Service Quality, Inventory and Competition: An Empirical Analysis of Mobile Money Agents in Africa

The use of electronic money transfer through cellular networks ("mobile money") is rapidly increasing in the developing world. The resulting electronic currency ecosystem could improve the lives of the estimated 2 billion people who live on less than $2 a day by facilitating more secure, accessible, and reliable ways to store and transfer money than are currently available. The development of this ecosystem requires a network of agents to conduct cash-for-electronic value transactions and vice versa. This paper examines how service quality, competition, and poverty are related to demand and inventory (of electronic credit and physical cash) where, in this setting, service quality consists of pricing transparency and agent expertise. Among our results, we find that average demand increases with both pricing transparency and agent expertise, and that agent expertise interacts positively with competitive intensity. We also find that competition is associated with higher inventory holdings of both cash and electronic value, and that agents in high-poverty areas hold greater amounts of cash but do not carry a smaller amount of electronic value indicating that they devote more capital to their inventory. These results offer insight to mobile money operators with respect to monitoring, training, and the business case for their agents. This paper furthers our understanding of service quality, competition and inventory, while developing a foundation for the exploration of mobile money by OM scholars.


Introduction
In the past 8 years, "mobile money" platforms have experienced explosive growth in the developing world, now with 255 active mobile money systems in 89 countries (Groupe Speciale Mobile Association 2015). These platforms, primarily built and managed by mobile network operators, allow money to be stored in the form of digital currency (hereafter referred to as e-float). In much the same way that text messages can be sent quickly and cheaply, e-float can be securely and instantly transferred across long distances at a near-zero marginal transaction cost. Mobile money platforms are of particular interest to the base of the pyramid (BoP) community-scholars and practitioners developing business models deliberately geared toward serving the population in poverty-because they have potential to connect millions of poor and "unbanked" people to the formal financial system. This has potential to provide several benefits: i) it can enable quicker recovery from economic shocks such as job loss or illness to the primary wage-earner (Jack and Suri 2014); ii) it can is the relationship between the level of poverty in an agent's catchment area and agent inventory decisions?
To answer these questions, we use a combination of agent network and demographic data sources from Kenya and Uganda, two East African countries at different stages of mobile money market development. We use an in-person survey of over 3,000 mobile money agents that operate in the two countries. We then combine the locations of the surveyed agents with the precise locations of over 68,000 bank branches, bus stands, and mobile money agents, as well as population and poverty estimates for each square kilometer of the countries. We find that agents who are more transparent with transaction pricing and agents who are more knowledgeable experience significantly (both statistically and economically) greater demand. Agents' pricing transparency does not interact in Balasubramanian and Drake: Service Quality, Inventory and Competition among Mobile Money Agents Article submitted to Manufacturing & Service Operations Management; 3 a spastically significant way with competitive intensity. Agents that provide more knowledgeable service, on the other hand, seem to reap greater rewards from their performance in the face of greater competition. This can be related to Hill's (1993) operations strategy framework of "orderqualifiers" and "order-winners": pricing transparency acts as an "order-qualifier" among a portion of the consumer-base, while expertise acts as an "order-winner"-a dimension along which agents compete. We also find that agents who face more competitive intensity stock more inventory of both cash and e-float. Lastly, we find that agents in high poverty areas stock more cash, but do not stock a correspondingly smaller amount of e-float.

Literature review
Our work relates to literature focused on service quality, competition, inventory management, and operations at the base of the pyramid (BoP).
Service quality has attracted academic interest in the past three decades. While quality has many definitions and dimensions (Reeves and Bednar 1994), we focus on two widely-recognized dimensions of service quality: credibility and competence (Parasuraman and Zeithaml 1988). We measure credibility and competence, respectively, through each agent's pricing transparency and their expertise with respect to transaction policies and procedures. While (to the best of our knowledge) there are no empirical studies that directly explore the effect of pricing transparency and a provider's expertise on demand, the framework of trust developed in the literature is relevant.
A generally accepted conceptualization of trust is two-dimensional: trust is composed of "benevolence trust" and "competence trust" (Singh and Sirdeshmukh 2000). Benevolence trust (the faith customers have in firms not to cheat them) is related to our measure of credibility (i.e., pricing transparency) and competence trust (the faith customers have in firms to be able to competently fulfill demand) is related to agent expertise. Both of these concepts and their applications are discussed further in Section 4 where we develop our hypotheses.
In addition to literature exploring service quality, there has also been interest in the effect of competition on inventory. Olivares and Cachon (2009) argue that the theory on this relationship is mixed. On one hand, greater competitive intensity will drive down price, predicting lower optimal inventory holdings. On the other hand, greater competitive intensity will force firms to compete on service level, driving inventory holding up. In the context of the US auto market, Olivares and Cachon (2009) find that greater competitive intensity among dealers results in greater inventory holdings. Our work explores this relationship, but in the context of an emergent financial service in the developing world. Balasubramanian and Drake: Service Quality, Inventory and Competition among Mobile Money Agents 4 Article submitted to Manufacturing & Service Operations Management; Finally, because mobile money has potential to dramatically lower the cost structure of providing financial services to those living in poverty, mobile money is fundamentally related to the emerging literature on serving the BoP (Prahalad and Hammond 2002). Research in the operations management community focused on the BoP is quite nascent (e.g., Sodhi and Tang 2011;Gold et al. 2013). Our work here contributes to this emerging stream by examining the relationship between poverty in catchment areas and inventory decisions.
Our work differs from existing literature. First, our context is very different from most studies examining service quality, competition, and/or inventory holding. Rather than analyzing competition and inventory in traditional sectors in the developed world, such as the US auto market, we focus on a financial service in East Africa. Second, we are able to analyze how competition moderates the relationship between service quality and demand, rather than only looking at service quality or competition individually. Third, our unique combination of rich survey and spatial data allows us to examine the relationship between poverty and inventory decisions which, to the best of our knowledge, has not been studied in any context.

Context
In this section, we provide an overview of mobile money's history, impact, and implications. We also describe the mechanics of a mobile money transaction, as well as challenges and opportunities confronting mobile money systems.

Mobile money motivation and history
The poor comprise the vast majority of the "unbanked", the nearly 2.5 billion people globally who do not have an account at a formal financial institution (Demirguc-kunt 2012). The poor and unbanked-roughly one-third of the world's population-rely mostly on physical cash when transferring money. Thus, the velocity of money is limited by how fast cash can be physically transported, by foot or by bus in most circumstances (Batista and Vicente 2013). This limitation is a critical disadvantage to the poor when money is needed most, such as in the aftermath of a negative economic shock (e.g., sickness or job loss) or a rare opportunity to climb out of poverty through investment (e.g., fertilizer or improved seed purchases) (Helms 2006). At these decisive moments, friends and family willing and able to transfer money must rely on expensive and/or unreliable methods such as bus money transfer services (Morawczynski 2009). Furthermore, saving for these pivotal moments is more challenging with inferior savings tools; to store and save money, most either hide cash in their homes (at risk of theft and ineligible for interest), or purchase relatively illiquid assets like gold or livestock (that are often sold at a loss in times of need) (Collins  (Wright and Mutesasira 2001).
Similarly, informal credit for investment opportunities and insurance options for risk mitigation are substandard among the unbanked; credit is often only available from moneylenders at usurious rates, and formal insurance is generally inaccessible, if it exists at all (Collins et al. 2009). The poor, especially those living in rural areas, remain unserved by formal financial institutions because their low balances and transaction sizes yield little revenue for banks (Mas 2010). Furthermore, because the rural poor live in definitionally low density and remote areas, these regions lack the scale to make the provision of traditional financial services an attractive proposition. Consequently, financial institutions have largely found it impractical to profitably serve the poor in the developing world, particularly those residing in rural areas (Kendall 2011).
The rapid growth of cellular networks in the developing world in the past decade lays the groundwork for a potential paradigm shift in financial services for the poor. According to the Boston Consulting Group (BCG), the number of unbanked in the developing world with access to mobile phones was estimated to be 2 billion in 2011, and that number is likely to continue to grow (BCG 2011). Recognizing this opportunity in 2003, the UK Department for International Development approached Vodafone and its Kenyan affiliate, Safaricom, about developing and piloting a new service in Kenya that would allow micro-credit institutions to disburse loans and receive repayments electronically. After the pilot project, Safaricom noticed that many people were using the service to repay the loans of others (generally as a means of settling a secondary transactional obligation).
Safaricom quickly realized the potential of its service as a tool for domestic money transfer. They branded the product "M-Pesa" ("pesa" means "money" in Swahili) and launched the service in  Growth of M-Pesa customers and agents in Kenya, adapted from Jack and Suri (2014)

Transaction mechanics and inventory challenges
As M-Pesa's tagline, "send money home", would suggest, the use of mobile money to remit money to a family member or friend from an urban area to a rural area grew quickly. The following is a typical use case: An urban laborer in Nairobi, Kenya gets paid in cash. He conducts a cash-in transaction with an urban agent in which he gives the agent cash and the agent credits the laborer's mobile money account with e-float (Stage 1 in Figure 2). For her role in executing the transaction, the agent receives a cash-in commission from the mobile money operator (e.g., Safaricom). (Notably, customers do not pay for cash-in transactions in any currently active major mobile money network.) The laborer, now with a balance of mobile money, sends this e-float to his family outside of Kisumu, Kenya with his phone (not necessarily a smartphone), in much the same way he might send an SMS message (Stage 2 in Figure 2). The operator collects a fee from the laborer for executing this person-to-person (P2P) transfer. Having instantaneously received the e-float onto her phone, the laborer's wife goes to the local agent outside of Kisumu to conduct a cash-out transaction. She gives the agent e-float in exchange for cash (Stage 3 in Figure 2), paying the mobile money operator a cash-out fee. Like the cash-in agent, the cash-out agent is also compensated with a commission from the operator for her role in executing the transaction. 1 In order to conduct a cash-in or cash-out transaction, the agent must have inventory of e-float or cash, respectively. Unfortunately, stockouts are an acute problem in mobile money networks. In Schematic of a typical urban-to-rural person-to-person transfer, adapted from Agrawal (2009) rural areas, agents often run out of cash. Our data show that even in urban settings, service levels can fall well below 80% for cash-out transactions. Agents also run out of e-float, but to a lesser degree (Intermedia 2013). Because stockouts of cash and e-float make it harder for customers to easily convert between the two forms of money, they degrade consumer confidence in the convenient convertibility of e-float.
When an agent stocks out of cash, the customer desiring to cash-out has two options: she can return to the same agent later with the hope that the agent has replenished his inventory of cash; or, she can travel to a different agent within the same operator network to cash-out (assuming that this second agent has cash available). When the customer desiring to cash-in experiences a stockout of e-float, the customer has the above two options, as well as a third option: to travel to an agent serving a different operator, assuming that the person he is sending money to also has an account with the competing operator. Note that because e-float is actual currency, it cannot be "created" on the spot by either the agent or the operator. Each unit of e-float an operator issues must be backed by traditional deposits at a prudentially regulated financial institution. Though moving e-float once it has been issued is clearly easier than moving cash, agents can still stock out of e-float if they have not been able to procure enough e-float to satisfy demand.

Hypothesis development
Here we develop hypotheses related to service quality-in terms of pricing transparency and expertise-and competitive intensity, including its interaction with service quality. We also describe hypotheses related to inventory management with respect to competitive intensity and poverty.

Service quality and demand
We posit that agent pricing transparency and agent expertise, two dimensions of service quality, are important influencers of demand among mobile money agents. In this case, both of these quality elements can be examined through a customer trust lens. As we mentioned previously, a generally accepted conceptualization of trust is two-dimensional: trust is composed of "benevolence trust" and "competence trust" (Singh and Sirdeshmukh 2000).
Benevolence trust is commonly defined as the "perceived willingness of the trustee to behave in a way that benefits the interests of both parties, with a genuine concern for the partner, even at the expense of profit" (Garbarino and Lee 2003). Posting of CICO pricing is both mandated by operators and expected by consumers. The absence of pricing transparency may serve as a warning sign to customers and therefore may degrade benevolence trust. This relates to emerging "disclosure" literature. The key difference is that, in the disclosure literature, the established norm is non-transparency-organizations have the decision whether or not to reveal traditionally unobservable information such as their environmental performance (e.g., Toffel and Reid 2009;Kalkanci and Plambeck 2012), corporate social responsibility (e.g., Dhaliwal et al. 2011;Gamerschlag et al. 2010), or operational processes (e.g., Buell and Norton 2011). In our setting, however, the established norm with respect to pricing is one of transparency-operators mandate that the tariff be posted, and the majority (over 90%) of agents comply with this mandate. We posit that violating the established transparency norm erodes benevolence trust and that non-compliance (i.e., non-transparency) will therefore be related to attenuated demand.
The second component of trust, competence trust, is generally defined as the perceived ability of the firm to deliver services reliably and without flaws (Garbarino and Lee 2003). A customer has competence trust in an agent if the customer believes that the agent has the knowledge and ability to properly conduct CICO transactions (e.g., the agent knows the correct daily transaction limits, identification requirements, and other operator policies regarding the use of mobile money).
Though expertise cannot be observed as easily as pricing transparency (presence of a posted tariff sheet), agent expertise (or lack thereof) can be assessed by customers if guidance from agents is either confirmed or discovered to be incorrect. Perceptions of expertise can also be shaped by the opinions and experiences of those in the customer's social network. Greater expertise would thus naturally lead to greater competence trust. Sun and Lin (2010) find that department store benevolence trust and competence trust both increase customer loyalty (measured on a scale that includes future repeat purchase intent). Because 9 transparency likely engenders benevolence trust in agents, and expertise likely engenders competence trust in agents, we posit that customers reward agents for pricing transparency and expertise, respectively, with higher demand.
Hypothesis 1. A) Customers reward agents for pricing transparency; demand increases with pricing transparency. B) Customers reward agents for expertise; demand increases with agent expertise.

Service quality, competition, and demand
Intuition suggests that in most settings competitive intensity would attenuate demand, as consumers have a choice between a firm and its competitor(s). Indeed, in their study of auto dealerships, Olivares and Cachon (2009) label this the "sales effect": increased competition decreases a dealer's sales. We posit that this effect is present in the mobile money context as well: increasing competitive intensity (increasing the number of proximate competitors) decreases each agent's demand.
Furthermore, we posit that agents who engender benevolence trust by being more transparent will be able to attract demand away from agents who do not engender such trust. We therefore hypothesize that the interaction between competitive intensity and pricing transparency is positive (i.e., the rewards for transparency increase with competitive intensity). Similarly, we posit that agents who engender greater competence trust by demonstrating expertise will be able to attract demand away from agents who do not. As with pricing transparency, we hypothesize that the interaction between competitive intensity and expertise is positive (i.e., the rewards for expertise increase with competitive intensity).

Hypothesis 2. A) Agent demand decreases in competitive intensity. B) Competitive intensity
increases rewards from pricing transparency. C) Competitive intensity increases the rewards from expertise.

Competition and inventory levels
Olivares and Cachon (2009) decompose the relation between competition and inventory into a "sales-effect" and a "service-level" effect. The "sales effect" refers to the notion that proximate competition depresses demand and sales (analogous to our hypothesis H2A) which in turn reduces desired inventory holdings. What they term the "service-level" effect refers to how competition affects service-level (the proportion of demand that is fulfilled with available inventory). They argue that the theory on this question is mixed. On one hand, greater competitive intensity will Hypothesis 3. Agent inventory holdings of cash and e-float increase with competitive intensity.

Poverty and inventory levels
The relationship between the level of poverty in an agent's catchment area and inventory holdings of cash and e-float is complicated. On one hand, the poor have definitionally lower net worth and cash-flow, which might suggest that their transaction sizes are smaller than the average mobile money customer. Controlling for the number of transactions, then, smaller transaction sizes would likely result in a decreased need for inventory of cash and e-float. However, given that the initial "killer app" of mobile money was domestic remittance ("send money home"), most often from areas of comparative wealth (largely urban areas) to areas of relative poverty (rural areas, in many cases), we might expect that there would be significantly more cash-out demand than cash-in demand in high-poverty areas. Indeed, Eijkman et al. (2009) observe that for five representative agents in rural areas in 2009, total cash-out transaction value was much higher than total cash-in transaction value. This imbalanced demand suggests that agents in high-poverty 2 areas may need to carry more cash inventory, because they do not have as much cash-in transaction demand that 11 would otherwise contribute to their inventory of cash. We thus posit that controlling for other factors, agents serving high-poverty areas carry more cash than agents serving lower-poverty areas.
Theory is mixed on the relationship between e-float inventory and poverty levels. On one hand, agents serving high-poverty areas sometimes experience high-magnitude demand for e-float (cashin). 3 These large cash-in demand arrivals would necessitate large e-float holdings to satisfy this demand. These arrivals may be attributable to the surprisingly complex financial lives of the poor.
Collins et al. (2009) note that many poor people have income that does not come in a steady stream but is "lumpy" (e.g. farmers have no income until they are able to harvest and sell their crops). This "lumpiness" necessitates savings in some form to smooth consumption over periods of no income. Economides (2015) finds evidence that Tanzanian mobile money accounts are used as a secure savings alternative to keeping cash at home. This could possibly explain some high-value cash-in behavior in high-poverty areas, and thus drive up the need for greater e-float holdings by agents. On the other hand, however, Intermedia (2013) notes that poor households are less likely to own a phone, a SIM card, and use mobile money, suggesting that the poor lag in adoption of mobile technology. It is likely, then, that adoption of e-float for usage as a means for purposes other than domestic remittance, such as utility payments, tax payments, and depositing into bank accounts is lower among the poor than the general population. This decreased demand would lead to agents in poor areas carrying less e-float. Furthermore, because agents in high-poverty areas likely have very limited budgets to invest in inventory of cash and e-float, and because we posit that agents in high-poverty areas carry relatively more cash, agents might be forced to carry less e-float. The notion that cash inventory might take priority over e-float inventory is supported by the fact that cash-out commissions are generally larger than cash-in commissions, generally by 50% or more.
Because agent budget constraints are likely to be very salient in this context, we posit that agents in high-poverty areas carry less e-float inventory than their peers who serve lower-poverty areas.
Hypothesis 4. Agent inventory holdings of cash increase with the level of poverty in the catchment area, and inventory holdings of e-float decrease with the rate of poverty in the catchment area.

Data and empirical specification
To test these hypotheses, we combine three data sources, each sampling Kenya and Uganda. First, we use data from an in-person, cross-sectional survey of over 3,000 mobile money agents. We combine this data with the precise locations of over 68,000 financial access and transportation points (including mobile money agents, banks, and bus stands) in the two countries. Finally, we integrate granular (per square kilometer) spatial estimates of population and poverty (with the latter defined as the number of people living on less than $2 per day) in each given square kilometer.

Agent network survey
We use data from large surveys conducted by the Helix Institute of Digital Finance, an organization that provides training and data to digital financial service providers in the developing world.
We Finally, in each district, a local team lead identified a representative enumeration area (EA). Once the EA was identified, surveyors were given mutually exclusive routes to walk, applying a "lefthand-rule" walk pattern, skipping a pre-determined number of agents on the left-hand side before attempting to interview the next agent. 4 This pre-determined number was based on the number of agents the given operator had in the district, which was derived from the spatial census data described below. Each survey lasted between 30-60 minutes. As with many in-person surveys, there was a very high response rate (roughly 95%). All data were point estimates of steady-state levels and were self-reported by the agents or observed directly by the surveyor. The interview covered a wide array of topics, including demographics and location (latitude and longitude), products and services offered, inventory management, revenue and commission structure, platform performance, training, monitoring, and support. From these data we glean independent variables P ricingT ransparency, and Expertise, as well as controls M ale, T ills, Sunday, Dedicated, and OperatorG. Each of these variables is described in detail in the econometric specification subsec-

High-resolution spatial demographics
Lastly, we draw upon spatial population and income data generated by WorldPop, an organization focused on creating high-quality maps for the humanitarian sector. In order to create these maps, WorldPop combines three sources of data: satellite imagery (Radarsat-1 country mosaics and Landsat Enhanced Thematic Maps); the Africover database containing geographic data on roads, land cover, and bodies of water; and country-level census data. The resulting integrated model generates precise population estimates for every square kilometer of Africa (Tatem et al. 2007). WorldPop used a similar method to develop a high-resolution spatial data layer of the population in povertythose living on less than $2 per day. To generate its poverty spatial data, WorldPop employs a process known as "Bayesian geostatistics" to integrate geocoded well-being surveys conducted by USAID (Demographic and Health Survey) and the World Bank (Living Standards Measurement Survey). The resulting data layer, like the general WorldPop data layer, has a resolution of 1 square kilometer (Tatem 2013). Figure 3 depicts a spatial data layer of the poor population in Kenya's Nyanza Province. The green circles depict buffers of 5, 10, and 15 kilometers (for ease of viewing) around bank branches. By combining the surveyed agent locations with the spatial demographic data, we generate control variable P opK1km (population in thousands within a 1km radius from the agent) and independent variable P ovRatio1km (the percent of the population within 1km who live on less than $2 a day) through buffer analysis.

Econometric specifications
We use two different OLS regressions to test our hypotheses: one to test our hypotheses related to demand, and the other to test our hypotheses related to inventory holding. due to stockouts (S), and the number of transactions denied due to system failure (F ). We define demand as D = T + S + F . To test our hypotheses related to demand (Hypotheses 1 and 2), we conduct an OLS regression on the natural logarithm of demand (log(Demand)).
Also from the survey data, we have agent level point estimates for each agent's cash and e-float inventory, in Ugandan Shillings and Kenyan Shillings, respectively. For comparability, we convert these estimates to United States Dollars (USD) using year-end 2013 exchange rates. To test our hypotheses related to inventory (Hypotheses 3 and 4), we conduct an OLS regression on the natural logarithm of cash inventory (log(AvgCashU SD)) and e-float inventory (log(AvgEF loatU SD)),

respectively.
We log demand and inventory for two principle reasons. First, demand and inventory cannot be negative in this context (to be deemed active, all agents must have at least 1 transaction per day, and agents generally cannot "borrow" cash or e-float). Secondly, using the log of demand and inventory, respectively simplifies the interpretation of estimated coefficients: a 1 unit change in an independent variable corresponds to a β*100% change in demand and inventory, respectively. to the question (greater expertise), and a 0 indicating that they did not respond correctly to the question (lesser expertise). DirectCompetition, our measure for competitive intensity, indicates the number of other agents primarily serving the same operator as a particular agent within 1 kilometer of that agent. To estimate the impact of competition on service quality, we include interactions between DirectCompetition, P ricingT ransparency, and Expertise.

Independent and interaction variables
For the second models examining inventory (one for cash holdings and the other for e-float holdings), we again use DirectCompetition as an independent variable. We also use P ovRatio1km as an independent variable, where P ovRatio1km is the ratio of the total population estimate within 1km of the agent who live on less than $2 a day. greater than 0 to indicate the number of agent tills (i.e., "virtual cash registers") that the agent operates. P opK1km is the population (in thousands) within a 1 kilometer radius of an agent.
IndirectCompetition is the number of agents primarily serving other operators in a 1 kilometer radius of an agent. Bank1km is the number of bank branches in a 1 kilometer radius of an agent; banks are a major resource in helping agents rebalance inventory (so agents can get cash and/or e-float). Bus1km is the number of bus stops in a 1 kilometer radius of an agent; this is a proxy for the difficulty of transit in the vicinity of the agent. U ganda is a binary variable that takes a value of 1 if the surveyed agent operates in Uganda, and a 0 if the agent operates in Kenya. Finally, OperatorG represents a set of 5 (+1) indicator variables for each operator/brand represented in the sample, which captures differences across operators not accounted for by our independent variables or the control variables above.
The second models, examining inventory holdings of cash and e-float, control for the average number of successful transactions T that an agent conducts in a day, as well as many of the control variables used in the first model that also could plausibly affect inventory holdings: P opK1km, IndirectCompetition, Bank1km, Busk1km, Dedicated, OperatorG, and U ganda.

Results
Summary statistics are presented in the appendix. Table A1 includes means, standard deviations, and differences in means. Just under half of the agents in the sample run a dedicated mobile money business. We also note that Kenyan agents in the sample locate on average in more densely populated areas than Ugandan agents, while Ugandan agents are much more likely to be located in areas of high poverty concentration than Kenyan agents. Finally, we see that Kenyan agents in the sample on average have at least twice as many bank branches and bus-stands within a 1 kilometer radius than Ugandan agents. This is reflective of the fact that Kenya is more economically developed than Uganda, and thus its financial and transportation infrastructure is more developed than its neighbor.
Examining our hypotheses relating to transaction demand, we present the results from four OLS regressions in Table A3. We include the set of control indicator variables representing operator brands (OperatorG) in regressions but exclude them from the tables to preserve operator confidentiality. The first model excludes interactions with DirectCompetition. The next two models each include a single interaction term, and the final model includes both interaction terms. 5 The discussion that follows is based on the full model unless otherwise noted. Robust standard errors are reported, as the Breusch-Pagan test indicated the presence of heteroskedasticity.
Examining our hypotheses relating to inventory holdings of cash and e-float, the results from two OLS regressions are presented in Table A4. Again, robust standard errors are reported, as the Breusch-Pagan test indicated the presence of heteroskedasticity.

Service quality, competition, and demand
Our first set of hypotheses focused on relationships between service quality (pricing transparency and expertise) and demand. The data support hypothesis 1A. Pricing transparency (nontransparency) is associated with an increase (decrease) in the agent's demand. The data also support hypothesis 1B. Agent expertise is associated with greater agent demand. The magnitudes 18 Article submitted to Manufacturing & Service Operations Management; manuscript no.  of these relationships are also notable; the presence of a tariff sheet is associated with an 12% increase in demand and the ability to answer a difficult question about mobile money policy is associated with a 8% increase in demand.
Regarding competitive intensity, we note that the coefficient on DirectCompetition is negative and statistically significant. The data support hypothesis 2A, that demand decreases in competitive intensity. The data indicate that competitive intensity also interacts with pricing transparency but not with expertise as the coefficient on DirectCompetition x P ricingT ransparency is not statistically significant. Hypothesis 2B is therefore not supported. Finally, the coefficient on DirectCompetition x Expertise is significant and positive, supporting hypothesis 2C: competitive intensity enhances the relationship between expertise and demand. These latter two relationships are illustrated in Figures 4a and 4b. In Figure 4a, we see the estimated log(Demand) values as P ricingT ransparency and DirectCompetition are varied. The main relationship between pricing transparency and demand, the gap between the two lines, is clearly visible, but there is not a statistically significant difference in the slopes of the two lines (which would be evidence of an interaction between competitive intensity and pricing transparency). Figure 4b illustrates estimates of logDemand as DirectCompetition and Expertise are varied. Analogously to the relationship between pricing transparency and demand, we clearly see the direct relationship between expertise and demand in the gap between the two lines. However in Figure 4b a statistically significant difference in the slopes of the two lines is also visible, evidence of an interaction between competitive intensity and expertise. Plots of predicted demand as a function of competitive intensity and service quality measures

Competition, poverty, and inventory levels
We now turn to our hypotheses relating to inventory holding. We note that the coefficients on DirectCompetition in both cash and e-float inventory models are statistically significant. Hypothesis 3 is thus supported: agent inventory holdings of both cash and e-float increase with competition.
With respect to poverty level in agent catchment areas, we note that the coefficient on P ovRatio1km is positive and significant in the model for cash inventory, while the coefficient of P ovRatio1km in the e-float inventory model is not significant. We thus find partial support for hypothesis 4: agent inventory holdings of cash increase with the poverty rate in the catchment area.
However, we do not observe evidence of a negative relationship between e-float inventory holding and poverty.

Robustness checks
To test whether our results are robust to the inclusion of observations with missing values, we use multiple imputation (predictive mean matching) to predict missing values of specific variables within observations. We then estimate log(demand), log(CashU SD), and log(EF loatU SD) respectively, using these augmented sets of survey observations. Results for these tests are presented in Tables A5 and A6 in the Appendix. While estimated coefficient values vary slightly relative to the base models, the results presented above largely hold (both directionally and in significance). The only exception is our finding that cash inventory holding increases with competitive intensity became slightly weaker in significance level in the multiple imputation model, though the relationship is still significant at the p < 0.1 level.

Discussion
Our results provide insight on the relationships between service quality, competition, inventory, and demand in a new and important operations context, particularly relevant to the base of the pyramid community.
This paper documents positive relationships between pricing transparency and demand as well as expertise and demand. In the mobile money context, agents who are transparent with their prices and agents who are more knowledgeable about mobile money policies experience greater demand.
We also add empirical evidence to the notion that demand decreases in competitive intensity, finding proximate competition to be associated with decreased demand in the mobile money context. We also find that expertise interacts with competition: the rewards for expertise increase in the face of greater competitive intensity. However, our data suggest that pricing transparency does not interact in a statistically significant way with competitive intensity. This suggests that pricing transparency may not be a dimension of competition in mobile money; rather, some fraction of the (potential) customer base simply may choose not do business with non-transparent agents, whether or not there are other competitors around. Using the operations strategy terminology coined by Hill (1993), pricing transparency in this context acts as an "order qualifier" for a segment of the market-these customers may not consider agents who are not transparent with prices. Our data suggest that expertise, on the other hand, acts as an "order winner" for a segment of customers-all else equal, these customers may choose more expert agents among their set of "qualified" agents.
This paper also documents positive relationships between competitive intensity and inventory of both cash and e-float in the mobile money context. This observation supports the finding of Olivares and Cachon (2009), who show that inventory of US auto-dealers increase in proximate competition. They postulate that there is a"service-level" effect -firms compete on service-level to attract business away from competitors. Our results provide support for such a "service-level effect" in the mobile money context as well, with agents carrying more inventory to increase their service level as another "order-winner". We also find that agents' cash inventory increases along with catchment area poverty level. While there may be many reasons for increased cash-holding, the most likely in this context is that remittances from lower-poverty areas (mostly urban) to higher-poverty areas (mostly rural) is a key driver of mobile money demand. Given that agents have limited budgets and that cash inventory increases with poverty level, we might expect that agents in high poverty areas would carry less e-float. We do not find support for this notion, however. This is possibly because agents in high-poverty areas might experience significant demand (at least at certain times) for e-float that is not always off-set by cash-out demand (which would generate e-float inventory). The consequence of this finding is important to agents, operators, and policymakers interested in the poor: because agents in high-poverty areas carry more cash and generally the same level of e-float, controlling for other factors (such as number of successful transactions) they have a less compelling business case. There may be some efficiency gains to be realized with better agent inventory control policies -this is the subject of future work by the authors.
Given that the health of the agent network is very important to the success of mobile money, mobile money operators must ensure that the business case for agents is compelling. Thus, this work can be useful in planning agent network expansion, as it yields predictions for demand and agents' inventory levels in a particular location. Operators can use these predictions to inform the optimal placement and number of agents in a defined area-enough to capture demand (and potentially to inspire higher service levels), but not so many to jeopardize agents' commercial viability. Our analyses also inform guidance and/or policies operators should put in place, such This analysis, like other analyses using survey data, suffers from potential bias: demand and inventory estimates were self-reported estimates by the agent. To the extent that agents thought their operators might see the results (even though they were promised confidentiality by the third party research firm, and this confidentiality was honored), agents' estimates of demand and inventory might be biased. However, there is no evidence to indicate that some agents might be more biased than others. There is also a limitation in regards to our spatial variables. First, the catchment areas are defined in terms of "as-the-crow-flies" distance, rather than true travel time. Secondly, the service quality of nearby competing agents very likely affects the relationship between a particular agent's service quality and her demand. Unfortunately, our dataset does not capture a quality metric such as the one that may have proven useful in this case. Finally, our dataset limited our measure of demand to an estimate of the raw count of transactions. An interesting additional dimension of analysis would consider the value of these transactions as well.

Conclusions
Mobile money is a rapidly growing industry that has potential to dramatically improve the lives of the poor in many ways. We begin to explore this industry by examining mobile money demand drivers. We find that agents who are transparent with transaction pricing experience relatively greater demand. Agents who are relatively more knowledgeable not only experience relatively greater demand, but also seem to reap greater benefit from their expertise in the face of competition.
We also explore the relationships between competition, poverty, and inventory holdings of cash and electronic value. We find that agents who face more competitive intensity carry more inventory of both cash and electronic value, while agents who serve high poverty areas hold more cash but not a smaller amount of electronic value.      Standard errors in parentheses + p < 0.10, * p < 0.05, * * p < 0.01 Standard errors in parentheses + p < 0.10, * p < 0.05, * * p < 0.01 Standard errors in parentheses + p < 0.10, * p < 0.05, * * p < 0.01