapter four, which is joint work with Bernd Frick. Built on fleet management panel data of an in-house hauler of a large European truck manufacturer, this paper analyzes the performance of 37 commercial truck drivers on a total of 6,326 individual trips within a threeyear time frame. Performance is measured via four different sets of variables that are recorded by GPS-based on-board computers: overall fuel consumption, driving behavior as reflected by several driving parameters, relative performance based on driving scores and, eventually, a computed overall performance evaluation. The organizational setting follows a quasi-experimental approach since the existing performance bonus system had been abolished by the hauler after the first two years of the three-year observation period. From Introduction 10 theory, two potential behavioral responses might be expected from employees. Classic contract theory assumes human beings to directly respond to extrinsic incentives in increasing effort as well as reducing shirking behavior (e.g. Lazear 2000b, Gibbons 1998, Gibbons, Roberts 2013). However, a competing theoretical approach doubts these benefits of extrinsic (monetary) incentives arguing that extrinsic incentives literally “buy off” an employee’s intrinsic motivation and, therefore, result in decreasing overall performance (e.g. Deci 1971, Bénabou, Tirole 2003, 2006, Gneezy et al. 2011). Empirical and experimental evidence for both approaches is discussed in the chapter. Interestingly, findings contradict economic theory in giving support for the crowding-out approach. Truck drivers are observed to display significantly lower effort and performance when incentivized extrinsically by means of a performance bonus. Estimations reveal that drivers behave less eco-friendly and less in accordance with company standards yielding an overall lower performance evaluation. This result implies that drivers’ intrinsic and social motives – e.g. image concerns, environmental believes – are crowded out by monetary rewards. As mentioned earlier, chapter five – again co-authored by Bernd Frick – analyzes behavioral responses of employees who are “promoted” from temporary to permanent contracts. This chapter is based on the identical fleet management panel data used in chapter 4. Again, the organizational setting follows a quasi-experimental design since it observes eight commercial truck drivers on 1,299 trips whose contract status changed from temporary to permanent. Considering advantages of permanent over temporary contracts (e.g. better working conditions, higher pay and stricter employment protection), temporary employees may be assumed to be highly incentivized to display high effort in order to qualify for a permanent contract. Empirical support for this argument is discussed within the chapter (e.g. Bradley et al. 2012). Findings suggest about half of the drivers to significantly increase fuel consumption after having signed a permanent contract with a strong initial effect and a long decline back to “normal” levels. This result indicates these drivers to behave as rational cheaters (Nagin et al. 2002). The same holds true for computerized performance evaluations. Subsequently, chapter 6 summarizes the main results of this thesis and provides both general conclusions as well as suggestions for future research. Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 11 2 WORK TEAM DIVERSITY AND THE OPTIMAL COMPOSITION OF TEAMS: A META-ANALYTIC APPROACH 2.1 Introduction These days, organizations increasingly implement teamwork to meet the demands of sustainable developments in economic, technological and societal environments (e.g. Devine et al. 1999). Given the growing importance of teams, management scholars have eagerly taken on the study of teamwork and its influence on business processes as well as employee performance (e.g. Sundstrom et al. 2000). As a consequence, teamwork research is nowadays established as one of the most relevant topics in management research, a position anecdotally described by Lazear (1999) who states that “it is impossible to pick up a business publication these days without reading about the wonders of teamwork” (Lazear 1999, p. C15). At the same time, the composition of the global workforce has changed in that worker heterogeneity has increased dramatically over the last decades. Multifaceted social and demographic developments may be held accountable for this phenomenon. Consider, for instance, the constantly growing proportion of women participating in the labor market (e.g. Ali, Kulik, Metz 2011). At the same time, women increasingly advance into managerial ranks (e.g. Elsass, Graves 1997) and into jobs formerly known to be all-male professions such as medicine or law (e.g. Norgren 2010). Furthermore, the individual working life span has prolonged since employees enter the labor market at a younger age and stay much longer before becoming eligible for retirement (e.g. Grund, Westergaard-Nielsen 2008). Eventually, cultural and ethnic background heterogeneity is evermore increasing in line with global migration and human capital mobility (e.g. Artuc et al. 2015). So far, there are no clear indications that the ongoing social and demographic trends will slow down or even reverse in the near future (e.g. Deadrick, Stone 2009, Tsui, Gutek 1999). Given the development towards a more heterogeneous workforce, it is of growing importance to understand how diversity issues affect processes and productivity within organizations and teams. According to Barsade et al. (2000), research on the “costs and benefits of diversity in the workplace has been going on at a vigorous pace over the last two decades and more” (Barsade et al. 2000, p. 802). As a consequence, the amount of studies on diversity approximately doubles every five years, for instance scholars published 134 Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 12 studies in 2003 compared to only 19 studies back in 1988 (Harrison, Klein 2007). However, the existing research on diversity in the workplace is confusing and disappointing in a way that “cumulative findings about the consequences of within-unit differences have been weak, inconsistent, or both” (Harrison, Klein 2007, p. 1199). Diversity research is mainly guided by two opposing research traditions that clearly demonstrate the ambivalent nature of diversity that is often considered a “double-edged sword” (Milliken, Martins 1996, p. 403). On the one hand, a substantial body of the literature states that within teams heterogeneity may induce social categorization that in turn implies sub-group formation (in-group vs. out-group) and team conflict and, as a consequence, result in lower team performance. Hence, homogeneity in teams is argued to be preferable to heterogeneity with regard to team performance. The most well-known contributions to this concept are expressed in similarity-attraction paradigm as well as social categorization theory (e.g. Tajfel 1982, Tajfel, Turner 1986, Turner 1987). On the other hand, scholars suppose that team production may benefit from heterogeneous cognitive abilities, experiences, skills and network ties of team members. These may be complements and, therefore, constitute a broad set of resources available to the team. As a result, heterogeneity in teams is reasoned to be advantageous over homogeneity regarding team performance. These considerations are summarized in information processing perspective (e.g. Hoffman, Maier 1961, Cox, Blake 1991) and economic theory alike (e.g. Lazear 1999). In this paper, we consider both theoretical approaches to derive our main research hypotheses on the diversity-performance relationship within organizational work groups. We aim at contributing new insights to the wide field of diversity research by means of a comprehensive meta-analysis. Moreover, we test for two potential moderating effects on the diversity-performance relationship – team size and team type. Both have been identified as being influential by previous research (e.g. Stewart 2006, Horwitz, Horwitz 2007, Devine 2002). For this purpose, we quantitatively summarize the last twenty years of empirical research on diversity in organizational work teams (from 1994 to 2014). Based on a thorough literature review, we identify 242 empirical studies on the link between diversity and performance in organizational teamwork settings. Following the tradition of previous meta-analyses on workgroup diversity (e.g. Bell 2007, Bell et al. 2011, De Dreu, Weingart 2003, Horwitz, Horwitz 2007, Stewart 2006, Webber, Donahue 2001), we apply strict in- Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 13 clusion criteria to manage the vast amount of literature. Eventually, we include a total of 66 individual samples from 63 studies into the meta-analysis. Overall, our findings confirm the ambivalent nature of diversity in teams. We identified the impact of diversity on team performance to depend primarily on the diversity attribute being either highly or less task-related. On the one hand, we observe a negative relationship between team performance and heterogeneity along less task-related attributes (gender and age). These findings largely support social category theory and similarity-attraction paradigm. On the other hand, we observe heterogeneity along highly task-related attributes (educational background) to be positively linked to team performance. This finding largely supports economic assumptions and information processing perspective. Furthermore, we observe both team size and team type to moderate the diversity-performance relationship. With increasing team size the influence of diversity on performance decreases for less task-related attributes and increases for highly task-related attributes. Concerning team type, we observe a positive moderating effect of top management teams (TMT) as well as research and development (R&D) teams on almost all diversity attributes irrespective of the task-relatedness. In contrast to this, work teams negatively moderate all diversityperformance relations. All results will be discussed and interpreted in detail throughout this study. This paper proceeds as follows. Section two will discuss the diversity construct, its dimensions, the most common heterogeneity measures as well as a definition of diversity as understood in this paper. Subsequently, section three will present the two competing theoretical concepts and derive the main research hypotheses of this paper. The meta-analytic procedure applied in this study will be described in section four. Eventually, section five presents the results of the meta-analysis while section six concludes and offers implications for practical use and future research. 2.2 Diversity Despite the relevance of diversity in the recent academic discourse, there is a lack of both a common understanding and a unique definition of the underlying construct (e.g. Guzzo, Dickson 1996). This becomes even more apparent considering the high number of synonyms that are used interchangeable for the umbrella term ‘diversity’ by scholars: “’dispersion’, ‘inequality’, ‘within-group variability’, ‘(dis)agreement’, ‘consensus’, ‘heterogeneity’, ‘homogeneity’, ‘deviation’, ‘difference’, ‘distance’, ‘relational demography’, Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 14 ‘sharedness’ and more” (Harrison, Sin 2006, p. 195). Since each separate definition is linked with a slightly different understanding and metric, the lack of a common diversity concept seems to be even more obstructive. Hence, any comparison across studies is complicated and needs particular attention due to the fact that researchers do not necessarily mean the same phenomenon when referring to diversity. In general, three main inconsistencies have been identified in diversity research: the lack of a unique constitutive definition, the high variety within diversity variables and the inconsistent use of diversity measures (Harrison, Sin 2006). This paragraph addresses these inconsistencies, first by giving a clear definition of diversity as understood throughout this paper, and second by discussing those measures that are most commonly used to assess work group heterogeneity. 2.2.1 Definitions of Diversity In many cases, studies on diversity focus exclusively on how diversity affects processes and outcomes of the units under observation. However, a controversial examination of the diversity construct itself is seldom found in the literature. Thus, it comes as no surprise that there is no joint agreement on how diversity can be defined. Instead, literature provides a plethora of divergent attempts that somehow dilute the understanding of diversity (Harrison, Sin, 2006). A pioneering definition of diversity in its psychological, sociological and economic sense has been proposed in a seminal work by Blau (1977). According to his understanding, heterogeneity on a given attribute depends on its distribution: “the larger the number of groups and the more evenly distributed the population is divided among them, the greater is heterogeneity” (Blau, 1977, p. 9). However, this definition requires variables to allow for rank-ordering and, thus, is strictly associated with status or power hierarchy. Scholars ever since have broadened this narrow definition by providing more extensive diversity concepts (Harrison, Sin, 2006). The fundamental question remains the same: how is maximum diversity defined. While the idea of minimum diversity simply reflects perfect homogeneity of a group in terms of a given attribute, maximum diversity seems to have a two-folded dimension. On the one hand, let us assume a group with all members having different characteristics along a particular attribute. This group may be assessed to reach the maximum degree of diversity. On the other hand, consider a group that can be divided into two subgroups that are clearly separated along a few but extremely opposing charac- Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 15 teristics. This group may be named maximal diverse as well. Moreover, one even might assume maximum diversity to appear in groups with one member clearly surpassing the others along a given characteristic. However, diversity literature has not put much attention on the clarification of these questions so far (Harrison, Klein 2007). In this meta-analysis, we understand diversity according to a definition recently proposed by Harrison and Klein (2007). Addressing the ambiguity of existing ideas on diversity, they distinguish between three types of group heterogeneity: separation, variety and disparity (see Figure 1.1). First, diversity as separation expresses horizontal differences in position within a group along attributes such as attitudes, values or opinions. Second, diversity as variety describes divergent categories among team members on a given attribute such as information or educational background. Third, diversity as disparity assesses vertical differences within a group on valued social assets such as pay or hierarchy. In general, the authors regard diversity as “the distribution of differences among the members of a unit with respect to a common attribute, X” (Harrison, Klein 2007, p. 1200). Figure 2.1: Diversity as Separation, Variety and Disparity Source: Harrison, Klein (2007), p. 1202. Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 16 2.2.2 Dimensions of Diversity Irrespective of its type, diversity is assumed to be attribute-specific. Put differently, a team is not diverse per se, but only regarding one or more clearly specified team member characteristics (Harrison, Klein, 2007). Any human being is a unique combination of features such as age, gender, cultural background, education, organizational tenure etc. In response to the large number of potential diversity attributes, scholars typically apply clustering in order to increase structure and manageability of data. The most fundamental cluster may be a distinction between easily observable attributes or surface level diversity (age, sex, and race) and less visible attributes or deep level diversity (tenure, education and function) (e.g. Harrison, Price, Bell 1998, Milliken, Martins 1996). Another commonly used sorting of diversity attributes has been used by Polzer, Milton and Swann (2002) who distinguish between demographic (age, sex, race, citizenship) and functional features (previous degree, previous job function, MBA concentration). A slightly different logic is proposed by Pelled (1996) as well as Pelled, Eisenhardt and Xin (1999): They classify attributes to be either task-related (education, organizational tenure, group tenure, functional background) or non-task related (age, gender, race). In this study, we follow their recommendation by differentiating between less task-related and highly taskrelated diversity attributes. 2.2.3 Measures of Diversity Usually, team diversity is assessed based on standardized heterogeneity measures at the group-level. The four most frequently used measures include the simple within-group standard deviation (s.d.), Blau’s (1977) index of heterogeneity, Allison’s (1978) coefficient of variation (CV) and Teachman’s (1980) index of heterogeneity. Within-group s.d. is one of the basic statistical measures to assess within-group variation in the distribution of a given attribute. It reaches its maximum in distributions that are bimodal, i.e. the sample split in half on a given attribute (Harrison, Sin 2006). The Blau (1977) index is one of the most widely used measures to assess diversity of categorical variables such as gender or age (e.g. Timmerman 2000). Maximum heterogeneity is reached when the proportions of all possible categories are equally distributed. Therefore, the Blau index minimizes with total homogeneity and approaches one with increasing heterogeneity (Harrison, Sin 2006). Teachman’s (1980) index may range from zero to positive infinity depending on the number of categories of any (categorical) variable. It is equal to Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 17 zero in situations where all team members belong to the same social category and increases with the number of different categories in the team (Harrison, Sin 2006). Allison’s CV (Allison 1978) is probably the most commonly used measure to assess within-group diversity (e.g. Williams, O’Reilly 1998). It is calculated by dividing the within-group standard deviation by the group mean. In contrast to other diversity measures, CV does not maximize when the variety of categories in the group increases but, likewise to s.d., grows with the magnitude of contrasts in the distribution of a given attribute. It reaches maximum values in teams with only one member being different to the others (Harrison, Sin 2006). The formulas for calculating these heterogeneity measures are presented in Table 2.1. However, all measures share a joint conceptual weakness as they are all prone to small team bias (Bedeian, Mossholder 2000) as well as distortions arising from divergent team sizes in only one sample (Biemann, Kerney 2010). Table 2.1: Formulas for the Calculation of Heterogeneity Measures Heterogeneity measure Formula Allison’s (1978) Coefficient of Variation 𝐶𝑉= √∑(𝑥𝑖− 𝑥)2 𝑁 𝑁 𝑖=1 𝑥 Blau’s (1977) Index 𝐵=1 − ∑𝑝𝑘2 𝑁 𝑘=1 Standard Deviation 𝑆𝐷= √∑(𝑥𝑖− 𝑥)2 𝑁 𝑁 𝑖=1 Teachman’s (1980) Index 𝑇= −(∑𝑝𝑘ln (𝑝𝑘)) 𝑁 𝑘=1 Note. N = total number of teams; xi = individual level characteristics; x = mean characteristics of the team; pk = proportion of team members in the kth category. Apart from the group-level heterogeneity measures presented above, there are other ways to conceptualize heterogeneity. First, relational diversity describes an individual’s (dis)similarity with his team mates and, thus, measures diversity at the individual level (e.g. O’Reilly, Caldwell, Barnett 1989, Tsui, Egan, O’Reilly 1992, Harrison, Price, Bell Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 24 teristics such as age, gender or cultural background to negatively moderate the link between group diversity and performance. Findings by Stewart (2006) support this argument. He confirms the positive moderating effect for top-management teams and project teams while observing a negative effect for production teams. Being aware of the cognitive nature of team processes in TMT and project teams in contrast to the repetitive and standardized work of production teams, these results come as no surprise. Based on these considerations, we predict: H2a: Team size negatively moderates the relationship between diversity and team performance for less task-related demographics. H2b: Team size positively moderates the relationship between diversity and team performance for highly task-related demographics. As discussed above, team size might moderate the diversity-performance relationship simply for statistical reasons since the most commonly used heterogeneity measures (e.g. Blau index of heterogeneity, coefficient of variation, Teachman index) are prone to biases arising either from small team size (Bedeian, Mossholder 2000) or from team size varying within the sample (Biemann, Kerney 2010). The necessity to test for statistical artefacts in the diversity-performance relationship affirms us in assessing team size as a potential moderator of the diversity performance relationship. Team type It is a well-established fact in the team literature that team performance and the efficiency of teamwork is highly dependent on the type of the team (e.g. Cohen, Bailey 1997, Thylefors, Persson, Hellström 2005). At the same time, there is broad evidence that team type moderates the diversity-performance relationship (e.g. Horwitz 2005, Stewart 2006). Team research applies multiple ways to distinguish between team types. For instance, Joshi and Roh (2009) differentiate teams along the longevity of cooperation. They assume visible diversity (e.g. age, gender) to be of less significance in teams designed for long-term cooperation (such as TMT or work groups deeply established in the organization), while invisible diversity (e.g. value, personality) is positively linked to length of cooperation (Harrison et al. 2002). Another strand of literature distinguishes teams along the tasks they perform (Sundstrom et al. 2000). In this study, we follow a team typology proposed by Devine (2002). First, intellectual work teams (such as TMTs) are assigned complex tasks in often uncertain situations and Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 25 are running non-standardized processes. Second, design teams (such as R&D) require creativity and technical innovation to fulfill their tasks. Third, physical work teams (such as production teams, performance teams or service teams) simply follow standardized tasks. We aim at testing for moderating effects along these three divergent team types. In general, TMTs and R&D teams are assumed to be heterogeneous on task-related attributes (i.e. functional and educational background) as the respective tasks require a broad base of varying knowledge and experience. At the same time, these teams are likely to be homogeneous on non-task related characteristics (i.e. age, gender, cultural background). In contrast to this, physical work teams are found to be more heterogeneous in their composition (e.g. Horwitz, Horwitz 2007, Devine 2002). As a consequence of these considerations, we derive two hypotheses on the impact of less task-related as well as highly task-related heterogeneity on team performance: H3a: The predicted negative relationship between less task-related heterogeneity and team performance will be weaker in R&D teams and TMT compared to other team types. H3b: The predicted positive relationship between highly task-related heterogeneity and team performance will be stronger in R&D teams and TMT compared to other team types. 2.4 Meta-Analysis Approach In the following, we use the meta-analysis approach to test for the above stated hypotheses on the diversity-performance relationship as well as to examine potential moderating effects of team size and team type. Literature review A thorough literature review was conducted in order to identify studies on workgroup diversity. We focus on research published between 1994 and 2014 since we aim at quantitatively summarizing the last 20 years of diversity research without replicating earlier metaanalyses that have already comprehensively reviewed earlier works (e.g. Bowers, Pharmer, Salas 2000, Webber, Donahue 2001). We located relevant studies on workgroup diversity for potential inclusion through several search strategies. First, we reviewed recent metaanalyses and narrative reviews by Bower, Pharmer, Salas (2000), Williams, O’Reilly (1998), Horwitz, Horwitz (2007), Stewart (2006), Webber, Donahue (2001), Bell (2007), Bell et al. (2011), De Dreu, Weingart (2003) and De Church, Mesmer-Magnus (2010). Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 26 Second, we electronically searched the online databases Scopus, Business Source Complete and PsycINFO with combinations of the following keywords: team, group, diversity, heterogeneity, homogeneity, diverse, (team) composition, demography, demographic diversity, age, gender, female, male, tenure, education, function, culture, race, ethnicity, performance, outcome, effectiveness and top-management-team. Third, the database search was supplemented with a manual search for in-press articles in both top-tiered journals and journals with a particular focus on group research including Administrative Science Quarterly, Organization Science, Journal of Applied Psychology, Academy of Management Journal, Academy of Management Review, Personnel Psychology, Group and Organizational Studies/Management, Organizational Behavior and Human Decision Processes, Journal of Organizational Behavior, Journal of Management, Small Group Research, Group Dynamics and Human Relations. We only consider peer-reviewed, Englishspeaking journal articles, thus excluding dissertations, conference papers, manuscripts or working papers. We are aware of the risk that this approach might put our results into question due to information loss as a consequence of publication bias. Yet, we proceed as described to be able to deal with the garbage-in-garbage-out issue (e.g. Egger, Smith, Sterne 2001) that seems of particular importance to us in the wide and complex field of diversity research. However, since published findings feature positive as well as negative relationships and do also report non-significant results at some instances, we feel confident to catch the substantial contributions without losing relevant information. Study abstracts were reviewed for suitable research on the relationship between demographic diversity variables and both team performance and creativity. This literature search yields a total of 242 studies that were further examined for potential inclusion in the meta-analysis. In order to be included, a study has to meet several inclusion criteria. First, we focus exclusively on empirical research that studies at least one of the hypothesized relationships between demographic diversity variables and workgroup productivity. In other words, we exclude studies that observe outcomes such as creativity, innovativeness, satisfaction or group cohesion. Additionally, we refrain from including aggregated measures of diversity such as demographic diversity, which is often used to combine age, gender and cultural background/ethnicity in only one variable. Moreover, we do not consider studies that observe psychological diversity variables such as cognitive ability or personality. Second, we only include studies that understood work groups as teams that meet the following criteria: Team members need to share a common goal, have similar or even identical working tasks Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 27 and be working interdependently. Moreover, teams should be embedded in an overall social system such as an organization and be regarded as a team by its members as well as by external persons (Guzzo, Dickson 1996). Third, since mixed levels of analysis are inappropriate for calculating sample-weighted effects (Beal et al. 2003), we focus on studies which assess diversity on the team-level and drop those studies that report findings on the individual level. Furthermore, we follow suggestions by Katzenbach and Smith (2005) and limit the maximum team size to twenty-five persons. Members of larger teams are usually not subject to task interdependency and do not adequately interact as a team, yet both factors are necessary to study diversity effects on team performance. Minimum team size is two persons. Fourth, we consider only studies that assess diversity based on objective measures such as within-group s.d., Blau’s (1977) index of heterogeneity, Allison’s (1978) CV and Teachman’s (1980) index of heterogeneity. Subjective measures of group heterogeneity are most likely biased due to individually differing perceptions of current group settings or cultural variations and are, therefore, not included in this meta-analysis. Additionally, we exclude studies that understand diversity either as relational diversity or as faultlines since both approaches are not comparable to the diversity concept applied here. Fifth, studies have to stem from real-life organizational settings. This excludes experimental research, case studies and simulation research design. Moreover, we concentrate on teams in working environments. Therefore, we exclude studies that are based on student work groups, sport teams or other non-task related organizational settings. We do so in order to guarantee high practical relevance of implications drawn from this meta-analysis. Eventually, we exclude studies that we asses to suffer methodological weaknesses and/or are not comparable to the overall set of studies. Applying these strict inclusion criteria is necessary to guarantee between-study comparability. A total of 179 studies out of 242 initially identified studies failed to meet the inclusion criteria. This leaves a final number of 63 studies to be included in our analysis. Two studies contribute two or more samples. As a result, we yield a final data set including 66 samples with information on a total of 12,478 teams. Diversity variables The independent variables investigated in this meta-analysis include the demographic features of age, gender, cultural/ethnical background, organizational tenure, functional and educational background as well as education level. To meet the inclusion criteria, diversity variables have to be aggregated to the team-level by means of an appropriate measure of Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 28 group-level heterogeneity, e.g. within-group s.d., the standard measures Blau’s (1977) index of heterogeneity, Allison’s (1978) CV and Teachman’s (1980) index of heterogeneity. Still, we allow for other methods to assess diversity objectively – such as the Herfindal index, the Gini index or the simple percentage share – as long as its choice seems reasonable and comparability with other measures is given. Table A.1 in the appendix lists the frequency distributions of diversity measures used in primary studies. Following Pelled (1996) in clustering diversity attributes along the dimension of taskrelatedness, we categorize age, gender and cultural background/ethnicity (referred to as culture in the following) as less task-related, while we rate organizational tenure, functional and educational background as well as education level to be highly task-related. While the metric of age, gender, culture and organizational tenure diversity is quite selfexplanatory and only varies slightly between primary studies (e.g. age and tenure measured continuously or in categories), three diversity variables need further consideration. Functional diversity is understood as functional experience gained during the career, the dominant function, organizational roles and the primary professional orientation. Educational background diversity is conceptualized as the major level or the specialization of studies. Education level diversity is assessed either as the number of years of (formal) education or as the highest educational level achieved. Outcome variables The dependent outcome variables include divergent types of team performance assessed along either objective or subjective performance measures. Objective performance measures include financial key figures such as return on equity (ROE), return on assets (ROA) and return on sales (ROS) as well as measures of team effectiveness such as productivity, profitability and goal achievement. Subjective performance measures are based either on a team’s self-assessed performance appraisals or on external ratings from managers, supervisors or other persons outside the teams. Moderator variables Eventually, in order to test for potential moderating effects in the diversity-performance relationship we include information on average team size and team type. We include mean team size when reported in the initial study – usually measured as simple headcount of team members. Studies that do not report average team size are excluded from the moderator analysis. Regarding team type, we cluster teams along Devine’s (2002) typology either Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 29 to be TMT, work team, R&D team or mixed teams. While TMT (executive) and R&D teams (design) match Devine’s (2002) intellectual work team cluster, work teams refer to the physical work team cluster. We understand “work teams” to include those teams introduced as work teams in the initial studies, but also teams that continuously work together on the same persisting tasks, e.g. teams labeled as school staff, sales teams or branch teams. Studies on work teams that are engaged in specific types of cooperation (e.g. virtual teams) or subject to task profiles that change over time (e.g. venture or project teams) are classified as “mixed teams”. The same is true for studies which include divergent team types without reporting results separately. Again, studies that do not provide the appropriate information on team type are excluded from moderator analysis at this instance. Meta-analytic procedure For this meta-analysis, we chose the product-moment correlation coefficient (r) to be the primary effect size index, because we exclusively focus on observational research and do not include experimental findings. However, product-moment correlation coefficients may entail serious statistical difficulties such as problematic standard error formulation (Lipsey, Wilson 2001) and excessive Type I error rates (Alexander, Scozzaro, Brodkin 1989). In response to these undesirable statistical properties, we follow suggestions by Lipsey and Wilson (2001) as well as Hedges and Olkin (1985) and correct correlation coefficients by Fisher’s Zr-transformation. With regard to model specification, we chose random effect models as introduced by Hedges and Olkin (1985) since this approach allows heterogeneity in the effect size to arise from two sources, subject level sampling error and variability of effects randomly distributed along studies (Lipsey, Wilson 2001). To test for the degree of precision of the mean effect size estimates, we calculated the 95% confidence interval (CI) around the population correlation. A CI that does not include zero indicates the mean effect size to be significant (Lipsey, Wilson 2001, Whitener 1990). If the 95% CI does not include zero, we follow Lipsey and Wilson (2001) in calculating z-tests to check whether the mean population effect size is significant at p≤α (e.g. 10%-significance). In order to test for the homogeneity of the effect size, we apply chi-square distributed Q statistics (Hedges, Olkin 1985). In general, homogeneity of effect sizes would imply that individual effect sizes differ from the estimated population mean effect size only by sampling error (Lipsey, Wilson Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 30 2001). 9 In other words, rejecting Q statistics would denote heterogeneous effect size distributions with two main implications. First, heterogeneity favors the use of random effects rather than fixed effects. Since we observe the majority of Q statistics to confirm heterogeneity in our data, our initial choice of random effect models is statistically supported. Second, heterogeneity suggests that randomly distributed variables may moderate the variance in individual effect sizes. As a consequence, we test for the proposed moderating effects of team size and team type by applying the DerSimonian and Laird (1986) random-effects meta-regression methods approach with Knapp and Hartung (2003) tests of effect estimates variance. QE is the corresponding residual heterogeneity statistic that indicates model significance (Lipsey, Wilson 2001). 2.5 Results We start the discussion of our findings by examining the relationship between diversity variables and team performance as hypothesized in H1a to H1g (Table 2.2). Subsequently, we present our findings with respect to the moderating effects of average team size (H2a, H2b, Table 2.3) as well as team type (H3a, H3b, Table 2.4). For each analysis, we report the number of correlations from independent samples included in the meta-analysis (k), the total number of teams (N), the effect size given as the corrected population correlation (rz) (sample-size weighted through Fisher’s Z-transformation), the standard error of the corrected population correlation (rz(se)), the lower and upper bounds of the 95% confidence interval (95% CI), Cochran’s Q as homogeneity statistics (Q) with the respective degree of freedom (df), and eventually the I2 statistic which presents the percentage of between-study variation attributable to heterogeneity (I2). Main results on the diversity-performance relationships (hypotheses H1a to H1g) are reported in Table 2.2. Forest plots that illustrate the correlations a study contributes to the meta-analysis are presented for each diversity variable in Figures A.1 to A.7 in the appendix. In support of Hypothesis H1a, we observe a negative effect size regarding the link between age heterogeneity and team outcome (rz = -0.052, 95% CI: -0.109 | 0.005). Since zero is included in the 95% CI, we calculate the z-test of effect size (ES) which allows to 9 As an alternative to Q-statistics we calculate the I2 measure that gives the proportion of total variation in mean effect size attributable to between-study heterogeneity (Higgins, Thompson 2002). It is calculated as 𝐼²=100% 𝑥 𝑄−𝑑𝑓 𝑄. As rule of thumb, one can say that an I² value around 25% indicates low heterogeneity, while values around 50% and 75% can be translated to moderate and high levels of heterogeneity, respectively (Higgins et al. 2003). Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 31 reject the null-hypothesis (ES=0) and, thus, confirm H1a at the 10%-level of significance. In other words, the more diverse team members are with regard to age the lower is overall group productivity. Hypothesis H1b predicted a negative relationship between gender heterogeneity and team performance. Our findings confirm the expected link (rz =-0.036, 95% CI: -0.076 | 0.004). Again, z-tests of effect size (ES) allow for confirmation of H1b (at the 10%-level). This indicates that gender heterogeneity in teams is detrimental to team performance. Regarding the relation between cultural diversity and team performance, we do not find any support for the expected negative relationship (H1c). Cultural diversity does not affect team productivity significantly. In summarizing findings on the less task-related diversity variables, results for both age and gender provide support for social categorization theory as well as the similarity-attraction paradigm since diversity along these attributes is identified to affect overall team performance negatively. Table 2.2: Relationship between Diversity and Overall Team Performance Diversity k rz rz(se) 95% CI Q df I2 Lower Upper Age 25 -0.052* 0.029 -0.109 0.005 88.79*** 24 73.00% Gender 30 -0.036* 0.021 -0.076 0.004 57.69*** 29 49.70% Culture 20 0.003 0.033 -0.061 0.067 105.98*** 19 82.10% Function 27 0.041 0.039 -0.035 0.117 369.37*** 26 93.00% Tenure 25 0.007 0.031 -0.054 0.069 178.85*** 24 86.60% Education Background 10 0.064* 0.038 -0.011 0.138 48.85*** 9 81.60% Level 12 -0.006 0.052 -0.108 0.097 43.23*** 11 74.60% Note. k = total number of correlation coefficients meta-analyzed; rz = corrected population correlation (sample-size weighted based effect size (ES) on Fisher’s z transformed correlation coefficients with significance test for ES=0; rz = standard error of the corrected population correlation; 95% CI = lower and upper bound of the 95% confidence interval; Q = homogeneity statistics; df = degree of freedom; I2 = percentage of between-study variation due to heterogeneity. ***p<.01; **p<.05; *p<.1 With regard to the expected positive relationship of heterogeneity in highly task-related attributes on team productivity, we only find support for educational background diversity as stated in H1f (rz = 0.064, 95% CI: -0.011 | 0.138). A significant z-test allows us to confirm H1f at the 10%-level of significance. This finding suggests teams to perform significantly better when drawing on a broad set of diverse educational backgrounds. However, there is no support for the predicted positive relationship between team performance and organizational tenure, functional and education level heterogeneity (H1d, H1e and H1g). Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 32 Therefore, overall support for the information processing perspective is weak and only based on findings with respect to educational background heterogeneity. In order to investigate the relationship between team heterogeneity and team performance in more depth, we separate primary studies depending on the way they assess team performance (Table A.2 in the appendix). While there are no outcomes other than those discussed above in the fields of age, gender, cultural, functional and educational background diversity, we gain additional insights for tenure diversity as well as education level diversity. In partial support of Hypothesis H1d, we observe tenure diversity to be positively related to team performance at least when performance measures are based on subjective external ratings, e.g. by supervisors (rz = 0.095, 95% CI: 0.021 | 0.106). Contrary to our expectations as expressed in hypothesis H1g, we found a negative relationship between educational level diversity and performance measured on outsider ratings (-0.132, 95% CI: - 0.267 | 0.002). All in all, we are able to confirm the expected negative link between team performance and both age diversity (H1a) and gender diversity (H1b). Furthermore, we find support for the hypothesized positive relationship between educational background diversity (H1f) and team performance. Additionally, we can provide partial support for the expected positive relation between team performance and tenure diversity (H1d). At the same time, we have to partially reject H1g as we observe a negative relationship between education level and performance rated by outsiders. Subsequent to the discussion of the main results of the meta-analyzed diversity-performance relationship, the second set of analyses centers on the potential moderating effects of average team size (as stated in H2a and H2b) and team type (as stated in H3a and H3b). Team size In order to test for the moderating effect of team size on the diversity-performance relationship, we applied random-effects meta-regression. Results are reported in Table 2.3. As stated in H2a, we expect team size to negatively moderate the link between heterogeneity in the less task-related demographic attributes of age, gender and culture. Meta-regression findings largely support team size to be the predicted negative moderator. In particular, the correlation coefficients for the diversity-performance relationship reduce with each additional unit member (for age by -0.001, for gender by -0.013 and for culture by -0.021). Highly significant values of QE confirm the significance of the regression models. These Work Team Diversity and the Optimal Composition of Teams: A Meta-Analytic Approach 33 findings may suggest that coordination and communication among team members are complicated with increasing team size. Moreover, social categorization and conflict may be more pronounced in larger teams. As predicted in H2b, we observe positive moderating effects on the diversity-performance relationship for the task-related demographic attributes tenure (0.003), functional (0.013) and educational background (0.063) as well as education level (0.016). These findings suggest that additional team members may contribute additional (complementary) knowledge, skills and experience. This seems to be beneficial to group performance at least for task-related heterogeneity. Overall, we are able to confirm hypotheses H2a and H2b. Table 2.3: Team Size as Moderator of the Performance-Diversity Relationship Diversity k rz rz(se) 95% CI QE df I2 Lower Upper Age 19 -0.001 0.008 -0.019 0.017 62.31*** 17 72.72% Gender 21 -0.013 0.005 -0.024 -0.001 35.54** 19 46.54% Culture 11 -0.021 0.013 -0.051 0.008 48.11*** 9 81.29% Function 15 0.013 0.016 -0.021 0.048 79.24*** 13 83.59% Tenure 14 0.003 0.022 -0.045 0.051 56.19*** 12 78.65% Education Background 6 0.063 0.049 -0.074 0.2 16.03*** 4 75.05% Level 11 0.016 0.026 -0.042 0.073 32.86*** 9 72.61% Note. k = total number of correlation coefficients meta-analyzed; rz = corrected population correlation (sample-size weighted based ES on Fisher’s z transformed correlation coefficients with significance test for ES=0; rz = standard error of the corrected population correlation; 95% CI = lower and upper bound of the 95% confidence interval; QE = homogeneity statistics; df = degree of freedom; I2 = percentage of between-study variation due to heterogeneity. ***p<.01; **p<.05; *p<.1 Team type Likewise to team size, we use meta-regression to test for moderating effects of team type on the diversity-performance relationship. Results are reported in Table 2.4. In hypothesis H3a we expected the negative relationship between less task-related heterogeneity and team performance to be weaker in TMT and R&D teams compared to other team types. For all three less task-related diversity attributes (age, gender and culture), we observe the predicted negative diversity-performance relationship for work teams and mixed teams. However, to our great surprise, results for TMT and R&D teams do not show the expected slightly weaker negative relationship but, instead, even suggest a positive relationship of age, gender and culture with team performance. In other words, for TMT and R&D teams Employee Absenteeism: Determinants in the International Context 40 overall working life span (e.g. Grund, Westergaard-Nielsen 2008). On this account there is a great need for reducing existing workplace burdens that might provoke stress or illnesses in order to sustain employees’ health in the long run. Detailed knowledge on the determinants of absence may help tackling these issues. In general, absence can be defined as non-attendance at work when scheduled to (e.g. Kristensen et al. 2006). This idea allows to distinguish between involuntary (e.g. sickness) or voluntary (e.g. shirking) absence behavior (e.g. Johns 1997, Sagie 1998). In lack of a unique theory on absence, several attempts to conceptualize absenteeism have emerged over the years with the well-known process model of Steers and Rhodes (1978) being the most commonly cited. In their tradition, absence is usually understood as being subject to various personal, organizational, social and environmental variables that influence both motivation and ability to attend work. Much of the (personnel) economics work on absenteeism relates to either neoclassical labor supply models (Allen 1981a/b, Dunn, Youngblood 1986) or efficiency wage theory (Shapiro, Stiglitz 1984). Theory is backed up by experimental and empirical results on demographics (e.g. Voss, Floderus, Diderichsen 2001; Barmby, Ercolani, Treble 2002), group absence norms (e.g. Bradley, Green, Leeves 2007; Bamberger, Biron 2007), work satisfaction (e.g. Kristensen et al. 2006), working conditions (e.g. Dionne, Dostie 2007), national sickness benefit regulations (e.g. Frick, Malo 2008; Johansson, Palme 2005), employment protection (e.g. Riphahn, Thalmaier 2001; Ichino, Riphahn 2005) as well as economic environment and unemployment (e.g. Leigh 1985) – yet at some instances with inconsistent findings. In this article, we empirically investigate why absence figures of blue-collar employees vary considerably between different production sites of a large corporation although employees face similar circumstances in terms of working conditions, manufacturing processes and final product at each facility. In particular, we are interested in the role of economic and social determinants in determining employee absence. Moreover, we examine the influence of worker characteristics and working conditions on absence. Using a hitherto unavailable data set covering 160 blue-collar work units at four international production sites of a large European automobile manufacturer, we offer a comprehensive analysis of potential determinants of absenteeism including:  social peer influences (team size, team turnover, the share of temporary employed workers and the share of workers suffering health impairments) Employee Absenteeism: Determinants in the International Context 41  economic influences / incentives (sick pay regulations, employment protection, (un)employment and prosperity level)  employee characteristics (age, gender, tenure and health)  working conditions (shift system) Following the data analyses, we conducted a total of eighteen interviews with on-site experts to further interpret our empirical findings. Experts include HR managers, line managers, worker representatives and shop floor staff. The results presented in this paper indicate that absence at the respective corporation is a multifaceted phenomenon that only shows few statistical patterns that are valid at all four plants under observation. Instead, all determinants examined are differently affecting employee absence at each of the four production sites – details are discussed throughout chapter 3.5. In a nutshell, we observe a positive link between absence and social peer influences such as unit size and unit turnover. Furthermore, a high share of temporary workers increases absence of permanently employed workers while we do not find any absence effect of the share of workers suffering health impairments. The results regarding economic influence suggest strict employment protection laws as well as a favorable prosperity level to significantly increase absence, whereas findings for the national (un)employment situation remain inconclusive. Results for national sickness benefits legislation fail to reach statistical significance. With regard to worker characteristics, we observe unit absence to increase with the share of female employees. Additionally, we detect a positive link between age and both involuntary as well as voluntary absence. Evidence on the relation between absence and tenure as well as individual health is somewhat inconclusive. Eventually, results on working conditions imply that absence is higher in a two-shift and three-shift systems compared to the standard one-(day-)shift system. In chapter 3.6, practical implications for attendance management are considered based on the empirical findings and insights gained during expert interviews. Our study contributes to the existing literature in several regards. First, we add to the growing research on determinants of absenteeism by offering new perspectives from inside an organization. In particular, we benefit from unique data that allows for comprehensive empirical testing. Second, we appear to be among the first researchers who are able to work with international data from within only one organization. Since all data is reported on company standards, we can neglect organizational and international differences in re- Employee Absenteeism: Determinants in the International Context 42 cording methodology that usually limit international comparison (e.g. European Foundation for the Improvement of Living and Working Conditions 2010). Third, our data allows us to examine three absence measures. Thus, we are able to empirically proxy for voluntary vs. involuntary absence. Furthermore, by using unit-level absence data we address a weakness of previous absence research that has focused almost exclusively on individuallevel absenteeism and only account for unit-level effects in patches (e.g. Rentsch, Steel 2003). However, since absence is usually modeled as a social phenomenon (e.g. Steers, Rhodes 1978, Kaiser 1998) it should, by definition, be influenced by unit-level peers. Finally, our study adds to the growing insider econometrics approach that has been first introduced to personnel economics in seminal papers of Ichniowski, Shaw and Prennushi (1997) as well as Lazear (2000b). The remainder of this paper is organized as follows. The upcoming section provides both a discussion of the absenteeism concept as understood in this paper as well as a review of the existing absence research. Section three will present the data set and the peculiarities of the organizational setting under observation. Section four discusses the estimation strategy while section five presents the empirical results. Eventually, section six concludes and offers implications for both practitioners and future academic research. 3.2 Theoretical Framework Research on absenteeism is multidisciplinary. It mainly originates from management literature but more recently arouse interest in (labor) economics and social psychology (Kaiser 1998). However, despite a substantial body of experimental and empirical evidence on absenteeism, a comprehensive understanding of the absence concept and its functioning is still missing. So far, research has merely agreed on a common definition of employee absence that includes all types of non-attendance at work when originally scheduled to, yet excluding scheduled absence agreed on with the employer such as holidays or flextime leaves (e.g. Kristensen et al. 2006). In other words, absenteeism is usually assigned to selfcertified or medically-certified sickness absence and recognized as such by the employer (e.g. Whitaker 2001). In this study, we follow this definition of absence. A “conceptual breakthrough” (Kaiser 1998, p. 81) in defining a theoretical framework of psychological absence research is the process model introduced in a now seminal article by Steers and Rhodes (1978). In their model, absence is understood as an individual decision Employee Absenteeism: Determinants in the International Context 43 based on the ability and the motivation to come to work (see Figure 3.1). In their model, the authors address weaknesses of previous contributions in questioning the assumptions of job satisfaction being the primary cause for absence (a comprehensive overview is given by Kaiser (1998)). In response, Steers and Rhodes (1978) model absence to be depending on both the motivation as well as the ability to attend work which in turn both are directly influenced by personal characteristics (e.g. age, gender, tenure). Additionally, the motivation to attend is further subject to the individual job situation (e.g. job scope, job level, work group size) as well as various attendance pressures (e.g. economic conditions, work group norms). The authors assume the ability to attend (e.g. sickness, family responsibilities) to moderate the effect of attendance motivation on absence. Later, the authors modified their first model into an advanced diagnostic model by including economic and social psychological factors that have not been considered in their initial model (Rhodes 1990). 13 Figure 3.1: Steers and Rhodes (1978) Process Model of Employee Absence Source: Illustration of Steers and Rhodes process model (Steers, Rhodes 1978, p.393). In economic research, the understanding of absence differs from the psychological concepts. Much of the economic absence research relates to two main theories. First, neoclassical labor supply models assume absence to occur from individual day-to-day labor- 13 Two alternative theories of absenteeism are discussed in the psychological absence literature. First, the withdrawal model which assumes absence to be withdrawal from unfortunate work circumstances (e.g. Hanisch, Hulin 1990, Porter, Steers 1973). In other words, absence may provide employees with stressrelief and allows them to return to work more productive (e.g. Bachler 1995). Second, the socialpsychological approach that emphasizes the importance of absence norms within groups (e.g. Kaiser 1998, Rentsch, Steel 2003). Employee Absenteeism: Determinants in the International Context 44 leisure choice decisions (e.g. Allen 1981a/b, Brown, Sessions 1996, Dunn, Youngblood 1986). The idea is simple: the individual utility is maximized when the marginal rate of substitution between income and leisure equals the wage offered by the employer. However, due to imperfect labor markets and job search costs, one can assume individuals to accept a job offer even if this constraint is not fulfilled at the contracted number of work hours. If at a given wage the contracted number exceeds the desired number of work hours, the individual has an incentive to be absent from work (Allen 1981a/b). Yet, this only holds true as long as the utility gains of being absent exceed the individual costs of absenteeism. These costs may include the gap between regular pay and sick pay. In other words, employees may use absence to maximize their individual utility (Dunn, Youngblood 1986). Therefore, absenteeism can be seen as “a desirable nonpecuniary element of the compensation package” (Allen 1981b, p. 207). Second, efficiency wage theory as discussed by Shapiro and Stiglitz (1984) assumes absence to arise from moral hazard and shirking. Therefore, absence may serve as a measure of individual effort level choice decisions and worker productivity and is widely used as such in the research (e.g. Ichniowski, Shaw 2013). The Shapiro-Stiglitz (1984) model is based on the assumption that the value of a present state (i.e. an employment relation) consists of the net current return as well as the expected future return. By being absent from work an employee risks dismissal and, as a consequence, the loss of the lifetime utility associated with the employment relation. Both neoclassical labor supply and efficiency wage theory share the assumption that absence, at least partially, is an employee’s individual decision under some constraints imposed by the employer. Hence, absence needs to be considered appearing either involuntary (e.g. as consequence of sickness or injury) or voluntary (e.g. to maximize individual utility). While involuntary absence spells are not expected to be influenced by motivation, voluntary absence can be modeled as utility maximization or shirking (e.g. Barmby, Sessions, Treble 1994, Brown, Sessions 1996, Chadwick-Jones, Brown, Nicholson 1973, Sagie 1998). Moreover, in the tradition of Steers and Rhodes (1978) absence can be interpreted as a social phenomenon with absence decisions never made solely by the individual, but instead being subject to certain constraints imposed by peers, subordinates, superiors and the organization’s overall absence culture (e.g. Chadwick-Jones, Nicholson, Brown 1982; Rentsch, Steel 2003). Nevertheless, a great part of absence research focus exclusively on individual level absence (e.g. Kaiser 1998, Rentsch, Steel 2003). Employee Absenteeism: Determinants in the International Context 45 Considering a potential sorting of determinants of absence variance we are loosely inspired by Kaiser (1998). He suggests absenteeism to be determined by personal characteristics and individual responses to influences stemming either from the work or non-work environment. Adding our own reading of the literature, we identify four fields of potential determinants of employee absenteeism: social peer influences in teams, economic influences / incentives, worker characteristics and working conditions. In what follows, we briefly discuss the determinants of absence along this classification in order to derive our main research hypotheses. 14 3.2.1 Social Peer Influences in Work Teams Turnover It has often been pointed out in absenteeism research that absence variance is small within and large between groups (e.g. Xie, Johns 2000). Usually, this phenomenon is attributed to social absence norms that exist either among groups, organizations or cultural spheres (e.g. Bamberger, Biron 2007, Rentsch, Steel 2003). Absence norms as understood in this paper can be defined as “set of absence-related beliefs, values and behavioral patterns that are shared among members of a work group or organizational unit” (Gellatly, Luchak 1998, p. 1086). Although there is comprehensive empirical evidence on the link between peer absence and individual absence (e.g. Martocchio 1994, de Paola 2010, Ichino, Maggi 2000, Rosenblatt, Shapira-Lishchinsky, Shirom 2010), little is known on the social mechanisms behind this relationship (Johns 1997). 15 Since team members are rational utilitymaximizers, we assume absence norms to be most likely in favor of the employees and, thus, increase absence. Now, consider a team whose members share a common belief about the legitimacy of absence. Within this team, the enforceability of the shared absence norm is subject to changes in the group composition since new workers are not aware of any existing absence norm or might not agree on participating (e.g. Miners et al. 1995). As a consequence, team-level absence can be assumed to decrease with team turnover. This argument leads us to hypothesis H1: H1: “Absence on the unit-level decreases with turnover.” 14 We limit our review of the literature to those determinants that will actually be surveyed within this study. For a comprehensive overview on absence research see, for instance, review articles and metastudies by Beemsterboer et al. (2009), Duijts et al. (2007) as well as Farrell and Stamm (1988). 15 One possible argument is built on self-categorization (Turner 1987) and social identification theory (Taijfel 1982) since individuals prefer working with similar peers and, thus, adopt a particular peer behavior. Employee Absenteeism: Determinants in the International Context 46 Unit size Referring to neoclassic labor supply models as well as efficiency wage theory, one can assume that some workers stay away from work without being genuinely sick. Given this assumption, researchers often model absenteeism as a potential indicator of shirking (e.g. Barmby, Orme, Treble 1995, Riphahn 2004, Bradley, Green, Leeves 2007). Economic theory suggests that the share of shirking employees usually depends on the probability of being detected as well as the potential consequences once being caught (e.g. Alchian, Demsetz 1972). In this context, the effect of group size on shirking is straightforward: Shirking can be deterred by appropriate actions either by employers (e.g. organized monitoring, incentive design) or by co-workers (e.g. mutual monitoring, social sanctions). 16 However, any of these actions will be harder to enforce with increasing group size. For instance, the ability to mutually monitor co-worker behavior is getting more difficult with each additional team member and, as a result, the effectiveness of social sanctions that discipline workers who are caught shirking declines. Likewise, monitoring by supervisors is getting more complicated and/or costly with increasing team size (Baron, Kreps 1999). Furthermore, from Bandiera, Barankay and Rasul (2005, 2009) we know about the importance of social ties and familiarity among team members in determining productivity. The magnitude of worker familiarity, however, can be assumed to decrease with team size. Ultimately, one can predict shirking to be facilitated in larger teams due to reduced monitoring and less familiarity. These predictions are expressed in hypothesis H2: H2: “Absence on the unit-level increases with team size.” Temporary workers It is widely recognized in economic research that temporary workers are less absent than permanent workers. Yet, as soon as taken on a permanent contract they significantly increase their absence behavior (e.g. Bradley, Green, Leeves 2007, Engellandt, Riphahn 2005). This phenomenon is mainly attributable to incentivizing contract characteristics since temporary agents seek to qualify themselves for permanent contracts that are usually associated with better working conditions (e.g. Aronsson 1999, Paoli, Merllié 2001) and higher pay (e.g. Mertens, Gash, McGinnity 2007). However, it remains an open question how temporary agents influence their permanently employed co-workers with regard to absence behavior. Two potential arguments may be considered. First, assume that permanently employed workers are aware of the incentivizing nature of temporary contracts. 16 On the importance of monitoring to reduce shirking in team settings recall Holmström (1982). Employee Absenteeism: Determinants in the International Context 47 Being rational utility-maximizers, permanent agents might allow themselves a more “generous” absence behavior as they suppose their temporary co-workers to be incentivized to fill in the gap left by their own absence. This implies that a high share of temporary workers increases the absence of the permanently employed co-workers due to shirking. Second, as mentioned above, temporary employment is usually associated with worse working conditions. Suppose, for instance, that temporary agents are appointed to the most burdensome working conditions, e.g. by being excluded from job rotation. In turn, this would allow permanently employed workers to stay on less demanding work tasks and suffer less work-related strains. Following this argument, a high share of temporary workers might decrease permanent workers’ absence. Since we know that temporary agents are appointed on the same tasks and meet identical working conditions than permanent workers at the observed company, we believe the latter argument to be less convincing in our study design. Therefore, we base our thoughts on shirking behavior and hypothesize: H3: “Absence of permanently employed team members increases with the share of temporary team members.” Workers with health limitations It is common practice at the studied organization that workers suffering from particular certified health limitations might be subject to temporary or permanent health impairments such as the prohibition to carry heavy loads or work overhead. As a result, these workers are exempt from performing tasks that would further deteriorate their health condition. Yet, it remains an open issue if the presence of employees with health impairments affects overall unit absenteeism. This question is mainly based on two considerations. First, it seems to be quite a straightforward argument that health-restricted employees are more prone to absence due to their physical (or mental) condition. Medical research largely emphasized the great significance of chronical and pre-existing health impairments in determining absence (e.g. Dewa, Lin 2000, Kessler et al. 2001). Second, the easier jobs in the team’s operations may often be occupied by workers with health impairments and, as a consequence, are excluded from unit-intern job rotation. In this situation, initially healthy employees might face additional burdens as they need to staff those workplaces that their team mates with health impairments cannot cover. These higher burdens may increase their own absence. Therefore, we expect a positive relationship between absence and the share of team members with health-impairments. Employee Absenteeism: Determinants in the International Context 48 H4: “Absence on the group-level increases with the share of team members with health impairments.” 3.2.2 Economic Influences / Incentives Sickness benefits It is a stylized fact in economics that the institutional framework determines employees’ absence behavior. Early evidence is given by Buzzard and Shaw (1952) who state that absence increase with higher sick pay replacement rates as the costs of absence decrease for employees. A corresponding labor supply model is offered by Brown and Sessions (1996). They argue that higher sickness benefits raise incentives of workers to be absent from work. These assumptions are supported by Frick and Malo (2008). Based on the European Survey of Working Conditions they observed a significant increase in absenteeism with more generous sickness benefits within the EU-14. Over the years, several EU governments initiated changes to national sickness legislation in order to address the burdens of sick pay for social security systems and employers. Absence research often benefits from these legislative changes that allow for natural experiments, e.g. Johansson and Palme (2002, 2005) as well as Voss, Floderus and Diderichsen (2001) for Sweden and Puhani and Sonderhoff (2010) as well as Ziebarth and Karlsson (2010, 2013) for Germany. These studies provide broad support of the positive link between statutory sickness benefit regulations and employee absence. We, therefore, predict: H5: “Absence will increase with more generous sickness benefit regulations.” Employment protection The same line of argument holds true for employment protection. Focusing on a legislative change in Sweden, Olsson (2009) observes that a less favorable labor protection law in terms of seniority dismissal protection significantly decreases absence behavior in organizations that are affected by the change. Based on German data, Riphahn and Thalmaier (2001) find that absenteeism increases after probation periods end and mandatory labor protection sets in. Further evidence is given by Ichino and Riphahn (2005) who observe the same behavior for Italian bank employees. Bradley, Green and Leeves (2012) report a significant raise in absenteeism when temporary workers are taken on permanently. They argue that the low employment protection associated with temporary employment incentivizes workers to show high effort in terms of low absence. As soon as labor protection ap- Employee Absenteeism: Determinants in the International Context 49 plies to their employment relation, the formerly temporary agents reduce effort and no longer refrain from shirking. This argument leads to hypothesis H6: H6: “Absence will increase with strict employment protection legislation.” Unemployment and prosperity level In addition to the legislative framework, the economic environment and unemployment rates are often linked to absence behavior beginning with a seminal paper by Leigh (1985). Using data from the time of the US recession in the 1970s, he observed employees to refrain from absenteeism when unemployment is high as they assume employers to first lay off workers who have proven absence-prone. Askildsen, Bratberg and Nilsen (2005) report similar findings for the Norwegian labor market: An increase in local unemployment rates significantly decreases absence. This result even holds when controlling for workforce composition which is often claimed to be responsible for cyclical variation in absenteeism. Knutsson and Goine (1998) observe the link between absence and unemployment for Swedish males. With regard to the economic situation, Virtanen et al. (2005) observed for Finish public sector employees that a constantly poor local economy has a decreasing impact on self-certified sickness. Audas, Goddard (2001) confirm the link between business cycle and absenteeism for the US. Thus, we predict: H7: “Absence will decrease with increasing unemployment.” H8: “Absence will increase with prosperity level.” 3.2.3 Worker Characteristics Non-work related worker characteristics (gender, age) The link between non-work related worker characteristics (i.e. gender and age) and employee absenteeism is well established. While for gender it is a stylized fact that females exhibit higher sickness absence than males (e.g. Barmby, Ercolani, Treble 2002, Mastekaasa 2000), results for age are less conclusive (e.g. Rhodes 1983). According to Bekker, Rutte and van Rijswijk (2009), there are three main reasons for gender differences in absenteeism. First, due to physical differences women suffer more from reproduction-related health issues such as pregnancy and menstruation (e.g. Alexanderson et al. 1996, Sydsjö, Sydsjö, Alexanderson 2001). Second, females face other daily obligations than males such as the double burden of job and family since child care is traditionally more associated with women (e.g. Åkerlind et al. 1996, Bratberg, Dahl, Risa 2002, Employee Absenteeism: Determinants in the International Context 56 Our study design has numerous advantages in addressing issues raised by previous research on absenteeism. First, starting with early work by Chadwick-Jones, Nicholson and Brown (1982) absence is not solely modeled as an individual phenomenon but as being subject to social peer influences. Given that within-unit absence varies only little whereas between-unit absence varies widely one can assume unit-level influences to play an important role in determining absence (e.g. Rentsch, Steel 2003, Harrison, Martocchino 1998). Thus, it seems surprising that relatively little work is done on absenteeism at the group level (e.g. Rentsch, Steel 2003) with only few exceptions (e.g. Xie, Johns 2000, Kristensen et al. 2006). We address this gap in absence research by offering new insights on sickness absence at the team level. Second, absence research mainly relies on selfreported sickness data, e.g. gathered via questionnaires, since register data is often unavailable to researchers. However, the response sensitivity of self-reported data in comparison to register data varies from high (e.g. 91% (Voss et al. 2008), 88% (Burdorf, Post, Bruggeling 1996)), to medium (e.g. 82% (Aguis et al. 1994)) and low (e.g. 55% (van Poppel et al. 2002)). In line with the latter, Johns (1994) finds employees to be absent actually twice as much as self-reported. 21 Since we are able to use register data from company records we can neglect concerns about the reliability and validity of our data. Third, since our data set includes four production plants from Germany, Spain and the UK we are able to control for varying institutional frameworks across these countries. As discussed earlier, regulations for sickness benefits and employment protection are found to significantly determine absence behavior in the international context (e.g. Frick, Malo 2008, Riphahn, Thalmaier 2001). However, international comparability of absence statistics usually proves difficult due to widely varying reporting standards for specific cases such as family care or maternity (European Foundation for the Improvement of Living and Working Conditions 2010). For this study, we can dispel any doubts about international comparability since all absence data is recorded based on company standards. Finally, we are able to apply multiple measures of absence as dependent variables to account for different aspects of absenteeism such as voluntary and involuntary absence. All absence measures used in this study are discussed in detail in the following section. 21 Moreover, accuracy of self-reported sickness absence is found to decrease with time (Severens et al. 2000) and with increasing number of sickness days (Ferrie et al. 2005, Johns 1994). Employee Absenteeism: Determinants in the International Context 57 3.4 Estimation Strategy Our empirical analyses are based on three different absence measures to consider absenteeism from various perspectives. We use absence KPIs reported on company recording standards. The most important absence figure is a unit’s monthly absence rate. It is defined as the percentage of time originally scheduled for work within the unit missed due to selfcertified and medically-certified sickness. All company-intern discussion on absence is entirely centered on this measure. Absence rates vary across units and locations, the overall statistics are presented in Table 3.2. Moreover, we follow Hensing et al. (1998) 22 in constructing further absence measures since working with register data does not allow for a clear separation between involuntary and voluntary absence. We, therefore, apply two monthly proxies for both absence types proposed among others by Sagie (1998). First, one can assume involuntary absence spells to be usually longer since sicknesses and injuries are seldom cured completely after one or two days. Hence, we calculate the mean duration of sick leaves of a team rounded to integer values (Hensing et al. 1998). Second, voluntary absence is widely recognized as short-lasting but more frequent. That is why frequency measures serve well as proxy for voluntary absence (e.g. Chadwick-Jones et al. 1971). We follow Hensing et al. (1998) in calculating a frequency measure by dividing the number of absence spells of a given team by the number of persons in that particular unit. Still, a definitive separation between involuntary and voluntary absence remains unclear since both proxies might include the other absence behavior as well. In general, our estimations are based on the following models with ABS representing the three absence measures discussed above as dependent variables: ABS = α + βPEER + γ ECON + δWORK + φ COND + 𝜗CONT + ε where α is the constant, PEER is a vector of social peer influences on the team level, ECON is a vector of economic influences / incentives, WORK is a vector of worker characteristics, COND is a vector describing working conditions and CONT is a vector of further controls, β, γ, δ, φ and ϑ are the estimated coefficients and ε is the error term. All variables are discussed in more detail below. 22 Hensing et al. (1998) propose frequency, length, incidence rate, cumulative incidences and duration of sick-leave as measures of absenteeism. Although we are not able to construct all measures for our sample due to unavailable or missing data, we use their frequency and duration measures. Employee Absenteeism: Determinants in the International Context 58 Social peer influences (PEER) We model social influences of workers’ peers to study their impact on employee absence. One important peer influence is focusing on group absence norms (e.g. Bamberger, Biron 2007, Rentsch, Steel 2003). These norms might change with turnover (Miners et al. 1995). With regard to hypothesis H1 we proxy for turnover by adding all new arrivals to and all exits from the team in order to create one overall turnover variable. The idea is simple: we use this quite intuitive measure of variance in team composition since absence norms should be influenced by both new and leaving workers alike. In addition, we address research on shirking in team settings (e.g. Dunn, Youngblood 1986) by stating that with increasing team size absence should be more pronounced due to facilitated shirking options (H2). To incorporate this idea in our estimations we include monthly team size for all units. We measure team size as the simple headcount of team members in a given month. As expressed by hypothesis H3, the share of temporary workers might increase the absence behavior of permanent workers. In order to control for this argument, we include the percentage of temporary workers in a unit in our estimations. Unfortunately, plant C did not report any information on the share of temporary workers. In hypothesis H4 we state that the share of team members with health impairments should increase unit-level absence for two reasons. First, most obviously these workers can be expected to be more prone to absence. Second, healthy workers might suffer additional workload. We take this hypothesized relationship into account by including the percentage of team members that have any kind of certified health impairment in a given month. Economic influences / incentives (ECON) In order to study potential economic influences and incentive effects on employee absence, we extend the company data set with environmental variables identified by previous research to determine employee absenteeism. At this point, we benefit from the international context of our study since we are able to compare different social security systems and employment protection laws as well as varying national economic environments. This allows us to consider potential economic influences along the four main topics introduced in H5 to H8. First, as proposed among others by Riphahn and Thalmaier (2001) as well as Ichino and Riphahn (2005) employment protection may have a significant impact on employee absence behavior. To account for varying legislative settings in Germany, Spain and the UK we incorporate the OECD’s Indicators of Employment Protection (OECD 2014a). Three synthetic indicators rate the strictness of a country’s regulations on individ- Employee Absenteeism: Determinants in the International Context 59 ual and collective dismissal as well as on fixed-term and temporary work agency contracts on a scale from weak=0 to very strict=6. Second, we consider suggestions stating that absenteeism increases with more generous sickness benefits (e.g. Frick, Malo 2008, Johansson, Palme 2002, 2005) in following an approach proposed by Frick and Malo (2008). They use the so-called MISSOC reports (Mutual Information System on Social Protection) published by the European Commission that describe national sickness benefits concerning their coverage, waiting period, maximum duration and replacement rate (European Commission 2014a). We convert the report’s textual descriptions into a scale from weak=0 to generous=6 sickness benefits in order to match the OECD’s Indicator of Employment Protection scales. Third, following previous evidence (e.g. Leigh 1985, Askildsen, Bratberg, Nilsen 2005) we expect high unemployment rates to significantly influence employee absence behavior. Hence, we control for national (un)employment in two ways. On the one hand, we use seasonally adjusted monthly national unemployment rates. On the other hand, we include the share of persons employed in manufacturing (NACE Classification D) in percent of total employees to model workers outside options. Data is based on quarterly Eurostat statistics for both unemployment (Eurostat 2014a) as well as employment information (Eurostat 2014b). Finally, we take into consideration research stating a link between national economic situation and employee absence behavior (e.g. Virtanen et al. 2005, Audas, Goddard 2001) by incorporating Worldbank information on annual percentage changes of GDP by country (Worldbank 2014). Worker characteristics (WORK) To account for the influence of worker characteristics on absence we chose four main potential determinants of absenteeism. On the one hand, we include the classical non-work- related demographic features of age and gender to test hypotheses H9 to H11. First, age is included as unit mean age. Although the measure of age on the group level seems somehow not ideal since similar age values in groups may be caused by totally different age distributions (e.g. Kristensen et al. 2006), we believe that controlling for the coefficient of variation (CV) of age should level out potential biases. Second, we use the share of males to model the gender distribution within a team. On the other hand, we aim at addressing potential influences of work-related individual demographics on absence. We, therefore, use unit mean tenure and the respective CV as proxies for professional experience. Following this logic, a high mean tenure indicates units with workers that have accumulated high firm-specific human capital and, thus, should have better strategies to cope with work- Employee Absenteeism: Determinants in the International Context 60 related stressors. Although age and tenure are highly correlated (0.698), results do not change when leaving tenure out from the estimations. Therefore, it seems that age and tenure variables are actually measuring two different phenomena. This is in line with among others Barmby, Ercolani and Treble (2002) who observe a tenure effect on absence even when age is controlled for. Unfortunately, CVs were calculated neither for age nor for tenure from plant D. Eventually, we address research on acute health condition with influenza and ILI being among the most stated reasons for staying away from work (e.g. Tsai, Zhou, Kim 2014, Keech, Scott, Ryan 1998). 23 As suggested in hypothesis H13, we control for ILI and influenza by using influenza recordings from the World Health Organization (WHO). These statistics assess influenza activity from no activity to widespread outbreak on a national level (WHO 2014). We convert weekly textual evaluations to monthly ordinal scales of no activity=0, sporadic=1, local outbreak=2, regional outbreak=3 and widespread outbreak=4. Working conditions (COND) In H14 we address research stating that shift work usually increases absence due to higher burdens concerning sleeping habits, disordered diurnal rhythm and complicated family and social life (Slany et al. 2014, Morikawa et al. 2001, Fekedulegn et al. 2013). By taking a normal seven hour one-shift system (only morning shift) as reference category, we model additional shifts based on a two-shift (morning and day shift) or three-shift systems (morning, day and night shift). Control variables (CONT) In order to further exclude confounding effects on employee absence behavior, we add further controls to the company data set. First, we address evidence on the influence of weather on absenteeism (e.g. Shi, Skuterud 2015) by including information on monthly mean temperature, precipitation and sun hours. 24 Second, we use month and year dummies to control for potential seasonality or other timing effects of absence. 23 Yet, we do not have information on the individual health condition of employees or work units. 24 Since we aim to ensure data comparability by relying on single-sourced data provided from the international network of the German Meteorological Service (DWD 2014), we only have a limited choice of meteorological stations close to the plants’ locations. Therefore, weather information was gathered at meteorological stations 130km away from sites at maximum. We do not assume this to be a problem. Employee Absenteeism: Determinants in the International Context 61 3.5 Empirical Results As a first step, we conduct one-way ANOVAs in order to check for significant mean variance in absence rates between and/or within organizational units. While we find betweenunit variance in absence rates to be highly significant at all plants (plant A (F(5, 192)=6.56, p=.000), plant B (F(7, 256)=8.60, p=.000), plant C (F(103, 3564)=9.67, p=.000) and plant D (F(41, 1419)=4.28, p=.000)), we observe within-unit absence to vary significantly at only 69 out of 160 units over time (43%). This phenomenon is in line with previous research that found absence variance to be great between units but smaller within units (e.g. Harrison, Martocchio 1998, Xie, Johns 2000). In response to this finding, we chose fixed effects regression models to be preferable over random effects models since there seems to be some group specific effects that remain constant over time. Fixed effects may be of psychological nature but may also be attributable to varying workplace characteristics that might increase ergonomic burdens of work. We, therefore, estimate fixed-effect models on all three absence measures to test for potential determinants of employee absenteeism. 25 In this process, we follow a two-staged estimation strategy. On the one hand, we pool data of two plants (both volume car production sites and both luxury car production sites, respectively) and apply pair-level estimations by using interaction terms for economic and social peer influences. This process seeks at avoiding potential biases arising from the fact that variables may not only vary between plants but may determine absence behavior at each plant in different ways. On the other hand, we apply plant-level estimations for each location independently to further control for plant-specific effects that might get lost in pair-wise regressions. For all estimations we use various model specifications including different sets of explanatory variables to control for the robustness of our findings. Subsequent to the empirical testing, we conducted a total of 18 on-site expert interviews at both international sites (B, C) in order to double-check our own experience from national plants. Experts are working in different organizational fields including HR experts, worker representatives, shop floor managers and shop floor staff. Their operational experience and practical knowledge helped us to interpret our findings comprehensively and derive implications that are feasible and realizable in organizational implementation. 25 Other researchers often use count data methods such as (zero-inflated) poisson or negative binomial regressions to study absence duration. Unfortunately, our duration data represents a unit’s mean and, therefore, does not meet the integer value assumption necessary for both regression models. Employee Absenteeism: Determinants in the International Context 62 In the following, we start the detailed discussion of our findings by first presenting estimation results on the effects of social peer influences on worker absence (H1 to H4). We then go on to present our findings with respect to economic influences / incentives (H5 to H8), worker characteristics (H9 to H13) and working conditions (H14). Within each paragraph we discuss pooled pair-level as well as plant-level results and further distinguish our findings along the three main dependent variables absence rate, absence spell duration and absence frequency. Overall findings suggest that determinants of absence vary by plant. In other words, we only find few patterns in the data that are common at all four plants. Main results of pooled interaction estimations are reported in Tables 3.3 to 3.5, main results of plant-level estimations are reported in Tables A.1 to A.3 in the appendix. Estimations in all tables are numbered for a better orientation and referred to throughout the text. Due to reasons of brevity we focus on presenting the main results in all tables. Social peer influences (PEER) Starting with peer effects, we have analyzed four potential determinants: turnover as a proxy for group norms, unit size as a proxy for shirking, the share of temporary workers and the share of workers with temporary or permanent health impairments. Results will be discussed in the specified order. In hypothesis H1, we predict absence to decrease with turnover because the enforceability of group norms – that we believe to increase voluntary absence – may be limited with variance in team composition. Yet, we find only weak support for this argument as the relationship between turnover and absence spell duration is only significant at plant D (estimation VI-4). However, since duration measures are usually used to proxy for involuntary absence, we refrain from interpreting this finding in support of H1. Instead, plant level estimations for location D (V-4) as well as interaction term estimations for locations C and D (III-4, -5, -6) indicate that high turnover increases absence frequency. Results for absence rate do not reach statistical significance. The great majority of local experts confirmed the empirical findings as they observe absence to be usually higher in units with high employee turnover. According to their experience, this behavior is attributable to employees feeling unsettled and uncertain rather than to changing group norms. Moreover, turnover and absenteeism are both identified to be strategies workers may choose in response to individual job dissatisfaction (theory of exit, voice and loyalty, e.g. Hirschman 1970, Farrell 1983). Thus, high turnover may reduce the workers’ need of other responses to job dissatisfaction such as absence. As a consequence, absence can be assumed to be low in high turnover situations. In total, we have to reject hypothesis H1. Table 3.3: Fixed-Effects Interaction Terms on Absence Rates (I) luxury sites volume sites I-(1) I-(2) I-(3) I-(4) I-(5) I-(6) VARIABLES Plant A (GER) Plant B (UK) Plant A (GER) Plant B (UK) Plant A (GER) Plant B (UK) Plant C (ESP) Plant D (GER) Plant C (ESP) Plant D (GER) Plant C (ESP) Plant D (GER) Mean age (in years) -0.949 -0.611 -0.206 0.703* -0.396 -1.512 -0.476 1.487 0.021 0.501** 1.601 0.257 (7.203) (2.612) (0.494) (0.385) (7.600) (2.699) (1.501) (1.908) (0.163) (0.223) (2.635) (2.762) Mean age² (in years) 0.00949 0.0170 --- --- 0.00260 0.0193 0.0059 -0.0134 --- --- -0.022 0.00293 (0.0994) (0.0287) (0.105) (0.0299) (0.0201) (0.0236) (0.035) (0.0338) Share of males -19.74 5.273 -17.50 7.464 -15.52 -0.192 -2.78* -0.297 -6.988*** 0.972 -6.751*** -0.364 (in %/100) (18.96) (8.150) (15.06) (7.947) (20.10) (9.218) (1.429) (4.916) (2.406) (6.67) (2.342) (6.735) Mean tenure (in years) 2.533 -2.223* -3.027** -0.344 -0.967 -1.472 0.0492 0.5573 0.2374 -0.5294 0.1902 1.457 (3.193) (1.223) (1.378) (0.226) (4.925) (1.015) (0.4074) (1.1011) (0.1652) (0.4025) (0.5842) (1.520) Mean tenure² (in years) -0.405 0.0521 --- --- -0.131 0.0426 0.0012 -0.0445 --- --- -0.0002 -0.0761 (0.260) (0.0336) (0.379) (0.0289) (0.0148) (0.0401) (0.0218) (0.0553) Unit size (in persons) 0.186** -0.0174 0.160*** -0.0322 0.136* -0.0264 --- -0.122 0.0464 0.0595 0.0611 0.0647 (0.0624) (0.0541) (0.0491) (0.0521) (0.0698) (0.0502) (0.148) (0.06004) (0.168) (0.058) (0.158) Influenza 0.0602 0.101 0.144 (0.248) -0.328 -0.5543 -0.816* (0=no activity,…) (0.275) (0.255) (0.291) (0.454) (0.460) Employment protection 6.857 6.423 --- 1.924 0.867 --- legislation (0=weak,…) (3.882) (3.743) (3.875) (8.749) Employment in 106.2 88.02 --- 128.565 153.555 --- manufacturing (in %/100) (180.0) (202.2) (91.666) (1.981) Unemployment rate -11.46 -30.00 --- 3.3119 13.334* --- (in %/100) (96.23) (96.09) (14.264) (26.688) Prosperity level (annual 10.56* 10.06* --- 6.1809** 3.495 --- change in GDP in %/100) (5.490) (5.673) (2.541) (4.9687) 2-Shift system (yes=1) 1.442 1.281 0.826 1.921 2.697 1.608 (1.083) (1.136) (1.229) (1.262) (1.699) (1.577) 3-Shift system (yes=1) n/a n/a n/a 1.208 1.787 1.261 (1.152) (1.303) (1.266) Constant 3.501 -13.69 50.39 -21.86 -8.137 -16.24 (60.35) (35.07) (60.45) (25.73) (24.844) (35.19) Observations 471 471 471 4,878 2,128 2,128 R-squared 0.189 0.182 0.157 0.042 0.07 0.068 Number of units 14 14 14 146 146 146 Robust standard errors in parentheses | *** p<0.01, ** p<0.05, * p<0.1 + all estimations including month and weather dummies & age and tenure coefficients of variation | only main results presented | n/a= not available Employee Absenteeism: Determinants in the International Context 63 Table 3.4: Fixed-Effects Interaction Terms on Absence Spell Duration (II) luxury sites volume sites II-(1) II-(2) II-(3) II-(4) II-(5) II-(6) VARIABLES Plant A (GER) Plant B (UK) Plant A (GER) Plant B (UK) Plant A (GER) Plant B (UK) Plant C (ESP) Plant D (GER) Plant C (ESP) Plant D (GER) Plant C (ESP) Plant D (GER) Mean age (in years) -0.276 -5.748 -0.274 0.0202 -0.540 -6.717 -2.24 0.950 0.0628 0.653** 2.566 -1.882 (2.570) (4.142) (0.388) (0.384) (2.407) (4.601) (2.345) (2.588) (0.2221) (0.302) (3.603) (3.091) Mean age² (in years) 0.000915 0.0673 --- --- 0.00411 0.0837 0.0301 -0.00550 --- --- -0.0347 0.0307 (0.0355) (0.0470) (0.0337) (0.0527) (0.0317) (0.0317) (0.0484) (0.0379) Share of males -21.42* 4.607 -24.3** 6.889 -22.85* 4.248 -2.721 -4.406 -5.2565* -4.135 -5.017* -5.513 (in %/100) (11.35) (6.781) (9.569) (7.457) (12.78) (6.197) (1.789) (6.571) (3.0911) (8.806) (3.027) (8.841) Mean tenure (in years) -5.607 -0.563 -1.594 0.102 -5.164 -1.798 -0.1032 -0.604 0.4505** -0.734 -0.0047 1.354 (4.232) (1.034) (1.612) (0.222) (4.623) (1.029) (0.4351) (1.583) (0.2127) (0.566) (0.722) (2.377) Mean tenure² (in years) 0.295 0.0179 --- --- 0.283 0.0462 0.013 0.000880 --- --- 0.0167 -0.0799 (0.278) (0.0285) (0.292) (0.0287) (0.0174) (0.0594) (0.0291) (0.0926) Unit size (in persons) 0.0475 -0.0163 0.0735* -0.0612 0.0691* 0.00821 --- -0.0781 0.2712*** 0.399* 0.299*** 0.428* (0.0362) (0.0888) (0.0399) (0.0704) (0.0373) (0.0927) (0.146) (0.0845) (0.236) (0.0833) (0.232) Share of temporary -0.306 -0.112 -1.233 --- --- --- workers (in %/100) (1.902) (1.704) (2.073) Employment protection -4.271 -4.687 --- 4.748 1.586 --- legislation (0=weak,…) (3.690) (3.655) (5.002) (10.27) Unemployment rate -124.6 -133.0 --- 9.535 -23.96 --- (in %/100) (105.3) (102.0) (18.03) (29.68) Prosperity level (annual 1.868 3.022 --- 7.583** 2.167 --- change in GDP in %/100) (4.328) (4.416) (3.558) (6.486) 2-Shift system (yes=1) 0.321 0.460 1.037 3.652** 3.606* 3.606* (0.765) (0.695) (0.764) (1.714) (2.144) (2.144) 3-Shift system (yes=1) n/a n/a n/a 2.503* 2.593* 2.593* (1.371) (1.499) (1.499) Constant 99.98 14.80 120.3* -4.780 -12.28 -15.59 (63.79) (38.29) (58.64) (37.53) (30.78) (45.73) Observations 471 471 471 4,817 2,129 2,129 R-squared 0.141 0.134 0.123 0.033 0.052 0.049 Number of units 14 14 14 143 143 143 Robust standard errors in parentheses | *** p<0.01, ** p<0.05, * p<0.1 + all estimations including month and weather dummies & age and tenure coefficients of variation | only main results presented | n/a= not available Employee Absenteeism: Determinants in the International Context 64 Table 3.5: Fixed-Effects Interaction Terms on Absence Frequency (III) luxury sites volume sites III-(1) III-(2) III-(3) III-(4) III-(5) III-(6) VARIABLES Plant A (GER) Plant B (UK) Plant A (GER) Plant B (UK) Plant A (GER) Plant B (UK) Plant C (ESP) Plant D (GER) Plant C (ESP) Plant D (GER) Plant C (ESP) Plant D (GER) Mean age (in years) -0.191 0.202* 0.00238 0.0199** -0.176 0.180 -0.00762 0.0188 0.0024 -1.10e-06 0.0237 0.0461 (0.188) (0.110) (0.0118) (0.00886) (0.193) (0.109) (0.0244) (0.04497) (0.0027) (0.00428) (0.0384) (0.0590) Mean age² (in years) 0.00261 -0.00205 --- --- 0.00238 -0.00198 0.00011 -0.00025 --- --- -0.00028 -0.00056 (0.00267) (0.00121) (0.00273) (0.00122) (0.00032) (0.00054) (0.0005) (0.00069) Share of males (in %/100) 1.074 0.0559 1.206* 0.0700 1.152 -0.0761 -0.056** 0.0593 -0.087** 0.04906 -0.081** 0.0201 (0.626) (0.301) (0.598) (0.327) (0.653) (0.293) (0.0256) (0.0795) (0.0368) (0.113) (0.0362) (0.107) Mean tenure (in years) 0.402** -0.0723 -0.0490 -0.0112* 0.293* -0.0609 0.0044 0.0442 0.0018 -0.0002 0.0072 0.0370 (0.157) (0.0612) (0.0517) (0.00594) (0.165) (0.0577) (0.0055) (0.0371) (0.00094) (0.00781) (0.00797) (0.0419) Mean tenure² (in years) -0.0331** 0.00173 --- --- -0.0249** 0.00168 -0.0002 -0.00139 --- --- -0.00027 -0.00141 (0.0115) (0.00174) (0.0114) (0.00166) (0.00021) (0.00139) (0.00031) (0.00152) Unit size (in persons) 0.00392* -0.00297 0.00211 -0.00176 0.00263 -0.00304 --- -0.00105 0.00104 -0.00228 0.00104 -0.00233 (0.00220) (0.00176) (0.00190) (0.00138) (0.00234) (0.00175) (0.0033) (0.00094) (0.00340) (0.00094) (0.00335) Turnover 0.000184 0.000300 0.000199 0.0011* 0.00157** 0.00179** (arrivals & exits) (0.000365) (0.000412) (0.000365) (0.00064) (0.00082) (0.000814) Share of temporary 0.105 0.0847 0.157* --- -0.0114 -0.00426 workers (in %/100) (0.0768) (0.0920) (0.0815) (0.0598) (0.0565) Employment protection 0.161** 0.159** --- -0.0984 -0.0122 --- legislation (0=weak,…) (0.0638) (0.0680) (0.0782) (0.1597) Employment in 6.746 4.146 --- 3.081 5.511** --- manufacturing (in %/100) (5.214) (4.832) (1.909) (3.0594) Unemployment rate -0.359 -0.905 --- -0.0436* 0.6342 --- (in %/100) (1.980) (2.365) --- (0.2449) (0.641) Prosperity level (annual 0.245 0.179 0.000921 0.3009*** --- change in GDP in %/100) (0.156) (0.172) (0.000630) (0.1112) 3-Shift system (yes=1) n/a n/a n/a 0.0201 0.0290* 0.0268 (0.0155) (0.0174) (0.0173) Constant -3.049 -1.400 -1.242 (2.060) -0.1187 -0.8429* -0.595 (1.977) (1.045) (0.4345) (0.5814) (0.611) Observations 471 471 0.239 471 4,810 2,123 2,123 R-squared 0.272 0.250 14 0.053 0.064 0.061 Number of unit 14 14 143 143 143 Robust standard errors in parentheses | *** p<0.01, ** p<0.05, * p<0.1 + all estimations including month and weather dummies & age and tenure coefficients of variation | only main results presented | n/a= not available Employee Absenteeism: Determinants in the International Context 65 Employee Absenteeism: Determinants in the International Context 72 vide support for hypothesis H13 – there is no link between absence and acute illness measured by influenza statistics detectable in our data. Working condition variables (COND) Following empirical results by Slany et al. (2014), Morikawa et al. (2001) and Fekedulegn et al. (2013), we predict in H14 that working hours other than the usually applied one-shift systems (day) increase absence due to negative health and social impacts coming along with irregular and/or atypical working hours. In our estimations, we observe two-shift systems to increase absence rates in comparison to the reference one-shift system at both international plants B and C by 2.8 (IV-2) and 2.3 (IV-3) percentage points. For plant A, we find a significant effect in the opposed direction (IV-1). However, we believe this result to be highly biased since at plant A only three units worked a one-shift system for only six months, while operating on a two-shift system for the remaining observation period. Concerning absence spell duration, pooled estimations for both volume car producers C and D reveal statistically significant effects of two-shift and three-shift systems (II-4, -5, -6). Plant-level estimations for two-shift systems at plants B (V-2) and C (V-3) and for a threeshift system at plant C (V-3) support this finding. Referring to absence frequency, we find a weak support for the negative impact of a two-shift system at plants B (VI-2) and C (VI- 3) and a three-shift system at plants C and D (III-5). On-site experts acknowledge that night shifts are usually the most exhausting working hours for the majority of the work force as they are opposing the human sleeping pattern. In particular, the transition from night shift to the subsequent day shift is exceptionally burdening. Interestingly, experts report that absence is lowest in night shift weeks since employees do not want to forfeit night shift premiums which can amount up to 30% of their normal pay (see Stein (2015) for empirical support). Still, we are able to partially confirm our hypothesis H14. Additional shifts significantly increase employee absence at some plants. Control variables (CONT) Although no hypotheses were derived for date and weather controls, our estimations reveal some interesting insights. 31 At both volume car producers, absence rates and spell duration are significantly lower during most summer months from April to September in comparison to the reference period (January) – even when influenza activity and weather is controlled for. However, this phenomenon is most likely attributable to the summer holidays 31 Results for control variables are not tabulated in this paper. Employee Absenteeism: Determinants in the International Context 73 as absence is usually not reported while being on holiday. Findings for absence frequency are weak and inconsistent with no clear pattern. Concerning weather, we find support for arguments by Shi and Skuterud (2015) who state that employees adapt their absence behavior to the weather conditions. At the Spanish site, we find absence rate, duration and frequency all increase significantly with temperature. For instance, unit-level absence spell duration increases by 0.47 days with each degree Celsius at the 1%-level of significance. This finding is in line with the impressions of on-site experts at this plant. Same results are found at least for absence spell duration at plant D (0.2 days, 1%-level). No other effects of temperature, sunshine hours and precipitation are found. In other words, we observe a link between weather and absence at some plants indicating that employees may adapt their absence behavior to favorable weather conditions. This behavior can be interpreted as shirking. In order to control for the robustness of our findings, we re-estimate our models using quantile regression. 32 Although this approach delivers additional significance at some quantiles, overall results remain the same. Furthermore, we pool our data over all four plants. Yet, we fail to find overall statistical patterns. Eventually, we limit our data to units working in the assembly line since assembly line production is the only division that is observed in all four plants. Yet, this does not deliver further insights on potential determinants of worker absenteeism. Unfortunately, our study entails limitations at some instances. First, by focusing on aggregated data we gain evidence on team-level determinants of absence, yet at the expense of limited insights on individual-level influences. Second, although using data on absence spell duration and frequency to proxy for voluntary and involuntary absence, our study is limited in so far that both measures may contain their respective counterpart as well (Thomson, Griffiths, Davidson 2000). Third, we are not able to control for individual or group-level health conditions of employees other than by applying nation-wide WHO influenza outbreak statistics. In doing so, we are quite confident to at least proxy for employee health conditions concerning acute flu pandemics. However, we do not have an appropriate measure for other illnesses such as musculoskeletal disorders or stress-induced illnesses provoked by monotony, repetitiveness and cycle time. Yet, we assume physical as well as mental sicknesses to be distributed homogeneously among shop-floor teams after 32 Results are not reported in this paper. Employee Absenteeism: Determinants in the International Context 74 controlling for age and gender. Fourth, by working on monthly intervals we only provide evidence on determinants that are relatively stable over time (Harrison, Martocchino 1998). Conversely, this means that daily or other short-term effects on absenteeism are leveled out over the aggregated monthly periods and, therefore, could not be studied in this paper. 33 Fifth, we are not able to control for educational and cultural influences on absence behavior. This would have been particularly interesting since discussions with on-site experts have shown that absence cultures vary widely between cultural spheres. Additionally, we are not able to include more detailed workplace characteristics other than the implemented shift system to assure the anonymity of teams. Eventually, since this paper is built on insider data, findings might suffer from a lack of generalizability to other working conditions and organizations. 3.6 Conclusion and Implications Using a hitherto unavailable data set gathered from insider econometric data of four international manufacturing plants of a large European automobile company, we are able to provide evidence on determinants of absenteeism. In total, we investigate team-level absence of blue-collar employees organized in 160 work units (approx. 2,150 workers). Based on a comprehensive literature review we derive fourteen hypotheses covering four main areas of potential determinants on absence: social peer influences, economic influences / incentives, worker characteristics and working conditions. We focus our analysis on three unit-level dimensions of absence: absence rate (percentage of scheduled time lost due to absence), absence spell duration (average number of days lost per absence occasion) and absence frequency (number of absence incidents divided by unit size). Our empirical analyses follow a two-stage research strategy that complements insider econometric data analyses with interviews with on-site experts. Interestingly, we observe effects to vary by production site, however, with only very few consistent findings. Since a detailed description of results by plant has been given throughout chapter 3.5, we refrain from recapitulating effects by site in what follows, but instead summarize the overall effects. Considering social level determinants, we predicted turnover to decrease absence since group norms are less enforceable when team composition is changing. Yet, results are 33 Although empirical evidence is missing at this point, we were able to gain some anecdotal insights by means of the experts’ discussions. For instance, experts confirmed findings by Rosenblatt, Shapira- Lishchinsky and Shirom (2010) of increased absence around school holidays and weekends (“Monday morning flu”). At one plant, experts could also identify higher absence rates at match days of the local soccer team in the affected shift and, thus, support arguments by Thoursie (2004). Employee Absenteeism: Determinants in the International Context 75 pointing in the opposite direction, presumably due to a feeling of uncertainty and the lack of an overall team spirit in teams with high turnover. In accordance with our expectations, we observe absence to increase with unit size. Although results may suggest shirking, we follow on-site experts in arguing that line managers’ attendance management may be more caring and familiar in smaller teams. With regard to peer effects, the share of temporary team mates is positively related to voluntary absence of permanent workers, presumably due to shirking. Surprisingly, the share of workers with health impairments fails to reach statistical significance in all model specifications. Evidence on economic influences and incentives reveals the expected positive link between employment protection laws and absence. In other words, workers tend to increase their absence behavior when feeling protected by strict employee rights. Moreover, we hypothesized a negative relationship between absence and outside options expressed by national (un)employment figures. Surprisingly, we observe this relation to be inconclusive. Furthermore, we are able to confirm the expected positive link between prosperity level and absence behavior. This effect seems to be attributable to a certain feeling of job security since on-site experts observed the inverse relationship during the financial crises of the previous years. We were not able to properly study the effects of national sickness benefit systems on absence since the company at hand grants additional sick pay to compensate for weaknesses in statutory sickness benefits. As a consequence, differences in national legislation are leveled out. Regarding worker characteristics, results suggest that absenteeism increases with the share of female employees within a unit. According to on-site experts, this effect is to a large degree attributable to the double burden of work and family obligations which is even today substantially more pronounced for females than for males. For some sites, we were able to confirm the predicted positive relationship between age and absence rate and spell duration. Thus, it seems that meeting workplace burdens is substantially more demanding with age. Yet, against our expectations we observe weak evidence that voluntary absence, too, increases with age. Results for mean organizational tenure and acute health conditions expressed by national influenza outbreak statistics are inconclusive. Consistent with the hypothesized relationship, we find two-shift as well as three-shift systems to induce significantly higher absence than the standard one-shift system. According to on-site experts this effect is largely attributable to the non-correspondence of working hours and human bio-rhythm as well as social life. All findings are robust on various specifications and survive a large number of robustness checks. Employee Absenteeism: Determinants in the International Context 76 Subsequent to the data analyses, we conduct a total of 18 expert interviews at both international plants with shop floor staff, worker representatives, line supervisors and HR managers in order to double-check our own experience from the national plants. These interviews helped us to further interpret our results and to learn more about the instruments of attendance management in practice at other locations. Based on our empirical findings and the practical experience of on-site experts, we yield some important implications that may help to increase both overall health conditions and attendance of blue-collar workers. The results on unit size seem to suggest that small units are advantageous over large units in terms of employee attendance. As discussed, we predicted this phenomenon to arise from better shirking options in larger teams. However, we learned from experts that quality of and familiarity within the manager-employee relationship is moderated by team size. We, therefore, suggest managers to invest in developing trustful and sincere social ties with their subordinates based on a fair and open culture. Of course, building up intensive relations within teams requires time line managers usually do not have. Thus, we call for HR and plant management to support line managers by granting a sufficient amount of time to invest in socializing with their teams. We believe the long-term outcomes in terms of reduced absenteeism will leverage any initial time loss. With regard to economic influences it seems to be infeasible for a company to take actions since politics (e.g. sickness benefits, employment protection) as well as overall economic situation (e.g. unemployment, prosperity level) are usually out of the range of influence of a company. Yet, it might be a reasonable strategy to openly communicate the current economic situation of the company, in particular the competitiveness on the global markets, to the employees to increase their awareness of potential menaces. This could motivate employees who feel too secure. As mentioned above, both international plants compensate for weaknesses in statutory sick pay by offering additional sickness benefits. However, at one site employees with more than three non-work related absence occasions per 12 months are getting an official warning and are withheld additional sickness benefits on their next absence incident (repeated absence arising from known long-term sicknesses e.g. cancer is not taken into account). According to on-site experts, this policy has proven to be successful in reducing absence frequency. Overall, empirical evidence on environmental determinants is pointing in the direction that drawing international comparisons of absence figures within the company might be biased due to national variance in laws and economic situations. In other words, critical assessments of international locations are not necessarily Employee Absenteeism: Determinants in the International Context 77 accurate and true and, as a consequence, policy decisions based on international comparisons may be misleading. In response to empirical evidence confirming the predicted positive link between age and absence, we emphasize the importance of harnessing all possible health-promoting potentials in ergonomic workplace design and occupational health care (OHC). We acknowledge that at all four plants working conditions are constantly improving in terms of ergonomics. Additionally, it seems important to mention that medical attendance of musculoskeletal illnesses by OHC is very comprehensive and high in quality. The same is true for medical check-ups offered to employees for free in the company’s on-site OHC centers. However, we learned from local experts that work-related and non-work related mental stressors are of increasing importance even in manufacturing. Experts reveal that on-site OHC centers still lag in developing a strategy to treat stress-induced illnesses as comprehensively as musculoskeletal illnesses. Moreover, some findings emphasize the advantageousness of teams that are mixed with regard to age since younger worker can benefit from older ones and vice versa. It, therefore, seems reasonable to yield a heterogeneous team composition. Considering empirical evidence suggesting females to suffer from the job-family double burden more than men it seems important to support families in dealing with the reconciliation of work and family life. From on-site experts we learned that a different set of supporting activities is in practice at each site. Instruments that enhance individual flexibility have proven particularly helpful in increasing attendance of parents. We, therefore, suggest offering parents high flexibility given organizational constrains. For instance, at one location parents can work reduced hours even when working shifts. Here, two persons on reduced shifts share one job. In situations where both parents are employed on the shop floor, they are offered opposing shifts if necessary to guarantee child care around the clock. At some plants, employees can use dependency leave (or third-party leave) when children or elders are ill and need intensive care. Although third-party leave has proven to be very useful in allowing parents to care for their sick children without having to fake own illnesses, employees report that making use of this instrument is often criticized by supervisors and HR. Greater understanding seems to be necessary at this point. Eventually, our findings on shift systems prompt the thought that shift-work should be avoided as we observe two-shift and three-shift systems to increase absenteeism in com- Employee Absenteeism: Determinants in the International Context 78 parison to the normal one-shift (day) systems. Yet, from local experts we learn that it is sometimes preferable to add a second or third shift to better distribute workload instead of using overtime. If shift work is unavoidable, it seems important to follow the latest scientific insights on healthy shift work design. In general, the findings of this work depict an important contribution to (personnel) economics absence research. In particular, we benefit from exceptional and unique insider data that allow for comprehensive econometric analyses. However, in this study we only observe automobile workers. As a consequence, we might suffer from selection bias that entails endogeneity problems – yet a common issue in insider econometric studies. We, therefore, encourage other researchers to further study absenteeism in different organizational contexts in order to gain a comprehensive idea of the functioning of employee absenteeism and to refine ways to manage attendance. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 79 4 ON THE ROAD AGAIN: CROWDING-OUT EFFECTS OF EXTRINSIC MOTIVATION IN COMMERCIAL TRUCKING 4.1 Introduction Today, the use of variable compensation schemes in the form of exogenous financial incentives has become a common practice in organizations to align interests of principals and agents. 34 In this paper, we seek to study how extrinsic rewards may influence employee effort choice decisions and, thus, worker productivity. We analyze performance under varying incentive schemes based on unique data on commercial truck drivers. The trucking industry appears particularly suitable for this purpose since the carrier-driver relationship represents the classic principal-agent setting as modeled by economic theory (e.g. Oyer, Schaefer 2011) with agents producing their output far beyond the monitoring range of the principal. 35 Additionally, truck driver performance is entirely achieved on an individual level and, therefore, is not prone to any kind of team bias. 36 To solve this very specific principal-agent problem, asset ownership is often considered a particularly cheap and efficient monitoring instrument. Yet, its applicability is limited since truck drivers are assumed to be risk averse and limited in their access to capital (e.g. Arrunada, Gonzalez- Diaz, Fernandez 2004, Nickerson, Silverman 2003, Sheikh 2007). Recently, modern GPS- based fleet management systems advance driver monitoring from only observing the simple output (e.g. arrival time and goods carried) to more comprehensive real-time surveillance of driver performance while on the road (e.g. Baker, Hubbard 2004, Barla et al. 2010, Hubbard 2000). This allows for new approaches in incentive design for drivers. Trucking is, by and large, an increasingly competitive business environment for haulers as well as drivers. Nowadays, fuel costs account for around 30% of the total life-cycle costs of a heavy duty truck (e.g. Schittler 2003). In this situation, findings that only slightly increase operational efficiency may leverage profitability and, therefore, are appreciated by practitioners. At the same time, the road transportation industry is a substantial consumer 34 Exogenous material incentives as understood in this paper to comprise any wage configurations that at least partially involve variable remuneration such as bonus payments, piece rates and pay for performance schemes. Although not existent in the organizational setting at hand, there might be non-monetary extrinsic incentives as well (see, for instance, seminal papers on tournaments by Rosen (1986) and career concerns by Fama (1980)). 35 Vernon and Meier (2012) discuss the peculiarities of principal-agent settings in trucking in more detail. 36 Still, exogenous influences determine driving behavior to a large extent, e.g. weather, road conditions and traffic density. Research that does not include these confounding impacts most likely provides incorrect and fallacious conclusions. Thanks to our extensive data, we are able to control for various external influences. All control variables are discussed in the remainder of this paper. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 80 of fossil fuels. Within the EU, CO2 emissions of heavy duty vehicles – trucks, buses and coaches – have increased by 36% between 1990 and 2010 and now account for about 25% of overall CO2 emissions and 5% of all greenhouse gas emissions from road transport (European Commission 2014a). Thus, motivating commercial truck drivers to adapt fuel efficient driving provides a promising opportunity to contribute significantly to the reduction of both fuel consumption and CO2 emissions for economic and environmental benefits. This is where monetary incentives for drivers come into play. Academic theory knows two opposing effects of exogenous incentives on motivation. Classic personnel economics theory assumes exogenous incentives to increase employee effort choice and productivity by aligning the interests of principals and agents (e.g. Gibbons, Roberts 2013, Oyer, Schaefer 2011). There is broad empirical and experimental evidence on this issue in varying organizational settings (e.g. Kahn, Silva, Ziliak 2001, Lavy 2002, 2009, Lazear 2000b, Shearer 2004). Nevertheless, behavioral incentive theory assumes exogenous incentives to crowd out intrinsic or social motives that are found to be equally important in determining employee effort choice and productivity (e.g. Bénabou, Tirole 2003, 2006, Gneezy, Meier, Rey-Biel 2011). As a consequence, overall employee performance might be lower when being monetarily incentivized. Again, there is comprehensive support for these arguments in the literature (e.g. Ariely, Bracha, Meier 2009, Frey, Oberholzer-Gee 1997, Gneezy, Rustichini 2000a, 2000b). In response to the concurrent theoretical assumptions and the conflicting findings on the effects of extrinsic incentives, this study aims at analyzing employee effort choice decisions under varying incentive schemes given the specific organizational setting of the trucking industry. We use hitherto unavailable data on the performance of 37 commercial truck drivers. Data is gathered from the GPS-based fleet management system of an inhouse hauler of a large European truck manufacturer. The particular features of this research are twofold. First, in contrast to most studies on extrinsic incentives we do not focus on the introduction of a bonus pay scheme, but on its abolition. Second, our analyses are based on four performance measures recorded by in-vehicle computers of the hauler’s fleet management system that collect real-time information on truck positioning, technical parameters and driving behavior. This data allows us to contribute to the ongoing discussion on the benefits and pitfalls of extrinsic incentives. If classic personnel economics theory holds true, we expect drivers to perform worse after the incentives are abolished. In contrast, if behavioral incentive theory holds true, we expect driver performance to increase On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 81 although drivers are no longer extrinsically incentivized. The empirical evidence of this work reveals a significant increase in driver performance after incentives have been abolished, independent of the performance measure. Thus, our findings confirm assumptions stated by behavioral incentive theory as extrinsic incentives are shown to reduce employee performance. Following previous research (e.g. Frey 1994, Frey, Oberholzer-Gee 1997, Gneezy, Meier, Rey-Biel 2011), we attribute this phenomenon to the crowding-out effect of extrinsic incentives. Other important intrinsic motives – such as internal competition or drivers’ environmental believes – are literally bought off by monetary incentives. With this study, we contribute to the literature in several regards. First, we add to the current discussion on the advantages and drawbacks of extrinsic incentives on worker productivity by extending the empirical literature on employees’ effort choice decisions under different incentive schemes. Our findings confirm the crowding-out approach suggested by behavioral incentive theory. Second, while numerous studies analyze the switch from fixed pay to performance pay (e.g. Lazear 2000b, Shearer 2004), little evidence is provided on situations where incentives are completely abolished. One of the few exceptions is a study by Freeman and Kleiner (2005) who study the change from piece rate pay to hourly wages. As a result, they observe a significant productivity loss. We add to their research by offering a comprehensive analysis of employees’ behavioral responses to the abolition of extrinsic incentives. Third, to our knowledge we are the first to use truck driver performance data gathered by modern GPS-based fleet management systems to analyze employee behavior. This seems surprising since this data consists of broad objective performance information. We, therefore, hope to pave its way into personnel economics. Finally, our work contributes to the emerging insider econometrics approach introduced in seminal papers by Ichniowski, Shaw and Prennushi (1997) as well as Lazear (2000b). This approach applies elaborate econometric methods to (panel) data of only one organization. Thus, it delivers high-quality findings on a micro-perspective, yet at the expense of a limited generalizability. This paper proceeds as follows. The next paragraph discusses the two competing theories on the impact of extrinsic incentives on employee behavior. Section three describes the unique organizational setting of this study while section four explores the estimation strategy as well as the results of our econometric analyzes. Eventually, section five offers an interpretation and final concluding remarks. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 88 formance and productivity after the monetary incentives are abolished. However, if behavioral incentive theory proves true, we expect to find a reduced performance while bonus pay is in practice due to crowding-out. In other words, we should observe performance to increase after the incentives are abolished since drivers might draw their motivation from other intrinsic or social motives when no longer financially incentivized. 4.3 Data Set and Descriptive Statistics In this paper we seek to analyze behavioral responses of workers to shifts from a bonus pay scheme to a non-incentive environment. Therefore, we use a hitherto unavailable panel data set that includes trip-wise performance information of commercial long-haul truck drivers of an in-house hauler of a large European truck manufacturer. The data is based on information recorded by GPS-based in-vehicle computers of the hauler’s fleet management system. Performance data covers an extended period of thirty-six months (January 2011 to December 2013). With the beginning of the observation window, the hauler installed a monetary performance premium to reward good driver performance. However, after two years in practice, the hauler decided to disestablish the incentive scheme again. Within the organizational structure of the parent company, the hauler realizes just-in-time delivery of intermediate goods from one production plant of the truck manufacturer to another. While on the road, drivers travel roughly 1,250 kilometers across four Northern European countries. To comply with drivers’ mandatory resting periods 44 , the total distance is split up into two stages with a fixed location for the drivers to rest. At this location, the hauler operates a premise that provides sleeping, cooking and sporting facilities for drivers during their rest. Upon their arrival at the location, drivers pass on their trucks to another driver who continues the journey on the very same truck. In other words, driver A arrives at the resting location, where driver B has just finished his mandatory rest. Driver B takes over the truck and immediately drives on to the final destination, whereas driver A stays for his rest and awaits the arrival of the next truck to continue his journey. Thus, drivers travel a round-trip with four stages on four different trucks before returning home. This procedure secures just-in-time supply of goods without transferring the company’s ware- 44 Due to European road haulage regulations drivers are not allowed to exceed 9 hours of driving per day with a subsequent mandatory rest period of at least 11 hours. There are few exceptions enabling drivers to break these regulations in order to react to current traffic conditions or short-term workload peaks (European Commission 2014b). On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 89 house on to the streets. Obviously, driver-truck sorting in this setting is totally random since drivers are required to take the next truck available without having any choice. Our initial data set includes 81 commercial truck drivers resulting in n=11,439 driver-trip observations. 45 We focus exclusively on drivers on permanent contracts to avoid confounding effects attributable to incentives arising from temporary contract characteristics as observed in other studies (e.g. chapter five of this thesis, Bradley, Green, Leeves 2012). All trucks are reported to be deployed almost exclusively on long-haul assignments from one production facility to the other. To fully exclude short-haul assignments we restrict the data to trips with a length between 520 kilometers and 720 kilometers. We selected this range since the two production facilities are 570 km respectively 670 km away from the resting facility. 46 Moreover, we exclude those observations with a reported fuel consumption outside the range of minus two times and plus four times the standard deviation around the mean. Finally, we only include those drivers that we can observe at least on 80 trips during the observation period (i.e. approximately four to five months of regular driving) and that have experienced the incentive scheme for at least three months. These constraints reduce the initial dataset thirty-seven truck drivers and n=6,825 observations. 47 We merged the trip-wise performance information with demographic data of drivers (age, gender and tenure). It appears from the summary statistics presented in Table 4.1 that the sample is dominated by male drivers, a fact that comes as no surprise in the trucking industry. Still, three female drivers are included in the data accounting for 580 observations (8.5% of total observations). Reported driver age covers almost the entire working life span from 22 to 64 years (mean=44.08, s.d.=11.07). This variance allows for a thorough analysis of potential age effects. Tenure as a proxy for driving experience is included on a monthly level given the exact date a driver joined the company (mean=25.59 months, s.d.=12.02). Since we exclude drivers from the sample that are employed on a temporary contract, employment status is similar for all drivers. Moreover, we could neglect any cul- 45 The panel data set is unbalanced due to driver and vehicle turnover during the observation period. For instance, at this company trucks are replaced after a total mileage of 1m kilometers. 46 We extended both distances by +/- 50km to account for drivers who exceed the maximum allowed driving time before reaching their destination or need to take alternative routes due to construction work or traffic congestion. 47 However, not all models could be estimated with the full data set since not all vehicles are equipped with the latest version of the relevant on-board communication units. Hence, incomplete information produces missing values at some instances. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 90 tural differences due to the fact that all drivers originate from the country domestic to the hauler. Table 4.1: Summary Statistics - Driver Demographics Variable mean s.d. min. max. Driver Age (in years) 44.08 11.07 22 64 Driver Gender (male=1) .915 - 0 1 Driver Tenure (in month) 25.59 12.02 1 64 Bonus (Bonus=1) .58 - 0 1 Number of drivers: 37 Number of driver-trip observations: 6,825 The GPS-based in-vehicle computers provide the hauler with extensive real-time information offering considerable opportunities to manage and monitor fleet activities including truck positioning, fuel consumption, driving behavior, truck handling and vehicle status. Table 4.2 presents the fuel consumption in liters per 100 km (mean=27.89 l, s.d.=3.02 l) as well as a shortlist of the most important driving variables (e.g. brake applications, speeding, average speed, maximum speed). Table 4.2: Summary Statistics - Driving Style Variable mean s.d. min. max. Fuel Consumption (in l/100 km) 27.89 3.02 21.2 39 Average Speed (in km/h) 75.32 3.34 42.1 84.7 Idling (in % of engine running time) .023 .02 .002 .238 Coasting (in % of engine running time) .14 .05 .007 .399 Speeding (in % of engine running time) .079 .11 0 .892 Brake Applications (# per 100km) 12.62 7.83 .2 149.1 Harsh Brake Applications (# per 100km) .11 .25 0 10.1 Harsh Accelerations (# per 100km) .07 .27 0 4.3 Maximum Vehicle Speed (in km/h) 96.23 5.55 84 114 Number of drivers: 37 Number of driver-trip observations: 6,825 On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 91 The descriptive data already reveal interesting insights. First, idling – which is often labeled to be one of the most important determinants of fuel combustion (e.g. Lutsey et al. 2004) – is not of major relevance in the organizational setting at hand (mean=2.3% of engine running time). This might be due to the fact that drivers spend most of their time on highways and rarely face stop-and-go traffic in cities. Second, drivers are required by national legislation of the transit countries to drive 80 km/h at max. However, maximum values for average speed (max=84.7 km/h) and maximum vehicle speed (max=114 km/h) suggest that drivers override national speed limits occasionally. This might indicate that at times drivers need to catch up for traffic-induced delays to secure just-in-time supply. Third, the simple min-max comparison of several driving parameters reveals a high heterogeneity (e.g. brake applications vary from .2 to 149.1 per 100 km). This indicates that driving style might to a large degree be influenced by exogenous effects such as traffic density or weather conditions. To address these issues we add external information to the data that enables us to control for exogenous conditions. All controls will be discussed in detail in the next section. Based on the information recorded by its in-vehicle devices, the fleet management software is programmed to compute scores that assess driver performance in relation to target values predefined by the hauler. These scores cover drivers’ anticipation behavior, choice of gear, use of brakes and hill driving behavior. 48 In addition, the software uses these scores as well as further driving parameters to assess overall driver performance for each trip as being either “good”, “mediocre” or “poor”. Expressed by a green-yellow-red traffic light visualization, the trip evaluation is reported back to the hauler via the fleet management system. For transparency reasons, all results are accessible for all drivers via an online tool. Although fuel consumption is not directly evaluated by the software, a “good” driving behavior is congruent with an eco-friendly driving style. It appears from Table 4.3 that driving scores on average display high means varying from 63% to 81%. Yet, we question the choice of gear score since a mean of 97% appears to arise from the usage of automatic transmission vehicles in the majority of the fleet. The limited number of observations on this score fits this picture. Concerning the traffic light evaluation scores, exactly half of all trips are evaluated as being “good” (50%), while slightly fewer trips are reported as being “poor” (45%). Only a small number is rated as being “mediocre” (5%). 48 The underlying algorithms that assess driver performance and generate scores are corporate secrets and were not revealed to us. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 92 Table 4.3: Summary Statistics - Traffic Light Performance Evaluation Tool Variable mean s.d. min. max. Observations* Driving Scores Anticipation (Score in %/100) .81 .16 0 1 5,251 Choice Gear (Score in %/100) .97 .14 0 1 3,673 Use Brakes (Score in %/100) .77 .19 .1 1 5,064 Hill Drive (Score in %/100) .63 .25 0 1 5,028 Trip Evaluation Overall (1=green, 2=yellow, 3=red) 1.94 .97 1 3 6,801 Good (yes=1) .50 - 0 1 - Mediocre (yes=1) .05 - 0 1 - Poor (yes=1) .45 - 0 1 - Number of drivers: 37 *Varying number of observations since not all on-board computers report complete data at all instances. For an overview of mean, standard deviation, as well as within and between variance of all dependent variables listed separately for the period with incentives in practice and the period after incentives have been abolished, see Table A.6 in the appendix. Our study design has numerous advantages and contributes to personnel economics in several regards. First, the organizational setting allows studying the effects of the abolition of incentives. As mentioned before, the hauler installed a performance bonus as of January 1st 2011. Based on the evaluation scores reported by the fleet management software, drivers were rewarded financially for good performances. We like to point out again that low fuel consumption is only indirectly rewarded. The maximum amount a driver could earn on incentives did not exceed 5% of his monthly fix pay. Two years later, the bonus system was abolished by the hauler at December 31st 2012. Interestingly, the hauler installed the performance bonus without telling the drivers for the first three months. Despite the delayed disclosure, we doubt that the introduction of the incentive scheme remained completely unnoticed by drivers without any rumors spreading or premature information leaks. This would bias any findings. Thus, we focus our main estimations exclusively on the abolition of the bonus at the end of December 2012. This seems justified since there is broad evidence on productivity effects following the installation of incentives (e.g. Lazear 2000b, Shearer 2004), but to our knowledge we are among the first researchers who work on the On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 93 abolition of an incentive scheme in practice, with Freeman and Kleiner (2005) being a rare exception. Our set up, therefore, might offer new insights on the functioning of incentives that have not been covered by academic research so far. Second, comparisons among truck drivers are often hampered due to inconsistent exogenous influences coming along with different assignments, destinations and routes. Since the round-trip in the organizational setting at hand is identical for all drivers and constant over time, we can assume each driver to be exposed to comparable exogenous influences while on the road (e.g. road condition, constructions, detours etc.). This allows for comprehensive analyses of driver performance. Third, driver-truck sorting is totally random in the organizational setting of this study since drivers cannot influence which truck is the next available at either stage of their round-trip. Thus, we are able to control for good or poor driver-truck matches. Fourth, we can neglect any selection bias of agents sorting into specific incentive schemes since most drivers were already employed prior to the installation of the incentive scheme. Fifth, detailed date and time information for each trip allows us to control for exogenous effects (e.g. weather conditions) and temporary traffic density due to holiday traffic or daily rush hours. Both weather and traffic density can be assumed to highly determine truck handling and driver performance. Finally, vehicle load factors need to be similar over time to avoid biases arising from unequal vehicle maneuverability. Data by the European Commission (2014c) suggest that almost one quarter of all vehicle-kilometers of trucks within the EU is run empty. 49 However, the hauler reports average load factors above 90%, thus we can assume similar cargo weights and equal maneuverability of trucks for all trips. This allows us to neglect this issue for our data. 4.4 Empirical Models and Estimation Results Our data offers broad opportunities to measure truck driver performance. In total, our estimations are based on four dependent variables that proxy for drivers’ effort choice and productivity. First, we use the average fuel consumption in liters per 100 kilometers. Although low fuel consumption is not directly incentivized, all target values set with regard to driving parameters and performance evaluations are based on eco-friendly driving behavior and result in low fuel use. Since drivers receive individual fuel efficiency training on a non-regular basis, we can assume them to be well informed on how to meet company 49 According to the European Commission (2014c) this is mainly attributable to strict cross-border transit regulations. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 94 standards. 50 Drivers deviating from these standards, therefore, may be assumed to have chosen low effort levels. Thus, we are convinced that fuel consumption used as dependent variable may serve as a good proxy for drivers’ effort choice decisions. In our sample, average fuel consumption is 27.89 l/100 km with a standard deviation of 3.02 liters (see Table 4.2). Second, we deploy the most important driving parameters such as speeding, brake applications and average speed (see Table 4.2). Due to fuel efficiency training, drivers know how to meet company standards. Differing values, therefore, may again be an appropriate proxy for driver effort choice. Third, we use the four main driving scores (anticipation, use of brakes, choice of gear, hill drive) recorded by the in-vehicle computers (Table 4.3). These scores directly influence trip evaluation and are highly relevant in determining the amount of bonus paid. Full online excess to the fleet management data base should raise drivers’ awareness on how they performed on each score. Hence, we assume drivers to know their individual strengths and weaknesses in driving behavior quite accurately. Otherwise put, drivers are aware of how they can increase individual bonus pay by improving on those scores that do not yet meet company standards. In our estimations, we model low driving scores to proxy for low effort choices of drivers. Fourth, we deploy internal traffic light evaluation results based on software calculations (see Table 4.3). Again, drivers are offered full online excess to the evaluation reports that rate trips as either “good”, “mediocre” or “poor”. In other words, drivers know about their individual performance and should be aware of any misconduct. 51 Since they have both the ability and the knowledge to drive according to company standards, we believe trip evaluations other than “good” to result from low effort choice decisions. All four performance measures are insensitive to any subjective rater bias since they are recorded and assessed by in-vehicle computers automatically. While fuel consumption, driving parameters and evaluation scores are categorical variables, trip evaluation is of ordinal nature. The estimated models have the following general form (with FUELC being the fuel consumption, DRIVPAR representing the driving parameters, SCORE representing the evaluation scores and EVAL being the traffic light evaluation reports): 50 According to van Mierlo et al (2004) (car) drivers know by intuition how to drive eco-friendly. Thus, it does not really matter that we have no information neither on the frequency nor the exact dates of individual fuel efficiency training since we believe any training effect to arise from recollection instead of new insights. 51 According to Huang et al. (2005) and Roetting et al. (2003), truck drivers appreciate feedback from technology. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 95 FUELC DRIVPAR SCORE EVAL = α + βDRIV + γVEH + δTRAF + φWEA + ωCONTR + ε (1) (2) (3) (4) where α is the constant, DRIV describes driver performance as expressed by equation (5), VEH controls for vehicle effects, TRAF describes a vector of external traffic conditions, WEA describes a vector of weather conditions, CONTR is a vector combining further controls, β, γ, δ, φ and ω are the estimated coefficients and ε is the error term. The independent variables are discussed in more detail below. Driver performance (DRIV) is modeled as a function of the following form: DRIV = f ((AGE + HUMC) * EFCH) (5) with EFCH = Intrinsic Incentives + Extrinsic Incentives (6) where AGE measures driver age and HUMC proxies driver human capital as driving experience with the hauler (monthly tenure). Age and experience are important personal characteristics in determining driver performance. On the one hand, age has been proven to be detrimental to important abilities relevant to driving. For example, older drivers display an extended response time as well as lower visual and psychomotor abilities (e.g. Llaneras et al. 1998). On the other hand, experience is found to compensate these age-related impairments (e.g. Brock, Llaneras, Swezey 1996, Guest, Boggess, Duke 2014). The importance of experience is even more pronounced when taking accident risk into account since younger and less experienced drivers have a higher risk to be involved in accidents (e.g. Häkkänen, Summala 2001, McCall, Horwitz 2005, Rodriguez, Targa, Belzer 2006). Thus, any analysis without including age and experience into the estimations would produce biased results. In lack of information on skills and qualification, we interpret age and experience as the overall performance potential of employees. However, the performance potential is at any moment subject to a drivers’ individual effort choice decision (EFCH). Performance potential and effort choice eventually translate into actual driver performance. We model EFCH to be the joint effect of both intrinsic and extrinsic performance incentive (see equation (6)). While in the organizational setting at hand intrinsic incentives might include motives such as social preferences, altruism or eco-friendliness, extrinsic incentives comprise the monetary performance pay incentives. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 96 We use further variables to control for exogenous effects that might influence truck handling and driving behavior as modeled in equations (1) to (4). First, we use a vehicle dummy variable VEH to control for different truck configurations. A different set of technological equipment might facilitate truck handling at some instances or influence fuel consumption solely for technological reasons. Second, we use a vector of date and time variables TRAF that proxy for traffic density while on the road. Concerning the time of the day, we assume rush hours in the morning (7 to 9 a.m.) and evening (4 to 6 p.m.) to impose high traffic density, while drivers being on the road at night might benefit from open highways. Dummy variables control for rush hours and trips at night alike (trips from 8 p.m. to 7 a.m.). Additionally, we control for day of the week effects since we expect e.g. Mondays and Fridays to be prone to congested roads due to weekend commuters. Furthermore, we include a dummy variable covering a three-day window around international holidays since we believe roads to be congested due to holiday traffic at these dates. 52 Third, a vector of variables, WEA, controls for weather conditions during each trip. This seem necessary since previous research found weather to highly affect driving behavior, general traffic conditions and crash risks (e.g. Kilpeläinen, Summala 2007, Norrman, Eriksson, Lindqvist 2000). We, therefore, add control variables for temperature, precipitation, snow depth and wind speed. 53 Eventually, we apply dummy variables for all thirty-six months of the observation period to account for possible trend or timing effects. Likewise, we use dummy variables for all fifty-two weeks of a given year to account for seasonality. Taken together, this broad range of control variables excludes the major sources of potential exogenous influences on driver performance. Thus, we are confident that any performance effect that we will detect in our estimations is entirely attributable to drivers’ effort choice decisions as a response to the incentive scheme. In the following, we discuss results on fuel consumption as modeled by equation (1) first, then report our findings on driving parameters as modeled by equation (2). Afterwards, we present results on driving scores as modeled by equation (3) and eventually discuss the estimations on trip evaluation as modeled by equation (4). The key estimations are reported in Tables 4.4 to 4.7, respectively. Overall, our results support the crowding-out effect of extrinsic incentives since truck 52 Only those holidays that are mandatory in all four countries drivers pass through during their round trips are considered (e.g. New Year’s Day, Easter, Christmas, etc.). 53 Since we do not have detailed truck positioning data, we gathered weather information at a meteorological station approximately halfway the total distance. This should level differing climate conditions at the starting and arrival points. Data was provided by the German Meteorological Service (DWD 2014), a German state institution. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 97 drivers’ effort (and as a consequence their performance and productivity) is significantly lower with monetary performance incentives in practice. First visual evidence on divergent levels of fuel consumption with and without incentives in practice is presented in Figure A.4.1 in the appendix. Comparing the kernel density functions for both incentive environments, it appears that fuel consumption shifts to the left after incentives have been abolished. This indicates lower fuel use without incentives in practice. As a first empirical test for significant behavioral responses, we run a two-way ANOVA on the thirty-seven drivers from April 2011 to December 2013 to examine the effect of incentives on fuel consumption as a proxy for driver performance and productivity (see Table A. 7 in the appendix). We focus on a time frame starting in April 2011 since the first three months of the observation window might be biased due to the secret introduction of the performance pay. 54 We found a significant effect of the bonus pay on driver’s fuel consumption indicating that significantly different quantities of fuel are used depending on the existence and non-existence of monetary incentives, F(1, 6,252)=4.31, p=.0379. Since the ANOVA does not report any significant differences in the simple comparison of drivers (p=0.5263) or the driver-bonus interaction (p=0.4610), we are confident that any performance effect estimated in this study may be explained by behavioral responses entirely attributable to the (non-)existence of the incentive scheme. Moreover, this finding supports our argument that drivers adapt their individual effort choice to the incentive environment. To follow these indications, we apply elaborate empirical estimations to our data. We start with fixed-effect specifications since we are particularly interested in within-driver behavioral responses to the disestablishment of the incentive scheme. Moreover, fixed-effects models account for present driver fixed-effects that are not reported in the data. These effects can be either time-invariant (e.g. talent, prior driving experience, etc.) or quasi time-invariant without any or only slight modifications over time (e.g. driving style, environmental friendliness, professional ethos, etc.). To account for these unobserved effects that are seldom uncorrelated with explanatory variables, fixed-effect specifications are preferable (e.g. Wooldridge 2013). Since the incentive scheme was introduced without telling the drivers for the first three months we restrict our estimations to a period from April 2011 to December 2013 to avoid any bias arising from premature information leaks. 54 The secret installation might not have been as secret as intended by the hauler since we assume rumors to have circulated among the drivers prior to the introduction of the new remuneration scheme. On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 104 statistically significant influence on driver performance in any model specification. Yet, tenure exhibits a minor but significant impact on evaluation parameters with anticipation scores reducing by .4 percentage points and use of brakes score reducing by .5 percentage points with each month a driver is employed at the hauler. This result opposes theoretical considerations on human capital as an important determinant of employee performance (e.g. Rodriguez, Targa, Belzer 2006). We believe decreasing attentiveness as a result of routine to be one possible explanation since truck driving is a very monotonous occupation – especially in settings with constant routes and on highways. However, the limited observation window of only 33 months as well as the overall limited tenure (max of 64 months) might bias our findings at this point. Vehicle dummies only seldom reach statistical significance. Coefficients for weather condition as well as date and time dummies are inconsistent for all four variables. Therefore, we refrain from reporting these results in detail. In a final step, we estimate models on the traffic light trip evaluation assessed by in-vehicle computers (variable EVAL in equation (1)). We use trip evaluation as a proxy for driver effort choice decisions (EFCH) and, thus, are confident to detect potential performance effects of incentive. Following Gujarati and Porter (2009) as well as Wooldridge (2013) we estimate ordered logit and ordered probit regression models to address the ordinal nature of trip evaluation data. However, since probit regression only constitutes a minor modification of logit regression and yields barely different findings (Greene 2003), we focus on logit regression results in this paper due to reasons of brevity. Table A.10 in the appendix displays results for the ordered logit estimations with trip evaluation being the dependent variable. It becomes obvious that evaluations are significantly worse when performance incentives are in practice. Post-estimation probabilities of trip evaluation with and without incentives in practices are given in Table 4.7. Table 4.7: Estimated Probabilities for Trip Evaluation in both Incentive Environments Variables Incentives in Practice Incentives Abolished Change 95% CI for Change Trip Evaluation Good .4034 .6894 .2860 -.3421 -.2298 Mediocre .0619 .0513 -.0106 .0072 .0139 Poor .5347 .2594 -.2754 .2219 .3289 On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 105 With incentives in practice the probability of a trip being assessed as “good” by in-vehicle computers is 40.34% compared to 68.94% after the abolition of incentives. This is a difference of 28.6 percentage points or 70.89%, respectively. At the same time, the probability of a trip being evaluated as “poor” dropped by 27.54 percentage points or 51.5% from 53.47% to 25.94%. Probabilities for “mediocre” trip evaluations vary only slightly by 1.06 percentage points or 17.2% after the abolition of the incentives. In other words, with performance incentives in practice the probability of a good truck driver performance is markedly lower compared to a situation without any extrinsic incentives. This indicates drivers to choose lower levels of effort when monetarily incentivized. Similar to the other results of this paper, this finding supports the crowding-out assumption of extrinsic incentives backfiring on the very behavior they intend to prevent. All estimations represented in Tables 4.5 to 4.7 are based on the full model specification. To control for biases arising from control variables we re-estimated all models under different specifications. Results proved to be robust for all estimations with little variation in coefficients. The same holds true when excluding the choice of gear score from the SUR models. Findings for the other dependent variables remain robust. We further check for the robustness of the results in two ways. First, we re-estimate our models excluding the three women from the data set. Second, we include all temporary drivers in the estimations. Although we expected an incentive effect deriving from the contract characteristics (chapter five of this thesis, Bradley, Green, Leeves 2012), the baseline results vary only marginally in coefficients. Thus, we are very confident that our findings are robust. Still, our study entails some limitations as a consequence of its organizational setting. First, we do not have detailed information neither on the algorithm that determines performance evaluations nor on the underlying mechanisms that determine the total amount of incentives. It would have been interesting to work on the incentive scheme in more detail, e.g. by observing only those drivers at the verge to a higher incentive level. Second, the incentives were introduced for the first three month without informing the drivers. Still, there is the possibility that the information might be spread premature. Since we expect the results for this period to be biased due to information leaks, we exclude the respective three months window from our baseline estimations. Yet, checking for the introduction of incentives after three months supports our overall findings (see Table A.8 in the appendix). Third, the bonus payments incentivize low fuel consumption only indirectly by rewarding an eco-friendly driving style. Therefore, one might question the validity of our findings on On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 106 the link between incentive scheme and fuel consumption. However, all target values of the driving parameters and evaluation scores set by the company aim at reaching eco-friendly driving behavior that yields at low fuel-consumption. Moreover, our findings on fuel consumption are entirely supported by estimations for raw performance indicators such as driving parameters and evaluation scores. Thus, we are confident to address doubts on the indirect relationship of incentives and fuel consumption. Finally, our data set includes only three women. Although there is support for a gender effect in self-selection and performance under incentives (e.g. Croson, Gneezy 2009, Dohmen, Falk 2011), we avoid working on this issue to account for the unequal gender distribution among drivers. Yet, the low share of women is industry specific and, therefore, should not bias or impact our overall findings. 4.5 Conclusion With this study we offer new insights to the ongoing debate on the benefits and drawbacks of extrinsic incentives. While classic economic theory assumes employees to react to extrinsic (monetary) incentives by increasing performance and productivity, behavioral incentive theory suggests extrinsic rewards to crowd out other important intrinsic and social incentives that motivate people. There is broad experimental and empirical support for both theoretical assumptions. To contribute to this discussion, we use a hitherto unavailable data set on the performance of commercial truck drivers collected from the internal fleet management system of the inhouse hauler of a large European truck manufacturer. Truck drivers receive bonus payments for good performance within the first twenty-four months of the thirty-six months observation period. Subsequently, management decided to abolish the incentives. We empirically measure the effect of the abolition of the incentive scheme on driver effort choice decisions by analyzing four separate sets of dependent variables of driver performance: fuel consumption, driving parameters, driving scores and trip evaluation. Consistent with behavioral incentive theory, we find comprehensive support for crowding-out effects of the bonus pay scheme applied by the hauler. In other words, we observe drivers to display significantly lower effort and performance when extrinsically incentivized compared to the non-incentive environment. These findings are in line with previous research (e.g. Deci 1971, Gneezy and Rustichini 2000a). More precisely, fuel consumption is significantly higher when incentives are in practice, adding up to additional 530 liters used per driver On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 107 and year which translates to extra fuel costs of €800 per driver and year. We interpret this amount as hidden costs of incentives. Combined with the net bonus payments, the hidden costs constitute a serious financial burden to the company’s baseline results – especially in a business as competitive as the trucking industry. When narrowing the observation period around the lagged information on the introduction and the final abolition of the bonus pay, we observe strong initial effects for both events. Further analyses of driving parameters suggest that drivers put less effort in adapting an eco-friendly driving style. This results in significantly lower values on driving scores: nine percentage points worse on hill driving, thirteen percentage points worse on anticipation behavior and even sixteen percentage points worse on use of brake. Concerning trip evaluation data, we observe a more than 70% higher probability of a trip being rated as “good” subsequent to the abolition of incentives. In short, our results indicate that drivers respond to incentives not the way intended by the hauler. Instead, it is the abolition of the incentive system that significantly increases driver performance. These findings remain robust on all four performance measures under various specifications and robustness checks. The results clearly support crowding-out effects of extrinsic rewards as stated by behavioral incentive theory. In this sense, the empirical evidence is in clear contrast to classical economic arguments on the benefits of monetary incentives (e.g. Gibbons 1998, Oyer, Schaefer 2011). We believe our findings to arise from monetary incentives crowding out other important intrinsic and social motives that seem to play a major role in truck driver motivation. In other words, drivers are for some reasons intrinsically motivated to choose a certain level of effort. Though, an extrinsic incentive – such as a financial reward – seems to diminish or even disperse their intrinsic motivation. As a consequence, overall net effort choice and performance may be lower than without receiving financial rewards. There are some intrinsic and social motives that might serve as interpretation at this point. First, research identified individual environmental awareness and concerns to be essential for adapting pro-environmental behavior (e.g. Eagly, Kulesa 1997, Fransson, Gärling 1999). Therefore, we assume drivers’ environmental beliefs to be a major intrinsic motive to adapt ecofriendly driving. 57 In line with this argument, extrinsic incentives are observed to crowd out pro-environmental behavior (e.g. Thøgersen 1994). Second, prior research has identified social image, status and reputation as powerful intrinsic incentives to people’s motiva- 57 For an overview on pro-environmental behavior of truck drivers see Schweitzer, Brodrick and Spivey (2008). On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 108 tion (e.g. Ariely, Bracha, Meier 2009, Ellingsen, Johannesson 2008). Hence, we assume drivers to have some social concerns on having an eco-friendly image and reputation, in particular as they are working in a high-polluting industry. This might be even strengthened as the country domestic to the drivers is far below OECD average on CO2 emissions (OECD 2014b). Third, peer competition among drivers may represent an intrinsic incentive, too. According to company staff, the first thing drivers do when returning from a round-trip is to check their traffic light evaluation relative to peer-group performance. Theory and empirical evidence suggest that peer-group competition functions as social incentive. For instance, effort choice is higher when relative performance rankings are reported (e.g. Charness, Masclet, Villeval 2013). We believe that monetary incentives – at least partially – bought off the game-like character of peer competition. The same holds true for classic peer effects that motivate drivers to good performance via social pressure (e.g. Falk, Ichino 2006, Mas, Moretti 2009). Eventually, introducing incentives in the first place might be interpreted by drivers as an act of distrust. According to Sliwka (2007), feeling distrust might turn a person into selfish behavior. This effect could even be intensified by the secret introduction of the incentive scheme that most likely was seen as a lack of trust by drivers. Acting selfish might reduce truck drivers’ concerns about both the environment and the competitiveness of their employer. Instead, they might increase their own utility by putting less effort in economic driving behavior. Furthermore, our findings support prior evidence on people responding strongly to both well-structured and badly designed incentives (e.g. Roberts 2010). In particular, incentives that reward performance with an amount too small may counteract the intended behavior (e.g. Gneezy, Rustichini 2000a, 2000b). By offering less than 5% of the monthly net income, the hauler might have badly designed the incentive by choosing an amount below the incentivizing threshold. This especially holds true given that drivers are working in a high-paying company in a country with a high wage level and, therefore, may anecdotally be ranked among the topearning truck drivers in Europe. However, translating our findings into clear implications to practitioners proves to be difficult. It seems that principals need to abandon the idea that extrinsic incentives serve as a panacea to motivational issues. Instead, monetary incentives do not work as intended per se and employee performance and productivity depend on other intrinsic and social motives, too. Introducing extrinsic incentives without seriously reflecting on these issues might prove detrimental to overall productivity in the end. Thus, we highly recommend On the Road Again: Crowding-Out Effects of Extrinsic Motivation in Commercial Trucking 109 considering the internal and external fit of incentive schemes and general HRM instruments. At this point, we follow Ichniowski, Shaw and Prennushi (1997) who argue that incentive plans work best when introduced with complementary HRM practices. Therefore, we like to encourage practitioners to carefully monitor incentive systems and to take appropriate actions if incentives appear to have counter-productive effects. Nevertheless, it seems to be good news that mistakes once made in incentive introduction or design seem reversible since employees quickly respond to new settings. Eventually, the very particular organizational set up and the insider econometric approach of this paper might raise doubts on the generalizability of our findings. However, given that our results support theoretical assumptions stated by behavioral incentive theory and are in line with previous research, we are confident that our evidence is applicable to similar situations at least to some extent. Yet, we are not able to provide clear evidence on the generalizability of our findings to other organizational settings. We therefore encourage researchers to further contribute to the ongoing debate on both crowding-out and price effect of extrinsic incentives. Studies examining different organizational settings apart from insider econometrics might widen general insights on incentive effects on employees’ effort choice decisions. In response to the so far inconsistent findings on extrinsic incentive effects on employee effort, research on moderating and mediating effects of incentives and the fit between intrinsic and extrinsic incentives seems to be highly appreciated. Behavioral Consequences of the Transition from Temporary to Permanent Employment 110 5 BEHAVIORAL CONSEQUENCES OF THE TRANSITION FROM TEMPORARY TO PERMANENT EMPLOYMENT 5.1 Introduction Temporary employment contracts have become a common practice in modern organizations to circumvent labor market regulations considered as restrictive and to flexibly react to changing staff requirements that are due to changes in product demand (e.g. Houseman 2001). Moreover, temporary employment offers firms opportunities to reduce labor costs as well as administrative complexity (De Cuyper et al. 2008). According to the most recent Eurostat Labor Force Survey, about 14% of all employment contracts in the EU-28 in 2013 were temporary in nature 58 (Eurostat 2014c). Among those on temporary contracts, younger workers are heavily overrepresented (OECD 2014c). In this paper we define temporary contracts to be different from standard contracts with respect to the following three dimensions: limited duration, various statutory drawbacks (no or little employment protection, no minimum wage etc.) and the different organizational setting (temporary contracts are often with temporary-work agencies (e.g. De Cuyper et al. 2008)). 59 Temporary employment is not only used to avoid labor shortages and reduce labor costs but is often considered a “stepping stone” (Booth, Francesconi, Frank 2002) or a “port of entry” into permanent employment (Berton, Devincienti, Pacelli 2011, Buddlemeyer, Wooden 2011). Recent research emphasizes the nature of temporary employment as a screening device allowing employers to test the quality of a particular job match without the risk of long-term contractual obligations (e.g. Güell, Petrongolo 2001, Gagliarducci 2005). Moreover, a growing body of literature demonstrates that contract characteristics significantly affect employees’ behavior (e.g. Ichino, Riphahn 2005, Guadalupe 2003) in the sense that e.g. temporary and permanent workers are found to differ significantly in their choice of effort levels (e.g. Engellandt, Riphahn 2005). Although permanently employed workers are supposed to perform better than temporary workers because they receive more (firm-specific) training and report higher work satisfaction, empirical studies found the opposite to be true by demonstrating that temporary workers sometimes outper- 58 In the country where the headquarters of the hauling company is located, the percentage of temporary employees is slightly above EU-28 average (OECD 2014). 59 In this study, the term “temporary employment” refers to all kinds of contingent, fixed-term, casual or non-permanent employment relations with an employer characterized by an ex ante defined limited contract duration. Behavioral Consequences of the Transition from Temporary to Permanent Employment 111 form permanent workers (e.g. Livanos, Zangelidis 2013, Amuedo-Dorantes 2002). We believe the reasons to be twofold. On the one hand, employers very often use temporary employment as a screening device. On the other hand, employees consider motivation and effort as strong signals to an employer that are likely to increase their probability of being promoted to a permanent contract. Both effects incentivize temporary workers to outperform permanent workers in terms of effort levels. However, this holds true only as long as temporary work is associated with inferior job characteristics compared to permanent employment. So far, temporary positions have been found to be associated with less favorable working conditions (e.g. Paoli, Merllié 2001), lower wages (e.g. Mertens, Gash, McGinnity 2007), less training (e.g. Arulampalam, Bryan, Booth 2004) and lower levels of job satisfaction (e.g. Boyce et al. 2007). Moreover, the uncertainty about future job prospects can be a mental burden to temporary workers resulting in health problems (Sverke, Hellgren, Näswall 2002). Finally, temporary workers do not benefit from mandatory job protection legislation to the same extent as permanent workers. Given these disadvantages, temporary workers should have strong incentives to demonstrate motivation and effort to “qualify” for a promotion from a temporary to a permanent contract. These considerations inevitably lead to the question of whether and to what extent formerly temporary workers adjust their effort level after having signed a permanent contract. One might expect significant behavioral responses from employees who have been offered more favorable working conditions, including dismissal protection legislation. Using an unbalanced panel of truck drivers from the GPS-based fleet management system of an inhouse hauler of a large European truck manufacturer we are among the first to address this issue with individual productivity information from employees whose contract status was changed from temporary to permanent. Our results confirm a significant reduction in commercial truck drivers’ effort levels resulting in higher fuel consumption and a poorer overall driving performance after being promoted from temporary to permanent employment. Our study contributes to the literature in several regards. First, we add to the growing research on the influence of employment contract characteristics on employee behavior. In particular, we extend the literature on workers’ effort choices when contract characteristics become more employee-friendly (e.g. Ichino, Riphahn 2005). Second, we are among the first to use data from a modern GPS-based fleet management system to analyze driver behavior. This usually extensive source of data has not yet made its way into the personnel Behavioral Consequences of the Transition from Temporary to Permanent Employment 112 economics literature. Third, we follow the insider econometrics tradition, which has enjoyed increasing popularity among personnel economists since the publication of the seminal papers by Ichniowski, Shaw and Prennushi (1997) on HRM practices in U.S. steel plants and Lazear (2000b) on monetary incentives for windshield installers. The insider econometrics approach applies sophisticated econometric methods to a (panel) data set gathered in one or a few companies. The resulting nano-perspective allows detailed insights into the behavior of individuals upon changes in incentives, working conditions, etc. The remainder of this paper is structured as follows. The next section provides both a review of the literature on the incentive effects of temporary employment as well as on employees’ behavioral responses to changes in contract characteristics. The specific organizational setting and the data set are described in detail in section three. Section four presents the estimation strategy as well as the results while section five concludes. 5.2 Literature Review and Theoretical Framework It is a stylized fact in the personnel economics literature that temporary workers receive less training by employers due to the usually rather short duration of the employment relation (e.g. Arulampalam, Bryan, Booth 2004, Hoque, Kirkpatrick 2003, Forrier, Sels 2003, Arulampalam, Booth 1998). As a consequence, temporary workers accumulate less firmspecific knowledge than permanently employed workers. Since firm-specific skills are critical for performance (e.g. Hatch, Dyer 2004, Hitt et al. 2001) permanently employed workers are expected to outperform temporary workers. However, there is also evidence showing that despite their lower levels of training temporary workers often outperform employees with a permanent contract. The performance differential is likely due to temporary workers’ higher levels of motivation and work effort. Using the Swiss Labor Force Survey Engellandt and Riphahn (2005), for example, find that the probability of working unpaid overtime is 60% higher for temporary compared to permanent workers. Moreover, in a large data set from Australia Bradley, Green and Leeves (2007) find evidence of significantly lower levels of absenteeism (their preferred measure of employee effort) among temporary workers. This latter result is confirmed by Livanos and Zangelidis (2013) for the EU, Arai and Thoursie (2005) for Sweden and Amuedo-Dorantes (2002) for Spain. Moreover, using data from a Spanish survey Guadalupe (2003) finds that other things equal temporary workers have a five percentage points Behavioral Consequences of the Transition from Temporary to Permanent Employment 113 higher accident probability which may be due to lower investments in temporary workers’ human capital as well as their pronounced incentives to demonstrate high levels of motivation and effort in the sense of taking higher risks. The incentive effects of temporary contracts that are assumed to result in higher motivation and effort levels of temporary employees are due to their specific screening and signaling properties (Spence 1973). First, temporary contracts are often used as a screening device to test newly hired workers before they are being offered a permanent contract (e.g. Green, Leeves 2004). Second, temporary work may be considered a “stepping stone” into permanent employment (Booth et al. 2002). Consistent with the latter argument Güell and Petrongolo (2001) find that temporary workers often sign a permanent contract before their initial temporary contract had expired. This suggests that firms use temporary employment as a screening opportunity to discover workers’ “true” qualities and offer permanent contracts as soon as the required abilities have been identified (see also Buddlemeyer and Wooden (2011) with comparable evidence for Australia and Gagliarducci (2005) for Italy). Summarizing, these studies suggest that temporary contracts are used as a screening instrument that enables employers to reduce information asymmetries about workers’ motivations and effort choices without the risk of long-term contractual obligations. Summarizing, the characteristics of temporary contracts should provide strong incentives for temporary workers to choose high effort levels to produce those signals that maximize the probability of being hired on a permanent contract. This holds true, however, only as long as permanent employment proves to be advantageous over temporary employment. Indeed, temporary employment is usually associated with poorer working conditions and lower pay levels for equal performance. On the one hand, temporary workers have been found to benefit less from investments in health care or ergonomics leading to higher levels of job dissatisfaction, fatigue and muscular pain (Benavides, Benach 1999, Benavides et al. 2000). On the other hand, they are exposed more often to repetitive tasks and movements as well as heavy loads (Paoli, Merllié 2001), enjoy lower levels of work autonomy and report themselves the least well-informed about their work environment (Aronsson 1999). Furthermore, temporary workers often consider themselves subject to “status stigmatization” leading to lower levels of well-being, job satisfaction, and performance (Boyce et al. 2007). Generally speaking, temporary jobs are found to be of an overall lower job quality (Green, Kler, Leeves 2010). The wage penalty of temporary workers is well documented in the literature: Mertens, Gash, McGinnity (2007) for example find that the Behavioral Consequences of the Transition from Temporary to Permanent Employment 120 It is important to bear in mind that the traditionally used measure “fuel consumption per 100 km” is not included in the automatic evaluation of the individual trips. However, driving behavior that leads to a “good” evaluation is characterized by an eco-friendly driving style (no harsh accelerations, no speeding, etc.). It appears from Figure A.1 in the Appendix, that fuel consumption and trip evaluation are uncorrelated: Average fuel use for trips evaluated as “good” is 28.1 liters per 100 km, for trips evaluated as “mediocre” it is 28.2 liters and for trips evaluated as “poor” the respective value is 27.9 liters (F=0,49; not significant). Moreover, it appears from Table 5.2 that slightly more than half of all trips (54%) are assessed as “good” and 43% as “poor”. Only 3% of all trips are evaluated as “mediocre”. Thus, apart from the standard performance measure “fuel use per 100 km” we use in our estimations a second performance measure (“quality of trip as assessed by the company’s fleet management system”) that is uncorrelated with the first one. Our study design has a number of advantages avoiding major weaknesses of previous research in the economics of trucking. First, drivers can usually not be compared as they work under highly variable conditions, experiencing different exogenous effects of different routes and destinations (e.g. road conditions, constructions, detours, etc.). Due to the organizational setting of our study we are able to assume equal road and environmental conditions since all drivers use the same predefined route. Second, based on detailed information on date and time of each trip, we are able to control for weather as well as for date- (e.g. holiday) or time-of-the-day related (e.g. rush hours) peaks in traffic density since “timing” is likely to influence driving performance. Third, drivers are observed on different trucks without having any influence on truck selection. Thus, we can account for vehicle specific effects and eliminate any impact of driver-truck matches. Fourth, almost one quarter of all vehicle-kilometers of trucks in European cross-border traffic occurs on empty trucks (European Commission 2014c). Any study that is unable to control for load factors is thus likely to suffer from an omitted variable bias as full and empty trucks differ in maneuverability and fuel consumption. Since the hauler whose data we use here reports average load factors of more than 90% on its out as well as its return journeys, this issue can be neglected. Behavioral Consequences of the Transition from Temporary to Permanent Employment 121 5.4 Empirical Models and Estimation Results Recall that our empirical analyses are based on two different, yet somewhat related performance measures that can be considered proxies for choice of effort. First, we use in our estimations average fuel consumption per 100 km as the dependent variable. It appears from Table 5.2 that mean fuel consumption is 28.0 liters/100km with a standard deviation of 2.9 liters. Since all drivers receive individual fuel efficiency training in irregular intervals, we consider them well-informed on how driving behavior and style affects fuel consumption. 65 Thus, we suggest any differences in fuel consumption to derive from drivers’ individual effort choices. Second, we use the reports of the internal trip evaluation tool assessing each trip as either “good”, “mediocre” or “poor” (see Table 5.2). Since drivers are offered full online access to all evaluation reports they are aware not only of their performance relative to their peers, but also of any kind of “misconduct” in driving behavior. Interestingly, publicly available performance evaluations have recently been found to eliminate social loafing and increase performance at least in group work settings (Lount, Wilk 2014). Further research indicates feedback to be highly appreciated by truck drivers with feedback from advanced technology being less desired than feedback from supervisors (Huang et al. 2005). Thus, drivers receiving performance feedback and still not performing in accordance with the employer’s expectations are assumed to drive with reduced effort. While fuel consumption is a categorical variable, the trip evaluation variable is ordinal. Both outcome variables are computed by the in-vehicle computers and are not prone to any subjective rater bias or manipulation by the drivers. Our models have the following general form (with FUELC denoting fuel consumption per 100km and EVAL being the composite evaluation measure): FUELC = α + βDRIV + γVEH + δTRAF + φWEA + ωCONTR + ε (1) EVAL = α + βDRIV + γVEH + δTRAF + φWEA + ωCONTR + ε (2) where α is the constant, DRIV represents the driver performance (see equation (3) for details), VEH is a set of vehicle dummies, TRAF a vector of traffic conditions, WEA a vector of weather conditions and CONTR a vector of further controls, β, γ, δ, φ and ω are the es- 65 Although we have no information on the frequency or the exact dates of individual fuel efficiency training we assume this not to be a problem since recent research has convincingly demonstrated that (car) drivers seem to know by intuition how to drive eco-friendly (van Mierlo et al. 2004). Therefore, we assume that fuel efficiency training does not fundamentally change driver behavior, i.e. fuel consumption goes down after training but returns to its pre-training level rather quickly. Behavioral Consequences of the Transition from Temporary to Permanent Employment 122 timated coefficients and ε is the error term. All variables are discussed in more detail below. Driver performance (DRIV) is modeled as follows: DRIV = f ((AGE + HUMC) * EFCH) (3) where AGE is driver age and HUMC represents a drivers’ human capital in the form of past driving experience with the current employer (tenure in three-month intervals). 66 Previous research provides evidence for the negative effect of age on various abilities relevant to driving, e.g. response time, visual and psychomotor abilities (e.g. Llaneras et al. 1998). Other findings, however, suggest that older drivers can compensate for age-related impairments by experience (e.g. Guest, Boggess, Duke 2014). Additionally, young and less experienced drivers are found to have a higher probability of being involved in accidents (e.g. Häkkänen, Summala 2001), which further underlines the importance of experience in trucking. Hence, we include age as well as driving experience in our model to account for these effects. Both variables taken together represent the performance potential of an employee, which is then subject to drivers’ current effort choice (EFCH). Depending on the choice of effort, performance potential eventually translates into actual driver performance. In our estimations EFCH is represented by a dummy variable taking a value of 1 of a driver is on a temporary contract and 0 if he is on a permanent contract. In addition, we control in our estimations for exogenous effects that may influence truck driver performance. VEH is a series of vehicle dummies accounting for differences in technological configurations of trucks that might facilitate driving and/or decrease fuel consumption. TRAF is a vector of variables controlling for traffic density on the trips in two different ways. On the one hand, we include variables to control for the time of the day when the streets are presumably less congested (e.g. during night time, between 8 p.m. and 7 a.m. the next day) or more congested (e.g. during rush hours, between 7 a.m. and 9 a.m. in the morning as well as between 4 p.m. and 6 p.m. in the afternoon) when traffic density is particularly high due to commuter traffic. On the other hand, we include the day of the week as we believe some days to be systematically different (e.g. Mondays or Fridays) from others. Furthermore, we control for a three-day window around international 66 We are well aware that tenure on a monthly basis would be preferable for the first few months; however, we chose three-month intervals because during our observation period of 36 months any effect of monthly tenure is likely to disappear quickly as time elapses. Behavioral Consequences of the Transition from Temporary to Permanent Employment 123 holidays during which we suppose traffic density to be higher due to holiday traffic (public holidays in all countries the drivers pass on their trips, e.g. New Year’s Day, Easter, Christmas, etc.). WEA is a vector of weather dummies including wind speed, precipitation, snow depth, and temperature 67 because previous research has found driving behavior, traffic conditions as well as crash risks to be associated with weather conditions (e.g. Kilpeläinen, Summala 2007). Moreover, we include dummy variables for all thirty-six months to account for possible trend or timing effects. Finally, to control for season effects we include dummies for all fifty-two weeks of a particular year. Given the wide range of control variables, we are confident to have excluded most exogenous influences on driver performance and are thus able to study the “clean” effect of drivers’ behavioral responses to a change in their contract status. We start the discussion of our findings by looking first at the impact of contract status on fuel consumption (equation (1) above) in the short-term, followed by a detailed analysis of its impact in the long-run. We then go on to present our findings with respect to the impact of contract status on trip evaluation (equation (2) above). The main results are reported in Tables 5.3, 5.4, and 5.5 suggesting that commercial truck drivers significantly change their behavior following changes in their contract status, i.e. reduce their effort levels after having been promoted from a temporary to a permanent contract. The baseline results of our fixed-effects estimates with fuel consumption as the dependent variable are displayed in Tables 5.3 and 5.4. We use various specifications including different sets of explanatory variables to control for the robustness of our findings. To identify the impact of a change in contract status on fuel consumption in the short-run, we narrow the observation period to a six-month window and compare for each driver his performance in the last three months under the temporary contract with that during the first three months under the permanent contract. It appears from Table 5.3 that the immediate consequence of promoting a driver from a temporary to a permanent contract is a statistically significant and economically relevant increase in fuel consumption of more than 10% (3.04 l/100km). While this finding supports our assumption of a strong initial behavioral response in the first few months after having been promoted to a permanent contract, it is 67 Weather information was collected from a meteorological station located approximately in the middle of the round-trip route. This admittedly crude weather proxy helps to overcome the problem that we cannot control for truck position in detail. Thus, we assume that weather information from half the distance adequately reflects the slightly differing climate zones of the trips’ starting and arrival points. Data was provided by the German Meteorological Service (DWD, 2014). Behavioral Consequences of the Transition from Temporary to Permanent Employment 124 not yet clear whether this initial reaction remains constant over time or whether drivers change their behavior once more after some time has elapsed. Table 5.3: Impact of Contract Status on Short-Term Fuel Consumption Variables (1) (2) (3) Temporary Contract (yes=1) -1.887** -1.887** -3.035** (.628) (.628) (.903) Age (in years) .480 1.079 (4.215) (4.188) Tenure (in three-month intervals) -1.445 (.872) Vehicle Dummies YES YES YES Date Dummies YES YES YES Time Dummies YES YES YES Weather Dummies YES YES YES Driver Fixed-Effects YES YES YES Constant 37.31 25.46 -1.461 (72.72) (170.0) (171.5) Number of observations 308 308 308 Number of drivers 8 8 8 R-squared .343 .343 .350 Robust standard errors in parentheses *** p<.01, ** p<.05, * p<.10 Table 5.4 displays the results of our second estimation, identifying the long-run behavioral responses of commercial truck drivers to changes in contract status. According to the point estimates, drivers on average reduce their individual effort level upon a change in contract (e.g. from temporary to permanent). The coefficient of contract status in our preferred specification (model 3) is -.475 (standard error = .150). This implies that, other things equal, a driver uses .475 l/100km less fuel when employed on a temporary contract. On average, drivers cover an annual distance of 125,765 km and use 35,185 liters of fuel. The estimated coefficient translates into an increased fuel use of 600 liters per year after a change in contract status. Assuming a fuel price of €1.50 per liter, this amounts to additional fuel costs of around €900 per driver and year for everybody who is promoted from a temporary to a permanent job. Thus, these costs can be interpreted as “hidden costs of contract conversion” (from temporary to permanent). Behavioral Consequences of the Transition from Temporary to Permanent Employment 125 Table 5.4: Impact of Contract Status on Long-Term Fuel Consumption 68 Variables FE RE (1) (2) (3) (4) (5) Temporary Contract -.397* -.415** -.475** -.487** -.487*** (yes=1) (.180) (.161) (.142) (.150) (.151) Tenure (in three-month -.0740 -.0026 intervals) (.208) (.209) Age (in years) -.604 -.0740 (1.550) (.027) Date Dummies YES YES YES YES YES Vehicle Dummies YES YES YES YES Time Dummies YES YES YES Weather Dummies YES YES YES Driver Fixed-Effects YES YES YES YES Constant 25.51*** 25.65*** .118 39.11 0 (1.085) (1.126) (25.37) (64.72) (0) Number of observations 1,299 1,299 1,299 1,299 1,299 Number of drivers 8 8 8 8 8 R-squared .090 .116 .128 .128 Robust standard errors in parentheses *** p<.01, ** p<.05, * p<.1 To further control for the influence of age and tenure we estimated a random-effects model, the results of which are displayed in column 5 of Table 5.4. Contrary to previous research (e.g. Llaneras et al. 1998), we find driver performance to be unaffected by age and/or experience. This is perhaps surprising as the drivers in our data set are very heterogeneous in terms of age (19 to 61 years) and tenure (up to 38 months). The fact that a driver’s human capital (in terms of experience) seems to be irrelevant for his style of driving renders the impact of motivation and choice of effort (EFCH) even more important. Perhaps surprisingly, none of the thirty-eight vehicle dummies comes close to statistical significance (coefficients not displayed in table to save space). This finding might suggest that vehicle technology is not as important as driver behavior in determining fuel consumption. Yet, we believe this finding to be due to a more straightforward explanation. The hauler sorts out vehicles from its fleet after they have been used for about one million kilometers. Since this threshold level is usually reached after about two to three years, the hauler’s fleet consists of almost (brand) new trucks and all drivers are equipped with state- of-the-art technology. Therefore, it seems reasonable that our dummies representing differences in vehicle technology fail to reach statistical significance. Interestingly and in contrast to previous research (Kilpeläinen, Summala 2007), the only statistically significant 68 Age and tenure effects were estimated individually but fail to reach statistical significance. Behavioral Consequences of the Transition from Temporary to Permanent Employment 126 weather variable in terms of fuel consumption is wind speed (+.11 l). This is again plausible as trucks are prone to windage due to their size and unfavorable aerodynamic features. None of the remaining weather variables, e.g. temperature, snow fall and precipitation came close to statistical significance. In particular, we expected high temperatures to have a noticeable influence on fuel consumption since air conditioning systems are widely known to increase fuel use. With respect to the time dummies (months and weeks), no clear pattern emerges. On Tuesdays fuel consumption levels are higher (plus .77 liters per 100 km) higher than on the reference day (Mondays). No other day of the week effect can be identified. 69 The results of our fixed-effect estimations still hold after including a dummy variable representing presence of a bonus regime during the first two years of the observation period. 70 The bonus regime might be an additional factor influencing driver behavior via choice of effort (EFCH). In particular, the bonus regime might impact the motivational effects that we attribute exclusively to behavioral responses following the change in contract status. However, we fail to find any significant effect of the bonus regime (see Table A.11 in the appendix with estimations including the bonus regime). Thus, we are confident that the behavioral responses identified are entirely attributable to changes in motivation and effort choice triggered by a change in contract status. Since we are particularly interested in the timing as well as the persistence of the individual drivers’ behavioral response, we included in our estimations a trend variable counting the number of months after contract conversion for each driver. The respective coefficient indicates a statistically significant negative effect (-.140), suggesting an increase of fuel consumption after the change in contract status (see Table A.12 in the appendix). The coefficient of the squared time trend, however, is positive and significant (.004) indicating a ushaped pattern (with the turning point after 18 months). This, in turn, suggests that truck drivers in the long run return to their initial performance levels. Apparently, utilitymaximizing truck drivers seem to no longer feel the need of displaying high levels of effort 69 The coefficients for trips at night (-.008 l) and during evening rush hours (+.26 l) have the expected signs but fail to reach statistical significance. Surprisingly, the coefficient for morning rush hours (-.11 l) is negatively signed, but also fails to reach statistical significance. 70 The bonus regime was in practice from the beginning of the observation period in January 2011 until December 2012 when the hauler decided to abolish it. Although a low fuel consumption was not directly rewarded by the incentive system, drivers were financially incentivized to perform according to the measures of the on-board evaluation tool. Behavioral Consequences of the Transition from Temporary to Permanent Employment 127 after having been signed to a permanent contract. Indeed, individuals reduce their effort levels substantially and performance deteriorates considerably once a transition in contract status has occurred. However, with respect to fuel consumption this change in behavior seems to get weaker over time. We now present the findings of our second estimation using the automatically generated trip evaluation as a measure for driving performance and proxy for driver choice of effort. To account for the ordinal nature of the data, we estimate a series of ordered logit models as suggested by Gujarati and Porter (2009) as well as Wooldridge (2013). It appears from Table A.12 in the appendix that there are significant behavioral responses in terms of higher levels of effort depending on contract status (temporary vs. permanent). Table 5.5 presents the post-estimation probabilities for individual trip evaluations. Table 5.5: Estimated Probabilities for Trip Evaluation Variables Temporary Permanent Change 95% CI for Change Trip Evaluation Good .7471 .4140 -.3331 -.4837 -.1825 Mediocre .0404 .0559 .0155 -.0045 .0355 Poor .2124 .5300 .3176 .1682 .4671 According to our estimations the probability of a trip being evaluated as “good” is 74.7% when a driver is on a temporary contract and 41.4% when a driver is permanently employed. Thus, the probability of a performance that is in accordance with the employer’s expectations is 33.3 percentage points (or 80.4%) higher in the case of the former workers. An inverse pattern appears for trips evaluated as “poor”. Here the probability for temporary workers is 21.2% compared to 53.0% for drivers on permanent contracts. This yields a difference of 31.8 percentage points or 150%, respectively. This result confirms our main finding that drivers on temporary contracts have strong incentives to signal motivation and high levels of effort. After having signed a permanent contract, drivers quickly reduce their effort levels leading to a lower overall driving performance evaluation. However, a closer look at the data reveals that when comparing both fuel consumption and trip evaluation before and after promotion to a permanent contract, only half of the drivers perform worse after they have been promoted. Thus, while some drivers show (negative) behavioral reac- Behavioral Consequences of the Transition from Temporary to Permanent Employment 128 tions to a change in contract status, others do not (this finding is in line with e.g. Nagin et al. 2002 as well as Riphahn and Thalmeier 2001). We admit that due to the organizational setting of our study we have to be aware of some limitations that preclude generalization of our findings. First, the number of drivers (n=8) is quite small. However, the data set consists of at least 100 observations per driver resulting in a total number of observations of 1,299. We are, therefore, confident that our results are robust and reliable. Second, due to the hauler’s data privacy restrictions of drivers’ personnel records, we are unable to include in our estimations measures of driver qualification and former experience with other employers. Yet, since we control for age and tenure with the current employer, we believe to have included sufficient proxies for qualification and past driving experience. Moreover, the coefficients of both, age and tenure failed to reach statistical significance indicating a negligible importance of qualification and (former) driving experience. 5.5 Conclusion Using a hitherto unavailable data set compiled from the fleet management system of an inhouse hauler of a large European truck manufacturer, we provide robust evidence on employees’ behavioral responses to changes in contract status (i.e. following the transition from a temporary to a permanent contract). We analyze commercial truck drivers’ effort choices before and after they have been signed to a permanent contract using two performance variables – fuel consumption per 100 km and an automatically generated trip evaluation measure. In line with previous research (Bradley et al. 2012), our findings suggest that commercial truck drivers choose their effort levels depending on the nature of their contracts, i.e. they exert more effort (in the sense of using less fuel and better trip evaluations) when on temporary contracts. After having been promoted drivers adjust their effort levels, i.e. use more fuel and receive more “bad” evaluations. These findings are robust on various specifications and survive a number of robustness checks. Drivers on temporary contracts (or more general: workers) – being aware of or just assuming screening activities to be used by the hauler (or more general: firms) – choose high levels of effort to signal motivation and dedication. However, as soon as drivers (workers) are signed to permanent contracts, they reduce effort back to “normal” levels. When employed under a permanent contract, fuel consumption of drivers is 600 liters higher per year (the additional costs per driver are Behavioral Consequences of the Transition from Temporary to Permanent Employment 129 around 900 € per year) than when employed under a temporary contract. We interpret this as the hidden costs of changes in contract status (from temporary to permanent). This negative (expensive) behavioral response is particularly strong in the first months after contract conversion but then declines to reach the initial level again after about 18 months. Moreover, for drivers on temporary contracts the probability of a trip being evaluated by the trucks’ computer system as “good” is 55% higher than for drivers on permanent contracts. Yet, we find only half of the drivers to display these behavioral changes when being promoted from a temporary to a permanent contract. The other half does not respond to the new contract status but instead perform at their initial effort levels. This finding is in line with Nagin et al. (2002) who also find that only a fraction of employees behave as “rational cheaters” when given the opportunity to do so while many others resist that temptation. Where do the incentive effects of temporary contracts come from? Permanent contracts have been found to be associated with better working conditions, higher pay, more job security and higher levels of employment protection. Most temporary workers want to recommend themselves for permanent contracts to enjoy increased levels of job security and dismissal protection legislation. This is particularly true in our case as the country where the headquarters of the hauler is located has most recently been ranked below average on OECD’s “Indicators of Employment Protection” for workers on temporary contracts and above average for workers on permanent contracts (OECD 2013). Moreover, we are not aware of any differences in working conditions for temporary and permanently employed drivers at this particular firm. What are the practical implications that can be derived from our results? From an organizational point of view, the results seem to suggest keeping employees on temporary contracts as long as possible to maximize the returns from higher effort levels. Yet, labor laws rule out recurring extension of temporary contracts. It may, therefore, appear an even more promising strategy to employ temporary workers only. However, in the absence of promotion opportunities (from temporary to permanent contracts) temporary workers lack the necessary incentives to choose high levels of effort. Dolado and Stucchi (2008) find that firms with a high contract conversion rate (by promoting workers from temporary to permanent contracts) display higher levels of labor productivity. Thus, it seems a reasonable strategy to openly communicate the screening nature of temporary contracts since then temporary workers will be motivated to choose high levels of effort. Moreover, offering properly designed monetary incentives during the weeks and months after contract conver- Summary and Future Outlook 136 author’s knowledge – has not yet been used in personnel economics research. Since availability and quality of this objective real-time data is overwhelming, its study may offer broad advantages in analyzing employee behavior. This thesis, therefore, makes a relevant claim to include this innovative and promising type of data into personnel economics. Second, the use of objective performance measures as indicators of worker productivity depicts a distinct advantage over most existing research that is often based merely on subjective (self-) rated performance. This is particularly true for absenteeism research since register absence data is often unavailable to researchers. As a consequence, most existing findings on absence are based on self-reported absence figures that are known to be biased due to underestimation with actual absence being twice as high (Johns 1994). Hence, the objective absence data used in chapter three depicts a clear advantage. Similarly, analyses in chapters four and five benefit from computerized performance evaluations that are collected automatically and objectively rate performance based on predefined algorithms. In general, the use of high quality objective performance measures constitutes a major advantage of the dissertation at hand. It addresses weaknesses which can be found in parts of the existing productivity research that often rely on subjective performance measures. Based on objective performance data, the presented findings complement existing evidence and, thus, contribute to the progress of the discussion in personnel economics. Third, despite the fact that incentives are often framed to be the core of personnel economics (e.g. Lazear 2000b) and as such are probably one of the most intensively studied fields in HRM research, very little is known about responses of employees to the abolition of incentives. To the best of the author’s knowledge, the abolition of incentives has not been studied that often – Freeman and Kleiner (2005) representing a rare exception. By studying employees who experience the abolition of an existing incentive scheme, this work may contribute to a better understanding of the general functioning of incentives. Eventually, this thesis contributes to absence research in three ways. First, as mentioned above the use of register data in absence research depicts a clear advantage over most of the existing research that lacks access to objective absence figures from inside a company. Second, the absence data used in this study is advantageous in that it was compiled in different international plants of the same company. Hence, the data is coherently recorded using identical corporate standards at all plants. This allows for international comparisons without facing the common drawbacks of varying reporting methods usually associated with international absence data (European Foundation for the Improvement of Living and Working Conditions 2010). Third, absence here is primarily evaluated at the group level. Despite the fact that absence Summary and Future Outlook 137 is a social concept and absence behavior is to a large extent influenced by worker’s peers, personnel economic knowledge on absenteeism is to a great part based on individual-level data (e.g. Rentsch, Steel 2003). Chapter three addresses this gap in absence research. In total, the findings of this thesis contribute to the ongoing discussion of the social and economic determinants in personnel economics by offering important insights on teamwork and incentives based on unique data. Despite its valuable contribution to personnel economics the work at hand faces some limitations originating from both econometric methods and data. In general, meta-analyses face two particular challenges (e.g. Egger, Smith, Sterne 2001): the publication bias as well as the garbage-in-garbage-out issue. First, studies with positive and significant findings are more likely to be published in peer-reviewed journals. As a result, meta-analyses that focus solely on published articles only cover significant findings and leave out controversial or inconclusive results, leading to the emergence of a publication bias. Second, the quality of any meta-analysis is dependent on the quality of the studies reviewed. To account for this pitfall – referred to as garbage-in-garbage-out effect – only studies published in high-quality peer-reviewed journals are included in the meta-analysis presented in chapter two. Albeit the focus on peer-reviewed articles might provoke a publication bias, this procedure seems to be particularly crucial to identify the most important contributions to the extensive literature on team diversity. Moreover, any publication bias is mitigated by the fact that the findings included in the meta-analysis in chapter two show positive as well as negative correlations. A common issue of insider studies is centered on its single-firm data sources. Although this allows for high-quality research at the micro-perspective, the generalizability of results is limited (Ichniowski, Shaw 2013). Moreover, this research is likely to suffer from selection bias and, therefore, it also suffers from endogeneity in the choice of workers and managers as it focuses exclusively on a specific industry. Results based solely on data and information originating from the automotive industry may not be presumed to be universally valid and, therefore, should not serve as a blueprint for other industries without further investigation. A distinctive limitation to all non-experimental data – and thus to insider econometrics as well – is that the choice of a treatment by means of a particular management practice is not random but instead is the result of a maximization decision taken by the company. Summary and Future Outlook 138 In conclusion, the evidence and implications discussed in this thesis provide important contributions to the strands of insider econometrics and personnel economics, in particular to the fields of incentive design and work organization in teams. With regard to the methods applied and topics studied, the work at hand concludes by pointing out promising suggestions for future research. First, although having a long tradition in other scientific fields such as medicine, the meta-analytic approach is only slowly gaining credence in economic research. However, results of this thesis demonstrate that meta-analyses are a useful means to overcome limitations of small sample sizes and gain an overview on topics with extensive, but so far inconclusive empirical evidence. In particular, topics that can be assessed as being “over-researched” may benefit more from meta-analyzed conclusions than from additional empirical findings. Second, modern times are changing the circumstances of firms and employees alike as complexity of cooperation and processes increases. Therefore, economic models need to be adapted and refined by incorporating information and data from within companies. On this account, insider studies are a key to improving the quality of both theoretical models and implications for practitioners derived from research. Thus, further contributions to insider econometrics are highly appreciated. Third, in order to gain a comprehensive understanding of an economic phenomenon, it is crucial to analyze it in all its facets. For instance, in response to the well-studied field of incentive introduction, this work analyses the behavioral consequences of its abolition. As a consequence, the presented findings may contribute to an overall better understanding of incentives. Therefore, more studies that depart from familiar paths and shed light on hitherto unstudied aspects of well-known subjects are needed. Fourth, as mentioned earlier, the use of data from GPS-based fleet management systems has proven to contribute significant findings to personnel economics. Thus, scholars should keep their eyes open to new and innovative data sources that so far have not been taken into account for scientific research. The digital age can be assumed to offer further “treasure chests” in terms of computerized data. Eventually, this dissertation emphasizes the advantageousness of the cooperation between academia and businesses in the context of insider econometrics. The author not only acknowledges the great opportunity to work with unique and otherwise unavailable data from inside a company but also highly appreciates the chance to discuss findings with internal experts in order to refine interpretations and derive more comprehensive implications. The author of this thesis, therefore, invites researchers as well as practitioners to join their forces in the search for valuable insights that benefit both scientific progress and companies’ bottom-lines. Appendix IX APPENDIX Table A.1: Heterogeneity Measures Listed by Frequency of Use Diversity measure Blau CV Teachman s.d. Other* Age 4 14 2 5 3 Gender 14 1 5 1 8 Culture 13 1 3 - 2 Function 14 - 5 - 3 Tenure 2 16 - 4 1 Education Background 7 - - 3 3 Level 4 3 2 - 1 * Other heterogeneity measures include Herfindal index, Gini index, mean, percentage share and authors’ own modifications. Figure A.1 : Forest Plot – Age Diversity and Overall Performance Source: Own calculations. Appendix X Figure A.2: Forest Plot – Gender Diversity and Overall Performance Source: Own calculations. Appendix XI Figure A.3: Forest Plot – Culture Diversity and Overall Performance Source: Own calculations. Appendix XII Figure A.4: Forest Plot – Tenure Diversity and Overall Performance Source: Own calculations. Appendix XIII Figure A.5: Forest Plot – Functional Background Diversity and Overall Performance Source: Own calculations. Appendix XIV Figure A.6: Forest Plot – Educational Background Diversity and Overall Performance Source: Own calculations. Appendix XV Figure A.7: Forest Plot – Education Level Diversity and Overall Performance Source: Own calculations.