Essays on Changing Nature of Work and : Implications for the U.S. Labor Market

by

Hye Jin Rho

B.A., Northwestern University (2006) S.M., Massachusetts Institute of Technology (2016)

Submitted to the Sloan School of Management in partial fulfillment of the requirements for the degree of

Doctor of Philosophy at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

September 2018

Massachusetts Institute of Technology 2018. All rights reserved.

Signature of ASignatureu io redacted Hye Jin Rho MIT Sloan School of Management Signature redacted July19,2018 Certified by: Thomas Kochan George Maverick Bunker Professor of Management Signature redacted Thesis Co- Certified by: Paul Osterman NTU Professor of and Management Signature redacted Thesis Co-Supervisor Accepted by: Catherine Tucker OF TECHNOLOY Sloan Distinguished Professor of Management Chair, MIT Sloan School of Management PhD Program SEP 27 2018 1 LIBRARIES AKk;IV t:. Essays on Changing Nature of Work and Organizations: Implications for the U.S. Labor Market

by Hye Jin Rho Submitted to the MIT Sloan School of Management on July 19, 2018 in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy in Management

Abstract

This dissertation examines how the changing nature of work and organizations has altered the U.S. labor market to influence outcomes for seekers (1) in alternative work arrangements and (2) of different genders. The first essay describes recent developments in the labor market for nonstandard workers, that is, an increase in the variety of pathways through which nonstandard workers are assigned to work. I suggest that changes in the regulatory environment, the rhetoric around competition, and technological developments have shaped inter-organizational relationships and norms in the industry to bring about a very different system of labor markets than was traditionally understood. I contend that such a multifaceted employment model with a diverse set of exchanges among multiple actors has profound implications for the future of IR research.

The second essay examines the "multi-layered labor contracting" structure in which the of nonstandard workers is outsourced to an intermediating , who then selects workers from a group of competing suppliers. Drawing on power-dependence theories, I examine the link between these new contractual relationships and economic outcomes for lead firms and workers. Using proprietary data from employment records of nonstandard workers in Fortune 500 firms, I find that an additional contracting layer between lead firms and workers is associated with higher returns to firms and lower returns to workers. The loss from an additional contracting layer is reduced when workers gain bargaining power through pre-existing relationships with the firm.

The third essay addresses how interactional processes between employers and job seekers at an initial recruitment phase online influence gender sorting ofjob seekers. We use unique data from a field study and (Study 1) a field experiment (Study 2) of online job postings to test two distinct interactional mechanisms: gendered language (as experienced by job seekers) and in-group preferences (as exercised by job seekers). We mostly find support for our predictions that, compared to male job seekers, female job seekers are more likely to show interest in and apply to a job when the job is described using more stereotypically feminine words or by female recruiters.

Thesis Committee

Thesis Co-Chair: Thomas Kochan Title: George Maverick Bunker Professor of Management

Thesis Co-Chair: Paul Osterman NTU Professor of Human Resources and Management

Emilio J. Castilla NTU Professor Management

2 Acknowledgements

I still remember the day I met my advisor Thomas Kochan at a Labor and Employment Relations Association (LERA) meeting before my time here at MIT. Inspired by his work and the way he teaches the community of business leaders, I decided to pursue the long journey of getting a doctoral degree. Ever since, he has always supported and encouraged my work, pushed me to become a better scholar, and never hesitated to help me every step of the way. He was a critical part of each of my learnings and successes; I am forever indebted to him for standing by me through finish. I am deeply thankful.

Paul Osterman has been a tremendous source of inspiration and knowledge from day one at MIT. I learned so much from the way he integrates academic research and policy. He always pushed me to sharpen my ideas in ways that I could not accomplish on my own. After my meetings with him where I was challenged and questioned, I frequently saw my research improve to a significant degree. Importantly, I have to say that his sense of humor and occasional jokes in seminars, meetings, and classes brought me composure at times of stress and tension. Thank you and I will miss them.

I benefited enormously from the of Emilio Castilla. We occasionally had long and extended meetings in his office, at a coffee shop, or over Skype where we brainstormed and discussed our research. Throughout the process, I learned to become a better thinker and a more meticulous empiricist. He worked with me through different analytical methods and new ideas that I will continue to use as a source of guidance going forward. He also taught me how to collaborate effectively with those outside of academia, which will forever be an asset. Thank you.

I am also indebted to faculty members including Susan Silbey, Erin Kelly, Ezra Zuckerman, Michael Piore, Robert McKersie, Richard Freeman, Rosemary Batt, Jody Hoffer Gittell, and David Weil. I thank Eileen Appelbaum and John Schmitt for their guidance and encouragement from Washington D.C. before and during my time here. I appreciate the assistance provided by those at my research sites, RecruitCo and ContingentCo. I am especially thankful to the president at ContingentCo, as this dissertation would not have been possible without his trust.

My friends and colleagues have always supported me throughout the process. Gokce Basbug, the other me at IWER, and Christine Riordan, I really could not have done this without you. I look forward to continuing to support each other throughout our . I also thank the IWER community, including Ben Rissing, Alan Benson, Mahreen Khan, William Kimball, Alexander Kowalski, Duanyi Yang, George Ward, and Claire McKenna. In particular, I thank Andrew Weaver for always encouraging me and reassuring me about my work when in doubt.

Eunbin Whang, your emotional support throughout the years was precious. Our countless hours of conversations and spontaneous meetups really helped me get through. Thank you. Hyejun Kim, my dearest three-year roommate, Donghee Jo and Doojoon Jang from MIT, I will always remember our Pocha days. I am so glad that we were able to do this together. Erica Lee, I am so grateful that you were here to see me both start and finish my program. Alex Cho, thank you for always being there throughout my successes and failures. Michelle Sodam Park, I owe a special thank you.

I dedicate my dissertation to my family, Seong Cheol, Ae Ran, Esther H., and Brooke H. Rho. Thank you God.

3 Table of Contents

Chapter 1. A Multifaceted Model of Employment Relationship for Nonstandard Work...... 6 1.1 Introduction...... 7 1.2 Developments in the Labor Market for Nonstandard Work ...... 10 1.2.1 Classic Em ploym ent M odel for N onstandard W ork...... 13 1.2.2 Mid-90s and beyond: The Rise in "Joint Employment" and Misclassification Lawsuits 16 1.2.3 Centralization and Outsourcing of Nonstandard Workforce Management ...... 18 1.2.4 Technological Advances in the 2000s and Legitimatization of Competitive Bidding .. 21 1.2.5 Em ergence of the "Hum an Cloud" Fram ework...... 25 1.3 A Case Study of M ultifaceted Employment Relationship...... 27 1.4 Discussion and Future Research...... 32 1.5 References...... 35 1.6 Tables and Figures...... 38 Chapter 2. Multi-layered Labor Contracting and Distribution of Power: Evidence from Em ploym ent Records for Nonstandard W ork...... 44 2.1 Introduction...... 45 2.2 Multiple Layers of Contracting and Power...... 47 2.3 Background...... 49 2.3.1 Buyer Power, Profitability, and W age Determ ination...... 52 2.4 Data and M ethod...... 58 2.4.1 Description of Variables ...... 60 2.5 Results ...... 64 2.5.1 M ain Analysis...... 65 2.5.2 Supplier Selection...... 68 2.5.3 Heterogeneity in Skills Dem anded ...... 70 2.6 Discussion...... 74 2.7 References...... 78 2.8 Tables and Figures...... 82 2.9 Appendix ...... 89 Chapter 3. Language and Gender in the Online Job-Matching Process ...... 91 3.1 Introduction...... 92 3.2 Interactional Mechanisms at Work: The Context of Online Job Matching ...... 95

4 3.2.1 Gendered Language of Job Descriptions ...... 96 3.2.2 In-Group Preference between Job Seeker and Recruiter ...... 99 3.3 Study 1 - Field Study of Online Job Postings...... 101 3.3.1 Study D esign of Study ...... 101 3.3.2 D ata of Study ...... 103 3.3.3 Results of Study ...... 110 3.4 Study 2 - Field Experim ent Study of Online Job Postings ...... 116 3.4.1 Study D esign of Study 2...... 116 3.4.2 D ata of Study 2...... 120 3.4.3 Results of Study 2...... 121 3.5 Lim itations and Future Research...... 124 3.6 D iscussion...... 127 3.7 References...... 131 3.8 Tables and Figures...... 136 3.9 Appendix ...... 146

5 Chapter 1. A Multifaceted Model of Employment Relationship for Nonstandard

Work

Abstract

Today, the industry for recruiting, selecting, and workers in alternative work arrangements reveals a much more complex and diverse picture than what industrial relations (IR) scholars have documented since the 1990s. Despite the increase in the variety of pathways through which nonstandard workers get hired to a work assignment-and its potential implications for labor market outcomes-these changes have not been documented. This paper makes a long overdue update of recent developments in the labor market for nonstandard workers. In the process, I develop a conceptual framework for understanding the labor market outcomes by providing potential explanations of how different forms of nonstandard work relationships have come to exist today. I suggest that changes in the regulatory environment, the rhetoric around competition, and technological developments have shaped inter-organizational relationships and norms in the industry to bring about a very different system of labor markets, even for the same type of nonstandard worker. I contend that the employment system today is moving towards a multifaceted model that encompasses a diverse set of exchanges among multiple actors and that these new developments have profound implications for the future of IR research.

6 1.1 Introduction

Work arrangements alternative to full-time regular employment have been rapidly rising over the last several decades. Katz and Krueger (2016), in their recent update of the U.S. Bureau of Labor

Statistics (BLS) Contingent Worker Survey (CWS), found that in 2015, 15.8 percent of workers

(compared to 10.1 percent in 2005) were employed under some type of alternative work arrangement as temporary help agency workers, on-call workers, contract workers, and independent contractors or freelancers. These alternative work arrangements challenge existing paradigms used to understand the human resource practices of firms and their internal labor markets (ILMs) that have predominantly defined the postwar employment system. Grounded in institutional thinking, industrial relations (IR) scholars in the past half-century have studied the internal personnel/human resource management rules that formalized job ladders and promotion systems in union firms and instituted similar rules in non- union firms in the absence of unions. IR scholars viewed ILMs as having insulated employment from external market forces, thus providing some degree of employment security and voice for a particular set of workers, some of which were outcomes that differed from those predicted by traditional labor economics (Doeringer and Piore 1971; Osterman 2011; Osterman and Burton 2004).

Developments in recent scholarship recognize that most alternative work arrangements today instead are based on employer-worker relationships that are outside the typical boundaries of a firm

(see, e.g., Bernhardt 2014; Bernhardt et al. 2016). While the traditional ILM was developed on the premise that relationships between the worker and the purchaser of the labor are bilateral and internal, the new norm no longer assumes such a direct relationship. New relationships involve third-party employment agencies, vendors, or online platforms that help facilitate the relationship between the hiring "lead firm" (i.e., firms with temporary labor demand) and the worker. Existing literature therefore focuses on labor market intermediaries that make up "triadic" employment relationships in particular and their role in how work is organized (for a review, see Bonet, Cappelli, and Hamori

2013) and how rewards are distributed within organizations (for a review, see Bidwell et al. 2013).

7 As this paper will demonstrate, however, the new empirical reality reveals that there are many additional building blocks that make up the increasingly "fissured" (Weil 2014) modem labor market.

Today, the labor market for nonstandard or contingent workers frequently involves not only multiple intermediating organizations with competing interests, but also a substantial variety of the job- matching process, even for the same type of nonstandard work. Consider, for example, the different pathways through which independent contractors (one type of nonstandard worker) are hired to a specific work task, which range from involving no intermediaries to many. In one scenario, independent contractors connect to a lead firm directly for a specific work project using their own social networks or pre-existing working relationships with the firm. Alternatively, they can use an intermediating technology platform online such as Upwork, Taskrabbit, or Mturk, where they can directly set the terms of their work contract. Both are relatively direct relationships and these workers are typically considered "independent" and self-employed. In a third scenario, the hiring lead firm decides to use a staffing agency to manage for independent contractors who either come through an online platform or are hired directly by the firm through a previous working relationship.

In such a case, these workers also differ from agency temps in that they are responsible for their own

self-employment taxes and do not receive any employer-provided benefits that agency temps

occasionally receive from the staffing agency. In a fourth scenario, the lead firm outsources its

management of staffing agencies to a third-party business process outsourcing agency. The process

outsourcer then, in place of the lead firm, selects independent contractors from a wide range of staffing

companies that have found workers on- or off-line and supplies them to the lead firm. These examples

are simplified for illustrative purposes, but show how each of these scenarios may create different

labor market characteristics and outcomes for the independent contractors.

8 Then, depending on how organizations hire, or rather, "onboard"' workers to work assignments (and similarly, how workers are assigned to work with specific lead firms), the structure of the employment relationship starts to exhibit a very different system of labor markets than traditionally understood, even within the same type of nonstandard work. The new work relationships, which I refer to as the multifaceted employment relationships, have profound implications for employment relations research. They fundamentally alter the ways in which profits flow along a labor supply chain and subsequently alter the business and regulatory environments. And importantly, these multifaceted employment relationships affect how workers perceive their work tasks, manage time, and cope with workplace conflicts.

The main goal of this article is to fill the gap in our understanding of the labor market for nonstandard work by providing an overview of these developments. I develop a conceptual framework for analyzing labor market outcomes for workers in these new alternative employment relationships by providing potential explanations for how different forms of nonstandard work relationships today have come to exist. An examination of why certain strategies for hiring and managing nonstandard workers are adopted would be helpful in understanding the impacts of different hiring pipelines on workers. I focus on the role of organizations in their relationships with other stakeholders as the main contributor to the increasing inter-organizational complexity in this labor market. Taking an institutionalist perspective, I contend that the construction of a fissured labor market does not simply delineate a breakdown of ILMs, but rather, an active reconfiguration of personnel rules and a deliberate separation from the traditional ILMs. Firms create (or are exposed to) an alternative space for managing a set of workers that was once considered difficult to manage, not

1The term "onboarding" is frequently used by industry practitioners to refer to the "process of bringing a worker into a position with a goal of providing all necessary tools to be productive as soon as possible" (Staffing Industry Analysts 2017: p.27). Since it is not commonplace for firms to "hire" nonstandard workers directly as their own employees, the term "onboarding" will be used throughout to describe the recruitment processes that enable alternative work arrangements.

9 only to maximize efficiency, but also to address the manage increasing regulatory risks and political struggles with intermediating organizations.

In subsequent sections, I outline variations in employment relationships governing alternative work arrangements today, examine the roles of new organizational and institutional constituents, and document how they have evolved over time driven by the changes in (1) the regulatory environment,

(2) the rhetoric around competition, and (3) the technological developments. To portray the realities of the developments in the industry, I use a proprietary dataset of contingent workforce onboarding for a small subset of Fortune 500 firms. In the process, I seek to make a long overdue update of our conception of the industry so that we can begin to understand the new labor market in which today's contingent workers are deeply embedded.

1.2 Developments in the Labor Market for Nonstandard Work

Over last three decades, the increase in the variety of organizations that intermediate the job- matching processes between the lead firm and the worker has fundamentally altered the landscape of the labor market for nonstandard workers. Especially for large firms that make extensive use of nonstandard workers, the hiring of nonstandard workers is no longer a simple strategic personnel decision made by HR and/or line managers, where reliance on one staffing company or a set of contracting companies fulfilled most short-term based labor needs. Rather, such hiring is now an intermingled process of human resource and procurement strategies accounting for how to meet labor needs, negotiations with a wide range of intermediaries and workers to set prices and responsibilities, and a search for available technical tools for talent management that are tailored to firm-specific needs.

Table 1 provides an overview of how new intermediating organizations have been introduced in the labor market over the last several decades to perform distinct human resource functions and roles that were previously the domain of a single employer (to be elaborated in detail throughout the

10 paper). The first row represents an employment model for direct-hire temps or independent

contractors who are in direct contact with the lead firm without any intermediaries; the second row

represents the traditional triadic model for agency temps. Each row represents an addition, and not a

substitution, of new relationships and arrangements. In other words, by mid to late 2010s, not only do

we continue to see the hiring and managing of nonstandard workers through traditional models, but we

also observe many varieties of new external vendors and technological platforms that were introduced

and developed over time with respect to the specific role they play. For example, an external vendor

(MPS: Managed Service Provider), as the sole standing intermediary in the mid-i 990s (Row 5)

entered the market to become an official employer of record and provide most of the traditional HR

functions in place of the lead firm. By the 2000s, some MSPs decided to split off their HR functions

to other intermediating organizations, including the staffmg agencies working with the MSPs

(identifying, attracting candidates, and payrolling) and the technological platforms (management of

hiring process, on- and off- boarding from work), while the MSPs themselves focused on selecting and

recruiting candidates based on the lead firms' needs.

[Table 1 about Here]

What explains the rise of such fragmentation in the industry? Under the standard economic

accounts of competitive markets, the creation of a fissured labor market for nonstandard workers-the

introduction of new actors and their changing roles -should be driven by an efficient

allocation of human capital, where rational organizations develop relationships and adopt labor market

strategies that are mutually beneficial. Relatedly, studies point to cost reduction and/or flexible hiring

strategies as key drivers in firms' growing use of nonstandard workers. Some evidence suggests that there is a penalty for agency temps (Kilcoyne 2005; Peck and Theodore 2006; Segal and

Sullivan 1997), particularly for low-skilled , but it is unclear whether the reduction of ultimately leads to cost savings to the firm after a portion of the savings is paid to the intermediaries.

The flexible hiring explanation (both numerical and functional flexibility), on the other hand, is based

11 on the view that organizations need to be able to adjust the number of workers flexibly to meet occasional labor shortages (especially in seasonal industries such as agriculture (McLaughlin and

Coleman-Jensen 2008)) and cyclical sensitivity in demand (Houseman and Heinrich 2015). This view, however, is insufficient in explaining how a variety of hiring patterns emerged when multiple

LMIs entered the market.

Alternatively, a different strand of research emphasizes the distinct role of organizational constituents shaped by distinct political and social processes, such as social norms, competing interests, and institutional arrangements. The temporary help services (THS) industry, for example, is seen to have actively reinvented the image of the temporary as "the good temp" (Smith and

Neuwirth 2008) while portraying the permanent worker as costly and profit-hurting (Hatton 2011;

Smith and Neuwirth 2008). These scholars argue that it was these deliberate social and cultural processes that established the meaning and legitimacy of nonstandard work throughout much of the late 20' century. Staffing agencies are also described as an "active agent" of labor market liberalization by Peck and his colleagues, who argue that staffing agencies were able to identify and exploit new deregulating labor markets to broaden their presence and to shape new labor market practices and norms (Peck, Theodore, and Ward 2005). In a related study, Autor (2003) finds that the changes in the legal environment between the 1970s to 1990s in the United States, where employers were blocked from unjust of workers, are in part responsible for the rapid rise in the temporary help industry during that period.

According to Osterman, "institutional considerations are more than a backdrop against which is played out: they shape how it is used, that is, the rules that govern its

implementation" (2011:647). Consistent with the view that takes into account organizational and

institutional considerations, I show that, while some organizations choose to embrace the market- based cost optimizing strategy in creating a fissured labor market for nonstandard work, others make

deliberate HR and procurement decisions driven by three major developments that have shaped the

12 norms and rhetoric in the industry. The developments I identify are first, the rise in lawsuits over

"joint-employer" liabilities and misclassification of employees that increased the need for risk

management; second, the rhetoric about "vendor neutrality" as part of a legitimization process for

instituting competitive system of bidding workers to assignments; third, the technological advances

enabling a wide range of firms' strategic choice.

1.2.1 Classic Employment Model for Nonstandard Work

The traditional employment model for nonstandard work consists of work arrangements that

either (1) involve a single labor market intermediary such as staffing agencies or vendors who assume employment liabilities in place of the lead firm (agency temps or contractors) or (2) are based on a direct employer-worker relationship without any employment-related benefits (independent contractors). The temporary help services (THS) employment more than doubled as a percentage of the total U.S. workforce from 1990 to 2008 (Luo, Mann, and Holden 2010). By 2015, 4.7 percent of the U.S. labor force were working with contract firms or temporary help firms, and 8.4 percent were independent contractors (Katz and Krueger 2016).

The most popular characterization of temporary workers goes back to the late 1940s and

1950s, when the iconic image of "Kelly Girl" was used for supplying temporary low-wage, administrative-and mainly female-workers to businesses (Hatton 2011). The temporary help services (THS) industry that once focused primarily on service workers expanded in the latter part of the 2 0 ' century to encompass a large share of blue-collar workers (Houseman and Heinrich 2015;

Smith and McKenna 2014) and others in high-skilled occupations (Barley and Kunda 2011; Smith and

Neuwirth 2008). Temporary help or staffing agencies typically serve as an intermediary organization between lead firms and workers seeking either or permanent work through a temporary stepping stone. This triadic exchange structure has become the archetype of the "new" employment relationship, as such relationship also frequently defines the recruitment of regular, full- time workers today with the prominence of LMIs such as firms, outplace services,

13 and professional employer organizations (PEOs) (for a review, see Bidwell et al. 2013; Bonet et al.

2013).

In the case of alternative work arrangements, such triadic exchange structure also holds true when the lead firms use on-site vendors ("vendor-on-premises") or contract firms to hire a group of contractors to work on a particular task or project at the host's location (Row 3 of Table 1). The contractors typically sign a contract, often in the form of statement-of-work ("SOW") to govern the terms of a work project.2 The SOW workers are different from temporary agency workers in that they

are typically hired for their services and solutions, with a clearly defined scope of work (work details,

schedules, and deliverables). While temporary workers are paid based on the number of hours

worked, SOW workers are usually paid a for the project from the contract firm (official

employer of record). They are therefore often considered contracted consultants or third-party

workers as they work for the lead firm's outsourced vendor.3 Figure 1 illustrates these triadic

employment relationships, which were the dominant form of staffing nonstandard workers until the

mid-1990s.

[Figure 1 about Here]

These labor market intermediaries (LMIs) have changed the boundaries of the firm and

externalized labor management functions to outside of the firm. With the introduction of LMIs,

traditional internal human resource activities--e.g., identification of labor needs, attraction and

selection of workforce, hiring or onboarding, and development, and payrolling-that were

primarily the domain of internal hiring managers in HR are shared with LMIs. While internal HR and

team managers in a business unit may still participate in the selection and onboarding of candidates

2 Any external labor services in the realm of consulting, professional services, and outsourced services that require a SOW is considered to be a SOW-based project. 3 IQ Navigator White Paper. 2013. "Defining and Managing SOW Services".

14 (see e.g., Bidwell and Fernandez-Mateo 2010), staffing agencies are the primary suppliers of labor by

seeking and identifying talent and providing the initial screening of workers.

Sociological studies on brokerage suggest that because they connect workers and firms that would otherwise not interact, staffing agencies as brokers may create or capture economic value through the job-matching process (Marsden 1982). These agencies sometimes have better information about the candidates they source by having worked with some of them for multiple assignments

(Bidwell and Fernandez-Mateo 2010; King, Burke, and Pemberton 2005) or by building relationships that firms value more (Fernandez-Mateo 2007), which gives them a competitive advantage when deciding how much to bill the lead firms or compensate the workers. They are also found to influence the quality of candidates as they have an incentive to, on one hand, guarantee expertise in identifying good talent or, on the other hand, meet efficiency goals by providing quick sorting of candidates

(Cepin 2012; Khurana 2002). Staffing agencies therefore have considerable brokerage power and discretion in the hiring process, not just in terms of setting the price and compensation structure, but also in influencing other labor market outcomes such as gender segregation during screening

(Femandez-Mateo and King 2011).

Further, staffing agencies under this scenario become the legal employer of record and bear the regulatory responsibility when it comes to managing , employment taxes, and, sometimes, fringe benefits.4 This means that not only do the lead firms no longer carry the administrative burden of hiring a permanent worker, but they are not legally liable for workplace conflicts, such as workplace-related injuries or an unfair termination of workers from the job. Fins that use independent contractors are similarly not liable for employer responsibilities. While independent contractors occasionally work with staffing agencies (Evans, Kunda, and Barley 2004), most typically

4 There are some but limited number of agencies today who choose to be high-road by offering a more complete benefits package including , benefits, severance pay, and workers' compensation benefits.

15 form the direct exchange relationship with their service-purchasers (often by signing a SOW) and are not subject to any legal employment relationship as they are commonly considered to be self-sufficient free agents (Barley and Kunda 2004). For the lead firms, then, the institutional arrangements that govern the rules of employment become an important consideration in their decisions to use alternative workforce arrangements (Autor 2003).

1.2.2 Mid-90s and beyond: The Rise in "JointEmployment" and Misclassification Lawsuits

A problem arises, however, when the lines between nonstandard workers and regular employees at the lead fim start to blur in daily working relationships. As nonstandard workers continue to conduct tasks at the lead firm's physical location under the supervision of the lead firm's manager, these workers may meet the "common law test" to fall under the category of an employee of the lead firm. The U.S. Internal Revenue Service (IRS) follows the common law test and considers 20 factors related to the degree of financial and supervisory control to identify whether there is an employer-employee relationship. If the relationship between the lead firm and nonstandard worker qualifies as an employer-employee relationship under this test, the lead firms can then be identified as the co- or joint-employer, meaning that they will also become liable for providing employment protections. In addition, employers are also subject to lawsuits over misclassifications of employees as independent contractors, based on previous legal interpretations of the Fair Labor Standards Act

(FLSA) on the definition of an employee.

These regulatory risks began to surface as a key consideration in hiring nonstandard workers in the mid-1990s when the IRS began probing into cases of misclassification and joint employment.

Following an IRS audit of Microsoft in early 1990s, Microsoft had to pay back payroll taxes for independent contractors who were viewed by the IRS as employees. Then in 1993, nonstandard workers, composed of both independent contractors and staffing agency temps at Microsoft, filed a

16 class action lawsuit against Microsoft to claim benefits that were available to regular employees.5 In

this 1996 landmark case, Vizcaino v. Microsoft, the U.S. Ninth Circuit Court of Appeals in San

Francisco applied the common law test to side with workers and Microsoft was required to pay $97

million. According to Justice Stephen Reinhardt of the Ninth Circuit, "Large corporations have

increasingly adopted the practice of hiring temporary employees or independent contractors as a means of avoiding payment of , and thereby increasing their profits" (Vizcaino v.

Microsoft Corporation, 1996). This much-publicized decision came as a shock to industries where hiring of temp workers and independent workers had become increasingly common. The Microsoft decision was quickly followed by a series of similar lawsuits, one of which included a $5.5 million

settlement by Time Warner to the U.S. Department of Labor (DOL) in 2000 for denying coverage of retirement and health benefit plans to misclassified independent contractors and temporary employees.6

Almost 20 years later, Fedex settled a misclassification case for $228 million for having misclassified 2,300 California drivers as independent contractors in 2015, and settled another similar case for $227 million across 19 states in 2017. A series of lawsuits, more recently including "gig workers" (e.g., Uber, Lyft, Instacart, and Shyp), led up to more concrete measures from the IRS and

DOL. During the Obama administration, in 2011, the IRS and the DOL signed a memorandum of

s Greenhouse, Steven. 2000. "TECHNOLOGY; Temp Workers at Microsoft Win Lawsuit." The New York Times, December 13. Retrieved January 1, 2018 (http://www.nytimes.com/2000/12/13/business/technology- temp-workers-at-microsoft-win-lawsuit.html); Aquent. 2009. "Managing Co-Employment Risk When Using a Staffing Agency." The Aquent Blog, February 7. Retrieved January 1, 2018 (https://aquent.com/blog/managing- co-employment-risk-when-using-a-staffing-agency); FindLaw. "Vizcaino v. Microsoft Raises the Stakes on Worker Classification; Other Major Changes Give Guidance to Employers." Retrieved January 1, 2018 (https://corporate.findlaw.com/human-resources/vizcaino-v-microsoft-raises-the-stakes-on-worker- classification.html). 6 Tarnoff, Steve, ed. 2000. "Time Warner Will Pay $5.5 Million to Settle DOL Lawsuit." HR Hub.com, November 20. Retrieved January 1, 2018 (https://www.hrhub.com/doc/time-warner-will-pay-55-million-to- settle-dol-0001). ' Wood, Robert W. 2015. "FedEx Settles Independent Contractor Mislabeling Case For $228 Million." Forbes, June 16. Retrieved July 1, 2017 (https://www.forbes.com/sites/robertwood/2015/06/16/fedex-settles-driver- mislabeling-case-for-228-million/#42c9b6fI c22e); Norup, Kimball. 2017. "Another FedEx Worker Misclassification Case Settled for $227 million." Talentwave Blog, May 9. Retrieved July 1, 2017 (http://www.talentwave.com/fedex-worker-misclassification-case-settled-for-227-million/).

17 understanding (MOU) in a joint effort to increase worker misclassification audits. Further, the DOL began signing MOUs with individual states to conduct joint investigations and to enhance enforcement

(see Table 4 for a list of federal and state initiatives to combat worker misclassification). While a

2018 report from the Treasury Inspector General for Tax Administration has found that the IRS has not effectively implemented the 2011 MOU to date (Treasury Inspector General for Tax

Administration 2018), the effort to address the misclassification issue is still on-going.

[Table 2 about Here]

Then in 2015 and 2016, David Weil, the DOL's Wage and Hour Division administrator at the time, issued Administrator's Interpretations concerning both independent contractors and joint employment (for relationships involving staffing agencies). Occupational Safety and Health

Administration (OSHA) at the DOL in 2015 also started an initiative to protect temporary workers, by recommending that both staffing firms and the lead firms assume joint responsibility for ensuring worker safety and health.

In response, by the mid to late 1990s, firms began adopting strategies to avoid potential lawsuits and joint-employer responsibilities. Lead firms quickly turned to working more closely with staffing agencies to limit their liability in the language of the contract and stating explicitly that the agency is solely responsible for workers' compensation claims. A more sophisticated strategy, however, involved actively distancing itself from the workers by working with other types of labor market intermediaries that came to exist around the same time (elaborated further below).

1.2.3 Centralization and Outsourcing of Nonstandard Workforce Management

By the mid-1990s, firms began recognizing both an increasing need to better manage regulatory risks, as well as a need to better manage their nonstandard workforce in general. As the nonstandard labor market grew, the simple traditional triadic model (Figure 1) was considered

inefficient for many of the largest firms as it became nearly impossible to track all the nonstandard

18 workers across their business units coming from hundreds of different staffing firms. Individual business unit hiring managers usually had personal, long-term relationships with different suppliers

(staffing agencies and vendors) that ranged from large, general agencies to smaller, industry-specific agencies. When vacancies for temporary staff opened up, hiring managers typically tapped into one or two familiar staffing firms of their choice; this was not only costly, but also amplified regulatory concerns at the time of increasing lawsuits since it was extremely difficult for the firm to track who was being on- and off-boarded across different functions of their business. At the same time, an industry report notes that lead firms, without the transparency of how much the staffing firms pay the workers, started to become increasingly dissatisfied with the procurement model in which they relied on a limited pool of staffing agencies (SAP Fieldglass 2016). Lead firms acknowledged that staffing agencies, consistent with predictions in the brokerage literature, were becoming profitable due to their ability to match parties from both sides and to leverage their relationships with client firms and workers they supply.

One popular solution to address these concerns was to create a centralized system of

Contingent Workforce Management (CWM) as part of the firm's human resources or procurement department program. The idea was to create a single streamlined process for the entire firm, including its establishments, so that the firm could have better visibility into its nonstandard labor force instead of relying on the system that manages regular workers in the human resources department. In the process, lead firms increasingly pressured staffing agencies by creating preferred vendor lists and controlling the agencies' price-setting (SAP Fieldglass 2016). Staffing agencies were concurrently exposed to the rise in new online job boards (e.g., Monster.com) which reduced their competitive advantage of matching workers to firms. The intensifying political struggles involving competing interests among client firms, staffing agencies, and online intermediaries led to the rise in a new alternative intermediary referred to as the Managed Service Provider (MSP), a conglomerate of suppliers, which emerged to become one of the most prominent influencers in the market.

19 Many firms turned to working with the MSPs as they were successfully introduced to the market as "experts in managing the complexities of the entire staffing process, not just filling job requisitions" (SAP Fieldglass 2016:4). Firms who saw the need for a centralized system basically outsourced their Contingent Workforce Management program to the MSPs so that these MSPs would manage temporary staffing programs of a client firm. An MSP would oversee and manage the client firm's contingent workforce program, and the MSP's typical responsibilities, according to Staffing

Industry Analysts-one of the largest staffing industry-based research groups-would include "overall program management, reporting and tracking, supplier selection and management, order distribution and often consolidated billing" (Staffing Industry Analysts 2017a:25). They are typically found working with large Fortune 1000 companies, such as Apple, Coca Cola or Citibank, who have many establishments and units that hire nonstandard workers around the world and thus have a need to manage nonstandard workers effectively.'

An MSP then becomes what I term a secondary intermediary between hundreds, or even thousands, of staffing agencies (i.e. primary intermediaries) and the lead firm in need of nonstandard workers. An MSP in this case is portrayed to be a secondary intermediary since they add a new layer to the existing triadic employment relationship model described earlier. As Figure 2 illustrates, the

MSP acts as the middleman between the primary intermediaries (staffing agencies and vendors) and the lead firm, which means that the firm and the worker doing work assignments for the firm is now a step further removed. This creates a polyadic employment structure: there are now two intermediaries that perform traditionally HR functions with the ability to determine how marginal productivity gains generated by the worker are distributed between the worker and the intermediaries.

[Figure 2 about Here]

8 OnContracting. 2014. "What is MSP or VMS in Staffing? August 11. Retrieved March 1, 2017 (https://www.oncontracting.com/article/what-is-msp-vms-in-temp-staffing.html).

20 There are several advantages to the MSP model identified by the industry practitioners (for example, see Laurano 2013). Not only does it implement a centralized firm-level system that is both financially and functionally visible across its business units or establishments, but an MSP also reduces the administrative burden related to the recruiting and managing of nonstandard workers

(selecting, hiring, on- and off- boarding, and payrolling). Importantly, MSPs are seen to have legal expertise and management experience in handling employment-related legal requirements. An

Aberdeen Group research found that in 2013, 67 percent of 192 firms that use an MSP listed mitigation of risks ("internal review process for compliance to federal labor and tax policies") as an advantage of working with them (Laurano 2013). The MSPs are also found beneficial for proper on- and off- boarding (71 percent), visibility into spending (47 percent), clearly defined goals (42 percent), and the ability to forecast use of contingent labor (36 percent).

This new industry structure, as a byproduct of political processes between hiring firms and intermediaries in the context of a changing regulatory environment, helps large firms move toward holistic talent management for a contingent workforce. It however adds complexity to the traditional models used to understand employment relationships, with firms actively distancing themselves from the workers. Employment-related legal responsibilities are now dispersed among even more stakeholders, complicating legal interpretations of employers that can potentially be categorized as joint by the court. It also complicates the wage structure as there are now more actors involved in the recruitment and the management of nonstandard workers, each with some power to set or influence bill rates and pay rates (wages) for workers. The use of MSPs however continues to grow: Staffing

Industry Analysts (2017b) found that 63 percent of 145 large firms surveyed used an MSP in 2017, up from 42 percent in 2009.

1.2.4 Technological Advances in the 2000s and Legitimatization of Competitive Bidding

While the MSPs were able to grow by identifying and addressing the needs of the lead firms, the use of an MSP also meant that firms had to rely solely on a single vendor for sourcing of

21 nonstandard workers with the assumption that the MSP would provide the best candidates at competitive rates. Yet, due to the structure in which the MSP was initially founded, that is, a conglomeration (or a division') of staffing agencies, staffing agencies affiliated with MSPs were often favored over other competing agencies. Since the initial founding of the model, the MSP decided to further subcontract out their supplier function, a traditional domain of staffing agencies, and focus on the core competencies of managing the relationships between the lead firm and the staffing agencies and take what is termed as the "vendor neutral approach." Despite nearly a decade of an industry- wide push towards an integration of a perfectly vendor neutral model, however, some MSPs were still directly affiliated with one or more staffing agencies or even owned by a staffing agency. In such a case, an MSP is considered the "master supplier" that supplies its own pool of qualified candidates or a "hybrid MSP" that engages in both the supply of candidates as well as the management of other staffing agencies as an outsourced agency (Staffing Industry Analysts 2014; 2016). This gave MSPs substantial bargaining power over the lead firms because such a structure often naturally leads to

MSPs favoring their own in-house staffing services by charging higher bill rates and passing undesirable assignments to secondary suppliers.

At this time, technological advances began influencing the industry, giving firms an alternative to working with an MSP. Vendors that provided a cloud-based technology platform, referred to as the Vendor Management Systems (VMS), entered the market as a technology intermediary, specifically designed to manage complexities of the contingent workforce. The VMS allowed the lead firms to manage their contingent workforce management in-house instead of relying on the MSP to centralize the process for them. Since all the worker requisition requests as well as onboarding and off-boarding of workers are made to go through this technology platform, a VMS reduces contingent labor costs by simplifying the payment system, while increasing visibility of labor

9 OnContracting. 2014. "What is MSP or VMS in Staffing? August 11. Retrieved March 1, 2017 (https://www.oncontracting.com/article/what-is-msp-vms-in-temp-staffing.html).

22 spend to inherently provide rich data for analytics and forecasting of labor demand. Some of the most popular VMS providers include SAP Fieldglass, Beeline, IQNavigator, Peoplefluent, and PRO

Unlimited, among others. They are most popularly known to have instituted a "competitive bidding" system where hundreds or thousands of staffing agencies can enter the cloud to bid their candidates for a vacancy of the lead firm.

This rhetoric around vendor neutrality has since dominated the industry and provided a normative framework from which this competitive bidding system gained legitimacy. At the core, it exposes workers to a bidding system that resembles those in product supply chains (Hahn, Kim, and

Kim 1986) and online marketplaces (Wise and Morrison 2000) which are found to produce an adversarial environment. Yet, the normalization of negative rhetoric around "vendor bias" and therefore an argument for instituting the more "fair" or "neutral" competitive bidding system prevailed in the industry with the growing dissatisfaction from the lead firms. The term vendor neutrality appears in most industry reports and articles that advertise VMS integration and is marketed as a huge benefit to all the parties involved.

Most MSPs also now use the VMS, and this integrated MSPNMS model also allows the MSP to adopt competitive bidding processes and separate out any vendor bias. A perfect vendor neutrality exists when neither MSP nor the VMS provides staffing services and are financially independent from any staffing providers they work with. Moreover, a perfect market-based competition may exist when lead firms decide to internally manage their CWM by purchasing VMS on their own and when the

VMS does not favor any staffing agencies. Today, while most firms still work with the MSPs and their VMS integrated model, many are also increasingly turning to the Internally Managed Program

(IMP) model, where they work directly with the VMS (Hiretalent 2016; SpendMatters 2017).

Combined, Staffing Industry Analysts (2017b) have found that 73 percent of the firms they surveyed used a VMS in 2017 from just about 16 percent in 2006. While it may seem as though a movement towards the use of VMS is solely out of efficiency concerns to maximize profit, it was in fact achieved

23 through constant power struggles between organizations in this market. As an industry report from a

VMS vendor states, "Right or wrong, vendor neutrality is a concept essentially born from the mistrust of staffing suppliers" (SAP Fieldglass 2016).

And while the holistic approach to contingent workforce management seems ideal, it is rarely achieved in reality. Even when a business unit and its HR or procurement manager uses the

MSP/VMS integration to streamline their entire contingent workforce, other business units within their umbrella organization may not sign up for it and continue to contract directly with their favoring supplier. This is especially common in some of the largest multinational firms with hundreds of establishments around the world-despite their ongoing efforts to centralize the system with technological tools made available today, the management of contingent workforce is often left fragmented, if not unmanaged.

Further adding to this complexity, there have been two major developments with the growth of

SOW-based outsourcing. First, the contract firms are increasingly supplying workers through temporary agencies themselves to create another layer of intermediaries (Ezratty 2010). Second, there is an increasing integration of SOW projects into the MSP/VMS model. Depending on the strategic choices made by the lead firms and the contract companies, there can be many different paths to establishing a work arrangement. In the most extreme yet not uncommon case, for example, a worker registered with a staffing firm would be onboarded by a contract firm who supplies labor for a SOW- based project, which goes through a competitive bidding process in the VMS system that an MSP for the lead firm uses (see Figure 3).

[Figure 3 about Here]

There are debates as to whether the integration of the MSP/VMS model to SOW projects is

efficient since the SOW project management is fundamentally different from temporary labor

24 onboarding into a standardized technology platform.10 Yet, such integration is widely used and is

promoted today for managing regulatory risks and for increasing visibility into the "hidden temps" that

are masked under the SOW projects to circumvent temporary worker headcount and restrictions of

tenure (Ezratty 2010). These developments reveal that the rhetoric around regulatory risk has become

an important part of the industry that dictates the development of new types of institutional

arrangements.

1.2.5 Emergence of the "Human Cloud" Framework

Most recently, there has been a rise in a new technological channel for engaging independent

contractors and freelancers in particular-the Freelancer Management Systems (FMS)-that is distinct

from the VMS. It was not until late 2013 and 2014 that this term started to appear among practitioners to describe a type of "work intermediation platforms"" through which an organization can engage in direct sourcing and managing of independent contractors and freelancers without the involvement of

staffing agencies. This pool of independent contractors and freelancers are considered a part of the

"human cloud", which Staffing Industry Analysts define as the following:

"An emerging set of work intermediationmodels that enable work arrangementsof various kinds to be establishedand completed (includingpayment of workers) entirely through a digital/onlineplatform. In many cases (though not always), the platform also supports "the enactment and management" of the work (to a lesser or greaterdegree). Job boards (like Monster) and social networks (like LinkedIn) do not fall within our definition of human cloud; while those two types of talent exchangeplatforms may support the sourcingand recruiting part of work arrangementsbeing established, such platforms do not further enable or support work arrangementsthrough to their completion (includingpayment of workers, taxfilings, etc.) " (Nelson 2015).

10 Vempati, Lalita. 2015. "Why Your Vendor Management System Can't Manage SOW-Based Projects." DCR Workforce Blog, September 10. Retrieved July 1, 2017 (https://blog.dcrworkforce.com/vendor- management-system-cant-manage-sow-based-projects). 1 A term coined by Spend Matters, a blog and social media website focusing on the topics of procurement and supply chain.

25 The FMS is fundamentally different from other online freelancer marketplaces, such as Upwork, Guru, freelancer.com, and PeoplePerHour, that became active in the late 2000s (Row 9 of Table 1). While online freelancer marketplaces allow individual exchange of transactions to facilitate a direct legal relationship, the FMS bears more resemblance to the VMS. The FMS offers a cloud-based technology platform to directly engage internal hiring managers in sourcing and managing of workers by privately sourcing them into the FMS system under strict qualification requirements (Karpie 2015). Upwork, for example, has launched its own FMS-Upwork Enterprise in September 2013 for large enterprises, and Upwork Pro in February 2014 for mid-sized businesses-to better serve organizations with large contingent labor needs (Karpie 2016). Hence, while the online marketplaces have previously only offered a platform on which workers and lead firms can find each other, the FMS engages in the function of matchmaking as well as of managing the labor they supply.

Figure 4 summarizes the new multifaceted model for alternative work arrangements today. In the last couple of years, firms or MSPs with VMS capabilities have begun to implement FMS to draw independent workers and freelancers (see Figure 4). For example, in February 2015, the VMS provider IQNavigator and the FMS provider Work Market formed an alliance to allow IQNavigator's clients to access Work Market's pool of independent workers and freelancers." In June 2016, the

MSP provider Randstad Sourceright acquired the FMS provider Twago in order to help MSP clients better integrate their freelance labor into their existing VMS." As firms are ultimately seeking to adopt a holistic view of contingent worker management, these integrative systems are expected to expedite the process for many of the firms that already depend on a large and increasing number of contingent workers. This development is in many ways in direct competition with online

12 IQNavigator. 2015. "IQNavigator and Work Market Form Strategic Alliance to Transform How Global 2000 Companies Source and Manage Contingent Labor." PRNewswire, February 3. Retrieved July 1, 2017 (https://www.prmewswire.com/news-releases/iqnavigator-and-work-market-form-strategic-alliance-to-transform- how-global-2000-companies-source-and-manage-contingent-labor-300029797.html). 13 Randstad Sourceright. 2016. "Randstad Sourceright Launches 'twago' Freelance Management Platform." PRNewswire, July 4. Retrieved July 1, 2017 (https://www.prnewswire.com/news-releases/randstad-sourceright- launches-twago-freelance-management-platform-585418501.html).

26 marketplaces: it came out of a deliberate effort by industry constituents to institute a system of management for independent agents that often exclude intermediaries in the space of online freelancer marketplaces. Pre-existing institutions are seeking to branch out and expand their business scope in areas where they see fit, even further augmenting the inter-organizational complexity governing alternative work arrangements.

[Figure 5 about Here]

1.3 A Case Study of Multifaceted Employment Relationship

While there have been some efforts to better understand the current climate of the new economy, one major challenge has been the lack of publicly available data that tracks changes in the composition of nonstandard workforce at the individual level, as well as at the establishment- or the firm-level, on the utilization of the nonstandard workforce. At the individual-level, the 1995 and 2005

Contingent Work Supplements (CWS) to the Current Population Survey (CPS) conducted by the U.S.

Bureau of Labor Statistics (BLS) have been the major source of data on the nonstandard workforce, and more recently, an updated version of the CWS by Katz and Krueger (2016) as part of the RAND

American Life Panel (ALP). Studies examining the establishment- or firm-level usage of the nonstandard workforce have used the National Employer Surveys (NES) by the Census Bureau in

1994, 1997, and 2000 (Cappelli and Keller 2013), the National Organizations Surveys (Houseman

2001; Kalleberg 2003), and the 1995 W.E. Upjohn survey establishments (Houseman 2001).

While these data sources have expanded our knowledge of nonstandard employment practices, they have two important limitations. First, because most of the data sources on nonstandard work come from surveys of individuals and managers at establishments, answers could vary widely depending on respondents' perceptions of the terms and categorizations used in survey questionnaires.

For example, individual workers may identify their work arrangements based on their own predetermined ideas about their tasks and work schedules. Meanwhile, HR managers answering

27 surveys about their establishments may not have-or even wish to disclose-information about the firm's overall usage of nonstandard labor. Second, these surveys are insufficient in capturing different modes of entry into the workforce by nonstandard workers. This is primarily because most workers are unaware of the details of their own path to work. For example, a temp worker who eventually ends up working for a business unit at a large firm may never find out whether his or her staffing agency had entered into a competitive bidding process through the MSP/VMS model, or the staffing agency had worked directly with an HR manager at the firm. In a similar vein, an HR manager may not be aware of how other HR managers or team leaders search for nonstandard labor when there is a need-as elaborated in detail earlier, most large firms do not have a consistent and fully-integrated labor onboarding system across their business units and even across teams within the business units.

Further, information on labor onboarding strategies for the contingent workforce is highly sensitive, which makes it even more difficult to rely only on survey data. Studies on the variations of nonstandard work arrangements and their effects are therefore extremely limited in scope.

Instead, I examine the nonstandard workforce and their individual pathways into work using unique and proprietary data from a North American company (henceforth, "ContingentCo"), a consulting firm which primarily collects contingent worker hiring data from 49 Fortune 500

companies that use "contingent worker management" programs in their human resources or procurement departments. To my knowledge, this is the first available data that provides detailed

information on how contingent workers are sourced and hired into lead firms' workforces. Further, the strength of the data lies in its rich coverage across different industries, occupations, and skill-

levels. Although it is unable to capture every detail of the multi-faceted employment relationships

described in earlier sections, this is the only source of data to date that provides at least a glimpse into

the new developments nonstandard workers have faced over the last decade. ContingentCo is able to

collect such information using its client firms' data within the Vendor Management System (VMS),

the technology platform described earlier, since all 49 firms make use of the VMS. However, the data

28 does not capture other variations within firms if there are other onboarding practices that occur outside of the firm's VMS system. This is because some firms do not require all business units or teams to use VMS when onboarding contingent workers-for example, some units within a firm may still continue to work with one long-trusted temporary help agency.

Figure 5 provides an overview of the process through which the job candidates in the

ContingentCo dataset enter the nonstandard labor market and are eventually onboarded to work for the lead firm. Those looking for employment opportunities enter the labor market first by working with the primary intermediaries, such as staffing agencies or vendors. These primary intermediaries are actively on the VMS platform to look for open requisitions that may fit any of the job candidates they work with. When there is a match, upon discussions with the potential candidates about the position and the terms of the contract, the suppliers submit the candidates for the position. If the requisition were first created by an "internal program management office" of a firm, the internal office would simply review the candidates that are submitted by multiple, or even hundreds, of suppliers in the

VMS, and make an onboarding decision based on the information available. If a successfully filled requisition were first created by an MSP that works as an outsourced agency for a firm, the candidates are then streamlined through multiple intermediaries without their knowledge. In this case, the MSP would review the candidates submitted in the VMS by the primary intermediaries, and may or may not approve the onboarding depending on the final approval from the lead firm. Here, the primary intermediaries (staffing agencies or vendors) identify and attract potential candidates, then submit qualifying or matching candidates for an open position on a technology platform intermediary (VMS), where the secondary intermediary (MSP) makes an initial attempt in selecting one or more final candidates. These final candidates go through remaining approval processes and actual onboarding by the lead firm.

[Figure 5 about Here]

29 During discussions about the exact details of the contract either before the candidate is submitted for the position or after the candidate is approved for onboarding, they also decide who becomes an official employer of the record. The "agency worker" would officially list the staffing agency that he or she initially worked with as an employer, while the "sub-vendor" would have vendors (or contract firms) as an employer. The "independent contractor" would have no official employer of record--even if the employment is found through the staffing agency who will be paid a fee for the service, both parties mutually agree that the staffing agency nor the lead firm will be liable for the worker's employment. As noted previously, the staffing agency under this scenario would still manage payroll for the contractor. The "employee" would have the most traditional employment relationship as he or she signs an official contract with the lead firm, even when the employment is found through intermediaries, but such a case is uncommon. Table 2 shows that for the sample of 49

Fortune 500 firms in the ContingentCo data, 18.4 percent of the firms use the internal office to manage contingent workforce, while 81.6 percent of the firms use the MSP model of outsourcing their CWM program to a third party. Since this data is based on information on the VMS platform, all of the firms use this technology platform intermediary to find candidates supplied by staffing companies or vendors. On this platform, there are 2,222 unique staffing firms or vendors from the period of 2000 to

2017 (the most reliable data comes from the years 2012 through 2017) that supplied about a million candidates, each of whom are categorized as an agency worker (97.9 percent), independent contractor

(0.67 percent), sub-vendor (1.14 percent), or employee (0.3 percent).

[Table 3 about Here]

The most common type of alternative work arrangement reflected in the data is an agency worker who is officially employed by a staffing firn that connects, through VMS, the worker and the firm that has outsourced its contingent worker management to the MSP (see Table 4). Out of the candidates classified as an agency worker (the largest type of contingent worker at 97.9 percent), 76.3 percent were submitted to assignments posted by an MSP while 23.7 percent were submitted to

30 assignments posted directly by the firm through its own internal program management office. In

addition, among those that were channeled through the MSP/VMS model, 90.1 percent were exposed to competitive bidding, in which all the candidates submitted by different staffing firms were theoretically given an equal chance regardless of any preexisting relationships between the staffing

firms and the MSP or the lead firm. Other notable relationships include the master supplier model where, as described earlier, the MSP or one staffing company becomes the primary supplier of staffing for the firm before the requisitions are distributed to secondary suppliers. The pre-identified workers are those that are identified by managers pre-competition, such that the VMS or MSP/VMS only serves to manage certain functions related to the assignment, like on- and off-boarding. The data shows that a large share of sub vendors in the MSP/VMS model are pre-identified-this is most likely because when vendors are involved in contract work assignments, the vendors and MSPs frequently have pre-existing relationships such that the MSPs would request a group of workers for a particular project. Most notably, it can also be seen that among agency workers that were channeled through the

InternalNMS model (that is, without outsourcing to an MSP), 92.8 percent were competitively bid candidates. The Internal/VMS model combined with competitive bidding is the least fragmented and the most competitive model out of all the models that employers can choose to adopt. Of the firms, then, who choose to internalize the staffing system, most prefer a market-based, competitive selection of candidates.

[Table 4 about Here]

The table also shows that only about 0.67 percent of candidates working with intermediaries contract to work as independent contractors (refer to Table 3). For both types of channels of onboarding, internal program management and the MSP model, 77.2 percent of independent contractors went through competitive bidding, and about 20 percent of independent contractors on average came through a pre-identified channel. A significant share of independent contractors coming through a pre-identified channel reflects the referral service described in the earlier section provided

31 by staffing agencies. Because staffing agencies typically "refer" these workers, many of them naturally go through the non-competitive channel where the host company manager (or the MSP) and the staffing company have already sorted out a mutual agreement about a certain set of contract workers. The lack of independent contractors in this data, however, is one important limitation in this data as they make up a significant share of nonstandard workforce today. This is because many independent contractors use online marketplaces or the FMS today-once firms start to integrate the

FMS practices into their MSP/VMS capabilities, it would increase the possibility of collecting more broadly the use of all different types of nonstandard workers.

This case study illustrates the multiple pathways from which nonstandard workers find work in 49 large Fortune 500 firms. While not much is known about why a firm may choose to adopt one strategy over the other for managing nonstandard workers, we can see that depending on firm strategy, an identical worker with same set of skills can get exposed to a quite different set of intermediating organizations and the system of recruitment without their own discretion.

1.4 Discussion and Future Research

Today, the industry for recruiting, selecting, and onboarding alternative work arrangements reveals a much more complex and diverse picture than what labor scholars have documented starting in the 1990s. The growth of the nonstandard workforce has facilitated changes in the institutional structures and the types and number of labor market intermediaries involved. The typical triadic exchange that involves a firm, a staffing agency, and a worker is no longer the dominant form of the

"new" employment relationship. Rather, we are moving towards a multifaceted employment model with a wide range of exchanges, from a complex, polyadic employment relationships involving more than three stakeholders, to a legally non-existent employment relationship between the firm and the independent contractors. This paper fills an important gap in our understanding of these multifaceted employment relationships by documenting the changes in the labor market for nonstandard workers. I show that the developments in the labor market structure were facilitated by the changes in the legal

32 environment, the rhetoric around competition, and the technological advances. In the process, organizations with competing interests struggled to constantly weigh options and make strategic choices about the most fitting management system for their growing nonstandard workforce. I argue that both market-based perspectives and the need to address legal and political concerns influenced these decisions.

The importance of shifting the focus to the role of institutions and organizational decision- making processes in adopting different contingent worker management programs is amplified by its potential effect on worker outcomes. While there is practically no research to date that documents the constantly evolving picture of the industry or that seeks to understand implications of such change on worker outcomes, there is some evidence that short-term employment relationships influence the distribution of rewards within organizations and across different stakeholders (for a review, see

Bidwell et al. 2013). With the growing fragmentation of the labor market for nonstandard workers, the differences in the onboarding processes and the organizational structure that allows for such variations are expected to affect employment outcomes for workers. By studying how onboarding processes occur to create inter-organizational complexity, the varieties of employment relationship structure can lead to very different outcomes, even for workers in the same type of alternative work arrangements. Future research should clearly identify new developments in the fissured labor market today and study how they affect workers as well as employers and other intermediaries involved.

I will begin to take up a subset of these issues in the next essay. There is a lack of clear consensus on whether the rise in alternative work arrangements has benefited or hurt employees and there is very little empirical research on how it may have systematically affected the distribution of wealth between the stakeholders, mainly due to the lack of available data. Using the ContingentCo data, I will specifically test the hypothesis that, holding other factors constant, the number and type of intermediaries involved in the on-boarding process (the rise in institutional complexity) not only affect

33 wages for a worker onboarded through an alternative work arrangement, but also influence economic returns to both the lead firm and the staffing suppliers.

34 1.5 References

Autor, David H. 2003. "Outsourcing at Will: The Contribution of Unjust Dismissal Doctrine to the Growth of Employment Outsourcing." Journal ofLabor Economics 21(1):1-42.

Barley, Stephen R. and Gideon Kunda. 2004. Gurus, Hired Guns, and Warm Bodies: ItinerantExperts in a Knowledge Economy. Princeton, NJ: Princeton University Press.

Bernhardt, Annette. 2014. Labor Standards and the Reorganizationof Work: Gaps in Data and Research. 100-14. IRLE Working Paper.

Bernhardt, Annette, Rosemary Batt, Susan Houseman, and Eileen Appelbaum. 2016. "Domestic Outsourcing in the U.S.: A Research Agenda to Assess Trends and Effects on Job Quality."

Bidwell, Matthew, Forrest Briscoe, Isabel Fernandez-Mateo, and Adina Sterling. 2013. "The Employment Relationship and Inequality: How and Why Changes in Employment Practices Are Reshaping Rewards in Organizations." The Academy ofManagement Annals 7(1):61-121.

Bidwell, Matthew and Isabel Fernandez-Mateo. 2010. "Relationship Duration and Returns to Brokerage in the Staffing Sector." OrganizationScience 21(6):1141-58.

Bonet, Rocio, Peter Cappelli, and Monika Hamori. 2013. "Labor Market Intermediaries and the New Paradigm for Human Resources." The Academy ofManagement Annals 7(1):341-92.

Cappelli, Peter H. and JR Keller. 2013. "A Study of the Extent and Potential Causes of Alternative Employment Arrangements." ILR Review 66(4).

Cepin, Geremy C. 2012. "Choose Wisely When Hiring a Search Firm to Recruit for Key Positions." CPA Prac. Mgmt. F. 8:8.

Doeringer, Peter and Michael J. Piore. 1971. InternalLabor Markets and Manpower Analysis. Lexington, Mass.: Heath and Company.

Evans, James, Gideon Kunda, and Stephen R. Barley. 2004. "Beach Time, Bridge Time, and Billable Hours: The Temporal Structure of Technical Contracting Author(s): James A. Evans, Gideon Kunda, Stephen R. Barley Reviewed Work(S):" Administrative Science Quarterly 49(1):1-38.

Ezratty, Jason. 2010. "The Other SOW." Contingent Workforce Strategies 34-37.

Fernandez-Mateo, Isabel. 2007. "Who Pays the Price of Brokerage? Transferring Constraint through Price Setting in the Staffing Sector." American SociologicalReview 72(2):291-317.

Femandez-Mateo, Isabel and Zella King. 2011. "Anticipatory Sorting and Gender Segregation in Temporary Employment." Management Science 57(6):989-1008.

Hahn, Chan K., Kyoo H. Kim, and Jong S. Kim. 1986. "Costs of Competition: Implications for Purchasing Strategy." JournalofPurchasing and MaterialsManagement 22(3):2-7.

Hatton, Erin. 2011. The Temp Economy: From Kelly Girls to in PostwarAmerica. Philadelphia: Temple University Press.

Houseman, Susan N. 2001. "Why Employers Use Flexible Staffing Arrangements: Evidence from an Establishment Survey." ILR Review 55(1).

35 Houseman, Susan N. and Carolyn J. Heinrich. 2015. Temporary Help Employment in and Recovery. W.E. Upjohn Institute.

Kalleberg, Arne L. 2003. "Flexible Firms and Labor Market Segmentation: Effects of Workplace Restructuring on Jobs and Workers." Work and Occupations 30(2):154-75.

Katz, Lawrence and Alan Krueger. 2016. The Rise and Nature ofAlternative Work Arrangements in the United States, 1995-2015. w22667. Cambridge, MA: National Bureau of Economic Research.

Khurana, Anil. 2002. "Professional Growth of IT Industry Role of Cost and Benefits Technology Usage and Personnel Dimensions."

Kilcoyne, Patrick. 2005. "Occupations in the Temporary Help Services Industry." in Occupational Employment and Wages. Washington, DC: US Department of Labor.

King, Zella, Simon Burke, and Jim Pemberton. 2005. "The 'bounded': An Empirical Study of Human Capital, Career Mobility and Employment Outcomes in a Mediated Labour Market." Human Relations 58(8):981-1007.

Laurano, Madeline. 2013. Contingent Labor Management: Strategies and Solutionsfor a Flexible Workforce. Aberdeen Group.

Luo, Tian, Amar Mann, and Richard Holden. 2010. "During the 1990-2008 Period, Employment in the Temporary Help Services Industry Grew from 1.1 Million to 2.3 Million and Came to Include a Larger Share of Workers than before in Higher Skill Occupations; Employment in This Industry Has Been Very Volatile Because Temporary Workers Are Easily Hired When Demand Increases and Laid off When It Decreases." 14.

Marsden, Peter V. 1982. "Brokerage Behavior in Restricted Exchange Networks." Social Structure and Network Analysis 7(4):341-410.

McLaughlin, Diane K. and Alisha J. Coleman-Jensen. 2008. "Nonstandard Employment in the Nonmetropolitan United States." Rural Sociology 73(4):631-59.

Osterman, Paul. 2011. "Institutional Labor Economics, the New Personnel Economics, and Internal Labor Markets: A Reconsideration." ILR Review 64(4):637-53.

Osterman, Paul and Diane M. Burton. 2004. "Ports and Ladders: The Nature and Relevance of Internal Labor Markets in a Changing World." in The Oxford Handbook of Work and Organization, edited by S. Akroyd, R. Batt, P. Thompson, and P. S. Tolbert.

Peck, J. and N. Theodore. 2006. "Flexible Recession: The Temporary Staffing Industry and Mediated Work in the United States." Cambridge JournalofEconomics 31(2):171-92.

Peck, Jamie, Nik Theodore, and Kevin Ward. 2005. "Constructing Markets for Temporary Labour: Employment Liberalization and the Internationalization of the Staffing Industry." Global Networks 5(1):3-26.

SAP Fieldglass. 2016. Vendor Accountabilityfor Contingent Workforce Management. SAP Fieldglass Whitepaper. SAP Fieldglass.

Segal, Lewis M. and Daniel G. Sullivan. 1997. "The Growth of Temporary Services Work." Journal ofEconomic Perspectives 11(2):117-36.

36 Smith, Rebecca and Claire McKenna. 2014. Temped Out: How the Domestic Outsourcing of Blue- CollarJobs Harms America's Workers. National Employment Law Project.

Smith, Vicki and Esther B. Neuwirth. 2008. The Good Temp. Ithaca, NY: Cornell University Press/ILR Press.

Staffing Industry Analysts. 2017a. The Global Lanugage of the Workforce Solutions Ecosystems. Crain Communications Inc.

Staffing Industry Analysts. 2017b. Workforce Solutions Buyers Survey 2017 - Contingent Workforce ProgramFunding Models and Rates.

Treasury Inspector General for Tax Administration. 2018. Additional Actions Are Needed to Make the Worker MisclassificationInitiative With the Department of Labor a Success. 2018-IE-R002.

Weil, David. 2014. The Fissured Workplace. Harvard University Press.

Wise, Richard and David Morrison. 2000. "Beyond the Exchange-the Future of B2B." Harvard Business Review 78(6):86-96.

37 1.6 Tables and Figures

Table 1. Development of New Labor Market Intermediaries Managing Alternative Work

Arrangements

Worker Legal Employer LMIs LMIs (technology platform) Type Direct-hire, Lead firm n/a n/a n/a n/a ICs Staffing Classic Models Staffing agency n/a n/a n/a Agency agency temps Contract firm Contract firm n/a n/a n/a Contractor Online Lead firm n/a n/a job n/a All board 1990s (mid- MSP* MSP n/a Online job n/a All late) board Staffing Online job Staffing agency MSP n/a All agency board Staffing Staffing agency n/a VMS** n/a agency All Staffing 2000s Staffing agency MSP VMS n/a All agency n/a (self- Online n/a employed) n/a freelancer n/a ICs marketplace*** (self- 2010s (early) n/a n/a n/a employed) FMS**** n/a ICs

2010s (mid- empl- n/a n/a VMS FMS ICs late)emlyd n/a (self- Staffing MSP VMS FMS All employed) agency

Notes. *MSP: Managed Service Provider; **Vendor Management System; ***Online freelancer marketplace refers to websites including Upwork, Guru, Freelancer, PeoplePerHour; ****Freelancer Management System

38 Table 2. Summary of Federal and State Initiatives for Misclassified Independent Contractors

Year Federal and State Initiatives for Misclassified Independent Contractors

2012 Memorandum of Understanding with 12 states, IRS, DOL

2012 federal budget includes $23M for "Misclassification Initiative"

California Senate Bill 459 goes into effect

Solicitor of Labor indicates that misclassified labor is one of agency's top priorities

DOL FLSA suits reach all-time high in 2012

2013 Iowa and New York signs MOU

IRS re-starts Questionable Employment Tax Practices program

Affordable Care Act goes into effect

Payroll Fraud Prevention Act of 2013 introduced in U.S. Senate

2014 Alabama, Wyoming, New Hampshire, Massachusetts, and Wisconsin signs MOU

Payroll Fraud Prevention Act of 2014 introduced in Congress NY JETF Audit of 12,000 Audits 133K Misclassified Workers

Executive order adds compliance requirements for federal contractors California AB-1897 passed

2015 Florida, Rhode Island, Idaho, Alaska, Vermont, Hawaii, Arkansas signs MOU

2016 Fed Budget focuses on IC misclassification

CA finds Uber driver to be employee

DOL Directive Dr. Weil Administrator's Interpretation No. 2015-1

Order 2016 New York expands existing task force on employee misclassification (Executive No.1590)

2018 New Jersey establishes a task force on employee misclassification (Executive Order No.25)

39 Figure 1. Traditional Business Model for Nonstandard Workforce

Firm Intermediary Worker Intermediary Firm

@0 @000 Staffing Agency )

Staffing Agency 3 00 @0

I NI Temporary Worker/ Business Units SOW Contractors

Figure 2. New Business Model for Alternative Work Arrangements in mid-90s and 2000s

Worker intermediary (Primary) Technology Intermediary Firm Platform (Secondary) S@ 0 Staffing Agency 1 under "Vendor model mMSPunderMSP e oNeutral"

DrcSVendooru

Vendor 2 TempWor ker/ Business Units SOW Contractors I d (0TI MSP under "Master Supplier" model or "Hybrid" modelI IV V Direct Source Contractors

Figure 3. New Business Model for Alternative Work Arrangements: an Extreme Case

Worker Tcnlg nemdayFr SOW Contractors Intermediaries (Primary) Pltfcnoy (nSerondaryFr Platform (Secodary)

i(ContBas Urnit

Business Units

40 Figure 4. New Mutifaceted Model for Alternative Work Arrangements Today

Worker Intermediary (Primary) Technology Int ermedlary Firm Platform (Se condary) Staffing Agency I

Staffing Agency 2 Staffing Agency 3 --- I Vendor 2 Temporary Worker/ Business Units )W Contractors S I /7

Direct Source Contractors "Human Cloud" (Freelancers and Independent Contractors)

Figure 5. Path to Work for Alternative Work Arrangements: Case Example from the ContingentCo data

Worker Intermediary Technology Intermediary Firm (Primary) Platform (Secondary) 0. 0.Internal Program 8989 Management Office

0 0Staf fing Agency 2

Staffing Agency 3

Agency worker - Independent contractor Sub-vendor Employee

41 Table 3. Descriptives of Alternative Work Arrangements in the ContingentCo data (-2012 to 2017)

(number) (%) Firms 49 InternalProgram Management Office 9 18.4 Outsourcingto Managed Service Provider(MSP) 40 81.6

Job Seekers 955,169 Agency Worker 934,966 97.88 Independent Contractor 6,412 0.67 Sub Vendor 10,928 1.14 Employee 2,863 0.3 Unique Suppliers (Staffing agencies or vendors) 2,222 Source: Author's analysis of ContingentCo data Notes: All 49 firms (including 40 MSPs) in this sample use VMS, the technology platform intermediary

42 Table 4. Descriptives of Alternative Work Arrangements in the ContingentCo data, by Source Type

(-2012 to 2017)

Candidates (n=955,169) Firm Source Type Agency Independent Sub Worker Contractor Vendor Employee (Total) 23.67 63.02 61.92 n.a. Competitive Bid 92.79 47.59 77.21 n.a. InternalProgram Management Office Master Supplier 0.65 n.a. 0.35 n.a. Pre-identified 6.14 52.41 22.43 n.a. Single Source 0.21 n.a. n.a. n.a. (Total) 76.33 36.98 38.08 100 Competitive Bid 90.14 72.71 30.09 n.a. Master Supplier 1.91 0.08 0.17 n.a. Outsourcingto Managed Service Pre-identified 6.75 16.2 68.97 n.a. Provider(MSP) Single Source 0.82 4.98 0.17 n.a. Direct Sourcing 0.3 0.8 0.6 n.a. SOW 0.02 5.23 n.a. n.a. Employee n.a. n.a. n.a. 100 Source: Author's analysisof ContingentCodata Notes: All 49firms (including40 MSPs) in this sample use VMS, the technology platform intermediary

43 Chapter 2. Multi-layered Labor Contracting and Distribution of Power: Evidence from Employment Records for Nonstandard Work

Abstract This article examines an important yet little-understood phenomenon governing alternative work arrangements: the rise in the "multi-layered labor contracting" structure in which the recruitment of nonstandard workers is outsourced to an intermediating organization who then selects qualified workers from a group of competing suppliers. I make the very first scholarly attempt to examine the link between such multi-layered contracting arrangements and subsequent economic outcomes for both the hiring lead firms and the workers. Using power-dependence theory, I argue that the lead firm's deliberate attempt to alter the recruitment process by involving multiple intermediating organizations compresses supplier power, which then incentivizes suppliers to transfer the competitive price burden to workers. Using proprietary data from employment records of about a million workers seeking nonstandard work at 49 large firms, I find that an additional contracting layer between the lead firm and the worker is associated with higher returns to the firms and lower returns to the workers. When workers gain bargaining power, however, through a pre-existing firm-worker relationship, the results show that the loss from an additional contracting layer is significantly reduced. The results hold even when controlling for supplier fixed effects to control for unobservable supplier characteristics, as well as when controlling for detailed measurement of skill requirements for nonstandard jobs that may instead dictate the price-setting process.

44 2.1 Introduction

Multi-layered labor contracting-that is, a firm-worker relationship that extends beyond a formal and dyadic exchange with multiple intermediating organizations-has become noticeably common. Substantial scholarship has documented a rapid increase in employment arrangements alternative to , such as temporary, contract, or subcontracted work (Katz and

Krueger 2016), and the resulting impact on workers in various forms (for a review, see Bernhardt et al.

2016). Much of this research assumes that the recruitment processes enabling these new relationships are triadic, with third-party organizations such as staffing agencies, headhunters, contractors, or even online platforms brokering transactions between the firm and the worker (for a review, see Bidwell et al. 2013; Bonet, Cappelli, and Hamori 2013). However, recruitment of nonstandard workers today instead frequently resembles what is referred to as "multitiered contracting" (Weil 2014:101) arrangements in production supply chains, in which work that has been contracted out is further subcontracted in the form of hyper-subcontracting. For nonstandard workers, this means that it sometimes takes multiple steps (or "layers") of contractual relationships, often without their knowledge, to be matched to a work assignment at the firm where they perform tasks.

In this study, I focus on one major but to date little understood or examined development in the industry that dictates recruitment of nonstandard workers-the rising outsourcing of a management function devoted to recruiting and managing of nonstandard workers, to a new labor market intermediary referred to as the Managed Service Provider (MSP). When the hiring "lead" firm uses an

MSP, an additional layer of labor contracting is created since the MSP, in place of the lead firm, selects from a group of competing staffing agencies or contractors their supply of qualified candidates in the sequential process of matching workers to firms. While the phenomenon has become widespread, virtually no research has closely examined this trend, the variation in firms' use of intermediating organizations, and its potential economic impact on the lead firm and/or the workers seeking employment in nonstandard jobs. Not only is there a lack of publicly available data that tracks

45 how firms vary in their hiring of nonstandard workers, but previous research on nonstandard work has mostly looked within a triadic (or dyadic) arrangement, whether it be for a single- firm, industry, or

skill level job. However, detailed case-based narratives in these studies occasionally allude to more than one pathway through which nontraditional workers and firms find each other, even for the same type of the job. Evans, Kunda, and Barley (2004), for example, identifies two markets for technical

contractors, one in which they find work by negotiating directly with the lead firms as independent

agents, the other in which they have staffing agencies find work for them. Because these agencies as job-matching brokers have some discretion in the price setting process (see Fernandez-Mateo 2007),

the prices agencies charge the firm and pay the workers will fundamentally differ from the prices

negotiated between the firm and the workers directly; or from the prices negotiated between three or

more actors in a contractual relationship.

The question that arises is then: does an additional contracting layer between the buyer (i.e.,

the lead firm) and the seller (i.e., the worker) alter the price setting behavior of the labor suppliers (i.e.,

the staffing agencies or the vendors)? Drawing on power-dependency theory, I develop a theoretical

framework for predicting how an additional labor market intermediary (LMI) dictating prices in a

labor contract may alter the distribution of power among different actors, to influence the supplier's

ability to control prices charged to the lead firm and prices paid to the worker. An additional LMI can

be thought to reduce the buyer power since the buyer's growing distance from the seller limits its

access to information about the seller (see e.g., Gould and Fernandez 1989; Marsden 1982 for relevant

literature on brokerage). However, even when the distance between the buyer and the seller increases,

the buyer may gain power advantage if other available options develop in such a way to reduce the

buyer's dependence on intermediating organizations (see e.g., Emerson 1962; Ryall and Sorenson

2007; Simmel 1950 for relevant literature on power-dependence theory).

Then within an institutional context where the job-matching process of nonstandard workers is

not confined to one dominant arrangement with a limited number of LMIs, the distribution of

46 bargaining power should start to look different. Specifically, I argue that the lead firm's attempt to alter the recruitment process to their advantage by involving multiple ILMs compresses supplier power, which then incentivizes suppliers to reduce rent shared with workers. To test this argument, I begin by describing from my preliminary fieldwork the different roles each LMIs exhibit as well as their relationships with other actors in the nonstandard labor market. Then, I take advantage of a year- long negotiated access to proprietary data from the employment records of nonstandard workers in 49 large firms in the Fortune 500 companies list. The unique nature of its individual-level hiring and pricing data (for close to a million nonstandard workers) across firms that work with varying types and numbers of intermediating organizations allows me to provide, for the first time in the literature, an analysis of multi-layered contracting arrangements in nonstandard labor market. Using this data, my main findings suggest that having an additional layer of contract-multi-layered labor contracting- increases prices charged to the firm (bill rates) while decreasing the prices paid to the worker (pay rates).

Ultimately, this research helps extend our understanding of how inter-organizational processes influence price-setting behavior of organizations involved in the job-matching process for nonstandard workers. Scholars have called for more research on outsourcing-induced changes in intra- and inter- organizational processes that influence economic outcomes: "many of the most important consequences of outsourcing on workers are mediated through the effects of outsourcing on the nature of work and organizations" (Davis-Blake and Broschak 2009:322). This paper suggests that outsourcing of recruiting and managing a particular category of the workforce alters interorganizational processes between the lead firm and the labor suppliers, in a way that influences workers. Further, I contribute to the literature on nonstandard work by providing evidence for the need to look beyond triadic relationships in recruitment in order to better understand the price and the wage-setting processes in this constantly changing labor market.

2.2 Multiple Layers of Contracting and Power

47 Power advantage, according to power-dependence theory, comes from asymmetrical-or, different levels of--dependence between exchange partners in a relationship (Emerson 1962).

Different bodies of scholarship have conceptualized how distribution of power between actors in layers of contractual relationships may influence both firm profits and wage-setting processes in organizations. Literature on supply chains, for example, examines bilateral buyer-supplier relationships in the context of buyer (i.e., the lead firm) power and their impact on increased buyer profits (see e.g., Galbraith and Stiles 1983; Gosman and Kohlbeck 2009), even when suppliers achieve productivity gains from the relationship (Kim and Wemmerl6v 2015). Extending this work to wage- setting processes, Wilmers (2018) shows, although limited to a dyadic buyer-supplier relationship, that suppliers' worker wages are adversely affected when their dependence on large buyers increase (i.e., greater buyer power).

Similarly, when relationships are more complex, for example involving three or more parties along a supply chain, the relative degree of power among different actors are found to determine firm profitability (see e.g., Cool and Henderson 1998; Lanier, Wempe, and Zacharia 2010). While research on complex relationships in supply chains and their wage effects is thin, Weil (2014) contends that in such hyper-subcontracting arrangements, the lead firm can distance itself from a set of employer- related liabilities and reduce labor costs when it maintains more bargaining power over the contractors

(suppliers) and in turn, their subcontractors. The lower tier suppliers within such structure are exposed to greater competition and fluctuating demand, Weil argues, to ultimately influence worker outcomes.

Contrary to the research on supply chains, brokerage literature on triadic relationships in the context of recruitment portrays suppliers as having considerable bargaining power defacto as they provide market intermediation, or brokerage, connecting two or more parties to create or capture economic value. Brokers are considered to have better access to information and an ability to provide a match between actors that would otherwise not interact (see e.g., Autor 2008; Gould and Fernandez

1989; Marsden 1982; Osterman 2004). Research shows that staffing agencies as brokers gain

48 bargaining power to set prices by building relationships that are long in duration (Bidwell and

Fernandez-Mateo 2010) or more valued by the firms (Fernandez-Mateo 2007). These relationship- building mechanisms reduce the number of alternatives for other actors and increase their dependence on staffing agencies, which gives brokers more bargaining power (Ryall and Sorenson 2007; Simmel

1950).

However, in the context of the labor market for nonstandard workers elaborated below, I show that when the relationships between staffing agencies and the lead firms are mediated by an additional intermediating organization, staffing agencies' ability to build relationships is significantly reduced to change the ways in which they set prices for both the buyers and the sellers.

2.3 Background

The modem labor market for alternative work arrangements is no longer confined to a triadic exchange relationship where one staffing agency (broker) mediates relationships between the worker

(seller) and the lead firm (buyer). A triadic exchange relationship has been studied as the one dominant model in the staffing industry, where a staffing agency or a vendor supplies temporary labor for the "lead" firm (i.e., firms with temporary labor demand) to provide, at times, better information about the pool of candidates they source and to identify qualified candidates. The competitive advantage of the staffing agencies is that they work with both the buyer and the seller of labor for multiple assignments through repeated exchanges (Bidwell and Fernandez-Mateo 2010; King, Burke, and Pemberton 2005) and have incentives to provide qualified candidates to firms with their guarantee of staffing expertise (Cepin 2012; Khurana 2002).

Starting in the mid-1990s, however, firms began adopting new forms of exchange relationships with the rapidly growing use of alternative work arrangements, such as temporary help agency workers, on-call workers, contract workers, and independent contractors. These arrangements have accumulated to account for almost 16 percent of the entire labor force in 2015 according to Katz

49 and Krueger (2016). With this growth, the human resources and procurement professionals in many large firms started to recognize the need for centralized management of nonstandard labor.

Traditionally, individual business unit hiring managers made decentralized decisions to hire short-term workers, often based on personal relationships with a small number of suppliers, thereby making it difficult to track all individuals who were on- and off-boarded for short-term assignments (Figure 1 a).

Firms however became increasingly unconvinced that staffing firms were serving their best interests, with the absence of transparency in pricing and the growth in regulatory concerns related to co- employment and employee misclassification lawsuits. The popular sentiment in the industry was,

"We don't feel that the staffing firms are policing themselves, therefore we will need to do

something." 4

In addressing this challenge, large firms began adopting strategies referred to as the

Contingent Workforce Management (CWM) that diverged in one of two ways: (1) creation of CWM

in-house; and (2) outsourcing of CWM to a third-party management agency referred to as the

Managed Service Provider (MSP). In the first case, firms create a specialized HR/procurement

function to centralize the management of alternative work arrangements (Firm A in Figure lb). This

model fundamentally differs from a traditional, personalized triadic relationship (Figure 1 a) since it

reflects a deliberate attempt by large firms to organize and centralize the HR system for nontraditional workers in-house, typically referred to as the Internally Managed Program (IMP). Alternatively, firms

outsource that entire Contingent Workforce Management to a third-party Managed Service Provider

(MSP) (Firm B in Figure lb). In 2017, 63 percent of 145 large firms surveyed in an industry study

used an MSP, up from about 42 percent in 2009 (Staffing Industry Analysts 2017). The MSP adds an

additional layer to labor contracting as it brokers relationships between the lead firm and hundreds, or

" As quoted in "Vendor Accountability for Contingent Workforce Management." 2016 SAP Fieldglass Whitepaper.

50 even thousands, of staffing agencies and vendors who supply temporary labor.15 While some MSPs

provide its own staffing services, most subcontract out staffing services to smaller staffing agencies,

and instead focus on management functions, such as supplier selection, on- or off- boarding, contract

negotiation, time keeping, and compliance oversight. When such management functions are

outsourced, the MSP becomes the primary contact for all the suppliers, which reduces both

administrative and regulatory burdens on the lead firm.

[Figure 1 about Here]

Enabling such a large-scale management of alternative work arrangements, both the lead firms

(Firm A) and MSPs (working with Firm B) contract with technology vendors that provide a cloud- based platform (or Vendor Management System, VMS)16 to compare and select workers from a large number of suppliers in the system. An employer survey has reported that about 73 percent of firms surveyed had used a VMS in 2017, a rapid growth from just about 16 percent in 2006 (Staffing

Industry Analysts 2017). The platform has enabled an important feature in this nonstandard labor market that was previously non-existent: a "competitive bidding" system, where staffing agencies bid prices for their available workers for a given job requisition. The lead firms (in the case of Firm A) or the MSPs (in the case of Firm B) then select, from a number of comparable options, candidates for further hiring processes. In fact, the widespread adoption of this practice is considered by industry professionals to provide the MSPs or the lead firms a way to minimize "bias" in selecting favorable suppliers, which allows suppliers to compete for a contract on a level playing field.17

" They first emerged through an organized effort by smaller staffing firms that saw the need to leverage their resources better to their advantage. See: OnContracting. 2014. "What is MSP or VMS in Staffing? August 11. Retrieved March 1, 2017 (https://www.oncontracting.com/article/what-is-msp-vms-in-temp-staffing.html). 16 Commonly referred to as "Vendor Management System" (VMS), this technology platform is often an internet- enabled, cloud-based application software marketed and sold by vendors that specialize in technology solutions specifically for nontraditional work arrangements. Some of the most popular VMS providers include SAP Fieldglass, Beeline, IQNavigator, Peoplefluent, PRO Unlimited. 17 In a purely competitive bidding model, the MSP does not have an affiliation with any particular staffing agency nor supply labor on its own. Most MSPs that use this technology as well as independent technology vendors frequently use words such as "vendor neutrality" or "vendor accountability" to describe their products: "Our clients will tell you that the vendor-neutral approach consistently delivers the highest quality talent at 51 2.3.1 Buyer Power, Profitability, and Wage Determination

In the multi-layered contractual relationship intermediated by an MSP (Firm B in Figure 1b), staffing agencies have limited bargaining power for two main reasons. First, the use of a cloud-based technological platform as described earlier intensifies competition among suppliers through its competitive bidding process. When suppliers compete on a level playing field within the cloud, they are unable to extract and capture value that typically arise through the relationship building mechanisms. While competitive bidding is new in the nonstandard labor market, procurement managers have long made use of competitive bidding in product markets. Studies have commonly characterized competitive bidding as adversarial in product supply chains (Hahn, Kim, and Kim 1986) as well as in online marketplaces (Wise and Morrison 2000), since it encourages suppliers to focus on short-term costs instead of long-term capabilities. Further, even a positive aspect of competitive bidding like simplification in the supplier selection process can be to be associated with inefficient outcomes. In industries such as lean construction, contracts awarded based on a plan with complete

specifications at the time of competitive bidding are found to create inefficiencies since it is common for project designs and methodologies to change in the long-run (Elfving, Tommelein, and Ballard

2005).

Second, the suppliers of labor (staffing agencies or vendors) are one-step away from the buyer

(lead firm), as they are tied to the MSP that acts as a focal broker between the buyer and the suppliers.

This limits a supplier's ability to build direct exchange relationships with the buyer. Many contractual

arrangements between an MSP and a lead firm in fact forbid a direct communication between the firm

market-driven competitive rates" (Andrew Popler, quoted in Andrew Karpie and Nick Heinzmann, "Linking a Combined MSP and VMS with Pure Vendor Neutrality: Advantages and Lessons Learned" Spend Matters, April 5, 2017).

52 and the suppliers. 18 Suppliers are therefore forced to bid competitively for an assignment posted by the MSP and not the buyer.

In this setting, the buyer has many available options, which reduces dependence on other actors and increases its power, according to Emerson (1962). The buyer can either use other MSPs to find equivalently well-qualified suppliers or easily bring the service back in-house with the help of technology.' 9 Instead of depending on an MSP, firms can adopt the same technological platform in- house (Firm A in Figure Ib), which opens the door to an equivalently broad selection of suppliers and

2 their workers. ' The MSP therefore has both an incentive and a tool to find suppliers that can charge lower bill rates to the buyer than when the buyer manages nonstandard workers in-house. In sum, contrary to the assumption in the literature that a broker gains power advantage by building personal relationships with the buyer, I establish that the buyer gains bargaining power to set prices for the rest of the supply chain members when an MSP brokers relationship between the buyer and the supplier in this nonstandard labor market. In other words, an additionalcontracting layer between the buyer (i.e. the leadfirm) and the seller of labor (i.e. the worker) is associatedwith higher returns to the buyer as the buyer gains bargainingpower over others in the supply chain (Propositionla).

From the perspectives of the suppliers, the reduction of their bargaining power over the lead firm due to the MSP-tie amplifies the need to effectively control prices in ways that can still generate profit. A supplier typically profits by charging a bill rate to the firm for workers it supplies, minus the

18 Schmidt, Kyle, "9 Points of Debate Between Travel Nursing Companies over the MSP Model" BluePipes Blog, 2013. 19 In fact, despite the widespread externalization of nonstandard workforce management, the more recent debate has focused on the potential benefits of bringing the management back in-house, accompanied by the same technological platform (as in Firm B in Figure ib). See, for example, recent industry articles on this topic: Guest Contributor. "Why Are Clients Moving Away from MSPs to Internally Managed Programs (IMPs) to Optimize Contingent Labor?" SpendMatters, March 3, 2017; Reagan, Peter. "Going Internal? Consider all costs" Staffing Industry Analysts, November 18, 2015. 20 This internalization would similarly influence supplier competition with competitive bidding, but the mere fact that the buyer directly overseas, manages, and chooses suppliers in the system gives an opportunity for relationships and preferences to form in ways that can give suppliers some bargaining power less available in the MSP-mediated relationships.

53 pay rate that goes to the workers as an hourly compensation (supplier-profitA = billjrate - pay._rateA). For work assignments that an MSP manages for the buyer, the supplier profit gets reduced by the fee paid to the MSP for its services per transaction (supplier-profitB = bill-rateB - payrateB - MSP.- ee).2 1 However, since suppliers tied to an MSP are unable to charge higher bill rates to the buyer to offset the additional cost incurred, standard economics would predict that suppliers under perfect competition should be willing to take lower profits holding constant the pay distributed to workers of same quality and skills. It may be the case that there are unobservable benefits (e.g., large volume of job openings, sharing of risks) of working with the MSPs to compensate for the lower profits. Alternatively, the suppliers will out of the subcontracting arrangement and instead resort to working directly with the lead firms to generate profit, unless they can "transfer" the additional price burden from the MSP to the workers (Fernandez-Mateo 2007) as a broker that has bargaining power over workers. Femandez-Mateo (2007) has suggested that such transferring of burden is in fact possible when there is some stickiness to the relationship between suppliers and workers; that is, when workers are uncertain about the quality of other suppliers. Within an institutional structure that involves MSPs, workers likewise have little or no understanding of how their work assignments are sourced and priced. For example, an online community of temporary nurses discussed the issue when a problem surfaced; and a more experienced nurse said the following:

"It is not unethical for the recruiter not to tell you upfront about vendor managers [MSPs or technology vendors]. ... It is complicated to explain, and there is no decision you would have to make with [the vendor managers]. They don't tell you what the bill rate is, or how much they pay for liability insurance, or what a recruiter makes. On the other hand, most any good agency will tell you if you ask (and the recruiter knows), but unless you are an advanced [temporary nurse], you will not know what you can do with that information so there is not much point to it." 22

21 This pricing model is commonly referred to as the "supplier-funded model" where buyers of staffing services, that is, the lead firms, do not pay any upfront fees to use the MSP. Instead, the MSP fee (typically ranges from about 2% to 5% of bill rate) is incorporated in the bill rate per transaction. 22 Anonymous contributor in allnurses.com, Feb 18, 2012.

54 While this set of temporary workers has a community online where they can at least get some information on the pricing complexities, most workers in alternative work arrangements have little insight into such information. Most publicly available information on staffing agencies (through easily accessible outlets such as staffing agency or association websites, blogs, or news articles) do not state clearly how staffing agencies source their labor, i.e. their connections to specific MSPs or the usage of technology platform. In the absence of transparent information, therefore, workers do not have a strong reason to seek out alternatives even when they are paid at a rate that could otherwise be negotiated higher by a staffing agency that works with a lead firm directly. In sum, an additional contracting layer between the buyer (i.e. the leadfirm) and the seller of labor (i.e. the worker) is associatedwith lower returns to the seller as the buyer gains bargainingpower over others in the supply chain (Proposition1b).

Worker Power and Wage Determination

In this institutional context, is there a way for workers to gain power and subsequent financial gains despite their position at the bottom of the supplier chain? I have argued that an additional contracting layer benefits the lead firm while reducing worker pay, primarily because the intensified competition encourages suppliers, who have more bargaining power over workers, to transfer the price burden to workers. Put differently, this should mean that absent the suppliers' negotiating power to control prices to their advantage over workers, they would have less financial gain in exchange for higher worker pay, and take even lower profits when the distance from the firm increases. Since brokers gain power to negotiate prices when they become less replaceable with the value gained from access to private information (Marsden 1982), their negotiating power would likewise be reduced if their capacity to provide value to the given party is limited by design. For suppliers of labor, their primary role as brokers is to provide a matching function between two parties, where their access to information about the matching parties, or their repeated success or failure in providing quality matches, would dictate the extent to which they can negotiate prices to their advantage. Then, in an

55 institutional context where suppliers mediating do not provide a match between firms and workers,

they would no longer have the discretion to lower worker wages in return for higher profits.

In my setting where fums actively decide to centralize the management of nontraditional

workers, firms may also find workers using their own internal resources independent of the suppliers,

and manage these workers through the centralized system. In other words, firms sometimes contract

with suppliers solely for the purposes of "payrolling" workers that are "pre-identified", rather than to

find a suitable match. Often used interchangeably with payrolled or directly sourced workers, pre-

identified workers are typically workers that have already been identified by a hiring manager at the

lead firm: i.e., previously hired nontraditional workers, interns, alumni, retirees, and referred workers by current employees at the firm. For this set of workers, while suppliers still mediate the relationship

between the buyer and the workers, their role is significantly reduced, and so is their power to

negotiate prices, as they are not engaged in selecting and recruiting workers for the buyer. These

suppliers instead provide a plainly administrative function, such as the W-2 payroll processing,

timekeeping, and tax administration, to become the legal employer of record in place of the lead firm.

In comparison to competitively bid workers whose suppliers match them to firms, the pre-

identified workers are therefore relatively free from the influence of suppliers as they have direct, pre-

existing relationships with the buyer. Further, by design, because suppliers managing pre-identified

workers do not provide matching services, the supplier fee charged solely for their payroll services

should be lower than traditional supplier markup rates." In other words, suppliers should be able to

pay these workers higher wages or charge buyers lower bill rates, compared to when the same

suppliers compete in the bidding system. Previous studies suggest that these pre-identified workers,

that is, both previously hired workers and referred workers, may benefit from working directly with

the firm. This type of pre-existing firm-worker relationship is an example of what Uzzi (1997)

2 Carre, Margaret, "Are Pre-identified Workers Costing You More Than Assumed?" Allegis Global Solutions Whitepaper.

56 referred to as the "third-party referral networks and previous personal relationships" that make up an

embedded tie, often considered to be more efficient than an arm's length tie (e.g., firm relationship

with competitively bid workers) (see e.g., Ingram and Roberts 2000; Uzzi 1997). Research shows that

compared to nonreferrals, employee-referred candidates had more appropriate applications

(Fernandez, Castilla, and Moore 2000; Fernandez and Weinberg 1997), were more likely to be hired

(Fernandez and Weinberg 1997), and had higher initial wages at the point of hire (Brown, Setren, and

Topa 2016; Galenianos 2013; Montgomery 1991). Even when brokers provide the matching function, longer firm-worker relationships developed through a broker's repeat placement of its workers with the same firm are found to reduce the share of broker's gains from the transaction (Bidwell and

Fernandez-Mateo 2010).

Pre-identified workers then have considerable bargaining power over the buyer unlike typical workers who build relationships with the lead firms, because they are "named requests" which is indicative of their value for the lead firm. Although the savings from a direct match may equally benefit both the buyer and the workers, the buyers should be willing to spend more to hire these workers. In fact, industry experts have discovered that:

"Although, in theory, the final cost of a payrolled pre-identified worker should be lower than those sourced through the supply base, without proper oversight in place, this is almost never the case. Hiring managers have limited insight into market pricing for contingent labor and tend to be poor negotiators of wages for these workers."24

This leads to the following proposition about a relationship that is with minimal labor contracting layers. A pre-existing buyer-seller relationship, when it resembles a direct relationshipwithout contractinglayers, is associatedwith lower returns to the buyer (i.e. the leadfirm) and higher returns to the seller (i.e. the worker) than when the relationshipis competitively determined (Proposition2).

24 Ibid., p.3.

57 Moreover, the theoretical buildup for propositions 1 and 2 should predict similar results for

lead firms and pre-identified workers in multiple contracting layers of employment. A lead firm that

externalizes nonstandard workforce management using an MSP hires pre-identified workers through

suppliers that the MSP selects for purposes of payrolling or have selected for a previous assignment.

While the role of suppliers is still limited to payrolling function, the MSP continues to engage in the

management function by selecting suppliers and helping to negotiate contracts. In other words, for the buyer, the MSP-tie will help them negotiate lower prices for pre-identified workers than when the buyer negotiates directly with the payroll suppliers. On the other hand, the MSP-tie that forces

suppliers to transfer price burden to workers in competitive bidding should not influence wages of the pre-identified workers the same way. The payroll suppliers do not have bargaining power over the pre-identified workers as they are unable to extract value from matching. Therefore, I expect to find that a pre-existing buyer-seller relationshipbenefits the buyer (i.e., the leadfirm) and the seller of

labor (i.e., the worker) by moderating the effect of an additional contractinglayer between them

(Proposition3).

2.4 Data and Method

I use unique and proprietary data of employment records for nonstandard workers provided by

a North American company (hereafter, "ContingentCo"), a labor market analytics company

specializing in a segment of workers that they classify as "contingent". To my knowledge, this is the

first available data that provides detailed information on how nonstandard workers are hired (or

onboarded") to work assignments across multiple large firms in the United States: specifically, a

detailed list of suppliers and their types, information on the lead firms that workers get assigned to,

assignment characteristics, and their prices. The data consists of hiring information of nonstandard

workers from 49 large firms in the U.S. in the Fortune 500 companies list that have designated

25 The term "onboarding" is used by industry practitioners to refer to the "process of bringing a worker into a position with a goal of providing all necessary tools to be productive as soon as possible" (Staffing Industry Analysts 2017: p.27). For purposes of this paper, I will be using the term "hire" consistent with literature.

58 resources specifically for a Contingent Workforce Management (CWM) program in their human resources or procurement departments.2 6 In the past, one major challenge in understanding alternative work arrangements has been that there is a lack of publicly or privately available data that captures in detail both the multiple pathways that workers go through to get matched to a firm, as well as the precise rates at which the workers are paid and billed by suppliers of labor. Data on the establishment- or the firm-level use of nonstandard workforce is unable to capture individual-specific hiring information (see e.g., Cappelli and Keller 2013; Houseman 2001; Kalleberg, Reynolds, and Marsden

2003), while surveys at the individual-level are unable to capture both the modes of entry into work and how much the lead firms benefit from hiring these workers. The ContingentCo data is able to capture the entirety of the labor market in which these nonstandard workers as well as buyers and suppliers of labor operate along the supply chain and across different industries, occupations, and skill-levels.

Importantly, the data has information on whether these large firms have outsourced their

CWM programs to a third-party agency, a Managed Service Provider (MSP). Previous studies on outsourcing has relied on approximate and indirect measures of outsourcing, either through staffing or service firm industry classifications (Dube and Kaplan 2010) or input-output data from the U.S.

Bureau of Economic Analysis (BEA) that uses intermediate inputs of the outsourcing industry as a measure of outsourcing (Yuskavage, Strassner, and Medeiros 2008). Firm-specific data on whether a particular service frequently done in-house, in this case, managing of a nonstandard segment of workforce, is outsourced, is an important measure of outsourcing that has not been available in previous studies. Finally, in studies of nonstandard and contingent work, it has often been difficult to tease out the supply-side selection bias, as workers can choose to work in alternative work

26 It is important to note that the ContingentCo has been able to collect such information since all 49 firms make use of a technology platform referred to as the Vendor Management System (VMS) for an efficient processing and documenting of their contingent workforce. Even though I expect that the use of the VMS should lead to differential outcomes for firms compared to those that do not use the same system (which is still common for smaller firms or firms that use small number of nontraditional workers), it is beyond the scope of this paper.

59 arrangements or outsourced agencies for reasons not causally determined. It may be the case that staffing agency workers in manual jobs, for example, sort into these jobs with greater preferences for flexibility and autonomy in exchange for lower wages to reflect compensating differentials. To understand defacto the role of inter-organizational processes independent of supply-side preferences, one must hold constant the type of workers and their choice of their own legal employers (suppliers).

I exploit the fact that the ContingentCo data has information on individual job candidates, all of whom have already sorted into working with a brokering organization to find work.

My main empirical strategy is to estimate the prices that suppliers charge the lead firm and wages that workers receive for each assignment, depending on the types of buyer-seller or buyer- supplier relationships that alter bargaining power between stakeholders, controlling for finn, state, and occupation variations.

2.4.1 Description of Variables

The final dataset contains information on a sample of 955,169 nonstandard workers, 395,879 of whom have successfully been hired into a job assignment (either completed or currently onboarded to an assignment; 41.5% of all candidates in the data) across 49 large firms in the United States, mostly between years 2009 to 2017. As can be seen in Table 1, these workers are successfully hired into 158,652 job assignments across 23 industries.

[Table 1 about Here]

Dependent Variables. The main dependent variables of this study are the logged hourly bill rates that suppliers charge the buyers per worker matched to an assignment; and the logged hourly pay rates that workers receive per assignment from a supplier (their legal employer of record). The average bill rate and the pay rate for workers successfully hired are $46.3 and $29.9 per hour in constant 2017 dollars, respectively. I use information on the year in which the job vacancies were filled and adjust the prices to constant 2017 dollars using the CPI-U-RS from the Bureau of Labor

60 Statistics. I exclude observations that had incorrect specification of years (0.02% of observations,

which gives 395,803 out of 395,879 successfully hired workers with adjusted bill rates). Since all

suppliers have to bill the lead firm for every candidate they submit to the system, bill rates are

available for every candidate-level observation. However, not all suppliers are obliged to report to the

lead firm how much they pay the workers: in ContingentCo data, only about 48.3% of the successfully hired workers had pay rates reported to the lead firms by the suppliers (191,340 workers). Studies have found that disclosure of wage information resulted in higher wages for workers in Britain

(Rosenfeld and Denice 2015) and compressed wages among top managers in California (after a 2010 mandate to post online; Mas 2017).

By using ContingentCo data to attain wage information for nonstandard workers for those that disclose them, I may therefore be selecting suppliers who provide higher than average nonstandard worker wages. To supplement the analyses, I also use the logged hourly markup rates that suppliers receive from having successfully matched a worker to a lead firm. The profit for a supplier from this transaction should reflect the difference between the bill rate and the pay rate, minus fees paid to the

MSP (if the supplier works with one); however, the dataset does not specify the share of bill rate that gets paid to the MSP. Thus the markup rates in this data are the percentage difference between the bill rate and the pay rate, simply reflecting the share of price that flows to the labor market intermediaries as a whole. 27 The average markup rate for workers successfully hired are 34.2%.

Measure ofLabor ContractingLayer (Outsourcing). A firm is coded as having outsourced its

Contingent Workforce Management (CWM) if it works with a Managed Service Provider (MSP) to source its nonstandard labor. Otherwise, a firm is coded as having an in-house system if they use an internal program management office to manage nontraditional workforce. In the sample, 40 firms have outsourced its management of nonstandard labor and 9 firms manage nonstandard labor in-house.

27 The markup rates for suppliers providing workers for Firm A and Firm B are the following, subsequently: markup-rateA (% -rateA-pay-rate^ * 100; markuprateB() = billrateB-pay rateB * 100 + MSP-fee bill-rateA bil-rateB 61 I operationalize labor contracting layer as the distance between the worker and the lead firm; that is, the total number of labor market intermediaries involved in hiring a worker for an assignment. If an individual works for a lead firm that is coded as having outsourced its nonstandard worker management to an MSP, the labor contracting between the worker and the firm is considered to have one additional layer than when the MSP is not involved.

Firm characteristics. Besides the status of outsourcing CWM, I use one additional firm-level variable as a control. The lead firms report their financial spending on the CWM program, which can be used as a measure of their program size. 14.3 percent of the firms spend less than $5 million and

14.3 percent of the firms spend more than $250 million on managing nonstandard workforce (others distributed between $5-25 million (28.6%); $26-50 million (20.4%); $5 1-100 million (14.3%); and

$101-250 million (8.2%) in spending).

Job characteristics. ContingentCo categorizes nonstandard jobs as belonging to 23 "classes ofjobs" which reflect industries the jobs belong to: Administrative/Clerical, Business Professional,

Change/Implementation, Clinical/Scientific, Compliance, Risk and Audit, Customer Service,

Education, Engineering/Design, Finance/Accounting, Food and Beverage, Healthcare, Hospitality,

Human Resources, Industrial, Information Technology, Legal, Manufacturing, Marketing/Creative,

Operations, Procurement, Retail, Sales, and Other. Some of the largest industries include: Information

Technology (20.4%), Industrial (14.1%), and Healthcare (11.7%). ContingentCo also categorizes nonstandard jobs into 342 "job titles," which I use as an occupation classification. The top 10 occupations in the sample are composed of: Nurse (8.9%), Engineer (6.4%), General Laborer - Light

(5.1%), Software Developer (3.1), Administrative Assistant (3.0%), Technical Support Analyst

(2.8%), QA Associate (2.6%), Technician (2.5%), Assembler (2.4%), and General Laborer - Heavy

(2.44%). In addition, there is a variable that ContingentCo has coded to reflect the level of a job position: Entry (18.9), Intermediate (22.4%), Senior (11.6%) and Expert (4.7%). However, because

62 42.5 percent of the jobs do not have this information, this variable is used for robustness check. As can be seen, jobs in the sample represent different levels of skills, occupations, and industries.

Exposure to competition. I operationalize the exposure to competition using the different types of candidate sourcing models that vary by suppliers. In the sample, there are three major types of candidate sourcing models that suppliers adopt the most and are most relevant for the purposes of this paper. First, workers may be in the "competitive bidding" system where all candidates submitted by different suppliers are exposed to open competition and are theoretically given an equal chance, regardless of any pre-existing relationships between the exchange partners. It is the most common model in the industry today that has been facilitated by the adoption of a cloud-based technology platform (80.5% of successfully hired workers). Competitive bidding has actively been promoted to help minimize precisely the kinds of economic value capture as described by the brokerage scholars and to provide a level playing field for all the suppliers in the system. Second, workers may also be sourced as a "pre-identified" worker (13.5% of successfully hired workers). Pre-identified workers are those that have been identified by the lead firm without competition, which is indicative of a pre- existing relationship between the lead firm and the worker. Finally, workers may work with a "master supplier" where a supplier becomes the primary supplier of staffing for a certain type of workers- e.g., light industrial and clerical-before the requisitions are distributed to secondary suppliers in case they fail to find qualifying candidates. I use master supplier status (3.4% of successfully hired workers) to conduct further test of exposure to competition.

Worker Characteristics. For most of the main analyses, I use an assignment status variable to limit the sample to workers who are currently on a job assignment or have completed an assignment

(41.5% of the full sample). This is so that my dependent variables (the bill rates and the pay rates) reflect the final negotiated price for a given assignment. Other candidates in my sample are either in the selection or interview stages, or are rejected by the firm or the candidate after a bid has been made

(58.6% of the full sample). Most workers in the sample are agency workers (96.7%); that is,

63 nonstandard workers who have staffing agencies as formal employers on record, regardless of the specific terms of their contract. While I do not have demographic data on these individuals to control for their skills, I use the "worker quality" variable as an approximate measurement of their skills or quality. Out of those who are currently on a job assignment or have completed an assignment, 33.9% of workers have successfully completed an assignment, 3.0% of workers were hired full time, 1.2% of workers did not start the assignment as client needs had changed, and 7.2% of workers were terminated from the assignment or left early (54.7% workers unknown).

2.5 Results

Basic descriptive statistics in Table 1 (Panel A) show that the lead firms using internal CWM program are billed on average $51.44 per hour per nonstandard worker they hire, whereas the lead firms outsourcing its CWM program to an MSP are billed on average $44.30 per hour per nonstandard worker they hire. There is also a wage gap between nonstandard workers hired through an internal

CWM program and those hired through the MSP. The wage penalty for going through an additional layer of contracting is around $9.11 on average ($36.74 and $27.63, respectively) and the difference is significant (p<0.001). Table 1 (Panels B and C) shows that the share of the bill rate the workers receive for their pay varies depending on the type of suppliers workers contract with, as well as the types ofjobs they apply for. The raw numbers show that the workers hired into competitively bid jobs received 61.9 percent of the bill as their wages, while the pre-identified workers received 76.6 percent of the bill as their wages. Workers hired into jobs with high skills, such as engineer, quality assurance

(QA) associate, and technical support analyst positions, received higher share of the bill (between about 92% to 98%), whereas workers hired into jobs with low skills, such as general laborer and administrative assistant positions, received lower share of the bill (about 71%). Similarly, workers hired for expert-level jobs received the highest share of bill as wages (8 8.5%) while workers hired for entry-level jobs received the lowest share of bill as wages (72.3%).

64 Prior to presenting the main results, I begin by testing the core assumption of my argument

that when the buyer outsources the contingent workforce management to the MSP, the buyer has bargaining power to drive down bill rates and further, wages. To test this, I need to show that the

MSP compresses share of the bill that goes to the workers by finding suppliers that can charge lower bill rates to the buyer than when the buyer finds suppliers using the in-house management system.

Table 2 presents logits models in which the dependent variable is whether the candidate has been hired by the buyer. Model 1 first shows that the hiring probability is greater if the bill rates are lower and if the candidate goes through an MSP, controlling for the type of work arrangements, worker quality,

exposure to competition, firm-specific program size, state, job industry, and occupation. Model 3 then

examines whether the hiring probability given the bill rates varies with the MSP-tie by adding the

interaction term, outsourced (to MSP) x bill rate (logged). This term is negative and statistically

significant, which indicates that candidates going through MSPs are more likely to be hired for projects that have lower bill rates and less likely to be hired for projects that have higher bill rates, in

comparison to candidates going through an internal system at the lead firm. Even when controlling for the level ofjobs (entry, intermediate, senior, or expert), the results hold (Models 2 and 4). The bill rate coefficient turns positive in Model 4, but interaction term is still negative and significant,

indicating that the likelihood of hiring is not determined solely by the decreasing bill rate but is influenced by the MSP-tie that increases the likelihood of hiring with decreasing bill rates.

[Table 2 about Here]

2.5.1 Main Analysis

Table 3 provides the results from my main set of analysis. I use ordinary least-squares regressions to estimate the magnitude of which the outsourcing of nonstandard worker management to an MSP influences firm profitability (bill rates) and worker wages (pay rates). All regression models control for the type of work arrangements (agency worker, independent contractor, or sub-vendor), worker quality (successfully completed an assignment, hired full time, assignment incomplete as client

65 needs changed, or terminated), and the lead firm spending on CWM. While there is a lack of worker-

specific human capital information typically used to control for their fit for the job, I use a direct measure of their quality attained post-assignment to control for how they actually performed at the job.

Controlling for the lead firm spending on CMW is also important, as the use of a third-party agency to manage a program can simply be attributed to the firm or overall size of nonstandard workforce at the firm. The wage regressions also include fixed effects for the state (54 dummies including District of

Columbia, Puerto Rico, Virtual, and Unrecorded), job industry (23 dummies), and occupation (342 dummies), as it might be the case that some states or occupations are more susceptible to multiple layers of labor contracting. For example, there might be more workers in occupations with low-skills contracting with suppliers with an MSP-tie.

Consistent with Proposition 1 a, Model 1 in the baseline specification shows that outsourcing of nonstandard worker management to an MSP, that is, an additional contracting layer between the buyer and the seller of labor, is associated with the buyer's ability to receive lower bill rates from the

suppliers (0=-0.023; p<0.001). This means that an additional contracting layer allows the lead firm to

lower the bill rate by 2.3%, a reduction of about $1.06 per hour for an assignment billed on average

$46.3 per hour in constant 2017 dollars. Model 2 controls for the level ofjobs to account for some of the skills required for the job with a reduced sample, since not all jobs specify a level or are coded as having one. For example, ajob posting for an "administrative assistant" or a "computer programmer" position does not typically specify whether the job is for an entry- or an expert- level. The results stay

consistent with about a 4.2% reduction in the bill rate, which means that the lead firm benefits

financially from outsourcing its management to an MSP. Models 5 and 6 are used to test Proposition

Ib, that outsourcing of nonstandard worker management to an MSP is associated with lower wages for

the seller, as buyers gain bargaining power to set prices over suppliers. The negative and significant

coefficients (P=-0.081 and -0.045, subsequently; p<0.001) show that an additional contracting layer

between the buyer and the seller of labor decreases worker pay by 8.4% (Model 5), and by 4.6%

66 (Model 6) when controlling for the level ofjobs. Specifically, it refers to a reduction of about $2.51

per hour (and $1.38 per hour from Model 6) for an assignment paid on average $29.9 per hour.

[Table 3 about Here]

Table 3 also provides a test of shifting power dynamics between stakeholders by looking at

when the suppliers' negotiating power is absent, that is, when workers are pre-identified by the lead

firm (to resemble a direct relationship). The coefficients for pre-identified workers in Model 5 show

that compared to workers who are sourced through suppliers in the competitive bidding system, the pre-identified workers are likely to see their wages go up (P=0.237; p<0.001). In the reduced sample,

controlling for the level of jobs (Model 6), workers that have pre-existing relationships with the lead

firm (buyer) in the form of having been pre-identified is likely to see their wages go up about 8% compared to those that enter competitive bidding, which supports proposition 2. In Models 7-8, I add the interaction term, CWM outsourcedxpre-identified worker, to first find that the baseline coefficients for both outsourcing and pre-identified worker conditions remain similar to those in

Models 5-6 (5.2% wage reduction in the outsourcing condition and 5.6% wage gain in the pre- identified worker condition, when controlling for job-level in Model 8). The positive interaction term in Model 8 shows that the reduction of worker wages associated with additional labor contracting layer (0=-0.05 1; p<0.001) decreases by 70.6% for pre-identified workers compared to those sourced through competitive bidding.

The lead firms, on the other hand, are likely to see their bill rates go up when hiring workers that are pre-identified as compared to those that are competitively bid (0=0. 119; p<0.001; Model 1).

While the significance goes away when controlling for job-level dummies, Models 3-4 continue to find support for declining buyer profits per pre-identified worker hired. However, the negative and significant interaction term in Model 4 (0=-0.021; p<0.001) shows that controlling for job-level dummies, an increase in bill rates associated with hiring pre-identified workers disappears when CWM is outsourced as compared to when in-house (declines by 105.2%). The interaction term also means

67 that a reduction in bill rates associated with additional labor contracting layer (P=-0.038; p<0.001) is further amplified when hiring workers that are pre-identified as compared to those that are competitively bid (a further reduction by 55.3%). I therefore find support for proposition 3 that a pre- existing buyer-seller relationship, in the form of buyer identifying a known seller of labor, benefits the workers by moderating the negative effect of outsourcing CWM. Such relationship is found to benefit the lead firms, however, only when they outsource CWM for an MSP to manage the selection of payroll suppliers.

2.5.2 Supplier Selection

Unlike workers, suppliers have transparent information about how they relate to different

actors in the system; hence, the relationship between suppliers and buyers are less sticky. Suppliers that work with an MSP (Firm B in Figure Ib) as opposed to the lead firm directly (Firm A in Figure

lb) may then select into such an arrangement due to some unobservable supplier characteristics that

drive down the prices for workers rather than the power-reducing distancing. For example, the MSP-

tied suppliers may simply be low-road employers that pursue savings on labor costs (for a review of

literature on high- and low-road employment practices, see Osterman 2018); or, they may be smaller

and newer staffing agencies that lag behind in resources or information transfer, thereby affecting the

rate and volume at which workers are matched to firms. I therefore must show that for the same

supplier that submits candidates to both the MSP and the lead firm directly, its ability to set bill rates

for the buyer and pay rates for the seller should vary depending on their adding of an additional

contracting layer with an MSP.

Table 4 presents results from bill rate and pay rate regressions that replicate results from Table

3, but this time with supplier fixed effects, in order to control for unobservable supplier characteristics

that may influence its ability to charge prices. Models 1-2 in Table 4 replicate results from Models 2

and 4 in Table 3 and Models 3-4 in Table 4 replicate results from Models 6 and 8 in Table 3. As can

be seen, the results stay consistent throughout with supplier fixed effects. An additional contracting

68 layer of labor is associated with 6.1% reduction in the bill rate and 7.5% reduction in the pay rate

when controlling for both suppliers and job-level in Models 1 and 3 (support for proposition 1).

Specifically, it means that for an average bill rate of $46.3 per hour (in constant 2017 dollars), a lead

firm pays the supplier $2.82 less per worker hired when sourced through an MSP as opposed to an

internal system. For an average pay rate of $29.9 per hour (in constant 2017 dollars), a worker gets

paid $2.23 less when sourced through a supplier tied to an MSP. Consistent with propositions 2 and 3,

pre-identified workers are more likely than competitively bid workers to make higher wages per hour

(3.8%), and their working with multiple intermediaries have less negative effect than competitively

bid workers (the interaction term in Model 4 refers a reduction in the negative CWM outsourcing

coefficient by 69.6%). In addition, the lead firms hiring pre-identified workers are more likely to see their bill rates go up even when controlling for suppliers, but such a rise is moderated by the CWM outsourcing (the interaction term in Model 2 refers to a reduction in the positive pre-identified coefficient by almost 50%).

[Table 4 about Here]

As an additional analysis, I look at the "master supplier sourced" condition where workers have little bargaining power in the supply chain: in such a case, the supplier acts as a primary supplier of labor for a particular type of worker that are often lower skilled (e.g. light industrial and clerical).

This arrangement naturally creates a buyer-supplier relationship that suppresses, similar to multiple layers of labor contracting, worker wages while benefiting the buyer. In Tables 3 and 4, I find that workers sourced though master suppliers are billed at a reduced rate of about 1.5% (Model 1 of Table

4) and make about 2.4% less (Model 3 of Table 4) compared to when they are competitively bid, controlling for both job-level and supplier dummies. When an MSP is involved in selecting a master supplier for the lead firm, the interaction term CWM outsourcedx master suppliersourced worker in

Tables 3 (Model 4) shows that a reduction in bill rates associated with the buyer-supplier tie almost disappears. I find similar results using supplier fixed effects in Table 4 (Model 2), yet with small

69 significance (0=0. 130; p

2.5.3 Heterogeneity in Skills Demanded

In previous sections, I have used job-level dummies to account for variations in skills required for the job. Unlike the occupation classification used in the U.S. Census Bureau, real jobs differ in their required tasks at work depending on the level of the job. However, not all jobs have such a clearly defined description of a skill level at which tasks need to be performed; and, even within the same level of the job, the actual content of the job may require different types of skills. For example, an entry-level accountant position at one firm may require more cognitive skills such as mathematical and analytical skills, while the same position at a different firm may require social skills such as an ability to communicate with clients. It may be the case that the effect of additional contracting layer found in previous sections on prices is driven by heterogeneity in skills demanded when the lead firm hires workers through an MSP as opposed to internally. Some lead firms may work with an MSP to fill temporary vacancies that require low skills for peripheral tasks; and others may create an internalized system to fill vacancies that require high and complex skills, necessitating a direct assessment of candidates.

I develop a measure of skill requirements for nonstandard work, by coding specific skills demanded for each of about 200,000 job descriptions. Despite that numerous research have used job descriptions to extract job-specific information such as the skills (Deming and Kahn 2018) or words that are discriminatory (Kuhn and Shen 2013) or stereotypical (Castilla and Rho 2018; Gorman 2005), and that job descriptions are often accessible on-line, reliable job descriptions for nonstandard work are rare. This is because job advertisements online for nonstandard work are usually vague exemplary descriptions of jobs for which recruiters claim to help find job seekers eventually; hence, these job ads

70 typically exclude job- or firm- specific information. On the other hand, job descriptions for alternative

work arrangements from ContingentCo are real-worldjob descriptions that lead firms post on their

cloud-based platform for suppliers to view and submit applicants. These descriptions can therefore

serve as a reliable source of skills demanded at the lead firm when coded.

Using coded data on skill requirements demanded by the lead firms, I first test to see whether

my main proposition on the effect of additional contracting layer on bill rates and pay rates still holds

controlling for required skills. Then, I examine whether skills have any predictive power on setting

bill rates and pay rates, independent of additional contracting layer through the use of an MSP. In

fact, growing evidence points to the predictive power of detailed measurement of skills on wages in both economics and sociology literature. Deming and Kahn (2018) look at firms' differential demands

for both cognitive and social skills for standard workers as an explanation for wage inequality-they

find that firm variability in skill demand accounts for substantial variation in pay across firms, controlling for worker quality and occupations. Looking at long-term trend, Liu and Grusky (2013)

study rising returns to even more detailed measurement of skills and find that rising wage payoffs to analytic skills were the most substantial. I follow a set of studies that use a detailed measurement of skills to come up with ten detailed skill categories nested in four broader skill categories: cognitive

(analytical, quantitative, reading and writing), social (interpersonal, managerial, service), manual

(physical, routine), and technical skills (see e.g., Deming and Kahn 2018; Handel 2016; Liu and

Grusky 2013; Weaver and Osterman 2017 for related schemes).2"

28 Appendix Table 1 shows specific skill categories that I coded using a content analysis software, Linguistic Inquiry and Word Count (LIWC; see, e.g., Pennebaker, Booth, and Francis 2007) after having identified words that belong to each of the ten skill categories. These words draw on descriptions and task requirements of each skill in the Occupational Information Network (O*NET) (U.S. Department of Labor) data that are conventionally used in studies of skills and tasks (e.g. Autor, Levy and Murnane 2003; Acemoglu and Autor 2011; Liu and Grusky 2013). I use a bag-of-words approach to code a job description as having a particular skill requirement if it has at least one of the words listed for each skill, a binary value assignment following Deming and Kahn (2018). Similar to the Burning Glass Data for standard worker job vacancies that they use, ContingentCo also has unique keywords and phrases they have developed for each job-I also use their keywords for robustness check and for some of the skills where they had better coded information (technical skills and minimum level of ). As can be seen in Appendix Table 2, cognitive, social, and manual

71 I also code a set of complex (or "hybrid')skills to better understand the increasing

complementarity between different categories of skills. While earlier studies on skills have focused on

cognitive skills that were found to have generated increasing returns with growing complexity ofjobs

(Autor, Katz, and Kearney 2008; Autor, Levy, and Murnane 2003), recent work suggests that there is rising complementarity between cognitive and social skills (Borghans, Ter Weel, and Weinberg 2014;

Deming 2017; Deming and Kahn 2018; Weinberger 2014). The social skill premium is considered to have risen partly due to the growing demand for difficult-to-computerize skills in the midst of technological change (Autor 2014; Borghans et al. 2014) and also due to organizational changes such

as the rise in team-based and multitasking-demanded work design (Bloom and Van Reenen 2007;

Caroli and Van Reenen 2001; Ichniowski and Shaw 2003). For nonstandard jobs, numerous studies have illuminated the importance of social skills and social capital when navigating the labor market, in particular for high-skilled contractors. They have to develop relationships with hiring managers in

staffing agencies for better information and negotiation (Barley and Kunda 2006), while managing reputation for a constant flow of jobs (Evans et al. 2004).29 I therefore examine returns from both

cognitive and social skill demands for nonstandard workers, as well as a mix of social skills with

cognitive, technical, and manual skills. Importantly, I examine how much the firm is willing to spend

on hiring workers with these skills, which is different from the actual share of the pay that workers receive in the labor market for nonstandard work.

Table 5 presents results from several bill rate and pay rate regression estimates controlling for

four broad skill categories (cognitive, social, manual, and technical), as well as creativity, language

skills, personal traits, and minimum level of education required, not presented in the table for space

skills were required by 64%, 79.2%, and 18.8% of jobs, subsequently. Interpersonal skills were required the most in these jobs when looking at a fine-grained skills measurement, followed by service and analytical skills. A hybrid of cognitive and social skills were required by 56.1% of jobs, while much smaller share of jobs required technical and social skills (3.5%) or manual and social skills (14.1%). 29 Barley and Kunda (2006), studying a group of contractors with technical skills, have observed, "even the most valued expertise was useless, unless the contractor could repeatedly rise above the clamor of the market, often on short notice, to attract the attention of hiring managers [with social skills]" (Barley and Kunda 2006: 50).

72 limitations. Across all estimates, I include full set of controls used in Table 3, except for the job-level

dummies. I find that even when controlling for broad skill requirements (Models 1 and 4),

outsourcing of nonstandard worker management to an MSP continues to be associated with lower bill

rates for the lead firms (2.6%) as well as lower pay rates for the workers (8%). The magnitude of the

effect increases significantly when I run the same analyses using supplier fixed effects in Models 2

and 5 (12.5% decline in bill rates; 18.3% decline in pay rates). The coefficients indicating a buyer-

seller relationship --hiring of workers pre-identified-- also continues to be positive and significant for

both the bill and the pay rates; and the coefficients indicating a buyer-supplier relationship --hiring of

workers using a master supplier-- continues to be negative and significant for both the bill and the pay

rates.

[Table 5 about Here]

Independent of the outsourcing condition, the coefficients for skill requirement controls show

results that are consistent with studies that find higher wage premium for high skills required for

standard employment. Models 1-2 show that jobs with cognitive, social or technical skill

requirements are associated with higher bill rates, which indicates that firms are willing to spend more

money to hire workers for nonstandard jobs requiring high skills. Models 4-5 show that jobs with

equivalent skill requirements are associated with higher pay rates (social skill coefficient is only

significant with supplier fixed effects). On the other hand, job requirement for manual skill, which is

typically considered a lower skill, are associated with a decline in the bill rate by 6.1% and in the pay

rate by 3% (Models 1 and 4). I then evaluate returns to complex skills and find complementarity between cognitive and social skills even for nonstandard jobs. The pay rate coefficients for complex

skill requirements, that is, cognitive and social skill, technical and social skill, or manual and social

skill requirements in Models 6 are all positive and significant, controlling for supplier fixed effects

(P=0.089 and P=0.074, subsequently; p<0.00 1). For bill rates, the amount the firm is willing to pay to hire nonstandard workers with complex skills, the coefficient is positive and significant only for

73 cognitive and social skills (0=0.089; p<0.001). However, the coefficient for technical and social skill requirement is not significant; and, the manual and social requirement is associated with negative bill rates, though the magnitude of the effect is small (0=-0.008; p<0.05).30

2.6 Discussion

The last several decades have seen a rise in the outsourcing of nonstandard workforce management to a third-party organization, the MSP, who further contracts out hiring of workers to

staffing agencies or contractors. In this paper, I make the first scholarly attempt to examine the link between such multi-layered contracting arrangements governing employment for nonstandard jobs and

subsequent economic outcomes for both the hiring lead firms and the workers. Using power-

dependence theory, I argue that the lead firm's deliberate attempt to alter the recruitment process to

involve multiple intermediating organizations compresses supplier power, to eventually incentivize

suppliers to transfer the competitive burden to workers. In the process of doing so, I find that an

additional contracting layer between the lead firm and the worker is associated with higher returns to the buyer, a reduction of bill rate of about 2 to 6 percent, depending on model specification. An

additional contracting layer, on the other hand, is associated with lower returns to the workers, a

reduction of pay rate of about 4 to 8 percent, depending on model specification.

To strengthen my argument about how relationships may change in such a way to influence

the distribution of bargaining power among stakeholders, I further analyze a pre-existing buyer-seller

relationship where workers gain more power over both the suppliers and the lead firms. I find that the

"pre-identified" workers were associated with higher wages than those who were competitively bid,

and their loss from an additional contracting layer was less. For the lead firms, while their hiring of

pre-identified workers were associated with lower returns per transaction than when they hire

30 In a separate analysis, I run regression models predicting markup rates, the profit retained by suppliers. I find positive and significant markup rates associated with matching workers for jobs with all three types of complex skills.

74 competitively bid workers, the higher returns from an additional contracting layer were amplified

when they hired pre-identified workers instead of competitively bid workers.

One important puzzle from these results is: why do suppliers connected to the lead firms

through the firm's internal VMS technology pay workers above the reservation wage, where, in the

world of competitive bidding, they are equally likely to pay low wages to workers as the suppliers that

contract with the MSPs? One explanation may be due to the lack of transparency. While suppliers in

the competitive bidding system are bidding based on the bill rates offered by comparable suppliers,

they lack information about how other suppliers pay their workers. In the ContingetCo data, only

about half the candidates (49.5%) had pay rates recorded by the suppliers in the system. Considering

that some of the information are hand coded by suppliers post hoc, it is most likely the case that less

than half the suppliers revealed the actual pay rate in the system to the lead firms and other suppliers.

Without knowing how much others pay their workers, suppliers hoping to attract qualified candidates

will continue to pay their workers above the reservation wage (or as much as they can). While

suppliers are most likely to find out about the competitiveness of their pricing as they continue to

interact with their candidates through different assignments, because the candidates themselves also

lack transparency about the channels through which they get assigned to jobs (as elaborated above),

suppliers will often have a distorted picture of how much nonstandard workers typically get paid. One

logical step in my future research is to build on this explanation and test against other potential

explanations using more qualitative data.

In addition, I conduct several robustness checks to tease out alternative explanations. During the wage-setting process, there may be some unobservable supplier characteristics that influence their

decisions to set wages. It could also be the case that firms using different hiring arrangements for nonstandard workers may systematically demand different types of skills. I control for both

alternatives, first by using supplier fixed effects, and second by coding and controlling for a detailed

75 measurement of skill demands. Even with these specifications, I still find consistent and significant results.

In running models using the skill requirements for nonstandard workers, I also show that, consistent with the literature on skills, the requirements for higher skills such as cognitive, social or technical skills were associated with a higher wage premium, while the requirements for lower skills such as manual skills were associated with a lower wage premium. Further, complex skill requirements (cognitive and social, technical and social, or manual and social skill requirements) also produced greater returns to workers. Due to the rare availability of data on skill requirements for nonstandard jobs, there has been no attempt to date to make a detailed assessment of how skill requirements for nonstandard jobs may contribute to wages. While numerous studies have examined how labor market outcomes for nonstandard workers tend to be polarized by the skill level

(Houseman, Kalleberg, and Erickcek 2003), such work has been limited to single- firm or occupational analysis and they do not measure the firm demand in skills.

Future research should extend this work on skill requirements for nonstandard workers to see whether they differ from requirements in standard work, in order to gain a better understanding of how firm employment strategies may vary depending on the types of skills they demand. For example, the

MSP is often promoted for its ability to address the "skills gap", an idea that employers have difficulties finding workers with the right set of skills (Carnevale, Smith, and Strohl 2013; Manyika et al. 2011). Firms may decide to use nonstandard workers for skills that are more "expensive" to fill using a permanent labor force, and further, use an MSP to facilitate the process.

Findings from this paper make contributions to the literatures of organizational theory and

social structural inequality using the theoretical constructs of power and price-setting. I address how

inter-organizational processes, through multiple layers of contracting arrangements in recruiting

nonstandard workers, shift distribution of power among stakeholders to influence both firm profits and

worker wages per assignment. While I am unable to identify whether the same individual working

76 with varying numbers of intermediaries produce different outcomes (the key limitation of this paper), I speak to how workers hired to similar jobs with the same skill requirements may end up with unequal labor market outcomes depending on the layers of employment contracts. Importantly, I make an advancement in our understanding of how nonstandard workers are hired in today's labor market by opening the "black box" of institutional processes which are much more complicated than previously examined.

There are important implications for both human resource management and public policies from this work. The findings highlight that the recruitment for nonstandard workers increasingly resembles production supply chains with the use of multiple layers of contracting and competitive bidding practices. While this may lead to lowering of bill rates per transaction, the resulting economic consequences for workers should not be undermined. Recently, in fact, industry professionals have actively been discussing whether managing of nonstandard work in-house might be a better way to move forward, especially with the new technological availability to self-manage a large number of workers at once." Whether there are overall savings to firm profits from adopting different strategies is beyond the scope of this paper but should be an important area of future research. It is also important to note that many large firms have moved to centralizing nonstandard worker management and contracting out this function for the benefit of managing regulatory risks. The increase in lawsuits and DOL (Department of Labor) investigations related to misclassification and co-employment of nonstandard workers have incentivized firms to keep better track of external workers to minimize regulatory risks. An understanding of this new development, therefore, is crucial in advancing policy agenda for nonstandard workers.

3 Guest Contributor, "Integrated Workforce Solutions: A New Alternative in the MSP-IMP Debate" SpendMatters. June 22, 2017; Marsh, Leslie, "IMP vs Third Party Programs" Hiretalent. February 2016.

77 2.7 References

Autor, David. 2014. Polanyi's Paradoxand the Shape of Employment Growth. Working Paper. 20485. National Bureau of Economic Research.

Autor, David. 2008. The Economics of LaborMarket Intermediation:An Analytic Framework. w14348. Cambridge, MA: National Bureau of Economic Research.

Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2008. "Trends in U.S. Wage Inequality: Revising the Revisionists." Review ofEconomics and Statistics 90(2):300-323.

Autor, David H., Frank Levy, and Richard J. Mumane. 2003. "The Skill Content of Recent Technological Change: An Empirical Exploration." Quarterly JournalofEconomics.

Barley, Stephen R. and Gideon Kunda. 2006. "Contracting: A New Form of Professional Practice." Academy of Management Perspectives 20(1):45-66.

Bernhardt, Annette, Rosemary Batt, Susan Houseman, and Eileen Appelbaum. 2016. "Domestic Outsourcing in the U.S.: A Research Agenda to Assess Trends and Effects on Job Quality." IRLE Working Paper.

Bidwell, Matthew, Forrest Briscoe, Isabel Fernandez-Mateo, and Adina Sterling. 2013. "The Employment Relationship and Inequality: How and Why Changes in Employment Practices Are Reshaping Rewards in Organizations." The Academy ofManagement Annals 7(1):61-121.

Bidwell, Matthew and Isabel Fernandez-Mateo. 2010. "Relationship Duration and Returns to Brokerage in the Staffing Sector." OrganizationScience 21(6):1141-58.

Bloom, Nicholas and John Van Reenen. 2007. "Measuring and Explaining Management Practices Across Firms and Countries." The QuarterlyJournal ofEconomics 122(4):1351-1408.

Bonet, Rocio, Peter Cappelli, and Monika Hamori. 2013. "Labor Market Intermediaries and the New Paradigm for Human Resources." The Academy ofManagement Annals 7(1):341-92.

Borghans, Lex, Bas Ter Weel, and Bruce A. Weinberg. 2014. "People Skills and the Labor-Market Outcomes of Underrepresented Groups." ILR Review 67(2):287-334.

Brown, Meta, Elizabeth Setren, and Giorgio Topa. 2016. "Do Informal Referrals Lead to Better Matches? Evidence from a Firm's Employee Referral System." Journal ofLabor Economics 34(1):161-209.

Cappelli, Peter and Jr Keller. 2013. "Classifying Work in the New Economy." Academy of Management Review 38(4):575-96.

Carnevale, Anthony P., Nicole Smith, and Jeff Strohl. 2013. Recovery: Job Growth And Education Requirements Through 2020. Georgetown University Center on Education and Workforce.

Caroli, E. and J. Van Reenen. 2001. "Skill-Biased Organizational Change? Evidence from A Panel of British and French Establishments." The QuarterlyJournal ofEconomics 116(4):1449-92.

Castilla, Emilio and Hye Jin Rho. 2018. "Language and Gender in Online Recruitment Process." UnpblishedManuscript.

78 Cepin, Geremy C. 2012. "Choose Wisely When Hiring a Search Firm to Recruit for Key Positions." CPA Prac. Mgmt. F. 8:8.

Cool, Karel and James Henderson. 1998. "Power and Firm Profitability in Supply Chains: French Manufacturing Industry in 1993." StrategicManagement Journal 19(10):909-26.

Davis-Blake, Alison and Joseph P. Broschak. 2009. "Outsourcing and the Changing Nature of Work." Annual Review of Sociology 35(1):321-40.

Deming, David J. 2017. "The Growing Importance of Social Skills in the Labor Market*." The Quarterly JournalofEconomics 132(4):1593-1640.

Deming, David and Lisa B. Kahn. 2018. "Skill Requirements across Firms and Labor Markets: Evidence from Job Postings for Professionals." Journal ofLabor Economics 36(S 1).

Dube, Arindrajit and Ethan Kaplan. 2010. "Does Outsourcing Reduce Wages in the Low-Wage Service Occupations? Evidence from Janitors and Guards." ILR Review 63(2):287-306.

Elfving, Jan A., Iris D. Tommelein, and Glenn Ballard. 2005. "Consequences of Competitive Bidding in Project-Based Production." JournalofPurchasing and Supply Management 11(4):173-81.

Emerson, Richard M. 1962. "Power-Dependence Relations." American SociologicalReview 27:31-41.

Evans, James, Gideon Kunda, and Stephen R. Barley. 2004. "Beach Time, Bridge Time, and Billable Hours: The Temporal Structure of Technical Contracting Author(s): James A. Evans, Gideon Kunda, Stephen R. Barley Reviewed Work(S):" Administrative Science Quarterly 49(1):1-38.

Fernandez, Roberto M., Emilio Castilla, and Paul Moore. 2000. "Social Capital at Work: Networks and Employment at a Phone Center." American Journal ofSociology 105(5):1288-1356.

Fernandez, Roberto M. and Nancy Weinberg. 1997. "Sifting and Sorting: Personal Contacts and Hiring in a Retail Bank." American Sociological Review 62(6):883.

Fernandez-Mateo, Isabel. 2007. "Who Pays the Price of Brokerage? Transferring Constraint through Price Setting in the Staffing Sector." American Sociological Review 72(2):291-317.

Galbraith, Craig S. and Curt H. Stiles. 1983. "Firm Profitability and Relative Firm Power." Strategic Management Journal4(3):237-49.

Galenianos, Manolis. 2013. "Learning about Match Quality and the Use of Referrals." Review of Economic Dynamics 16(4):668-690.

Gorman, Elizabeth H. 2005. "Gender Stereotypes, Same-Gender Preferences, and Organizational Variation in the Hiring of Women: Evidence from Law Firms." American Sociological Review 70(4):702-28.

Gosman, Martin L. and Mark J. Kohlbeck. 2009. "Effects of the Existence and Identity of Major Customers on Supplier Profitability: Is Wal-Mart Different?" JournalofManagement Accounting Research 24.

Gould, Roger V. and Roberto M. Fernandez. 1989. "Structures of Mediation: A Formal Approach to Brokerage in Transaction Networks." Sociological Methodology 19:89.

79 Hahn, Chan K., Kyoo H. Kim, and Jong S. Kim. 1986. "Costs of Competition: Implications for Purchasing Strategy." JournalofPurchasing and MaterialsManagement 22(3):2-7. Handel, Michael J. 2016. "What Do People Do at Work?: A Profile of U.S. Jobs from the Survey of Workplace Skills, Technology, and Management Practices (STAMP)." JournalforLabour Market Research 49(2):177-97.

Houseman, Susan N. 2001. "Why Employers Use Flexible Staffing Arrangements: Evidence from an Establishment Survey." ILR Review 55(1).

Houseman, Susan N., Arne L. Kalleberg, and George A. Erickcek. 2003. "The Role of Temporary Agency Employment in Tight Labor Markets." Industrialand Labor Relations Review 57(1).

Ichniowski, Casey and Kathryn Shaw. 2003. "Beyond Incentive Pay: Insiders' Estimates of the Value of Complementary Human Resource Management Practices." Journal ofEconomic Perspectives 17(1):155-80.

Ingram, Paul and Peter W. Roberts. 2000. "Friendships among Competitors in the Sydney Hotel Industry." American JournalofSociology 106(2):387-423.

Kalleberg, Arne L., Jeremy Reynolds, and Peter V. Marsden. 2003. "Externalizing Employment: Flexible Staffing Arrangements in US Organizations." Social Science Research 32(4):525-52.

Katz, Lawrence and Alan Krueger. 2016. The Rise and Nature of Alternative Work Arrangements in the United States, 1995-2015. w22667. Cambridge, MA: National Bureau of Economic Research.

Khurana, Anil. 2002. "Professional Growth of IT Industry Role of Cost and Benefits Technology Usage and Personnel Dimensions."

Kim, Yoon Hee and Urban Wemmerl6v. 2015. "Does a Supplier's Operational Competence Translate into Financial Performance? An Empirical Analysis of Supplier-Customer Relationships: Does a Supplier's Operational Competence Translate into Finical Performance?" Decision Sciences 46(1):101-34.

King, Zella, Simon Burke, and Jim Pemberton. 2005. "The 'bounded'career: An Empirical Study of Human Capital, Career Mobility and Employment Outcomes in a Mediated Labour Market." Human Relations 58(8):981-1007.

Kuhn, Peter and Kailing Shen. 2013. "Gender Discrimination in Job Ads: Evidence from China *." The QuarterlyJournal ofEconomics 128(1):287-336.

Lanier, Danny, William F. Wempe, and Zach G. Zacharia. 2010. "Concentrated Supply Chain Membership and Financial Performance: Chain- and Firm-Level Perspectives." Journal of Operations Management 28(1):1-16.

Liu, Yujia and David B. Grusky. 2013. "The Payoff to Skill in the Third Industrial Revolution." American JournalofSociology 118(5):1330-74.

Lu, Qian. 2015. "The End of Polarization? Technological Change and Employment in the Us Labor Market."

Manyika, James et al. 2011. An Economy That Works: Job Creation and America's Future. McKinsey Global Institute.

80 Marsden, Peter V. 1982. "Brokerage Behavior in Restricted Exchange Networks." Social Structure and Network Analysis 7(4):341-410.

Mas, Alexandre. 2017. "Does Transparency Lead to Pay Compression?" JournalofPolitical Economy 125(5).

Montgomery, James D. 1991. Social Networks and Persistent Inequality in the Labor Market. Center for Urban Affairs and Policy Research Evanston. Osterman, Paul. 2018. "In Search of the High Road: Meaning and Evidence." ILR Review 71(1):3-34.

Osterman, Paul. 2004. "Labor Market Intermediaries in the Modem Labor Market." Workforce Intermediariesfor the Twenty-First Century 155-69.

Rosenfeld and Patrick Denice. 2015. "The Power of Transparency: Evidence from a British Workplace Survey." American SociologicalReview 80(5):1045-68.

Ryall, Michael D. and Olav Sorenson. 2007. "Brokers and Competitive Advantage." Management Science 53(4):566-583.

Simmel, Georg. 1950. The Sociology of Georg Simmel. Vol. 92892. Simon and Schuster.

Staffing Industry Analysts. 2017. Workforce Solutions Buyers Survey 2017 - Contingent Workforce ProgramFunding Models and Rates.

Uzzi, Brian. 1997. "Social Structure and Competition in Interfirm Networks: The Paradox of Embeddedness." Administrative Science Quarterly 42(1):35.

Weaver, Andrew and Paul Osterman. 2017. "Skill Demands and Mismatch in U.S. Manufacturing." ILR Review 70(2):275-307.

Weil, David. 2014. The Fissured Workplace. Harvard University Press.

Weinberger, Catherine J. 2014. "The Increasing Complementarity between Cognitive and Social Skills." Review ofEconomics and Statistics 96(5):849-61.

Wilmers, Nathan. 2018. "Wage Stagnation and Buyer Power: How Buyer-Supplier Relations Affect U.S. Workers' Wages, 1978 to 2014." American SociologicalReview 83(2):213-42.

Wise, Richard and David Morrison. 2000. "Beyond the Exchange-the Future of B2B." Harvard Business Review 78(6):86-96.

Yuskavage, Robert E., Erich H. Strassner, and Gabriel W. Medeiros. 2008. "Domestic Outsourcing and Imported Inputs in the US Economy: Insights from Integrated Economic Accounts." BEA Papers 0090. Bureau of Economic Analysis."

81 2.8 Tables and Figures

Figure 1. Changes in nonstandard workforce management

(a) Traditional triadic model

Workers Suppliers

Firms

(b) Two types of new employment model

Suppliers Workers (in cloud)

Firm A

Firm B MSP

82 Table 1. Summary of Descriptive Statistics

PANEL A: Firm Characteristics (49 firms) Contingent Workforce Management (CWM) program

In-HouseIn-House OutsourcedMSP) (to Difference (n=9) (n=40) Bill rate $51.44 $44.30 $7.13 (0.129) (0.067) (0.134)*** Pay rate $36.74 $27.63 $9.11 (0.119) (0.072) (0.142)*** Markup rate (share) 0.282 0.363 -0.081 (0.0005) (0.0004) (0.0007)***

PANEL B: Worker Characteristics (955,169 candidates) Pay rate (%) Bill rate ($) ($) Share bill rate Hired (on Assignment or completed) 41.5 46.3 29.9 0.646 Otherwise (interview , rejected, etc) 58.6 52.7 36.1 0.685

If Hired==1 (395,879 candidates) Exposure to Competition Competitive Bidding 80.5 45.9 28.4 0.619 Pre-identified 13.5 57.0 43.7 0.766 Master Supplier 3.4 16.4 11.9 0.730 Other 2.7 41.1 15.8 0.386 Type of Work Arrangements Agency Worker 96.7 44.6 29.5 0.661 Independent Contractor 0.7 92.6 89.0 0.962 Sub Vendor 1.1 110.2 81.1 0.736 Employee 1.6 66.4 n.a. n.a. Worker Quality Completed successfully 33.9 53.1 25.3 0.477 Hired full time 3.0 42.1 27.6 0.656 Canceled (client) 1.2 58.1 29.9 0.515 Terminated or left early (candidate) 7.2 39.4 25.3 0.641 Other 54.8 43.0 32.1 0.747

Mean SD Min Max Bill rate ($) 46.3 37.9 8.0 804.3 Pay rate ($) 29.9 27.1 5.7 685.1 Markup rate (share) 0.34 0.13 0.0 2.1 Notes. All in constant 2017 dollars calculated using CPI-U-RS and the year the lead firm released assignment request to suppliers. Standard errors in parenthesis. ***Statistically significantat the 0.001 level (two-tailed tests).

83 PANEL C: Job Characteristics (207,170 jobs) Pay rate (%) Bill rate ($) ($) Share bill rate

If Hired==l (158,652 jobs) 76.6 Job Level Entry 18.9 23.5 17.0 0.723 Intermediate 22.4 48.0 35.2 0.735 Senior 11.6 70.3 53.2 0.756 Expert 4.7 100.9 89.2 0.885 n.a. 42.5 47.2 26.3 0.557 State (Top 5) California 21.0 61.3 36.6 0.596 Minnesota 18.4 52.5 27.5 0.523 Washington 7.1 61.9 48.4 0.782 Texas 5.0 31.6 24.7 0.782 New Jersey 5.0 55.3 47.7 0.863 Job Industry (Top 5) Information Technology 20.4 68.7 58.6 0.853 Industrial 14.1 20.5 15.6 0.761 Healthcare 11.7 73.6 49.1 0.667 Engineering/Design 10.5 58.2 54.8 0.942 Manufacturing 7.4 26.7 17.8 0.669 Occupation (Top 10) Nurse 8.9 78.1 38.5 0.493 Engineer 6.4 52.7 51.4 0.976 General Laborer - Light 5.1 21.3 17.2 0.810 Software Developer 3.1 80.9 70.4 0.870 Administrative Assistant 3.0 29.5 21.1 0.715 Technical Support Analyst 2.8 31.2 28.6 0.919 QA Associate 2.6 31.9 29.6 0.928 Technician 2.5 38.2 22.2 0.581 Assembler 2.4 21.9 17.4 0.796 General Laborer - Heavy 2.3 15.0 10.7 0.714

Mean SD Min Max Unique candidatesperjob 4.6 11.9 1 1,135 Unique suppliers perjob 2.0 2.7 1 38 (Suppliers in total: 2,222) Notes. All in constant 2017 dollars calculated using CPI-U-RS and the year the lead firm released assignment request to suppliers

84 Table 2. Logit Regression Models Predicting Whether the Candidate is Hired by the Lead Firm

Bill Rate interacted with Outsourcing Model 1 Model 2 Model 3 Model 4 Ln(Bill Rate) -0.439*** -0.049*** -0.139*** 0.450*** (59.572) (4.483) (13.327) (29.880) Outsourced (to MSP) -0.963*** -0.537*** 0.690*** 2.074*** (96.330) (45.169) (16.432) (37.006) Outsourced x Ln(Bill Rate) -0.442*** -0.689*** (40.476) (47.727) Job-level dummies Yes Yes

Constant 2.826*** 1.152*** 1.812*** -0.503+ (11.039) (4.321) (6.928) (1.813)

N candidates 951,001 644,803 951,001 644,803 Notes. All regressions include dummies for type of work arragements, worker quality (reason for closing of an assignment), exposure to competition, firm-specific program size (spending on CWM), state, job industry, and occupation. Standard errors are in parenthesis. +p<0.10, *** p<0.001 (two-tailed).

85 Table 3. Bill Rate and Pay Rate Regressions for Nonstandard Workers Hired

Ln (Bill Rate) Ln (Pay Rate)

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8 ~0.045*** ~0.l36*** ~0.05l*** CWM: Outsourced (In-house as reference) -0.023*** -0.041 *** -0.024*** -0.038*** -0.081*** -0.045*** -0.136*** -0.051*** (0.003) (0.002) (0.003) (0.003) (0.003) (0.003) (0.004) (0.004)

Worker: Pre-identified (Competitively bid as reference) 0.119*** 0.003 0.091 *** 0.019*** 0.237*** 0.077*** 0.069*** 0.055*** (0.002) (0.002) (0.004) (0.004) (0.003) (0.003) (0.005) (0.004) Worker: Master Supplier Sourced (Competitively bid as reference) -0. 118*** -0.036*** -0.183*** -0.128*** -0.104*** -0.028*** -0.086*** 0.145*** (0.004) (0.005) (0.013) (0.012) (0.004) (0.006) (0.021) (0.019) Outsourced x Pre-identified 0.039*** -0.021*** 0.238*** 0.036*** (0.005) (0.004) (0.005) (0.005) -0.014 -0.190*** Outsourced x Master Supplier Sourced 0.071 *** 0.111*** (0.014) (0.013) (0.022) (0.020) Job-level (Entry as reference) Intermediate 0.194*** 0.193*** 0.219*** 0.221*** (0.002) (0.002) (0.003) (0.003) Senior 0.399*** 0.398*** 0.431*** 0.433*** (0.002) (0.002) (0.003) (0.003) 0.769*** Expert 0.635*** 0.633*** 0.765*** (0.003) (0.003) (0.005) (0.005)

2.723*** 2.647*** Constant 3.094*** 3.137*** 3.111*** 3.167*** 2.666*** 2.668*** (0.056) (0.045) (0.056) (0.045) (0.072) (0.062) (0.071) (0.062)

118,444 191,340 118,444 N Hired Workers 395,803 238,852 395,803 238,852 191,340 0.751 0.833 0.753 0.833 n .715 0822.usteu 0715.k-sqUIU 0.822

Notes. All regressions include dummies for type of work arragements, worker quality (reason for closing of an assignment), firm-specific program size (spending on CWM), state, job industry, and occupation. Standard errors are in parenthesis.

*** p<0.001, ** p<0.01, * p<0.05 (two-tailed)

86 Table 4. Bill Rate and Pay Rate Regressions for Nonstandard Workers Hired, with Supplier Fixed Effects

Ln (Bill Rate) Ln (Pay Rate) Model 1 Model 2 Model 3 Model 4 CWM: Outsourced (In-house as reference) -0.059*** -0.056*** -0.072*** -0.079*** (0.004) (0.004) (0.006) (0.006)

Worker: Pre-identified (Competitively bid as reference) 0.036*** 0.055*** 0.070*** 0.037*** (0.002) (0.005) (0.003) (0.005) Worker: Master Supplier Sourced (Competitively bid as reference) -0.015* -0.144* -0.024*** -0.062 (0.006) (0.071) (0.006) (0.072) Outsourced x Pre-identified -0.027*** 0.055*** (0.005) (0.006) Outsourced x Master Supplier Sourced 0.130+ 0.040 (0.071) (0.072)

Supplier FE Yes Yes Yes Yes Job-level Control Yes Yes Yes Yes

Constant 3.298*** 3.295*** 2.361*** 2.367*** (0.273) (0.273) (0.275) (0.275)

N Hired Workers 238,852 238,852 118,444 118,444 Adjusted R-squared 0.858 0.858 0.865 0.865 Notes. All regressions include dummies for type of work arrangements, worker quality (reason for closing of an assignment), firm- specific program size (spending on CWM), state, job industry, occupation, job-level, and supplier. Standard errors are in parenthesis. *** p<0.001, ** p<0.01, * p<0.05, + p

87 Table 5. Bill Rate and Pay Rate Regressions for Nonstandard Workers Hired, with Supplier Fixed Effects and Skill Requirements

Ln (Bill Rate) Ln (Pay Rate) Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 CWM: Outsourced (In-house as reference) -0.026*** -0. 116*** -0.118*** -0.077*** -0.170*** -0.168*** (0.003) (0.004) (0.004) (0.003) (0.005) (0.005)

Worker: Pre-identified (Competitively bid as reference) 0.125*** 0.055*** 0.054*** 0.240*** 0.094*** 0.093*** (0.002) (0.003) (0.002) (0.003) (0.003) (0.003) Worker: Master Supplier Sourced (Competitively bid as reference) -0.l10*** -0.055*** -0.054*** -0.106*** -0.062*** -0.062*** (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Skill Requirements Cognitive 0.082*** 0.052*** -0.007** 0.060*** 0.058*** 0.014*** (0.002) (0.002) (0.003) (0.002) (0.002) (0.003) Social 0.018*** 0.022*** -0.012*** 0.004 0.012*** -0.021*** (0.002) (0.002) (0.002) (0.002) (0.002) (0.003) Manual -0.059*** -0.058*** -0.052*** -0.030*** -0.048*** -0.054*** (0.002) (0.002) (0.003) (0.003) (0.003) (0.004) Technical 0.020*** 0.016*** 0.025*** 0.039*** 0.028*** 0.010 (0.003) (0.003) (0.007) (0.005) (0.004) (0.008) (Complex skills) Cognitive and Social 0.089*** 0.074*** (0.003) (0.004) Technical and Social -0.011 0.026** (0.008) (0.010) Manual and Social -0.008* 0.012** (0.004) (0.005)

Supplier FE No Yes Yes No Yes Yes

Constant 3.077*** 3.623*** 3.635*** 2.689*** 2.787*** 2.792*** (0.055) (0.320) (0.319) (0.071) (0.130) (0.130)

N Hired Workers 395,803 395,803 395,803 191,340 191,340 191,340 Adjusted R-squared 0.723 0.804 0.805 0.757 0.817 0.818 Notes. All regressions include dummies for type of work arrangements, worker quality (reason for closing of an assignment), firm- specific program size (spending on CWM), state, job industry, and occupation. Standard errors are in parenthesis.

*** p<0.001, ** p<0.01, * p<0.05 (two-tailed)

88 2.9 Appendix

Appendix Table 1. Skill Categorization of Job Descriptions

Skill and task requirements Coded words Cognitive skills cognitive (cognitive skill*); analytical (analytic*); critical thinking (critical thinking, think* critical*, deductive, inductive, reasoning); problem solving (problem solving, problem-solving, Analytical solution*, cost-effective, cost effective, propos* option*, prepar* proposal*, complex problem*, identif* problem*); decision making (decision making, decision-making); research (research, analys*, analyz*, interpret*, organiz* information, gather* information, information gathering)

Quantitative basic (math* ability to add, subtract*, multiply, multiplication, arithmetic); advanced (advanced math* stati*, algebra*, geometry, trigonometry, calculus, quantitative, modeling, data model) Reading and Writing basic (basic read*, basic writ*); other (reading, to read, writ*)

Soft skills social (social skill*); interpersonal (interpersonal); communication (communicat*, oratory, oral, Social/interpersonal oral skill*, speak* contact* interact* facilitat*, conversat*); ability to work in teams (team* collaborat* build trust, conflict, group, help other, ability to relate, relationship, diplomatic); negotiation (negotiat*); presentation skill*, to present (presentation) Managerial lead*, motivate, project management, manage project*

Service care work (customer*, patient*, care for, caring); other service work (sales, complaint*, client*)

Manual skills Physical manual labor, lift*, move, unload*, climb*, bend*, kneel*, long period, long hour* Routine routine*

Technical skills (ContingentCo keywords) Basic and advanced computer e.g. word, excel, python, stata, java, and many more

Other skills Creative original, originality, ingenuity, inventive, creativ*, innovat*, initiate, initiation, self-start*, self start* conscientiousness (conscientious*, organized, organiz* skill*, organizational and, competen*, ambitio*, attitude*, inspir*, , professionalism, attentive, persiste*, self-motivated, self motivated, self driven, self-driven, motivated, hardworking, hard-working, hard working, Personal Traits dependable, proactive*, pro-active*, detail*, accura*, meticulous*, diligent*, thorough*); extraversion (outgoing, sociable, friendly, enthusiastic, energetic, assertiv*, self-confiden*, self confiden*, confident, confidence, adventurous); agreeableness (trustworthy, trusted); multi-task (multi task*, multitask*, multi-task*); time-management (time management, on time, on-time, meet deadline, meeting deadline, priorit*, set goal) language bilingual arabic cantonese chinese creole dutch english french german gujarati hindi Language italian japanese korean latin mandarin polish portuguese russian spanish vietnamese (ContingentCo keywords)

Education, minimum level required high school, associate, bachelors, masters, doctor

Organization- and job-specific aspects of work no oversight, minimal oversight, little oversight, no supervision, minimal supervision, little Autonomy supervision, no guidance, minimal guidance, little guidance, independen*, take charge, ownership, self-discipline*, self discipline* Unionization collective bargaining agreement*, union contract* union agreement*

Irregular shift flexible hour*, flexible shift*, irregular hour*, irregular shift*

89 Appendix Table 2. Share of Job Advertisements for Nontraditional Jobs with Each Skill Requirement

Mean SD Share of Job Ads with Following Skill Requirements (n=168,01 4; Min=0; Max= 1) Cognitive 0.640 0.480 Analytical 0.396 0.489 Quantitative 0.096 0.294 Reading and Writing 0.385 0.487 Social 0.792 0.406 Interpersonal 0.612 0.487 Managerial 0.229 0.420 Service 0.422 0.494 Manual 0.188 0.391 Physical 0.142 0.349 Routine 0.057 0.233 Technical 0.043 0.203 Hybrid Cognitive and Social 0.561 0.496 Technical and Social 0.035 0.184 Manual and Social 0.141 0.348 Other Skills Creative 0.085 0.279 Language 0.166 0.372 Personal Traits 0.415 0.493 (n=35,307; Min=O; Max=30) Work Experience (avg. year) 4.8 3.1 Education (min. required) (%) High School 23.0 Associate 2.7 Bachelors 10.9 Masters 0.9 Doctor 0.2 Not specified 62.3

90 Chapter 3. Language and Gender in the Online Job-Matching Process

Co-authored with Emilio J. Castilla*

Abstract

Research on gender segregation has tended to focus on either on the role of employers' preferences during employee hiring or the role ofjob seekers' choices during theirjob search. Far less understood, however, is how interactional processes between recruiters and job seekers that operate from the employer side also influence gender sorting of the job seekers in the online job-matching process. In this article, we move beyond the well-studied effects of the gender-typing of jobs and industries to identify and test two distinct (complementary) interactional mechanisms-gendered language (as experienced by job seekers) and in-group preferences (as exercised by job seekers)-that may differentially affect how job seekers make application decisions at the initial recruitment phase when jobs are described by recruiters. Using unique data from both a field study (Study 1) and a field experiment study (Study 2) of online job postings, we first investigate the gendered language mechanism by testing whether the specific words recruiters use to advertise otherwise identical jobs to candidates (masculine/feminine words) differently affect female and male job seekers' interest in and application to the job. We then investigate the in-group preference mechanism by testing whether job seekers are more likely to inquire about and apply to otherwise identical jobs when the gender of the recruiter is similar to their own. We conclude by discussing the implications of our study for understanding the role language and employer recruitment efforts play in sustaining gender segregation in today's labor markets.

* We are grateful for financial support received from the James S. Hardigg (1945) Work and Employment Fund and the MIT Sloan School of Management. We thank Harshini Jayaram for providing research assistance at the early stages of our field study. We also thank many of our colleagues at MIT for their feedback on earlier versions of this paper, specifically that of Roberto Fernandez, Thomas Kochan, Paul Osterman, Ben A. Rissing, Susan Silbey, and Lotte Bailyn. We also thank Peter Cappelli, Matthew Bidwell, Kate Parsons, and Nancy Rothbard for their helpful suggestions. We also benefited from the feedback of seminar participants in the Institute for Work and Employment Research and the Economic Sociology Working Group at MIT. An earlier version of this study was presented at the 2017 ASA Annual Meeting, the 2014 LERA Conference and the 2014 Seventh Wharton People and Organizations Conference. The views expressed here are exclusively those of the authors.

91 3.1 Introduction

Gender segregation research considers employer-job seeker interactions during the job-matching process-both a demand (employer) side and a supply (job seeker) side process-to influence how men and women are allocated to different positions at work (see Reskin 2000; Ridgeway 1997).

Starting at an initial point of contact, employers frequently interact with job seekers to find the best candidates for available jobs. On the one hand, employers directly or through the use of recruiters and other labor market intermediaries (i.e., online recruitment agencies) advertise job openings on company and external sites such as newspapers, job boards, and most frequently now, online, with the goal of reaching a large pool of talented candidates. On the other hand, job seekers then learn of, inquire about, and apply to available jobs based on the recruiters' descriptions ofjobs. As a result, the recruitment process potentially shapes job seekers' preferences and choices, discouraging/encouraging women and men to choose certain jobs, with eventual implications for the persistence of gender segregation in today's labor markets (see, e.g., Correll 2001, 2004; Fernandez and Friedrich 2011;

Barbulescu and Bidwell 2012; Wynn and Correll 2017). Yet, we still know little about how during these interactions, employers' gender hiring preferences get revealed to shape job seekers' behavior depending on their gender-a joint process that is both demand and supply sided (for a related exception, see Fernandez and Sosa 2005).

During the job-matching process, interactions enable employers to automatically categorize workers by gender and their stereotypes and to develop preferences about the "ideal worker" to fill their job vacancies (Ridgeway 1997, 2011). These implicit biases get embedded in organizational structures and procedures, such as the language in job descriptions, to make gender salient during the job search. Due to the interactional nature of the job-matching process, job descriptions should then influence both the decisions of employers to hire workers and the choices that job seekers make about their applications. While research shows that interactional processes from the demand side influence gender differentiated hiring of job seekers (see, e.g., Gorman 2005), less understood is how such

92 interactional processes shape the supply side: the application decision of job seekers by gender.

Mostly because of the difficulties inherent in observing and collecting data on how employers word

these job postings and how job seekers react to them in the real-world setting, prior empirical research

has not generally paid attention to the joint employer-job seeker interactional process (Gorman 2005;

Pager, Western, and Bonikowski 2009).

We address these past research difficulties by studying in-depth how employers' gendered

recruiting processes at the recruitment phase may affect the ways in which men and women are

differently sorted into jobs in today's labor markets, particularly those online. Specifically, we

identify and test two key distinct (complementary) interactional mechanisms-gendered language (as experienced by job seekers) and in-group preferences (as exercised by job seekers) when jobs are being described by recruiters-that potentially affect how job seekers make application decisions at the early job-matching process, beyond and above the effects of other well-studied gender-segregating mechanisms such as the gender-typing of jobs, tasks, and/or industries. First, drawing on prior research on the gender job stereotyping, we investigate the gendered language mechanism.

Specifically, we test whether the stereotypically masculine/feminine words recruiters use to describe jobs to prospective candidates differently affect female and male job seekers' interest in applying for what are otherwise identical jobs. Second, since job seekers' interest in jobs may depend on the individual recruiters advertising the job, we assess the in-grouppreference mechanism by testing whether job seekers are more likely to inquire about otherwise identical jobs when the gender of the recruiter matches their own.

To investigate how these gendered language and in-group preference mechanisms may operate during recruitment, we analyze unique data from two different studies-a field study of a recruitment company's job search platform (Study 1) and a field experiment study of an online job posting (Study

2). In Study 1, we use the dataset of an online labor market from a global company-"RecruitCo"

(henceforth)-that provides web-based employment recruiting services. In particular, we conduct a comprehensive content analysis of the entire population of almost 150,000 jobs posted by

93 approximately 25,000 different recruiters reaching more than half a million job seekers in the United

States using the RecruitCo online job search platform. On this platform, recruiters for corporate or staffing firms present themselves with short self-descriptions, names, titles, companies they work for, and job vacancies they are looking to fill. Job seekers have access to these recruiter job postings and are able to inquire about specific jobs through an informal forum. This forum allows us to observe directly one of the earliest possible interactions between recruiters and job seekers in the labor market, that is, at the pre-application stage, by tracking the inquiries job seekers make before they formally apply for a job. These job postings also allow us to measure the share of stereotypically masculine and feminine words used when describing real jobs online (more detail is provided later in article). In

Study 2, we conduct a field experiment study of one online job posting in order to further test our two key mechanisms while addressing some key limitations of the field study (Study 1). Using a between- subject experimental design, we advertised a realjob online to a population of 4,150 survey participants for a period of one month. We manipulated both the gendered language used to describe one identical job of "project assistant" (that is, feminine wording, masculine wording, or gender- neutral wording of the job posting) and the gender of the recruiter posting the job.

Beyond highlighting the significance of job postings and other early recruitment employer efforts for understanding gender sorting, this study makes a number of contributions to the employment, organizations, and gender segregation literatures. First, although these two proposed mechanisms may complement each other, each operates independently in many organizational settings and across jobs/occupations/industries, with potentially major consequences for gender sorting in the labor market. To our knowledge, these two interactional mechanisms have not been empirically tested before as experienced and exercised from the job seeker side to shape their real application decisions

(and not simply attraction to a job), mainly because of the extraordinary data requirements necessary for studying them (see Fernandez and Sosa 2005). Second, since these mechanisms have never been studied simultaneously across a wide-range of real jobs, occupations and industries, it is possible that when we think that we observe gendered sorting because of the gender stereotypicality of the job,

94 occupations, or industries, as the main mechanisms leading to gender segregation, we are actually seeing the evidence of the distinct role of employers in revealing their hiring preferences through these two (complementary) mechanisms. In this regard, job seekers of different genders may not only react differently to how stereotypically gendered the job itself is, but also to the stereotypically gendered language used in the job descriptions, as well as the gender of the recruiters advertising the job.

Investigating the role of these mechanisms thus has implications for furthering our understanding of how gender segregation is sustained in current labor markets. Finally, by examining these two mechanisms during recruitment, this article addresses research calls to study the potential micro-level processes that affect gender disparities in the workplace (see, e.g., Stainback, Tomaskovic-Devey, and

Skaggs 2010; Castilla 2011; Rivera 2015).

3.2 Interactional Mechanisms at Work: The Context of Online Job Matching

The supply-side interactional processes are inevitable in a two-sided job matching (both pre- application and post-application as summarized in Figure 1) where the demand-side gender stereotyping makes gendered criteria salient in the job ad. While job seekers typically begin the search with an exposure to job descriptions that are potentially gendered, little empirical evidence exists regarding whether and how job descriptions and recruiters affect the ways in which men and women sort themselves to different jobs, especially at this very early phase of the job-matching process. In this article, we start addressing this gap by studying how employers' choice regarding the language used to advertise jobs and the gender of the recruiters posting the jobs may affect the way job seekers of different genders react to job postings. Below, we identify and test two key distinct

(complementary) interactional mechanisms-gendered language (as experienced by job seekers) and in-group preferences (as exercised by job seekers) when jobs are being described by recruiters-that may differently shape the pre-application behavior ofjob seekers depending on their gender in today's labor markets.

[Figure 1 about Here]

95 3.2.1 Gendered Language of Job Descriptions

Job advertisements are crucial starting points at which employers express their pre-defined ideas about who ought to fill the vacancies. In the United States, explicit language in job advertisements that reveals a preference for employees of a particular gender is prohibited under Title

VII of the 1964 Civil Rights Act. Prior to that Act, explicit gendered hiring was common, with many employers' job ads in newspapers directly encouraging gender-specific candidates to apply (see, e.g.,

Bem and Bem 1973; Pedriana 2004). Despite the initial widespread noncompliance with Title VII, explicit gender naming preferences in job ads eventually ended in the United States by 1973 (Pedriana and Abraham 2006). In places where communicating gender hiring preferences is legally permissible, however, job ads continue to be gender-targeted especially for lower-skilled positions, even as such discriminatory language is largely absent from postings for higher-skilled positions (e.g. see Kuhn and

Shen 2013 for a study in China).

While legal barriers in the United States today prevent employers from specifying any gender preference in job advertisements, the process behind the preparation of ads may nonetheless provide an opportunity for the revelation of implicit gender preferences. Research traditions on the micro foundations of demographic inequality, in both sociology and psychology, have long emphasized the role of cognitive and implicit bias influencing the ways in which employment-related decisions are made (see, e.g., Bielby 2000; Heilman, Martell, and Simon 1988; Reskin 2000). These cognitive processes such as gender stereotyping are found to be embedded in the language employers use to describe jobs. Gaucher, Friesen, and Kay (2011), for example, argue that job ads carry subtle gendered wording that-whether intentionally or not-may ultimately affect gender segregation in the labor market. In their analysis of real-world job postings, they find that job ads contain more wording commonly perceived as masculine in male-dominated occupations than in female-dominated occupations. Masculine language in the job descriptions include words like "determined,"

"analytical," "independent," "decisive," "persistent," "ambitious," and "assertive," while feminine

96 words include "committed," "interpersonal," "," "compassionate," "honest," and

"understanding" (see, e.g., Gaucher et al. 2011).

The gendered language used to describe jobs to job seekers then is also an indirect reflection of employers' organizational schemas-that is, images of the "preferred or ideal" worker-which are culturally pre-inscribed within different organizational contexts (Ridgeway 1997). These preferred images guide gender stereotypes that are made cognitively available to subconsciously distort judgements about those that interact (Blair and Banaji 1996; Kunda and Spencer 2003). Studies show that gender stereotyping by employers influence employers' hiring decisions (Rudman and Glick

1999, 2001) and the of worker performance (Biernat and Kobrynowicz 1997; Heilman

2001). Further, Gorman (2005) finds that gender stereotypical words in job ads, as schemas formed through incumbents of an organizational position, influence the gender composition of hires in law firms. The key assumption in her analysis, which we build on in our study, is that language, as reflected in job postings, serves as a tool for the organizational schemas to shape the gender division of labor by employers. When job postings (or selection criteria in them) contain more stereotypically masculine criteria, Gorman reports a smaller share of women among new hires in law firms, whether the position is for recent law school graduates or for those with experience at another firm. She finds a larger share of women for job postings with more stereotypically feminine criteria. In turn, the employer-induced gender stereotyping in job ads may also contribute to gender composition of new hires when gendered language and embedded schemas influence job seekers' application decisions.

Gendered Expectations about the Job-Match. The implicit communication of gender preferences through job ads influence the ways in which job seekers interpret the gendered norms of the job. This early interaction allows job seekers to form expectations about the success of the search

(and potential success at the job) based on culturally shared gender beliefs. Expectation states theory posits that cultural beliefs about the differences in the competence of women and men allow actors to form unequal performance expectations of one actor compared to another in social relational settings

(Berger, Cohen, and Zelditch 1972). These biased expectations shape gender differentiated behaviors,

97 as they assess their own competencies based on the anticipated evaluations by others (Ridgeway and

Smith-Lovin 1999; Wagner and Berger 1997, 2002). Correll (2004) demonstrates that such self- assessments can differentially influence the career choice processes for women and men; she finds in a lab experiment that when participants were led to believe men on average are better at a particular task, male participants perceived their ability to be higher than female participants did and had higher aspirations for completing such task. Similarly, in a survey of job-seeking MBA students, Barbulescu and Bidwell (2013) find that female job seekers were less likely to expect to get offers from jobs such as finance that are considered stereotypically masculine, and that this expectation reduced their propensity to apply for the job.

When language of job ads shapes stereotypicality of jobs, we can then expect job seekers to develop expectations about their success at the job search and such expectations to guide application decisions. Focusing specifically on the use of stereotypical words in job descriptions, Gaucher et al.

(2011) uses a lab experiment to find some evidence that potential job seekers find jobs less appealing when the job description used words that stereotypically describe the opposite gender. While limited to a handful of gender-typed occupations, this work suggests that job seekers of different genders may not only react differently to how stereotypically gendered the job itself is, but also how gendered the language in the job description is. This is important because job seekers may simply be more attracted to jobs that have histories of being dominated by workers of their own gender, which would imply that their gendered expectations are only influenced by the socially constructed norm of the occupation itself rather than the norm (schemas) that organizations seek to display. At the core of our argument is the distinct role of organizations in revealing their stereotyped preferences to influence the ways in which job seekers select into occupations.

We should then expect the wording in job advertisements used by recruiters to influence gendered expectations ofjob seekers about the success of a job application, even when we holdjobs and industriesconstant. While we acknowledge that the use of both feminine and masculine words in job descriptions may make the job more appealing to any job seeker (ultimately increasing job

98 seekers' chances of applying for the job regardless of their gender), here we are interested in assessing

the differential effect of the gendered language on female and male job seekers' job search behavior.

This gendered language mechanism in the interactional process ofjob matching suggests that gender

conformity between the job seeker and the language used in the job ad would increase job seekers'

interest in the job. So in the context ofjob ads, our first and preliminary prediction is that compared to

malejob seekers, female job seekers would be more attractedto a given job when recruiters use more

(less) stereotypicallyfeminine (masculine) words to describe it, ceterisparibus (Proposition 1 a).

We then look at whether stereotypically gendered language has any influence on the job

seekers' actual application to the job in real-world settings. Due to the limitations of lab experiments in previous studies, there has been little understanding of how language, independent of the gender- dominance of an occupation, influences the real application to the job. The gendered language mechanism suggests that the gender conformity between the job seeker and the language used in the job ad would increase job seekers' real application to the job. So in the context ofjob ads, our second main prediction is that compared to male (female) job seekers, female (male) job seekers would apply more to a given job when recruiters use more stereotypicallyfeminine (masculine) words to describe it, ceterisparibus(Proposition lb).

3.2.2 In-Group Preference between Job Seeker and Recruiter

An additional supply-side interactional mechanism potentially accounting for gender segregation in labor markets is the job seeker preferences toward communicating with those that are similar to them (members of their in-group). Because job seekers hold a more positive view of, and feel more connected to, members of their own social group (Perdue et al. 1990), they may also be influenced by the gender of the recruiter, as likely one of their first individual contacts with the hiring organization. Along these lines, the homophily principle in sociology literature suggests that individuals prefer to interact with and trust members of their own demographic group (Brewer and Brown 1998;

McPherson et al. 2001; for a recent review, see an online appendix to Greenberg and Mollick 2017).

99 Studies have stressed the importance of such in-group preferences in predicting key career outcomes

such as hiring and promotion of women (see, e.g., Gorman 2005; Gorman and Kmec 2009) and gender wage gap (see, e.g., Cohen and Huffman 2007; for evidence against this argument, see Maume 2011;

Srivastava and Sherman 2015). In some of these studies, recruiters' gender made a difference in

steering job seekers of the same gender into their available jobs (see, e.g., Fernandez and Sosa 2005;

Gorman 2005, 2006).

Further, preference towards those in the same social group is seen to intensity the effect of

gender during a social network-based job search (Stoll, Raphael, and Holzer 2004). Stoll and

colleagues, in their study ofjob seekers' racial homophily, posits that for black job seekers, seeing black recruiters in positions of authority could be a signal that there might be less discrimination in

hiring at the organization. Moreover, it could also be a signal that there are more promotion

opportunities at the hiring firm with a less hostile environment towards blacks. Similarly, when

gender is salient during the job-matching process as employers and job seekers interact, job seekers

may perceive success at the search when working with recruiters of their in-group. Pre-existing

research found mixed evidence on the role of recruiter-applicant gender conformity when job seekers

were surveyed post face-to-face interviews: while job seeker interviews with recruiters of their own

gender were positively related to job attractiveness (Turban and Dougherty 1992), such interaction did

not result in a positive fit perception with the organization (Cable and Judge 1996). Interview

processes, however, make it difficult to tease out the effect of non-gender related attraction that are,

for instance, based on cultural similarities such as tastes, experiences, leisure pursuits, and self-

presentation styles (Rivera 2012). We focus on the pre-application stage of when job seekers first find

out about the gender of the recruiter, which, to our knowledge, has largely been left out in the studies

of recruiter-applicant gender conformity.

In the context of online job ads, recruiters often make their gender known to job seekers

through pictures or clearly gendered names before any interactions occur. In these instances, we

would expect that the gender of the recruiter may trigger a gender preference by job seekers, even

100 when we hold jobs, industries, and the language of the job advertisements constant. Again, because we are interested in assessing the differential effect of the gender of the recruiter on female versus male job seekers' job search behavior, our prediction is that compared to malejob seekers, female job seekers would be more (less) attractedto a given job when the recruiter'sgender is female (male)

(Proposition 2a). Further, the in-group preference mechanism suggests that gender conformity between the job seeker and the recruiter would increase job seekers' actual application to the job. So our prediction is that compared to male job seekers, female job seekers would apply more (less) to a givenjob when the recruiter'sgender is female (male) (Proposition 2b).

3.3 Study 1 - Field Study of Online Job Postings

3.3.1 Study Design of Study 1

To test our two sets of propositions, we first analyze a unique dataset from one web-based recruiting platform where job seekers frequently learn about job opportunities posted online by recruiters on behalf of hiring organizations. For the purpose of this field study, we collected a large dataset from a global company, RecruitCo. Based in North America, RecruitCo provides a web-based recruiting platform to recruiters and job seekers worldwide. When recruiters from staffing agencies or corporations post jobs on this company's recruiting platform, these jobs are automatically syndicated to online job search boards-for example, SimplyHired and Indeed-and manually broadcasted to social networking services-for example, Twitter and Facebook-if recruiters choose to do so. Job seekers arrive at the RecruitCo recruiting platform through these "job channels" and proceed to the job application stage within the same platform.

Two features of the RecruitCo recruiting platform are worth noting here. First, recruiters have discretion not only to post specific job ads but also to personalize the job search experience for job seekers by presenting themselves as professional yet approachable recruiters. Recruiters can also post their profile pictures and names, which likely reveal their gender. Further, they can draft personalized comments about themselves, the job, or the company they work/recruit for (e.g., "I have five years of

101 experience as an HR specialist. My goal is to find top talent for our company" or "I am a very positive and proactive IT Recruiter looking to find top talent at our IT consulting firm").

A second important feature of the RecruitCo recruiting platform is that it enables job seekers to directly inquire about a specific job posted by a particular recruiter before they formally apply for the position. Indeed, the RecruitCo platform is designed to easily allow job seekers to contact the recruiter directly (and quickly) by simply hitting the "Inquire Here" button, which opens a page where the job-seeker can enter information into a message box that is sent directly to the recruiter. These inquiries range from a short, simple question about the content of the job (e.g., "What types of product would I be selling in this position?") to a lengthy message revealing an intention to apply (e.g., "I believe I am qualified for the position for [XYZ] reasons and would like to learn more about it"). The inquiries from RecruitCo are then an important real-world measure of job seeker attraction to the job.

Consequently, this online setup offers a rare research opportunity to examine the recruitment process at typically the earliest point of recruiter-job seeker contact. Figure 2 illustrates what job seekers see online when using RecruitCo's online platform. As seen in Figure 2, job seekers also have the option of submitting a resume when inquiring about a job, which, when combined with a formal message, would look similar to a traditional online application.

[Figure 2 about Here]

The nature of the data we collected and coded from RecruitCo3 2 allows us to assess our first proposition by testing whether the odds of afemale job seeker (versus a malejob seeker) inquiring about-orapplying to-an otherwise identicaljob increase when there are more stereotypically feminine than masculine words in job advertisements, ceteris paribus (Hypothesis 1 in Study 1). We compute the net proportion of stereotypically feminine words by calculating the difference between

32 As explained later, our sample includes the population of job seekers who inquired about a particular job during our study period. We therefore are able to analyze the odds of a female job seeker inquiring versus a male job seeker, as our dependent variable. Study 2 uses a different dependent variable given that were able to collect unique data from a population of field experiment participants at risk of inquiring about/applying for our real job posted online. More detail is provided later.

102 the share of stereotypically feminine words and the share of stereotypically masculine words used

when describing real jobs based on prior research (as compiled by Gaucher et al. 2011). To assess our

second proposition, we test whether the odds of afemale job seeker (versus a male job seeker)

inquiringabout--or applying to--an otherwise identicaljob is higher when the recruiterposting the

job isfemale, ceterisparibus (Hypothesis 2 in Study 1). More detail is provided in the next sub-

sections.

3.3.2 Data of Study 1

RecruitCo shared with us detailed information on the population ofjobs recruiters posted on

their online platform and whether job seekers inquired about them between October 2010 and August

2012. Since our study looks at how the gendered language ofjob postings may influence an

applicant's decision to inquire about a job, we only study jobs that have at least one job inquirer and

recruiters who posted at least one job during our period of study. In addition, we restrict our analysis

to jobs that are both located in the United States and posted by U.S.-based companies.33 This allows

us to only study companies operating in an institutional environment that prohibits the use of

discriminatory language when hiring employees (see Kuhn and Shen 2013). We also select these jobs

to avoid any potential cultural differences in the use of language (see Maass et al. 2006). Further, we

exclude job ads that have fewer than ten words in their descriptions to conservatively filter out jobs

with incomplete information (that is, approximately one percent of the entire population of jobs).

Our final dataset contains information for a full sample of 150,270 jobs posted by 24,146

employers and recruiters from corporate or staffing firms in the U.S. between October 2010 and

August 2012 (henceforth referred as "full sample"). These jobs generated 566,409 job seeker inquiries

during that same period." Recruiters posted on average 6.2 jobs (ranging from 1 to 1,509), and

" We include jobs located in the United States even when posted by companies in unknown locations and also jobs without location information when posted by companies in the United States. 1 This full sample of job seeker inquiries is about 62.6 percent of all the inquiries included in the original data provided by RecruitCo. The original RecruitCo data consisted of the entire population of 423,270 jobs that were posted by 101,565 recruiters from 100,990 corporate or staffing firms in the world (mostly from countries like

103 received on average 3.8 job inquiries (ranging from 1 to 1,699). Each job posting had an average of

3.8 inquiries.

Job Seeker Characteristics. One major challenge in using the RecruitCo data is that we are unable to identify the number of unique job seekers applying for specific jobs of interest.35 As a

result, the main dependent variable of this field study takes the value of 1 if the job seeker who

inquired about a specific job is female, and 0 if male. Because 56.3 percent of all job inquiries

included the job seekers' resume, we analyze this particular sub-sample of inquiries as those who have

made an application decision. On the RecruitCo online platform, job seeker inquiries resembled

traditional job applications when job seekers attached resumes-these job seekers chose to reveal

information about their job qualifications without conversing with recruiters first. We therefore used

the submission of resumes as a proxy for the job seeker's application to the job.3"

We used the GPeters database-a program that uses Google to analyze common patterns for

first names and calculate whether the names are popularly used for a male or a female person-to code

the gender ofjob seekers using their first names (the only names available to us in the data set, for

confidentiality reasons).3 7 Key descriptive statistics are reported in Table 1. Gender coding was

conducted only for those who decided to disclose their names and those who disclosed identifiable

names (approximately 86.2% of all 566,409 job seeker inquiries in our full sample). 56.2% ofjob

seekers are male, 30% are female; for 13.8% of all job seekers, the gender is unknown.

the United States, India, and the United Kingdom) between October 2010 and August 2012. RecruitCo introduced the online job posting platform at the beginning of October 2010. There were 904,609 job inquiries in total in the original data. " Unfortunately, because we could not identify unique job seekers in our data, we could not cluster errors by job seeker id nor include job seeker fixed effects when estimating our models. We are able to address this limitation in our field experiment (Study 2). 36 Unfortunately, RecruitCo did not have reliable data on who eventually applied to (and got) the job after interacting with the employers. That information is only known to the RecruitCo-client employers. 37 See http://www.gpeters.com/names/. The GPeters database has been used in studies including Flory, Leibbrandt, and List (2015) and Leibbrandt and List (2014), among others. The GPeters database has been used in studies including, among others. As an additional check, they use data from the Social Security Administration (SSA) to calculate a weighted gender probability for each name. Their check indicates that for names that appeared in both databases, the gender ratios from the GPeters database closely resemble the gender ratios from the SSA database.

104 [Study 1 - Table 1 about Here]

We also have data on how these job seekers first found information about the jobs on the recruiting platform (that is, job channels). Only 0.02% ofjob inquiries were made by seekers who came directly to the platform to find out about the job, whereas 37.6% came from popular social networking sites such as LinkedIn, Twitter, and Facebook. Almost 60% of job inquiries came from job boards, including SimplyHired and Indeed, and other sources such as emails or the Google search engine.

Recruiter Characteristics. Using the same gender coding procedure for job seekers, we coded the gender of the recruiters who disclosed identifiable names (76.5% of all 24,146 recruiters in our full sample). The first names of recruiters were unavailable for coding if a recruiter posted an email address or the company's name instead. The gender distribution of recruiters is the following: 37.8% of recruiters are male and 38.7% of recruiters are female.3

Job Characteristics. To measure the extent to which the job advertisements contain what has been termed in prior research "stereotypically feminine" or "masculine" words, we conducted content analysis of all 150,270 job ads (that is, our full sample) posted on the recruiting platform. Following

Gaucher et al. (2011), we use their compilation of masculine and feminine words taken from five separate empirical studies (that is, Bartz and Lydon 2004; Bem 1974; Hoffman and Hurst 1990;

Rudman and Kilianski 2000; Schullo and Alperson 1984). The stereotypically masculine words include agentic (e.g., individualistic, competitive) and masculine trait (e.g., ambitious, assertive) words, whereas the stereotypically feminine words include communal (e.g., supportive, interdependent) and feminine trait (e.g., compassionate, understanding) words (for a similar approach, see Madera, Hebl, and Martin 2009).

38 More than half of the recruiters (56.2%) posted their pictures (or, very rarely, pictures that represent their companies, such as company logos), which are visible to job seekers. Pictures were also posted by recruiters whose gender was unidentifiable (53.8% out of all gender-unidentified recruiters) or missing (33.5% out of all recruiters with missing first names). This means that the job seekers may still be able to identify the recruiters' gender by looking at their pictures and their names. Regrettably, for confidentiality reasons, RecruitCo did not allow us to access any pictures.

105 As summarized in Appendix Table 1, there are in total 42 masculine traits and 39 feminine traits in this literature-not counting those that share the same first few letters as listed (for example, self-confidence and self-confident is counted as one characteristic). Based on these categories, we used the content analysis software called Linguistic Inquiry and Word Count (LIWC; see, e.g.,

Harrison and Dossinger 2017; Pennebaker, Booth, and Francis 2007) to calculate the percentage of stereotypically masculine and feminine words in each job advertisement in our dataset. Used widely by social psychologists, LIWC allows users to create a custom dictionary to count the number of specified words divided by the total number of words in the text (see Harrison and Dossinger 2017;

Neuendorf 2011 for a review of studies using LIWC). To illustrate how this coding works using

LIWC, consider the following example of a job advertisement:

"We are looking for a committed financial assistant with 3-7 years of experience. Candidate must possess strong interpersonal skills and polished oral and written communication skills. Ability to work independently, manage and set priorities. We offer excellent benefits and work-life balance."

In this job posting, once again following Gaucher et al. (2011), we see two stereotypically feminine words ("committed" and "interpersonal") and one stereotypically masculine word ("independently").

LIWC allows generating two variables that represent the share of feminine words (number of feminine words/total number of words = 2/39 = 0.0513; 5.1%) and masculine words (1/39 = 0.0256; 2.6%). In real-world settings, job postings are typically longer and more detailed, leading to a small mean (but high variance) in the share of feminine or masculine criteria. To further illustrate how this coding works, other examples ofjob postings and the values for the two key variables are provided in

Appendix Table 2.

Panel C of Table 1 shows that on average, the share of stereotypically feminine words is 1%

(ranging from 0 to 15.4%) and the share of stereotypically masculine words is 0.98% (ranging from 0 to 16.7%) in our full sample of 150,270 job ads. For 24% ofjob ads, there are no stereotypically feminine words; while 26.6% have no stereotypically masculine words. For 13.6%, there are no stereotypically feminine nor masculine words (below, we refer to these as neutral job postings).

106 The main independent variable in our field study is the net proportion of stereotypically feminine words in the job advertisement.3 9 This net proportion measure is computed by calculating the difference between the share of stereotypically feminine words and the share of stereotypically masculine words in a given job advertisement, so that positive values reflect stereotypically feminine wording, and negative values reflect stereotypically masculine wording. We use the share of stereotypically masculine/feminine words in our analyses to adjust for the length of the job postings when measuring the salience of the gendered wording of the job posting.4 0 In our full sample, the average net proportion of stereotypically feminine words is 0.03% (ranging from -16.7% to 15.4%).

To account for the gender neutrality of job postings, we included a neutral job posting dummy variable where jobs with no stereotypically feminine nor masculine words are coded as 1 (0 otherwise); our findings do not change substantively whether we include or exclude such dummy variable in our regression models.

The above job ad example also shows that companies often choose to disclose information about their organizational policies, especially those relating to work-life balance. To account for potential gender differences in the effect such work-life balance wording may have on the dependent variable of interest, we created a work-life balance dummy variable where the jobs are coded as 1 if the job description contained the following words: great (or fantastic, friendly, or amazing) work (or working) environment, work-culture, work-life, work-family, paid leave, (day), maternity

(or paternity) leave, family friendly, flexible, flexibility, and dependent care (0 otherwise). Out of all jobs, 10.8% advertised in our full sample (16,292) emphasize such company work-life balance policies.

3 Substantially similar results are found when including the share of stereotypically masculine words and the share of stereotypically feminine words as two separate independent variables in our main models (instead of including the net proportion of feminine words); these additional analyses are available upon request. For simplicity, we only report the analyses including the net proportion of stereotypically feminine words in the job posting. 40 The salience of one stereotyped word in a job posting that has 500 words, for example, would be very different from one in a job posting that has 50 words.

107 Our regression models also control for the type of employment: In particular, whether jobs are full-time, part-time, contract, contract-to-hire, temporary, or intern jobs. The majority of jobs are full- time (60.5%) while a major proportion are contract jobs (25%); this large number of advertised contract jobs makes sense considering the share of staffing firms using the RecruitCo online search platform. These jobs are categorized into 53 different industries, with the top five industries being technology (19.5%), healthcare (8.8%), finance and banking (6.9%), computer software (5.3%), and manufacturing (4.8%). The top five states where these jobs are located are California (12.7%), Texas

(11.4%), New York (9.8%), Illinois (6.0%), and Massachusetts (5.6%). For additional information, see Table 1, Panel C.

Additional Data Collection Efforts/Coding. While the data on the industry and the state where these jobs are located were provided by RecruitCo, such data had limited information on the occupations these jobs belonged to. To control for occupation fixed effects in our models, we used available data on job titles manually entered by recruiters and categorized/coded them into occupations using the 2010 Standard Occupational Classification (SOC) provided by the Bureau of Labor Statistics

(BLS) (for a similar approach, see Pager and Pedulla 2015).41 Because these job titles usually target specific job seekers and provide very fine-grained detail, we used the BLS's most detailed categorization of jobs, which yielded 840 occupations.4 2 This categorization process led to a total of

6,593 SOCPC-defined direct match titles that are categorized into the 840 SOC occupations.

We matched job titles from the dataset to the SOCPC-defmed direct match titles and coded

41 Pager and Pedulla (2015) use a version of the SOC codes released in 2000 and code the job titles using their first and second tier classification systems, which reflect 23 major occupational groups (tier one) and a subdivision of 96 minor groups (tier two). We use the 2010 SOC code's fourth tier classification system, which contains 840 fined-grained occupations. 42 This categorization process was facilitated by the Direct Match Title File created by the SOC Policy Committee at the BLS in 2000 to address the issue of an occupation potentially having multiple associated job titles (Cosca and Emmel 2010). The Direct Match Title File provides examples of job titles for each occupation that can only be categorized by one occupation. This prevents job titles such as "painter" from being categorized into two separate SOC occupations: "fine artists, including painters, sculptors, and illustrators" and "painters, construction and maintenance." Instead, the direct match for the "painters, construction and maintenance" occupation includes bridge painter, facilities painter, and industrial painter, among others, providing a more elaborate way of categorizing occupations.

108 them based on whether they match exactly or whether RecruitCo job titles included exact job titles from SOCPC titles. For example, we coded job titles that said "financial officers" and "financial officers needed immediately" into an SOC occupational category of "financial managers" (2010 SOC code: 11-3031) as one of its direct matches for the particular category includes "financial officer."

Using this matching scheme, we were able to assign 840 SOC occupations to a smaller sample (sub- sample A) of 34,460 jobs that were posted by 10,526 recruiters and had 146,597 job seeker inquiries during our study period. This covers about a quarter of the full sample (22.9% of all 150,270 jobs).

The top five most frequent occupations based on the SOC categorization are "management analysts"

(11.2%), "accountants and auditors" (7.8%), "software developers for applications and systems software" (6.3%), "data entry keyers" (5.0%), and "web developers" (3.8%).

In our analyses reported below, we use both the full sample (150,270 jobs) without the occupational categorization and the smaller sample of jobs with the occupational categorization as elaborated above (34,460 jobs for sub-sample A). The purpose behind analyzing the smaller sample is to examine whether the gendered language mechanism holds even after controlling for occupational fixed effects (that is, variations in how gender-dominated an occupation originally is). Occupational categorization thus allows us to tease out the effect of occupation-specific idiosyncrasies from the effect of the gendered wording of the job ad on the job seeker behavior. When the main descriptive statistics for this sub-sample A are compared with those of the full sample without the occupational information, we can observe that sub-sample A is very similar to the full sample (see Table 1). The only major difference we observe is that the percent of female job seekers is slightly larger (35.6%) in sub-sample A when compared to the full sample (29.9%). We then further match 840 SOC occupations to the 2010 Census occupational classification, which has 539 detailed occupations, using a crosswalk provided by the U.S. Census Bureau to create a variable for the actual distribution of gender across all the detailed occupations from their 2012 Current Population Survey (CPS) March

109 Annual Social and Economic Supplement (ASEC) data."

3.3.3 Results of Study 1

The GenderedLanguage of Job Advertisements

To test the gendered language mechanism, we estimated several logistic regression models predicting the odds that job seekers who inquire about jobs are female (rather than male) depending on the net proportion of stereotypically feminine words in the job ads, ceteris paribus. Those without gender information, as well as those that inquired about job postings without industry information are dropped for this particular analysis, resulting in 482,771 job seeker inquiries (85.2% of the full sample) and 271,035 job seeker applications with resumes attached (47.9% of the full sample). 4

[Later in this section, we discuss additional estimated models, whose results are consistent with the ones presented here.]

Table 2 presents logistic regression models with our main dependent variable. In Columns 1-

3, we analyze all "job inquiries" that job postings attracted, including more serious inquiries from job seekers who also sent resumes to the recruiter when inquiring about the job: Once again, the dependent variable takes the value of 1 if the gender of the job seeker is female and 0 if male. In

Columns 4-6, we analyze the sub-sample ofjob inquiries with attached resumes, that is, all "job applications": In these models, the dependent variable now takes the value of 1 if the job seeker submitting a resume is female and 0 if male.

Models 1 and 2 (Columns 1-2 and 4-5) of Table 2 present logistic models using thefull sample. All models control for individual and job features such as employment type, company type, and the channels from which jobs seekers obtained information online. We also include state and

4 To learn more about the average number of job inquiries for most female/male/gender-integrated occupations in our full sample and the share of female/male job seekers for each occupation in comparison to those in the 2012 CPS March ASEC data, see Appendix Table 3. 44 We also ran a set of multinomial models with a dependent variable that takes a value of 0 if the job seeker gender is male, 1 if female, and 2 if unknown/unavailable and found substantially similar results to the ones presented here (available upon request).

110 industry fixed effects to account for state and industry heterogeneity in the way that job postings are worded and the types ofjobs and workforce that employers are looking for. As shown in both Models

1 (Columns 1 and 4), we find support for Hypothesis 1 with a positive and significant coefficient for the variable "net proportion of stereotypically feminine words" (that is, the difference in the share of stereotypically feminine and masculine words): The odds that a job seeker who inquired about a job is a woman increases as the net proportion of stereotypically feminine words used in job ads increases

(P=0.047, p<0.001; see Column 1 of Table 2 under the "Job Inquiries" sub-heading). In addition, the odds that a job seeker who attached a resume when inquiring about a job is a woman also increase as the net proportion of stereotypically feminine words used in job ads increases (0=0.048, p<0.001; see

Column 4 of Table 2).

[Study 1 - Table 2 about Here]

Each Model 1 (Columns 1 and 4) also allows for testing whether the odds that a job seeker who inquired about a job is female (rather than male) are higher when job ads have explicit language regarding flexibility and family-friendly employer policies. Even after controlling for employment type, we find that the odds of a female job seeker inquiring about a job increase when job ads show that the employer supports work-life balance arrangements (p<0.001).

In-Group Preference between Job Seekers and Recruiters

To test the in-group preference mechanism, our additional logistic models include two dummy variables that account for the gender of the recruiter: One variable takes the value of 1 if the recruiter is female (0 otherwise). A second variable takes the value of 1 if the gender of the recruiter is unknown (0 otherwise). So after the inclusion of these two variables, the reference category in our models is male recruiter.

As shown in Model 2 (Column 2), the odds of a female job seeker (rather than a male job seeker) inquiring about a job increase when the job is posted by a female recruiter compared to when the job is posted by a male recruiter (P=0.234, p<0.001; see Column 2 of Table 2 under the "Job

Inquiries" sub-heading). Also note that the odds of a female job seeker inquiring about a job increase

111 when the job is posted by a recruiter whose gender identity is unknown or whose name was not gender-identifiable (oftentimes, this involves a name of foreign origin) compared to a male recruiter

(0=0.050, p<0.001; see Column 2 of Table 2). The main findings remain substantively the same when analyzing job applications (see Column 5 of Table 2 under the "Job Applications" sub-heading).

Alternative Key Explanations andAnalyses

In this sub-section, we address potential alternative explanations that may account for the results supporting our propositions. We also provide additional analyses aimed at ensuring the robustness of our findings. First, we investigated whether the results from Models 1 and 2 of Table 2

still hold when we introduce occupation fixed effects (to control for occupational-level gender

segregating factors). One may suspect, for example, that female- (or male-) dominated jobs on average are simply more likely to receive a higher number of applications from female (or male) job

seekers. If job seekers were driven only by the gender dominance of a job, it would make it difficult to tease out the role of organizations in portraying their schemas in job advertisements. To account for this possibility, as explained earlier, we worked with a smaller sample that allows us to use occupation

as a variable (sub-sample A).45

In both Models 3 (Columns 3 and 6) of Table 2, we repeated the same analyses using

occupation fixed effects in addition to all the control variables included in Models 2. We find that the

significant positive effect of the difference in the share of stereotypically feminine and masculine

words in job postings on the odds of the job seeker being female remains, all being equal (P=0.0 15, p<0.01; see Column 3 of Table 2 under the "Job Inquiries" sub-heading). We therefore continue to

find support for Hypothesis 1 in our occupation fixed effect models: Even when we control for

occupational heterogeneity, the odds of female job seekers responding to job postings increase with a

higher net proportion of stereotypically feminine words. In addition, we continue to find support for

Hypothesis 2: The female recruiter effects remain substantially the same (p=0.077, p<0.001; see

" We replicated Models 1 and 2 of Table 2 using sub-sample A and found that our results remain substantially similar.

112 Column 3 of Table 2). Similarly, the odds of female job seekers applying to jobs increase with a

higher net proportion of stereotypically feminine words (P=0.016, p<0.05; see Column 6 of Table 2

under the "Job Applications" sub-heading) and when the job is posted by female recruiter (0=0.077,

p<0.001; see Column 6 of Table 2).

Second, although occupation fixed effects control for unobserved differences across

occupations that may affect the gendered pool ofjob seekers, they do not account for potential gender

differences within occupations-that is, different gender inquiry and application patterns as a result of differences in skill level or complexity of the particular occupations under study. It could be possible that, for example, the less skilled or complex a job is, the larger the share of female job seekers applying for such a job (Kuhn and Shen 2013). This would indicate that the hierarchical gender status beliefs that men are more competent than women (see, e.g., Gorman 2006; Ridgeway 2011) may influence application behaviors by job seekers, independent of the intentions of the hiring organization.

To test whether within-occupation variation may influence job seekers' choices, we created a narrower sub-occupational category that is more detailed than the occupational category included in sub-sample A. Often, RecruitCo job titles indicated the level of skills or complexity required for the job (i.e., words such as "assistant," "associate," and "senior"). The skill-level information was not captured in the occupational category of sub-sample A, since the SOCPC-defined direct match title file used for creating the category does not make such a distinction. We manually identified these cases and coded them into a sub-occupational category. For example, occupations under an SOC- matched "food service manager" category are now further divided into "assistant restaurant manager,"

"assistant restaurant general manager," "assistant senior restaurant manager," "entry level restaurant manager," "restaurant manager," "senior restaurant manager," "food and beverage manager," etc. Due to the difficulties of manually coding all 840 occupational categories into different skill levels, we only coded gender-integrated occupations where the share of female workers is between 45 and 55 percent according to the 2012 March Current Population Survey (CPS). This helps us to run a

113 conservative test of our propositions since the salience of gender is expected to be less prominent than in female-dominated or male-dominated occupations (Ridgeway 2011) (See Appendix Table 3 for the list of gender-integrated occupations in the full sample). This resulted in a smaller sub-sample of

16,432 job seekers (sub-sample B).

Table 3 reports the results of the logistic regression models using sub-sample B. To check the robustness of our new sub-sample, both Models 1 in Table 3 have the same specification as Models 2 in Table 2 while both Models 2 in Table 3 have the same specification as Models 3 in Table 2. In both cases, the coefficient for the net proportion of stereotypically feminine words predicting the odds that the gender of the job seeker inquiring is female remains positive and significant (p<0.00 1 for Model 1; p<0.0I for Model 2 under the "Job Inquiries" sub-heading). Finally, as shown in Models 3, even after controlling for detailed-occupation fixed effects, the coefficient for the net proportion of stereotypically feminine words remains positive and significant (P=0.043, p<0.01; see Column 3 of

Table 3 under the "Job Inquiries" sub-heading). Table 3 also provides support for Hypothesis 2

(P=0.132, p<0.01; see Column 3 of Table 3 under the "Job Inquiries" sub-heading). The same holds true for the likelihood of applying to the job-despite the decline in statistical significance, the results remain positive and significant (p<0.05 for both Hypotheses 1 and 2; see Column 6 under the "Job

Applications" sub-heading).

[Study 1 - Table 3 about Here]

Finally, we have chosen to report the models where individual inquiries and applications are the unit of analysis. That said, in an additional set of multivariate models, we also used the job ads as the unit of analysis to directly study the impact of the gendered language of the job posting on the proportion of inquiries coming from women versus men. In a particular set of models, for example, we used the proportion of female job seekers who inquired about the job as our main dependent variable and estimated a series of fractional logit models proposed by (Papke and Wooldridge 1996).

The results of these models remain substantively similar to the results from the logistic models presented here for both Hypotheses 1 and 2: Concerning Hypothesis 1, we find that the net proportion

114 of stereotypically feminine words predicting the proportion of female job seekers is positive and significant (0=0.018; p<0.001). Concerning Hypothesis 2, the proportion of female job seekers also significantly increases when the job is posted by a female recruiter compared to when the job is posted by a male recruiter (0=0. 165, p

Within the limits of Study 1's (non-experimental) design, the patterns in our analyses provide some evidence in support of the two distinct mechanisms behind the decision to inquire about a job-the gendered language and the in-group preference mechanisms in the context of online job advertisements. Our results stress that the (consciously or unconsciously) gendered words recruiters use to advertise jobs differently affect female and male job seekers' inquiry and application patterns for otherwise identical jobs. They also show that the odds of the job seeker being female are higher when the recruiter is female compared to when the recruiter is male, ceterisparibus.

There are three main research limitations of our field study (Study 1), however. First, in our population of job inquiries and applications, we can only study changes in whether the interest for a particular job is coming from female job seekers versus male job seekers. In this regard, the positive and significant coefficient for the net proportion of feminine words means that an additional feminine word either increases the chances of a job inquiry and application being made by a female job seeker and/or decreases the chances of a job inquiry and application being made by a male job seeker. Given the characteristics of the data set, regrettably, we cannot differentiate between these two potential gender segregating outcomes. Second, in Study 1, we are only able to analyze job seekers who inquired about and applied to the job-that is, we had no access to those individuals who may have seen the job posting but decided not to take action. [This is why our main dependent variable in Study

46 While the results do not change, these models should be interpreted with caution, as we do not know the number of male or female job seekers who read the job advertisement but did not apply. This makes it difficult to compare how female and male job seekers behave differently. For example, it could be the case that we are seeing a positive effect on the number of female job seekers when the net proportion of stereotypically feminine words are greater first simply because fewer female job seekers clicked on that particular job posting to begin with. Second, we may still see the positive effect on the number of female job seekers even when the proportion of female job seekers that applied versus those that did not apply is low, since we do not know who saw the job posting and did not apply. Study 2 was designed to help us address this limitation.

115 1 was the odds that the job seeker who inquires and applies is female.] Without knowing the gender of the job seekers at risk of applying (i.e., who saw the job ad to begin with), our field study results may be biased (Heckman 1976a, 1976b, 1979). We do not know, for example, whether job seekers of any particular gender were more likely to apply to every job posting available. Third, there could still be other unobservable features of the job, company, and recruiter that could affect the chances of inquiring about the job for which we could not control given the available data provided by RecruitCo.

3.4 Study 2 - Field Experiment Study of Online Job Postings

3.4.1 Study Design of Study 2

To address the field study limitations and further investigate the findings of Study 1, we conducted a unique field experiment study that directly tests the magnitude of how gendered words and recruiter gender affect job seeker behavior. Experimental methodologies have been used in real employment settings (Flory, Leibbrandt, and List 2015; Pager and Pedulla 2015). With the goal of assessing discrimination, researchers have used matched pairs of trained "auditors" who pose as applicants in real job application settings (e.g., Pager 2007), while others have sent out fictitious resumes of matched pairs to real employers (see Bertrand and Mullainathan 2004). For the purpose of our study, we developed an innovative field experiment study design that differs from the traditional approach where employers, instead ofjob seekers, become the "auditors" by posting fictitious job postings online. In our study, we therefore acted as a real employer and advertised one realjob online: the "project assistant" position. We chose this particular occupation since it is considered to be a gender-neutral job: according to the American Community Survey in 2012 by the Census Bureau, about 50.69 percent of workers were female in an occupation category, "Miscellaneous life, physical, and social science technicians, including social science research assistants (SOC-code: 1940YY)", that most closely resembles ajob we advertised. When advertising our job online, we manipulated both the use of gendered language to describe it (masculine, feminine, or neutral wording) and the gender

of the recruiter (male or female) posting it (between subjects design).

116 In particular, we conducted a two-stage field experiment study design. At the first stage, we gathered information about potential job seekers (i.e., individuals at risk of inquiring) using an online survey. Participants were paid to complete this survey. At the second stage, once the survey was completed, we prompted the same individuals with the availability of a real job. The advantages of the two-stage field experiment setup are twofold. First, we are able to gather information about job seekers who saw the job posting but did not show interest; this information has largely been absent in the typical one-stage field experiment study design. Second, we are able to manipulate only gendered words and recruiter gender holding all other aspects of the job ad identical for all six (3x2) manipulated conditions. This addresses a key limitation of existing field experiment study designs used in real employment settings, where the matched pairs of trained auditors and fictitious resumes are almost-but not entirely-identical in their non-manipulated characteristics (e.g., identical resumes that vary only in job seeker names cannot be submitted to the same employer).4 7

To recruit potential job seekers, we used Amazon Mechanical Turk (mTurk) and posted a short online survey twice48 for a period of one month. The purpose of the survey at the first stage was to attract individuals who work on tasks posted on mTurk for monetary reward who may or may not be interested in applying for the real project assistant position we advertised right after the survey completion (the second stage). The participants were told that their answers were completely anonymous and that their answers could not be connected to their identity. We paid them for participating in this initial online survey, which took about five minutes on average to complete. The survey asked questions related to their experiences with job searching or organizations (see footnote

47 One notable exception is (Flory et al. 2015) where, in their two-stage study design, they randomly assign job postings that are identical in their non-manipulated features to job seekers who showed interest in their first- stage job ad. Even in this design, however, the information about those who did not select themselves into the pool of interested job seekers is unavailable. 48 The first stage survey initially asked a set of questions related to job search (titled "Job Search Survey"). Due to the potential risk of such survey biasing responses when participants are prompted with a real job, we ran the field experiment the second time, this time asking a number of questions irrelevant to job search. We do not find any substantial difference in the main findings reported here when we analyzed the data separately (analyses are available upon request). In this study, we report results from both rounds of the field experiment while controlling for the timing of the experiment.

117 21), as well as their demographic information. For this survey task, we recruited a sample of 4,150 participants ages 18 to 65 who completed our two surveys over a period of one month (i.e., individuals at risk of applying for our real job).

At the end of the survey, all participants were thanked for taking the time to complete the survey. We wrote: "You will now get paid for your participation." Then we included the mTurk validation code that would ensure that participants got paid and exited the survey. Typically, this validation code signals the successful completion of a given task. However, at the end of the page, we included the logo of our educational institution and asked them the following question: "Looking for a

Real Job?" This question was followed by the below short paragraph (henceforth, the Short Ad):

"Our research team at [OUR SCHOOL NAME] is always looking for project assistants to help us with our research projects on employment and careers. No previous experience is necessary for the job. No relocation is required as work can be completed online. Would you be interested in learning more about the position? If interested, please select "yes" to proceed."

This was the critical part of our field experiment (i.e., the second stage), because all participants could proceed to reading more about the real job and subsequently apply to be considered for the position or leave the site at this point. Immediately following the last paragraph, each participant could choose,

"Yes, tell me more" or "No, please end the survey." Of all 4,150 survey participants, 48.3 percent

(2,003 participants) expressed interest in learning more about the real job and proceeded to the next page where they obtained additional job information. The complete job posting (as seen by the participants) is included in Figure 3.

[Figure 3 about Here]

Key to our experiment design is that all 2,003 participants who agreed to learn more about our project assistant job were then randomly assigned to one of the three gendered wordings of the same job and one of the two genders of the recruiter-that is, a 3x2 experimental design, between subjects.

These six job advertisement conditions are summarized at the bottom of Figure 3. For the feminine- wording manipulation, we used the following seven common words in the main text of the job description: "committed," "supportive," "sensitive," "interpersonal," "cooperatively,"

118 "compassionate," and "honest." For the masculine wording manipulation, we used: "determined,"

"assertive," "outspoken," "analytical," "independently," "decisive," and "persistent."4 9 For the neutral wording manipulation, we did not use any of the above-mentioned words. Instead, the job advertisement simply read as follows (the Long Ad):

Job Description Our research team at the [OUR SCHOOL] is looking for a candidate willing to gain experience as a research assistant. An ideal candidate should care about the success of our projects.

Your main task will be to provide assistance sorting and inputting text data that will be used for a research project; to help coordinate project logistics; to help launch research surveys; and to perform other duties for the team.

Job Requirements No previous experience required as we will provide training. High school degree is required. Work can be completed online. No relocation is required.

Candidate must possess polished oral and written communication skills. Must be able to demonstrate the ability to manage assignments and set priorities. Anyone with superior organizational skills is encouraged to apply.

For more information, contact us at [RECRUITER EMAIL]. We will answer any questions you may have.

In order to manipulate the gender of the recruiter, we used the name "John" for the male recruiter condition, and the name of "Jennifer" for the female recruiter condition. Once the participants were randomly assigned to one of the six versions of the job ad (3x2), they were asked to choose among the following options: "Send me more information about the position," "Apply," or "Not interested." The participants who chose one of the first two options then could further communicate with the recruiter and/or apply for the position (specifically, they were given the option of providing their name, email address, resume, and any messages to the recruiter) whereas those who chose "Not interested" did not proceed any further. Using this information, we are able to measure both the attraction to the job and an actual application to the job.

" We decided to include seven gendered words in both the masculine wording and feminine wording conditions with the goal of being as realistic as possible in the description of our real job. Future research should consider altering the number (and type) of gendered words included in the job posting.

119 To assess the first proposition in our field experiment (Study 2), we specifically test whether compared to male job seekers, female job seekers are more likely to show interest, and applied to, an otherwise identicaljob when feminine words (ratherthan masculine words) are used to advertise such job, ceteris paribus (Hypothesis 1 in Study 2). To assess our second proposition, we test whether compared to male job seekers, female job seekers are more likely to show interest, and applied to, in an otherwise identicaljob when the recruiterposting the job is female (ratherthan male), ceteris paribus (Hypothesis 2 in Study 2).

3.4.2 Data of Study 2

Of all 4,150 participants who agreed to take the initial online survey, 48.3 percent (2,003 participants) indicated interest in learning more about the project assistant job after having been exposed to the Short Ad. These individuals were then randomly assigned to one of the three gendered wordings of the same job and one of the two genders of the recruiter at the second stage-that is, a

3x2 experimental design, between subjects (summarized in Figure 3).

Job Seeker Behavior and Characteristics. One dependent variable is whether the survey participant is interested in the job after having been exposed to the Long Ad (1=Yes; 0=No). This variable was coded as 1 if the participant, when prompted by a question about the job, chose "Send me more information about the position" or "Apply" and coded as 0 if the participant chose "Not interested." A total of 1,648 individuals (39.7 percent of all participants and 82.3 percent of participants who clicked on the Short Ad) indicated interest in the job after being exposed to the Long

Ad. While a measure of attraction, this variable is interpreted with caution due to the high number of interest the job has generated on Mturk. Our main dependent variable of interest is whether the survey participant applies to the job after having been exposed to the Long Ad, our experimental conditions.

Once participants chose the either "Send me more information about the position" or "Apply", they were directed fill out a real application to the job, including their name, email address, resume, and the . While about 69 percent of the interested participants filled out their email address, only

120 about 27 percent of the interested participants submitted a resume for review. We therefore coded a

variable for job application as 1 if the participant has submitted a resume, which is indicative of their

seriousness about applying to the job.

We also coded the gender of the participant: 48.6 percent of all participants were female, and

52.2 percent of those who indicated interest in the job were female. Participants had on average 11.9

years of work experience (12.2 years of work experience for those who indicated interest in the job).

The main descriptive statistics for the sample are included in Table 1 of Study 2.

[Study 2 - Table 1 about Here]

Manipulation 1: GenderedLanguage of the Job Posting. To test Hypothesis 1, we manipulated the wording used to describe the project assistant job. Of the 2,003 individuals interested in learning more about the job after reading our Short Ad, 33.9 percent were randomly assigned to the feminine wording of the job, 32.1 percent to the masculine wording, and the rest to the neutral wording.

Manipulation2: Gender of the Recruiter. To test Hypothesis 2, we manipulated the name of the recruiter listed at the end of the job posting. We first reinforced the recruiters' gender by using their first names as part of the email address that job seekers could use to inquire about the job and also by using a gender pronoun for job seekers' inquiries. Of the 2,003 individuals interested in learning more about the job after reading the Short Ad, 50.8 percent were randomly assigned to the female recruiter (Jennifer). The rest were randomly assigned to the male recruiter (John).

3.4.3 Results of Study 2

We start by describing the preliminary bivariate results of our field experiment comparing gender differences by condition for the 1,503 individuals interested in learning more about the job after reading the Long Ad. Since the field experiment allows us to tease out the behaviors of female and male job seekers separately, we are able to examine whether job seekers of each gender respond differently to the same manipulations. Compared to male job seekers, we find that female job seekers

121 are more likely to show interest in the project assistant job when the job is worded using feminine words (86.5% of female job seekers versus 78.6% of male job seekers; p=0.007, two-sided test) or when the recruiter is female (85.9% of female job seekers versus 80.6% of male job seekers; p=0.022, two-sided test). Additionally, because we manipulated both the gendered wording of the same job and the gender of the recruiter at the same time, we also examine whether there are any interaction effects on the probability of being interested in the job. We find that compared to male, female job seekers are more likely to show interest in the job when a female recruiter posts the job using feminine wording (88.4%; p=0.026, two-sided test). We however find partial support of our hypothesis: in general, we find that female job seekers are more interested in the job than male job seekers across all manipulations and male job seekers' interests were not in gender conformity.

We then look at whether there are differences in job seekers' application to the job by gender.

Contrary to the gendered language mechanism (Hypothesis 1), we find that female job seekers applied more to the project assistant job when the job is worded using masculine words (29.4% of female job seekers versus 21.35 of male job seekers; p=0.019). On the other hand, we find support of the in- group preferences mechanism (Hypothesis 2): compared to male, female job seekers are more likely to apply when a female recruiter posts the job (26.9% of female job seekers versus 17.9% of male job seekers; p=0.001, two-sided test). This study also allows us to tease out the behavior of female and male job seekers and examine them separately. We find that female job seekers are less likely to apply when a female recruiter posts the job (26.9%) compared to when a male recruiter posts the job (21.7%; p=0.053); male job seekers are more likely to apply when a male recruiter posts the job (22.4%) compared to when a female recruiter posts the job (17.9%; p=0.075). Because these differences may be due to other key differences in individual job seeker features, we estimated several binary regression models predicting the likelihood that a job seeker applies to the real job, controlling for job

seeker's characteristics, such as age, level of education, marital status, employment status, years of work experience, intentions for job search, student status, and region. Once again, the dependent variable takes the value 1 if the participant applies (0, otherwise). These results are summarized in

122 Table 2. For simplicity, we only report the main coefficients from the models in Table 2 of Study 2.

[Study 2 - Table 2 about Here]

We describe the results of our logistic models estimated using the sample of all participants who clicked on the Short Ad at the completion of the survey (N=2,003). Model 1 first includes the main effects for the gender of the job seeker, the gendered language condition, and the recruiter gender condition: We find that female job seekers are more likely to apply to the job than male job seekers and less likely to apply when the job is describe by using feminine wording compared to masculine wording. However, to directly test our hypotheses on the differential effect of gendered language and in-group preferences on female and job seekers' job search behavior, we are interested in the interaction term between the feminine language condition and the female job seeker variable as well as the interaction term between the female recruiter and the female job seeker.

As shown in Model 2 of Table 2, we find that the feminine wording of the job ad has negative effect for female job seekers than for male job seekers-i.e., there is a significant interaction between the feminine wording of the job ad and the job seeker female dummy variable (0=-0.488; p<0.10, two- sided test). Again, this result is inconsistent with our Hypothesis 1. Relating to the testing of our in- group preference mechanism, Model 2 shows that the gender of the recruiter has a strong effect on the applicant gender-i.e., there is a significant interaction between female recruiter and the female job seeker variable (0=0.603; p

4) to directly test whether job seekers of each gender respond to same manipulation differently.

Model 3 shows that female job seekers are less likely to apply when the job is described using feminine wording and more likely to apply when the job is posted by female job seekers. We do not find any significant results for male job seekers.

In our field experiment study (Study 2), while we find evidence in support of the in-group

123 preference mechanism behind the decision to inquire about a job, we find inconsistent support for the gendered language mechanism. We find that compared to male, female job seekers are more likely to show interest in the job when the job is described using feminine wording, but they are less likely to make a real application decision when the job is described using the feminine wording as opposed to masculine wording. The findings suggest that gender conformity between the job seeker and language would not always lead to an application decision for female job seekers.

There may be several explanations to our findings. First, the way we have introduced words into identical job ads may have influenced the perception of the job and the work. For example, the words "supportive" (feminine) and "assertive" (masculine) may lead workers to believe that a project assistant position to resemble jobs that require more support, such as administrative jobs, or assertion, such as managerial jobs. The words such as "interpersonal" (feminine) and "analytical" (masculine) may also lead workers to expect that the job will equip them with different set of skills relevant for future career aspirations. The job we advertised was "part-time" and 46 percent of the participants who showed an initial interest in a real job opening (clicked on the Short Ad) said they are looking for a part-time job. Since temporary work is frequently viewed as a stepping stone into stable employment (see Autor and Houseman 2010), it may be the case that when it comes to a decision about submitting a real application that requires time commitment, job seekers assess their expectations about the job differently, depending on their gender. Second, although we chose an occupation considered to be gender-neutral in the Census Bureau's broad categorization of

"Miscellaneous life, physical, and social science technicians, including social science research assistants", the occupation may be considered gender-typed. Job seekers may behave differently based on how stereotypical language is used to describe gender-typed jobs. We do not know, for example, whether female job seekers are more interested in applying to female-dominated jobs if the job is described using more stereotypically masculine or feminine wording.

3.5 Limitations and Future Research

124 Our research could be productively extended in several ways. First, in the previous section, we have explained that while the job and pay may be constant, the wording itself may also be meaningful in that it could describe relevant tasks and work contexts. The use of language by recruiters to portray the "ideal worker" can come in different forms. Therefore, when different tasks and work environments are a consideration for a job seeker, he/she may no longer be comparing the different language ofjob ads but also jobs that might be quite different in the tasks and the work environment.

Thus, future research should experiment with other aspects (and operationalization) of language in job postings used by recruiters and employers, specifically when tasks and work contexts are not affected, to further assess their effects on gender sorting. Building from our field experiment, studies that directly test the extent to which the two mechanisms here proposed operate in clearly gender-typed occupations will also be useful in deepening our understanding of the language effect.

Second, related to this, in this study, we have taken the job seekers' perspective to examine how gendered language and the gender of the recruiter may affect their interests in otherwise identical job postings. A promising area of further inquiry consists of studying recruiters to understand their motivations and expectations for engaging in particular kinds of gendered language at the time of describing available jobs (similar to empirical studies aimed at assessing why and when individuals bring employment opportunities to the attention of their friends and acquaintances; see, e.g., Marin

2012; Smith 2005).

Third, with regard to the measurement of our key independent variable, even though we employ prior methods used in the literature for the construction of "masculine" versus "feminine wording" (see Gaucher et al. 2011), we think such measurements could be further improved in future research. One concern is whether the feminine versus masculine words in the job posting have limited face validity, especially when the same word may have a different connotation depending on the context in which it is being used (possibly in a way that may matter for perceptions of femininity and masculinity). Along similar lines, there is the concern that the mere count of the masculine/feminine words in the job description, while efficient given the volume of job postings we had to code as part of

125 this research project (Study 1), may be a simple way of measuring how masculine or feminine a job posting is. In this article, we tried to address these concerns by keeping the job and the context constant in our field experiment (Study 2) by simply altering the posting by using 7 (masculine or feminine) words to describe the same real job. However, one promising future approach is to continue conducting carefully designed field experiment studies in which job seekers are randomly assigned to gendered job postings using an alternative set of masculine and feminine words with controlled differences in meaning and connotations or using other measurement methods.

Fourth, because of our research design and available data, we could not directly examine why job seekers ultimately did (not) apply to certain jobs. In particular, we could not explore the many potential explanations accounting for the finding that female job seekers were more likely show interest when jobs are described using feminine words, including job seekers' expectations relating to perhaps being more hirable, better performers in the job, and/or more successful in such jobs post-hire, for instance (see, e.g., Barbulescu and Bidwell 2013; Cejka and Eagly 1999; Ridgeway and Correll

2004, among many, for the theoretical coverage of such explanations). Also from our field experiment, while we provide potential explanations as to why we find that female job seekers applied less to our job when it was described using feminine words than masculine words, we are not able to make a conclusive statement about our results. We therefore see promise in experimental research distinguishing among some of these potential theoretical explanations behind the language mechanism explored in this study.

Finally, by examining the role of gendered language and the gender of the recruiters who post jobs, this article took a first step at addressing recent calls to study the potential micro-level processes that sustain gender disparities in labor markets (see, e.g., Castilla 2011; Rivera 2015; Stainback,

Tomaskovic-Devey, and Skaggs 2010). While our study finds that the gendered language and the gender of the recruiter/hiring manager influence the gender sorting when inquiring about jobs and applying for jobs, we see promise in further examining the impact of the two proposed mechanisms

(as well as any potential interactions between the two of them) beyond recruitment to other

126 applicant/employee-recruiter/manager interactions, for example, during job candidate interviews, employee performance appraisals, employee training evaluations, and career-related discussions with both internal and external prospective employee recruits-see, for example, research efforts to study language in letters of recommendation (e.g., Schmader, Whitehead, and Wysocki 2007; Trix and

Psenka 2003) or in evaluations (e.g., Correll and Simard 2016; MacNell, Driscoll, and Hunt 2015).

3.6 Discussion

The job-matching process is inherently two-sided and social relational, as both employers and job seekers consistently interact to form impressions and make evaluations of the other throughout the search (Ridgeway 2011). While employer-side interactional mechanisms have been tested before

(Gorman 2005), little understood was whether these mechanisms are experienced and exercised by the job seekers of different gender in such a way to shape their application behavior. We are the first to fill this gap by directly investigating the differential gender effects of the gendered language (as experienced by job seekers) and the in-group preference (as exercised by job seekers) mechanisms on both the job seekers' attraction and application to a real job posted by a recruiter in a real-world setting, ceteris paribus.

Using unique data from both a field study (Study 1) and a field experiment study (Study 2) of online job postings, our findings reveal that the words recruiters of different genders use when posting jobs affect how female and male job seekers behave at the initial recruitment phase. First, in our field study, we find that the higher the net proportion of stereotypically feminine words in the job advertisements is, the higher the odds that a female job seeker (rather than a male job seeker) will inquire about jobs and apply to jobs, ceterisparibus. This is evidence of what we term the gendered language mechanism, especially for female job seekers (Propositions la and Ib). We also find that female job seekers are more attracted to jobs, and apply to jobs, that are posted by female recruiters than male job seekers are, all else equal. This finding is consistent with the in-group preference mechanism, especially for female job seekers (Proposition 2a and 2b). By holding occupations

127 constant, we are able to control for a pre-existing perception about the gender-dominance of an

occupation that may also influence applicant behavior. This is even more relevant given the

proliferation of online job postings.

Moreover, our Study 2 allows us to make a causal claim using a real job posting, as potential job seekers were randomly assigned to job posting conditions where only the gendered wording and

the gender of the recruiter were manipulated. In our field experiment study, we find strong evidence

for the in-group preference mechanism: compared to male, female job seekers showed more interest

and applied more to jobs posted by a female recruiter (Propositions 2a and 2b). We however find

mixed evidence for the gendered language mechanism (Propositions l a and Ib) in that compared to

male, female job seekers showed more interest in the job when the job was described using feminine

wording, but they made less real application decision when the job was described using the feminine

wording as opposed to masculine one.

Our study expands our understanding of employers' gender segregation mechanisms operating

during the online job-matching process, especially during recruitment, in a number of ways. First, we

were able to observe, code, and analyze, for the first time, how language and recruiter gender at the

recruitment stage may affect the sorting of women at the very first phase of the job-matching process, by observing both their initial attraction to the job and their decisions to submit real applications.

Such gender segregation mechanisms have not been studied before as experienced and exercised from job seekers' perspectives, likely because job descriptions posted by real organizations and the

corresponding job inquiries/job applications sent by job seekers have been typically unavailable to

researchers interested in studying gender segregation in today's labor markets. Our study of language

in job descriptions reveals the significant differential impact of gendered language on female versus

male job seekers' job search behavior. To our knowledge, this differential impact has not been

assessed before in the literature.

Second, our results provide evidence of the role of two additional mechanisms accounting for

why gender segregation at the time of application may still exist in current labor markets, even in an

128 institutional environment like the United States that prohibits the use of discriminatory language when

hiring employees. In this regard, we argue that it is possible that when we observe evidence of sorting based on the gender stereotypicality of the particular job/occupation/industry, we are actually seeing evidence of these other two (complementary) mechanisms at work. In our analyses of the field study and the field experiment data, we provide some evidence for both mechanisms, even after accounting

for two possible alternative mechanisms supported in the gender segregation literature for why female job seekers show interest in certain jobs/occupations/industries and not others (namely, occupational source of segregation and skill source of segregation). First, while women may sort themselves into occupations or industries that are more female- or less male-dominated (or typed) (Fernandez and

Friedrich 2011), it could also be that female- or male-dominated occupation descriptions simply contain more stereotypically feminine or masculine words (as shown in Gaucher et al. 2011), in the end affecting gender sorting. In our analyses of field study data, we still find support for our two mechanisms even after controlling for job-/occupation-/industry-based gender segregating mechanisms. Second, jobs that demand higher or lower skills may also vary in the kinds of language used to attract applicants. Once again, our results still hold, even after accounting for skill variation.

Further, our study expands our knowledge of the role that new labor market intermediaries

(LMIs) play in shaping gender segregation. LMIs today are key institutions facilitating matches between workers and employers (see, e.g., Osterman 2004; Autor 2008). Indeed, a proliferation of web-based recruitment services may serve as a clearinghouse for workers and employers (Nakamura et al. 2009; Stevenson 2009). While recent work has already stressed how LMIs may contribute to labor market inequality (e.g., Fernandez-Mateo and King 2011; Fernandez-Mateo and Fernandez 2016), little is known about how LMIs may contribute to gender sorting at the early recruitment stage of the job-matching process across a variety of jobs, companies, and industries. In a web-based recruiting platform, a new LMI form where hiring companies go to attract talented individuals, job seekers are not only able to communicate freely with recruiters before applying, they are also exposed to additional information about the recruiters and employers. Our study reveals that how recruiters of

129 different genders present themselves and advertise jobs online potentially shapes job seekers' job interests.

To end, our study has relevant implications for management and public policies. Our findings highlight that organizational practices concerning the creation and use of job postings (by employers and recruiters) matter for gender segregation in today's labor markets. Specifically, how jobs are described (use of language), who advertises such jobs (characteristics of the recruiter), and where such jobs are posted (online platform where gender is revealed) may account in part for persisting levels of gender segregation in current labor markets. In particular, our study suggests that altering the gendered language and the gender of the recruiter when advertising a particular job can increase women's interest in the job. So, if an employer action as subtle as changing the gendered words in the description of its jobs can affect labor market outcomes such as gender segregation, we see value in getting employers to experiment with language, recruiters' profiles, as well as other early organizational recruitment practices to overcome gender-typing effects for occupations and industries driving applicant self-selection. It is also our hope that future research continues unpacking the effects of the two underlying key mechanisms (and their interactions) tested here at the time job seekers

search for jobs. We also see great merit in extending our research beyond the recruitment stage and beyond gender to studying how the use of certain language in workplace interactions (as well as demographic similarity) between managers/ and employees may operate at later stages of employees' careers-for example, by examining how encouragement, feedback, advice, and other developmental opportunities to advance in the workplace are presented by managers and experts to employees of different genders.

130 3.7 References

Autor, David. 2008. The Economics of Labor Market Intermediation:An Analytic Framework. w14348. Cambridge, MA: National Bureau of Economic Research.

Autor, David H. and Susan N. Houseman. 2010. "Do Temporary-Help Jobs Improve Labor Market Outcomes for Low-Skilled Workers? Evidence from 'Work First."' American Economic Journal: Applied Economics 2(3):96-.

Barbulescu, Roxana and Matthew Bidwell. 2013. "Do Women Choose Different Jobs from Men? Mechanisms of Application Segregation in the Market for Managerial Workers." Organization Science 24(3):737-56.

Bartz, Jennifer A. and John E. Lydon. 2004. "Close Relationships and the Working Self-Concept: Implicit and Explicit Effects of Priming Attachment on Agency and Communion." Personalityand Social Psychology Bulletin 30(11):1389-1401.

Bem, Sandra L. 1974. "The Measurement of Psychological Androgyny." Journal of Consultingand Clinical Psychology 42(2):155.

Bem, Sandra L. and Daryl J. Bem. 1973. "Does Sex-Biased Job Advertising 'Aid and Abet' Sex Discrimination?l ." Journalof Applied Social Psychology 3(1):6-18.

Berger, Joseph, Bernard P. Cohen, and Morris Zelditch. 1972. "Status Characteristics and Social Interaction." American Sociological Review 37(3):241-55.

Bertrand, Marianne and Sendhil Mullainathan. 2004. "Are Emily and Greg More Employable than Lakisha and Jamal? A Field Experiment on Labor Market Discrimination." The American Economic Review 94(4,):991-1013.

Bielby, William T. 2000. "Minimizing Workplace Gender and Racial Bias." Contemporary Sociology 29(1):120.

Biernat, Monica and Diane Kobrynowicz. 1997. "Gender and Race-Based Standards of Competence: Lower Minimum Standards but Higher Ability Standards for Devalued Groups." Journalof Personalityand Social Psychology 72(3):544-57.

Blair, Irene V. and Mahzarin R. Banaji. 1996. "Automatic and Controlled Processes in Stereotype Priming." Journalof Personality and Social Psychology 70(6):1142.

Brewer, Marilynn and Rupert Brown. 1998. "Intergroup Relations." Pp. 554-94 in Handbook ofSocial Psychology, edited by D. T. Gilbert, S. T. Fiske, and G. Lindzey. New York: Oxford University Press.

Cable, Daniel M. and Timothy A. Judge. 1996. "Person-Organization Fit, Job Choice Decisions, and Organizational Entry." OrganizationalBehavior and Human Decision Processes 67(3):294-311.

Castilla, Emilio J. 2011. "Bringing Managers Back In: Managerial Influences on Workplace Inequality." American SociologicalReview 76(5):667-94.

Cejka, Mary Ann and Alice H. Eagly. 1999. "Gender-Stereotypic Images of Occupations Correspond to the Sex Segregation of Employment." Personalityand Social Psychology Bulletin 25(4):413-23.

131 Cohen, Philip N. and Matt L. Huffman. 2007. "Working for the Woman? Female Managers and the Gender Wage Gap." American Sociological Review 72(5):681-704. Correll, Shelley J. 2001. "Gender and the Career Choice Process: The Role of Biased Self- Assessments." American JournalofSociology 106(6):1691-1730.

Correll, Shelley J. 2004. "Constraints into Preferences: Gender, Status, and Emerging Career Aspirations." American SociologicalReview 69(1):93-113. Correll, Shelley and Caroline Simard. 2016. "Research: Vague Feedback Is Holding Women Back." HarvardBusiness Review 3. Cosca, Theresa and Alissa Emmel. 2010. "Revising the Standard Occupational Classification System for 2010." Monthly Labor Review 133(8):32-41. Fernandez, Roberto M. and M. Lourdes Sosa. 2005. "Gendering the Job: Networks and Recruitment at a Call Center." American JournalofSociology 111(3):859-904. Fernandez, Roberto M. and Colette Friedrich. 2011. "Gender Sorting at the Application Interface: Gender Sorting at the Application Interface." Industrial Relations: A Journal of Economy and Society 50(4):591-609. Fernandez-Mateo, Isabel and Zella King. 2011. "Anticipatory Sorting and Gender Segregation in Temporary Employment." Management Science 57(6):989-1008. Fernandez-Mateo, Isabel and Roberto M. Fernandez. 2016. "Bending the Pipeline? Executive Search and Gender Inequality in Hiring for Top Management Jobs." Management Science 62(12):3636-55. Flory, Jeffrey A., Andreas Leibbrandt, and John A. List. 2015. "Do Competitive Workplaces Deter Female Workers? A Large-Scale Natural Field Experiment on Job Entry Decisions." The Review of Economic Studies 82(1):122-155. Gaucher, Danielle, Justin Friesen, and Aaron C. Kay. 2011. "Evidence That Gendered Wording in Job Advertisements Exists and Sustains Gender Inequality." JournalofPersonality and Social Psychology 101(1):109-28. Gorman, Elizabeth H. 2005. "Gender Stereotypes, Same-Gender Preferences, and Organizational Variation in the Hiring of Women: Evidence from Law Firms." American Sociological Review 70(4):702-28. Gorman, Elizabeth H. 2006. "Work Uncertainty and the Promotion of Professional Women: The Case of Law Firm Partnership." Social Forces 85(2):865-90. Gorman, Elizabeth H. and Julie A. Kmec. 2009. "Hierarchical Rank and Women's Organizational Mobility: Glass Ceilings in Corporate Law Firms." American Journalof Sociology 114(5):1428-74. Greenberg, Jason and Ethan Mollick. 2017. "Activist Choice Homophily and the Crowdfunding of Female Founders." Administrative Science Quarterly 62(2):341-74. Harrison, Spencer H. and Karyn Dossinger. 2017. "Pliable Guidance: A Multilevel Model of Curiosity, Feedback Seeking, and Feedback Giving in Creative Work." Academy ofManagement Journal60(6):2051-2072.

132 Heckman, James J. 1976a. "A Life-Cycle Model of Earnings, Learning, and Consumption." Journalof PoliticalEconomy 84(4, Part 2):S9-S44.

Heckman, James J. 1979. "Sample Selection Bias as a Specification Error." Econometrica 47(1):153- 61.

Heckman, James J. 1976b. "The Common Structure of Statistical Models of Truncation, Sample Selection and Limited Dependent Variables and a Simple Estimator for Such Models." Pp. 475-492 in Annals of Economic and Social Measurement, Volume 5, number 4. NBER.

Heilman, Madeline E. 2001. "Description and Prescription: How Gender Stereotypes Prevent Women's Ascent Up the Organizational Ladder." JournalofSocial Issues 57(4):657.

Heilman, Madeline E., Richard F. Martell, and Michael C. Simon. 1988. "The Vagaries of Sex Bias: Conditions Regulating the Undervaluation, Equivaluation, and Overvaluation of Female Job Applicants." OrganizationalBehavior and Human Decision Processes 41(1):98-1 10.

Hoffman, Curt and Nancy Hurst. 1990. "Gender Stereotypes: Perception or Rationalization?" Journal of Personalityand Social Psychology 58(2):197.

Kuhn, Peter and Kailing Shen. 2013. "Gender Discrimination in Job Ads: Evidence from China *." The QuarterlyJournal ofEconomics 128(1):287-336.

Kunda, Ziva and Steven J. Spencer. 2003. "When Do Stereotypes Come to Mind and When Do They Color Judgment? A Goal-Based Theoretical Framework for Stereotype Activation and Application." PsychologicalBulletin 129(4):522-44.

Leibbrandt, Andreas and John A. List. 2014. "Do Women Avoid Salary Negotiations? Evidence from a Large-Scale Natural Field Experiment." Management Science 61(9):2016-2024.

Maass, Anne, Minoru Karasawa, Federica Politi, and Sayaka Suga. 2006. "Do Verbs and Adjectives Play Different Roles in Different Cultures? A Cross-Linguistic Analysis of Person Representation." Journalof Personality and Social Psychology 90(5):734-50.

MacNell, Lillian, Adam Driscoll, and Andrea Hunt. 2015. "What's in a Name: Exposing Gender Bias in Student Ratings of Teaching." Innovative Higher Education 40(4):291-303.

Madera, Juan M., Michelle R. Hebl, and Randi C. Martin. 2009. "Gender and Letters of Recommendation for Academia: Agentic and Communal Differences." Journal ofApplied Psychology 94(6):1591.

Marin, Alexandra. 2012. "Don't Mention It: Why People Don't Share Job Information, When They Do, and Why It Matters." Social Networks 34(2):181-92.

Maume, David J. 2011. "Meet the New Boss... same as the Old Boss? Female Supervisors and Subordinate Career Prospects." Social Science Research 40(1):287-98.

McPherson, Miller, Lynn Smith-Lovin, and James M. Cook. 2001. "Birds of a Feather: Homophily in Social Networks." Annual Review of Sociology 27(1):415-44.

Nakamura, Alice 0., Kathryn L. Shaw, Richard B. Freeman, Emi Nakamura, and Amanda Pyman. 2009. "Jobs Online," in David H. Autor, ed., Studies of Labor Market Intermediation, The University of Chicago Press, 2009, chapter 1, pp. 27-65.

133 Neuendorf, Kimberly A. 2011. "Content Analysis-A Methodological Primer for Gender Research." Sex Roles 64(3-4):276-289.

Osterman, Paul. 2004. "Labor Market Intermediaries in the Modem Labor Market." Workforce Intermediariesfor the Twenty-First Century 155-69.

Pager, Devah. 2007. "The Use of Field Experiments for Studies of Employment Discrimination: Contributions, Critiques, and Directions for the Future." The Annals of the American Academy of Politicaland Social Science 609(1):104-133.

Pager, Devah and David S. Pedulla. 2015. "Race, Self-Selection, and the Job Search Process." American Journalof Sociology 120(4):1005-54.

Pager, Devah, Bruce Western, and Bart Bonikowski. 2009. "Discrimination in a Low-Wage Labor Market: A Field Experiment." American Sociological Review 74: 777-99.

Papke, Leslie E. and Jeffrey M. Wooldridge. 1996. "Econometric Methods for Fractional Response Variables with an Application to 401 (k) Plan Participation Rates." Journal ofApplied Econometrics 11(6):619-632.

Pedriana, Nicholas. 2004. "Help Wanted NOW: Legal Resources, the Women's Movement, and the Battle Over Sex-Segregated Job Advertisements." Social Problems 51(2):182-201.

Pedriana, Nicholas and Amanda Abraham. 2006. "Now You See Them, Now You Don't: The Legal Field and Newspaper Desegregation of Sex- Segregated Help Wanted Ads 1965-75." Law & Social Inquiry 31(4):905-38.

Pennebaker, James W., Roger J. Booth, and Martha E. Francis. 2007. Operator's Manual: Linguistic Inquiry and Word Count: LIWC2007. Austin, TX: LIWC.net.

Perdue, Charles W., John E. Dovidio, Michael B. Gurtman, and Richard B. Tyler. 1990. "Us and Them: Social Categorization and the Process of Intergroup Bias." Journal ofPersonalityand Social Psychology 59(3):475-86.

Reskin, Barbara F. 2000. "The Proximate Causes of Employment Discrimination." Contemporary Sociology 29(2):319.

Ridgeway, Cecilia L. 2011. Framedby Gender: How Gender Inequality Persistsin the Modern World. New York: Oxford University Press.

Ridgeway, Cecilia L. 1997. "Interaction and the Conservation of Gender Inequality: Considering Employment." American SociologicalReview 62(2):218.

Ridgeway, Cecilia L. and Shelley J. Correll. 2004. "Unpacking the Gender System: A Theoretical Perspective on Gender Beliefs and Social Relations." Gender & Society 18(4):510-31.

Ridgeway, Cecilia L. and Lynn Smith-Lovin. 1999. "The Gender System and Interaction." Annual Review of Sociology 25:191-216.

Rivera, Lauren A. 2015. "Go with Your Gut: Emotion and Evaluation in Job Interviews." American Journalof Sociology 120(5):1339-89.

Rivera, Lauren A. 2012. "Hiring as Cultural Matching: The Case of Elite Professional Service Firms." American SociologicalReview 77(6):999-1022.

134 Rudman, Laurie A. and Peter Glick. 1999. "Feminized Management and Backlash toward Agentic Women: The Hidden Costs to Women of a Kinder, Gentler Image of Middle Managers." Journalof Personalityand Social Psychology 77(5):1004.

Rudman, Laurie A. and Peter Glick. 2001. "Prescriptive Gender Stereotypes and Backlash toward Agentic Women." JournalofSocial Issues 57(4):743-762.

Rudman, Laurie A. and Stephen E. Kilianski. 2000. "Implicit and Explicit Attitudes toward Female Authority." Personalityand Social Psychology Bulletin 26(11):1315-1328.

Schmader, Toni, Jessica Whitehead, and Vicki H. Wysocki. 2007. "A Linguistic Comparison of Letters of Recommendation for Male and Female Chemistry and Biochemistry Job Applicants." Sex Roles 57(7/8):509-14.

Schullo, Stephen A. and Burton L. Alperson. 1984. "Interpersonal Phenomenology as a Function of Sexual Orientation, Sex, Sentiment, and Trait Categories in Long-Term Dyadic Relationships." Journal of Personalityand Social Psychology 47(5):983.

Smith, Sandra Susan. 2005. "'Don't Put My Name on It': Social Capital Activation and Job-Finding Assistance among the Black Urban Poor." American JournalofSociology 11 1(1):1-57.

Srivastava, Sameer B. and Eliot L. Sherman. 2015. "Agents of Change or Cogs in the Machine? Reexamining the Influence of Female Managers on the Gender Wage Gap." American Journal of Sociology 120(6):1778-1808.

Stainback, Kevin, Donald Tomaskovic-Devey, and Sheryl Skaggs. 2010. "Organizational Approaches to Inequality: Inertia, Relative Power, and Environments." Annual Review ofSociology 36(1):225-47.

Stevenson, Betsey. 2009. "The Internet and Job Search." Pp. 67-86 in Studies of Labor Market Intermediation. Edited by David Autor. Chicago: University of Chicago Press.

Stoll, Michael A., Stephen Raphael, and Harry J. Holzer. 2004. "Black Job Applicants and the Hiring Officer's Race." Industrialand Labor Relations Review 57(2):23.

Trix, Frances and Carolyn Psenka. 2003. "Exploring the Color of Glass: Letters of Recommendation for Female and Male Medical Faculty." Discourse & Society 14(2):191-220.

Turban, D. B. and T. W. Dougherty. 1992. "Influences of Campus Recruiting on Applicant Attraction to Firms." Academy ofManagement Journal 35(4):739-65.

Wagner, David G. and Joseph Berger. 2002. "Expectation States Theory: An Evolving Research Program." in New Directions in ContemporarySociological Theory, edited by J. Berger and M. Z. Jr. Lanham, Md: Rowman & Littlefield Publishers.

Wagner, David G. and Joseph Berger. 1997. "Gender and Interpersonal Task Behaviors: Status Expectation Accounts." Sociological Perspectives40(1):1-32.

Wynn, Alison T. and Shelley J. Correll. 2017. "Gendered Perceptions of Cultural and Skill Alignment in Technology Companies." Social Sciences 6(2):45.

135 3.8 Tables and Figures

Figure 1. Early Stage of Recruitment in the Job-Matching Process

*0 *

Pre-application (Recruitment) Post-application (Selection and Hiring)

-Job seeker

136 Figure 2. Online Job Posting Interface in the Field Study (Study 1)

Computer Programmer Location: New York, NY Compensation: Competitive Full Time Employment

Profile We are looking for a local candidate with an in John Smith depth knowledge of ASP.NET, C#, JmascripL. HTML, CSS, SSRS, SQL Server 2008, Sessions, HR Specialist at Fun Caching. Co. New York, NY a "I have five years of Candidates must have proven ability to work in experience as a HR team environment but function independently specialist. My goal is to without close supervision. Anyone with superior find top talent for our organizational skills, personal and programming company." ability to create clear, modularized code are encouraged to apply.

SNUR lR

INQUIRE ABOUT THIS JOB

Name: Send John Smith a Profile In - in eesdinob #mO: John Smith I"Computer Programmer" at Email: your company "Fin Co." HR Specialist at Fun Please contact me about the Co. Phone: position. New York, NY "I have five years of experience as a HR specialist. My goal is to Attach a resume (optional): find top talent for our Choose File company."o

137 Figure 3. Online Job Posting and Two Main Manipulations in the Field Experiment (Study 2)

Project Assistant Job Number: 15990 Department: [Educational Institution] Employment Type: Part-Time Employment Category: Non-Exempt Compensation: $14 per hour

Job Description Our research team at [Educational Institution] is looking for a [FIRST WORD] and [SECOND WORD] candidate willing to gain experience as a research assistant. An ideal candidate should be [THIRD WORD] and care about the success of our projects.

Your main task will be to provide assistance sorting and inputting text data that will be used for a research project; to help coordinate project logistics; to help launch simple research surveys; and to perform other duties for the team.

Job Requirements No previous experience required as we will provide training. High school degree is required. Work can be completed online. No relocation is required.

Candidate must possess strong [FOURTH WORD] and polished oral and written communication skills. Must be able to demonstrate the ability to work [FIFTH WORD], manage assignments, and set priorities. Anyone with superior organizational skills, [SIXTH WORD], and [SEVENTH WORD] is encouraged to apply.

For more information, contact our recruiter [RECRUITER NAME] at [RECRUITER EMAIL]. [GENDER PRONOUN] will answer any questions you may have.

Gendered Language Manipulation (3 Conditions - Between Subjects Design)

Feminine Masculine Neutral Wording Wording Wording First Word Committed Determined No Word Second Word Supportive Assertive No Word Third Word Sensitive Outspoken No Word Fourth Word Interpersonal Analytical No Word Fifth Word Cooperatively Independently No Word Sixth Word Compassionate Decisive No Word Seventh Word Honest Persistent No Word

Gender of Recruiter Manipulation (Conditions - Between Subjects Design)

Male Female Recruiter Recruiter

Name John Jennifer

138 Study 1: Table 1 Summary of Descriptive Statistics

Sample Sub-Sample Variables Full Mean (Std. Dev.) Percent Mean (Std. Dev.) Percent Panel A: Job Seeker (Inquiry) Characteristics n=566,409 n= 49,597 Gender (ofJob Inquiry) Male 56.2 50.4 Female 30.0 35.6 Not available 13.8 13.9 Information Channel Direct 0.02 0.03 Linkedln 32.7 25.7 Twitter 4.6 3.9 Indeed 3.0 4.2 Simply Hired 40.3 48.1 Facebook 0.3 0.3 Google 2.3 1.8 Other (Trovit, Juju, Emails, etc.) 16.8 15.9 Job Inquiries with Resume Attached 56.3 54.8

Panel B: Recruiter Characteristics n=24,146 n=10,526 Gender Male 37.8 37.5 Female 38.7 41.1 Not available 23.5 21.4 PicturePosting (ofSelf or of Company Representing), all 56.2 63.9 Recruiters with unidentifiable gender 53.8 62.3 Recruiters with missing gender 33.5 39.8 Company Type Corporate 36.9 29.6 Staffing 61.0 68.3 Other 0.1 0.1 Missing 2.1 2.0 State of Company (5 most common in full sample) California 11.8 10.7 Texas 11.2 11.8 New York 7.9 8.2 New Jersey 5.8 4.9 Illinois 5.8 6.1 Job Postingsper Recruiter (min=1; max=1,509) (min=1; max=1,489) 6.2 (18.6) 3.3 (16.1) Job Inquiriesper Recruiter (min=1; max=1,699) (min=1; max=1,699) 3.8 (12.3) 4.1 (18.0) Job Inquiries with Resume Attached per Recruiter (min=1; max=369) (min=1; max=369) 2.8(4.3) 3.1 (5.9)

139 Study 1: Table 1 (Continued) Summary of Descriptive Statistics

Full Sample Sub-Sample A Variables Mean (Std. Dev. Percent Mean (Std. Dev.) Percent Panel C: Job Characteristics n=150,270 n=34,460 Share of Stereotypically Feminine Words (min=0; max=15.38) (min=0; max=10) 1.01 (0.94) 0.96 (0.90) Share of Stereotypically Masculine Words (min=0; max=1 6.67) (min=0; max=14.29) 0.98 (1.01) 1.04 (1.09) Net Proportionof StereotypicallyFeminine Words (min=-16.67; max= 15.38) (min=-10.53; max=10) (share of feminine words - share of masculine words) 0.03 (1.33) -0.08 (1.33) Neutral Job Posting (1 if no feminine or masculine words; 0 otherwise) 13.6 15.9 Work-life Balance (I if any mention; 0 otherwise) 11.0 14.0

Employment Type Full Time 60.5 61.0 Contract 25.0 21.4 Contract-to-hire 7.9 8.6 Part Time 0.9 1.8 Temp 0.7 1.0 Intern 0.1 0.1 Missing 4.9 6.1 Industry (5 most common in full sample) Technology 19.5 14.3 Healthcare 8.8 11.3 Finance & Banking 6.9 7.4 Computer Software 5.3 3.3 Manufacturing 4.8 5.5 State (5 most common in full sample) California 12.7 11.3 Texas 11.4 11.9 New York 9.8 9.7 Illinois 6.0 6.0 Massachusetts 5.6 5.2 Occupation (5 most common in sub-sample A) Management Analysts N/A N/A 11.2 Accountants and Auditors N/A N/A 7.8 Software Developers N/A N/A 6.3 Data Entry Keyers N/A N/A 5.0 Web Developers N/A N/A 3.8 Job Inquiriesper Job (min=1; max= 1699) (min=1; max=1699) 3.8 (8.4) 4.3 (12.6) Job Inquirieswith Resume Attached per Job (min=l ; max=369) (min=1; max=369) 2.9 (4.4) 3.2 (5.7)

140 Study 1: Table 2 Logistic Regression Models Predicting Whether the Gender of the Job Seeker is Female (Full Sample and Sub-Sample A)

Job Inquiries Job Applications

Model 1 Model 2 Model 3 Model I Model 2 Model 3 (Sub- (Sub- (Full Sample) Sample A) (Full Sample) Sample A) Net Proportionof Stereotypically Feminine Words 0.047*** 0.046*** 0.015** 0.048*** 0.047*** 0.016* (share of feminine words - (0.002) (0.002) (0.006) (0.003) (0.003) (0.007) share of masculine words)

Employment Type (Full-time as reference) Contract 0.071*** 0.075*** 0.017 0.066*** 0.068*** 0.008 (0.009) (0.009) (0.021) (0.012) (0.012) (0.028) Contract to Hire 0.197*** 0.186*** -0.017 0.220*** 0.208*** 0.013 (0.012) (0.012) (0.026) (0.016) (0.016) (0.034) Part Time 0.647*** 0.622*** 0. 124* 0.709*** 0.688*** 0.270** (0.029) (0.029) (0.063) (0.044) (0.044) (0.095) Temp 0.490*** 0.453*** 0.032 0.468*** 0.440*** -0.111 (0.034) (0.034) (0.064) (0.048) (0.048) (0.091) Intern 0.226*** 0.225*** -0.530** 0.228** 0.228** -0.460+ (0.063) (0.063) (0.202) (0.078) (0.077) (0.248) Company Type (Corporate firm as reference) Staffing firm -0.031 *** -0.020* -0.002 -0.024* -0.012 0.002 (0.008) (0.008) (0.019) (0.011) (0.011) (0.025) Other 0.604*** 0.678*** 0.293+ 0.667*** 0.761*** 0.199 (0.078) (0.078) (0.173) (0.107) (0.107) (0.244) Work-lfe Balance (1 if any mention; 0 otherwise) 0.196*** 0.192*** 0.054* 0.163*** 0.156*** 0.027 (0.010) (0.010) (0.022) (0.014) (0.014) (0.030) Recruiter Gender (Male as reference) Female 0.234*** 0.077*** 0.235*** 0.077*** (0.007) (0.016) (0.010) (0.022) Unknown 0.050*** 0.008 0.073*** 0.011 (0.009) (0.020) (0.012) (0.027) State FE Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Occupation FE No No Yes No No Yes

Constant -2.312*** -2.414*** -3.005*** -2.650*** -2.714*** -2.504*** (0.131) (0.131) (0.356) (0.253) (0.254) (0.637) Pseudo R-squared 0.096 0.097 0.227 0.080 0.082 0.21 Number of Inquiries 482,771 482,771 125,607 271,035 271,035 68,534

Notes. This table shows the regression coefficients with standard errors in parentheses. Undisclosed in this table, due to space limitations, are coefficients for dummies concerning the job seekers' information channel. The models also control for those postings with neutral wording (i.e., those postings with no feminine or masculine words). The dependent variable takes the value of 1 ifjob seeker who inquired about the job is female and 0 otherwise. + p

141 Study 1: Table 3 Logistic Regression Models Predicting Whether the Gender of the Job Seeker is Female (Sub-Sample B, with Detailed Occupation Fixed Effects)

Job Inquiries Job Applications Model Model I Model 2 Model 3 Model 1 Model 2 3 Net Proportionof Stereotypically Feminine Words 0.052*** 0.047** 0.043** 0.054** 0.052* 0.045* (share of feminine words - share of masculine words) (0.015) (0.016) (0.016) (0.019) (0.020) (0.021) Recruiter Gender (Male as reference) Female 0.251*** 0.134** 0.132** 0.231*** 0.114* 0.117* (0.041) (0.043) (0.044) (0.053) (0.055) (0.056) Unknown 0.084 0.043 0.006 0.038 -0.008 -0.035 (0.053) (0.054) (0.055) (0.066) (0.069) (0.070)

State FE Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes Occupation FE No Yes No No Yes No Occupation FE (more detailed, by skill-level) No No Yes No No Yes

Constant -1.321*** -3.215* -0.844 0.094 -0.868 0.794 (0.512) (1.446) (0.574) (1.351) (1.818) (1.483) Pseudo R-squared 0.048 0.076 0.084 0.047 0.079 0.086 Number of Inquiries 14,753 14,753 14,753 9,013 9,006 8,999

Notes. This table shows the regression coefficients with standard errors in parentheses. Undisclosed in this table, due to space limitations, are coefficients and odds ratios for dummies concerning the job seekers' information channel, as well as employment type and company type. The models also control for those postings with neutral wording (i.e., those postings with no feminine or masculine words). The dependent variable takes the value of I if job seeker who inquired about the job is female and 0 otherwise. + p

142 Study 2: Table 1 Summary of Key Descriptive Statistics

Clicked on Short Interested in the All Partcipants Ad Job After Reading Applied Long Ad

VariablesVaibls(S.D.) Mn Percent (S.D.) Percent (S.D.) Percent (S.D)Ma Percent Gender: Female 48.6 50.4 52.2 55.2 Employment Status Full-Time 56.9 51.3 50.1 45.5 Part-Time 25.8 30.2 31.1 32.7 Unemployed 17.3 18.5 18.8 21.8 Lookingfor a Job Full-Time 21.4 24.8 26.0 36.7 Part-Time 54.9 46.0 43.6 29.1 No 14.8 18.0 18.1 17.3 Maybe 8.9 11.2 12.3 16.9 Student (in school) 12.1 11.9 12.4 17.6 Level of EducationAchieved Some High School 0.4 0.4 0.4 - High School 9.0 9.2 8.9 8.3 Some College 23.4 23.7 24.2 22.1 Associate Degree 12.0 12.2 12.5 13.3 College Degree 41.2 41.1 40.7 44.6 Master Degree 1.1 1.0 0.7 0.5 Professional School Degree 0.2 0.3 0.4 0.2 Doctorate 11.0 10.5 10.7 9.2 Other 1.6 1.6 1.5 1.8 MaritalStatus Single, Never Married 41.7 42.0 42.1 41.4 Married 38.6 37.0 36.7 35.4 Living with Partner 10.7 11.1 11.5 14.0 Divorced 6.8 7.6 7.5 7.9 Region New England 4.3 4.4 4.2 3.2 Middle Atlantic 15.0 15.9 15.8 18.0 East North Central 15.3 14.9 14.9 16.4 West North Central 5.7 5.3 5.8 4.5 South Atlantic 22.8 23.0 22.9 21.4 East South Central 5.4 5.2 5.5 5.6 West South Central 9.4 9.3 9.6 10.1 Mountain 6.3 6.1 5.8 5.4 Pacific 15.8 15.8 15.4 15.3 Years of Work Experience 11.9 12.2 12.1 11.5 (6.4) (6.5) (6.5) (6.4) Age 36.0 36.6 36.4 35.1 (10.3) (10.3) (10.2) (9.9) Total Number of Participants (a) 4,150 (a) 2,003 (b) 1,648 (c) 444 (d)

Notes. (a) This is number of valid participants (i.e., with gender information) who completed the survey, and therefore were at risk of clicking on the short ad. (b) This is number of participants who clicked on the short ad (i.e., interested in learning more about the position when exposed to short job posting). This is 48.3 percent of (a). (c) This is number of participants who clicked on short ad and who were interested in learning more about the job. This is 39.5 percent of (a). (d) This is

143 number of participants who clicked on short and applied to the job by submitting a resume. This is 10.7 percent of (a).

144 Study 2: Table 2 Binary Regression Models Predicting Whether Participant Applied to the Job

Clicked on Short Ad Model 1 Model 2 Model 3 MAIN MODEL Job Seeker Gender (Male as Reference) Female Job Seeker 0.296* 0.501* 0.19 (0.117) (0.195) (0.227) Gendered Language of Job (Masculine as Reference) Stereotypically Feminine Wording -0.222+ 0.04 0.028 (0.135) (0.200) (0.200) Stereotypically Neutral Wording -0.426** -0.356+ -0.367+ (0.139) (0.206) (0.207) Recruiter Gender (Male as Reference) Female Recruiter 0.037 0.04 -0.283+ (0.112) (0.113) (0.166) Job Seeker Gender X Gendered Language of Job: Female Job Seeker x Feminine Wording -0.488+ -0.481+ (0.272) (0.272) Female Job Seeker x Neutral Wording -0.118 -0.115 (0.279) (0.280) Job Seeker Gender X Recruiter Gender: Female x Female Recruiter 0.603** (0.226) Constant -0.796* -0.886** -0.743* (0.435) (0.442) (0.445)

Pseudo R-squared 0.0663 0.0679 0.0713 N. of Total Participants (Clicked on Short Ad) 2,003 2,003 2,003 Notes. This table shows the regression coefficients with standard errors in parentheses. Due to space limitations, the table does not present the coefficients for controls measuring job seekers' key characteristics such as age, level of education, marital status, employment status, years of work experience, job search status, student status, and region (these coefficients are available upon request). + p

145 3.9 Appendix

Appendix Table 1 (Study 1) Stereotypically Masculine and Feminine Words Used to Code Job Advertisements

Masculine Words Feminine Words

Active Headstrong Affectionate Modesty Adventurous Hierarch* Child* Nag Aggress* Hostil* Cheer* Nurtur* Ambitio* Impulsive Commit* Pleasant* Analy* Independen* Communal Polite Assert* Individual* Compassion* Quiet* Athlet* intellect* Connect* Respon* Autonom* Lead* Considerate Sensitiv* Boast* Logic Cooperat* Submissive Challeng* Masculine Depend* Support* Compet* Objective Emotiona* Sympath* Confident Opinion Empath* Tender* Courag* Outspoken Feminine Together* Decide Persist Flatterable Trust* Decisive Principle* Gentle Understand* Decision* Reckless Honest Warm* Determin* Stubborn Interdependen* Whin* Dominant Superior Interpersonal Yield* Domina* Self-confiden* Kind Force* Self-sufficien* Kinship Greedy Self-relian* Loyal*

Notes. The asterisk denotes the acceptance of all letters, hyphens, or numbers following its appearance. Source: This coding procedure was adopted from Gaucher, Friesen, and Kay (2011)

146 Appendix Table 2 (Study 1) Examples of Job Advertisement Coding

[JOB ADVERTISEMENT 1] .NET Developer

Job Description:

We are looking for a local candidate with an in depth knowledge of ASP.NET, C#, Javascript, HTML, CSS, SSRS, SQL Server 2008, Sessions, Caching. Candidates must have proven ability to work in a team environment but function independently without close supervision. Anyone with superior organizational skills, personal and programming ability to create clear, modularized code are encouraged to apply.

Word Count: 59 Share of Stereotypically Masculine Words: 3.39 Share of Stereotypically Feminine Words: 0

[JOB ADVERTISEMENT 2] Clinical Project Manager

Job Description:

This is a full-time position in XXX to work as a primary resource for clinical applications project planning, work flow analysis, management, development, and support with special focus on high- end clinical solutions and applications.

- Evaluates, recommends, configures, manages, and installs clinical technology to meet customers' needs. - Works on complex implementation and support projects using clinical and IT skills. - Manages multiple resources and projects to effectively implement standard systems and services.

Five years of hands-on Clinical or IS Implementation informatics experience. Bachelor's degree required. Master's Degree in Health Care or Health Care informatics preferred. Formal project management certification or CPHIMS certification preferred.

Excellent written and verbal communication skills with the ability to make formal presentations required. Exhibits strong interpersonal skills. Requires use of electronic mail, time and attendance software, learning management software and intranet. Must be able to read, write legibly, speak and comprehend English.

Word Count: 147 Share of Stereotypically Masculine Words: 0 Share of Stereotypically Feminine Words: 1.36

147 Appendix Table 3 (Study 1) Summary of Descriptive Statistics by Female/Male dominated and Gender-Integrated Occupations (Full Sample)

Job Inquiries in Full Sample (Study 1) % Female workers Female Male (2012 CPS Total Mean SD (%) (%) March) Female-DominatedOccupations 30,963 78.7 21.3 Data entry keyers 6,351 0.205 0.404 86.7 13.4 95.6 Health practitioner support technologists and technicians 1,137 0.037 0.188 87.6 12.4 94.0 Paralegals and legal assistants 955 0.031 0.173 86.0 14.0 93.4 Human resources assistants, except payroll/timekeeping 1,828 0.059 0.236 84.4 15.6 91.6 Word processors and typists 782 0.025 0.157 66.7 33.3 90.1 Medical assistants 699 0.023 0.149 78.1 21.9 89.8 Licensed practical and licensed vocational nurses 705 0.023 0.149 80.2 19.8 86.8 Office clerks, general 3,698 0.119 0.324 67.7 32.3 85.8 Secretaries and administrative assistants 580 0.019 0.136 67.8 32.3 85.0 Bookkeeping, accounting, and auditing clerks 747 0.024 0.153 79.1 20.9 84.1 Registered nurses 717 0.023 0.150 79.7 20.3 82.1 Meeting, convention, and event planners 11,971 0.387 0.487 76.3 23.7 77.5 Nurse practitioners 793 0.026 0.158 83.7 16.3 77.3

Male-DominatedOccupations 26,936 10.8 89.2 Mechanical engineers 2,335 0.087 0.281 9.9 90.2 3.8 Sales engineers 1,412 0.052 0.223 12.0 88.0 5.0 Construction managers 546 0.020 0.141 7.7 92.3 6.4 Electrical and electronics engineers 1,640 0.061 0.239 13.1 86.9 8.8 Computer network architects 2,291 0.085 0.279 9.4 90.6 9.6 Roustabouts, oil and gas 7,812 0.290 0.454 4.8 95.2 9.7 Architectural and engineering managers 707 0.026 0.160 9.4 90.6 10.3 Civil engineers 1,133 0.042 0.201 15.9 84.1 13.1 Transportation, storage, and distribution managers 636 0.024 0.152 11.5 88.6 15.4 Industrial engineers, including health and safety 2,005 0.074 0.262 13.1 86.9 18.0 Industrial production managers 835 0.031 0.173 10.6 89.5 18.2 Software developers, applications and systems software 4,749 0.176 0.381 21.5 78.6 21.3 Chefs and head cooks 835 0.031 0.173 5.0 95.0 23.8

Notes. Occupations are coded as female-dominated if the share of female employment was greater than 75% in 2012 based on the Current Population Survey (CPS) data; occupations are coded as male-dominated if the share of female employment was less than 25%; occupations are coded as gender-integrated if the share of female employment was less than 62.5% and greater than 37.5%.

148 Appendix Table 3 (Study 1) (Continued) Summary of Descriptive Statistics by Female/Male dominated and Gender-Integrated Occupations

Job Inquiries in Full Sample (Study 1) % Female workers (2012 Total Mean SD Female (%) Male (%) CPS March) Gender-IntegratedOccupations 49,664 36.0 64.0 Training and development specialists 617 0.012 0.111 29.0 71.0 62.0 Pharmacists 9,305 0.187 0.390 41.9 58.1 61.8 Biological scientists 511 0.010 0.101 50.1 49.9 59.9 Credit analysts 1,548 0.031 0.174 45.0 55.0 58.2 Database administrators 1,032 0.021 0.143 51.8 48.2 56.2 Operations research analysts 654 0.013 0.114 26.7 73.3 55.7 Athletes, coaches, umpires, and related workers 3,281 0.066 0.248 49.3 50.7 55.5 Editors 1,444 0.029 0.168 36.4 63.7 55.0 Purchasing managers 1,737 0.035 0.184 48.1 51.9 54.3 Food service managers 1,408 0.028 0.166 45.8 54.2 54.0 Market research analysts and marketing specialists 518 0.010 0.102 47.7 52.4 51.6 Artists and related workers 1,051 0.021 0.144 22.3 77.7 50.4 Writers and authors 1,146 0.023 0.150 21.0 79.0 48.3 Technical writers 8,317 0.167 0.373 26.7 73.3 46.4 Chemists and materials scientists 526 0.011 0.102 51.1 48.9 45.8 Designers 631 0.013 0.112 27.8 72.2 43.9 Marketing and sales managers 2,106 0.042 0.202 31.6 68.4 43.1 Accountants and auditors 13,128 0.264 0.441 31.3 68.7 42.2 Management analysts 704 0.014 0.118 52.5 47.5 41.0

Notes. Occupations are coded as female-dominated if the share of female employment was greater than 75% in 2012 based on the Current Population Survey (CPS) data; occupations are coded as male-dominated if the share of female employment was less than 25%; occupations are coded as gender-integrated if the share of female employment was less than 62.5% and greater than 37.5%.

149 Appendix Table 4 (Study 1) Additional Multivariate Models to Further Assess Hypotheses 1 and 2: Jobs Advertisements as the Unit of Analysis (Full Sample)

Job Inquiries Job Inquiries with Resume Attached

Fractional Logit Negative Binomial Fractional Negative Binomial Model Regression Models Logit Model Regression Models

DV: Proportion of DV: Number DV: Number DV: Proportion of DV: Number DV: Number Female Job Female Job Male Job Female Job Female Job Male Job Seekers Seekers Seekers Seekers Seekers Seekers (Model 1) (Model 2) (Model 3) (Model 4) (Model 5) (Model 6)

Net Proportionof Stereotypically Feminine Words 0.018*** 0.023*** -0.022*** 0.013** 0.018*** -0.024*** (share of feminine words - share of masculine words) (0.004) (0.005) (0.003) (0.005) (0.004) (0.003) Recruiter Gender (Male as reference) Female 0.165*** 0.167*** -0.064*** 0.170*** 0.183*** -0.065*** (0.011) (0.012) (0.009) (0.014) (0.013) (0.009) Unknown -0.004 0.074*** 0.020+ 0.006 0.060*** -0.014 (0.013) (0.015) (0.012) (0.018) (0.015) (0.011)

State FE Yes Yes Yes Yes Yes Yes Industry FE Yes Yes Yes Yes Yes Yes

Constant -2.410*** -1.263*** 0.977*** -2.596*** -0.714* 1.506*** (0.186) (0.187) (0.148) (0.362) (0.299) (0.176)

AIC 0.974 0.951 Pseudo R-squared 0.049 0.032 0.044 0.023 Number of Job Advertisements 149,813 149,813 149,813 84,241 84,241 84,241

Notes. This table shows the regression coefficients with standard errors in parentheses. Undisclosed in this table, due to space limitations, are coefficients for dummies concerning the job seekers' information channel. The models also control for those postings with neutral wording (i.e., those postings with no feminine or masculine words). For the fractional logit models, the dependent variable is the proportion of female job seekers inquiring about a job, which ranges from 0 to 100 (with a mean value of 32.0 and a standard deviation of 38.5). For the negative binomial regression models, the dependent variable is the count of female/male job seekers. + p<0.10, * p<0.05, ** p<0.01, *** p<0.001 (two-sided tests)

150