Eunsoo 1
Eunsoo 1
Eunsoo 1
by
Eunsoo Kim
Doctoral Committee:
Professor Puneet Manchanda, Chair
Associate Professor Anocha Aribarg
Associate Professor Yves F. Atchadé
Professor Professor Peter Lenk
Assistant Professor Eric Schwartz
Eunsoo Kim
eunsoo@umich.edu
ORCID iD: 0000-0001-9332-5653
c Eunsoo Kim 2017
Dedication
This dissertation is dedicated to my dear family for all their love and unconditional support;
to Lisa and Rosa, whom I shared most of my time with during this journey; and to all of
those who cherish relationships with others.
ii
Acknowledgements
First and foremost, I would like to express my utmost gratitude to my advisor, Puneet Man-
chanda for his years of guidance and sresupport. He not only helped me develop the skills
and attitudes needed for research, but also provided me emotional support. His mentorship
led me to be a better researcher and a better person. I would also like to thank Anocha
Aribarg with whom I’ve been working from the beginning of mt Ph.D, program to the very
end. I am greatly inspired by her positive drive, thoughtfulness and persistence, and I am
very thankful for her for standing by my side. Moreover, I truly appreciate my dissertation
committee members, Anocha Aribarg, Peter Lenk, Yves F. Atchadé and Eric Schwartz for
taking their precious time providing me helpful comments and feedback to move the disser-
tation forward. Moreover, I would like to thank marketing faculty members and my fellow
Ph.D. students in the doctoral program at the Ross School of Business for all of their sincere
encouragement.
Last but not least, I would like to thank my parents, sister and brother for all their
endless love. Without their support, I would not have reached anywhere near where I am
right now.
iii
Contents
Dedication ii
Acknowledgements iii
List of Tables vi
List of Figures viii
Abstract x
Chapter 1 Social versus Economic Factors in Network Formation: An
Empirical Analysis of the Multi-level Marketing Industry 1
1.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Institutional Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.1 The Multi-level Marketing Industry . . . . . . . . . . . . . . . . . . . 6
1.3.2 Social Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3.3 Economic Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.3.1 Distributors’ Characteristics (New Joiners’ Economic Return) 9
1.3.3.2 New joiners’ Characteristics (IBOs’ Economic Return) . . . 13
1.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4.1 Network Formation via Matching . . . . . . . . . . . . . . . . . . . . 14
1.4.2 Matching Related Variables . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.2.1 Social Factors: Demographic Homophily & Geographic Dis-
tance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4.2.2 Distributor Economic Signifiers for New Joiners . . . . . . . 16
1.4.2.3 New Joiner Economic Signifiers for Uplines . . . . . . . . . 17
1.5 Modeling Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5.1 Two-sided matching model . . . . . . . . . . . . . . . . . . . . . . . . 17
1.5.2 Two Sides, Utilities, and Equilibrium . . . . . . . . . . . . . . . . . . 18
1.5.2.1 Two Sides . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.5.2.2 Latent Match Utilities (Relationship Value) . . . . . . . . . 19
1.5.2.3 Equilibrium Outcome . . . . . . . . . . . . . . . . . . . . . 22
1.5.3 Empirical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
iv
1.5.3.1 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.5.3.2 Bayesian inference using Gibbs sampling . . . . . . . . . . . 26
1.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.6.1 Model Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.6.1.1 Social Factors . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.6.1.2 Economic Factors . . . . . . . . . . . . . . . . . . . . . . . . 29
1.6.2 Social versus Economic Factors in Network Formation . . . . . . . . . 30
1.6.3 Social versus Economic Factors and Future Business Behavior . . . . 31
1.7 Conclusion and Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.8 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.9 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Chapter 2 The Impact of Existence of Others on Inactive Behavior in the
Multi-level Marketing Industry 48
2.1 Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.3 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.3.1 Turnover and Costs in MLM . . . . . . . . . . . . . . . . . . . . . . . 53
2.3.2 Prior Literature on Turnover . . . . . . . . . . . . . . . . . . . . . . . 54
2.3.3 Three Different Dimensions of Others . . . . . . . . . . . . . . . . . . 56
2.3.3.1 Individual . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3.3.2 Network Family . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.3.3.3 Proximity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
2.4.1 Individuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.4.2 Network Family . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4.3 Proximity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.5 Existence of Others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.5.1 County-level Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.5.2 Individual-level Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.5.2.1 Time-independent hazard model . . . . . . . . . . . . . . . 68
2.5.2.2 Time-varying hazard model . . . . . . . . . . . . . . . . . . 68
2.5.2.3 Results and Discussion . . . . . . . . . . . . . . . . . . . . . 69
2.6 Prediction Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . . 73
2.6.1 Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.6.2 Data for Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . 75
2.6.3 Analysis and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
v
2.6.3.1 Approach One – Incidence Prediction . . . . . . . . . . . . . 76
2.6.3.2 Approach Two – Duration Prediction . . . . . . . . . . . . . 77
2.6.3.3 Approach Three – Duration Prediction . . . . . . . . . . . . 79
2.7 Conclusion and Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.8 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.9 Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Bibliography 101
vi
List of Tables
vii
2.17 Contingency Table for Approach Two (Out of Sample Prediction) . . . . . . 93
2.18 Variable Importance Based on Approach Two . . . . . . . . . . . . . . . . . 94
2.19 Contingency Table for Approach Three (Out of Sample Prediction) . . . . . 95
2.20 Variable Importance Based on Approach Three . . . . . . . . . . . . . . . . 96
viii
List of Figures
ix
Abstract
This dissertation contributes to our knowledge of how social factors (e.g., homophily, the
existence of others) influence individual business behaviors. To explore the influence of
social factors, the Multi-level Marketing (MLM) Industry is especially suitable because social
and business networks are interrelated. MLM is a major type of direct selling wherein
Independent Business Owners (IBOs) recruit others, who become IBOs themselves. IBOs in
MLM are pivotal because they both serve end-customers and expand and maintain the IBO
network. However, as these IBOs are not firm employees, understanding the relationships
among them is critical. Two important questions are addressed. How is the IBO network
created (Essay One)? How do IBOs influence the activity of other IBOs (Essay Two)?
Essay One, “Social versus Economic Factors in Business Network Formation: An Em-
pirical Analysis of the Multi-level Marketing Industry,” explores how IBOs form networks
between IBOs and new joiners. Unlike social networks, where relationships are mainly driven
by social motives, IBOs and new joiners in MLM also expect economic returns. Thus, I in-
vestigate network formation in terms of social and economic factors. Moreover, I investigate
how the relationship value between the IBO and the new joiner affects the new joiner’s fu-
ture business activities. By employing a structural Bayesian two-sided matching model, I
find that both factors significantly affect network formation. However, economic factors are
more important than social factors, playing a significant role in determining the network
relationship. Also, social factors are significant in forming the network relationship, but are
less influential driving new joiners to engage in business activities in the future. Finally, new
joiners who are driven primarily by economic factors at the formation stage show significantly
more business activity.
Essay Two, “The Impact of Existence of Others on Inactive Behavior in the Multi-level
Marketing Industry,” examines one specific aspect of IBOs’ turnover behavior. There is a
vast literature on turnover, yet its main focus is on traditional hierarchical organizations.
MLM has a unique industry environment, which suggests that previous findings may not
apply. Specifically, relationships among IBOs are highly emphasized; they are often seen
as personal, yet competition among IBOs remains despite the importance of coordinating
efforts. Given these unique features, I focus on how IBOs’ inactivity can be quantitatively
x
understood in relation to three different dimensions of others (Individual, Network family,
and Proximity). One focal aspect is status, which is awarded by MLM firms to successful
IBOs to motivate IBOs. Analyzing county-level and individual-level data, I find consistent
evidence that being near high status IBOs decreases inactivity — a protective effect. The
results hold true at the network family level and for proximity. Also, IBOs with a larger
family tended to stay active. However, having more IBOs in proximity increased inactivity,
indicating that competition in proximity seems inevitable. Lastly, neural network is applied
to predict inactivity, which reveals a similar conclusion about the role of high status IBOs.
These findings have implications for the MLM industry, which has an interest in maintaining
the IBO sales force.
xi
Chapter 1 Social versus Economic Factors in
Network Formation: An Empirical Analysis
of the Multi-level Marketing Industry
1.1 Abstract
Multi-level Marketing (MLM) networks are unlike other business networks, where relation-
ships are mainly driven by economic benefits. According to the existing understanding of
MLM networks, relationship formation between distributors (IBOs) and new joiners is based
mainly on social aspects. I leverage a novel dataset from a large MLM firm to examine
the importance of economic returns versus social aspects in determining the formation of
networks. I also examine the role of social and economic factors, which contribute to the
value of relationships between the IBOs and the new joiners, in explaining the new joiners’
future business activity. T consider the disparate perspectives of the current IBOs and the
new joiners, I use a two-sided matching model using hierarchical Bayesian estimation. The
results suggest that network formation is determined by potential economic returns on both
sides. Business-support-related information matters in business relationship formation for
new joiners. IBOs, on the other hand, place importance on information that could signal
the future potential income from new joiners. As social factors, demographic homophily and
geographic distance also had a significant impact on business relationship formation, yet
their importance was about one third that of economic factors. Using the recovered social
and economic factors, I also share evidence that the new joiners facing business relationships
with greater emphasis on economic factors tend to engage actively in the business.
1.2 Introduction
Individuals are part of both personal and professional (or business) networks. Business net-
works are different from purely social networks in that there is economic and social exchange
between individuals in these networks (Grayson 1996). In other words, both the static and
1
dynamic structure of relationships has an impact on economic flows between individuals in
these networks. A very large and important industry where such networks determine eco-
nomic outcomes is the Multi-level Marketing (MLM) industry. In this industry, the main
mechanism of distribution is an organized network of independent distributors (typically
referred to as Independent Business Owners or IBOs). Firms in this industry do not employ
a dedicated sales force, relying instead on the network of IBOs to market and sell their
products. Thus they prefer to give up direct control in exchange for lower cost distribution
and advertising/retailing costs (Biggart 1989, King and Robinson 2000, Jain et al. 2015).
For the IBOs, the economic returns come from two major sources - the margin from selling
products to end customers and bonuses generated based on the selling activities of recruited
sub-distributors (Grayson 1996, Peterson and Wotruba 1996, Jain et al. 2015). This creates
an incentive for IBOs to grow (and maintain) the network by recruiting sub-distributors.
The prevalent belief among participants in this industry is that IBOs should leverage their
social networks to find sub-distributors (Clothier 1997, Grayson 1996). However, recent
industry publications have begun to emphasize that IBOs need to focus less on social con-
nections or intimates, noting that most successful IBOs do not have a significant proportion
of family and/or friends in their networks (Lilyquist 2016a). Other researchers have noted
that depending on beliefs about the friendship, existing social ties can either hurt or improve
business effectiveness (Grayson 2007). However, there is little research documenting the role
of economic versus social aspects in the development of these business networks.
In this research, I examine business network formation in the MLM industry, focusing
on the relative roles of economic versus social factors. This specific industry provides a
good environment to examine the formation of the individual business relationship, when
there can be a conflict or co-presence between the social and economic aspects. I leverage a
novel dataset from a large MLM company spanning almost two years of network formation
among its IBOs. Specifically, I quantify the role of economic returns, which I term the
economic factor, and social homophily (demographic and geographic), which I term the social
factor, in determining network relationships. Through this, I shed light on how IBOs select
individuals to partner with or vice versa, and I investigate whether the prevalent belief about
using social connections to build networks has empirical support. Such social-professional
relationships differ from other non-individual-based business relationships, which mainly
emphasize business potential, such as partners that can generate significant values with
better sourcing process abilities (Ni and Srinivasan 2015) or experiences (Sorensen 2007). In
addition, using the recovered social and economic factors determining the relationship value
between an IBO and a new joiner, I explore whether these factors affect the new joiner’s
business activity. There has been some work focusing on the MLM industry in general. Some
2
of the topics include the motivation to leave or join the business (Jain et al. 2015, Wotruba
and Tyagi 1991), institution setting comparison (Brodie et al. 2002), determinants of business
outcomes post relationship formation (Frenzen and Davis 1990), and size and profitability of
the network (Coughlan and Grayson 1998). However, business relationship formation has not
been a main focus. According to the MLM Executives Industry Survey, the average IBO uses
76% of his or her total work hours in building business networks and managing an existing
network of IBOs as opposed to engaging in retail selling (Coughlan and Grayson 1993, 1998).
This indicates the importance of examining how IBOs form business relationships with one
another. There is some evidence that whom one chooses as a business partner is critical
for success (Coughlan and Grayson 1998, Crittenden and Crittenden 2004, Allen 2016), and
that IBOs and the new joiners engage in purposive seeking behavior (King and Robinson
2000). Also, in order to grow the distributor network from the MLM company’s perspective,
it is crucial to understand what type of sponsoring relationship leads the new joiners to
actively engage in the business later on. I conduct an exploratory analysis to examine the
question. Due to lack of behavioral data, a vast majority of previous works on the MLM
industry relied on simulation-based approaches (Legara et al. 2008, 2009), survey data or
experiments (Coughlan and Grayson 1998, Jain et al. 2015, Grayson 2007) or interviews
(Albaum and Peterson 2011, Grayson 1996).
The MLM industry is controversial, especially as many of its critics confuse it with
pyramid schemes or Ponzi schemes (Bonoma 1991). The latter are illegal while the former is
legal, as clarified by the United States Federal Trade Commission in 1979 (see also Albaum
and Peterson (2011), Keep and Vander Nat (2014), Peter J. Vander Nat (2002) and Bosley
and McKeage (2015)). While it is not easy to distinguish one from the other (Taylor 2011),
the main distinction comes from the compensation rule vis-a-vis recruiting. Specifically, if
performance bonuses are generated from the legitimate sales of goods to consumers, or to
oneself, or from the sales generated by people who are recruited, as opposed to the mere act
of recruiting other distributors, then this activity is legal (Xardel 1993, Albaum and Peterson
2011, Legara et al. 2008, Taylor 2011).1 The industry is also seen as unpleasant to work in
due to the high pressure of recruiting new people, the difficulty of selling products and the
alienation of family and friends (Grant 2012). Though understandable, there are a lot of
misconceptions regarding these aspects of the industry, and other observers have tried to set
the record straight (Coughlan 2012). Another misconception is that this industry is based
on a “flawed” business model and is therefore doomed; however, the industry in the United
1
Xardel (1993) explains in detail the criteria to distinguish legitimate MLM companies from illegal pyra-
mid schemes - see Appendix 8. Most MLM companies provide specific guidelines that clarify the legality of
their business, and some note explicitly that bonuses are a function of a threshold level of business revenues
generated by an IBO’s network.
3
States has been operating for 50 years and continues to grow, recording a 4.8% increase in
sales in 2015 relative to 2014, reaching $36 billion and accounting for 37% of the retail sales
generated globally in the MLM industry (Tortora 2015). A recent estimate from the World
Federation of Direct Selling Association (WFDSA 2016) suggests that there were more than
103 million agents all across the globe in 2015 — an increase of more than 20% from just
five years ago.
In the MLM industry, IBOs do not receive fixed compensation from the firm. The
firm gives IBOs freedom to build their networks (i.e., hire sub-distributors), does not set
fixed quotas and allows them to decide on the amount of time they spend on recruiting
versus selling. IBOs therefore have direct control over the time and effort they put into the
business and also the consequent returns from the business. However, the firm sets clear
incentives, rewarding IBOs with performance bonuses based on the product sales resulting
from network activity. Thus, network formation is a central issue in this industry (King and
Robinson 2000). Having larger networks is critical, as it allows IBOs to grow the business,
“multiplying” their efforts by penetrating new markets through advertising, distribution and
sales (Peterson and Wotruba 1996, Jain et al. 2015). Also, the challenge of IBOs is to build
and maintain one’s own network of IBOs and thereby benefit from the work of those recruited
(Brodie et al. 2002).
As noted above, both industry insiders and outsiders believe that growth of this network
is best achieved via leveraging the IBO’s social network because these people are the easiest
to reach and convince (Clothier 1997). However, other researchers have also pointed out
that leveraging social relationships can create potential conflicts, since business dealings
can distort the existing relationship and the partners can have different role expectations
(Bloch 1996, Grant 2012, Grayson 2007). Other research notes that selecting the “right”
individuals, maintaining their motivation, and developing their skills are crucial to success
in this industry (Crittenden and Crittenden 2004). This argues for choosing IBOs less for
social reasons than for business reasons. There is also evidence that the network structure
contains information and therefore can potentially affect indivudals’ actions (Centola 2010).
In order to investigate these issues, I specify a two-sided matching model and take it to
the data. Two-sided matching models are often used to model the underlying preferences
driving the observed match between two sides (with mutual agreement). Such models have
been applied in various settings, e.g., the marriage market (Logan et al. 2008, Hsieh and Lee
2012), the education market (Boyd et al. 2013), the mutual fund market (Park 2008, Chen
2013), the sourcing market (Ni and Srinivasan 2015) and the sports recruiting market (Yang
and Goldfarb 2015). The data comprise matches (between existing IBOs and new joiners) in
a specific geographic area of the Unites States over a twenty-two month period. I estimate
4
the model using hierarchical Bayesian methods.
In brief, results show that while social factors (measured via demographics, income sim-
ilarity, and geography) do play a role, both parties are impacted significantly by economic
considerations in finalizing matches. New joiners face a higher relationship value when join-
ing IBOs that can provide higher levels of business support. IBOs tend to prefer new joiners
who are likely to be more motivated to make economic returns. Interestingly, contrary to
expectations, the level of the distributor in the network hierarchy does not influence the
preference of new joiners. To quantify the relative importance of the economic and social
factors from the realized matches, the social factors had about one-third the importance
of the economic factors at the aggregate level. Also, based on my exploratory analysis of
whether economic or social factors explain the new joiners’ future business activity, I find
that new joiners whose relationships have higher economic value engage more actively in the
business.
Essay One adds to the literature on business networks, network formation, and the MLM
industry on multiple dimensions. First, I am able to model both economic and social factors
that go into network formation. While the industry and previous research argues that
social factors are the primary force in network formation, I find evidence that economic
factors matter substantially as well. The two-sided matching approach isolates the different
considerations on both sides of the match. Second, I also examine the role of network
structure as a signal enabling network formation. Third, using the recovered social and
economic factors, which constitute the relationship value between IBOs and new joiners at
the relationship formation stage, I am able to draw an inference on the new joiner’s MLM
business engagement. Finally, to the best of my knowledge, this is one of the first papers to
use actual behavioral data from the MLM industry.
The remainder of the paper is organized as follows. In section 1.3, I describe the institu-
tional setting as well as the conceptual and empirical industry background and conceptual
framework used to guide the empirical analysis. I describe the research setting and data in
section 1.4. Section 1.5 describes the modeling framework and results. I discuss findings in
section 1.6. Section 1.7 concludes with a discussion of the study’s limitations and suggestions
for future research.
5
1.3 Institutional Background
6
sell products made or selected by the MLM company. Unlike employees in the traditional
business domain, IBOs have a lot of freedom in how they conduct their business. For instance,
the IBOs can sell in any territory or region, using their own operations without any direct
supervision, receiving the appropriate commission as long as they operate within the policies
and procedures defined in the MLM company regulations. Some of the examples of these
regulations include product claims that can be made, the type and amount and content of
advertising, specific locations where sales can or cannot be made, as well as the specifics of
the compensation plan (Biggart 1989, King and Robinson 2000).
New joiners become part of the business through multiple routes. They can meet their
future sponsor distributor if they are already end customers, or if they are friends or relatives
or business acquaintances. They may be invited to a meeting organized to explain the
sales and marketing plan (some IBOs, typically the successful ones, arrange these meetings
periodically). Others travel great distances to meet people and explain the MLM setting to
them (Xardel 1993). Potential joiners can also leverage the company’s website, which reveals
distributors’ contact information. If a new joiner is interested in the sales and marketing
plan of the MLM, s/he agrees to be sponsored by an existing distributor. Figure 1.1 depicts
the network formation process. In Figure 1.1(a), each distributor in the business is labeled
by a number according to his/her position in the structural hierarchy. In industry parlance,
this number is denoted as a “level.” When the new joiner joins the business, his/her level
is determined by the level of the sponsoring distributor. For example, if in Figure 1, Red’s
level is 5, all the IBOs who are directly recruited by Red (both Blues) are one level below –
represented numerically as level 6. Level does not necessarily reflect how long the distributor
has been in the business as it is solely dependent on the level of the sponsor. The Blues are
now denoted as Red’s “direct downlines” while Red is each Blue’s “upline.” The new joiner
can be connected only to one upline and in most cases cannot change his/her upline. All the
IBOs who are directly or indirectly connected to Red from below (e.g., Blues and Oranges)
are denoted as Red’s “total downlines.”
Each distributor has two potential sources of income. The first source is via the sales
of the products directly to the end consumer, by selling products at a margin. The second
source is via the sales of his or her recruits, the sales of the people they in turn recruit,
and so on. For example, in Figure 1.1(a), Red’s second source of income arises from the
sales of his total downlines (Blues and Oranges) as a form of bonus. As noted earlier, this
second source of income makes MLM different from single-level direct selling. This second
source also provides the impetus and motivation to each distributor to manage and expand
his/her network. In general, the relative contribution of the two sources of income varies by
status, hierarchical position within the network and experience in the business. Typically,
7
the higher the status, the higher the hierarchical position and the longer the experience, the
higher the contribution of income from the network.3 I now detail the social and economic
factors that could influence new joiners and potential uplines in terms of network formation
in MLM settings.
1. Demographics:
Network formation based on common observables has been documented in both em-
pirical and theoretical papers (Christakis et al. 2010, Currarini et al. 2009). As noted
above, distributors are “pushed” to interact with friends and family in the hope that
these social contacts will become part of their business networks (King and Robinson
2000). It is also easier to interact with others who are similar to oneself (Harrison
and Klein 2007), and such interactions may lead to a future business partner. Thus, I
expect that common observable demographic characteristics will increase the potential
to build these relationships, especially as previous research has documented that this
commonality helps to build trust (Narayan et al. 2011).
3
For example, in the estimation sample, the high status IBOs get 80% of their income from the network,
while the average distributor (across all status levels) gets 31% of his/her income via the network.
8
2. Geographic Distance: Along with demographics, geographic distance has often been
incorporated in the matching settings as a proxy for homophily as similar people are
likely to be in close physical proximity in terms of home and/or work (Boyd et al.
2013, Geweke et al. 2003). Also, geographic proximity has been identified as one of
the main factors in engendering word of mouth towards product or service adoption
(Nam et al. 2010, Bell and Song 2007, Albuquerque et al. 2007, Choi et al. 2010). New
joiners in the MLM industry rely heavily on word-of-mouth information before taking
the plunge (Legara et al. 2009). Thus, if the new joiners are located close to IBOs, they
may have a higher chance of hearing product testimonies. Closer physical proximity
also increases the probability that the distributor is known to the new joiner.
New joiners are often advised not only to look for the right company with the right product
(Collamer 2013, Lilyquist 2016a), but also to engage in thorough research to find the right
person to work with as a sponsoring upline distributor (Allen 2016). In the MLM industry,
the key leadership relationships are those between IBOs and the member who recruited them
into the organization (i.e., their uplines). As mentioned earlier, each distributor has a lot of
motivation to help the new joiners to succeed in their business, mainly because the distributor
9
receives a commission based on the sales made by his/her recruits and all the recruits the
new joiners will eventually sponsor. In addition, in most MLM companies, the distributor
Rules of Conduct require all uplines to train and supply IBOs they have recruited/registered,
as well as offer multiple ways to participate (Xardel 1993, Crittenden and Crittenden 2004).
The more a distributor is willing to build and succeed in the business, the more she will
focus not only on recruiting, but also on training the new recruits on best business practices
(King and Robinson 2000). As a result, new joiners will experience higher relationship value
when partnering with IBOs with established reputations and demonstrated business success
(Bonabeau 2002). The challenge for joiners is to assess the level of business support prior to
choosing which distributor to link with, based on physical contact, word of mouth or online
research (Lilyquist 2016a). They typically do this by examining the following characteristics
of existing IBOs.
1. Status from Business Performance: As the founder of a leading MLM firm said, “We
have two forms of reward in this world, one is recognition, and the other is dollars. I
employ them both in the business” (Klebnikov 1991). Consistent with the academic
definition of status (Hu and Van den Bulte 2014), status in this industry reflects social
stratification based on economic success as well as reputation and recognition. Specif-
ically, status is awarded based on two metrics — one’s own financial activity and that
of one’s connected downlines. All the IBOs who devote enough time and effort to the
business are qualified to be ranked in the higher status, reflected by different signifiers.4
Based on their financial achievement, IBOs not only earn additional financial rewards,
but also receive a significant amount of social recognition (Xardel 1993, Amway 2015).
Status does not follow from the formal lines of authority as in conventional organiza-
tions, but is a function of ability and performance, creating a social hierarchy that is
different from network structural hierarchy (Pratt 2000). The recognition manifests
itself in being highlighted at company conventions and meeting, being asked to instruct
other IBOs who are not as successful, etc. In general, the higher the status, the more
the time spent by an distributor in personally teaching new joiners how to build a
successful business (Pratt 2000).
Thus, from a new joiner’s perspective, status of the uplines is likely to be a signal
about the the uplines’ degree of success, and how much the distributor is recognized
and respected in the business. The higher the status, therefore, the more desirable the
connection.
4
This signifier varies by firm. One large MLM firm uses job titles to signal status. Another one uses
precious gemstones as signifiers.
10
2. Duration in the Business: The MLM industry is known for a relatively high turnover
rate (Wotruba and Tyagi 1991, King and Robinson 2000). There could be multiple
reasons why this is the case. First, compared to other industries, there is a high degree
of freedom for each distributor and there are no barriers to exit. Second, as more
than 90 percent of IBOs are engaged in the business as a part-time job supplementing
their main income (DSA 2015b, Dir 2017), it may be unnecessary for them to stay
in the business for a long period of time. Third, some IBOs, after joining, question
the income potential in the MLM industry and exit. In this situation, long tenure by
an existing distributor signals that s/he has probably exceeded the income threshold
that is required for him/her to stay with the MLM company. In other words, s/he
is financially successful (Wotruba and Tyagi 1991). Thus, the longer a distributor
remains in the business, the more likely s/he is to attract new joiners. On the other
hand, IBOs who are relatively new to the business may be more enthusiastic about all
aspects of the business (Legara et al. 2008), and this could be valued highly by the
new joiners. Thus, it is an empirical question to examine how and to what degree long
tenure in the business generates higher relationship value in the new joiners’ network
formation decisions.
3. Relative Level Based on the Network Structure: Level is a measure reflecting one’s hi-
erarchical location in the tree-like network structure or network hierarchy (see Figure
1.1). The structure of an MLM network is hierarchical, which may suggest that IBOs
located at a higher level would earn more. However, given the compensation system,
the profit opportunity is technically the same for each IBO, regardless of where s/he is
located in the network hierarchy (Lilyquist 2016b). In addition, as explained earlier,
an IBO’s level is determined by his or her sponsoring IBO when s/he joins the business,
and it does not change over time. Thus, distributors can even possibly earn more than
those who bring them into the organization (King and Robinson 2000).
So while this suggests that network hierarchy should not play a role after controlling
other correlated aspects, it may be hard for new joiners to understand this ex ante.
Prior literature has also suggested that the network structure contains (perceived)
information on power, knowledge dissemination and innovation within firms and can
affect adoption (Albert and Barabási 2002, Van den Bulte and Wuyts 2007). The
predictions for how network structure affects outcomes are mixed, with Legara et al.
(2008) showing (via an analytical model) that earnings are independent of location in
the network hierarchy, while Daquis et al. (2013) show (via simulation) that higher
levels can make more profits. None of these predictions are tested with behavioral
data or in terms of network formation, which gives us an opportunity to do so in our
11
approach.
12
joiners prefer to be attached to a distributor who has successfully built the business
multiple levels below by transmitting the business practices with high management
skills, when the downlines could be perceived as possible competitors?5
While the number of direct downlines and the existence of indirect networks can signal
the IBO’s focus and their ability in terms of recruiting and managing/expanding the
network, the number of retail customer proxies for the IBO’s ability to sell the products
to registered retail customers.
For the success of the business, selecting the right individual matters not only for new
joiners, but also for IBOs (Crittenden and Crittenden 2004). The MLM Executives
Industry Survey shows that an average IBO uses 44% of his or her total business
hours in managing the existing network of IBOs — the most time-consuming activity
in recruiting and retailing selling (Coughlan and Grayson 1993, 1998). Managing
network involves teaching, training and motivating downline distributors to both sell
the product and recruit new IBOs. Preserving and building the business network
not only takes time but also requires psychological commitment (King and Robinson
2000, Pratt 2000). Thus, IBOs need to maintain a good balance among selling the
product, maintaining the current network, and recruiting new direct downlines (King
and Robinson 2000). Because of the time commitment and psychological investment
required, the ability to find the right prospective IBOs has been recognized as one of
the keys to success in the business (Lilyquist 2016b). Thus, IBOs will try to seek for
the most suitable direct downlines to recruit given their circumstances.
Whereas new joiners have abundant information to judge the possible uplines, current
uplines do not have much information to judge whether the new joiner will be pro-
ductive enough to contribute to their income steam in the long term. Individuals join
the MLM industry for various reasons, which in turn leads to different levels of time
and effort allocation in the business (DSA 2015c). For example, some people enter the
industry to make it their main source of income, while others enter it to obtain supple-
mental income. This variation is one of the reasons the turnover rate is especially high,
and it accounts for the relatively small average earnings in the industry (Albaum and
Peterson 2011). Given that the IBO’s commission is based on the downlines’ product
5
A similar argument holds for the number of total downlines — the number of direct downlines. But
unlike the number of direct downlines, which IBOs know, there is uncertainty on information about the
entire network below them. The result is consistent with what I report in the results section.
13
selling and recruiting activities and that they need to put a lot of effort into motivating
and training new joiners (Coughlan and Grayson 1998), IBOs will prefer new joiners
who are more committed to the business.
1.4 Data
Data come from a leading global MLM firm that has been a pioneer in this industry.6 This
firm has more than 3 million IBOs operating in 100 countries. The main product categories
that the firms sells via its IBOs are nutrition, health & beauty, bath & body and home decor.
The data cover a twenty-two-month period from January 2006 to October 2007, with activity
recorded at the monthly level. The data include detailed information on each distributor’s
identification number, hierarchy and status, network-related characteristics, demographic
variables (age, gender, marriage status, language, ethnicity) and Zip code of the distributor.
The time-series data allow me to map out the network formation and growth for distributor
over this twenty-two-month period. I first explain how I set up the data for the model
presented in §3 and then present descriptive statistics.
14
indicated in the column ‘Transitioned IBOs.’ Among the 941 new joiners, 7.43% successfully
added a new direct downline within the same month.
Note that in any month, an upline can match with one or more downlines. The most
common type of match is a one-to-one match, where an upline matches with only one joiner.
For example, if Gray would like to match an existing distributor for the network in Figure
1.1-a), she can match with four potential uplines - Red, Blue, Orange 1 and Orange 1. If
she chooses Red, then the network looks like Figure 1.1-b). However, it is possible for Red
to acquire more than one downline in a month. Table 1.2 shows the number of downlines
matched with uplines in each month. While the majority of matches are one-to-one, 6% of our
matches are one to more than one. The modeling approach therefore needs to accommodate
these types of matches as well.
I have data on three demographic variables - age, language and ethnicity. Table 1.3 shows the
proportion of language users and ethnicity in the estimation sample, including both uplines
and new joiners. Three languages (English, Spanish and Korean) and seven ethnic groups
(European, Hispanic, Asian, African American, Non-Orient Asian/Polynesian, Jewish and
Arab) account for a majority of data. The age distribution across our sample can be seen in
Figure 1.3.
In order to see whether networks in this industry showed homophily, I looked at patterns
of these three variables across uplines and downlines. There is significant and positive corre-
lation (0.34) between the age of uplines and downlines. The contingency table for language
and ethnicity is shown in Table 1.4 and Table 1.5. The cells showing the highest percentage
from the downline’s perspective are marked with ?. These data patterns on language and
ethnicity suggest that there is evidence for demographic homophily in the MLM network.
To incorporate this into the model, for each distributor-joiner pair, I used binary variables
to denote whether the pair shared the same language and ethnicity. For age, I used the
absolute difference in age between each distributor-joiner pair. As I focus on the homophily
effect based on the clear pattern in the data, I use the absolute difference, but I acknowledge
that the specification does not capture possible age asymmetry effects, for instance, whether
seniority plays a role in the choice of business partner.
Lastly, I incorporate dyad-level income for following three reasons. First, people may
interact at the work site and learn about supplementary income opportunities. According
to a survey data from Raymond and Tanner Jr. (1994), although most of the consumers
15
made product purchases at home, the workplace was the second-most-frequent location for
the business interaction, accounting for 39.5% of respondents. Second, it helps to capture
members of the social inner circle with similar income levels. Third, people with similar
income level are highly likely to live nearby in the same neighborhood. I transformed nine
income groups shown Table 1.3 into a dyad-level binary variable indicating membership in
the same income group or not, to capture income similarity.
Even within the Chicago area, geographic proximity is likely to affect both potential
new joiners’ and uplines’ preferences. I compute the centroid based on distance between the
new joiner and the existing uplines based on Zip-code information. The histogram on the
right-hand side in Figure 1.3 shows the realized distance distribution in our sample from the
new joiners’ perspective. The figure shows that most new joiners and uplines who formed a
relationship were physically fairly close to each other.7
Now let’s turn our focus to the variables that signal the attractiveness of distributors to new
joiners on economic grounds. I begin with the descriptive statistics of the high status in
Table 1.6.
In addition to the level in the network — a structural hierarchy, other measures of
hierarchy exists, which is the status. High status signifies success, typically based on earnings
and other contributions to the firm. The firm provided data on the most important signifier,
denoted here as “High status.” This is an extremely difficult signifier to earn as it is reserved
for the most accomplished IBOs. For example, among 884 uplines in our sample, only 7.5%
of IBOs currently hold (or have held) high status.8 As a result, achieving this status is
highly aspirational for all IBOs. The earnings of high status IBOs were much higher than
those of low status IBOs with an average monthly payment of $2669 versus $17 (based on
2007-08 data provided to us by the firm). High status IBOs also get a “recognition pin” and
are featured in distributor publications, get to participate in the annual growth-incentive
program and are invited to an all-expense-paid conference to obtain useful resources and
connections to help them grow as leaders.
Duration in the business is measured as the number of months the distributor has been
part of the business. I see a wide variation in the distributor focus related variables. The
7
As noted earlier, I have censored the distance at 35 miles, covering about 80% of network formation in
the Chicago area. Without censoring, the distance (in miles) between each pair is distributed as follows:
minimum = 0, 1st quartile = 1.9, median = 6.1, mean = 130.7, 3rd quartile = 24.320, max = 7371.
8
Status is evaluated every fiscal year, and the data provided do not record the year-by-year status of
distributor. However, I have data on whether the distributor ever achieved high status and his/her status
at the end of the data. In addition, a new joiner cannot have high status. Using these features, I am able to
obtain the status of a distributor quite precisely.
16
number of direct downlines and the number of total downlines reflect the degree to which
the distributor has been focusing on building network activity, and the number of retail
customers indicates that of product-selling activity (Figure 1.4). Among 884 uplines, 19.5%
do not have any prior recruiting experience before the match I observe (having zero direct
downlines) and 63.8% of uplines do not have any retail customers.
In contrast to the prevailing beliefs in the MLM industry, Table 1.7 shows that being a
high status member is relatively highly correlated with the distributor focus variables (the
number of total downlines, direct downlines, and retail customers) but not very correlated
with level (-0.05). This implies that location in the network hierarchy is not a predictor of
business success. As expected, I see a high correlation between experience and the number
of direct downlines. In terms of level, people who are located in the lower part of the tree
generally have less experience in the business.
In order to proxy for a new joiner’s economic potential, I use data on his or her annual
financial earnings filed as part of his or her application. Using this variable, uplines can
infer the degree to which a downline relies on this business and “hunger” to succeed in the
MLM business. Table 1.3 shows that about a quarter of all applicants have very low annual
incomes (less than $15,000), while about 40% earn above $50,000 annually.
17
product caused by other consumers’ choices, as the exact same product can be offered to
the consumer. In the two-sided matching model, however, the business relationship decision
of an IBO or a new joiner affects how others make their decisions, because no business
partners are exactly the same, and the formation of one relationship limits others’ choice
sets. More importantly, unlike a shopping setting where consumers choose the product,
here, both IBOs and new joiners need to mutually agree in order to form a match. Second,
unlike the discrete choice model, where the choice set is usually exogenously pre-specified,
in this model, a decision maker (an IBO or a New joiner) cannot take the choice set as
given as (potentially) not all decision makers on the other side may be willing to match with
him/her. This issue has been termed in various ways as the endogenous choice set issue
or the simultaneity issue in the two-sided matching literature (Echenique et al. 2010, Hsieh
and Lee 2012). What this implies is that the preferences of the IBOs and the new joiners
determine the willing partners for each IBO and new joiner, which constitute a choice set or
an opportunity set, as defined in a later section). In our approach, I use Bayesian Markov
Chain Monte Carlo methods to simulate these unobserved opportunity sets.
The major components of the model are the IBOs’ and the new joiners’ latent match
utility, which is the relationship value that reflects preferences on both sides. Together
with the rules of the matching game, I recover the latent match utility and the associated
parameters that determine the equilibrium match, as reflected in the dataset. Using this
model, I can understand the determinants of the business relationships in the MLM industry.
9
18
1.5.2.1 Two Sides
Two separate sides are involved in the model — the IBOs’ side and the new joiners’ side -
and I model the relationship formation between the two sides. In forming a set of people
on each side, I have no information on a potential set of IBOs and new joiners who wanted
to form business relationships, but failed to do so. Therefore, the model is based on the
observed matches, as is often the case with empirical matching models. This means that the
realized relationship identifies the set of people on the IBO and new joiner sides, and nobody
on the IBO and new joiner side remains unmatched (Sorensen 2007, Akkus et al. 2014, Yang
et al. 2009, Boyd et al. 2013). In turn, I do not address what motivates an individual to
become part of the MLM industry or why a certain distributor decides to enter the recruiting
market at a specific time, as I condition on the set of actual matches.
Let It and Jt denote a set of IBOs (or potential uplines, from the new joiners’ perspective)
and a set of new joiners (or potential direct downlines from the uplines’ perspective) at time t,
which constitutes a matching market, where t = 1,2,....,T. The set of new joiners, Jt , consists
of all the newly joined business owners who are connected to one of the existing distributors
by the end of time t. The set of IBOs, It , includes those who formed a relationship with
the counterpart new joiners at least once in month t. It is composed of two types of IBOs:
the first type, the IBOs who entered the MLM business before month t, and the second
type, those who entered the business during month t as new joiners and in turn successfully
recruited other new joiners within the same month t, or so-called transitioned IBOs. Table
1.1 column ‘Number of transitioned IBOs’ indicates such IBOs.
Due to the one-to-many match, the unique aspect of this industry, the new joiners can be
connected to only one of the existing IBOs by joining the business (exclusivity in recruiting).
However, the existing IBOs can recruit several (sit ) new joiners in a given market t. The set
of all the potential matches is Mt = It × Jt , where a matching µt is a subset of Mt and the
realized match in the data. IBO i’s set of recruited new joiners is defined as µt (i), and new
joiner j’s connected IBO is denoted as µt (j). Following this notation, the match between an
IBO i and a new joiner j in market t can be expressed in three equivalent ways: ij ∈ µt ,
µt (i) = j, and µt (j) = i. I define the temporal duration of market t as one month to match
the granularity of the data. To keep the study tractable, I model the market as independent.
Each new joiner and distributor on each side seeks to form a relationship with an individ-
ual from the opposite side. Formation of the relationship is governed by the latent utility
of the relationship, or the relationship value for the set of all the potential matches Mt ,
19
which is represented in the Vt matrix. The number of rows in Vt matches the number of
IBOs in It (nIt ), and the number of columns matches the number of new joiners in Jt (nJt ).
Specifically, the relationship value between distributor i and new joiner j in month t is de-
noted as Vijt , which constitutes the Vt matrix. This Vt is unobservable, and it is what I am
trying to recover from the dataset. For ease of presentation, I suppress the month t notation.
V11 . . . V1j ... V1nJ
.. . . ..
. . ... ... .
V = Vi1 . . . Vij ... VinJ (1.1)
.. ... ..
. ... ... .
VnI 1 . . . VnI j ... VnI nJ
I assume that the latent utilities, or relationship values, are pair-dependent, and shared
between the two sides by the fixed sharing rule. In other words, IBO i’s utility of being
connected with new joiner j is the same as new joiner j’s utility of being connected with
IBO i - the variation across these relationship values explains the network formation. Thus,
increasing the utility of either the IBO’s or the new joiner’s side will generate a higher
relationship value. Even if the relationship values are defined at the pair level, IBO i will
prefer a match that confers relatively higher Vik as compared across the characteristics of
new joiners k ∈ J. Similarly, new joiner j will prefer a match that provides relatively higher
Vkj as compared across the IBOs k ∈ I.
In order to make the model tractable, I assume a complete information game on each
side, which is a common practice in the empirical matching model. It seems plausible that
new joiners who are somewhat interested in the business engage in gathering information
on the potential upline IBOs. Some companies reveal the existing IBOs’ information on the
website to facilitate the decision making. Also, from the IBOs’ perspective, in addition to
drawing future direct downlines from existing social relationships, there are different ways
to invite the new joiners into the business. One way is through a meeting organized to
explain the sales and marketing plan (Xardel 1993). Although I have evidence of purposive
searching between new joiners and IBOs (King and Robinson 2000, Allen 2016), as I do
not have information about the detailed matching process (e.g., whom a person on one side
approached on the other side in order to settle down with a final observed match), I rely
on this assumption. However, the opportunity set over which the decision makers maximize
their utility will be a subset of the people on the other side, which is explained in the following
20
section.
This pairwise utility is assumed to be shared between the two sides by the fixed sharing
rule. This assumption has been used in different settings in the empirical matching literature
involving unassortative matching (Sorensen 2007, Park 2008). Even if the relationship value
is specified at the pair level, the preference of each new joiner over the potential IBOs will
be affected only by the characteristics of the IBOs (including the pair-dependent character-
istics) and the error term, as the values related to the joiner’s own characteristics are fixed.
Similarly, the preferences of potential IBOs regarding the new joiners will be affected only
by the characteristics of a set of new joiners and the error term.
Methodologically, this use of pairwise latent utilities with a fixed sharing rule assumption
is known to handle the multiple equilibria issue, which is often raised in the two-sided nature
of the empirical matching model (Uetake and Watanabe 2012b, Hsieh and Lee 2012). Few
studies have carried out empirical estimation, suggesting broadly three different ways aside
from this approach. The most prevalent approach is utilizing a transfer between the two sides
in forming a relationship (e.g., a wage or price that gets transferred between the two sides in
creating the match). This transfer allows us to derive a joint utility, and it can be estimated
using non-parametric maximum score estimation, so-called the Fox estimator (Echenique
et al. 2013, Fox 2008, Yang et al. 2009, Pan 2014, Yang and Goldfarb 2015, Zamudio et al.
2013). Some rely on ex ante knowledge of how the match is formed, or assume a type
of equilibrium, or use a partial identification strategy (Uetake and Watanabe 2012b, Boyd
et al. 2013, Uetake and Watanabe 2012a). The last approach relies on a vertical preference
structure in the market, where everyone agrees on the preference ordering of the other side
and vice versa (an assortative matching) (Clark 2006, Chen 2013, Lee 2014, Agarwal and
Diamond 2014, Ni and Srinivasan 2015).
However, business relationship formation in the MLM industry does not involve explicit
transfer. I do not have information on the business relationship formation. Also, the social
variables in our study are pair dependent (e.g., demographic homophily, geographic distance),
which prohibits me from imposing vertical preference restrictions on both or one side of the
market. I therefore follow utility specifications that model the joint utility of both sides, as
if both sides are sharing the common utility generated by the relationship, which is often
encountered under matching without transfer (Sorensen 2007, Park 2008). Together with
the fixed sharing rule, the utility of two sides align, ensuring unique equilibria without the
need to assume transfer, an equilibrium selection rule, or a vertical market structure. Also,
the utility of a business relationship is often determined by the synergy of the characteristics
of the IBOs and the new joiners.
21
1.5.2.3 Equilibrium Outcome
In the structural two-sided matching model, the observed match is the optimal equilibrium
outcome. The solution concept I adopt is a pairwise stability, which is the common equi-
librium notion in the matching literature (Roth and Sotomayor 1992). According to this
concept, no IBO and new joiner prefer each other over their observed relationship in the
data. Or, a matching µ is stable if it cannot be improved by any IBO-new joiner pair (Roth
and Sotomayor 1992). In the college admission model, a stable matching always exists (Roth
and Sotomayor 1992), and the fixed sharing ensures that the model has a unique stable so-
lution. When an IBO and a new joiner prefer to abandon their existing relationships with
their own business partners, which is observed in the data, and form a new relationship with
each other, they form a blocking pair. In order for the match to be stable, such blocking
pairs should not exist. Thus, the pairwise stability is defined by a no-blocking-pair condition
and can be specified with a set of inequalities.10
The set of inequalities that reflects the non-blocking condition is based on intuitive logic.
To illustrate, let’s assume that an IBO i and a new joiner k are unmatched. The reason i and
k did not form a business relationship is that at least one side had an alternative business
partner generating higher utility, and the alternative partner is what we observe in the data.
An analogous example would be a marriage market. The reason a woman and a man did
not marry each other is that at least one person had an alternative partner generating a
higher utility, which led him or her to marry the alternative partner. Based on this logic, the
relationship value between unmatched i and k (Vik ) should be lower than that of i and k’s
formed business partners observed in the data from at least one side, which can then justify
the unmatched outcome. Thus the relationship values for all the unmatched relationships
will be restricted with upper bounds.
On the other hand, let’s assume that an IBO i and a new joiner j are matched with
each other, which we observe in the data. Based on i’s perspective, retrospectively, i must
have built a relationship with j, who offers the highest relationship value among potential
new joiners with whom i could have formed relationships but left behind. Now, as this is a
two-sided matching model, we also need to consider j’s perspective at the same time, as both
perspectives must have been satisfied in order to form a business relationship. Similarly, j
must have been connected to the IBO i, who offers the highest relationship value (Vij ) among
those potential IBOs with whom j could have formed a relationship. Thus, following this
10
In theory, the stability is defined by the following two conditions: individual rationality and the no-
blocking-pair condition (Roth and Sotomayor 1992). The meaning of individual rationality is that no upline
or new joiners form a match that is less preferable than being unmatched. However, individual rationality
does not apply in our set-up, as I assume that agents prefer to create relationships.
22
logic, the relationship values for all the matched relationships will be restricted with lower
bounds.
As the above illustration shows, a little complication comes from the one-to-many match-
ing characteristic; namely, an IBO can gather more than one new joiner, whereas a new joiner
can be connected to only one IBO. To present the non-blocking-pair condition mathemati-
cally, it is convenient to define the two concepts: V ij and V ij . I keep the notation consistent
with that used in previous literature. First, V ij is an upper bound when i and j are un-
matched pairs. V ij is the pairwise opportunity cost of deviating from one of their current
relationships that i and j have - observed in the data - and hypothetically creating a new
match with each other. Mathematically,
V ij ≡ max minj 0 ∈µ(i) Vi,j 0 , Vµ(j),j (1.2)
To explain, the first specification within the max bracket in equation (1.2) is from IBO
i’s perspective. If IBO i were to give up one of his/her existing relationships in order to
form a new match with new joiner j, it would be the one that provides him/her the lowest
relationship value. Again, as IBO i can recruit more than one new joiner, I need to compare
the current relationships IBO i has, in order to compute IBO i’s cost of deviation (the first
part within the max bracket). The second part within the max bracket is from new joiner j’s
perspective. Whatever s/he needs to give up in order to form a hypothetical relationship with
IBO i is the utility from his/her observed relationship, Vµ(j),j . After acquiring the abandoning
cost from both i and j’s perspective, the maximum value of the two is the opportunity cost
of deviating from their current match and forming a new match with each other.
Based on this explanation, for the observed unmatched pair ij, the relationship value Vi,j
must be lower than the opportunity cost of ‘creating’ the new relationship with each other
and abandoning their current match, V ij . If Vi,j were greater than V ij , the value that makes
it worthwhile for both i and j to give up the observed relationship, then i and j would have
an incentive to give up (one of) their current partners and form a relationship with each
other, causing a conflict with the no-blocking-pair condition.
Second, V ij is a lower bound when i and j are unmatched pairs. V ij is the pairwise
opportunity cost of i and j remaining together. In order to create the relationship ij ∈ µ,
i and j both must have given up all the other potential business relationships that they
otherwise could have formed. Mathematically, the remaining cost is computed as
V ij ≡ max maxj 0 ∈O(i) Vi,j 0 , maxi0 ∈O(j) Vi0 ,j , (1.3)
where O(i) and O(j) are opportunity sets for i and j. In the two-sided matching model, each
23
entity on each side chooses a relationship so as to maximize his or her relationship value over
the set of available opportunities, which is defined by opportunity set O(·).11 For instance, if
IBO i could have provided a higher relationship value than new joiner j”s observed business
parter based on data, µ(j 0 ), or Vi,j 0 ≤ Vµ(j 0 ),j 0 , j’ is in i’s opportunity set. In equation (1.3),
the first portion in the max bracket is the maximum relationship value that IBO i had to give
up for the relationship with new joiner j. The second portion is defined from the perspective
of new joiner j in a similar way.
As I mentioned before, there are two types of IBOs in It : the IBOs who entered the
market before month t, and those who entered during month t and successfully recruited
new joiner(s) within the same month. The transitioned IBOs appear both in It and Jt .
Although I do not observe the actual opportunity sets for both sides and thus rely on the
simulation, I do know that depending on when the new joiner applied to the business, the
maximum pool of IBOs to consider may change - the new joiners who became part of the
business after the transitioned IBO’s entry might have faced a larger pool of IBOs. Thus,
I adjust the opportunity sets accordingly, which is captured by the application order, d.12
Mathematically, O(·) is defined as below. The opportunity sets of IBO i and new joiner j
are:
O(i) = j 0 ∈ J : Vi,j 0 >Vµ(j 0 ),j 0 & (j 0 6= i)
(1.4)
O(j) = i0 ∈ I : (Vi0 ,j >minj 0 ∈µ(i0 ) Vi0 ,j 0 ) & (i0 6= j) & (di0 < dj )
(1.5)
In computing V ij , I need to consider only those who are part of the opportunity set, not the
pairwise utilities generated from all the agents on the other side. This is because each IBO
and new joiner have the opportunity to be paired only with anybody from the other side
who wishes to be connected with him/her, or prefers him/her to his/her current match, as
it is a mutual process.
Now using V ij and V ij , the non-blocking pair condition is characterized by a set of
inequalities following the two rules below (Sorensen 2007).
If all the pairwise utilities in V satisfy the two preceding equality conditions, no IBOs
11
Some researchers term it a feasible set or a choice set.
12
I acknowledge that the slight modification that I incorporate to include the IBOs who entered during
the month may introduce variations into the uniqueness condition. But instead of removing those IBOs who
entered during the month, I included them with the minor restriction to reduce the bias and reflect reality.
Observing more than twenty markets, along with using hierarchical Bayes estimation to pool information
from other markets in the inference can alleviate this potential problem.
24
or new joiners believe that they can improve their matches by dissolving their current rela-
tionship and forming a new one with each other, and the resulting match is stable. In other
words, these conditions require that each IBO match with the best new joiner among those
willing to match with the IBO, and that each new joiner match with the best IBO among
those willing to create the relationship.
where Xijt ∈ Rk is a vector of characteristics applied for IBO i and new joiner j to form a
paired utility, including [Durationi , DistributorF ocusi , Statusi , Leveli , EarningP otentialj ,
DemographicHomophilyij , Distanceij , IncomeSimilarityij ], reflecting both social and eco-
nomic variables. The list of variables included in the relationship value function is specified
in Table 1.8. The parameters that govern relationship values could vary by the composition
of IBOs and new joiners at a specific month. Thus, a vector of preference parameters αt is
defined to be month t specific. As the parameters are defined at the distributor-new joiner
pair level, αt cannot be individual specific. Also, αt cannot be defined at the pair level
because there is only one observation from each pair at a given month.
In order to estimate a vector of preference parameters, αt ∈ Rk , ij is assumed to follow
iid normal error with no loss of generality and to ease the computation complexity. The
implies the pairwise unobserved deviations from the mean utilities across the agents and
months.
1.5.3.1 Identification
The matching model can be seen as a variant of the discrete choice model; thus, in order
to identify the model, the standard assumptions need to be applied in the empirical setting
(McCulloch et al. 2000, Imai and van Dyk 2005, McCulloch and Rossi 1994). As is the case
in the discrete choice model, only differences of characteristics across IBOs or new joiners
have identifiable effects, as one’s relative utility (or preference) will not be affected by both
scale and level. Thus, in order to fix the scale, I standardized the scale of normal error, ,
to be 1.
In terms of identifying the IBOs and new joiners who did get their best match and who
25
did not, the variation in the pair-dependent variables allows us to identify those two groups
of people. If the new joiners care only about the IBO’s economic factors, then everybody
will have the same preference across the IBOs, and the preference over the potential IBOs
would be the same. The same holds for the potential IBOs, the assortative matching case
(Ni and Srinivasan 2015). What makes the preference different across each individual are
the pair dependent variables, such as geographic distance and demographic homophily, and
that identifies the IBOs who did get their best match and those who did not. Also, the
slots in the IBO’s side help us incorporate the situation where not everybody can possibly
be matched with their most preferred counterpart.
I employ Bayesian inference using a Gibbs sampling algorithm to estimate the two-sided
matching model. Bayesian estimation is highly beneficial when facing integration of high
dimensional latent variables (Sorensen 2007, Logan et al. 2008, Ni and Srinivasan 2015).
This is especially useful in a two-sided matching set-up, which involves high dimensional
integration (Sorensen 2007, Park 2008).
In the estimation process, I use data augmentation to recover the latent relationship value
Vt , which can nicely handle the difficulty of integrating high dimensional latent relationship
value variables (Geweke 1999, Sorensen 2007). Data augmentation can ease the complexity of
drawing Vt , which satisfies the stability condition of the college admission model. I impose
a conjugate normal prior for tractability in sampling the latent relationship value. Thus
αt follows normal distribution, αt ∼ N (µα , Σα ), µα ∼ N (µα , A−1 ), and Σα ∼ IW(ν, W ) .
Especially with the normal error assumption in the utility and the normal conjugate prior on
αt , the marginal conditional augmented posterior distributions for V in the Gibbs sampling
are truncated normal.
Also, in drawing latent pairwise utility V, it is necessary to simulate the simulating
opportunity set (Uetake and Watanabe 2012b). However, not every IBO or new joiner has
the same opportunities; the set of available opportunity varies for each person and depends
on the relationship value of the other person in the two-sided matching market.13 Thus,
unlike a general one-sided discrete choice model where the choice set is fixed and agents
are assumed to maximize the utility over a given choice set, a direct inference based on the
observed choice is not possible, as I do not observe the relevant choice set for each individual,
13
For instance, based on equation (1.4), new joiner j’s opportunity set is all the IBOs who could gain
higher utility if they were matched with new joiner j as opposed to their current match, µ(i). If new joiner
j proposes to form a relationship to one of the IBOs in O(j), the IBO will agree to do so as s/he can gain
higher utility compared to that of their current match. Thus, IBO i’s choice set is the opportunity set and
is dependent on the relationship value of the other side.
26
nor do I have supplemental data. Hsieh and Lee (2012) term this endogeneity in choice set.
The ML-based estimation approach proposed by Uetake and Watanabe (2012b) involves
approximating the opportunity set for each agent on one side based on all the preferences of
the other side through simulation, which determines the willing partners. However, Bayesian
estimation with Gibbs sampling can nicely handle such a complicated region of integration
and can embed such interdependency of relationship values when forming opportunity sets,
as I simulate the opportunity set conditioning on the previously updated αt and the values
from other pairwise V elements for each MCMC interation.
1.6 Results
I first focus on the role of social factors, represented by demographic homophily and ge-
ographic proximity. Except the age difference and geographic distance, all variables are
coded as binary dummy variables, either sharing the same characteristics (= 1) or not (= 0).
Our sample represents three types of language speakers, seven types of ethnicities, and nine
income groups, previously shown in Table 1.3. These variables capture pre-existing social
networks and homophily aspects. Based on the results, the coefficients for social factors,
represented by language, age difference, ethnicities, income groups and geographic distance,
are all significant.
In more detail, the coefficients for the same language and the same ethnicities are strongly
positively significant, and the same income group is marginally positively significant. This
indicates that IBO - new joiner pairs with similar backgrounds generated higher pairwise
utility, or relationship value, leading to a business relationship. The coefficients for age
difference and geographic distance are strongly negatively significant. IBO-new joiner pairs
exhibit a higher utility with a smaller age gap, implying that the distributor network is often
created among existing friends or individuals with similar backgrounds, who could enter into
a friend relationship, or who are already friends. When it comes to the geographic distance
27
at the Zip-code level, IBO-new joiner pairs who live close by tend to generate higher utility.
Although the estimated coefficients show how significant each variable is in explaining
the pairwise utility of the relationship value, it is unclear how to interpret the estimated
coefficient, which is an issue previously raised in the matching model. Therefore, following
Sorensen (2007) and Park (2008), I computed the marginal probability of the coefficient
to observe the marginal effect of each variable, holding other variables constant.14 This
shed lights on the impact of each variable on a choice probability based on a simple choice
scenario. To illustrate, let’s assume that there are two possible choice alternatives (i.e., two
new joiners). If the two new joiners have identical observed characteristics, the IBO’s choice
will depend entirely on the unobserved factors, and the probability that the IBO prefers one
new joiner to the other will be a random 50% (the same is true for a new joiner establishing
a relationship with one of the IBOs). Now, from the perspective of a new joiner, comparing
two otherwise identical IBOs, the probability that the new joiner prefers the IBO having the
same language as a business partner is 79%, 29% higher from the random 50%. Similarly, the
magnitude of the approximate marginal probability increase for those of the same ethnicity,
while other characteristics remain unchanged, is 28.8%. The marginal difference of a variable
in the choice scenario belonging to the social factors is reported in the right-most column in
Table 1.9. Based on this marginal probability, distance showed the highest absolute marginal
effects, followed by the same language.
The empirical results on social factors support the idea that unlike people’s general belief
about how the business relationship should be formed, those with similar backgrounds and
living close by tend to associate with each other as business partners. This result is consistent
with MLM industry practices whereby distributors tend to initiate business opportunities
from a list of acquaintances (Biggart 1989, King and Robinson 2000, Lilyquist 2016a), which
is a unique feature in this MLM industry. The result appears to be consistent with findings
from Frenzen and Davis (1990), who find that having social ties can increase the likelihood
of selling/purchasing products in this type of industry. However, unlike Frenzen and Davis
(1990), I offer empirical evidence that these social aspects play a role in business relationship
formation as well, which has not been examined before, although I do not observe actual
social ties. This reasonably follows from previous findings that the new joiners are often
exposed to business opportunities through purchasing products (Legara et al. 2009). My
empirical results indicate that both new joiners and distributors show a tendency to start a
business with someone who is probably familiar to them, possibly sharing a similar lifestyle,
14
Given new joiner j, the probability that matching with i (ij) is preferred to matching with i’ (i’j) is
0
P r(Xij α + ij > Xi00 j α + i0 j ). From this, I can derive the marginal effect, dP/dX, using the chain rule. For
dummy variables, marked with ? in the results in Table 1.9, I approximate the marginal effect by calculating
the difference, rather than using a derivative.
28
having a similar socio-economic background, or having potential commonalities.
29
relationship value of high status distributors as their sponsoring upline as compared to low
status distributors. Using the estimates, I find that the probability that a new joiner be-
comes a business partner with a high status IBO (all else being equal) is twice as high as
when the new joiner partners with a low status IBO (Table 1.9).
The results on the network structural hierarchy (level) suggest that, after controlling
for other possible confounding variables, such as the duration in the business, distributor’s
focus, status and other social factors, the level does not have a significant impact on network
formation. Even if the distributor’s compensation scheme does not depend on the network
location, previous literature shows mixed results on the impact of level on the distributor’s
own performance. At least for distributor network formation, the results seem to indicate
that there is no bias toward a specific network location.15
30
factor driving the relationship, I compute the aggregate value of each factor. Reordering
0
the equation 1.6 based on the social and economic factors gives us Vijt = Xsocialij α1t +
0
Xeconij α2t + ij . Using the draws of α1t and α2t , along with the associated X at each
market t, social and economic factors from the realized business relationships are computed.
0
From the equation, Xsocialij α1t is termed the social factor and Xecon0ij α2t is defined as the
economic factor. Next, I average the two factors across the markets to examine the aggregate
relative importance between the two.
The estimates suggest that the average social factor is 0.50, whereas the average economic
factor is 1.56 — social factors have approximately one-third the importance of economic
factors.17 So although social factors are important to a certain degree, economic factors
convey a significant portion of information in determining the relationship value, which
explains the business-relationship formation.
31
whereas 35% show the relationship value stems from the economic factor (positive economic,
negative social). The correlation between the two factors is -0.26. These percentages, along
with smaller percentages in both positive or both negative conditions, show that IBOs and
new joiners are creating business relationships that rarely satisfy both factors (in a relative
sense), and the business relationship is often formed emphasizing one factor or the other.
With the recovered social and economic factors, which explain the relationship value with
IBOs, I examine whether the two factors influence the new joiner’s future business activity.
I specifically focus on the new joiner’s number of direct downlines, registered customers, and
total downlines 6 months after joining the business, as these activities reflect the expansion
of the MLM network. I tracked these measures for the 582 new joiners who entered the MLM
business during the first 15 months within the sample. Table 1.11 shows the expansion of
the business after 6 months. Almost 50% of the new joiners did not engage in business
expansion activities, perhaps utilizing the MLM business only for their own consumption or
staying inactive.
The question is, what type of relationship at the relationship-formation stage leads to
higher business activity in the future? I regress the recovered social and economic factors
on the number of direct downlines, total downlines and registered customers. The results
are presented in Table 1.12. I find that the economic factor that explained part of the
relationship value in the formation stage influenced the new joiner’s engagement in MLM
business activities, in terms of engaging in recruiting direct downlines and registering cus-
tomers. Interestingly, social factors were significant only in forming the relationship; they
did not explain whether the new joiners actively engaged in MLM business activities.18 This
finding implies that emphasizing social aspects to create business relationships can hinder
the potential overall business growth.
32
could reflect the economic aspects stemming from the two key participants, new joiners
and IBOs, I am able to study the role of both social and economic factors in network
formation. Furthermore, I provide empirical evidence on whether people involved in business
are behaving irrationally by creating business relationships based solely on social factors for
the business purposes), or rationally by considering future economic aspects.
The research highlights one of the important behavioral aspects, distributor network
formation in the MLM industry, which has been understudied in the marketing domain. In
the MLM industry, expanding the distributors (IBOs) is an important issue, because IBOs
are the ones who advertise, distribute, and sell the product directly to the end user without
going through all the intermediaries. Unlike conventional firms that rely on a sales force, the
new joiners are recruited by the IBOs, not hired. So from a firm’s perspective, expanding
the network through the IBOs’ efforts is important, and both new joiners and IBOs have
good reasons to ‘select’ the right business partner.
Taken together, the results of the main analysis show strong evidence of social factors
playing a role, namely, demographic and geographic homophily. In addition, economic re-
turns on each side (new joiners and IBOs) affect the network formation. I examine under
what circumstances the relationship value (pairwise utility) becomes the highest, which re-
flects the nature of the distributor network formation. Also, by computing the marginal
effect of social factors, and economic factors on each side, I quantify the attractiveness lift in
terms of forming the relationship. This research is especially relevant in the currently chang-
ing recruiting/distributor signing environment, where there is more purposive searching and
exploratory courtship between the new joiners and the IBOs (King and Robinson 2000).
To highlight the findings so far, I find that on the new joiner’s side, IBOs having a high
status and being relatively new to the business, increase the relationship value between IBOs
and new joiners. Also, having a small set of direct downline and constantly engaging in retail
product selling increase the value of the relationship. All these IBO characteristics are driven
by the new joiners’ economic concerns. From the IBOs’ perspective, new joiners with high
income generated the lowest relationship value. Unlike previous mixed findings, my results
show that the network structure (e.g., where the distributor is located in the hierarchy) does
not add extra information to network formation after controlling for some of the possible
confounding variables, including experience, status, and the distributor’s focus.
Consistent with the descriptive analysis on the demographic homophily, I find that shar-
ing the same language, ethnicity and similar age increase the relationship value significantly,
while income similarity is marginally significant. Geographic proximity increases the likeli-
hood of being matched by generating higher relationship value.
Using the recovered social and economic factors, the follow-up analysis shows that in
33
explaining the network in MLM, although IBOs and new joiners place a high relationship
value on people like themselves (i.e., on social factors), economic factors played a larger
role in forming relationships - approximately three times larger. Finally, understanding the
type of relationship value between IBOs and new joiners that can lead to active business
participation is crucial in expanding the distributorship. In the follow-up analysis, I provide
evidence that new joiners with a high economic factor, which is a part of relationship value,
tend to show more active business activities.
Managerially, this research can help the MLM industry and IBOs in three ways. First,
at the industry level, this research can change people’s perceptions about how the business
works. The results suggest that people involved in this industry are motivated more by
rational (business) reasons than by social factors. Second, this research can help both current
IBOs and new joiners to understand how to better design and manage their business by
increasing relationship value. Specifically, it can help distributors to understand how strongly
they are positioned in terms of attracting new people, and what they can do about it if they
are not strongly positioned. Also, the new joiners can benefit by understanding what type
of IBOs may be suitable for their needs and for their future success. Both sides can now
understand how to increase the relationship value. Third, it can help the firm to recommend
the right business partners, especially for those who do not have a specific partner in mind
yet. The aim is not only to create the relationships with a high relationship value, but also
to generate new joiners’ active participation, which can expand the business.
Methodologically, in common with previous work on empirical matching models, my ap-
proach has the following limitations. First, I assume away search costs as in Sorensen (2007).
The results could be impacted if search costs exist and interact with the social and economic
factors that I consider. Second, I assume that each distributor is knowledgeable about all
agents (and their associated information) on the other side who are willing to form a match
as in Boyd et al. (2013). Therefore, during the estimation stage, the stability condition is
imposed on the all the agents on the other side, but only those who are interested in forming
the match are used in drawing V. Third, as is typical with such models, my estimation uses
observed outcomes. The data make it hard to identify ex ante the potential set of IBOs the
new joiner has approached and vice versa, as well the actual social network of all agents. I
acknowledge that the use of demographic variables and geographic distance is an imperfect
proxy for the network. Finally, utility specification is at the level of joint utility. Recent
studies note that in the empirical matching model, it is still hard to recover the true pref-
erence of each side separately (Hsieh and Lee 2012), unless there is a restrictive matching
setting in the market. Thus I am unable to distinguish which side - the IBO or the new
joiner - is receiving higher utility from the match. Additionally, the variables that are pair
34
dependent (e.g., demographic homophily or geographic proximity) could have an asymmetric
impact on each side, which I do not distinguish in this paper. I hope that future work can
relax these assumptions.
35
1.8 Tables
36
Table 1.2: Types of Match Observed in the Data
Month 1 to 1 match 1 to 2 match 1 to 3 match 1 to 4 match
1 23 0 0 0
2 19 0 0 0
3 12 0 0 0
4 17 1 0 0
5 16 0 0 0
6 21 0 0 0
7 52 1 0 0
8 43 2 0 0
9 58 2 1 0
10 43 2 0 1
11 29 4 0 0
12 41 5 0 0
13 52 3 0 0
14 47 2 0 0
15 50 4 0 0
16 43 3 1 0
17 58 2 0 0
18 61 2 0 0
19 58 8 1 0
20 41 2 0 0
21 49 3 1 0
Total 833 46 4 1
Percentage 94.23% 5.20% 0.45% 0.11%
37
Table 1.4: Contingency Table on Language (in Percentage)
Upline Language
Downline Language English Spanish Mandarin Korean Total
English 71.42* 2.01 0.24 1.25 74.93
Spanish 2.08 22.09* 0.01 0.02 24.19
Mandarin 0.04* 0 0.02 0 0.07
Korean 0.29 0 0 0.52* 0.81
Total 73.83 24.11 0.27 1.79 100
38
Table 1.5: Contingency Table on Ethnicity (in Percentage)
Upline Ethnicity
Downline Ethnicity Missing European Hispanic Asian African American Non-Orient Asian Jewish Arab Total
/Polynesian
Missing 1.79 4.23 1.92 0.63 0.4 0.61 0.11 0.14 9.82
European 4.09 31.92* 2.86 1.02 1.55 0.68 0.61 0.24 42.96
39
Hispanic 1.87 3.47 23.11* 0.37 0.25 0.68 0.1 0.08 29.94
Asian 0.51 1 0.21 5.98* 0.07 0.22 0.02 0.07 8.08
African American 0.47 1.93* 0.24 0.11 1.32 0.14 0.05 0.07 4.33
Non-Orient Asian/Polynesian 0.41 0.41 0.51 0.17 0.05 1.46* 0.01 0.06 3.08
Jewish 0.1 0.57* 0.07 0.02 0.03 0.02 0.07 0.01 0.88
Arab 0.13 0.22* 0.06 0.07 0.06 0.15 0.01 0.2 0.9
Total 9.36 43.76 28.97 8.36 3.73 3.97 0.98 0.87 100
Table 1.6: Summary Statistics on Business Support and Network Structure
High status Duration in the business (Month) # Direct Downlines # Total Downlines # Customers Level
Min. 0.00 0.00 0.00 0.00 0.00 15.00
1st.Q 0.00 1.00 1.00 1.00 0.00 28.00
Median 0.00 5.00 1.00 2.00 0.00 33.00
Mean 0.02 24.75 2.79 20.18 2.77 33.07
3rd.Q 0.00 26.00 3.00 8.00 2.00 38.00
Max. 1.00 344.00 25.00 1375.00 143.00 97.00
40
High status 1.00 0.28 0.41 0.54 0.32 -0.05
Duration 0.28 1.00 0.64 0.39 0.32 -0.22
# Direct Downlines 0.41 0.64 1.00 0.44 0.46 -0.17
# Total Downlines 0.54 0.39 0.44 1.00 0.14 -0.12
# Customers 0.32 0.32 0.46 0.14 1.00 -0.09
Level -0.05 -0.22 -0.17 -0.12 -0.09 1.00
Table 1.8: A List of Variables in the Matching Model
1. Social Factors
Demographic homophily
a. Age difference
b. Language
(Same = 1, else = 0)
c. Ethnicity
(Same = 1, else = 0)
d. Income group
(Same = 1, else = 0)
Geographic homophily
a. Zip code distance
2. Economic Factors
Distributor characteristics (New joiners’ Economic Return)
a. Duration in the business
(# months)
b. Distributor’s Focus
Network (# Direct downline, Existence of indirect network)
Product Selling (# Customers)
41
Table 1.9: Estimates of the Matching Model
Variable Mean Marginal probability
Language? 1.16 ** 0.29
Absolute age difference (1/10) -0.33 ** -0.09
Social Factors Ethnicity? 1.13 ** 0.28
Distance (miles/10) -1.65 ** -0.46
Income group similarity? 0.31 * 0.09
Duration(months/10) -0.68 ** -0.19
Direct downline (1/10) -2.96 ** -0.84
Economic Factors Indirect network? -0.14 -0.04
(Distributor characteristics) # Customers (1/10) 0.99 * 0.26
Status? 3.91 ** 0.50
Relative level (*10) -0.09 -0.03
Economic Factors Medium income? -0.21 -0.06
(New joiner characteristics) High income? -0.34 * -0.09
42
Table 1.10: Relationship Types Based on Estimated Social and Economic Factors (%)
Economic Factor
Positive Negative Total
Positive 19.66 31.99 51.65
Social Factor
Negative 34.64 13.71 48.35
Total 54.30 45.7 100
43
1.9 Figures
44
Figure 1.2: Approximate 35 miles Radius from Chicago
45
Figure 1.4: Number of Direct Downlines, Total Downlines and Customers
46
Figure 1.5: New Joiner’s Estimated Social and Economic Factors in Relationship Formation
Figure 1.6: New Joiner’s Relationship Types Based on Estimated Social and Economic
Factors
47
Chapter 2 The Impact of Existence of
Others on Inactive Behavior in the
Multi-level Marketing Industry
2.1 Abstract
Turnover behavior has long been a topic of interest in organizational studies. Understand-
ing the behavior in the context of multi-level marketing (MLM) poses a unique challenge,
however, due to the unique characteristics of the industry, which differs significantly from
traditional hierarchical organizations. MLM firms work closely with a network of distribu-
tors (Independent business owners, or IBOs), who either sell products to the end customers
or recruit others into the network. Although IBOs in the MLM industry are ‘independent’
business owners, their business activity can be subject to a significant influence from oth-
ers. This stems, in part, from the cooperative, social nature of business relationships, where
there is a high degree of exposure to the IBOs. To understand this influence, this paper
examines external social factors, namely, how the existence of different others affects an
IBO’s inactivity. In particular, I examine three different dimensions of others: Individuals,
Network family, and Proximity. Using both county-level analysis and individual-level sur-
vival analysis, I find evidence that the existence of others in proximity increased an IBO’s
inactive behavior. In contrast, when an IBO is associated with a larger sized network family,
IBOs remained active. Another factor influencing the inactivity is the existence of ‘success-
ful others’ (with high status) in proximity and network family dimensions, which have a
protective effect from inactive behavior. Machine learning analysis has also been applied to
predict future inactivity behavior. The model is found to be useful in distinguishing those
IBOs who will not experience the inactivity versus those who will, and also in detecting an
inactive period occurring within a quarter. Consistent with results from statistical analysis,
the relative importance of the variables (See Machine Learning section) obtained from ma-
chine learning reveals that the percentage of high status IBOs within a family still plays an
important role, even in predicting a future inactivity period.
48
2.2 Introduction
“A few companies are beginning to realize that sales turnover is a sleeping giant, incon-
spicuously swallowing a significant portion of their productivity and profits” (Futrell and
Parasuraman 1984). Despite the fact that the quote is from the early eighties, it is still
true today. Salesperson turnover has been a topic of interest in the organizational literature
since the late seventies for two main reasons: cost and the phenomenon itself (DeConinck
and Johnson 2009, Holtom et al. 2008). Cost is of interest because losing each employee
entails losses for the firm, which are estimated to vary from a few thousand dollars to more
than twice the employee’s salary, depending on the industry, the content of the job, the
availability of replacements and other factors (Hinkin and Tracey 2000). This includes the
cost involved with hiring and training a new employee (Griffeth and Hom 2001).
Apart from the issue of cost, interest in turnover in sales organizations has been ongoing,
making it a notable because the estimated turnover rate among salespeople is double that
for other jobs (Richardson 1999). Although some aspects of turnover may be positive, such
as the elimination of ineffective performers (Jolson et al. 1987), in general, researchers and
industry managers are eager to understand how to reduce it because of its impact on the
size and continuity of revenue generation (Jolson et al. 1987, Wotruba et al. 2005, Darmon
2008, Holtom et al. 2008). Thus, substantial research attention has been placed on identi-
fying correlates and predictors of employee turnover in traditional hierarchical organization
settings (Jolson et al. 1987).
Within the sales industry, one specific type of sales organization suffers from a turnover
rate that is even higher than that of the sales force of the U.S. manufacturers, the traditional
organizations (Coleman 1989): Multi-level marketing (MLM or Network marketing) orga-
nizations (Raymond and Tanner Jr. 1994, Wotruba et al. 2005). MLM is a type of direct
selling (the other type is Single-level marketing) (Biggart 1989, Brodie et al. 2002), which
accounts for more than 95 percent in terms of the number of firms, number of sales people,
and retail sales within the direct selling industry (DSA 2015b).1 Some well-known companies
that employ the MLM business model are Amway, Mary Kay and Herballife.
MLM organizations differ from traditional hierarchical organizations along several di-
mensions. In MLM, distributors (Independent Business Owners or IBOs) are not viewed
as employees, but rather as an independent contractors or entrepreneurs, with physical and
psychological independence from the firm (Jolson et al. 1987). As such, they are not subject
to the same kind of monitoring and control found in traditional sales departments in hier-
archical organizations, although they are required to follow the legal and ethical guidelines
1
Please refer to institutional background in essay one for detailed industry facts.
49
informally set forth by the company organization (Pratt 2000). IBOs not only sell products
but also recruit and supervise other IBOs who become part of the IBO network. This indus-
try setting is well reflected in the compensation plan, where IBOs do not earn a fixed income
as an employee in a hierarchical firm; rather, their earnings are based on commissions and a
personally set retail margin. Commissions are based on the product sales carried out by the
IBO and the connected network of IBOs, which is expanded though recruiting. This struc-
ture provides IBOs a flexible schedule with a better work-life balance, while also providing
motivation to expand the network (Biggart 1989). In order for the MLM firm to motivate
IBOs from the flexible working environment, monetary and non-monetary recognition are
utilized, which include the systematic commission plan and status.
Given the different industry settings, the turnover costs of the MLM is unique, even
differing from other sales organizations. In particular, the burden of the turnover costs,
discussed in more detail in the literature review, falls on the shoulder of both the MLM
firms and the IBOs. The risk to the firms is mainly due to the IBOs being the main retail
outlet to the end customers, while IBOs being ‘independent.’ In the 2016 10-K, Herballife
disclosed risks related to their business and their very first point was to acknowledge that
their distributors may voluntarily terminate their distributor agreements at any time: “Our
failure to establish and maintain Member and sales leader relationships for any reason could
negatively impact sales of our products and harm our financial condition and operating
results” (Herballife 2016). The risk from the IBOs’ side is due to their role in recruiting
and training efforts that themselves, rather than the firm, undertake and the unique mentor
relationship between recruiter-recruited IBOs. Unlike in a traditional firm, where employees
must work before they decide to quit (apart from a leave of absence), in an MLM, one
can remain inactive without permission from the manager or the company, since there is
no formal quitting procedure from the distributor and no official employment termination
notice from the company (Wotruba 1990b, Wotruba et al. 2005).
This concept is consistent with being one’s own boss as a self-employed independent
contractor, with the individual having a choice in terms of how much to engage with the
business. Thus, IBOs who are registered under the business, but at a certain point became
less interested without engaging in any activities, can be an indicator of turnover. Research
has shown that many MLM firms specify some length of inactivity as a potential indicator
of turnover (Wotruba 1990b, Wotruba et al. 2005). Despite the uniqueness of the MLM
and IBO structure, little academic literature has examined turnover behavior in the MLM
industry. The few existing studies on direct selling turnover rely on subjective survey data
to explore this phenomenon, which provides a room for quantitative data-driven research.
Utilizing a unique dataset from a leading MLM firm, I examine turnover, more precisely,
50
inactive behavior in the MLM industry. Previous literature on MLM turnover suggests a
strong relationship with job performance (Jolson et al. 1987, Wotruba 1990b), which can
be attributed to the business characteristics of an individual. Although the business char-
acteristics of an individual provide high explanatory power, other factors may play a role
in turnover. On the one hand, an MLM is social, encouraging cooperative relationships
with others (Sparks and Schenk 2006), which has even been described as cult-like behavior
(Bhattacharya and Mehta 2000). On the other hand, selling the same product from the firm,
IBOs may inevitably be influenced by other IBOs surrounding them. Jolson et al. (1987)
have attempted to examine the degree of perceived competition using a one item survey
but did not find evidence for relationships between competition and turnover. Aside from
this study, to the best of my knowledge, no previous literature documents how individual
turnover behavior can be affected by the presence of others in the MLM industry, either in
a cooperative or a competitive manner.
I examine the existence of others along three different dimensions based on MLM indus-
try characteristics: Individuals, Network family and Proximity. For the first dimension, I
examine two types of key individuals unique to the MLM industry—upline and high status
IBOs. An upline is an individual with whom a prospective IBO creates a business relation-
ship. When a prospective IBO joins the MLM, s/he is connected with only one upline, which
creates a mentor-type relationship. A high status IBO is a successful IBO approved by the
MLM firm who takes on the role of leader of a sub-group of IBOs. Unlike in a traditional
organization, having the high status does not imply having more authoritative power: there
is no subordinate relationship between a high status IBO and IBOs without that status
(Biggart 1989). The second dimension, the network family, is a motivational network or-
ganization usually set up by accomplished IBOs in the MLM business to help other IBOs
become successful. Lastly, the third dimension, geographical proximity, indicates those who
are located nearby. Across these three dimensions, I examine the influence of high status
IBOs on an IBO’s inactive behavior.
To answer the question, I rely on a county-level regression analysis and an individual-level
survival analysis. The two analyses show similar findings for the effects of others explain-
ing the focal IBO’s inactive behavior. Consistent with industry concerns (e.g., Msweli and
Sargeant (2001)), I find competition among close IBOs to be inevitable within a particular
area, as reported in the franchise literature (Pancras et al. 2012). However, in both the
descriptive county-level regression and individual-level survival analysis, I find the interest-
ing role of status related to inactive behavior, which is a feature unique to the industry
indicating of success. Exposure to high status IBOs decreased inactive behavior in general
(based on county-level analysis). Specifically, the percentage of high status IBOs along the
51
network family dimension and the proximity dimension decreased inactive behavior (based
on individual-level analysis). In other words, the more high status IBOs within the same
network family and within proximity, the more likely that IBOs would remain active in the
business—a protective effect. However, along the individuals dimension, the distance from
the structurally nearest high status IBO did not have a protective effect, whereas the dis-
tance from an upline sponsoring IBO was found to be positively related to active behavior.
In addition, IBOs associated with a larger network family tended to stay active in the busi-
ness, potentially by fulfilling the need for social connection (DSA 2015c), or by fulfilling a
desire to have access to potential network resources (Jolson et al. 1987). Moreover, I find
that when the market has a high penetration rate—relatively more people are involved in
this business—IBOs are less likely to face the inactivity hazard, potentially due to awareness
about the company or its products, which makes it easier for IBOs to do business activities
(Jolson et al. 1987).
Turning my attention from understanding the inactivity phenomenon to predicting a
period of IBOs’ future inactivity, I conducted a set of machine learning analyses, focusing
on neural networks. Based on one of the analyses, I detect and predict a duration for
encountering an inactive period in the future by quarter. The model is designed to allow us to
pinpoint the upcoming quarter when the IBO is going to be inactive. The model is sufficient
for distinguishing the IBOs in a specific month that do not follow an inactivity period, and
the IBOs that are currently under the inactivity period, attaining an overall accuracy of
81.44%. Based on one approach I take, the model is also sufficient for predicting the inactive
period occurring within a quarter. However, there exists some room for improvement for a
long term prediction—predicting IBOs in a specific month, experiencing the inactive period
after a quarter. Based on the results, the relative importance across variables shows that one
of the top five most important variables in the neural network is the percentage of high status
members within a network family. Variables related to ‘others’ also were ranked relatively
high, reinforcing my main claim that exposure to other IBOs has a role in explaining the
focal IBO’s inactive behavior. This results may be due to the nature of the industry, where
social interaction and status is highlighted, and industry features that help motivate IBOs
to be successful in the business.
This study contributes to the turnover literature, in particular, the MLM industry. In-
terestingly, some of the industry specific features that do not appear or do not exactly match
those of traditional hierarchical firms, such as status or network family, were found to have a
protective effect in preventing IBOs’ inactivity. MLM firms can potentially utilize the find-
ings in this research to reduce inactivity behavior. Methodologically, this study is the first
empirical study based on hard data to examine turnover in MLM; previous studies relied on
52
in-person surveys. This unique aspect allowed me to explore the environmental factors that
have not been thoroughly studied. Lastly, I examine the applicability of machine learning
techniques to the MLM turnover literature. I was able to draw similar conclusions using
both traditional methods and machine learning, showing that how the focal IBO is situated
among other IBOs, in particular, among successful others, plays a role in explaining and
predicting inactivity.
The rest of the essay is organized as follows. Section 2.3 provides a brief review of MLM
turnover cost, prior literature on turnover, and the three different dimensions, namely, indi-
viduals, network family, and proximity. Section 2.4 presents the data. Section 2.5 introduces
the county-level and individual-level analyses. I then move on to Section 2.6, which focuses
on inactivity prediction using machine learning. Section 2.7 concludes the second essay.
53
registered online, they are personally taken care of by their own personal IBO, which reveals
that firms are unable to track each end customer the IBO is serving. Raymond and Tanner Jr.
(1994) highlight the need to follow up with the customers of IBOs who no longer work for
the firm. Even if the customers appreciate the products the firm is offering, they may
have been poorly serviced as a result of thier IBO’s inactivity or leave. Then the firm
not only loses the current sales by losing the contact, but also risks future sales, if no
channel is provided for continuing the relationship. Another higher order issue that naturally
follows from dissatisfied customers is the unfavorable perceptions arising from the frustration
experienced when products are no longer easily available (Raymond and Tanner Jr. 1994).
From the IBO’s perspective, when a recruited downline remains inactive or leaves, con-
siderable investment must be made once more in the recruitment and training of new IBOs,
as opposed to focusing on selling the product. Unlike the traditional hierarchical firms, this
recruiting and training cost burden falls on the IBOs rather than on the firms, resulting in
an inefficient distribution outlet. In contrast, when a recruiter IBO stays inactive or leaves,
the recruited IBOs will be exposed to a bad example and fail to receive productive guidance,
which in turn leads to a potentially less successful distribution outlet.
54
sizes the role of others in turnover behavior. For instance, in a meta-analysis of turnover,
work environment factors appeared, such as leadership (leader member exchange, supervisory
satisfaction) and co-worker related variables (work group cohesion, and co-worker satisfac-
tion) (Griffeth et al. 2000). These factors indicate that one’s decision to stay in the business
is affected by other individuals with whom one highly interacts—supervisor or co-workers.
In addition, Holtom et al. (2008) pointed out from their meta-analysis that one of the trends
explaining the turnover behavior in organization studies is the emphasis on interpersonal
relationships, such as leader-member exchange or interpersonal citizenship behaviors. Based
on this general background, in the following section, I introduce additional literature on
traditional organizations that can be mapped onto the suggested three dimensions of others
in the MLM industry.
While focus has mostly been on turnover in traditional firms, including some evidence
for the role of others, some prior research has touched upon direct (or MLM) salespeople
turnover. Looking at various types of sales organizations2 , including salespeople serving in-
home consumers, Jolson et al. (1987) carried out an exploratory investigation survey of eighty
possible correlates of sales force tenure in general, looking at four different categories: individ-
ual non-work related, external economic factors, individual work related, and organizational
variables. Of the eighty, only four were shown to be significant: age (+), earnings (+),
popularity/acceptance of product/company (+) and sales force size (+). In addition to this
study, which comprehensively examines related factors, other factors were also highlighted
in other studies. In terms of job performance, the relationship between job performance and
inactivity-proneness is found to be negative only for the low performers (the poorer, not the
better, performers are more likely to become inactive) (Wotruba 1990b). The findings on the
relationship between job performance and turnover in direct selling is consistent with general
findings in the professional selling literature in traditional hierarchical organizations (Futrell
and Parasuraman 1984, Ladik et al. 2000, Jackofsky et al. 1986). Unmet expectations has
been examined both cross sectionally and longitudinally, along with job satisfaction, which is
a well-established relationship in general sales force settings (Wotruba and Tyagi 1991, Ko-
roth 2014). Involvement in the business has also been identified as an influencer of turnover
in MLMs. In particular, the work compared part-timers with full-timers. The finding is
that part-timers had greater job satisfaction, performed better measured by earnings per
hour worked, and were less likely to quit (Wotruba 1990a). How one perceives how others
view their job or selling behavior was also related with inactivity-proneness in direct selling
(Wotruba 1990b).
2
Examined four different sales force types depending on the main customer base: (1) resellers, including
wholesalers and retailers, household consumers, (2) household consumers contacted at their residences—
mainly direct sellers, (3) manufacturers, (4) other organizational end users (business, goverment, etc.).
55
These constructs are based mostly on an IBO’s individual characteristics and were exam-
ined using a qualitative survey approach. This research differs from the previous work in that
I examine the IBO’s connection to three different dimensions of others in explaining their
inactive behavior: Important individuals, Network family and Proximity. Although Jolson
et al. (1987) had a one-item survey question related to ‘perceived degree of competition’, it
proved to be insignificant and was dropped from further discussion.
In a slightly different context, turnover behavior can be linked to the customer churn
management literature because the goal of both turnover and customer churn management
is to make people keep connected with the company; whether that being employees, IBOs or
customers. In particular, in the MLM turnover, IBOs can be considered as consumers as well
as distributors, so previous findings on continuing relationship with existing consumers with
exposure to other consumers can contribute to understanding the MLM turnover behavior.
Haenlein (2013) examined social interactions in customer churn behavior in mobile phone
provider and found that when a focal actor has an out-going call relationship with other
individuals who have recently defected, the focal actor is significantly more likely to defect
from the provider. While much attention has been placed on social interactions in the
customer acquisition process (e.g., Iyengar et al. (2011)), based on the network data, research
shows the role of social interaction in churning behavior. Taking a similar stance, in my
second essay, I examine the role of the existence of three different dimensions of others on
inactive behavior in the MLM industry, based on empirical data.
In Section 2.3.3, I will explain the three different dimensions of others that can possi-
bly affect IBO’s turnover behavior in MLM: Important individuals, Network family, and
Proximity.
At the individual dimension, I examine two important individuals in the MLM industry who
potentially have an influence on the focal IBO’s inactivity behavior: an upline and a high
status IBO.
56
called uplines; joiners are referred to as direct downlines. Important here is that one new
joiner can be connected with only one upline.
The relationship between the upline and the recruited IBO is quite special. It is similar
to that of mentor-mentee, where each is an independent entrepreneur. Once the new joiner
becomes a part of the network under a certain IBO – a sponsoring upline, it is the incumbent
on the upline to help the new joiner initiate the business, maintain motivation, and provide
a long- lasting support (Crittenden and Crittenden 2004). Because of this intricate relation-
ship, new joiners are often advised to find the ‘right’ sponsoring upline to be successful in
the business (Allen 2016). I thus hypothesize that IBOs located near their own connected
upline IBOs show a tendency to stay active in the business.
IBOs usually receive a supplementary bonus for sales generated by recruiting on top of retail
margin, which motivates them to build networks and sell products – a key characteristic of
MLMs that distinguishes them from single-level direct selling (Wotruba et al. 2005). But
beyond the supplementary bonus, firms utilize additional motivational stimuli in the form
of high status as a designation of success. As the MLM firms do not have direct control over
the IBOs, firms award multiple tiers of status based on IBOs’ business activities that entail
not only financial benefits but also high social recognition (King and Robinson 2000). More
specifically, the status provides attention and respect from other IBOs and the MLM firm.
This status generates social recognition, which explains one of the motivations for being part
of the MLM business (Wotruba and Tyagi 1992, DSA 2015c). At the same time, these high
status IBOs are bound by additional responsibilities along with selling the products and
expanding the network. They play a pivotal role in teaching and motivating other IBOs at a
group level through leadership responsibilities and setting a good example of success in the
MLM by being an ‘admirable entity’ (Amway 2015, Biggart 1989). Based on the associated
characteristics of high status, I hypothesize that IBOs who are geographically close to a high
status IBO will tend to stay active in the business.
Network family is commonly known as the Business Support System or motivational or-
ganization, and is generally an independent organization initiated and funded by successful
IBOs to offer support, training, and professional development programs to other IBOs within
the same family. Structurally, imagine multiple mutually exclusive network trees. Due to
its independent nature, the MLM firm provides an accredited program. An IBO’s family
57
and an upline’s family may not necessarily be the same; however, it is rare for someone of
lower status to have a family different from that of his/her upline. In organizational studies,
organization structure (e.g., size, flat/tall hierarchy) is believed to affect the behavior of
members, such as turnover, absenteeism, even performance. However, the findings about
organization size are mixed (Dalton et al. 1980). Among different network family character-
istics that may affect one’s business activity, I examine two characteristics, namely, size of
the family and percentage of high status IBOs within the family.
Well-documented evidence exists for the friendly, personal, and cooperative nature of re-
lationships among IBOs (Biggart 1989, Sparks and Schenk 2006, Grayson 1996), and the
success of the MLM business model is largely dependent on IBOs’ forming cooperative re-
lationships (Sparks and Schenk 2006). IBOs may feel connected by becoming part of a big
group, providing a relatively higher family-based social satisfaction, which is an additional
utility IBOs receive by becoming part of an MLM business (Bhattacharya and Mehta 2000).
Previous literature on traditional organizational studies have supporting claims on the
role of socialization on turnover. Research on socialization tactics (i.e., seeking performance
feedback and information seeking, building relationships and networks) shows that the tactics
help the organization to embed employees into the organization. Utilization of such tactics
increases new joiners’ perceived level of accommodation and helps them to adjust to the
new work environment. Also, it increases the level of organizational commitment, which
is strongly related to job satisfaction, job performance and turnover. In particular, tactics
related to collectivism were positively related to job embeddedness, which mediates the
relationship between the socialization tactics and turnover (Allen 2006, Menguc et al. 2007).
Job embeddedness is found to be negatively correlated with turnover (Mitchell et al.
2001). The concept indicates the degree of attachment and inertia with the organization, and
is composed of three concepts: link (formal or informal connection within an organization),
sacrifice (perceived loss of material or psychological benefits by quitting the job) and fit
(compatibility with corporate and external environment) (Mitchell et al. 2001). Similar to the
job embeddedness construct, McPherson et al. (1992) specifically examine the social network
within the organization and find that the more social ties one has within the organization,
the less likely one is to leave the organization.
Hom et al. (2009) corroborate the prevailing views of the importance of interrelationship
in turnover that social exchange (e.g., shared investment, mutual trust, enduring relationship
among employee-organization relationships) and job embeddedness in employee-organization
58
relationships increase workforce commitment and loyalty, which reduces the intention to quit.
Although the institutional setting is different, based on the previous literature from
organizational studies, the connection can be made: IBOs with a larger family will have
more opportunities to be exposed to socialization tactics, to form social ties, and in general,
be connected to other IBOs within the same family. Thus, IBOs with a larger family tend
to stay active in the business.
In a previous survey on tenure in organizations, Jolson et al. (1987) find evidence from
sales people serving in-home customers showing that sales force size is positively related
to an IBO’s tenure. Their argument is that the larger the sales force, the more improved
the quality of the sales managers and the training programs, and the more effective the
sales support activities (Jolson et al. 1987). As mentioned earlier, unlike in traditional firms
where workers meet one another on a daily basis, as IBOs are not hired by the firm, they
have relatively limited contact with the firm or with other IBOs. Instead, IBOs (especially
high status IBOs) often organize informal collaborative work groups, usually through the
same family, to exchange information and selling tips, hold sales workshops to create shared
opportunities, and assist one another (Sparks and Schenk 2001). Of course, the finding was
based on a single-item survey, and uncertainty exists in terms of how the survey respondents
interpreted sales force size. To come to a more robust understanding of turnover, I examine
the number of others both within the family and in proximity, which I will explain in Section
2.3.3.3.
In general, the larger the family size, the greater the likelihood of having successful people in
the family (Jolson et al. 1987) showing a good example. Specifically, when those successful
people exist within the same family, their status will become more prominent to the other
IBOs. Identifying successful IBOs is done by the MLM firm.
As mentioned earlier, leadership-related variables such as supervisory satisfaction or
leader-member exchange were a component explaining the turnover behavior in traditional
organizations (Griffeth and Hom 2001). This institutional setting, along with the findings
in organizational studies, lead me to expect that the higher the percentage of high status
within one family, the less likely one is to become inactive because of the potential exposure
to direct and indirect supervision and possible interaction with successful members.
59
2.3.3.3 Proximity
Even if an IBO is associated with a network family and receives helpful support and motiva-
tion, his/her business activity will also greatly depend on the proximity of each individual.
Family connection is a global network aspect, while proximity is a local business environ-
ment. In an MLM, there are no constraints as to which market they will serve, unlike in
a traditional sales force with assigned territories. Nevertheless, IBOs located close to oth-
ers will have a higher chance of serving a similar customer base, with the possibility of an
overlapping social network. In addition, since the compensation plan is highly dependent
on how much an IBO and his/her recruitee sell the product, it is likely that IBOs located
close to others will face competition. Previous research on sales force turnover attempted to
examine the negative effect of perceived degree of competition on turnover, but the results
turned out to be insignificant (Jolson et al. 1987). However, Msweli and Sargeant (2001) and
Sparks and Schenk (2006) acknowledge the competition prevailing between IBOs, and the
importance of coordinating between IBOs in the MLM organization (Msweli and Sargeant
2001), despite the need to be cooperative in the MLM industry (Sparks and Schenk 2006),
a fundamental trait of IBO relationships (Biggart 1989, Grayson 1996). This concern of
competition indicates that empirically supported competition may exist among IBOs within
a proximate distance. I would thus expect to see local competition within a proximate area.
On the other hand, IBOs may stay active longer when high status IBOs from the same
network family are located in proximity, both because of the role of high status explained
earlier and the fact that the IBOs have a better chance of learning about their success stories
compared to high status IBOs not sharing the same family.
2.4 Data
Data is available from one of the leading MLM firms in the U.S. for 318,363 IBOs who joined
the MLM business from January 2006 until October 2007. Unlike previous churn literature
where a clear cut definition of churn exists, as shown in Table 2.1, the data itself does not
indicate whether an IBO dropped out of the business completely or not. Thus, I examine
the concomitants of inactive behavior, that is, those who did not engage in the business for
six consecutive months (i.e., two quarters).
More specifically, if six consecutive inactive data points are observed at the end of a
particular period, from month (t+1) to (t+6), the IBO is categorized as inactive from month
t. Some IBOs may not have engaged in the business for six consecutive months in the middle
of the sequence but returned later on. These returned IBOs were then considered to be active
60
Table 2.1: A Sample of Churn Related Literature
Industry Churner
when his/her subscription is not renewed within
Coussement and
Newspaper Subscrip- four weeks after the expiry date (Company doesn’t
Van den Poel
tion Services allow ending the subscription prior to the maturity
(2009)
date)
when a mature subscriber (who has been with the
Lemmens and U.S. wireless telecom-
company for at least six months) churns during the
Croux (2006) munications company
period of 31- 60 days after the sampling date
observe the month in which the customer was
Braun and
Telecom provider acquired and defined as churner if the customer
Schweidel (2011)
churned while under observation.
Haenlein (2013) Mobile phone industry when churn happened during the data period
when churn happened during the data period
Jamal and Buck- Direct-to-home satel- (drawn from among those customers who had sub-
lin (2006) lite TV provider scribed to the firm’s service during the past 12
months)
Neslin et al. Wireless telecom in-
Churn was measured in the fifth month
(2006) dustry
from the beginning of the data.3 Engagement in the business is defined according to the
following two criteria : (i) showing no point accumulation either through personal sales or
group level sales, or (ii) no evidence of earning bonus income from the firm. Based on a six-
month screening, 89,176 IBOs became inactive in the business, which is about 26 percent of
the joiners. Figure 2.1 shows the tenure of the IBOs in the business who eventually became
inactive.
Both a county-level and an individual-level analyses are presented in Section 2.5. In this
data section, the three dimensions for the individual-level analysis is discussed, except for
the variables that appear in both analyses. Data relevant to the county-level analysis is
presented under the county-level analysis section (Section 2.5.1).
2.4.1 Individuals
Two important individuals in an MLM have a potential influence on the focal IBO’s inactive
behavior: an upline and a high status IBO. For each IBO, an identification code indicates
the connected upline IBO and a relevant high status IBO (a sub-group leader). Based on
the very last data point, October 2007, among 606,551 IBOs, 3,901 — approximately 0.6%,
3
Active was defined as (personal sales point > 0 & non-missing value) OR (group sales point > 0 &
non-missing value) OR (earned bonus > 0 & non-missing value). One needs to be in inactive state for 6
consecutive months toward the end of the data.
61
a very small number were awarded high status .
To capture the influence of these two individuals, geographic distance between them is
computed based on zip-code information. The descriptive statistics for the upline distance
and the high status distance in miles is provided in Table 2.4. As can be seen, in general, the
high status IBO is located farther away from the focal IBO compared to the upline (mean
(High status distance) = 346.12 versus mean (Upline distance) = 149.891), perhaps because
MLM business relationships are often created with someone located nearby. Based on a
survey from Raymond and Tanner Jr. (1994), most consumers made product purchases at
home, followed by the workplace as the second most frequent location for business interac-
tions, which can lead to MLM business relationships.
Another to note is the correlation between the distance from a high status IBO and the
percentage of high status IBOs within the family or the percentage of high status IBOs
within the proximity. Tables 2.2 and 2.3 show the correlations. The two correlations are
directionally negative (2.2corr(distance from a high status IBO, percentage of high status
IBO within the family) = -0.005), Table 2.3 corr(distance from a high status IBO, percentage
of high status IBOs within proximity) = -0.131). However, the distance from a high status
IBO is not necessarily highly correlated with the percentage of high status IBO within the
family nor the percentage of high status IBOs within proximity, meaning that the distance
from the high status IBO is capturing a different aspect from others in different dimensions.
62
Tables 2.2 and 2.3, the correlation between the size of family and the number of high status
IBOs is 0.907; I thus examine the percentage of high status IBOs to capture the relative
degree of exposure to them. Figure 2.3 shows the percentage of high status IBOs in each
family at the very last data point. The small percentage implies that attaining high status
is an indicative of significant success, which involves a rather difficult process. Even within
the small percentage, however, there is variation across families. As an example, one of the
families has no high status members. Another notable characteristic is that the family size
and percentage of high status IBOs do not necessarily have a linear correspondence. For
example, family 37 has about half the number of IBOs compared to family 17, the largest
family; yet, the percentage of high status IBOs in family 37 is slightly greater.
Not only do the size and the composition of status differ in families, but also how the IBOs
within a family are distributed varies as well. Figure 2.4 shows the geographic distribution of
two exemplar families in October 2007 within the U.S. As can be seen, depending on where an
IBO is located, the exposure to own versus other family members varies considerably. Also
these families are dispersed across the U.S. and that the density or the highly populated
area may be different depending on the family,. Table 2.4 shows the summary statistics of
family-related variables using the time series data, using from January 2006 to October 2007.
2.4.3 Proximity
In order to determine the existence of other IBOs within a small boundary, proximity is
defined as a 30-mile radius from an IBO’s zipcode. Figure 2.5 shows the approximate 30-mile
radius from a zip-code. Proximity (own family ratio and the number of total IBOs within 30
miles) was computed using the individual-level family label and zip-code information. The
30-mile radius was drawn at the zip-code level, allowing us to possibly examine micro-level
competition among IBOs.
Table 2.4 shows the descriptive statistics of IBO’s 30-mile proximity at the individual
level. IBO’s zipcodes are used to calculate the number of IBOs within the same family, the
number of total IBOs regardless of the family, and the percentage of own family. Tables 2.2
and 2.3 show the correlation among the variables representing the two dimensions of others:
Network family and proximity. Since the high correlation between the number of an IBO’s
own family and the number of total IBOs within a 30-mile radius is highly correlated (corr=
0.77), the percentage of own family is used instead of the raw number for the individual
level analysis. Included is the percentage and the number of high status IBOs from the same
family for future usage.
The number of total IBOs varies widely. For instance, an area with the zip code 86502
in Arizona has only 5 total IBOs within a 30-mile radius (Figure 2.6 Left. Red boundary.),
63
whereas an area with the zip code 90631 in California has 44,791 IBOs (Figure 2.6 Right.
Red boundary.). Similarly, wide variation exists in the percentage of own family and the
number of total IBOs. For some IBOs, for example, less than 1 % of surrounded IBOs are
from the same family, while for other IBOs, 90 % of surrounded IBOs are from the same
family.
The row Market in Table 2.4 indicates the market penetration rate by the IBOs. It is
constructed from the Total number of IBOs/the number of labor force within a county. This
figure captures how popular and widely accepted the MLM firm and its product are. I have
supplemented the yearly county level labor force data from the Bureau of Labor Statistics.
Thus, IBOs within the same county will have the same figure. As the county level is the
granularity of data I have for the labor force, the total number of IBOs were also computed
at the county level.
64
December 2005 is 2,877 among from 3,220 counties in the U.S.. The dependent variable is
the percentage of IBOs active during a given month for each county. The effect of family
or proximity cannot be determined here as the data is available only at the individual-level
(see discussion in the follow-up analysis). Instead, an initial relationship is shown between
the existence of other IBOs and the overall activity within a county.
For each county, the number of IBOs (regardless of family) and the percentage of high
status IBOs are computed to examine the role of others in inactivity behavior. The reason
for using the percentage of high status IBOs instead of the raw number is due to a high
correlation between the number of high status within a county and the number of IBOs
(Table 2.6).
To capture market characteristics, the unemployment rate of the county and the penetra-
tion rate within the county are included. The penetration rate (Total number of IBOs/number
of Labor force) reflects the popularity/awareness of the firm or products (demand), as ex-
plained in the last part of the data section.
In addition, as mentioned in the literature review section, there is wide consensus on
the relationship between performance and turnover, both in the sales force tenure literature
and the MLM tenure literature. To control for this, the average estimated earnings from
products selling in a given a county were calculated. While each IBO can set his or her own
retail margin, absent the individual level mark-up information, it is assumed to be fixed at
a 20 percent, which corresponds to the minimum markup the survey respondents responded
in Coughlan and Grayson (1998).4 The reason for using the earnings from personal product
selling is because of the high correlation between the percentage of high status IBOs and
bonus earnings, and the percentage of high status IBOs and total earnings, as shown in
Table 2.6 (cor(% high status, bonus) = 0.731, cor(% high status, earning) = 0.583, cor(%
high status, personal selling) = 0.159).5
In addition to the influence of other IBOs, there could be influence from others who are
not necessarily associated with the same MLM firm; for example, competitors in a similar
business. To partially capture the degree of competitors working for other MLM firms, I
used competitors’ headquarters information, under the assumption that there would be many
IBOs working with the MLM company if the headquarters were located nearby. Using the
2016 Direct Selling News North America 50 List, which is created based on the North Amer-
ica 50 ranking for the 2016 DSN Global 100 (DSN 2016),6 I identified possible competing
MLM firms for 2006 - 2007. Among the 50 firms in the 2016 list, 39 competing firms existed
4
Based on Coughlan and Grayson (1998), the firm suggested markup is reported to be 40 - 50 percent,
and the average respondents’ markup was within the range.
5
The same analysis was also run using bonus instead, and the results were consistent.
6
ranked based on the 2015 revenues
65
during those years, which matched my data span. Next, I found the location of the head-
quarters and matched it with the county. From 39 competing firms’ locations, 27 counties
were identified as having other competing firms. To reflect the competition, two metrics were
used. The first is the shortest distance from one of these competing firms’ headquarters. The
distance is calculated based on the center of the county using fips information. Figure 2.7
shows the distribution of the shortest distance from the competitors’ headquarters in miles
(min = 0, Median = 170.300, Mean = 295.400, max = 725.300). The second is the number
of competing firms’ headquarters within a certain boundary, a similar approach to that of
Seim (2006). The results with both metrics are presented in Tables 2.7 and 2.8. Lastly, year,
month and county-level fixed effects were included. The descriptive statistics on variables in
the county-level analysis are reported in Table 2.5.
Results The results based on the two competitor related metrics are presented in Tables
2.7 and 2.8. The analysis shows that the counties with a smaller number of IBOs in the
previous month resulted in a higher percentage of IBOs staying active in the business. This
implies that even though IBOs are not assigned/constrained to a fixed geographic region in
expanding the network or serving end customers, IBOs are inevitably competing with one
another within a similar geographic region by selling the same product purchased from the
firm, as in other franchise business. Also, the percentage of active IBOs were found to be
higher when there are relatively more high status IBOs within the county, which may be the
case for two potential reasons. First, having a stratified status system based on an IBO’s
achievement can encourage other surrounding IBOs to stay in the business — showing a
good example of success. To put it differently, one can naturally think, “s/he achieved it,
why can’t I?.” Second, high status does not only mean receiving more financial benefit, but
also having a greater responsibility to guide others and become a role model. Receiving such
guidance and tips to run the business may encourage IBOs to stay active in the business.
In general, the higher penetration rate (number of total IBOs adjusted for the labor
force size of a county), the less likely to show inactivity. Thus, when a county has a high
demand either for the brand or the product, the percentage of IBOs staying active is higher.
Also, counties with higher average previous IBOs’ earnings from own product selling at (t-1)
showed a higher percentage of active IBOs at (t), consistent with the previous literature.
In terms of competitors from other MLM firms, the estimate for the shortest distance
from the competing firm’s headquarters is positively significant implying that the farther
away from the competing headquarters, the higher the active IBO percentage. Based on the
descriptive statistics of minimum distance from the competing firms’ headquarters (median
= 178.6, mean = 247.2), I also tried using the number of competing firm’s headquarters
66
within 200 miles for the robustness check, shown in Table 2.8. The overall pattern remains
the same. The number of competitor headquarters is negatively statistically significant as
expected (coef. = -0.069, t-value = -3.583), indicating that when a county is highly exposed
to competing firms’ business activity, the percentage of IBOs staying active is lower.
Although initial relationship can be seen between the inactivity behavior and the business
environment, in particular the existence of other IBOs, there are some limitations to this
approach. This analysis does not allow us to examine the different dimensions of others,
namely, network family and geographic proximity, at the individual level. Also, as it is an
aggregate level analysis, individual level heterogeneity cannot be controlled for, which leads
us to the individual-level analysis.
67
(cohort effects).
I first employed a time-independent Cox hazard model by extracting summary statistics from
the individual level time-series panel data to explain the duration of IBOs staying active in
the business. Unlike other parametric survival analysis models, the Cox hazard model is
known for not relying on distributional assumptions in modeling hazard (Cox 1992). Three
types of models were employed, with the mean, initial, and the last value of the covariates
for each IBO. If the results from these three models were consistent, then there would not be
not much need to incorporate the time varying nature of covariates, as the results could be
loosely generalized. Right censored observations were also incorporated in the model. Table
2.9 shows the results.
The results show that the sign and significance across the models are consistent for some
variables; for example, own family ratio within a proximity, distance from two important
individuals, penetration rate, and the shortest distance from the competitors’ headquarters.
In general, however, using the mean, initial and last value of the covariates for each individual
produced inconsistent estimates, with some even flipping the sign, which led to the need to
investigate a model that incorporates time-varying covariates.
Here, I employed a modified version of the Cox hazard model with time-dependent covariates,
also referred to as an extended Cox-model (Therneau et al. 2017). Specifically, for my
monthly data, a discrete-time hazard rate model was applied that incorporates the impact of
time-varying covariates (e.g., density of other IBOs, own time-varying characteristics) on the
likelihood of becoming inactive. It is similar to Gupta (1991), Nikolaeva et al. (2009) in the
sense that the nature of time dependence is incorporated; it differs, however, as it modifies
the Cox hazard model, which does not assume a specific distribution to the hazard rate.
A Cox model with time-independent covariates would compare the survival distributions
between those who do not become inactive (ever) to those who become inactive for each
month. An IBO’s status at the end of the period of data would determine which category
they were in for the entire duration. However, in the time-varying covariate case, the risk
of an event between becoming inactive versus not are compared at each event month, and
would re-evaluate which risk group each IBO belongs to based on whether IBOs became
inactive by that month.
Conditional on an IBO’s time-dependent covariate Xi (t), the Cox model for hazard is:
68
h(t; Xi (t)) = λ(t) × φ(Xi (t)), (2.1)
where φ(Xi (t) is a time-varying covariate function and λ(t) is the base hazard rate. Each
IBO’s hazard rate will depend on its characteristics (e.g., one’s own business activity) and
proximity market factors (e.g., competitive/supportive IBO density), some of which may
vary over time. Thus, the impact of covariates in IBO i ’s hazard rate in period t is captured
in function φ:
where Xi (t) are time-dependent covariates, β are the response coefficients. The inactivity
hazard at time t depends only on the value of the covariates at that time (Xi (t)); β is
constant over time. A positive β̂ implies that an increase in the value of the covariate by
one unit will increase the conditional probability of the IBO’s inactivity by exp(β̂) × 100%,
whereas a negative β̂ implies that there will be a decrease in the probability of inactivity by
exp(β̂) × 100% compared to the baseline hazard.
Suppose there are K distinct inactivity occasions and let (t1 , t2 , ..., tK ) indicate K ordered
inactivity starting months from the entry. The risk set R(tk ) denotes the set of IBOs who
are at risk for inactivity (still active) after tk months of entry.8 For a particular inactivity
at month tk , conditional on the risk set R(tk ), the partial likelihood can be specified as
N δ i
Y exp[β 0 Xi (tk )]
L(β) =
i=1
Σj∈R(tk ) exp[β 0 Xj (tk )]
where δi is the inactivity starting point indicator. Thus, at each event month, the model
compares the current covariate values of the IBOs who had became inactive at month tk to
the current values of all other IBOs who were at risk at that tk , or who had not yet become
inactive (Therneau et al. 2017, Xu 2016).
Table 2.10 shows the individual level survival analysis results incorporating the time varying
covariates. Each column from (1) to (7) indicates a model by gradually adding different types
of variables, where Model (7) is the full model. Model (1) is the baseline model including
only the time, state, family fixed effects and cohort effect (entry time). Model (2) adds
individual-level dummy-coded socio-demographic variables, and Model (3) adds individual
business-related characteristics. Model (4) includes market and competitor information.
8
Entry time of each IBO may be different.
69
Model (5) adds distance information from specific types of connected IBOs, namely upline
and high status. Last, models (6) and (7) include proximity and network family variables,
respectively. Across the models, the signs and the magnitude remain quite consistent.
Table 2.11 is based on Model (7), but reports more detailed results. I begin by discussing
the most comprehensive model (7). After discussing the model, I will compare Models (1)
– (7). Table 2.11, exp(coeff) reports the change in hazard rate from the baseline if there is
a change in a variable by one unit. Let me begin by examining how the existence of others
at the individual level affect an IBO’s inactive behavior. It is followed by network family,
proximity, and other variables that are worth examining.
Individual
As mentioned earlier, two individual IBOs are of interest based on the institutional setting:
Upline and high status IBO. An upline is the person an IBO created the business relationship
with. When a prospective IBO joins the MLM, s/he is connected with only one IBO, namely
an upline. A high status IBO is a successful IBO approved by the MLM firm who takes on
a role as the leader of a sub-group. The zip code distance between these two IBOs and a
focal IBO are measured, respectively. The results in Table 2.11 reveal that, consistent with
the hypothesis, the zip code distance from the sponsoring upline IBO slightly decreased the
inactivity hazard compared to the baseline. The percentage change in the hazard rate from
the baseline hazard is quite small, probably because the unit of the distance in this analysis
is one mile. On the other hand, when the IBO is closer to the high status IBO, the inactivity
hazard slightly increased, unexpectedly. This does not support the hypothesis raised in the
literature review, that IBOs who are geographically close to a high status IBO will tend to
stay active in the business. Further investigation of this issue will be needed to understand
the underlying mechanism. Generally, the absolute effect size is larger for the distance to the
upline distance rather than to the high status IBO, indicating that IBOs’ activity is more
largely influenced by the direct business sponsor.
Network Family
Based on the results in Model (7), I find that the larger the network family size, the lower
the inactive hazard. Compared to the baseline hazard, increasing the family size by one
unit results in a slight decrease in inactive hazard, 0.962 times less inactive compared to the
baseline. This is consistent with previous survey results on organization size and turnover of
salespeople serving in-home customers (Jolson et al. 1987), and literature based on general
organization turnover, in which workers tend to stay in the organization when there are
socialization tactics, high job-embeddedness, and high social ties within the organization.
70
Being associated with a larger network family seems to provide such socialization opportu-
nities with a larger group of other IBOs, in turn serving as an inspirational and motivational
group — having a protective effect.
The percentage of high status IBOs within the same network family is estimated to be
negative, which indicates that an increase in the percentage of high status IBOs within the
family by one unit will significantly decrease the hazard, 0.078 times less hazard compared
to the baseline. Based on the results, it seems that the high status IBOs are not regarded
as potential competitors but rather aspirational entities who motivate IBOs to be like them.
High status IBOs within a network family also show the protective effect.
Proximity
Having more IBOs within the 30-mile radius from the same family seems to increase the
inactivity hazard. This result is somewhat consistent with county-level results, where the
percentage of active IBOs and the number of total IBOs within the same county were found
to be negatively correlated. However, in the individual-level analysis, the total number of
IBOs within proximity did not significantly affect an IBO’s behavior. In other words, the
more IBOs ‘whom you have a high chance to know’ are around you, makes it worse. Although
the cooperative relationship is advised (Sparks and Schenk 2006, Msweli and Sargeant 2001),
IBOs may perceive other IBOs nearby as potential competitors.
In order to examine the role of high status IBOs within the 30 mile radius, I included
the percentage of high status IBOs in the full model.9 Adding this variable did not change
the sign and significance of other variables. Interestingly, the percentage of high status IBOs
who share the same family within the 30- mile radius reduced the IBOs’ risk of inactivity.
Not only having relatively more high status IBOs within the family as a whole network but
also having more of them physically near by seems to help motivate IBOs to engage in the
business.
Other variables
Other variables are also worth highlighting in Table 2.11. First, the penetration rate, which
captures the relative density of IBOs among the labor force within a county, was found to
be negative, moderately decreasing the hazard 0.812 times. This finding is consistent with
the results from the county-level analysis, where the counties with a high penetration rate
showed a higher percentage of active IBOs. Again, having more IBOs relative to the total
number of labor force may imply a stronger presence of a brand or firm in a market with
9
High status % = (# of high status IBOs from own family within 30mi/ Total # of IBOs within 30mi)
71
a higher awareness and familiarity either to end customers or future potential IBOs. This
creates an accommodating environment for initiating sales to end-customers or expanding
business networks (Msweli and Sargeant 2001). This finding also aligns with previous survey
findings that the popularity or acceptance of a product or company is positively related to
sales force tenure (Jolson et al. 1987).
In terms of competitors, the minimum distance to the closest competing MLM firm’s
headquarters is negatively significant, which is similar to the findings in the county-level
analysis. Touching upon the individual business related variables, I see that IBOs who are
successful in the business — either through financial earnings or business activities (i.e.,
building a network (# of frontline and/or # of downline), selling the product (# of retail
customer))- tended to stay active in the business. Although the high status variable was
insignificant, probably due to the correlation among other variables signifying success, the re-
sult is consistent with literature that sheds light on the relationship between job performance
and turnover in MLMs (Jolson et al. 1987, Wotruba 1990b).
Table 2.12 shows the gradual increase in the AIC, as more information was added to the
model. Due to the different number of samples, the difference between AIC of two consec-
utive columns are computed to approximate the overall value of information type added in
explaining the IBO’s inactivity behavior. Based on the ∆ AIC, individual business-related
information (e.g., earnings, business style reflected by the number of direct downlines, total
downlines and registered customers) had the highest contribution in explaining inactivity (∆
AIC = 17,556). Based on the solid literature on job performance and turnover in organization
studies, this finding is expected.
This is followed by the two important individuals related variables, namely the distance
from upline and high status IBOs (∆ AIC = 7,643). The improvement is significantly large
considering the number of variables added. The improvement may be because these two
individuals are either the most closely connected to an IBO, or the most closely observable
success story.
Next, the third largest increase is from the network family information, such as the size
of family and the percentage of high status IBOs within the family. Interestingly enough,
the gain is even slightly larger than that of the socio-demographic information, which has
a considerably larger set of variables. This improvement implies that, aside from individ-
ual business-related characteristics, the network related information — network size, and
network composition — has a positive influence on an IBO’s activity.
Lastly, and unexpectedly, proximity related information, (e.g., the total number of IBOs,
the percentage of own family, and the percentage of high status IBOs from own family
within 30 miles) had the least explanatory power reflecting inactivity behavior. Based on
72
these results, aside from the two individual IBOs, the existence of others at the network
family level appears to dominate the others nearby.
73
Whether an IBO stays in the business or becomes inactive, the precursor to turnover,
cannot be determined by a small set of variables. Thus, in this section, for prediction
purposes, more detailed information from the data, such as recency or frequency of business
activities, is included. In addition, because inactivity is most likely the outcome of the
interactions between multiple factors such as past business activity, network structure and the
market environment, the prediction of whether an IBO stays active in the business or not is
a function of non-linear relationships between multiple predictor variables. Machine learning
methods have thus been applied to our analysis, as they have proven to be appropriate for
such models (Bajari et al. 2015a). Specifically, I focus on one methodology, neural networks.
74
variable. Second is the output layer, which is the dependent variable or the desired output.
Last is the hidden layer(s), which connects the input layer to the output layer. This layer
determines the non-linear mapping relationship (Chen and Du 2009). Figure 2.8 shows an
example of a neural network with one input layer, one hidden layer and one output layer.
Multiple hidden layers can be incorporated to capture more complex data patterns. Previous
studies indicate that a neural network with even one hidden layer with a sufficient number
of neurons can approximate any continuous function (Kaastra and Boyd 1996). The number
of neurons in the hidden layer should fall somewhere between the input layer size and the
output layer size, but there is no rule of thumb. Each layer is connected with weights, which
I recover through algorithms. Figure 2.8 shows a partial set of weights, labeled w10 , ..., w13 ,
which connects the hidden layer and the output layer. Lastly, the activation function allows
us to generate actual network output, which is used to match with the desired value. Given
Figure 2.8, the activation function plays a role in neurons in the hidden layer and the output
layer, colored in orange. Different types of functions, such as sigmoid, linear, logit, hyperbolic
tangent (West et al. 1997), can be used as the activation function.
75
Recency (e.g., the elapsed time since last purchase or recruiting) and Frequency (e.g., the
number of months with prior purchases or with recruiting) are the new category added,
whereas Monetary value (Total amount of selling, Total number of people recruited) appears
under the Individual business-related category.
Lastly, other variables include time, month, network family and state fixed effects. The
same set of variables shown in Table 2.13 is used across the three different analyses presented.
This section compares neural network with other methods. Here, the purpose of the analysis
is to provide a way to pin-point the IBOs who are likely to become inactive in the following
month. The output of the algorithm is 1 indicating the IBO will become inactive following
month for exact six consecutive months, and 0 otherwise. Figure 2.9 shows the visual
representation of the dependent variable. The orange box indicates the period without
engagement for less than six months, thus treated as active. The purple boxes indicate
12
A false negative rate of 0.x means that I correctly identify(100-(100*0.x))%
76
the period without engagement six months or more. The Ynext is coded as 1 if the month is
followed by six consecutive no-engagement months. If an IBO later came back to the market,
just like the observation between the two purple boxes in Figure 2.9, the observation is again
treated as active until it becomes inactive the following month for exactly six consecutive
months.
In order to provide a comparison between the binary logistic model and the neural net-
work model, prediction using the traditional method was also implemented. The same set
of independent variables is used for the two models. The logistic model is widely used for
binary predictions. However, unless one pre-specifies interaction terms among variables,
the method is restricted in being able to capture the non-linear relationships among all the
variables. A neural network can naturally incorporate such non-linear relationships among
the variables through non-linear activation functions that connect neurons in different layers
(orange highlighted circles in Figure 2.8) and also by having multiple hidden layers.
For all the neural network models, I used H2O, which is a fast, scalable open source
machine learning tool (Candel et al. 2016). H2O can be used in R, Python, and other
platforms, and it is known to provide faster speed compared to other neural network packages
in R (e.g., nnet, neuralnet). For training, the data was randomly split into a training
set (70%), a validation set (15%), and a test set (15%). The validation set was used to
overcome the over-fitting problem in the model. Two hidden layers with 130 and 90 neurons
respectively, were used.
Table 2.14 shows a direct comparison of the prediction on the test data between the neural
network and binary logistic regression. Across all metrics, the neural network outperformed
the logistic regression results, particularly on the precision metric. Precision indicates among
the predicted inactivities how precisely the prediction reflects the true data. Thus, the neural
network outperformed in making a precise prediction. Table 2.15 shows the contingency table
based on the neural network.
The second approach is consistent with what was used in the previous sections. With no
evidence of engaging in the business for a consecutive 6 months (t+1) – (t+6), and having no
further indication of returning to the business, the observation right before the consecutive 6
months, (t) is identified as an inactive point. Following this approach, the dependent variable
is constructed based on the very last inactivity behavior for each individual. Figure 2.10
shows an example. The purple boxes indicate inactive period, where there is no engagement
for six months or longer. The orange box indicates no engagement, yet the duration is less
than six months. For each individual and for each month, the dependent variable is coded
77
as the number of quarters left until start of the inactivity (Yq 1 in Figure 2.10).
The reason for using the quarter-based dependent variable (Yq 1) instead of predicting the
exact month (Ym1 ) is to the reduce the unbalanced data problem. For example, 91% IBOs
were categorized as active. With so few observations facing the inactive period, categorizing
the duration by month, would leave only a small number of observations for each month
category to learn from. Lumping the remaining duration specified in month into a quarter
results in relatively more observations for each quarter category to learn from.
Combining the months into quarters seemed to be the best approach for now. These
quarters indicate when an IBO will face the inactive period, such as within one quarter, two
quarters, up to five quarters. They are treated as categorical variables, similar to solving
the classification problem. The output of the algorithm is a quarter-based prediction: how
many quarters left until the inactive month. An additional category, ‘active’ is also defined,
in order to keep the observations that are not followed by the inactive period.
For training, the data was randomly split into a training set (70%), a validation set
(15%), and a test set (15%,), consistent with the previous approach. The validation set is
used to overcome the over-fitting problem in the model. Ten-cross validation was used and
two hidden layers was specified with 100 and 30 neurons respectively. For the activation
function, rectifier was used. One issue that arise is that of the unbalanced sample,as shown
in column ‘Data %’ in Table 2.17. Such data imbalance leads to a bias of the classifier
towards the majority class. To handle this discrepancy, the training data has been balanced
so that the minority classes are over-sampled and the majority classes are under-sampled.
The out of sample prediction is shown in Table 2.17. Based on the second approach,
the accuracy of 79.9% was quite high. In general, the method did well in distinguishing
those observations without future inactive behavior from those with upcoming behavior
(Active vs the rest). In addition, the model was sufficient for predicting the inactivity
period beginning within a quarter: among the data with an inactive period starting within a
quarter, 77.9% were correctly identified. However, it did not perform well in distinguishing
the future quarters in which an IBO might become inactive. Lack of precision for the longer
term prediction could be due to the definition of inactivity, which leaves smaller number of
observations to learn the IBOs’ behavior for the longer term prediction.
Table 2.18 shows the top twenty-five relative importance of the input variables. The
relative importance can be approximated using the estimated weights. Even though the
impression is that often the neural network is a black box, there have been attempts to
recover the relative importance utilizing the estimated weights (Olden et al. 2004, West et al.
1997). The method reported here relies on Gedeon’s method (Gedeon 1997), implemented
by H2O. Based on the method, the importance of input i on output k is computed as the
78
product of connecting weights going through r, where r is all possible hidden neurons. The
relative importance is obtained by normalizing Qik s so that the max is one.13
For illustration purpose, when there are m input neurons and one hidden layer with n
hidden neurons, measure for the contribution of an input i to a hidden neuron j is defined
as Pij where
|wij |
Pij = Pm (2.3)
p=1 |wpj |
Although the model did not precisely predict when future inactivity would take place, it
is still meaningful to examine which variables determined the prediction the most. Interest-
ingly, along with the individual business characterstics (e.g., Activation days, # frontline)
and socio-demographic variables (e.g., age, income groups, which appear more often than
other socio-demographic variables), variables that are related to network family (high status
percentage within a family, network family fixed effects) and important individuals (distance
from high status and upline) are ranked relatively high. The variables related to others are
boldfaced. Although the statistical significance cannot be tested using machine learning
methods, and figuring out signs for variables are on-going question to answer, given the
results, how IBOs are exposed/connected to other IBOs seems to affect inactive behavior
based on the relative importance of the variables.
The third approach is to detect every inactive period equal to or longer than six consecutive
months. IBOs can go through multiple inactive periods. This approach differs from the
second one in how I treat such return to activity. In the second approach, if a return is
observed after the first inactive period, the first inactive period is categorized as active, until
the very last inactive period is observed, showing no evidence of a return to the business.
13
Based on the H2O Java code, the relative importance is computed based on the weights from the first
two hidden layers possibly for the sake of simplicity and speed. Thus, if the model contains more than two
hidden layers, it is recommended to compute the relative importance manually rather than relying on H2O.
79
However, in the third approach, each inactive period is treated as a separate occasion. Thus,
for each IBO, there may be multiple starting points of inactive periods. Similar to the second
approach, the dependent variable is coded as the number of quarters left until the start of
inactivity (Yq2 in Figure 2.10).
Unlike approach two, one additional category is added on top of quarters and active
category, which indicates the time when an IBO is going through the inactive period. In
Figure 2.10, such category is identified as zero.
Similar to approach two, the data was again randomly divided into a training set (70%), a
validation set (15%), and a test set (15%). In order to keep the consistency, the random seed
was remained unchanged. Ten-cross validation is applied with two hidden layers with 100
and 30 neurons, respectively. For the activation function, hyperbolic tangent is used. The
classification results are shown in Table 2.19. Prediction from the third approach is 81.4%,
slightly higher than that in the second approach (79.9%). Similar to the second approach, the
model is effective in selecting the observations (IBO at a specific month) within the inactive
period, or the observations that are not followed by inactive period (active); however, the
duration prediction is often confused with the active category.
A similar conclusion can be drawn from the relative variable importance. Based on
Table 2.20, the role of high status is more prominent: the percentage of high status within
a family, and in proximity both appeared to have relatively high importance. In addition,
the percentage of high status IBOs within the same family appeared in both approach two
and three, with a fairly high rank. Although a argument for causality may be hard to make,
there seems to be consistency in what was found in thee survival analysis, namely, that the
IBOs’ inactivity behavior is related to the surrounding others.
80
to infuse additional stimuli to keep IBOs engaged in the business. High status not only
comes with a financial benefit but also social recognition. High status IBOs sell products
and recruit others, just as no-status IBOs. Thus, they differ from the high level managers of
traditional organization, who have different roles, with stronger vertical authority. The high
status IBOs, however, face additional responsibility to gather a subgroup of IBOs.
Using both descriptive county-level analysis regression and individual-level survival anal-
ysis, I found consistent evidence that being near high status IBOs decreased the likelihood
of inactive behavior — a protective effect. The results held at the network family level
(percentage of high status within a family) and within a given proximity (percentage of high
status from the same family within 30 miles), although the distance from a high status IBO
did not necessarily decrease inactivity behavior. As for individuals, staying near a sponsoring
upline showed a decrease in inactivity hazard. In addition, IBOs associated with a larger
network family tended to stay active in the business, potentially by fulfilling the need for
social connection (Bhattacharya and Mehta 2000, DSA 2015c). Or, possibly a large network
family may have a larger number of successful IBOs, easily developing improved training
programs and more effective sales support activities (Jolson et al. 1987). Yet geographic
competition was inevitable within a certain boundary. One difference in the competition
effect in MLM is that only IBOs associated with the same network family increased the
inactive hazard (percentage of own family within 30 miles). Moreover, I found that when
the market has a high penetration rate, IBOs are less likely to face the inactivity hazard.
In terms of the relative importance across the different dimensions in explaining inactivity,
aside from individual business characteristics, which are known to highly predict turnover,
individual level others had the strongest explanatory power, followed by network family, and
then proximity.
Following the county and individual level analysis, I redirected the problem from an un-
derstanding of the inactive phenomenon to the prediction of inactivity (predicting inactive
period incident and duration). I applied one machine learning technique, a neural network,
to examine its applicability to this type of problem. The model did a decent job in terms
of prediction, resulting in roughly 81% accuracy; in particular, given a time point, it distin-
guished well which IBOs would stay active (approach two and three), or would fall during
inactive period (approach three). Based on approach one and two, the model was also suf-
ficient for predicting the inactive period occurring within a quarter. However, predictions
of when IBOs would experience an inactivity period in the distant future were inconclusive,
in particular, for inactive periods occurring after a quarter (approach three). The relative
importance across more than two-hundred variables showed that the variables related to the
existence of others ranked quite high, even in the neural network models. In particular,
81
the percentage of high status within a family consistently appeared as one of the top five
variables explaining future inactivity behavior.
This study contributes to the turnover literature of the MLM industry. Some of the
industry-specific characteristics of MLM do not exist or do not align with traditional hier-
archical firms. Findings such as status or network family having a protective effect against
IBOs’ inactivity can perhaps be used as a building block for the MLM firm’s strategy and
can shed light on how the MLM industry could preserve their workforce. Even if some
research touches upon MLM turnover (e.g.,Jolson et al. (1987), Wotruba et al. (2005)), to
the best of my knowledge, this paper is the first attempt to understand an IBO’s behavior
through the lense of intricate relationships with other IBOs, such as distance from impor-
tant individuals, connection in a network family, or exposure to other IBOs near by. In
addition, methodologically, this study fills a gap in the MLM turnover literature. Previous
studies have examined turnover behavior through in-person surveys, relying on a qualitative
approach. The quantitative analysis that has been provided in this study, however, allows
us to further exploit some of the factors mentioned above, such as important individuals,
network family, and proximity, which cannot be directly addressed through surveys. Based
on the empirical data, our study can provide data-driven suggestions for better understand-
ing turnover in MLMs. For example, despite the associated cost, MLM firms could consider
relaxing the high status standard to increase the number of high status IBOs, which would
managerially motivate existing IBOs to remain active. Of course, the firm may need to
consider the decrease in the value of scarcity and symbolic meaning of high status caused
by the increase of high status IBOs. Alternatively, MLM firms could take an initiative to
manage the network families — providing ways for small families to collaborate with one
another to share manpower, and give IBOs a sense of belonging to a big network family.
In addition, although the explanatory power was low, potential competition among IBOs
sharing the same family within a geographical region should be carefully scrutinized in order
to prevent inactivity.
Some limitations in this research can be addressed in future research. First, the model
used to run the individual analysis is a variant of the Cox model with time-varying covariates,
which perhaps is not as robust compared to other survival models relying on distributional
assumptions. Despite the drawbacks of using this model, I find the interesting evidence
for influence of others. Moving further, expanding the model using a parametric hazard
model (Gupta 1991, Nikolaeva et al. 2009) can potentially enrich the research. Extending
the survival model with a parametric baseline hazard, may allow me to directly compare the
time to inactivity from both methods. Second, the results are based on how the inactivity
was defined. Perhaps defining it differently might result in different implications. Trying
82
different ways to see if the results are robust will be helpful for generalizing the results. Third,
in defining proximity, 30-mile radius was used as the boundary. However, other potential
boundaries should be tested for future robustness check. Potentially, referring to how other
industries define the sales territory could help set the boundary for proximity. Lastly, moving
forward, finding a better way to predict the exact time of upcoming inactivity behavior will
be a question that remains to be answered.
83
2.8 Tables
84
Table 2.4: Individual-level Descriptive Statistics
Min. 1st Q. Median Mean 3rd Q. Max.
Upline distance 0 4.142 13.638 149.891 54.745 5747.027
Individual
High status distance 0 16.09 80.06 346.12 393.02 5091.19
Family size 248 20312 37616 42850 61199 100724
Family # High status 0 0 3 12.95 11 190
% High status 0% 0.691% 0.799% 0.865% 0.987% 2.058%
# Own family 1 122 405 1342 1310 15450
% Own family 0.008% 8.783% 19.451% 23.136% 33.884% 89.189%
Proximity # Total IBO 5 854 2770 6091 7273 44791
% High status 0% 0% 0.094% 0.215% 0.272% 13.864%
# High status 0 0 3 12.95 11 190
Market Penetration 0.000 0.003 0.004 0.004 0.005 0.024
85
Table 2.6: County-level Correlation Table
% Active IBO # High status % High status # Total IBO Unemployment Penetation Mean Personal Mean Bonus Mean earning
% Active IBO 1.000 -0.005 0.058 -0.039 0.000 -0.009 0.011 0.097 0.071
# High status -0.005 1.000 0.184 0.690 -0.029 0.111 0.069 0.169 0.149
% High status 0.058 0.184 1.000 0.060 -0.033 0.042 0.159 0.731 0.584
86
# Total IBO -0.039 0.690 0.060 1.000 -0.034 0.136 0.059 0.074 0.075
Unemployment 0.000 -0.029 -0.033 -0.034 1.000 -0.150 -0.089 -0.067 -0.084
Penetation -0.009 0.111 0.042 0.136 -0.150 1.000 0.088 0.109 0.105
Mean Personal 0.011 0.069 0.159 0.059 -0.089 0.088 1.000 0.453 0.363
Mean Bonus 0.097 0.169 0.731 0.074 -0.067 0.109 0.453 1.000 0.749
Mean earning 0.071 0.149 0.584 0.075 -0.084 0.105 0.363 0.749 1.000
Table 2.7: County-level Results (a)
Estimate Std. Error t-value P-value
(Intercept) 0.868 0.025 34.439 0.000
lag.Total IBO -0.00001 0.000 -5.405 0.000
lag.High status % 0.166 0.071 2.344 0.019
lag.Penetration rate 4.890 0.488 10.017 0.000
lag.UnempRate 0.001 0.001 0.972 0.331
lag.meanPersonal$ 0.000003 0.000 2.104 0.035
Competitor shortest distance 0.00027 0.000 3.583 0.000
Year dummy Y
Month dummy Y
County dummy Y
Adjusted R-squared 0.773
87
Table 2.9: Individual-level Analysis Results with Time Independent Covariates
mean X first X last X
Variable type X coef se(coef) coef se(coef) coef se(coef)
Family size/1000 -0.156*** 0.001 0.006*** 0.002 -0.137*** 0.001
Family
High status % 8.226*** 0.072 -0.087 0.073 5.016*** 0.062
Own family ratio 0.413*** 0.040 0.236*** 0.039 0.473*** 0.040
Proximity Total IBO/1000 -0.007*** 0.001 0.002 0.001 -0.005*** 0.001
High status % 0.027 0.017 -0.019 0.015 -0.002 0.019
Upline distance -0.0004*** 0.000 -0.0005*** 0.000 -0.0004*** 0.000
Individuals
High status distance 0.0002*** 0.000 0.0004*** 0.000 0.0003*** 0.000
TotalPenetCounty % -2.386*** 0.049 -0.259*** 0.048 -2.648*** 0.045
Market
UnempRate -0.027*** 0.005 0.011** 0.005 -0.025*** 0.005
Competitor shortestdistcomp -0.001*** 0.000 -0.0002*** 0.000 -0.001*** 0.000
High status -0.007 0.225 -0.017 2.552 -0.002 0.115
Earning 0.0002*** 0.000 -0.0003*** 0.000 0.0002*** 0.000
relative level (by LOA) 0.002 0.041 -0.673*** 0.037 -0.220*** 0.042
Individual business # Frontline -0.088*** 0.014 -0.317*** 0.018 -0.084*** 0.011
# Total downline -0.325*** 0.006 -0.317*** 0.009 -0.257*** 0.005
# Retail customer -0.131*** 0.005 -0.270*** 0.006 -0.065*** 0.004
Age Y Y Y
Language (4 groups) Y Y Y
Marital status (3 groups) Y Y Y
Socio-demographic
Gender (3 groups) Y Y Y
Ethnicity (8 groups) Y Y Y
Income level (10 groups) Y Y Y
Entry dummy Y Y Y
Other controls State dummy Y Y Y
LOA dummy Y Y Y
Log likelihood -557163.8 -624098.8 -544212
*** p-value < 0.01, ** p-value < 0.05, * p-value < 0.1
88
Table 2.10: Individual-level Survival Analysis Results
Variable type Variables (1) (2) (3) (4) (5) (6) (7)
Family size/1000 -0.039***
Network Family
High status % -2.540***
Own family ratio 0.315*** 0.262***
Proximity Total IBO/1000 -0.0008 0.001
High status % -0.0511*** -0.0442***
Upline distance -0.00006*** -0.00005*** -0.00005***
Individual High status distance 0.00002** 0.00003** 0.00003***
TotalPenetCounty % -0.238*** -0.236*** -0.269*** -0.202***
Market
UnempRate 0.014*** 0.014*** 0.011** 0.012***
Competitor shortestdistcomp -0.0002** -0.0002** -0.0003*** -0.0002**
High status -0.015 -0.015 -0.015 -0.015 -0.016
Earning -0.0001*** -0.0001*** -0.0001*** -0.0001*** -0.0001***
relative level (by LOA) -0.453*** -0.447*** -0.442*** -0.463*** -0.512***
Individual business characteristics
# Frontline -0.262*** -0.261*** -0.261*** -0.261*** -0.257***
89
# Total downline -0.235*** -0.235*** -0.235*** -0.235*** -0.238***
# Retail customer -0.118*** -0.118*** -0.118*** -0.119*** -0.122***
Age Y Y Y Y Y Y
Language (4 groups) Y Y Y Y Y Y
Marital status (3 groups) Y Y Y Y Y Y
Socio-demographic characteristics
Gender (3 groups) Y Y Y Y Y Y
Ethnicity (8 groups) Y Y Y Y Y Y
Income level (10 groups) Y Y Y Y Y Y
Entry dummy Y Y Y Y Y Y Y
Year dummy Y Y Y Y Y Y Y
Other controls Month dummy Y Y Y Y Y Y Y
State dummy Y Y Y Y Y Y Y
LOA dummy Y Y Y Y Y Y Y
Log-Likelihood -625211.1 -624571.4 -615207.3 -615375.7 -611551.9 -611516.4 -610849.7
N Obs. 1996546 1996546 1996453 1995957 1981810 1981810 1981810
*** p-value < 0.01, ** p-value < 0.05, * p-value < 0.1
Table 2.11: Individual-level Survival Analysis Results - Model (7)
Variable types Variables coef exp(coef) se(coef) z
Family size/1000 -0.039*** 0.962 0.001 -27.413
Network family
High status % -2.540*** 0.079 0.080 -31.674
Own family ratio 0.262*** 1.299 0.038 6.805
Proximity (30 mi) Total IBO/1000 0.001 1.001 0.001 1.159
High status % -0.0442*** 0.957 0.017 -2.539
Upline distance -0.00005*** 1.000 0.000 -4.362
Individual
High status distance 0.00003*** 1.000 0.000 3.632
Market County penetration rate % -0.202*** 0.817 0.043 -4.699
UnempRate 0.012*** 1.012 0.005 2.654
Competitor shortestdistcomp -0.0002** 1.000 0.000 -2.554
High status -0.016 0.985 1.467 -0.011
Earning -0.0001*** 1.000 0.000 -3.385
relative level (by LOA) -0.512*** 0.599 0.040 -12.748
Individual business
# Frontline -0.257*** 0.773 0.012 -22.237
# Total downline -0.238*** 0.788 0.005 -50.506
# Retail customer -0.122*** 0.885 0.004 -31.6
*** p-value < 0.01, ** p-value < 0.05, * p-value < 0.1
90
Table 2.13: Variables for Machine Learning
IV
Family size, lag family size, network growth from (t-1)
Network Family High status %, #
Min, max level, Depth of a family (min-max)
Own family %, #
Total IBO
High status from own family %, #
Proximity
Total %, # of high status
Network Family count
% of own high status among # of high status in proximity
Individual Distance from a high status IBO & Upline
TotalPenetCounty %, Labor force, UnempRate, # employed
Market (County) Diversity across the families withint a county (2)
TotalIBOcounty
The shortest distance from competing firms’ headquarters
Competitor # of competitors’ headquarters by each band (100mi - 500mi)
# of competitors’ headquarters within 300mi, 500mi
High status
duration from entry (running variable)
financial report (t): Total Earnings, Product sales(Point, $), Bonus(Point, $)
lag financial report (t-1)
Cumulative average of financial report until (t-1)
Missed value in finanical (t-1)
Activation of MLM business (30 day, 90day, 120 day)
Individual business
relative level (by LOA), absolute level
# Frontline, # Total downline, # Retail customer
Addition/removal of # Frontline, # Total downline, # Retail customer from (t-1)
Cumulative addition/removal since the beginning
Approximated bonus bracket at (t)
Bonus bracket attainment (first 3%, 9%, 18%, 25%)
Month since the bonus bracket attainment
Age
Language
Marital status
Socio-demographic
Gender
Ethnicity
Income level
Total no engagement months until t
Total engagement months until t
Continuous engagement # until t
Recency, Frequency # of engagement after the first inactive duration, if applicable
Cumulative average of no engagement
maximum rest until t
after the first inactive duration (dummy, count), if applicable
Entry dummy (cohort)
Year
Other controls Month
State
91
Family
Table 2.14: Model Comparison between NN and Logistic Model
Neural Network Logistic
Accuracy 0.859 0.839
Recall 0.844 0.843
Precision 0.738 0.699
Table 2.15: Contingency Table for Approach One (Out of Sample Prediction)
Prediction
0 1 Row sum % Data Row correct %
0 119736 (59.7%) 18610 (23.02%) 138346 69% 87%
Data
1 9717 (4.8%) 52510 (26.2%) 62227 31% 84%
Colum sum 129453 71120 200573
92
Table 2.17: Contingency Table for Approach Two (Out of Sample Prediction)
Prediction
Active Q1 Q2 Q3 Q4 Q5 Row sum Data % Row correct %
Active 108818 16013 215 1068 358 37 126509 62.66% 86.0%
Q1 14217 51853 93 302 67 8 66540 32.96% 77.9%
93
Q2 3549 2015 89 126 10 4 5793 2.87% 1.5%
Data
Q3 1409 390 14 363 27 0 2203 1.09% 16.5%
Q4 440 63 2 53 164 4 726 0.36% 22.6%
Q5 70 7 2 1 14 32 126 0.06% 25.4%
Column sum 128503 70341 415 1913 640 85 201897 100%
Table 2.18: Variable Importance Based on Approach Two
Variable Relative importance
1 Year 2007 1.00
2 High status % within a family 0.98
3 Activation of MLM (30 days) 0.95
4 Entry cohort effect (200603) 0.94
5 Age 0.92
6 Income group 6 0.88
7 # Frontline 0.88
8 Entry cohort effect (200604) 0.86
9 Income group 1 (the Lowest) 0.86
10 Income group 9 (the Highest) 0.85
11 # of family within proximity 0.84
12 Income group 8 0.83
13 Network family dummy (37) 0.83
14 Entry cohort effect (200601) 0.82
15 Distance from a high status IBO 0.82
16 Gender missing 0.82
17 Entry cohort effect (200602) 0.81
18 Income group 7 0.81
19 Level 0.81
20 Cumulative average of no engagement 0.80
21 # of competitors’ 200mi 0.79
22 Distance from an Upline 0.78
23 State fixed effect (Georgia) 0.77
24 Income group 4 0.76
25 Ethnicity missing 0.76
94
Table 2.19: Contingency Table for Approach Three (Out of Sample Prediction)
Prediction
Inactive Q1 Q2 Q3 Q4 Q5 Active Row sum Data % Row correct %
Inactive 52931 0 0 0 0 0 7890 60821 30.12% 87.03%
Q1 1 1383 59 16 2 0 11040 12501 6.19% 11.06%
95
Q2 0 80 52 14 1 0 1426 1573 0.78% 3.31%
Data Q3 0 9 9 34 1 0 333 386 0.19% 8.81%
Q4 0 6 0 13 6 0 70 95 0.05% 6.32%
Q5 0 1 0 1 3 1 6 12 0.01% 8.33%
Active 15001 1104 215 146 24 0 110019 126509 62.66% 86.97%
Column sum 67933 2583 335 224 37 1 130784 201897 100%
Table 2.20: Variable Importance Based on Approach Three
Relative importance
1 Engagement 1.00
2 Bonus bracket 0.54
3 Age 0.50
4 # Frontline 0.50
5 High status % within a family 0.48
6 State fixed effect (FL) 0.48
7 Income group 4 0.47
8 Entry cohort effect (200607) 0.46
9 Diff (#Downline IBO) 0.46
10 Ethnicity missing 0.45
11 Entry cohort effect (200604) 0.45
12 Network family dummy (2) 0.45
13 Income group 6 0.45
14 Activation of MLM (30 days) 0.45
15 Income group 3 0.45
16 State fixed effect (IL) 0.44
17 Entry cohort effect (200603) 0.44
18 Level 0.44
19 Network family dummy (37) 0.44
20 # of competitors’ 200mi 0.43
21 Entry cohort effect (200601) 0.43
22 High status % within a family in proximity 0.43
23 Network family dummy (21) 0.43
24 Engagement after inactivity 0.42
25 State fixed effect (NC) 0.42
96
2.9 Figures
97
Figure 2.3: Percent of High Status Members Given a Network Family - Oct. 2007
98
Figure 2.6: Example of the Lowest and the Highest Total IBOs in Proximity
99
Figure 2.8: Example of Neural Network with Three Layers
100
Bibliography
101
Bell, David R, Sangyoung Song. 2007. Neighborhood effects and trial on the internet: Evidence
from online grocery retailing. Quantitative Marketing and Economics 5(4) 361–400.
Bhattacharya, Patralekha, Krishna Kumar Mehta. 2000. Socialization in network marketing orga-
nizations: is it cult behavior? The Journal of Socio-Economics 29(4) 361–374.
Biggart, Nicole Woolsey. 1989. Charismatic capitalism: Direct selling organizations in America.
University of Chicago Press.
Bloch, Brian. 1996. Multilevel marketing: what’s the catch? Journal of Consumer Marketing
13(4) 18–26.
Bonabeau, Eric. 2002. Agent-based modeling: Methods and techniques for simulating human
systems. Proceedings of the National Academy of Sciences 99(suppl 3) 7280–7287.
Bonoma, Thomas. 1991. This snake rises in bad times. Marketing News 25(4) 16.
Bosley, Stacie, Kim K McKeage. 2015. Multilevel marketing diffusion and the risk of pyramid
scheme activity: The case of fortune hi-tech marketing in montana. Journal of Public Policy
& Marketing 34(1) 84–102.
Boyd, Donald, Hamilton Lankford, Susanna Loeb, James Wyckoff. 2013. Analyzing the deter-
minants of the matching of public school teachers to jobs: Disentangling the preferences of
teachers and employers. Journal of Labor Economics 31(1) 83–117.
Braun, Michael, David A. Schweidel. 2011. Modeling customer lifetimes with multiple causes of
churn. Marketing Science 30(5) 881–902. URL http://www.jstor.org/stable/23011984.
Brodie, Stewart, John Stanworth, Thomas R Wotruba. 2002. Comparisons of salespeople in mul-
tilevel vs. single level direct selling organizations. The Journal of Personal Selling and Sales
Management 67–75.
Candel, Arno, Viraj Parmar, Erin LeDell, Anisha Arora. 2016. Deep learning with h2o.
Centola, Damon. 2010. The spread of behavior in an online social network experiment. Science
329(5996) 1194–1197.
Chaplot, Devendra Singh, Eunhee Rhim, Jihie Kim. 2015. Predicting student attrition in moocs
using sentiment analysis and neural networks. AIED Workshops.
Chen, Gilad, Robert E Ployhart, Helena Cooper Thomas, Neil Anderson, Paul D Bliese. 2011. The
power of momentum: A new model of dynamic relationships between job satisfaction change
and turnover intentions. Academy of Management Journal 54(1) 159–181.
Chen, Jiawei. 2013. Estimation of the loan spread equation with endogenous bank-firm matching.
Structural Econometrics Models (E. Choo and M. Shum, eds.) 31 251–290.
Chen, Wei-Sen, Yin-Kuan Du. 2009. Using neural networks and data mining techniques for the
financial distress prediction model. Expert systems with applications 36(2) 4075–4086.
Choi, Jeonghye, Sam K Hui, David R Bell. 2010. Spatiotemporal analysis of imitation behavior
across new buyers at an online grocery retailer. Journal of Marketing Research 47(1) 75–89.
102
Christakis, Nicholas A, James H Fowler, Guido W Imbens, Karthik Kalyanaraman. 2010. An
empirical model for strategic network formation. Tech. rep., National Bureau of Economic
Research.
Clark, Simon. 2006. The uniqueness of stable matchings. Contributions in Theoretical Economics
6(1) 1–28.
Clothier, Peter J. 1997. Multi-level Marketing: A Practical Guide to Successful Network Selling.
3rd ed. Kogan page.
Coleman, Lynn G. 1989. Sales force turnover has managers wondering why. Marketing News 23(25)
6–7.
Collamer, Nancy. 2013. Can you really make money in direct sales? URL http://www.forbes.
com/sites/nextavenue/2013/04/01/can-you-really-make-money-in-direct-sales/2/
18246f154cf1. [forbes.com; Updated 1-April-2013].
Cotton, John L, Jeffrey M Tuttle. 1986. Employee turnover: A meta-analysis and review with
implications for research. Academy of management Review 11(1) 55–70.
Coughlan, Anne T. 2012. 7 things you should know about direct selling. Wordpress,
http://distribunomics.wordpress.com/2012/11/21/7-things-you-should-know-about-direct-
selling .
Coughlan, Anne T, Kent Grayson. 1993. Multi-level marketing executives industry survey summary
report. Unpublished report, Northwestern University, Evanston .
Coughlan, Anne T, Kent Grayson. 1998. Network marketing organizations: Compensation plans,
retail network growth, and profitability. International Journal of Research in Marketing 15(5)
401–426.
Coussement, Kristof, Dirk Van den Poel. 2009. Improving customer attrition prediction by inte-
grating emotions from client/company interaction emails and evaluating multiple classifiers.
Expert Systems with Applications 36(3) 6127–6134.
Cox, David R. 1992. Regression models and life-tables. Breakthroughs in statistics. Springer,
527–541.
Crittenden, Victoria L, William F Crittenden. 2004. Developing the sales force, growing the busi-
ness: The direct selling experience. Business Horizons 47(5) 39–44.
Currarini, Sergio, Matthew O Jackson, Paolo Pin. 2009. An economic model of friendship: Ho-
mophily, minorities, and segregation. Econometrica 77(4) 1003–1045.
Dalton, Dan R, William D Todor, Michael J Spendolini, Gordon J Fielding, Lyman W Porter.
1980. Organization structure and performance: A critical review. Academy of management
review 5(1) 49–64.
Daquis, John Carlo P, Angelique O Castañeda, Nelson D Sy, Ranier Joseph V Abgona. 2013.
Profitability and growth topology analysis of unilevel-type of network marketing structures.
The Philippine Statistician 62(2) 13–29.
103
Darmon, René Y. 2008. The concept of salesperson replacement value: A sales force turnover
management tool. Journal of Personal Selling & Sales Management 28(3) 211–232.
DeConinck, James B, Julie T Johnson. 2009. The effects of perceived supervisor support, perceived
organizational support, and organizational justice on turnover among salespeople. Journal of
Personal Selling & Sales Management 29(4) 333–350.
Deng, Wei-Jaw, Wen-Chin Chen, Wen Pei. 2008. Back-propagation neural network based
importance–performance analysis for determining critical service attributes. Expert Systems
with Applications 34(2) 1115–1125.
DSA. 2015a. Direct selling in 2015: An overview. URL http://www.dsa.org/docs/
default-source/research/dsa_2015factsheetfinal.pdf?sfvrsn=8.
DSA. 2015b. Dsa 2015 growth & outlook report: U.s. direct selling in 2014. URL http://www.
dsa.org/docs/default-source/research/researchfactsheet2007-2014.pdf?sfvrsn=0.
DSA. 2015c. Dsa 2015 motivation - in direct selling, success is different for dif-
ferent people. URL http://www.dsa.org/docs/default-source/advocacy/
dsa-successisdifferentfactsheetv4.pdf?sfvrsn=2.
DSN. 2016. 2016 dsn north america 50 list. URL http://directsellingnews.com/index.php/
view/2016_dsn_north_america_50_list1.
Echenique, Federico, SangMok Lee, Matthew Shum. 2010. Aggregate matchings .
Echenique, Federico, SangMok Lee, Matthew Shum. 2013. Partial identification in two-sided match-
ing models .
Fader, Peter S, Bruce GS Hardie, Ka Lok Lee. 2005. Rfm and clv: Using iso-value curves for
customer base analysis. Journal of Marketing Research 42(4) 415–430.
Failla, Don, Joe Hardwick. 1995. How to Build a Large Successful Multi-level Marketing Organiza-
tion. Multi-Level Marketing International, Incorporated.
Fox, Jeremy T. 2008. Estimating matching games with transfers. Tech. rep., National Bureau of
Economic Research.
Frenzen, Jonathan K, Harry L Davis. 1990. Purchasing behavior in embedded markets. Journal of
Consumer Research 17(1) 1–12.
Futrell, Charles M, A Parasuraman. 1984. The relationship of satisfaction and performance to
salesforce turnover. The Journal of Marketing 33–40.
Gale, David, Lloyd S Shapley. 1962. College admissions and the stability of marriage. American
mathematical monthly 9–15.
Gedeon, Tamás D. 1997. Data mining of inputs: analysing magnitude and functional measures.
International Journal of Neural Systems 8(02) 209–218.
Geweke, John. 1999. Using simulation methods for bayesian econometric models: inference, devel-
opment, and communication. Econometric Reviews 18(1) 1–73.
104
Geweke, John, Gautam Gowrisankaran, Robert J Town. 2003. Bayesian inference for hospital
quality in a selection model. Econometrica 71(4) 1215–1238.
Goodman, Charles S, Marvin A Jolson. 1973. Overdue attention to a neglected step-child of
marketing.
Grant, Kelli B. 2012. 10 things direct sales marketers won’t say. Wall Street Journal Market Watch
.
Grayson, Kent. 1996. Examining the embedded markets of network marketing organizations.
Grayson, Kent. 2007. Friendship versus business in marketing relationships. Journal of Marketing
71(4) 121–139.
Griffeth, Rodger W, Peter W Hom. 2001. Retaining valued employees. Sage Publications.
Griffeth, Rodger W, Peter W Hom, Stefan Gaertner. 2000. A meta-analysis of antecedents and
correlates of employee turnover: Update, moderator tests, and research implications for the
next millennium. Journal of management 26(3) 463–488.
Gupta, Sunil. 1991. Stochastic models of interpurchase time with time-dependent covariates. Jour-
nal of Marketing Research 1–15.
Haenlein, Michael. 2013. Social interactions in customer churn decisions: The impact of relationship
directionality. International Journal of Research in Marketing 30(3) 236–248.
Harrison, David A, Katherine J Klein. 2007. What’s the difference? diversity constructs as separa-
tion, variety, or disparity in organizations. Academy of management review 32(4) 1199–1228.
Herballife. 2016. Form 10-k, herbalife ltd. URL http://ir.herbalife.com/annuals.cfm.
Hinkin, Timothy R, J Bruce Tracey. 2000. The cost of turnover: Putting a price on the learning
curve. The Cornell Hotel and Restaurant Administration Quarterly 41(3) 144–21.
Holtom, Brooks C, Terence R Mitchell, Thomas W Lee, Marion B Eberly. 2008. 5 turnover and
retention research: A glance at the past, a closer review of the present, and a venture into the
future. Academy of Management annals 2(1) 231–274.
Hom, Peter W, Anne S Tsui, Joshua B Wu, Thomas W Lee, Ann Yan Zhang, Ping Ping Fu, Lan
Li. 2009. Explaining employment relationships with social exchange and job embeddedness.
Journal of Applied Psychology 94(2) 277.
Hsieh, Chih-Sheng, Lung-Fei Lee. 2012. A structural modeling approach for network formation
and social interactions with applications to students’ friendship choices and selectivity on
activities. Job market paper, Department of Economics, Ohio State University .
Hu, Yansong, Christophe Van den Bulte. 2014. Nonmonotonic status effects in new product adop-
tion. Marketing Science 33(4) 509–533.
Hua, Zhongsheng, Yu Wang, Xiaoyan Xu, Bin Zhang, Liang Liang. 2007. Predicting corporate
financial distress based on integration of support vector machine and logistic regression. Expert
Systems with Applications 33(2) 434–440.
105
Imai, Kosuke, David A van Dyk. 2005. A bayesian analysis of the multinomial probit model using
marginal data augmentation. Journal of econometrics 124(2) 311–334.
Iyengar, Raghuram, Christophe Van den Bulte, Thomas W Valente. 2011. Opinion leadership and
social contagion in new product diffusion. Marketing Science 30(2) 195–212.
Jackofsky, Ellen F, Kenneth R Ferris, Betty G Breckenridge. 1986. Evidence for a curvilinear
relationship between job performance and turnover. Journal of Management 12(1) 105–111.
Jain, S, B Singla, S Shashi. 2015. Motivational factors in multilevel marketing business: A confir-
matory approach. Management Science Letters 5(10) 903–914.
Jamal, Zainab, Randolph E Bucklin. 2006. Improving the diagnosis and prediction of customer
churn: A heterogeneous hazard modeling approach. Journal of Interactive Marketing 20(3-4)
16–29.
Jolson, Marvin A, Alan J Dubinsky, Rolph E Anderson. 1987. Correlates and determinants of sales
force tenure: an exploratory study. Journal of Personal Selling & Sales Management 7(3)
9–28.
Kaastra, Iebeling, Milton Boyd. 1996. Designing a neural network for forecasting financial and
economic time series. Neurocomputing 10(3) 215–236.
Keep, William W., Peter J. Vander Nat. 2014. Multilevel marketing and pyramid schemes in
the united states: An historical analysis. Journal of Historical Research in Marketing 6(2)
188–210.
King, C.W., J.W. Robinson. 2000. The New Professionals: The Rise of Network Marketing as the
Next Major Profession. Prima Soho. URL https://books.google.com/books?id=O6fWk-_
FbrwC.
Klebnikov, Paul. 1991. The power of positive inspiration. Forbes 9 244–251.
Koroth, Abdul Assis. 2014. Antecedents of distributors turnover in multilevel marketing. Indian
Journal of Commerce and Management Studies 5(1) 62.
Ladik, Daniel M, Greg W Marshall, Felicia G Lassk, William C Moncrief. 2000. The relationship
of satisfaction and performance to salesforce turnover: A replication and extension. American
Marketing Association. Conference Proceedings, vol. 11. American Marketing Association, 232.
Lee, Kwee-Fah, Teck-Chai Lau, Kai-Yin Loi, et al. 2016. Driving distributors’ satisfaction in mul-
tilevel marketing (mlm) companies. International Journal of Academic Research in Business
and Social Sciences 6(2) 105–122.
Lee, SangMok. 2014. Incentive compatibility of large centralized matching markets .
Lee, Thomas W, Terence R Mitchell, Brooks C Holtom, Linda S McDaneil, John W Hill. 1999. The
unfolding model of voluntary turnover: A replication and extension. Academy of Management
journal 42(4) 450–462.
Legara, Erika Fille, Anthony Longjas, Rene Batac. 2009. Competition in a social structure. Inter-
national Journal of Modern Physics C 20(01) 1–7.
106
Legara, Erika Fille, Christopher Monterola, Dranreb Earl Juanico, Marisciel Litong-Palima, Caesar
Saloma. 2008. Earning potential in multilevel marketing enterprises. Physica A: Statistical
Mechanics and its Applications 387(19) 4889–4895.
Lemmens, Aurélie, Christophe Croux. 2006. Bagging and boosting classification trees to predict
churn. Journal of Marketing Research 43(2) 276–286.
Lilyquist, Mindy. 2016a. 10 MLM success tips for home business success. URL http:
//homebusiness.about.com/od/homebusinessprofiles/a/10-Mlm-Success-Tips.htm.
About.com; Updated 05-April-2016.
Lilyquist, Mindy. 2016b. MLM success starter guide - tips, tools and resources to help you build
your mlm business. URL http://homebusiness.about.com/od/homebusinessprofiles/tp/
Mlm-Success-Starter-Guide.htm. About.com; Updated 16-May-2016.
Logan, John Allen, Peter D Hoff, Michael A Newton. 2008. Two-sided estimation of mate pref-
erences for similarities in age, education, and religion. Journal of the American Statistical
Association 103(482) 559–569.
Maciel, Leandro Santos, Rosângela Ballini. 2008. Design a neural network for time series financial
forecasting: Accuracy and robustness analysis. Instituto de Economı́a, Universidade Estadual
de Campinas, Sao Paulo-Brasil .
McCulloch, Robert, Peter E Rossi. 1994. An exact likelihood analysis of the multinomial probit
model. Journal of Econometrics 64(1) 207–240.
McCulloch, Robert E, Nicholas G Polson, Peter E Rossi. 2000. A bayesian analysis of the multino-
mial probit model with fully identified parameters. Journal of Econometrics 99(1) 173–193.
McPherson, J Miller, Pamela A Popielarz, Sonja Drobnic. 1992. Social networks and organizational
dynamics. American sociological review 153–170.
Menguc, Bulent, Sang-Lin Han, Seigyoung Auh. 2007. A test of a model of new salespeople’s
socialization and adjustment in a collectivist culture. Journal of Personal Selling & Sales
Management 27(2) 149–167.
Mitchell, Terence R, Brooks C Holtom, Thomas W Lee, Chris J Sablynski, Miriam Erez. 2001. Why
people stay: Using job embeddedness to predict voluntary turnover. Academy of management
journal 44(6) 1102–1121.
Mobley, William H. 1992. Employee turnover: Causes, consequences, and control .
Msweli, Pumela, Adrian Sargeant. 2001. Modelling distributor retention in network marketing
organisations. Marketing Intelligence & Planning 19(7) 507–514.
Msweli-Mbanga, P. 2004. Predicting turnover behaviour of direct salespeople. Southern African
Business Review 8(3) 14–25.
Nam, Sungjoon, Puneet Manchanda, Pradeep K Chintagunta. 2010. The effect of signal quality and
contiguous word of mouth on customer acquisition for a video-on-demand service. Marketing
Science 29(4) 690–700.
107
Narayan, Vishal, Vithala R Rao, Carolyne Saunders. 2011. How peer influence affects attribute
preferences: A bayesian updating mechanism. Marketing Science 30(2) 368–384.
Neslin, Scott A, Sunil Gupta, Wagner Kamakura, Junxiang Lu, Charlotte H Mason. 2006. Defection
detection: Measuring and understanding the predictive accuracy of customer churn models.
Journal of marketing research 43(2) 204–211.
Ni, Jian, Kannan Srinivasan. 2015. Matching in the sourcing market: A structural analysis of the
upstream channel. Marketing Science 34(5) 722–738.
Nikolaeva, Ralitza, Manohar U Kalwani, William T Robinson, S Sriram, et al. 2009. Survival
determinants for online retailers. Review of Marketing Science 7(1) 360.
Olden, Julian D, Michael K Joy, Russell G Death. 2004. An accurate comparison of methods for
quantifying variable importance in artificial neural networks using simulated data. Ecological
Modelling 178(3) 389–397.
Pan, Yihui. 2014. The determinants and impact of executive-firm matches. Available at SSRN
1571892 .
Pancras, Joseph, Srinivasaraghavan Sriram, Vineet Kumar. 2012. Empirical investigation of retail
expansion and cannibalization in a dynamic environment. Management Science 58(11) 2001–
2018.
Park, Minjung. 2008. An empirical two-sided matching model of acquisitions: Understanding
merger incentives and outcomes in the mutual fund industry. Unpublished manuscript available
at http://www. econ. umn. edu/˜ mpark/MergerPaper. pdf .
Peter J. Vander Nat, William W. Keep. 2002. Marketing fraud: An approach for differentiating
multilevel marketing from pyramid schemes. Journal of Public Policy & Marketing 21(1)
139–151. URL http://www.jstor.org/stable/30000715.
Peterson, Robert A, Thomas R Wotruba. 1996. What is direct selling? definition, perspective, and
research agenda. Journal of Personal Selling & Sales Management 16(4) 1–16.
Pratt, Michael G. 2000. The good, the bad, and the ambivalent: Managing identification among
amway distributors. Administrative Science Quarterly 45(3) 456–493.
Price, James L. 1983. Reviewed work(s): Employee turnover: Causes, consequences, and control.
Journal of Marketing Channels 506–507.
Raymond, Mary Anne, John F Tanner Jr. 1994. Selling and sales management in action: Main-
taining customer relationships in direct sales: Stimulating repeat purchase behavior. Journal
of Personal Selling & Sales Management 14(4) 67–76.
Richardson, Robert. 1999. Measuring the impact of turnover on sales. The Journal of Personal
Selling and Sales Management 53–66.
Rossi, Peter E, Robert E McCulloch, Greg M Allenby. 1996. The value of purchase history data in
target marketing. Marketing Science 15(4) 321–340.
108
Roth, Alvin E, Marilda Sotomayor. 1992. Chapter 16 twosided matching. Handbook of Game
Theory with Economic Applications, Elsevier, vol 1 485–541.
Seim, Katja. 2006. An empirical model of firm entry with endogenous product-type choices. The
RAND Journal of Economics 37(3) 619–640.
Sinha, Tanmay, Patrick Jermann, Nan Li, Pierre Dillenbourg. 2014. Your click decides your fate:
Inferring information processing and attrition behavior from mooc video clickstream interac-
tions. arXiv preprint arXiv:1407.7131 .
Sorensen, Morten. 2007. How smart is smart money? a two-sided matching model of venture
capital. The Journal of Finance 62(6) 2725–2762.
Sparks, John R, Joseph A Schenk. 2001. Explaining the effects of transformational leadership:
an investigation of the effects of higher-order motives in multilevel marketing organizations.
Journal of Organizational Behavior 22(8) 849–869.
Sparks, John R, Joseph A Schenk. 2006. Socialization communication, organizational citizenship
behaviors, and sales in a multilevel marketing organization. Journal of Personal Selling &
Sales Management 26(2) 161–180.
Taylor, Jon M. 2011. The case (for and) against multi-level marketing: The complete guide to
understanding the flaws – and proving and countering the effects – of endless chain “oppor-
tunity” recruitment, or product-based pyramid schemes. URL http://www.mlmwatch.org/
01General/taylor.pdf. Consumer Awareness Institute.
Therneau, Terry, Cynthia Crowson, Elizabeth Atkinson. 2017. Using time dependent covariates
and time dependent coefficients in the cox model .
Tortora, Andrea. 2015. Direct selling’s strength in the world’s billion dollar markets. Direct Selling
News .
Uetake, Kosuke, Yasutora Watanabe. 2012a. Entry by merger: Estimates from a two-sided matching
model with externalities. Available at SSRN 2188581 .
Uetake, Kosuke, Yasutora Watanabe. 2012b. A note on estimation of two-sided matching models.
Economics Letters 116(3) 535–537.
Van den Bulte, Christophe, Stefan Hendrik Katrin Wuyts. 2007. Social networks in marketing.
MSI Relevant Knowledge Series .
Vellido, Alfredo, Paulo JG Lisboa, J Vaughan. 1999. Neural networks in business: a survey of
applications (1992–1998). Expert systems with applications 17(1) 51–70.
West, Patricia M, Patrick L Brockett, Linda L Golden. 1997. A comparative analysis of neural
networks and statistical methods for predicting consumer choice. Marketing Science 16(4)
370–391.
WFDSA. 2016. Global direct selling - 2015 retail sales URL http://www.wfdsa.org/files/pdf/
global-stats/SalesReport2015v25-31-2016.pdf. World Federation of Direct Selling As-
sociations.
109
Wotruba, Thomas, Stewart Brodie, John Stanworth. 2005. Differences in turnover predictors be-
tween multilevel and single level direct selling organizations. The International Review of
Retail, Distribution and Consumer Research 15(1) 91–110.
Wotruba, Thomas R. 1990a. Full-time vs. part-time salespeople: A comparison on job satisfaction,
performance, and turnover in direct selling. International Journal of Research in Marketing
7(2-3) 97–108.
Wotruba, Thomas R. 1990b. The relationship of job image, performance, and job satisfaction to
inactivity-proneness of direct salespeople. Journal of the Academy of Marketing Science 18(2)
113–121.
Wotruba, Thomas R, Pradeep K Tyagi. 1991. Met expectations and turnover in direct selling. The
Journal of Marketing 24–35.
Wotruba, Thomas R, Pradeep K Tyagi. 1992. Motivation to become a direct salesperson and its
relationship with work outcomes. Journal of Marketing Channels 2(2) 41–56.
Wu, Desheng Dash, Zijiang Yang, Liang Liang. 2006. Using dea-neural network approach to evaluate
branch efficiency of a large canadian bank. Expert systems with applications 31(1) 108–115.
Xardel, D. 1993. The Direct Selling Revolution. Business: Blackwell, Blackwell Business. URL
https://books.google.com/books?id=-1pkQgAACAAJ.
Xu, Ronghui (Lily). 2016. Lecture 7 time-dependent covariates in cox regression URL http:
//www.math.ucsd.edu/~rxu/math284/slect7.pdf.
Yang, Yupin, Avi Goldfarb. 2015. Banning controversial sponsors: Understanding equilibrium
outcomes when sports sponsorships are viewd as two-sided matches. Journal of Marketing
Research 28(6).
Yang, Yupin, Mengze Shi, Avi Goldfarb. 2009. Estimating the value of brand alliances in profes-
sional team sports. Marketing Science 28(6) 1095–1111.
Zamudio, César, Yu Wang, Ernan Haruvy. 2013. Human brands and mutual choices: an investi-
gation of the marketing assistant professor job market. Journal of the Academy of Marketing
Science 41(6) 722–736.
Zoltners, Andris A., PK Sinha, Sally E. Lorimer. 2013. How to make sense of sales force
turnover. Harvard Business Review ¡a href=”https://hbr.org/2013/06/how-to-make-sense-
of-sales-for”¿.
110