BISE – RESEARCH PAPER

Empirical Evaluation of Fair Use Flat Rate Strategiesfor Mobile InternetTariffs constitute important decision making parameters in the marketing mix of mobilephone companies. Flat rates, as an example of such a pricing model, are decoupling thecustomers’ usage and the generated revenue. This leads to commercial risks fortelecommunication providers. The current price level for a data flat rate in conjunction withcurrent technologies and usage patterns leads to high production costs and negativecontribution margins. As an alternative concept, fair use flat rates lead to a limitation of usewhile also satisfying the typical customer usage patterns. Therefore, they are preferable totraditional flat rates. New production technologies such as LTE will not change this situationsince they do not make customers more willing to pay. Instead, the motivation for theintroduction of LTE is based on an improved cost situation for the telecommunicationprovider.

DOI 10.1007/s12599-011-0172-6

The Authors

Dipl.-Kfm., Dipl.-Komm.-WirtMarcel FritzDetecon Inc. (USA)128 Spear StreetSan Francisco, CA 94105USAmail.mfritz@gmail.com

Prof. Dr. Christian Schlereth (�)Assistant Professor for Marketingand Electronic ServicesFaculty of Economics and BusinessAdministrationGoethe University FrankfurtGrueneburgplatz 160323 FrankfurtGermanyschlereth@wiwi.uni-frankfurt.de

Dr. Stefan FiggeT-Mobile Stiftungsprofessurfür Mobile Business & MultilateralSecurityFaculty of Economics and BusinessAdministrationGoethe University FrankfurtGrueneburgplatz 160323 FrankfurtGermanysfigge@me.com

Received: 2010-08-30Accepted: 2011-04-15Accepted after two revisionsby Dr. Bub.Published online: 2011-08-13

This article is also available in Ger-man in print and via http://www.wirtschaftsinformatik.de: Fritz M,Schlereth C, Figge S (2011) Em-pirische Evaluation von Fair-Use-Flatrate-Strategien für das mobileInternet. WIRTSCHAFTSINFORMATIK.doi: 10.1007/s11576-011-0284-0.

Electronic Supplementary MaterialThe online version of this article(doi: 10.1007/s12599-011-0172-6)contains supplementary material,which is available to authorizedusers.

© Gabler Verlag 2011

1 Introduction

With yearly revenues of almost 150 bil-lion Euros, the information and telecom-munication sector is one of the largest in-dustries in Germany (BMWI 2010). Theavailability of broadband Internet accessover the mobile telephone network is es-pecially seen as a pillar of hope within thismarket. These services known as “mo-

bile Internet” are supposed to open upnew sources of revenue and as a resultwill compensate for the difficult situa-tion in the voice communication seg-ment, which has been affected by satura-tion and price competition.

However, the demand for mobile In-ternet initially fell short of the serviceproviders’ expectations for a while. Oneparticular reason for this was the priceskimming strategy used by network op-erators as well as the low data transferrates in relation to the fixed-line network(Delaney 2009). These adoption barrierswere only overcome with the introduc-tion of flat rate tariffs, a decline in theprice level, and the roll-out of more effi-cient transmission technologies. The de-mand for mobile Internet is now increas-ing rapidly. For the period from 2008 to2013 an average yearly growth of 46.8%in the number of users is expected. An av-erage growth of 89.3% a year is predictedin data traffic and thus network load forthe same period of time (Informa 2009).

This worldwide observed trend createsa fundamentally new situation for thenetwork operators: Based on the lim-ited capacity of the network infrastruc-ture and mobile phone frequencies, thesatisfaction of additional demand nowa-days also requires a constant expansionof the network capacity. As a result thenetwork operator faces direct productioncosts that depend on the consumed datavolume. The postulate of “marginal costsnear to zero” (Shapiro and Varian 1998)

Business & Information Systems Engineering 5|2011 269

BISE – RESEARCH PAPER

from Internet Economics is therefore nolonger given in the case of the mobile In-ternet.

In order to account for these produc-tion costs and to offer mobile Internetprofitably, the creation of a tariff port-folio (i.e., a menu of pricing plans) as aregulator between demand and costs isof central importance. Recently networkoperators are rising to the challenge, es-pecially with the offer of so-called fair useflat rates. In this article, we understandfair use flat rates as a tariff with a fixedmonthly price, but for which the trans-fer speed is restricted after a specified vol-ume limit (“cap”) is exceeded. As a re-sult, there are no extra charges for ex-ceeding the volume limit, but frequentusage and the cost risk to the provider isrestricted. Thus, at present the monthlyprice of a leading German telecommuni-cation company is €39.95 for a 5 GB vol-ume limit and a speed of 7.2 Mbit/s.

The determination of price, volume,and speed of a fair use flat rate is a verycomplex economic problem (Marn et al.2004). For example, a decrease in the vol-ume limit leads to a lower usage on av-erage, but also to a decrease in the num-ber of customers and revenue. With re-gard to the international introduction ofthe next generation of mobile networks,Long Term Evolution (LTE), further cen-tral questions are how the added value toconsumers due to the increased networkspeed as well as the improvement in theoperators’ cost situation should ideally bereflected in the price.

This paper aims to identify the optimalprice structure for fair use flat rates. Forthis purpose, we suggest the use of a sim-ulation model which demonstrates quan-titatively the effect of different strategieson the contribution margin as an aid todecision making. Unlike the existing and,in most cases, purely analytical work onthe determination of optimal prices (e.g.,Schade et al. 2009; Png and Wang 2010),which assumes that customer preferencesare known, this paper uses discrete choiceexperiments to empirically estimate con-sumers reaction to the offer of fair use flatrates.

Section 2 first presents the state of sci-entific research in the fields consideredin this paper. Section 3 investigates thetechnological changes of mobile Internetproviders due to the introduction of LTE.Section 4 presents the data collection andsimulation study as the core of the paper.Section 5 concludes the work with a crit-ical discussion of the results as well as anoutlook for further research questions.

2 State of Research

The theories relevant for this paper arederived from the interpretation of mo-bile broadband access as a form of aninformation and communication system(ICS) which is designed to be offered asa product on the market. Mobile broad-band access, as well as Software-as-a-service or other digital value-added ser-vices, therefore constitutes an instance ofthe class “ICS as a service on the mar-ket”. With the introduction of such ICS,problems with acceptance often becomeapparent which point to methodical defi-ciencies in the design (GI 2010). In orderto eliminate these acceptance problems,business information systems engineer-ing (BISE) generally focuses on the de-velopment and implementation of meth-ods for requirement analysis or require-ment engineering and ultimately also onbeing able to early measure and test theperceived benefit of the product from theviewpoint of the user before its launchonto the market (GI 2010).

Based on this understanding, the fol-lowing three scientific areas are of par-ticular relevance: (1) Interdisciplinaryexplanation approaches of the InternetEconomy, (2) the discussions about busi-ness models led by business administra-tion and BISE, and (3) the methods forcollecting data on customer preferencesand their willingness to pay as developedin the fields of marketing and psychology.

2.1 Internet Economy

The Internet economy deals with the eco-nomic implications of the commercialuse of the Internet, i.e. the production,distribution, and consumption of infor-mation goods (Shapiro and Varian 1998).The focus is typically on the non-accessbusiness, which is the provision of ser-vices and content. However, the optimalpricing strategy of mobile Internet accessis also an elementary component of theinformation system “Internet” and thusis to be analyzed in this context. The fun-damental works of Shapiro and Varian(1998) and Zerdick et al. (2001) iden-tify the central questions for this pur-pose. These include not only technical as-pects of implementation but also micro-economic discussions about its distribu-tion and consumption. The objective, interms of requirements engineering, is thetheory-based analysis and formulation ofmore appropriate business models as a

design pattern for use by market partic-ipants in the Internet economy.

In order to describe business models,the works of Stähler (2001) and Gordijnet al. (2005) develop suitable technicalterminology. There is widespread agree-ment on the fact that the central areas ofa business model represent the architec-ture of the added value, the range of ben-efits, and the revenue models. While thefirst is necessary to describe the produc-tion of information goods, the mergingof the offered benefits and the revenuemodel is a critical factor for the purchaseand utilization decision of customers. Inthe telecommunications industry, directrevenue models have prevailed over indi-rect revenue models funded by advertis-ing, so that customers have to choose be-tween different types of tariffs.

2.2 Revenue Models of ISPs

Until now fair use flat rates, as rele-vant tariff concepts of the telecommu-nications industry, have been largely ne-glected in research. Skiera (1999) inves-tigates the use of flat rates, pay-per-use,and multi-part tariffs consisting of a fixedfee and a price per unit. These tariffsare aimed at customers with varying de-mand patterns – for example, frequentand infrequent users – so that the paidmarginal price eventually depends on theusage and the chosen tariff. Lambrechtand Skiera (2006) show that customershave different preferences for these typesof tariffs and that they prefer flat ratesrather than multi-part tariffs even if theypay a higher price for the same service.On the one hand, the price payable atthe end of the month remains constantso that the costs for the service can beplanned. On the other hand, customersare unsure about the extent of usage (De-laney 2009) and flat rates can more eas-ily be compared than multi-part tariffs.Customers also mentally record the costof a flat rate at the beginning of the pe-riod. Since then no marginal costs are in-curred at the time of consumption, theconsumption that has already been paidfor can be enjoyed to a greater extent(Lambrecht and Skiera 2006).

In the case of high variable costs, theflat rate proves to be problematic becausethe customer has no incentive to limittheir use, which leads to a massive in-crease in network load in the mobile In-ternet, loss of service quality, and hencealso costs for the service provider.

Fair use flat rates combine the ben-efits of classical flat rates as perceived

270 Business & Information Systems Engineering 5|2011

BISE – RESEARCH PAPER

by customers with the possibility for theprovider to restrict user behavior. Thetariffs proposed by Altmann and Chu(2001) are closest to these as they aredifferentiated by price and quantity aswell as quality of service (QoS) – thus inthe case of the mobile Internet they aredifferentiated by the data transfer speed.However, the authors distinguish a freeversion of the service with a very low QoSand a paid for version with a high QoS,and it remains unclear how tariffs can bealigned with the heterogeneous demandbehavior.

2.3 Measurement of Customer DemandBehavior

We distinguish three empirical ap-proaches to determine customer behav-ior (Völckner 2006): The use of (1) expertopinion, (2) transaction data (e.g., Lam-brecht et al. 2007), and (3) survey data(e.g., Iyengar et al. 2008). Expert opinionis an important source of informationobtained from people who have extensiveexpertise and experience on a particularsubject. However, accuracy is not guar-anteed and results are often subjective.Expert opinion does nevertheless offerconsiderable benefits if the data are usedsynergetically to evaluate the recommen-dations from other data sources.

In contrast, transaction data have ahigh external validity because they arebased on real purchase decisions. How-ever, the price usually varies only slightlyand non-purchase decisions are not ob-served (Swait and Andrews 2003). Fur-thermore, they do not exist at the timeof launch of a service and the effects ofa slightly higher or lower price are hardto measure on demand. This in turn re-sults in the fact that the customers’ truewillingness to pay remains unknown.

If transaction data are not available,it is advisable to use survey data, e.g.,using discrete choice experiments (seeSect. 4.1). Survey data have the advan-tage over transaction data that the re-search subject being analyzed does nothave to exist in reality and observationscan be collected at low costs (Swait andAndrews 2003). In addition, the ana-lyst may exert greater influence on thedata generating process of the studies andcan investigate important central prob-lems in a more targeted manner. Surveydata are frequently criticized due to theirlower external validity since participants

only make hypothetical decisions whichhave no direct consequences on their ac-tions (Völckner 2006). Researchers cur-rently face this criticism by combiningincentive-compatible mechanisms, suchas the Becker-DeGrot-Marschack mech-anism (Wertenbroch and Skiera 2002),with the survey or they do not directly askthe participants about their willingnessto pay (e.g., Ding 2007). Instead, theymake choice decisions, such as in discretechoice experiments, which are compara-ble to real purchasing situations. Multi-part tariffs have only been investigated inthe study by Iyengar et al. (2008), whichevaluates tariffs with the help of surveydata.

3 Technological Conditions forthe Mobile Internet

The simulation study presented in Sect. 4is used as decision support in order tobring technology-related costs and tariff-moderated demand for mobile Internetaccess in line with each other. The tech-nological aspects relevant to this are pre-sented in the following.

3.1 Established TechnologicalEnvironment: UMTS and HSPA

The foundation for the currently avail-able mobile Internet access was laid downwith the roll-out of UMTS in 2003.The initial maximum possible data trans-fer rate with UMTS of 384 Kbit/s wasstill well below the speed of comparablefixed lines. This adoption barrier was fi-nally overcome with the worldwide ini-tiated roll-out of the technological stan-dard High Speed Packet Access (HSPA)in 2007. It enables transfer rates of upto 7.2 Mbit/s, comparable to a DSL fixedline (Dahlman et al. 2008).

German network operators are alreadyplanning the launch of the next mobilegeneration. First of all, a double in thepossible transfer rates to 14.4 Mbit/s isexpected with the upgrade from HSPAto HSPA+. All the aforementioned tech-nologies encompass the principles of re-source sharing and packet-based datatransfer. This means that the resourcesrequired for data transfer are not re-served exclusively for one connection,but rather all the information packetsare processed in parallel according to the

best-effort principle. Consequently, theavailable data rate per customer decreaseswith each addition to the network load.Thus, the current average transmissionperformance achieved with the use ofHSPA is about 1 Mbit/s and thereforewell below the theoretical maximum of7.2 Mbit/s.

In line with the decreases in capacityand speed, costs incur for the networkoperator when providing mobile Inter-net access. These costs can be directly at-tributed to the individual units of service.In the short term, opportunity costs in-clude the reduction in service quality andcustomer satisfaction. In the medium tolong term, investments for capacity ex-pansion of the radio access, backhaul,or core network should be consideredas step-fixed costs. In addition to thesecongestion-related network costs (capac-ity costs), electricity and maintenancecosts for the signal transfer (network-related operating costs) are recognized.In total, it is assumed that the productioncosts for the transfer of one megabyteof data volume over a mobile networkin Germany in 2009 amount to 2 to 3cents and are therefore ten times higherthan in the fixed line network (Informa2009).

After having had overcome differentadoption barriers the mobile Internet hasbeen able to report a huge growth indemand in Germany. In particular theuse of mobile computers such as lap-tops or netbooks has also greatly in-creased. Here, mobile access is used ei-ther as a supplement or as a completereplacement for the domestic fixed line.With the increasing dispersion, though,the established mobile technologies aremoving evermore towards their limits,in both technological and commercialterms.

3.2 The Pillar of Hope: LTE

While HSPA technology is based largelyon UMTS, the transition towards LongTerm Evolution (LTE) developed by the3rd Generation Partnership Project asan international standard (Dahlman etal. 2008)1 causes significant technolog-ical changes. Although LTE is designedto coexist with the previous technologies,there are still vast differences both in thearchitectural approach and in central ra-dio interfaces for mobile technologies. As

1See also http://www.3gpp.org/-Industry-White-Papers.

Business & Information Systems Engineering 5|2011 271

BISE – RESEARCH PAPER

Fig. 1 Relationshipbetween utility andselection decision

a consequence and opposed to its pre-decessors, LTE allows both a more flex-ible spectrum allocation and also betterspectrum efficiency. Thus, the capacity oftransmission towers equipped with LTEincreases by a factor of three comparedto HSPA. Capacity constraints can thusbe resolved cost-effectively and for thetime being by technologically upgradingexisting transmission towers. It shouldbe added that a simpler, more flexiblearchitecture reduces administrative ex-pense of an LTE network, compared toHSPA. Taken together, these effects areconsidered to cause a reduction in pro-duction costs of LTE-based mobile Inter-net access.

Besides these advantages in terms ofcosts, improvements also arise from acustomer perspective. With the maxi-mum data transfer rate, it is expectedthat there will be initial values of around20 Mbit/s with an increase in subsequentyears. Additionally, with LTE the latency,which is the waiting time between a callto Internet content and its transmission,is also decreased and thus the user experi-ence is brought closer to the one of a DSLfixed Internet line.

Since the auctioning of spectrums forLTE has already been conducted in manyEuropean countries, mobile operators arecurrently carrying out network expan-sion. It is likely that the commercial avail-ability of LTE in Germany will be an-nounced by the end of 2011 at the latest.While sufficient transparency can be as-sumed with regard to the achievable costsavings with LTE, the question remainshow much customers are willing to payfor LTE and what the impact is on the tar-iff design. This question is followed up inthe next section.

4 Data Collection and SimulationStudy

After having described the real market forthe mobile Internet in the preceding sec-tion, this section is devoted to the modelconstruction and analysis. For this pur-pose, first the methodology for the collec-tion of customer preferences is presentedand the results are explained. Followingthis, the preferences will be used in a sim-ulation study to identify the effects of dif-ferently structured fair use flat rates oneconomic success.

4.1 Discrete Choice Experiments

The discrete choice experiment (Louviereet al. 2000) has established itself as an im-portant data collection method for mea-suring the customer preferences in a vari-ety of disciplines, such as marketing, psy-chology, or health care. Discrete choiceexperiments have a firm foundation insociology and behavioral research and areknown for being able to explain actualpurchasing behavior very well (Swait andAndrews 2003).

The participants repeatedly choosetheir preferred alternative in a choice set(see Fig. 1), which is modeled on realdecision-making situations (such as tar-iff selection). A choice set consists of sev-eral profiles which are described by fea-tures and their characteristics. Thus, inevery choice set trade-off decisions mustbe made; this means choosing betweendifferent attractive combinations of fea-ture characteristics which in turn allow todraw conclusions about the preferencesof participants.

We employ random utility theory(Louviere et al. 2000) to analyze con-sumers’ choices: it assumes that a partici-pant h decides for the profile i that pro-vides him with the highest latent – i.e.,

not directly observable – utility uh,i. Theterm utility is used here as a quantita-tive measure of the satisfaction of needsand consists of a deterministic compo-nent vh,i and a stochastic component, theerror term εh,i. The deterministic com-ponent is calculated from the subjectiveutility βh ·Xi (vector of utility parametersof the participant h multiplied with thedesign vector of the i-th product) minusthe perceived costs �h · pi (price param-eter multiplied with price).

uh,i = vh,i + εh,i

= βh · Xi − �h · pi + εh,i

(h ∈ H, i ∈ I). (1)

Based on the commonly used assump-tion of the Gumbel-distributed errorterm (Louviere et al. 2000), the Logitmodel can be formed. According to Train(2009), the differences compared with al-ternative distributions, such as the nor-mal distribution, are negligibly low. Thestrength of the Logit model as opposedto, for example, the Probit model result-ing from the normal distribution, is thatit is mathematically simple and easy tointerpret. This allows the selection prob-ability Prh,i of the person h for profile i inchoice set Ca to be described by:

Prh,i = exp(vh,i)

×(

exp(vh,0) +∑j∈Ca

exp(vh,j)

)−1

(h ∈ H, i ∈ I), (2)

Equation (2) also takes into account anon-purchase option vh,0 = 0, in case acustomer h decides against all the tariffsoffered. The individual parameter dis-tributions for each participant that ex-plain the observed behavior can be es-timated with the help of the Hierarchi-cal Bayes model. The fact that this is

272 Business & Information Systems Engineering 5|2011

BISE – RESEARCH PAPER

Table 1 Characteristics of the features of mobile Internet access

Level 1 Level 2 Level 3 Level 4 Level 5 Level 6 Level 7 Level 8

Volume limit (in GB) 0.5 1.0 3.0 5.0 10.0 Unlimited

Speed (in Mbit/s) 3.6 7.2 20.0

Monthly price (in €) 10 15 20 25 30 35 40 45

Table 2 Results of the estimation

Value Importanceweight

Constants

Average 3.18

Std. deviation (3.86)

Volume limit (in GB) 0.5 1.0 3.0 5.0 10.0 Unlimited 45.22%

Average −6.72 −2.39 1.16 1.77 2.55 3.63

Std. deviation (3.04) (1.68) (0.96) (1.06) (1.57) (2.57)

Speed (in Mbit/s) 3.6 7.2 20.0 13.99%

Average −1.88 0.55 1.33

Std. deviation (1.32) (0.47) (1.03)

Monthly price 40.79%

Average 0.27

Std. deviation (0.17)

possible despite the low number of ob-servations (16 in the following study) isone of the main advantages of Hierarchi-cal Bayes compared to alternative meth-ods (see Online Appendix). For this pur-pose, more than several thousand itera-tions of behavior patterns of the popula-tion of participants are identified whichserve to enrich the data of an individualparticipant. Hierarchical Bayes is there-fore extremely computationally intensiveand has only been used since a few yearsbecause of the increased performance ca-pability of modern computers. A detaileddescription of the computation can befound in Gensler (2003) and the relatedassumptions are made in Chandukala etal. (2007).

4.2 Construction of the study

In order to develop the survey design, themost important characteristics of mo-bile Internet were identified (see Table 1)using a market analysis and a prelimi-nary study, which applied the “dual ques-tioning” technique (Myers and Alpert1968). While the speeds 3.6 Mbit/s and7.2 Mbit/s already exist on the mar-ket today with HSPA, 20 Mbit/s can beachieved solely through the introductionof LTE. The questionnaire of the main

study consisted of four sections: The firstsection collected information on the In-ternet usage patterns of the participants,in particular the average monthly dataconsumption, but also the number ofhours spent, for example, for surfing onwebsites, movies, online games, or mu-sic. In the second section, the choice deci-sions in the choice sets were made. In thethird section, we asked how much par-ticipants expected to exceed or fall be-low the monthly volume limit for two of-fers with a speed of 7.2 Mbit/s and a vol-ume limit of 1 GB and 5 GB. In the lastsection, we collected demographic andsocio-economic information.

A major challenge is the creation ofan efficient choice design (Street andBurgess 2007). For this purpose, the tech-niques in Street and Burgess (2007) wereapplied and a D-optimal (6 ·3 ·8) full fac-torial design with 18 choice sets was gen-erated. These designs are known for theirhigh efficiency and their suitability for adiverse range of research designs. Eachchoice set shows three different tariffs formobile Internet and a non-purchase op-tion (see Fig. 1). The observations from16 of 18 choice sets are included in theestimation and the remaining two choicesets are used to test the predictive validity.

4.3 Survey Results

An online survey conducted in Septem-ber 2009 generated 270 completed ques-tionnaires. The sample was collected byan online panel provider so that the com-position of the sample is representativefor the German population with respectto age and sex (see Online Appendix).The respondents showed moderate tostrong interest in mobile Internet and37.02% chose one of the top two cate-gories on a 5-point Likert scale.

According to the respondents’ self-assessment, usage is strongly orientedtowards the consumption of the offeredvolume limit. 74.39% (87.54%) of par-ticipants claimed that they would notexceed the volume limit of a fair use flatrate with a 1 GB (5 GB) cap. Within thisgroup, 44.64% (34.60%) of participantssaid they would align their usage exactlywith the limit. 20.76% (33.22%) wouldmake use of up to 200 MB (1 GB) less and9.00% (19.72%) would use considerablyless.

For the simulation, the customer pref-erences are estimated on an individ-ual level with the Hierarchical Bayesianmethod. All features are effect-coded, ex-cept the price, which is subject to a vec-tor model. A burn-in phase of 20,000 it-erations was chosen for the estimation sothat the system converges in the relevantfield of information. The estimated pa-rameters are based on the analysis of afurther 20,000 iterations.

The results of the estimation and theimportance weightings (see e.g., Gensler2003) of the properties are listed in Ta-ble 2. All values are plausible and have theexpected sign. The strictly monotonousmode in which utility in the volume lim-its and speeds increases, indicates highface validity. The most important fea-ture for the participants is the volumelimit (importance weight of 45.22%), al-most equivalent to the price (importanceweight of 40.79%). The speed has thelowest importance weight (13.99%). It isnoteworthy that an unlimited monthlyvolume limit only offers an insignificantadded value compared to the 10 GB limit.It is even more remarkable that the in-crease in benefit between the speed of3.6 Mbit/s and 7.2 Mbit/s is valued muchhigher than that between 7.2 Mbit/s and20 Mbit/s.

To assess the validity of the results, weconsult the share of choices which wereindividually correctly predicted with animputed first-choice model. The choicedecisions are correctly forecasted in

Business & Information Systems Engineering 5|2011 273

BISE – RESEARCH PAPER

Table 3 Tariffs and data consumption

Tariff A0

(Reference)Tariff A1 Tariff A2 Tariff A3 Tariff A4 Tariff A5 Tariff A6 Tariff B Tariff C

Simulated tariffs

Volume limit (in GB) 5.0 1.0 1.0 3.0 5.0 10.0 Unlimited 5.0 5.0

Speed (in Mbit/s) 7.2 3.6 20.0 7.2 20.0 7.2 20.0 7.2 3.6

Monthly price (in €) 40 20 30 30 40 50 60 40 25

Estimated data consumption

Average (in GB) 1.82 0.62 0.73 1.26 1.98 2.29 2.71

Std. deviation (1.33) (0.24) (0.26) (0.76) (1.42) (2.15) (3.13)

86.40% of all choice sets used for the esti-mation and in 71.63% of the two remain-ing choice sets not considered in the es-timation. Therefore, both values are wellabove the 25% chance criterion of ran-dom choice. Therefore, we conclude thatour model is an adequate one (e.g., Figgeand Theysohn 2006; Schlereth and Skiera2009) and that the data are suitable forthe following simulation.

4.4 Set-up of the Simulation

For demonstration purpose, differentstrategies for using fair use flat rates shallbe investigated on the basis of the col-lected customer preferences. In addition,the extent to which the reduction of vari-able costs, caused by innovations such asLTE, affects the optimal design of the tar-iff offer is to be analyzed as important in-formation for an investment decision.

For this purpose, a simulation modelis applied which integrates informationabout the major factors that influenceprice determination: Customer benefit,competitive prices, and costs of the com-pany (see Simon and Fassnacht 2009).The decision alternatives are thereforevalued while taking into account all directprice-relevant information at the sametime. Simulations of this kind in price re-search are considered to be particularlypowerful, but at the same time extremelychallenging (see Wiltinger 1998).

Formally, the model is described asthe functional relationship E = f (X,Y)

between objective criteria (E), decisionalternatives (X), and environmental pa-rameters (Y) (see Hanssmann 1993). Thestarting point of the model is telecom-munications provider A who wants tomaximize the contribution margin (E)

by adjusting their currently offered fairuse flat rate A0. As a secondary objec-tive, efforts will be made to increase thenumber of customers. The pricing strate-gies (X), which are exemplary evaluated,

Fig. 2 System relationships

are listed in Table 3 in the form of tar-iffs A1 to A6 (X).

The environmental parameters of themodel are the marginal production costsof the supplier, competitive offers, andthe distribution of data consumption toconsumers. The competitive situation inGermany is reflected more simply by con-sidering two static competing offers Band C. The distribution of data con-sumption is parameterized by the po-sition (“average per capita consump-tion”) and shape (“asymmetry”) of a log-normal distribution in the model.

The relationship between these modelparameters is illustrated in Fig. 2 (basedon Skiera (2010); for more complexstructural models see, for example, Reissand Wolak 2007). The offer of serviceprovider A determines the customer ben-

efit and thus the demand according to thetariff offered by competitors (1). The set-up of the tariff factors influences both theamount of customers and their Internetusage behavior (Iyengar et al. 2008). Thelatter occurs in two different ways: First,different tariffs address different typesof users (for example, frequent or infre-quent users). Second, the tariff ex postmoderates the usage behavior of cus-tomers as they adjust the amount theyuse to fit the tariff they have purchased.

The data traffic (2) is derived from theamount of customers and their usage be-havior, and it is moderated by the ex-pected distribution of the data consump-tion (Skiera 2010). Furthermore, the datatraffic is directly affected by the tariff fea-tures volume limit and speed (3): Thespeed in the tariff moderates the indi-

274 Business & Information Systems Engineering 5|2011

BISE – RESEARCH PAPER

Fig. 3 Simulation results

vidual data consumption per month, andthe volume determines its monthly up-per limit.2 The price-induced costs resultfrom the product of data traffic and pro-duction costs per unit volume (4). Theproduct of the number of customers (5)and the price (6) provides the achievedrevenue (Reiss and Wolak 2007).

The specific modeling of the revenueside and cost side is explained in the fol-lowing: The selection probability of thecustomer h for tariff Ai is calculated in (3)taking into account competitors B and C(Draganska et al. 2010) by:

Prh,Ai = exp(vh,Ai)

× (exp(vh,0) + exp(vh,Ai)

+ exp(vh,B) + exp(vh,C))−1

(h ∈ H,Ai ∈ I). (3)

The revenue is the product of the prob-ability of that tariff Ai is chosen and themonthly price: UAi = ∑

h∈H Prh,Ai ·pAi .An overall market potential of 10 millionconsumers was taken in order to scale thesample to the size of the total Germanmarket (Informa 2009).

To model the cost side, the subjectswere first arranged into an ordinal rank-ing according to the estimated consump-tion data on the use of fixed Internet onthe basis of their statements in the firstpart of the questionnaire. Then, the dataconsumption was determined by draw-ing random numbers from a log-normaldistribution using the principle of in-version of the probability transformation(Liebl 1995) and assigned to the sub-jects. The parameters of the log-normal

distribution were estimated as μ = 0.7and σ = 1.6. The resulting distributionscorrespond to previous empirical stud-ies on Internet usage (Hatton 2008; Pers-son 2010). In the third step, the data con-sumption of a subject concerning theiruse of mobile Internet was determinedas an individual percentage of their useof fixed Internet for each case. For this,we used the answers in the first andthird part of the questionnaire. In ad-dition, we assumed that a tariff with aspeed of 3.6 Mbit/s (20 Mbit/s) leadsto data consumption being ten percentlower (higher) than with 7.2 Mbit/s.

4.5 Simulation Results

The simulation results are summarized inFig. 3. The left half of the figure illustratesthe simulation results for the case of vari-able costs of 2 cents per MB. The righthalf of the diagram shows the sensitivityof the simulation results when the vari-able costs are altered.

Figure 3 shows that the starting tariffof operator A is not optimal and onlyachieves a slightly positive contributionmargin. Starting from the basic rate A0,the maximum contribution is achievedthrough a price reduction while reduc-ing the tariff volume at the same time(tariffs A1, A2, and A3). Thus, for exam-ple, by reducing the monthly price from€40 to 30 in conjunction with a reduc-tion in the volume limit from 5 GB to3 GB, the observed operator increasestheir contribution to€60 million per an-num, while achieving a 30 percent highermarket share (tariff A3).

The introduction of a higher transmis-sion speed of 20 Mbit/s will only lead toimproved results in the model if the fac-tors price and volume are adjusted at thesame time. Otherwise, the achieved con-tribution margin decreases (tariff A4),since consumers’ additional willingnessto pay for the higher transmission speedis not sufficient to compensate for the ex-tra data consumption. The offer of a trueflat rate (tariff A6) proves to be particu-larly insufficient in the simulation.

The right half of Fig. 3 shows the sen-sitivity of the achieved contribution mar-gin to the marginal cost of production forbasic rate A0, as well as the tariffs with thehighest (tariff A3) and lowest (tariff A6)

contribution margin in the scenario ofvariable costs of 2 cents per MB. Thisshows that tariff A3 achieved the high-est contribution margin of the three ob-served rates, almost independently of thelevel of the variable costs. Furthermore,it is evident that the unlimited flat rate(tariff A6) will be profitable at produc-tion costs below 1.6 cents per MB. Withvariable cost of less than 1.4 cents per MBit leads to a higher contribution marginthan the base strategy (tariff A0).

5 Conclusion

5.1 Discussion of Results

Based on the simulation results, threestrategic recommendations for telecom-munications providers can be derived.First, the model confirms that traditionalflat rates with unlimited usage cannot

2In the model we abstract from data traffic which is caused by the use of a reduced speed (64 kbps).

Business & Information Systems Engineering 5|2011 275

BISE – RESEARCH PAPER

AbstractMarcel Fritz, Christian Schlereth,Stefan Figge

Empirical Evaluation of Fair UseFlat Rate Strategies for MobileInternet

The fair use flat rate is a promising tariffconcept for the mobile telecommuni-cation industry. Similar to classical flatrates it allows unlimited usage at a fixedmonthly fee. Contrary to classical flatrates it limits the access speed oncea certain usage threshold is exceeded.Due to the current global roll-out of theLTE (Long Term Evolution) technologyand the related economic changes fortelecommunication providers, the ap-plication of fair use flat rates needs areassessment. We therefore propose asimulation model to evaluate differentpricing strategies and their contribu-tion margin impact. The key input ele-ment of the model is provided by so-called discrete choice experiments thatallow the estimation of customer pref-erences.

Based on this customer informationand the simulation results, the articleprovides the following recommenda-tions. Classical flat rates do not allowprofitable provisioning of mobile Inter-net access. Instead, operators shouldapply fair use flat rates with a lower us-age threshold of 1 or 3 GB which leadsto an improved contribution margin.Bandwidth and speed are secondaryand do merely impact customer pref-erences. The main motivation for newmobile technologies such as LTE shouldtherefore be to improve the cost struc-ture of an operator rather than using itto skim an assumed higher willingnessto pay of mobile subscribers.

Keywords: Mobile Internet, Discretechoice experiments, Fair use flat rates

be provided profitably at current marketprices and the current cost level of HSPA.Therefore, the cancellation of these tar-iff models is advisable for every telecom-munication provider and respective tar-iff decisions have already taken place onthe market. With a reduction in produc-tion costs to below 1.6 cents per MB, aswould be possible through the use of LTE,the simulation model shows a changein this situation. Based on such a costlevel, higher contribution margins can beachieved with an unlimited flat rate thanwith the reference tariff A0.

With regard to the current high vari-able costs, a second recommendation canbe made. That is to differentiate flat ratetariffs more than in the past with thehelp of the segmentation of different us-age levels. This statement is based on thesimultaneous improvement of the con-tribution and market share in tariff A1and tariff A3 (see Fig. 3) which addresscompletely different types of users. Tar-iff A1 with a competitive price pointand a lower transfer volume addressesthe sporadic users of the mobile Inter-net who only access information throughmobile access to a limited extent. Con-sequently, the offered data transfer per-formance is of secondary importance and3.6 Mbit/s is sufficient for this segment.Tariff A3 addresses the frequent users ofmobile Internet. Here, the model showsthat maximum data volumes and datatransfer services are not absolutely neces-sary. The model also shows that 3 GB al-ready addresses the data volume require-ments of most frequent users. The addi-tional offer of Internet access at speeds of7.2 Mbit/s and with a 3 GB volume limitthus promises the best way to increasemarket share and also contribution mar-gin. Despite the general homogeneity ofmobile Internet service, the simulationresults support the finding that a differ-entiation of the services positively affectsthe achieved contribution margin.

The third recommendation relates tothe low importance weighting of the datatransfer speed. In the study, it was foundthat the availability of an LTE enabled20 Mbit/s speed least impacted the pref-erence of customers. At the same timethe increase in speed from 7.2 Mbit/s to20 Mbit/s offers a much smaller addedvalue than the jump from 3.6 Mbit/sto 7.2 Mbit/s enabled by the introduc-tion of HSPA (see Table 2). Hence, thevalue of LTE is caused by the technolog-ical cost benefits for the telecommunica-tions provider rather than by a possible

increase in revenue due to a higher pro-portion of customer preferences relatedto data transfer speed. With the emer-gence of new mobile application fieldswith higher demands on the data trans-fer speed, a change in this situation islikely. Possible candidates for such pref-erence changing applications are, for ex-ample, cloud computing services, mobileinteractive TV, and video calls.

This raises the general question of thetemporal and geographic transferabilityof the results. In a temporal sense, this islimited by the fact that customer prefer-ence structures are instable and dependon the current establishment level of theexamined good (Teichert 2001). In geo-graphical terms, the benefit awarded toconsumers varies depending on the localmarket situation, for instance the pricelevel or availability of substitute goods.Due to very similar development histo-ries of the respective telecommunicationsmarkets, it can be assumed that the re-sults can be transferred to the European,and to a certain extent, the North Ameri-can markets.

For the three recommendations statedit is also to be considered that simplifyingassumptions were made, which shouldbe resolved gradually in future research.For example, static prices were assumedfor competing service providers. How-ever, in reality competition is expectedto be more dynamic (Draganska et al.2010), with the competition adjusting itsprices to newly introduced tariffs untila Bertrand-Nash price equilibrium (Dra-ganska et al. 2010) is achieved. Never-theless, this is not a trivial method ofmodeling, particularly in the context ofservices, since tariff choice and quantitydecisions have to be considered simul-taneously and also because perceptionsof the benefit can change over time. An-other limitation arises from the simpli-fied model of the cost side. For example,we abstracted from a possible cost reduc-tion with an increasing output quantity,we exclusively considered volume-basednetwork costs, and omitted the existenceof purely customer-related costs, such asfor sales commissions, in order to providea clearer presentation.

5.2 Outlook

The aim of this study is to identify theoptimal pricing strategy for fair use flatrates taking into account new technolo-gies such as LTE. Although fair use flatrates are currently mainly applied in the

276 Business & Information Systems Engineering 5|2011

BISE – RESEARCH PAPER

context of mobile Internet, they are alsosuitable for many other Internet-basedservices which have variable costs and canbe differentiated on the basis of QoS. Oneexample are commercial Web servicesthat use and combine fee-based Web ser-vices themselves as a way of creating amash-up service. These services could beoffered at attractive flat rate prices. How-ever, the reduction of certain QoS char-acteristics at high utilization will ensurethat the variable costs do not exceed theachieved revenue.

The use of empirically-supported sim-ulation models can assist in finding thebest possible service and tariff portfoliosand provide additional knowledge for in-vestment decisions. The study sees it-self as a contribution to answering thequestion of private financing of the in-frastructure necessary for an informa-tion society. The problem of intense pricecompetition in the marketing of homo-geneous information and communica-tion services, which is postulated througheconomic theory and can be observed onthe market, presents telecommunicationproviders with the challenge of develop-ing robust revenue sources for financ-ing. Market mechanisms can only makesure that the necessary investments aremade for the development of new in-formation infrastructure if adequate an-swers to these problems are found. Ini-tial problems of market mechanisms run-ning the risk of failure can already be ob-served. For example, experts understandthat the Italian telecommunication net-work is close to collapsing because of thelack of investment in capacity expansion.The same applies to Great Britain and theU.S. (Kort 2010).

Acknowledgements

The authors thank Andreas Albers andMike Radmacher for their helpful com-ments during the preparation of themanuscript. We also thank the threeanonymous reviewers and the editor UdoBub for the extremely constructive team-work and suggestions. The data collec-tion through a panel operator was fi-nancially supported by Detecon. Further-more, parts of the work were producedduring Christian Schlereth’s research stayat the Centre for the Study of Choice(CenSoc) at the University of Technologyin Sydney (UTS). The authors thank Jor-dan Louviere for comments on method-ology and the German research founda-tion for financial support for the stay in

the form of a research grant (GZ: SCHL1942/1-1).

References

Altmann J, Chu K (2001) How to chargefor network services – flat-rate or usage-based? Computer Networks 36(5–6):519–531

BMWI (2010) Informationstechnik undTelekommunikation. Bundesministeriumfür Wirtschaft und Technologie. http:www.bmwi.de/BMWi/Navigation/Wirtschaft/branchenfokus,did=197728.html. Ac-cessed 2011-03-15

Chandukala SR, Kim J, Otter T, Rossi PE, Al-lenby GM (2007) Choice models in mar-keting: economic assumptions, challengesand trends. Foundations and Trends in Mar-keting 2(2):97–184

Dahlman E, Parkvall S, Skold J, Beming P(2008) 3G evolution: HSPA and LTE for mo-bile broadband. 2nd edn. Academic Press,Oxford

Delaney J (2009) Mobile broadband: homeand away. European wireless and mobilecommunications. IDC, London

Ding M (2007) An incentive-aligned mecha-nism for conjoint analysis. Journal of Mar-keting Research 44(2):214–223

Draganska M, Klapper D, Villas-Boas S (2010)A larger slice or a larger pie: an empir-ical investigation of bargaining power inthe distribution channel. Marketing Sci-ence 29(1):57–74

Figge S, Theysohn S (2006) Quantifizierungdes Nutzenangebots eines werbefinan-zierten Geschäftsmodells für den Mo-bile Commerce. WIRTSCHAFTSINFORMA-TIK 48(2):96–106

Gensler S (2003) Heterogenität in der Prä-ferenzanalyse. Gabler, Wiesbaden

GI (2010) Requirements Engineering – Auf-gaben und Ziele. Gesellschaft für Infor-matik. http://www1.gi-ev.de/fachbereiche/softwaretechnik/re/ziele.htm. Accessed2011-03-15

Gordijn J, Osterwalder A, Pigneur Y (2005)Comparing two business model ontologiesfor designing e-business models and valueconstellations. In: Proceedings of the 18thbled eConference, Slovenia

Hanssmann F (1993) Einführung in dieSystemforschung, 4th edn. Oldenbourg,München

Hatton M (2008) Strategies for mobile broad-band pricing and packaging. Analysis Ma-son, Cambridge

Informa (2009) Content & services interactiveforecast tool. Informa Telecoms & Media,London

Iyengar R, Jedidi K, Kohli R (2008) A conjointapproach to multi-part pricing. Journal ofMarketing Research 45(2):195–210

Kort K (2010) Italiens Netz steht vor dem Kol-laps. Handelsblatt 2010-07-08. http://www.handelsblatt.com/unternehmen/it-medien/italiens-netz-steht-vor-dem-kollaps/3484828.html. Accessed 2011-03-15

Lambrecht A, Seim K, Skiera B (2007) Does un-certainty matter? Consumer behavior un-der three-part tariffs. Marketing Science26(5):698–710

Lambrecht A, Skiera B (2006) Paying too muchand being happy about it: existence, causesand consequences of tariff-choice biases.Journal of Marketing Research 43(2):212–223

Liebl F (1995) Simulation. ProblemorientierteEinführung. Oldenbourg, München

Louviere JJ, Hensher DA, Swait JD (2000)Stated choice methods: analysis and appli-cation. Cambridge University Press, Cam-bridge

Marn MV, Roegner E, Zawada CC (2004) Theprice advantage. Wiley, Hoboken

Myers JH, Alpert MI (1968) Determinant buy-ing attitudes: meaning and measurement.Journal of Marketing 32(4):13–20

Persson M (2010) The mobile broadbandconnectivity market. NGT Research Series2010, Götheburg

Png IPL, Wang H (2010) Buyer uncertainty andtwo-part pricing: theory and applications.Management Science 56(2):334–342

Reiss PC, Wolak FA (eds) (2007) Structuraleconometric modeling: rationales and ex-amples from industrial organization. Hand-book of Econometrics, North-Holland, Am-sterdam

Schade S, Frey T, Mahmoud N (2009) Sim-ulating discount-pricing strategies for theGSM-mobile market. Bus. Inf. Syst. Eng.doi:10.1007/s12599-009-0057-0

Schlereth C, Skiera B (2009) Schätzungvon Zahlungsbereitschaftsintervallenmit der Choice-Based Conjoint-Analyse.Schmalenbachs Zeitschrift für betriebs-wirtschaftliche Forschung 61(8):838–865

Shapiro C, Varian HR (1998) Information rules:a strategic guide to the network economy.Harvard Business School Press, Boston

Simon H, Fassnacht M (2009) Preismanage-ment, 3rd edn. Gabler, Wiesbaden

Skiera B (1999) Mengenbezogene Preis-differenzierung bei Dienstleistungen.Deutscher Universitäts-Verlag, Wiesbaden

Skiera B (2010) Differences in the ability ofstructural and reduced-form models to im-prove pricing decisions. Zeitschrift Market-ing – Journal of Research and Management6(1):91–99

Stähler P (2001) Geschäftsmodelle in der digi-talen Ökonomie. Merkmale, Strategien undAuswirkungen. Josef Eul Verlag, Lohmar

Street DJ, Burgess L (2007) The construc-tion of optimal stated choice experi-ments: theory and methods, 1st edn. Wiley-Interscience, Somerset

Swait J, Andrews RL (2003) Enriching scan-ner panel models with choice experiments.Marketing Science 22(4):442–460

Teichert T (2001) Nutzenschätzung inConjoint-Analysen. Theoretische Fun-dierung und empirische Aussagekraft.Deutscher Universitäts-Verlag, Wiesbaden

Train K (2009) Discrete choice models withsimulation, 2nd edn. Cambridge UniversityPress, Cambridge

Völckner F (2006) Methoden zur Messungindividueller Zahlungsbereitschaften: EinÜberblick zum State of the Art. Journal fürBetriebswirtschaft 56(1):33–60

Wertenbroch K, Skiera B (2002) Measuringconsumer willingness to pay at the pointof purchase. Journal of Marketing Research39(2):228–241

Wiltinger K (1998) Preismanagement in derunternehmerischen Praxis: Problemkreiseder organisatorischen ImplementierungGabler, Wiesbaden

Zerdick A, Picot A, Schrape K, Artope A,Goldhammer K, Lange UT, Vierkant E,Lopez-Escobar E, Silverstone R (2001) DieInternet-Ökonomie. Strategien für die digi-tale Wirtschaft. Springer, Heidelberg

Business & Information Systems Engineering 5|2011 277

1

Empirical Evaluation of Fair Use Flat Rate Strategies for Mobile Internet

Marcel Fritz, Christian Schlereth, Stefan Figge Business & Information System Engineering (2011) 3 (5)

Appendix (available online via http://springerlink.com)

2

Appendix 1 – Evaluation of the Assumptions for the Estimation The assumptions which are frequently made for the data analysis of discrete choice experi-

ments are summarized in various papers and books, for example, in Chapter 2 of Chanduka-

la et al. (2007). In the following, we critically examine the assumptions and summarize briefly

the extent to which we consider the assumptions to be fulfilled. For a detailed discussion we

refer to Chandukala et al. (2007).

Tab. A-1 Assumptions regarding the estimation und their critical evaluation

Assumptions Critical evaluation (i) Choices are conditionally independent.

The D-efficient design of the discrete choice experiment en-sures that the choice alternatives are independent. Thereby, we neglect potential dynamic cognitive processes (e.g., learning or fatigue effects when providing responses in each choice set).

(ii) Choice probabilities are driven by parameters that do not change over time.

Due to the fact that the discrete choice experiment is answered in a relatively short period of time (e.g., within 5 minutes), it can be assumed that preferences remain constant. Again, we ab-stract from dynamic cognitive. This assumption has to be criti-cally questioned in the analysis of transaction data with actually made decisions over a much longer period of time.

(iii) Demand is represented by zero’s and one’s, indicating no choice and choice.

This assumption has already been ensured when developing discrete choice experiments. The integration of a non-purchase option in every choice set also allows that consumers are not forced to make a decision if all the alternatives are unattractive.

(iv) There is an explicit set of choice alternatives included in the analysis.

This assumption has also already been ensured by the carefully constructed design of discrete choice experiments.

(v) There is an explicit function form for covariates.

The estimation carried out in this work omits the inclusion of covariates (e.g., age, gender, or income). For the attributes and levels shown in the selection alternatives an additive functional relationship is assumed.

(vi) Some of the coefficients are unique to the choice alter-natives, while others are con-stant across choice alterna-tives.

The utility function used in this study includes both a constant (which is identified through the non-purchase option) and coef-ficients which depend on the characteristics of a tariff.

(vii) The independence of irre-levant alternatives property is valid.

This assumption is discussed critically in the literature, and there are plenty of counter examples in which this property is not fulfilled (see for example the red bus, blue bus discussion in Train 2009). In this paper, this property should be fulfilled, be-cause only one service category (mobile Internet) is considered and the tariffs in the choice sets are clearly distinguished by the design.

3

Appendix 2 – Comparison of Estimation Methods This section compares two alternatives to the Hierarchical Bayesian method, which has been

employed in the study: the maximum likelihood and latent class estimation techniques. The

difference consists in the aggregation level at which the parameter values are determined. In

the case of discrete choice experiments, maximum likelihood methods are only able to accu-

rately determine a vector of parameter values which describes the preferences of all respon-

dents (aggregated level estimates). Therefore, the method is suitable if the respondents are

largely homogeneous in terms of their preferences. In contrast, finite mixture models assume

that there are several segments with different preferences, and that respondents belong to

one of these segments (latent segment level). The membership of a respondent to a seg-

ment as well as the parameter values that describe them are estimated simultaneously in this

approach. Individual parameters for each respondent can mostly only be estimated with Hie-

rarchical Bayes (individual level). A detailed overview of the functionality of the estimation is

provided by Gensler (2003).

The differences in model quality of the different aggregation levels are shown in the following

comparison applying all three methods of estimation. Thereby, we make use of the common-

ly applied first choice hit rate (percentage of the correctly predicted choices with the esti-

mated parameters) and the mean absolute deviation (MAD) between the predicted probabili-

ty of the selected and actually observed choices (e.g., Gensler 2003). These two ratios are

used to predict both the choice sets considered in the estimation (internal validity) and the

two holdouts (predicted validity). In the case of the finite mixture model, a discrete number of

segment classes, i.e., a number of groups with different selection behavior, must be specified

in advance. As an example, we have chosen the segment classes 4 and 8 in the following

comparison. The results of the comparison are reported in Table A-2.

Table A-2 Quality of the estimations

Internal validity Predicted validity First choice rate MAD First choice hit rate MAD Maximum likelihood 52% 0.61 57% 0.58 Finite mixture (4 segment classes) 70% 0.42 66% 0.46 Finite mixture (8 segment classes) 75% 0.36 62% 0.44 Hierarchical Bayes 86% 0.22 76% 0.27

The model that considers the largest degree of heterogeneity, i.e., Hierarchical Bayes, yields

the highest model quality. The differences with respect to the maximum likelihood estimates

are large (e.g., 86% vs. 52% and 76% vs. 57% hit rate). It can be concluded that the prefe-

rences of respondents are very heterogeneous and that a consideration of these preferences

during the estimation substantially improves the results’ internal and predictive validity.

4

Appendix 3 – Representativeness of the Survey Participants Regarding the German Population

In cooperation with a German survey panel provider a representative sample of the popula-

tion was collected. This sample consists of 49.83% male participants (49.00% share of men

in Germany) and an average age of 44.50 years (average age 42.10 years in Germany). A

detailed comparison by age groups can be found in Table A-3.

Table A-3 Comparison of age groups

Share \ Age 18-21 22-40 41-65 Population 5.7% 37.1% 57.1% Survey 4.8% 28.7% 66.4% * Population shares are taken from the Federal Statistical Office. Only the age groups between 18-65 years were considered.