pricing services subject to congestion: charge per-use fees or sell subscriptions by gerard cachon...

8/3/2019 Pricing Services Subject to Congestion: Charge Per-Use Fees or Sell Subscriptions by Gerard Cachon and Pnena Fel

1/17

MANUFACTURING & SERVICE

OPERATIONS MANAGEMENT

Vol. 13, No. 2, Spring 2011, pp. 244260issn 1523-4614 eissn 1526-5498 11 1302 0244

informs

doi 10.1287/msom.1100.0315 2011 INFORMS

Pricing Services Subject to Congestion:

Charge Per-Use Fees or Sell Subscriptions?Grard P. Cachon

Operations and Information Management, The Wharton School, University of Pennsylvania,Philadelphia, Pennsylvania 19104, [email protected]

Pnina FeldmanOperations and Information Technology Management, Haas School of Business, University of California, Berkeley,

Berkeley, California 94720, [email protected]

Should a firm charge on a per-use basis or sell subscriptions when its service experiences congestion?Queueing-based models of pricing primarily focus on charging a fee per use for the service, in part becauseper-use pricing enables the firm to regulate congestionraising the per-use price naturally reduces how fre-quently customers use a service. The firm has less control over usage with subscription pricing (by definition,with subscription pricing customers are not charged proportional to their actual usage), and this is a disadvan-tage when customers dislike congestion. However, we show that subscription pricing is more effective at earningrevenue. Consequently, the firm may be better off with subscription pricing, even, surprisingly, when congestionis intuitively most problematic for the firm: e.g., as congestion becomes more disliked by consumers. We showthat the absolute advantage of subscription pricing relative to per-use pricing can be substantial, whereas thepotential advantage of per-use pricing is generally modest. Subscription pricing becomes relatively more attrac-tive if consumers become more heterogeneous in their service rates (e.g., some know they are heavy usersand others know they are light users) as long as capacity is fixed, the potential utilization is high, and the twosegments have substantially different usage rates. Otherwise, heterogeneity in usage rates makes subscriptionpricing less attractive relative to per-use pricing. We conclude that subscription pricing can be effective even ifcongestion is relevant for the overall quality of a service.

Key words : service operations; operations strategy; pricing and revenue management; game theory; queueingtheory

History : Received: June 18, 2008; accepted: July 28, 2010. Published online in Articles in Advance January 21, 2011.

How should a firm price its service when con-gestion is an unavoidable reality? Customers dislikecongestion, so a firm has an incentive to ensure itprovides reasonably fast service. At the same time, thefirm needs to earn an economic profit, so the firmspricing scheme must generate a sufficient amount ofrevenue. Furthermore, these issues are closely linked:the chosen pricing scheme influences how frequentlycustomers use a service, which dictates the level ofcongestion; congestion correlates with the customers

perceived value for the service, and that determinesthe amount of revenue the firm can generate.A natural option is to charge customers a per-use

fee or toll: customers pay a per-transaction fee eachtime they withdraw money from the ATM; beautyshops and hair salons price on a per-use basis; andcar maintenance companies, such as Pep Boys Auto,charge each time a service is completed. Naor (1969)

began this line of research, and there have been manysubsequent extensions of his basic model, but nearlyalways with a focus on per-use fees. (See Hassin andHaviv 2003 for a broad survey of this literature.)

Although the emphasis in the queueing literaturehas been placed on per-use pricing, other pricingschemes are observed in practice. Most notably, somefirms sell subscriptions for the use of their service:a health club may charge an annual membershipthat allows a customer to use the facility withoutadditional charge for each visit; AOL, an Internetservice provider, initially charged customers per-useaccess fees but later switched to subscription pric-ing (a monthly access fee with no usage limitation);

Netflix, a retailer that provides movie DVDs forrental, also uses subscription pricing (a monthly feefor an unlimited number of rentals); Disney chargesan entry fee for its theme park without charging perride on the attractions; et cetera.

Despite the existence of subscriptions in practice,a subscription pricing strategy has a clear limitationin the presence of congestion effects: subscribers arenot charged per use, so it is intuitive that they seekservice more frequently (e.g., use the health club toooften), thereby increasing congestion and decreasing

244

Additionalinformation,includingrights

andpermissionpolicies,isavailableathttp://journals.informs.org/.


2/17

Cachon and Feldman: Pricing Services Subject to Congestion: Charge Per-Use Fees or Sell Subscriptions?Manufacturing & Service Operations Management 13(2), pp. 244260, 2011 INFORMS 245

the value all subscribers receive from the service. Asa result, in a setting with clear congestion costs (e.g.,in a queueing model) one might assume that sub-scription pricing would be inferior to per-use pricing.However, in this paper we demonstrate that subscrip-tion pricing may indeed be a firms better pricing

strategy despite its limitations with respect to conges-tion. We do so in two different capacity managementscenarios: (i) the firms service capacity is exoge-nously fixed; and (ii) the firm endogenously choosesits service capacity in addition to its pricing policy.

The next section reviews the extensive literature onpricing services, with an emphasis on models thataddress the issue of congestion. Section 2 details our

base model. Sections 3 and 4 compare the two pricingschemes under two different assumptions for how thefirms capacity is determined. Section 5 extends the

base model to consider the effects of heterogeneoususage rates. Section 6 summarizes our conclusions.

1. Related LiteratureOur work is primarily related to three streams of liter-ature: pricing in queueing models, the theory of clubs,and advance purchase pricing.

Queueing theory provides a natural framework formodeling congestion, and we adopt that frameworkas well. However, as already mentioned, the litera-ture on pricing of queues generally only considersper-use pricing (e.g., Littlechild 1974, Edelson andHilderbrand 1975, De Vany 1976, Mendelson 1985,Chen and Frank 2004). Per-use pricing is sufficient

for maximizing social welfare, but it is known that aprofit-maximizing firm does not choose the welfare-maximizing price (e.g., Naor 1969).

Randhawa and Kumar (2008) and Bitran et al.(2008) do consider additional pricing schemes inqueueing models. Randhawa and Kumar (2008) com-pare per-use pricing with subscription pricing thatimposes limits on usage, e.g., Netflix has a plan inwhich a customer can view as many movies as theywant as long as they do not possess more than fourDVDs at a time. They show that this constrainedsubscription plan may be better for the firm thanthe unconstrained per-use pricing because it reduces

the volatility of the demand process the firm experi-ences. We do not consider subscription pricing withlimitations, i.e., in our model a subscription pricingplan allows for unlimited usage. Furthermore, in theirmodel the two plans have the same revenue potential,whereas in our model a key difference is that sub-scription pricing can have a higher revenue potentialthan per-use pricing. Hence, the restriction on usagewith their subscription plan is necessary to create adistinction between the two pricing schemes. Bitranet al. (2008) study a two-part tariff that combines

both per-use and subscription pricing. Their focus isdifferent than ours: they do not compare per-use tosubscription pricing and instead emphasize how con-sumer uncertainty regarding service quality affectsthe dynamics of their system over time (in our modelconsumers have rational expectations, so we do not

explicitly model the learning process).There is a literature in economics on the pricing ofshared facilities (i.e., clubs) subject to congestion, suchas swimming pools and golf clubs (e.g., Berglas 1976,Scotchmer 1985). Just as in our model, customers pre-fer that the service/facility is used by fewer people,so that there is less congestion. These papers showthat a two-part tariff is optimal for the firm: a per-use fee is chosen to induce a usage level that maxi-mizes social welfare and a subscription fee is chargedto transfer all rents from customers to the firm. Theliterature on nonlinear pricing also studies the designof two-part tariffs for congestion-proned services (e.g.,

Clay et al. 1992, Miravete 1996, Masuda and Whang2006). They show that when consumers have hetero-geneous needs for the service, a menu of two-parttariffs may be optimal. Like Bitran et al. (2008), thesepapers do not compare per-use pricing to subscrip-tion pricing. Strictly speaking, according to our modelthe firm always prefers the two-part tariff over eithersubscription or per-use pricing (each is a subset of theset of two-part tariffs). However, we believe a com-parison between subscription and per-use pricing iswarranted. The queueing literature focuses on per-usepricing, and both per-use pricing and subscriptionsare observed in practice. In addition, a two-part tar-

iff may not be desirable for reasons that we do notmodel (nor are generally modeled); e.g., a consumermay dislike being charged twice for the same service,especially if they do not understand the motivationfor such a pricing scheme. Furthermore, firms mightprefer to forgo the additional revenues from using atwo-part tariff to save the transaction costs and theadministrative burden required for its implementa-tion. Disney, for example, initially charged consumers

both to get into the park and for specific rides withinthe park, but later it abandoned per-ride charges.

Barro and Romer (1987) demonstrate that per-usepricing can be equivalent to subscription pricing. Forexample, they argue that a ski slope could generatethe same revenue by charging a fee per ride or bycharging a daily lift-ticket price (which is analogousto a one-day subscription). However, in their modelthey assume that the number of ski-lift rides is fixedand fully utilized no matter which pricing scheme isused. Hence, a daily lift-ticket price can be chosensuch that usage is the same as with a per-ride price. Incontrast, in our model consumers regulate their usagedepending on the pricing schemesubscription pric-ing leads consumers to use the facility more than any




3/17

Cachon and Feldman: Pricing Services Subject to Congestion: Charge Per-Use Fees or Sell Subscriptions?246 Manufacturing & Service Operations Management 13(2), pp. 244260, 2011 INFORMS

positive per-use pricing schemeresulting in differ-ent utilizations of the server. Hence, in our model thetwo schemes are not equivalent.

Our subscription pricing scheme resembles ad-vance-purchase pricing (e.g., DeGraba 1995, Xie andShugan 2001). When consumers purchase in advance

of the service, such as buying a concert ticket weeks before the event, consumers are willing to pay theirexpected value for the service. In contrast, when con-sumers spot purchase, i.e., when they know theirvalue for the service, they are naturally willing topay only their realized value. When purchasing inadvance, consumers are more homogeneous relativeto the spot market, so the firm can earn more rev-enue by selling in advance than by selling just witha spot price: it can be better to sell in advance toevery customer at their expected value than to sell inthe spot market to a portion of consumers (i.e., thoseconsumers with a high realized value). In our model,subscriptions also have this ability to extract rents

because consumers are more homogeneous whenthey purchase subscriptions than when they purchaseon a per-use basis. However, we consider the impactof congestion, whereas the advance-purchase modelsdo not (i.e., consumers in those models do not regu-late their usage based on the pricing policy).

2. Model DescriptionA single firm provides a service to a market with apotential number of M homogeneous customers. Con-sumers are assumed to be infinitesimali.e., everyconsumer is small relative to the size of the mar-ket. Each customer finds the service to be valuableon multiple occasions, or service opportunities. Forexample, a customer may wish to occasionally use ateller at her bank, use the Internet repeatedly, or renta movie at least a couple of times per month. Thisstream of service opportunities occurs for each cus-tomer at rate . At the moment a service opportunityoccurs, a customer observes the value, or utility, V,she would receive if she were to receive the serviceto satisfy that opportunity. Service values for each

customer are independent and identically distributedacross opportunities, where the support of V is theinterval 0 v. Hence, we have a single market seg-ment of consumers, so differences between per-useand subscription pricing are not driven by a desireto price discriminate between segments, in contrast toEssegaier et al. (2002). We discuss multiple customersegments in 5.

Although customers value receiving the service,they prefer as fast a service process as possibleeachcustomer incurs a cost w per unit of time to complete

service (time waiting and in service).1 Finally, con-sumers neither receive utility nor incur disutilitywhen not in the service process and waiting for thenext service opportunity to arise.

The firm offers one of two pricing schemes: a per-use fee or a subscription price. The per-use fee, p, is a

charge for each service completion: e.g., a fee for with-drawing money from an automatic teller machine, afee for each visit to a health club, or a per-minute feefor accessing a database. A subscription price, k, isa fee per unit of time, which is independent of theamount of service the customer receives. (This defini-tion of a subscription is equivalent to a fixed fee, K,for a finite duration, d, with unlimited usage duringthat time, where k = K/d.) Where useful, we use pand s subscripts to signify notation associated withthe per-use and subscription schemes, respectively.

The servers processing rate is . In 3 we assume is exogenous, whereas in 4 the firm chooses sub-

ject to a fee that is proportional to the service rate.W is the expected service time. We use the termservice time to refer to the total time to complete theservice, i.e., it includes time waiting and in service.We assume that WM is sufficiently small relative to1/, where = M is the maximum possible arrivalrate of service opportunities (i.e., the arrival rate whenevery customer seeks service at every service oppor-tunity). This implies that a customers informationabout the queue length during one service occasionis of little use in predicting the waiting time for thenext service encounter. It also implies that for a fixedpotential arrival rate of service, , the potential pop-

ulation of customers, M, is large, they do not seekservice too frequently ( is small), and capacity issufficiently large that WM 1/. (For example,the interarrival time of services could be measuredin days, whereas service times could be measured inminutes.) Consequently, the arrival rate to the firmsqueue does not vary (approximately) with the queuelength (which is typically assumed in the queueingliterature), and there is little chance that a serviceopportunity arises while a customer is in the ser-vice process. For example, a customer does not receiveanother need to withdraw cash from an automaticteller machine while she is in the process of with-drawing cash. (See Randhawa and Kumar 2008 fora model of a closed queueing system in which thearrival rate to the queue depends on the number ofcustomers in queue.) Therefore, the expected servicetime depends only on the actual arrival rate, . Thefunction W is strictly increasing (W > 0) andconvex (W 0). (Thus, W 0 = 1/ because 1/

1 As in Afche and Mendelson (2004), it is possible to allow thewaiting cost to be linear in the value of the service, w = a + bv.A detailed analysis is available from the authors.




4/17


is a customers service time when there is no conges-tion.) Naturally, W is decreasing in .

When a service opportunity occurs, a customerdecides whether or not to seek service (i.e., jointhe firms service system). The decision is based onthree factors: the value of the service opportunity,

the cost associated with the expected time to com-plete the service transaction, and the firms pricingpolicy. Although the customer observes the value fora particular service opportunity before deciding toseek service or not, the customer does not observethe firms current queue length. However, the cus-tomer has an expectation for the average arrival rateof customers to the firms service, , and the customerknows the function that translates an arrival rate intoan expected service time, W.2 Thus, wW is theexpected cost to the customer of the time to receiveone service opportunity. We refer to wW as theexpected service-time cost or the expected congestion

cost. Note that a customer cannot balk (or, choosesnot to balk) from the queue after choosing to seekservice (otherwise, the customer would effectively beable to observe the queue length before the joiningdecision is made). Finally, the firms pricing policyclearly influences the customers decision. With eachservice opportunity, the customer decides whetherto seek service based on the amount of utility thatwould be earned from the opportunity relative to con-gestion costs and the firms per-use fee (which inthe case of subscription pricing is zero). Whether toadopt a subscription is based on the expected arrivalof service opportunities and their expected net utili-

ties. We assume that the impact of any one consumeron the average queue length is insignificant. Con-sumers are risk neutral and make choices based onthe average utility each option generates (rather thanthe discounted utility of each option). In addition,consumers make pure-strategy choices (join the ser-vice system or not, subscribe or not). Allowing mixedstrategy choices either favors subscription pricing orhas no impact on our results.3

To complete the definition of the model, we pro-vide some additional structure for the service value

2 In fact, the customer only needs to have an expectation of the

firms service time, and that expectation must be correct (i.e., theydo not need to know the functional form of W .3 Consumers need to decide with each service opportunity whetherto seek service or not. The optimal strategy for a consumer isalways a pure strategy conditional on the value of the serviceopportunity. Hence, including mixed strategies has no impact withthis decision. Regarding the subscription decision, we find that thefirms profit can be higher if mixed strategies are allowed in theexogenous capacity model. However, the firms profit is unchangedin the endogenous capacity model by the inclusion of mixed strate-gies. (Refer to 3 of the electronic companion, available at http://msom.pubs.informs.org/ecompanion.html, for a complete analysisof equilibria in mixed strategies under both capacity scenarios.)

distribution and the system-time function. Let F bethe distribution function and f the density func-tion of each service value: assume F is differentiable,F 0 = 0, and F exhibits an increasing failure rate(IFR). For some results we invoke one of the follow-ing additional assumptions related to the hazard rate,

hx = fx/ Fx, where Fx = 1 Fx:Assumption 1 (A1). hx/hx2 is decreasing.

Assumption 2 (A2). xhx is increasing.

(A2) holds for a power distribution with parameter > 1, whereas both (A1) and (A2) hold if F is uni-form on the support 0 v or Weibull with parameters 1 and > 0. (Note, a Weibull distribution with = 1 is an exponential distribution.) In both versionsof the model, we assume F is uniform on the sup-port 0 v to derive analytical comparisons betweenthe pricing schemes. Regarding the system-time func-tion, in the capacity choice model (4) we assume that

W = 1/ , which corresponds to the expectedtime in an M/M/1 queue with first-come-first-servepriority. Furthermore, we use that functional form tocompare the pricing schemes in the exogenous capac-ity model (3).

3. Exogenous CapacityIn this section we analyze a version of our model inwhich the firms service-processing rate, , or capac-ity, is exogenously fixed with either pricing scheme.This analysis is appropriate for a firm that has theshort-term flexibility to modify its pricing but does

not have the short-term ability to alter its capacity.For each pricing scheme we derive the firms equilib-rium arrival rate and optimal revenues, which allowsus to establish conditions under which one scheme ispreferred over another.

3.1. Per-Use PricingWith per-use pricing a customer observes the real-ized value of a particular service opportunity andthen requests service if the net utility is nonnega-tive, i.e., the value of that opportunity is greater thanor equal to p + wW. Given that p, w, and arecommon to all customers (they all have the same

expectations) and constant across time, there is somethreshold value, v, such that a customer seeks servicewhenever the realized value of an opportunity is vor greater, and otherwise the customer passes on theopportunity:

v = p + wWThe actual arrival rate to the service is then Fv.For expectations to be consistent with actual operat-ing conditions (i.e., = F v, the threshold v mustsatisfy

v = p + wW F v (1)




5/17


Given that W is decreasing, it follows that there is aunique solution to (1). Furthermore, the threshold isincreasing in the per-use fee, p.

The firms revenue is Rp = p, which can beexpressed in terms of the threshold v:

Rpv

=

F vv

wW

F v

That is, the firms maximization problem can be writ-ten as maxv Rpv. The following theorem establishesthat an optimal threshold, vp, exists and is unique(proofs are provided in the appendix).

Theorem 1. The per-use revenue function, Rpv,is quasi-concave and vp = argmaxv Rpv is uniquelydefined by

vp =wW F vp+w F vpW F vp+F vpf vp

(2)

To translate vp back into an actual price, the firms optimalper-use fee is

pp =F vpf vp

+ w F vpW F vp (3)

To understand the economic intuition behind (2),note that the first term in the right-hand side (RHS)is the customers waiting-time cost. The second termis the externality the customer imposes on other con-sumers due to the (infinitesimal) increase in arrivalrate when she decides to join. From a social welfareviewpoint, customers should join as long as their util-

ity from joining vp is larger than the sum of thesefirst two terms. However, the third term, F vp/fvp,is added by the profit-maximizing monopolist. Thiscorrection term implies that the per-use arrival rate inequilibrium is smaller than the social optimal arrivalrate. Furthermore, from (3), the optimal per-use fee isgreater than the welfare-maximizing per-use fee.

3.2. Subscription PricingWith a subscription scheme there is no explicit feecharged per transaction, e.g., the members of a healthclub can use the service whenever they wish with-out additional charge. However, a customer may not

take advantage of a service opportunity if her valuefor that opportunity is low relative to her expectationof congestion costs, and that expectation depends onthe number of subscribers and the frequency of theirusage. For now, we assume that all consumers sub-scribe and then we confirm that expectation is correct.As a result, if each consumer uses the threshold vs todecide whether or not to seek service, then the arrivalrate to the service is F vs is the arrival rateof service opportunities conditional that all M con-sumers are subscribers and F vs is the fraction of

service opportunities that generate a service request.In equilibrium, the value of the service opportunityat which a consumer is indifferent, vs , exactly equalsthe expected congestion cost:

vs = wW F vs (4)

Now consider whether to purchase a subscriptionor not. At the time this decision is made the customerdoes not know when future service opportunities willoccur or their values, but does know his/her thresh-old value, vs, for seeking service. Hence, as part of thepurchasing decision, a customer expects that a sub-scription generates the following net value per serviceopportunity,

F vsEV V vs vs

where F vs is the probability that a service oppor-tunity is sufficiently valuable to seek service, EV

V vs is the value received conditional that a serviceopportunity yields a value greater than the threshold,and the last term, vs , is the expected congestion cost(from (4)).

Given that service opportunities arrive at rate , itis optimal for the firm to set the subscription rate, k,equal to the value of a subscription per unit of time(net of system-time cost):4

k = F vsEV V vs vs

All consumers purchase a subscription even thoughthey are indifferent between doing so or not, whichconfirms our initial assumption that all consumerssubscribe.5 As a result, subscription pricing allowsthe firm to extract all consumer surplus, conditionalon the level of congestion that subscriptions gen-erate. The latter condition differentiates this workfrom the literature on advance selling (e.g., Xie andShugan 2001)in those models the potential con-sumer surplus is independent of the pricing scheme,whereas here it depends on how much congestionmaterializes.

The firms resulting revenue can be expressed interms of the threshold vs :

Rsvs = kM= F vsEV V vs vs

Note that although the threshold vp was a deci-sion variable for the firm with per-use pricing, the

4 Lowering k merely reduces revenue per customer without chang-ing demand, so that cannot be optimal. There is no demand witha higher k, so that is not optimal either.5 If mixed strategies are allowed for the consumer purchase deci-sion, then it is possible to show that the firm may be able to earnhigher revenue with subscription pricing than we report.




6/17


firm does not control the threshold vs with subscrip-tion pricingit is set by (4). In other words, withexogenous capacity and subscription pricing, the firmcannot control congestion even though it possessesan effective mechanism for maximizing revenue con-ditional on the systems congestion. That said, con-

gestion is subject to some self-regulationcustomersrequest service only for service opportunities whosevalues exceed their expected congestion cost.

3.3. Comparison Between Per-Use andSubscription Pricing

This section compares the revenues generated byper-use and subscription pricing with an exogenouscapacity and quantifies the upper bound on the rev-enue loss from using these schemes relative to theoptimal scheme.

To compare the revenue generated via these twoschemes, note that with subscriptions a consumerpays an amount that equals the average net value

of her service requests, EV V v wW F v,whereas with per-use pricing, she only pays forthe net value of the marginal service request,v wW F v. Consequently, if the congestion lev-els were the same with either pricing scheme (i.e.,the thresholds v were identical), then subscriptionpricing clearly generates more revenue. We refer tothis as the revenue-extracting benefit of subscriptionpricing. However, the level of congestion will not beidentical (in general) across the two schemes. As onewould expect, a comparison of (2) and (4) reveals thatper-use pricing results in less congestion (a higherthreshold) than subscription pricing: vp

vs . This

establishes the trade-off between these two schemes:subscription pricing is better at extracting revenue,

but per-use pricing is better at controlling conges-tion. The firms preference over these two schemesdepends on which of these two effects dominates. Forexample, in the special case in which consumers areindifferent to congestion (i.e., when w = 0), subscrip-tion pricing yields higher revenue than per-use pric-ing and generates more social value per unit time (byhaving larger usage rates). However, as congestion

becomes more costly to consumers (as w increase),per-use pricing may be more attractive.

To make these comparisons more explicit, we

assume in the rest of this section that V U 0 v andW = 1/ . We now define the set of param-eters for which the firm can earn nonnegative rev-enue. Although the firms problem is determined byfour parameters (w, , , and v), the next theoremindicates that the pricing schemes relative rankingsdepend only on two of them.

Lemma 1. The relative revenue between subscriptionand per-use pricing, Rs/Rp, can be expressed in terms of and , where = w/v and = /, and both revenuesare nonnegative for 0 1.

Whether per-use pricing or subscriptions are pre-ferred depends on (which measures the relativestrength of congestion costs to service values) and thepotential utilization rate of the system, .

Theorem 2. When = 0, subscription pricing alwaysyields higher revenue than per-use pricing. For each value

of > 0, there exists a unique such that subscrip-tion yields higher revenue than per-use pricing for < (recall, is the potential utilization, /. Other-wise, per-use pricing yields higher revenue. Moreover, is decreasing in .

From Theorem 2, per-use pricing is preferred oversubscription for highly congested systems. The keyissue is the degree of congestion needed for per-use pricing to be preferred. For various levels ofpositive congestion costs, > 0, Table 1 providesthe potential utilization rate, , at which the twoschemes yield the same revenue. It can be demon-

strated (Proposition 4 of the electronic companion)that lim0 = 2 and lim1 = 1. Thus, sub-scription pricing always generates higher revenuethan per-use pricing when the potential arrival rateto the queue is less than the processing rate. Sub-scription pricing can be preferred even if the potentialarrival rate is as much as 140% of the firms process-ing rate. Subscription pricing may also be preferredwhen the systems actual utilization rate, F v/, ishigh. Table 1 lists the systems actual utilization ratewhen the potential utilization rate is . For exam-ple, when = 001 and = 1411, subscription pric-ing yields the same revenue as per-use pricing even

though the actual utilizations are 968% and 648%,respectively. Of course, subscriptions performs bet-ter than per-use pricing under such high utilizationrates when the relative congestion cost parameter, ,is low. When is high, subscriptions will be preferredto per-use pricing if the actual utilizations are moremoderate. In addition, it can be shown that the actualutilization rates are increasing in (proof availablefrom authors). Thus, when = 001, subscription pric-ing is preferred whenever it yields an actual utiliza-tion rate that is lower than 968%. Hence, althoughsubscription pricing cannot control congestion well,

Table 1 Potential Utilization Rates, , That Yield Identical Revenue

with Per-Use and Subscription Pricing, as Well as Actual

Utilizations When the Potential Arrival Rate Is

Actual utilization (%) when =

Per use Subscription

0.99 1.002 0.3 0.5

0.75 1.069 7.1 13.8

0.50 1.153 16.4 31.3

0.25 1.264 30.5 55.5

0.01 1.411 64.8 96.8




7/17


Figure 1 Percent of Two-Part Tariff Revenue Attained by the Optimal Simple Scheme as a Function of Potential Utilization, , for Different Values of

60

70

80

90

100

0 11.153 1.411

2 3 4

Revenueoftheoptimalsimplescheme

relativetotwo-parttariff(in%)

= 0.01

=0.5

=0.99

it still generates higher revenue than per-use pric-

ing even in systems with a considerable amount ofcongestion.

To explore the strength of subscription pricing fur-ther, the next theorem characterizes revenues withextreme levels of potential utilization.

Theorem 3. The following limits hold: (i) lim0 Rs =v1 2/2 and lim0 Rp = v1 2/4.(ii) lim Rx = 0, x sp

Subscription pricing generates twice as much rev-enue as per-use pricing when capacity is unlimited = 0. Therefore, subscription pricing starts with aconsiderable advantage relative to per-use pricing. As

a result, congestion needs to be substantial in the sys-tem before the congestion-controlling benefits of per-use pricing dominate the rent-extracting capability ofsubscription pricing. Furthermore, revenue declinesin with all schemes, so per-use pricing dominatessubscription pricing only when revenues are in factlow. This suggests that per-use pricing can provideonly a modest absolute advantage relative to sub-scription pricing, but the absolute advantage of sub-scription pricing can be substantial. Taken together,these results indicate that from a practical perspective,subscription pricing can indeed be better than per-use pricing even if capacity is fixed and the system issubject to congestion-related costs.

3.4. Comparison to the Two-Part TariffAs mentioned in 1, the two-part tariff is the opti-mal pricing scheme in this setting. A two-part tariffcombines a per-use fee with a subscription rate. Thefirm can set a per-use price to achieve the social opti-mal congestion, and it can set a subscription priceto extract all customer welfare. Hence, social wel-fare is maximized and is fully extracted by the firm.A complete analysis of the two-part tariff is relegated

to Propositions 13 of the electronic companion. We

characterize the two-part tariff scheme, show that thecongestion under two-part tariff is lower than thatunder subscription pricing and higher than underper-use pricing, and find that the revenue generated

by the two-part tariff is well approximated by sub-scription when is low and by per-use pricing when .

Despite its ability to extract revenue, as discussedearlier, two-part tariffs may not be offered in prac-tice for reasons that we do not model. Nevertheless,we can evaluate the percent loss of using an optimalsimple scheme (per-use or subscription) relative tothe optimal two-part tariff.6 Figure 1 illustrates the

percent of two-part tariff revenue that is attained byoptimally setting one of the two simple schemes as afunction of potential utilization, , for different valuesof . For any fixed , the maximum percent revenueloss of the optimal simple scheme relative to the opti-mal two-part tariff occurs at the potential utilizationwhere both simple schemes are equally profitable (i.e.,at ). The revenue loss is lower at lower values of (subscription pricing is optimal) and higher values of (per-use pricing is optimal), vanishing at 0 andat . Moreover, the maximum percent revenueloss is decreasing in .

4. Capacity ChoiceIn 3 the firm can choose how to price, but not itscapacity, so the pricing decision results only in vari-ation in service time. In this section the firm chooseshow to price and its capacity, so the pricing deci-sion influences both the firms capacity and its ser-vice time. We assume that capacity is costlythe firm

6 Giridharan and Mendelson (1994) also quantify the loss of usinga suboptimal pricing scheme in a model with congestion.




8/17


incurs a cost at rate c for maintaining capacity ,where c > 0. Furthermore, we continue to assumeW = 1/ .4.1. Per-Use PricingThe consumers choice in this setting is the sameas in the fixed-capacity model. As a result, we can

express the firms profit function in terms of thethreshold value at which consumers are indifferent, v,and capacity, :

pv = Rpv c = Fv

v w Fv

c

The profit function is concave in , so it is straight-forward to determine that pv is the firms optimalcapacity for a given threshold, v, where

pv = Fv +

w Fvc

The firms profit rate, pvpv, can now be writ-ten as

pv = F vv c 2

F vwhere the constant is defined for convenience:

=

cw/

The firm solves the maximization problem, maxv pv. The following theorem establishes the uniquenessof the optimal per-use threshold.

Theorem 4. If v > c, there exists an upper bound psuch that for every < p there exists a unique optimal

threshold, vp = argmaxv pv, that yields positive profit,pvp > 0. This threshold is the smallest solution to theimplicit equation given by

vp =F vpf vp

+ F vp

+ c (5)

Furthermore, if (A1) holds, then there exist two solutionsto (5). Otherwise, there does not exist an optimal vp < v.The optimal capacity is

p = F vp +

w F vpc

(6)

and the firms per-use fee is

pp = vp w/p F vpThe bound in Theorem 4, p, merely states that

the firm can earn a positive profit only if capacity issufficiently cheap, customers are sufficiently patientand the market is sufficiently large. Observe that werestrict attention to the interesting case in which prof-its are positive. If > p (so that pvp < 0), the firm

is strictly better off not selling per-use, and if = p(so that pvp = 0), the firm is indifferent.

4.2. Subscription PricingWith subscription pricing and a fixed capacity thefirm has little control over congestion. However,the firm gains some control over congestion whenthe firm can choose its capacity. In particular, if is the firms capacity, then a consumer with value v =w/ F v is indifferent between seeking serviceor not. Instead of thinking in terms of the firm choos-ing , we can use that relationship to frame the firmsproblem in terms of choosing the threshold, v,

sv = Fv + w/v

The firms profit function can then be written as

sv = F vEV V v v c Fv + w/v

where the first term is the revenue the firm earns fromsubscriptions assuming the firm chooses the maxi-mum subscription fee that induces all consumers to

purchase a subscription, conditional on the expectedlevel of congestion. The firm solves the maximizationproblem, maxv sv.

Theorem 5. If E V > c, there exists an upperbound s such that for every < s , there exists anoptimal threshold, vs arg maxsv, that yields positive

profit, svs > 0. This threshold is implicitly defined by

vs =

1

F vs cfvs (7)

Furthermore, if (A2) holds, then there exist two solutions

to (7) and the smallest solution is the unique optimalthreshold. The optimal capacity is

s = F vs +w

vs (8)

and the firms subscription rate is

k = F vsEV V vs vs

As with Theorem 4, Theorem 5 indicates that a posi-tive profit occurs only when capacity is not too expen-sive, customers do not incur time costs that are toohigh, and there is a sufficient number of customers

in the market. However, the two bounds, p and s ,need not be the same. Observe also that as with Theo-rem 4, we restrict attention to the case in which profitsare positive. If > s (so that svs < 0), the firm is

better off not selling subscriptions, and if = s (sothat svs = 0), the firm is indifferent.

4.3. ComparisonIn this section we assume V U 0 v. Let vt and t

be the optimal threshold value and the service ratechosen when the firm uses the two-part tariff scheme.As with fixed capacity, by charging a two-part tariff,




9/17


the firm is able to achieve social optimal congestion(by setting an appropriate per-use fee) and to extractall consumer welfare (by setting a subscription ratethat makes consumers indifferent to subscribing), sothe two-part tariff is optimal for the firm.

As in the fixed-capacity case, we show that per-

use pricing leads to a system with less congestionthan optimal and that subscription pricing leads tomore congestion than optimal: vs < vt < vp. Further-more, the firm invests more in capacity than optimalwith subscription pricing (to control congestion some-what) and less with per-use pricing: p < t < s .Even though the firm invests more in capacity withsubscription pricing, congestion is also higher withthat scheme: usc > utc > upc,

7 where uxc is theactual utilization rate,

uxc = F vx/x x sp

As in the exogenous capacity case, when w = 0, con-gestion is not an issue. The firms choice is simple.Because there are no congestion costs, the firm has noincentive to provide fast serviceit will choose = ,and the capacity cost will be the same for both pric-ing schemes. That is, similar to the exogenous capac-ity case, when w = 0, subscription pricing dominatesper-use pricing for all values of c.

In the discussion that follows, we compare betweenthe two pricing schemes when the marginal conges-tion cost is positive (i.e., for w > 0). If capacity is inex-pensive, c = 0, subscription pricing performs strictly

better than per-use pricing,

svs c = 0 > pvp c = 0

without the concern of congestion, the revenue extrac-tion benefit of subscription pricing dominates. How-ever, subscription profits decrease at a faster rate withrespect to the cost of capacity,

svs

cupc for every IFR distri-

bution. The comparisons of the subscription threshold value withthe two-part tariff one was only proven for the uniform distribu-tion, but we observed numerically that the results resported in thissection hold for the Weibull distribution with 1 as well. (Thecumulative distribution function of a Weibull random variable X is

Fx = 1 ex/ , where > 0 and > 0.) A detailed description ofthe numerical analysis is available from the authors.

Figure 2 Profit Rates of the Two Pricing Schemes with Respect to the

Capacity Cost, c

(a)

1

6

4

Subscription

2

0

Profitrates

3

Per-use

c

C~

5 7

(b)

Subscription

6

4

2

0

Profitrates

Per-use

c

1 3 5 7

Note. The following parameter values are used: (a) w = 005; and(b) w= 05. (= 1 and v= 10 in both panels.)

pricing scheme x sp. Combining these results,one of two scenarios emerges: either cs cp or cp cs .

Figure 2 illustrates these scenarios. On the lefthand side, cs cp. There exists some c such thatthe two schemes earn the same profit, svs c =pvp c > 0. It follows that subscription yieldshigher profit than per-use pricing for c 0 c,whereas per-use is better for c c cp. Furthermore,for c

cs

cp, subscription pricing cannot earn a posi-

tive profit, whereas per-use pricing does. That is whatone might expect given that subscription pricing givesthe firm less control over congestionif capacity costsare sufficiently high, per-use pricing is preferable andmay be the only scheme that yields a positive profit.However, although per-use pricing can be more prof-itable than subscription pricing, it is only more prof-itable when capacity is sufficiently expensive, andthe absolute advantage of per-use pricing is generallysmall, whereas the absolute advantage of subscriptionpricing can be large.




10/17


The second scenario, cp cs is illustrated on theRHS of Figure 2. Subscription pricing is preferred ifc 0 cp, and subscription pricing is the only schemethat returns a positive profit if c cp cs. In otherwords, it is possible that subscription pricing is thepreferred scheme for any capacity cost that allows the

firm to make a profit. Furthermore, if capacity is suffi-ciently expensive, it is possible that subscription pric-ing can yield a profit, whereas per-use pricing cannot;in those situations capacity is sufficiently expensivethat per-use pricing is unable to extract enough rev-enue from customers to cover the cost of capacity.8

(Refer to Proposition 8 in the electronic companionfor a formal statement and proof of the result.)

The only difference between the two panels in Fig-ure 2 is that the right-hand side has a higher wait-ing cost: w = 05 instead of w = 005. In fact, it can

be shown (Proposition 9 of the electronic companion)that there exists a w such that for all w > w the secondscenario occurs, i.e., if the marginal waiting cost issufficiently high, subscription pricing dominates per-use pricing for all capacity costs that yield a positiveprofit. In other words, when congestion is most costly,in the sense that the service-time cost is high, thensubscription pricing can be better than per-use pric-ing even though it has less control over congestion. Toexplain, if congestion costs are high, a large capacitymust be chosen to minimize congestion, and this canonly be profitable when the pricing scheme is able toextract a sufficient amount of revenue.

Figure 3 illustrates the threshold structure thatidentifies the regions in which each pricing scheme

is more profitable. When the combination of capac-ity and congestion cost parameters is high (i.e., forwc points located above the solid line), both pric-ing schemes yield negative profits. Subscriptions arepreferred in the lower region and per-use pricing ispreferred in the region with low w and high c.

It is also illustrative to compare the pricing schemeswith respect to actual utilization. It can be shownthat the relevant metric is w/v. (Note that witha fixed capacity we use w/v for making compar-isons between the two pricing schemes, but now is endogenous and different across schemes.) Table 2provides the firms actual utilization under each pric-

ing scheme when capacity is c, i.e., when the marginalcost of capacity is such that per-use and subscriptionpricing yield the same profit. (If w were any higher,

8 This result provides an interesting contrast with the necessaryconditions for each pricing scheme to be profitable. Recall thatEV > c is necessary for subscription pricing, whereas the lessrestrictive v > c is necessary for per-use pricing. These are onlynecessary conditions, and not sufficient conditions, as we havedemonstrated. Therefore, it would be misleading to conclude fromthose conditions that a high capacity cost favors per-use pricing inall circumstances.

Figure 3 Profitability Regions in wc Space for Each Pricing Scheme

Negativeprofits

c

w

Subscription

Per-use

then subscription pricing dominates per-use pricingfor all utilizations that yield a positive profit, i.e., inthat case we enter the w > w regime.) We observenumerically that actual utilization is increasing in cwith each pricing scheme. Consequently, subscription

pricing is better than per-use pricing for all utiliza-tions that are lower than those indicated in the table.For example, when w/v = 003, subscription pric-ing is better than per-use pricing whenever it yieldsa utilization of 80% or lower. The table indicates thatsubscription pricing can be better than per-use pricingeven if the utilization rate is quite high (say, higherthan 98%). Therefore, as in the previous model, sub-scription pricing can be better than per-use pricingeven if it results in a highly utilized system.

5. Customers Heterogeneity in

Usage RatesIn our model, customers are heterogeneous in theirrealized values for service opportunities, but they areotherwise homogeneous. A natural relaxation of thismodel is to allow heterogeneity in service usage rates.For example, suppose there are two equal-sized seg-ments of consumers. One segment has service oppor-tunities that occur at rate h = + , and the othersegment has service opportunities that occur at ratel = , where 0 < , i.e., is the average arrivalrate. If consumers do not know a priori which seg-ment they belong to, then our analysis continues tohold because the consumers remain homogeneous in

their expectations. To create meaningful segments, itis necessary to assume that consumers know a priorito which segment they belong. This section considersa model with this assumption.

Table 2 Actual Utilization When Capacity Is Such That Subscription

Pricing and Per-Use Pricing Yield the Same Profit i.e., c = c

w/v 003 002 0005 0002 00005 00001 000001

up 614 672 819 881 938 972 991

us 802 837 917 948 974 988 996




11/17


Interestingly, the per-use results in 3 and 4 con-tinue to hold without any needed modification. Toexplain, with per-use pricing each customer makes adecision with each service opportunity, so the rate atwhich service opportunities occur has no impact onany one decision. All that matters is the aggregate rate

at which consumers use the serviceas long as isheld constant, it does not matter if there is one seg-ment ( = 0 or two segments ( > 0 or how far apartthe two segments are.

On the other hand, the existence of different seg-ments influences subscription pricing. We show thatwhen the service rate is fixed, heterogeneity in usagerates may increase or decrease subscription revenues,

but that heterogeneity always decreases profits whenthe firm chooses the level of capacity. With subscrip-tion pricing the firm has two choices. The first is toset the subscription rate such that all types purchase

kl=

l F vsl EV

V

vsl

vsl

wherevsl = wWMF vsl

The second is to set the subscription rate high enoughso that only the high-usage types purchase,

kh = h F vshEV V vsh vshwhere

vsh = wWMh F vsh/2Consider first the fixed-capacity model (3). Let Rland Rh be the firms revenue functions when sell-ing to both segments and to only the high-usage seg-ment, respectively. In our original model,

=0, the

subscription price is kl, all consumers subscribe, andthe firm extracts all social welfare. As the market

becomes more segmented (that is, increases) thefirm must reduce kl to capture both segments. Thelow-type consumers continue to earn zero surplus,

but now the high-type consumers are able to retainsome surplus, which increases as more segmentationis introduced. Hence, Rl0 > Rl 0 condi-tional on all customers subscribing, compared to theoriginal model, subscriptions are less attractive rela-tive to per-use pricingthe increase in results in adecrease in subscription revenue, but has no impacton per-use revenue.

This, however, need not be the case if the firm aban-dons the low segment and prices at kh to capture onlythe rent from the high types. By selling only to theheavy users, the firm is able to charge a higher priceand extract all social welfare, but it sells to only half ofthe consumers. These two countervailing effects mayresult in higher profits relative to the original model.

The next theorem establishes the condition forwhich heterogeneity in the consumers usage ratesmay lead to an increase in the firms profits undersubscription for V U 0 v and W = 1/ .

Theorem 6. Rh is quasi-concave. If > 1, thereexists a unique , such that Rh > Rl0 .Otherwise, Rh Rl0 and Rh is increasing.

From Theorem 6, if the potential utilization is high,selling subscriptions can become even more attrac-tive when customers are heterogeneous. To explain,

remember that although subscription is good atextracting revenues, it cannot control congestion. This

becomes more problematic as increases. By sellingto only the high-usage customers, the firm forgoesrevenue from low-usage consumers, but it is able tocontrol congestion somewhat, while at the same timeit extracts the entire customer surplus. In this case, if is high enough, this results in higher revenues thanthe revenues obtained by selling to all consumers at = 0.9

Let Rs be the firms optimal revenue under sub-scription with degree of heterogeneity i.e., Rs =maxRlRh. Rs can be fully characterized

by , , and = / 0 1 (the proof is simi-lar to Lemma 1 and is omitted). The next theoremcompares between subscription and per-use revenueswhen consumers are heterogeneous in usage rates forV U 0 v and W = 1/ .

Theorem 7. When = 0, subscription pricing alwaysyields higher revenue than per-use pricing does. For eachvalue of > 0 and every value of 0 1, there exists aunique , such that subscription yields higher rev-enue than per-use pricing for < (recall, is the

potential utilization, / and is a standardized measureof heterogeneity, /. Otherwise, per-use pricing yields

higher revenue.

Theorem 7, which is a generalization of Theorem 2,shows that when consumers are heterogeneous intheir usage rates, subscription is preferred over per-use pricing for low levels of and per-use is pre-ferred over subscription for high levels of . I t isinteresting to assess how the degree of congestionneeded for per-use to be preferred changes when con-sumers are heterogeneous. For various levels of pos-itive congestion costs, > 0, and different degrees ofheterogeneity, , Figure 4 illustrates the potential uti-lization rate, , at which the two schemes yield

the same revenue. Observe that 0 = 1. Thisfollows because when = 0, consumers are homoge-neous and Rl0 is equivalent to the revenue of theoriginal subscription model. When = 1, the levelof heterogeneity is so high that the revenue obtained

9 This result is analogous to the result obtained by allowing formixed-strategy equilibrium. Setting a higher subscription price tomake only a fraction of consumers join is equivalent to choosing thelevel of market segmentation, . And, in fact, we find that if (andonly if) > 1, allowing for a mixed-strategy equilibrium benefitssubscription even more.




12/17


Figure 4 Potential Utilization Rates, , That Yield IdenticalRevenue with Per-Use and Subscription Pricing When

Consumers Are Heterogeneous

0.6

0.8

1.0

1.2~p

1.4

1.6

1.8

0 0.2 0.4 0.6 0.8 1.0

=0.01

= 0.25

= 0.5

= 0.75

= 0.99

from selling only to the high-usage consumersi.e.,Rh1is equivalent to the revenue of the originalsubscription model. Observe also that the values of 0 = 1 (marked by the dashed lines in Fig-ure 4) correspond to the values presented in Table 1.Figure 4 demonstrates that the degree of conges-tion needed for per-use pricing to be preferred oversubscription when consumers are heterogeneous caneither increase or decrease. This result follows fromthe two countervailing effects that heterogeneity hason subscription revenues: when consumers are het-

erogeneous, the firm cannot extract all revenues byselling to all consumers, and thus the revenue accrued

by selling to all consumers decreases. However, whenthe potential utilization is high, the firm can benefitfrom selling to only the high-usage consumers if thedegree of heterogeneity is high (Theorem 6).

When capacity choice is endogenous (4), the firmthat sells subscriptions can control congestion notonly by setting a high subscription price so that onlyhigh-type consumers purchase, but also by choosingthe service rate. As in 4.2, instead of solving for ,we can frame the firms problem in terms of choos-ing the thresholds vl (if the firm decides to set a sub-

scription price so that all consumers purchase) andvh (if the firm sets a high subscription price so thatonly high-usage consumers purchase). Consider firstthe choice of capacity conditional on all consumerssubscribing. In this case,

l = F vl + w/vland the firms profit function can be written as

lvl = Ml F vlEV V vl vl c F vl + w/vl

The firm solves the maximization problem,maxvl lvl .

Theorem 8. If l/EV > c, there exists an upperbound l, such that for every l, there exists anoptimal threshold vsl arg maxlvl , that yields non-negative profit, lvsl

0. This threshold is implicitly

defined by

vsl =

1

l/ F vsl cf vsl (9)

Furthermore, if (A2) holds, then there exist two solutionsto (9) and the smallest solution is the unique optimalthreshold conditional on all consumers subscribing.

If, however, the firm decides to cater only to thehigh-usage consumers,

h = Mh F vh/2 + w/vh

and the firms profit function is

hvh = Mh F vhEV V vh vh/2 cMh F vh/2 + w/vh

The firm solves the maximization problem,maxvh hvh .

Theorem 9. If E V > c, there exists an upperbound h, such that for every h, there exists anoptimal threshold vsh argmax hvh that yields non-negative profit, hvsh 0. This threshold is implicitlydefined by

vsh =

Mh/2 1F vsh cf vsh

(10)

Furthermore, if (A2) holds, then there exist two solu-tions to (10) and the smallest solution is the unique opti-mal threshold conditional on only high-usage consumerssubscribing.

Finally, the firm sets the subscription price thatyields maximum profits, maxlvslhvsh.We show that when the firm chooses the levelof capacity, customer heterogeneity always decreasesprofits under subscription (Proposition 11 of the elec-tronic companion). In this case the firm is able tocontrol congestion by choosing the service rate, .Whereas selling only to heavy users had a congestioncontrol advantage in the fixed-capacity case, here thefirm loses from selling to fewer consumers. Hence,compared to the original model, subscriptions are lessattractive, in a relative sense, than per-use pricing.Moreover, the threshold capacity cost under whichsubscription profits are higher than per-use profits islower than in the original model (Corollary 1 of theelectronic companion).




13/17


We conclude that customer heterogeneity in usagerates generally makes per-use pricing more attractiverelative to subscription pricing. However, it can makesubscription pricing even more attractive if capacityis fixed and and are high.

Finally, we suspect that with heterogeneity in usage

rates, a single two-part tariff is generally no longeroptimal for the firmgiven that there are multi-ple segments, a single contract that is designed toserve both types will no longer be able to extractthe entire surplus. In this situation the firm will typi-cally improve its profits by designing a menu of two-part tariffs or nonlinear tariffs (e.g., Clay et al. 1992,Miravete 1996, Masuda and Whang 2006).

6. ConclusionUsing a queueing framework, we find that a firmmay prefer subscription pricing over per-use pricingeven if consumers dislike congestion. Furthermore,

subscription pricing may be preferable in situationsthat would a priori suggest a preference for per-use pricing: when customers strongly dislike the timeto complete the service, thereby making congestioncostly to the firm. Subscription pricing can dominatein these situations because (i) the firm must investin a considerable amount of capacity to reduce ser-vice times to a minimum and (ii) the firm can coverthat large capacity cost only if it can extract enoughrevenue from customers. Next, we find that the abso-lute advantage of subscription pricing can be con-siderable, whereas the absolute advantage of per-usepricing is generally modestper-use pricing gener-

ates higher revenue or earns higher profit only whenrevenue or profit is reasonably low. If customers areheterogeneous in their service rates, subscription pric-ing can become even more attractive if capacity isfixed and the potential utilization and market seg-mentation are high. Otherwise, heterogeneity in usagerates makes subscription pricing less attractive rela-tive to per-use pricing. Overall, we conclude that theemphasis on per-use pricing in the queueing literatureis misplacedwe provide evidence that subscriptionpricing can indeed be the preferable pricing strategyeven in services that experience congestion.

Electronic CompanionAn electronic companion to this paper is available onthe Manufacturing & Service Operations Management website(http://msom.pubs.informs.org/ecompanion.html).

AcknowledgmentsThe authors thank Albert Ha, the seminar participants atHong Kong University of Science and Technology, and con-ference attendees at the M&SOM conference in Maryland,the INFORMS Revenue Management and Pricing SectionConference in Montreal, and the INFORMS Annual Meetingin Washington, DC.

Appendix. ProofsLet g vs/v and l vp/v. The functions g and lwill be used in the proofs of Lemma 1 and Theorem 2 below.

Proof of Theorem 1. Let v = dRpv/dv, wheredRpv

dv= f vv wW F v

+ F v1 + wfvW F vThere is at least one maximum because 0 > andlimv v 0. By construction, vp, given by (2), satisfiesthe first-order condition. Take vp = 0 and rearrange terms

1 F vp

vpf vp= wW

F vpvp

+ wF vpW F vp

vp (11)

The RHS of (11) is decreasing in vp because Fv = PrV vis decreasing in v and W is increasing and convex. Theleft-hand side (LHS) of (11) is increasing in vp because Fis IFR. (Actually, the left-hand side is increasing even if Fhas an increasing generalized failure rate.) Thus, there is aunique v

pthat satisfies v

p=

0.

Proof of Lemma 1. For the per-use case, Theorem 1establishes that there is a unique optimal v, which is v whenRpvp 0. That condition simplifies to w/ v/, or 1.It follows that Rpvp 0 for all 0 1. With subscriptionpricing, revenue is Rsvs = F vsEV V vs vs. ForV U 0 v positive revenues occur vs < v. Note that thethreshold vs is given by vs = w/ F vs, where the LHSis increasing and the RHS is decreasing. Thus, for a nonneg-ative revenue to occur, we must have that (when evaluatingthe above condition), at v, v w/ 0, or 1. Next, weshow that the relative revenues are a function of and .For the uniform distribution, condition (4) can be written as

vsv =

1 1 vs/v (12)Note that g is a function of and only. Plugging vsinto the subscription revenue function, we obtain Rs = 1g2/2v. Similarly, for the pay-per-use case, we canexpress condition (2) as

vp

v= 1

2

1 +

1 1 vp/v2

(13)

Note that l is a function of and only. Plugging vp intothe per-use revenue function, we obtain

Rp = 1 l

l 1 1 l

v

Thus, Rs/Rp depends only on and .

Proof of Theorem 2. The fact that subscription alwaysdoes better than per-use pricing when w = 0 is immedi-ate. In that case, F vsEV V vs F vpEV V vp> F vpvp, where the first term is the revenue generated

by subscription pricing and the last term is the revenuegenerated by per-use. Note that the first inequality follows

because vs vp. The remainder of the proof establishes theresult for w > 0. Uniqueness: by implicitly differentiatingthe revenue functions, we get

Rs

= 1 g

2

+ 1 1 g2 v (14)




14/17


andRp

= 1 l

2

1 1 l2 v (15)

where g and l are shorthand notation for g and l.To show that there exists a unique such that Rs

>Rp < and Rs < Rp > , it is enough torequire that

for which R

s

> Rp

, Rs

< Rp

.Rs>Rp implies that

1 g221 l > l

1 1 l

Rearranging (14) and (15), requiring Rs < Rp, we must

have that

1 g221 l >

1 l + 1 1 g221 1 l2

Thus, to complete the proof, it is enough to show that

l 1 1 l >

1 l + 1 1 g221 1 l2 (16)

Plugging in l (from condition (13)) into the LHS of (16) andrearranging, condition (16) becomes

1 1 l + 1 l

1 1 g1 1 l

2< 1 (17)

To see that condition (17) holds, note that the first term issmaller than l (follows from the fact that vs < vp ,which implies that the term is less than g and that g < l)and that the second term is less than 1 l (because g < limplies that the squared term is less than 1).

The threshold utilization factor, , is implicitly defined by

= 1 g 2

2 1 l

l 1 1 l

= 0 (18)

Differentiating with respect to yields

d

d=

+

+

g g

+

l l

= 0

Note that the last term equals zero (from the first-order con-dition). Rearranging the terms, we find that

= / + /g g/

/

=

1 l/1 1 l + 1 gg/

1 l2

/1 1 l2

whereg

= 1

1 1 g + g > 0

Thus, we then get that / < 0. Existence: it was estab-lished in Proposition 4 of the technical appendix that athreshold exists for = 0 and = 1. Because / < 0,existence is guaranteed 0 1.

Proof of Theorem 3. (i) = 0: Substituting = 0in Equations (12) and (13) results in vs = v and vp =1 + v/2. Then, the following expressions for the revenue

rates are immediate:

Rs =v1 + 2

2 Rp =

v1 + 24

(ii) : Rearranging (12), we getvs

v

1

+ vs

v =

As there are up to two roots that solve the above.The larger of the two is a maximum. This implies that inthis case, we have

lim

vsv

= lim

1

= 1

Similarly, rewriting (13), we getvp

v 1

2

1

+ vp

v

2=

22

This implies that for all pricing schemes,

lim

vs

v =lim

vp

v =lim

1 =

1

Substituting into the revenue rates, we obtain for :lim Rs = lim Rp = 0.

Proof of Theorem 4. We prove by contradiction that ifthere is a maximum, it is unique. Suppose there exist v1and v3 such that v1 < v3, pv1 0, pv3 0, pv1 =0, pv3 = 0, i.e., both are local maxima with nonnegativeprofits. Our conditions imply for both v v1 v3 that

v c =Fvfv

+ Fv

v c2

Fvwhere the first condition is the first-order condition and thesecond condition ensures nonnegativity of profits. Combin-ing the two conditions, we have

Fvfv

Fv

(19)

Given that v1 and v3 are local maxima, there must be a localminima, v2, such that v1 < v2 < v3. There are two cases toconsider: pv2 < 0 and pv2 > 0.

Consider pv2 < 0. Analogous to (19), pv2 = 0 and

pv2 < 0 imply

F v2f v2

F v2f v2

>F v3f v3

(21)

andF v3

>F v2

>F v1

(22)




15/17


Combining (20), (21), and (22) yields

F v3f v3

0 for p. Otherwise, there does not exist an optimalvp < v.

The smallest solution to (5) is vp; although we cannotcompute the number of possible solutions to (5) for a gen-eral IFR distribution F, we show by contradiction that theoptimal vp is the smallest solution to (5). Note first thatp0 = c + 2 < 0 and that p0 > 0. Thus, the small-est solution to (5) is a local maximum. We have alreadyshown that if there exists a local maximum so that pv 0, it is unique. Suppose there exist two local maxima withnegative profits v1 and v3 such that v1 < v3, i.e., pv1 < 0,pv3 < 0,

pv1

=0, pv3

=0. Our conditions for both

v v1 v3 imply that

v c =Fvfv

+ Fv

v c2

F v2

(23)

Consider v1. Because F is IFR, v1 < v2 implies

F v1f v1

>F v2f v2

(24)

andF v2

>F v1

(25)

Combining conditions (23) and (25), we get

F v2f v2

>F v2

>F v1

>F v1f v1

which contradicts condition (24). Letting v2 < v3, however,a contradiction cannot be reached, which does not preclude

the existence of additional solutions to (5) in the negativerange for a general IFR distribution. Combining the results,we conclude that vp is the smallest v that solves (5).

Sufficient condition for at most two solutions to (5): Rear-ranging (5), we have

vp 1

hvp=

F vp+ c

The RHS is convex and increasing, and the LHS isincreasing. Taking the derivative of the LHS, we get 1 +hvp/hvp

2. Thus, if condition (A1) holds, there can beat most two solutions to (5), with the smallest one being themaximum.

Proof of Theorem 5. First note that s0 = and thatlimvv sv = 0. Differentiating sv, we obtain

dsv

dv= c

w

v2+ fv

F v

Equating to zero and rearranging terms, the result in (7)follows.

A maximum exists if there exists a vs < v such thatsvs = 0 and svs 0 Requiring that svs 0 is equiv-alent to having

sv

Fv = EV V v v c 2

v Fv 0

for some v. Assume = 0. Then, if EV > c, there mustbe a solution with positive profit. Let Ms svs.From the Envelope Theorem, we have Ms/ =2/vs < 0, which means that svs is decreasingin . Note that even though we have not ruled out the exis-tence of several local maxima vs , svs is decreasingin at every critical point. This implies that there existssome s such that svs for s . Otherwise, theredoes not exist an optimal vs < v.

Furthermore, denote the RHS of (7) by zv, i.e., zv =1 c hv1/2 F v1/2. We want to show that thereexists a unique vs that maximizes profit and solves




16/17


vs = zvs. Because F is IFR, zv is increasing. Differentiat-ing zv, we get

zv = 2

c hv1 c hv3/2 F v1/2

+ fv1 c hv1/2 F v3/2Plugging in (7), we get

zvs =1

2

hvsvs +

c hvsvs1 c hvs

A sufficient condition for zvs to be increasing is for bothterms in the brackets to be increasing. The first term is thegeneralized failure rate. It is increasing if F is IGFR. Thesecond term is increasing if hvsvs is increasing and F isIFR. Thus, under these conditions, zvs is increasing andconvex and there are at most two solutions to v = zv.Because s0 < 0 and

s0 > 0, the smallest solution is the

maximum.

Proof of Theorem 6. First note that Rh = Rl0. Exis-tence: differentiating Rh with respect to , we get

dRhvsh

d= Rhvsh

+ Rhvsh

vs dvsh

d

= M v vsh2

4v Mhv vsh

2v dvsh

d

where dvsh/d > 0. Let = dRh/d. Becauselim0 > 0 and the domain of is bounded above, thereexists at least one maximum. Uniqueness: rewrite vsh =wWMh F vsh/2 in terms of :

vsh =2 w/vsh

MF vsh (26)

Plugging the expression in the revenue function, we get

Rhvsh = wvsh EV V vsh vsh=

wvsh

v vsh

2

where the last inequality follows because of the uniformdistribution assumption. Solving for the FOC yields

vsh =

wv

(27)

which is unique. Thus, Rh is quasi-concave. PluggingEquation (27) in (26) and simplifying, we obtain that +vs = 2/M. This is achievable if > , and the resultsfollow. Otherwise, the maximum Rh is obtained for

=

and Rl0 > Rh 0 . Proof of Theorem 7. By implicitly differentiating the

revenue functions, we get

Rl

= 1

1 gl2 + 1 1 gl2

v (28)

Rh

= 1+

/221gh21+/2+11+/21gh2

v(29)

and Rp/, given by (15), where gl vsl/v and gh vsh/v (analogously to g and l). To complete the proof, it is

enough to show that (1) there exists at most one , suchthat Rl

> Rp < and Rl < Rp > ;(2) there exists a unique , such that Rh

> Rp < and Rh

< Rp > ; and (3) Rl=0 < Rh=0,then Rl

< Rh .

(1) Rl >Rp implies that

11gl2

21l > l 11l

Rearranging (28) and (15), requiring Rl < Rp we

must have that

11g221l >

1l+11g2211l2


l 11l >

1l+11g2211l2

which follows from the proof of Theorem 2.(2) Uniqueness: Rh

> Rp implies that

1

+/21

gh

2

21l > l

11l Rearranging (29) and (15), requiring Rh

< Rp, wemust have that

1+/21gh221l

>1l1++11+/21gh2

211l2


l 11l

>1l1++11+/21gh2

211l2

(30)

Plugging in l (from condition (13)) into the LHS of (30) andrearranging, condition (30) becomes

11l2 +1l

11+/21gh11l

2

1l1

11l2 < 1 (31)

To see that condition (31) holds, note that the first term issmaller than l (follows from (13)), that the second term issmaller than 1l (because gh < l implies that the squaredterm is less than 1), and the third term is negative. Exis-tence: it is established in Proposition 10 of the technical

appendix that a threshold h for which Rhh =Rph exists for =0 and =1. Because h/ 1/3. Thus, we needto show that Rl

1/3. Substitut-

ing the revenue functions and rearranging, it is enough toshow that

1+2

1gh2 > 11gl2Because gh gl, it is enough to show that 1+/2> 1 ,which holds for > 1/3.




17/17


Proof of Theorem 8. First note that l0= and thatlimvlv lvl=0. Differentiating lvl, we obtain

dlvl

dvl=c

w

v2l+fvl

Ml F vl


A maximum exists if there exists a vsl < v such thatlvsl=0 and lvsl0. Requiring that lvsl0 is equiv-alent to having

lvl

F vl= l

EVVvlvlc

2

vl F vl0

for some vl. Assume =0. Then, if l/EV>c , theremust be a solution with positive profit. Let Mllvsl. From the Envelope Theorem, we haveMl/=2/vsl < 0, which means that lvsl isdecreasing in . Note that even though we have not ruledout the existence of several local maxima vsl , lvslis decreasing in at every critical point. This implies thatthere exists some l such that lvsl for

l. Oth-

erwise, there does not exist an optimal vsl < v.Furthermore, denote the RHS of (9) by zvl, i.e., zvl=

l/ chvl1/2 F vl1/2. We want to show that thereexists a unique vsl that maximizes profit and solvesvsl =zvsl. Because F is IFR, zvl is increasing. Differentiat-ing zvl, we get

zvl =

2

chvl

l

chvl3/2

F vl1/2

+f vl

l

chvl1/2

F vl3/2

Plugging in (9), we get

zvsl =1

2

hvslvsl +c

hvsl vsl

l/ c hvsl

A sufficient condition for zvsl to be increasing is for bothterms in the brackets to be increasing. The first term is thegeneralized failure rate. It is increasing if F is IGFR. Thesecond term is increasing if hvslvsl is increasing and F isIFR. Thus, under these conditions, zvsl is increasing andconvex and there are at most two solutions to vl =zvl.Because l0 < 0 and

l0 > 0, the smallest solution is the

maximum.

Proof of Theorem 9.First note that h0= and thatlimvhv hvh=0. Differentiating hvh, we obtain

dhvh

dvh =c w

v2h +Mhf vh/2Mh F vh/2


A maximum exists if there exists a vsh < v such thathvsh=0 and hvsh0 Requiring that hvsh0 isequivalent to having

hvh

Mh F vh/2=EV Vvhvh c

Mh/2

2

vh F vh0

for some vh. Assume =0. Then, if EV>c, there mustbe a solution with positive profit. Let Mhhvsh.From the Envelope Theorem, we have Mh/ =

2/vsh < 0, which means that hvsh is decreasingin . Note that even though we have not ruled out the exis-tence of several local maxima vsh, hvsh is decreas-ing in at every critical point. This implies that there existssome h such that hvsh for h. Otherwise, theredoes not exist an optimal vsh < v.

Furthermore, denote the RHS of (10) by zvh, i.e., zvh

=/Mh/21chvh1/2 F vh1/2. Note that zvh=zv, where zv is the RHS of (7) from Theorem 5 and is a positive constant. Therefore, the remainder of the proofis the same.

References

Afche, P., H. Mendelson. 2004. Pricing and priority auctions inqueueing systems with a generalized delay cost structure. Man-agement Sci. 50(7) 869882.

Barro, R. J., P. M. Romer. 1987. Ski-lift pricing, with applications tolabor and other markets. Amer. Econom. Rev. 77(5) 875890.

Berglas, E. 1976. On the theory of clubs. Amer. Econom. Rev. 66(2)

116121.Bitran, G., P. Rocha e Oliveira, A. Schilkrut. 2008. Managing cus-

tomer relationships through pricing and service quality. Work-ing paper, Massachusetts Institute of Technology, Cambridge.

Chen, H., M. Frank. 2004. Monopoly pricing when customersqueue. IIE Trans. 36(6) 569581.

Clay, K. B., D. S. Sibley, P. Srinagesh. 1992. Ex post vs. ex antepricing: Optimal calling plans and tapered tariffs. J. RegulatoryEconom. 4 115138.

DeGraba, P. 1995. Buying frenzies and seller-induced excessdemand. RAND J. Econom. 26(2) 331342.

De Vany, A. 1976. Uncertainty, waiting time, and capacity utiliza-tion: A stochastic theory of product quality. J. Political Econom.84(3) 523542.

Edelson, M. M., D. K. Hilderbrand. 1975. Congestion tolls for Pois-

son queueing processes. Econometrica 43(1) 8192.

Essegaier, S., S. Gupta, Z. J. Zhang. 2002. Pricing access services.Marketing Sci. 21(2) 139159.

Giridharan, P. S., H. Mendelson. 1994. Free-access policy for inter-nal network. Inform. Systems Res. 5(1) 121.

Hassin, R., M. Haviv. 2003. To Queue or Not to Queue: EquilibriumBehavior in Queueing Systems. Kluwer Academic Publishers,Boston.

Littlechild, S. C. 1974. Optimal arrival rate in a simple queueingsystem. Internat. J. Production Res. 12(3) 391397.

Masuda, Y., S. Whang. 2006. On the optimality of fixed-up-to tar-iff for telecommunications service. Inform. Systems Res. 17(3)247253.

Mendelson, H. 1985. Pricing computer services: Queueing effects.

Comm. ACM 28(3) 312321.Miravete, E. J. 1996. Screening consumers through alternative pric-

ing mechanisms. J. Regulatory Econom. 9 111132.

Naor, P. 1969. The regulation of queue size by levying tolls.Econometrica 37(1) 1524.

Randhawa, R., S. Kumar. 2008. Usage restriction and subscriptionservices: Operational benefits with rational customers. Manu-

facturing Service Oper. Management 10(3) 429447.

Scotchmer, S. 1985. Profit-maximizing clubs. J. Public Econom. 27(1)2545.

Xie, J., S. M. Shugan. 2001. Electronic tickets, smart cards, andonline prepayments: When and how to advance sell. MarketingSci. 20(3) 219243.



pricing services subject to congestion: charge per-use fees or sell subscriptions by gerard cachon...

Documents