2.1 Introduction

In the prebroadband era, ¹ recovering a local telephone company's network costs for voice service was a relatively straightforward process. The users of the local network fell into two categories: end users who bought phone service and long-distance carriers who sold long-distance service to end users. The pricing arrangements were also relatively clear. The end user bought local service from the local telephone company and long-distance carriers such as AT&T and MCI paid the local phone companies for the use of their networks to complete long-distance calls. The relationships between the local and long-distance companies were clearly defined by Federal Communications Commission (FCC) rules and a standard tariff defined the services rendered by the local telephone company and the charges to long-distance carriers.

The broadband world has complicated network cost recovery considerably. In the broadband world, voice service is only one application traversing broadband networks, so it is no longer straightforward to separate it from other services. Accessing many applications over a broadband pipe produces diverse traffic patterns among users in comparison to traditional voice service, and to a large extent, usage is not closely tied to the prices they pay for broadband services. End users can buy a variety of broadband plans; the more expensive ones include higher speed access to the Internet, traditional voice, and cable TV. End users buying the same broadband plan often differ significantly in their use of the broadband network. The relationship between bandwidth use and price paid is far more tenuous among application providers. Google and Yahoo are heavy bandwidth users because they essentially take pictures of the Internet to improve their web site performance.² YouTube, Netflix, and other video providers take up perhaps one-third of the Internet's bandwidth in the United States.³ These applications providers pay a fraction of the bandwidth cost of an end user to the Broadband Service Provider (BSP). By contrast, small application providers such as a local restaurant face a usage/cost relationship close to those for end users.

Billing application providers is beyond the capabilities of many BSPs. If an application provider is not in a BSP's service area, it typically does not have or want a contractual relationship with the BSP. Without a contract, these application providers can terminate traffic free of charge. In case of default, BSPs recover their network costs from their customers. This is the so-called bill-and-keep arrangement.

A major policy issue is whether bill-and-keep is a welfare maximizing strategy from a government policy perspective. The answer to this question often depends on perspective. Application providers support bill-and-keep because they fear that BSPs could exercise monopoly power to raise terminating rates for their traffic. BSPs oppose bill-and-keep but face the almost impossible challenge of how to bill millions of application providers. The FCC supports a bill-and-keep system but recognizes that it needs a funding base to subsidize broadband service where private businesses cannot make a business case for offering broadband services at speeds widely available elsewhere. To raise support funds, the FCC recognizes that it may want to assess application providers and use the funds raised to make broadband service universally available. If, however, nothing is done to recover network costs from application providers, wireline BSPs may not build out to remote customers and may introduce usage limits such as the 10 GB per month consumption threshold already instituted by Verizon Wireless and AT&T Wireless more than which a customer pays an extra charge.

The focus of our paper is whether it is in the financial interest of application providers as a group to buy down end user charges. A reasonable case could be made that application providers, as a group, may be willing to subsidize end users if two conditions hold: First, end users are price sensitive to the cost of broadband speeds. If so, lowering the price of broadband service will encourage them to buy higher speed service and use the Internet more intensively. As a result, they would not throttle their use of Internet and will not be deterred from accessing applications with higher bandwidth requirements, which would foster new applications development. Second, subsidization of end users must be competitively neutral across providers and across networks. Research in the two-sided market literature shows that if subscribers to higher speed tiers are price elastic, it may be beneficial for content providers who do not enter in direct contractual relationships with end users to subsidize prices for high speed offerings through the public support system (e.g., [1]). Buying down of end user charges could be done through a universal broadband fund such as the newly established Connect America Fund (CAF). The intended goal of CAF is “to enable all US households to access a network that is capable of providing both high-quality voice—grade service and broadband that satisfies the National Broadband Availability target,” currently set at 4 Mbps downstream and 1 Mbpsupstream.⁴ To the extent that a private sector business case cannot be made to offer this minimum service, CAF funds should be available to fund the shortfall.⁵ In fact, the FCC is currently evaluating whether end users, content providers, or both should contribute to CAF, which is designed to foster universal broadband connectivity.⁶

Estimating price elasticities of demand has a second and important benefit to BSPs whether the buy down occurs or not. If demand is price elastic to speed, it will affect their pricing strategies. Lower prices may boost demand for broadband services at higher speeds. Unfortunately, the empirical economic literature offers mixed results for price elasticity of demand for broadband. Some studies find own-price elasticity of broadband demand to be larger [2–4], while others estimate smaller price sensitivities [5, 6]. One simple explanation for this variation in estimates is that the choice to access to the broadband differs from the choice of which bandwidth service level to subscribe to once the customer chooses to buy broadband service. Specifically, households may have different price sensitivities for gaining access to the Internet and the service level they choose. Varian [4] estimates willingness to pay for speed using experimental data collected in 1998 and 1999. The own-price elasticities of demand for bandwidth range from −1.3 to −3.1. He finds that lower service bandwidths are perceived as substitutes for the level of bandwidth chosen by participants in the experiment and reports low willingness to pay for higher bandwidths. In a more recent experimental survey design study, Rosston et al. [7] estimate households’ willingness to pay for speed and reliability. They find that households value broadband highly, but values for very fast speeds are only marginally higher than values for fast speeds. They explain that a typical household does not use applications requiring blazing speeds. Glass and Stefanova [8] find highly inelastic own-price elasticities of demand by looking at the total subscribers of a BSP and the lowest price for its introductory offer. The estimates are −0.66 using data from 2005 and −0.21 using data from 2009. These estimates are based on the introductory offer prices and can be interpreted as estimates of the price sensitivity of gaining access to the Internet. The lower price elasticity estimated from the 2009 dataset suggests that access to the Internet has become more of a necessity in the later period.

These studies suggest that once a household has subscribed to broadband, it may have different price sensitivity for access to the Internet as it relates to expected usage. A household may perceive bandwidth as a differentiated service, which is captured to some extent by the choice of speed tiers. Even a low speed tier connection can be used for browsing e-mail, You Tube video, and near-real time video,⁷ but customers who are heavy users, especially of near- and real-time video, are farmore likely to purchase a high speed tier. In the current study, we use a detailed demand and price data collected in 2010 to estimate a demand model for Internet speed tiers offered by rural BSPs and use the parameter estimates to calculate price elasticities of demand for three differentiated speed tiers—with downstream speeds up to 3 Mbps, between 3 and 6 Mbps and with downstream speeds above 6 Mbps.⁸

2.2 Model

Our goal is to estimate price elasticities of demand for different speed service packages purchased by broadband users. The generalized Leontief [9], the Rotterdam model [10], the Translog model [11], and the Almost Ideal Demand System (AIDS) [12] are among the most popular attempts to provide flexible, but theoretically consistent, functional forms to study consumer purchases of different products. We use an AIDS model for our estimation.

The AIDS model is specified in the usual way:

where is the total expenditure on broadband in a given market, is the price of the th service, is the share of the total expenditures spent on the th good, and is the price index defined as

We estimate the standard AIDS model specified above and an AIDS model with demographic variables. Following Heien and Wessells [13], demographic variables can be incorporated into the model by including them in the intercepts of the share equations.

where denotes the th demographic characteristic. , , , , and are parameters to be estimated. Parameter restrictions are added to ensure that the demand equations are consistent with the theory:

These conditions guarantee that the sum of the budget shares is unity. Homogeneity restrictions imply that the budget shares will not change if all prices and expenditures are multiplied by the same positive constant. The symmetry restrictions require that compensated demand effects be symmetric to be consistent with consumer theory.

In order to have the budget shares for each market add up to 1 in the model with demographic variables, the adding up restriction was replaced by

For a complete system of demand equations, the variance covariance matrix is singular because of the adding up property, so we drop one of the tiers and estimate the model using maximum likelihood. It is shown that maximum likelihood estimates are invariant to which product is dropped in the estimation [14].

Expenditure elasticities are computed at mean shares using the following expression,

Price elasticities are given by

where 1 if and 0 otherwise.

2.3 Data

The data used in our analysis were obtained in 2010 from a nationwide survey of rate of return rural local exchange carriers (RLECs) participating in the National Exchange Carrier Association (NECA) pools and their affiliated Internet service providers. We collected data on Internet services purchased by rural households from these providers. Companies reported price, upload and download speed, and total number of subscribers for all Internet packages they offer. We used data for 194 broadband providers which reported at least 3 broadband services—one with download speed below 3 Mbps, one with download speed below 6 Mbps, and one with download speed greater than or equal to 6 Mbps. If a company offered more than one service in any of the three categories, we aggregated the services into the three representative service tiers.

2.4 Variable Descriptions

We aggregated the service tiers into three categories—services with downstream speed below 3 Mbps, services with downstream speed between 3 and 6 Mbps, and services with downstream speed above 6 Mbps. The resulting average speeds within these categories are 1.29 Mbps for the service tiers below 3 Mbps, 3.34 Mbps for the middle category, and 7.93 Mbps for the top category of services. Average prices increase with the speed offered and the average price of $75.37 paid per month for the top tiers is almost two times the average price of $37.93 paid per month for the lowest service tiers. The average number of subscribers for the tiers below 3 Mbps is 2016, while the average number of subscribers for the highest speed tiers is less than a quarter of that: 479. Average upload speeds also increase with the higher download speeds but are in general much lower than the download speeds. The average upload speed for the highest tier in our data of 6 Mbps and above is 1.29 Mbps. The average expenditure for all three service tiers per company is $135,335/month. The expenditure share for the lowest tier is 60% of total, for the middle tier is 28%, and for the top tier is 12%. Table 2.1 shows the summary statistics for the aggregated sample.

To derive the demographic variables, we matched the serving territories of the RLECs with census data as well as data published on the National Broadband Map web site.⁹ We include variables to control for age, education, income, race, and density of the population of the markets in our study. Percentage of population of age less than 19 years and age greater than 60 years will be interpreted relative to the percentage of population of age between 19 and 60 years old. Educational achievement is often found to be important in studies of broadband adoption. We include a variable for the percentage of population with bachelor's degree or higher. People living in markets with lower household density may find it more beneficial to subscribe to higher speed broadband, if available, because of a larger need to telecommute, use telemedicine applications, or do online shopping, compared to households living in more densely populated areas. Percentage of population with incomes below federal poverty levels is also included in the model. White and Asian households are often found to be positively correlated with broadband adoption when compared to non-White, non-Asian populations, thus we include the percentage of population of White and Asian race to our model.

2.5 Results

The explanatory variables include the logarithms of prices and expenditures for the three service tiers in both models and demographic characteristics of the markets in the second model. The parameter estimates are reported in Table 2.2. Most of the price coefficients achieve statistical significance at the 0.01 level, while the demographic variables in the AIDS model with demographics, are mostly insignificant. The exceptions are the negative coefficient on income below poverty levels for tier 3 and the negative coefficient for tier 3 on percentage of white population in the market.

Tables 2.3 and 2.4 reports the uncompensated own- and cross-price elasticities and expenditure elasticities for each service tier and corresponding asymptotic standard errors. All but two of the elasticities are statistically significantly different from zero at the 0.05 level.

All elasticities have the expected signs. Own-price elasticities of broadband services are negative and indicate elastic demand.¹⁰ The elasticity estimates increase with speed with services below 3 Mbps having the lowest estimate (−1.721 with the standard AIDS model and −1.746 whendemographics are included) and services above 6 Mbps having the highest estimate (−2.707 and −2.788, respectively). Subscribers to higher tiers, who we expect are more heavy users of the Internet (watching video, playing online precision games, etc.) are more sensitive to price, while lower tier subscribers, who we expect use the internet primarily for web browsing and checking e-mail, are less price sensitive.

Cross-price elasticities are positive, indicating that service tiers are substitutes. Cross-price elasticities of the higher speed tier with respect to the lower speed tier is always larger than the cross-price elasticity of the lower tier with respect to the higher speed tier. This observation is compatible with the idea that lower speed users are much less concerned by the monthly charges for higher speed services than higher speed users are concerned with the cost of a lower speed service tier. Estimated expenditure elasticities show that all broadband services are normal goods and the demand is positively correlated with income levels.

Note: Values in parentheses are standard errors. Asterisk denotes significance at 0.05 level. Log likelihood ratio test fails to rejects the null hypothesis of no demographic effects in the model at 0.05 level but rejects it at 0.1 level.

Note: Values in parentheses are standard errors. Asterisk denotes significance at 0.05 level.

2.6 Conclusions

This study utilizes cross-sectional data from rural rate of RLECs to estimate market demand functions for broadband service tiers. In addition to prices and number of subscribers by service tier, the data include demographic characteristics of the markets. We used two specifications of the AIDS to estimate the demand parameters and price and income elasticities. Our empirical findings indicate that DSL customers are price sensitive and price sensitivity increases for customers purchasing higher bandwidth packages. Together with our earlier study [8], the overall results suggest that DSL access has become a communications necessity, but customers are very price sensitive to the DSL service level they buy. Our results are consistent with other studies that find high household valuation for broadband but low incremental values for higher speeds. Moreover, the empirical results strongly suggest that application providers, as a group, would benefit from buying down the price of DSL for end users. Network providers would receive additional funds to expand network capacity and they would be less likely to introduce monthly capacity limits. Application providers, as a group, would benefit from greater use of the Internet and higher likelihood that end users would access applications with higher bandwidth needs.

References

1. 2. P. Hande, M. Chiang, A. R. Calderbank, and S. Rangan, “Network pricing and rate allocation with content provider participation,” Proceedings of IEEE INFOCOM, Princeton, NJ, 2009.
2. 2. D. Kridel, P. Rappoport, and L. Taylor “The demand for high-speed access to the Internet: the case of cable modems,” in D. G. Loomis and L. D. Taylor, eds., Forecasting the Internet: understanding the explosive growth of data communications, Kluwer Academic Publishers, Norwell, MA, 2001.
3. 3. A. Goolsbee, “Subsidies, the value of broadband, and the importance of fixed costs,” in R. Crandall and J. H. Alleman, eds., Broadband: should we regulate high-speed internet access?, Brooking Institution Press, Washington, D.C., 2002, pp. 278–294.
4. 4. H. R. Varian, The demand for bandwidth: evidence from the INDEX project, 2002. Available at http://www.sims.berkeley.edu/∼hal/Papers/brookings.pdf. Accessed March 30, 2014.
5. 5. R. K. Goel, E. T. Hsieh, M. A. Nelson, and R. Ram, “Demand elasticities for Internet services,” Applied Economics, 38, 2006, 975–980.
6. 6. K. T. Duffy-Deno, “Business demand for broadband access capacity,” Journal of Regulatory Economics , 24, 2003, 359–372.
7. 7. G. Rosston, S. Savage, and D. Waldman, “Household demand for broadband Internet service in 2010,” B. E. Journal of Economic Analysis and Policy Advances , 10(1), 2010.
8. 8. V. Glass and S. Stefanova, “An empirical study of broadband diffusion in rural America,” Journal of Regulatory Economics , 38(1), 2010, 70–85.
9. 9. W. E. Diewert, “An application of the Shephard duality theorem: a generalized Leontief production function,” Journal of Political Economy , 79, 1971, 481–507.
10. 10. H. Theil, “The information approach to demand analysis,” Econometrica , 6, 1965, 375–380.
11. 11. L. R. Christensen, D. W. Jorgenson, and L. J. Lau, “Transcendental logarithmic utility functions,” American Economic Review , 65 1975, 367–383.
12. 12. A. Deaton and J. Muellbauer, “An almost ideal demand system,” American Economic Review , 70, 1980, 312–326.
13. 13. D. Heien and C. R. Wessells, “Demand systems estimation with microdata: a censored regression approach,” Journal of Business and Economic Statistics , 8, 1990, 365–371.
14. 14. R. Pollak and T. Wales, “Estimation of the linear expenditure system,” Econometrica , 37, 1969, 611–628.