ABSTRACT
Does entry and exit of competitors from broadband services markets have large effects on the quality of broadband plans offered to consumers? Answers to this question inform design of subsidies to improved broadband in underserved areas, and antitrust policy. The authors investigate quality effects of entry (or exit) in 2014 legacy (cable and digital subscriber line) duopoly residential markets. Entry or exit by legacy technology providers had large impacts on 2018 service quality, though impacts diminished rapidly with continued entry. Entry by a single fiber Internet service provider (ISP) had no impact on available quality; entry by more fiber ISPs had modest but statistically significant effects.
In markets for differentiated products, quality1 is likely to be a critical dimension of competition. In addition to price, quality choices allow firms to differentiate product offerings from one another and to target different demand segments. In such circumstances, an analysis of how well a market performs cannot simply rely on the price and quantity observed as market outcomes but must also explicitly consider product quality. Because increases (decreases) in quality are consumer surplus enhancing (diminishing), ceteris paribus, the relationship between market competition and product quality is quite important from a policy perspective.
Firms can increase product quality in order to respond to competitors entering markets with established products. Product quality differences can thus work to lessen price competition (Shaked and Sutton, 1982). But firms in markets with low competitive pricing pressure may choose lower quality levels for their products than firms facing higher levels of competitive pressure in the marketplace (Tirole, 1988, section 2.1). If products have different quality tiers, there may be little price competition within each tier. In short, greater competition has ambiguous effects on product quality in the theoretical literature (Shapiro, 1983; Allen, 1984; Kranton, 2003).
In this article, we analyze the effects of local market structure on residential broadband service quality. Technological change stimulating new entry (exit) into (from) “legacy technology” broadband service duopolies created a unique opportunity to analyze an interesting and important question: How exactly does increased competition, measured as numbers of competing Internet service providers (ISPs) in a census block, affect quality of service, measured as the highest speed offered by legacy service providers in a local service market (census block)?
The U.S. residential broadband market is an ideal setting for a study of the causal relationship between market structure and quality. First, the available data suggest that nominal prices for residential broadband service plans in the United States have basically been relatively flat on average—changing little over the last decade. The major dimension for competition among major broadband ISPs has been improved service quality, particularly upgrades to download speeds. Second, over a relatively short time window, we document evidence of substantial entry and exit by firms within spatially disaggregated local markets. Third, technological innovation has significantly lowered the costs of quality improvements, so there is substantial variation in quality within a cross-section of spatial locations over time. Improvements in Internet service quality have the potential to bring large impacts on household welfare: most people now use the Internet, and they use it intensively.2
Using data from multiple sources, we have built a rich panel data set that covers all broadband services delivered to households by ISP, technology, and maximum download speed for each of approximately six million populated U.S. census blocks between December 2014 and December 2018. We summarize trends in market structure, quality, and technology utilization within this five-year period. Previous empirical analyses examining the effect of market structure on quality provision in the U.S. market (over earlier periods, using more aggregated data) relied on very noisy and sometimes inappropriate measures of quality. Additionally, earlier studies made strong assumptions about the way firms compete in the market, and strong statistical assumptions about explanatory variables.
Given the richness of the data set and a novel empirical identification strategy, we are able to address many of the limitations of previous studies. Identification of the causal effects of market structure on service quality must potentially account for the possible endogeneity of market structure measures. Our difference-in-difference (with fixed effects) framework addresses these problems.
The analytical framework deployed in this article is developed around a reduced form model of the intensive competitive margin, as firms choose the quality of their service offerings in the short to medium run, while facing fixed numbers of competitors.3 Firms using any broadband delivery technology face a tradeoff between number of users that can be served by their existing network, and speeds that can be offered to those users. Given market structure (the number of competitors in the market), how does the extent of competition from other ISPs affect the maximum service quality (maximum download speed) offered to consumers? We recognize that in the long run, entry and exit into a market is also endogenous—that is, the same factors that lead a firm to offer higher quality service may also lead others to enter (or exit) the market. We are interested in what we can learn about how market quality outcomes—in this case, the maximum quality/speed for broadband available from legacy technology service providers in a census block market—vary with market structure, that is, the numbers of different types of competitors, after conditioning on economically relevant demand and cost shifters.
While market structure is endogenous in the long run, a variety of national and state regulatory policies clearly affect market structure outcomes, and policy levers clearly can have an effect on market structure “treatment” variables. In Texas, for example, cities and municipalities are barred by state law from providing broadband service to residents, while this is not so in other states where municipal broadband service is legal. As another example, the Federal Communications Commission’s (FCC) new auction mechanism for universal service subsidies explicitly provides for possible subsidies to entrants wishing to serve “high-cost” census blocks with inadequate broadband availability from incumbent providers. Understanding the impact of market structure (number of competitors) on service price and quality is relevant to policy choices even if endogenous firm decisions interact with policy in determining market outcomes.
We estimate the effect of increasing competition on service quality using a Difference-in-Differences (DiD) framework. Our results suggest that entry into (or exit from) a 2014 “legacy duopoly” census block (about 40% of U.S. urban census blocks were legacy—cable and digital subscriber line [DSL]—terrestrial fixed broadband duopolies in 2014) by other legacy ISPs had a relatively large and statistically significant impact on maximum broadband quality offered to consumers by legacy ISPs serving a given block. Increased competition in local broadband markets from “overbuilding” by broadband providers using legacy technology seems to result in large quality improvements. In contrast, entry by a single fiber ISP had no statistically significant effect on offered service quality, while entry by two or more fiber ISPs had a modest but statistically significant effect in increasing legacy ISP speeds. We conclude by interpreting these results and identifying potential policy implications.
The remainder of the article is organized as follows. The section “Literature Review” describes related literature; the section “Empirical Context” briefly reviews essential background; the section “Data Sources” describes data sources and how we built our analytical dataset; the section “Identification Strategy Overview” discuss our identification strategy; the section “Difference-in-Difference Results” the details of the DiD model and empirical results; the section “Interpreting Our Results: Summary” interprets and summarizes results, and the section “Conclusion and Policy Implications” presents conclusions.4
Literature Review
There is a broad economic literature studying the effects of market competition on equilibrium outcomes, such as price and quality. One of the first empirical papers to address the effect of market structure on prices was Bresnahan and Reiss (1991). They measure the effects of entry on competitive conduct, with firms pricing closer to marginal costs as a metric for more competitive conduct. They find that almost all changes in competitive conduct happen after a second or third competitor enters the market. The previous empirical literature has largely not engaged with the effects of entry on dimensions of competition other than prices. In addition, a feature of the broadband industry, and the data we utilize, is that there are objective and directly observed measures of quality: advertised maximum download and upload speeds, which providers and consumers contract for, and we directly observe in the data.5
Turning to empirical studies, the FCC has produced voluminous reports on the residential broadband industry. For instance, every year, the FCC releases a Measuring Fixed Broadband Report analyzing the performance and quality of U.S. residential broadband service. Another example is the Broadband Deployment Report, assessing penetration of high-speed broadband connections among U.S. households. A broader analysis of the telecommunications industry is performed by the FCC every two years and released as the Communications Marketplace Report. In these reports, the FCC assesses the state of competition in the broader communication industry.6 In all these reports, the FCC has generally described steady progress in terms of increased broadband availability, affordability, and competition. In the 2020 Broadband Deployment Report, the FCC states “Available evidence demonstrates that the digital divide continues to narrow as more Americans than ever before have access to high-speed broadband” (FCC, 2020). Unfortunately, these FCC reports do not generally summarize the data in a consistent way from year to year, while data are sometimes retrospectively revised in significant ways, and the underlying data is not always available to replicate particular analyses.
There is an extensive literature studying the U.S. broadband industry. Previous work includes studies of service availability, broadband policy effectiveness, and the relation between competition, market structure, and quality. On geographic service availability, Gadiraju et al. (2018) study the deployment of broadband in the United States during the period 2014-2016. They use Form 477 census block data from the FCC to describe service availability and speed, and how they are related to demographic characteristics measured at the census block group level. Their main finding is that around three-fourths of households have access to two or more Internet providers. Beede (2014) describes broadband service availability in December 2013 across the United States. A novelty is that raw data (restricted and nonpublic) from the FCC’s Form 477 (at the census tract level) is used in the analysis, so it provides a unique overview of the whole industry. One main finding was that 98% of the population had a choice of at least two ISPs with speed of 3 Mbps or higher. Additionally, 37% of the population had a choice of at least two providers at speeds of 25 Mbps or greater, but only 9% had three or more choices. Xiao and Orazem (2011) use the framework developed by Bresnahan and Reiss (1991) to study entry and competition in the broadband industry. Their main finding is the negligible effect of the fourth incumbent entry on firms’ competitive conduct.
Neither of those papers considers the longer term evolution of the market, limiting their analysis to the market state at a given point of time. One of the goals of this article is to assess the impact of changes in market structure and concentration on broadband quality outcomes over the five years from 2014 to 2018. Flamm and Varas (2018) have observed that census tract level data significantly overstate the number of broadband providers from which an individual residential household in that census tract may be able to purchase service. That is why we construct our analysis at the census block level.
Several papers have studied the effects of policies that seek to promote broadband penetration in underserved areas. The focus of these papers is policy impact on broadband penetration and its effects on other industries. Boik (2017) uses household-level cable and satellite broadband subscription data from North Carolina to estimate how many currently unserved regions warrant an entry subsidy. He finds that the cost of connecting the 5% least dense areas of North Carolina is equivalent to $1519 per household per month. Bai et al. (2020) study the effect of the Broadband Initiative Program (BIP) in the agricultural sector. They find that that BIP funding had a small but short-term impact on per capita farm sales, but the effect is mainly through a regional spillover rather than a direct effect. Whitacre and Gallardo (2020) assess the impacts of state policies on total and rural broadband availability in the United States. Their results suggest that state-level funding programs have had a meaningful impact on broadband availability.7 Another example is Rosston and Wallsten (2020) who evaluate a low-income broadband program that Comcast implemented after the Comcast-NBCUniversal merger. They find that broadband adoption increased at a higher rate in areas served by Comcast compared with areas not served by Comcast. Zuo (2021) finds that discounts on broadband service for low-income families increased Internet use, employment rates, and earnings.
This article is also related to a previous literature that studies competition and market structure in the broadband service industry, and its possible effects on quality. Previous papers have found a positive relationship between increased competition and download speeds. Wilson (2022) finds that in competitive markets higher download speeds are provided compared to monopoly markets.8 Reed and Watts (2018) show that access to high-speed Internet increases as the number of providers increases in a county. This article differs from Wilson (2022) and Reed and Watts (2018) in using a more granular market definition. We study the quality-market structure relationship at the census block level, rather than at the city or county level. ISPs may not offer service in all parts of cities and counties in which they serve some customers. As it is discussed below, using broader market definitions tends to overestimate the number of ISPs available for consumers in any given area. This article also differs from Wilson (2022) and Reed and Watts (2018) in terms of the types of data used for the empirical analysis. Wilson (2022) uses speed test data as a quality measure. This data is likely to generate selective and biased results because users experiencing low speeds are more likely to run speed tests. Reed and Watts (2018) conduct their analysis based on cross section variation of market structure and service availability. In contrast, this article exploits panel variation over time, controlling for unobservable time-invariant market characteristics which may bias results in a cross-section analysis.
Prieger et al. (2015) study how broadband firms respond to the entry and quality decisions of their rivals. Estimating a discrete choice model of entry and quality, their preliminary results show that firms actually respond to quality choices of rival broadband providers, and firms’ responses are heterogeneous with respect to type of provider and quality. Wallsten and Mallahan (2010) is one of the few papers that analyzes the effect of market structure on prices. Using nonpublic versions of older FCC data, they find a strong correlation of number of providers with housing density, penetration with income, and show how different wireline technologies are more or less prevalent in areas with differing housing density. A key result is that lower prices, attributed to competition, are observed in markets with three providers, compared to markets with just two providers. This study establishes an analogous effect with respect to quality.
Molnar and Savage (2017) use field measures of download speed to analyze the effect of entry and market structure on product quality. Their results show that wireline speeds tend to be higher in markets with two or more wireline providers than with a single wireline provider. An important limitation in their analysis is their very limited sample of households. For instance, in many of their markets (defined as census block groups), they use an observed speed measurement derived from only a single volunteer household’s purchase of Internet services, and their data is available only in a single period. Additionally, they impose strong distributional assumptions on unobservables in their model, and the instruments they use to address potential endogeneity problems might plausibly be directly relevant to an ISP’s quality choice decisions and create doubts about instrument validity.9
A recent study closely related to our work is Kotrous and Bailey (2021). Their study focuses on the effects of new entrants on incumbent wireline provider service quality but limits the analysis to entry by new fiber ISPs. They find results similar in some respects to ours—that is, a single fiber entrant has no positive impact on competitor download speeds, while additional fiber entrants may stimulate a small improvement. Our analysis is broader insofar as it also analyzes the impacts of new entrants using legacy technologies as well as fixed wireless technology. It also differs substantially in terms of heterogeneity of units analyzed (mixing together all urban and rural census blocks in the lower 48 states, vs. our focus on effects of the evolution of market structure over time in a more homogenous set of 2014 urban legacy technology duopolies), numbers and types of control covariates used, and in statistical modeling assumptions.
A key element in any analysis of competition is market definition. Equilibrium outcomes, such as the number of firms in each market, price, or quality, may vary greatly depending on how broadly or narrowly markets are defined. Previous papers have used geographically larger spatial definitions of fixed broadband service markets, driven primarily by data availability. For instance, Reed and Watts (2018) use counties as markets; Flamm (2015), Xiao and Orazem (2011), and Grubesic and Murray (2004) use zip codes as markets; Molnar and Savage (2017) use census block groups, and Wallsten and Mallahan (2010), Prieger et al. (2015), and Denni and Gruber (2007) use census tracts. Certainly (and particularly for wireline networks), it seems clear that there is zero substitution possible on the demand side between broadband services providers whose networks pass a particular block, and service providers whose networks do not reach that block, even if both blocks lie in the same census tract or block group. A novel contribution of this article is to study how, when a residential broadband market is more plausibly defined spatially at the block level, observed numbers of competitors offering services to a residential household affects highest quality service availability.
Empirical Context
The existing literature has established a number of stylized facts about pre-Covid U.S. residential broadband service (Flamm and Varas, 2021). Broadband services were a significant expenditure for U.S. households. In most U.S. local broadband markets, there were very limited numbers of competing residential broadband ISPs. The year 2017 was a high water mark for competition, with declines in mean and median competitors per census block registered over 2017–2018.
Residential broadband service prices have been decreasing slowly over time in the United States. Nominal service plan prices have basically been flat, and quality-adjusted prices have fallen slowly by tech industry standards. Price competition among residential broadband ISPs seems to be relatively muted.
Quality competition, on the other hand, seems to be quite important in U.S. residential broadband. Small numbers of competitors notwithstanding, most urban census blocks have experienced great improvement in maximum speeds available to households. Gigabit-class download speeds (900+ Mbps) were available from at least one provider in 85% of urban census blocks in 2018, compared to just 5% of these blocks back in 2014. In (much larger spatially) rural blocks, gigabit availability rose to roughly 37% in 2018 from about 3% in 2014.
There was very limited net entry by new residential broadband service competitors. Over the 2014–2018 time period, mean ISPs per census block increased by .07 in urban areas; in rural areas, by .18. Technological innovation in wireless service provision powered much of the new entry visible over the 2014–2018 period, but the new competition from wireless ISPs was almost entirely restricted to lower tier quality broadband service. Furthermore, after 2017, there was considerable exit of fixed wireless ISPs from local residential broadband service markets. The period during which these technological improvements were observed coincides with the deployment of a new and evolving set of government programs designed to increase the quality of available broadband service in areas historically receiving “high-cost” telephone service subsidies. For that reason, we must worry about the confounding effects of subsidies to incumbent providers in affected areas that might be expected to have impact on available broadband service quality.
A small number of urban census blocks (as classified by the 2010 U.S. census), as well as rural areas, received Connect America Fund (CAF) subsidies.10 We do not attempt to translate this complex menu of transitional broadband subsidy mechanisms into specific impacts by subsidy mechanism. Acceptance of CAF subsidies by service providers was tied to future broadband service quality obligations in high-cost census blocks. Our strategy to control for the average impact of these subsidies on broadband speeds in affected urban census blocks is to create annual subsidy “vintage” variables. The “vintage” of subsidies affecting a rural census block is defined by the year in which the first claim for a broadband subsidy to a location in that census block was made.
We interpret the year the first broadband subsidy claim was made for a location within a census block as an indicator of the mix of subsidy mechanisms put into place when a rural census block was “converted” to the new CAF funding system. The FCC subsidy offers reflected how the FCC cost model in use at the time evaluated block characteristics (and profitability, at prices like those faced by a “reasonably comparable” urban household11); acceptance of the offer by an incumbent local exchange carrier (ILEC) presumably indicates that the ILEC expected service to be profitable using its own internal expectations of cost net of subsidy, and pricing. Subsidy acceptance was accompanied by an obligation to serve new locations in subsidized blocks if the improved service level could be provided “at reasonable cost.”12
Broadband subsidy funding vintages in years no later than 2014 (the first year of our sample) are effectively incorporated into a census block fixed effect. Census block vintages first associated with high-cost broadband-related subsidy claims filed in 2015 and 2016 most likely reflect the new Connect America Model (CAM)-based subsidies in census blocks served by so-called “price cap” incumbent carriers. High-cost census blocks first receiving broadband subsidy funding in 2017 seem to be a mix of CAM price cap blocks and an inaugural class of Alternative Connect America Model (ACAM) census blocks served by so-called “rate-of-return” carriers.13
A key point to remember about these subsidies is that, through 2018, they largely went to incumbent broadband and telephone service providers in exchange for contractual best efforts to improve broadband quality (as measured by download/upload speed targets) in their service areas.
Data Sources
The primary data source for this analysis is the Fixed Broadband Deployment data collected by the Federal Communications Commission (FCC) through its Form 477. All facilities-based fixed broadband service providers are required to report a variety of data for census blocks in which they offer Internet access service at speeds over 200 kbps in at least one direction (i.e., download or upload).14 The ISPs reporting service offers to a block are not required to actually serve any customers in the reported census block—creating a hypothetical “could” aspect in the reported offer of service that sometimes appears to create data issues, particularly for wireless service providers who in theory can provide service within a fixed line-of-sight radius of stationary wireless antennas (though maximum theoretical speed that can be provisioned will generally vary with distance). Such “phantom” wireless ISPs may not in practice actually market their service within all line-of-sight areas reported to the FCC as offered service.15
The Fixed Broadband Deployment Data is published twice per year (June and December). For each census block, broadband service providers are identified by their “Holding Company” (HOCO) name. For each technology deployed by a provider in that census block, the data contains the maximum advertised download and upload speed, and whether that service is sold to households or business.16 Technologies identified include various flavors of cable DOCSIS standards, various flavors of DSL, other wireline, fiber, satellite, and terrestrial fixed wireless. In its current format (at the census block level), this data was available from December 2014 to December 2018. Within that period all variables are comparable across years. We have reclassified all flavors of wireline DOCSIS networks as “cable,” and all flavors of wireline DSL networks as “DSL.”17 In our analysis we also use the FCC’s Mobile Deployment Data.
Broadband locations receiving high-cost CAF subsidies, at the census block level, are available from the Universal Service Fund Administrative Company’s public data portal.18 We created census block-level binary (0,1) indicator variables taking on value zero if no CAF high-cost broadband funding had previously been received in the census block, 1 in any year thereafter. These binary “vintage of CAF” variables were created for the years 2015 through 2018.19
Economic and demographic information at the census-block level is obtained from three sources: (i) the American Community Survey (ACS), (ii) FCC-staff annual estimates of population, housing unit, and household counts for census blocks, and (iii) U.S. Census, Origin-Destination Employment Statistics (LODES). Examples of ACS economic and demographic variables are population, percentage male population, ethnicity distribution, mean income, age, education, housing value, and percentage population without telephone service at the census block group level. Each ACS five-year estimate is assigned to the last year of the estimation period. For instance, the estimates for the five-year period 2014–2018 are assigned to the year 2018.
In addition, the Census Bureau creates estimates of the characteristics of workers residing in census blocks, at the census block level, based on administrative records. The measured “residence area characteristics” (RAC) of workers filling jobs residing in a census block include the total number of jobs, as well as counts of jobs by age, education, industry classification of employment, race, ethnicity, and sex classifications. Because an employee can work multiple jobs, the counts of socioeconomic characteristics are compiled by job, and we therefore measure these job characteristics as the share of total census block jobs filled by workers in each respective socioeconomic category.
We use Costquest estimates of fiber network cost per connected household, assigned to one of six categorical ranges for most (but not all) U.S. census blocks.20,21 We interact base year estimates of network costs with time dummy variables to capture national shifts over time in wireline network costs, by baseline cost category group. Finally, we use geospatial data measuring terrain characteristics. We borrow the geographical terrain ruggedness data developed by Nunn and Puga (2012). In particular, we use their Terrain Ruggedness Index (TRI).22 The TRI data developed by Nunn and Puga are calculated at the level of 30 arc-second cells on a regular geographic latitude/longitude grid covering the entire planet.23 This continuous terrain ruggedness measure is fixed and time-invariant. We also interact this variable with time dummies to capture differential changes over time in costs that vary with terrain ruggedness.
Data processing details
From the FCC fixed deployment data, we extracted information relevant to residential consumer service provision, since competition in the urban residential broadband services market is the focus of this article. We also dropped observations on satellite broadband service providers, for two reasons: first, the Form 477 service data from these ISPs lacks granularity and is provided to the FCC at the state level.24 Second, the satellite data does not reflect either active local marketing efforts at the individual census block level, or capacity constraints that may limit satellite service availability in particular local markets.25
For each provider-census block combination, we calculate the maximum advertised upload and download speeds offered by the provider using either a fixed wireless technology or a wireline technology (or both, when a mix of technology sets is utilized by the provider).26 For each time period, we then calculate three measures of market structure at the census block level that were previously described: (i) number of providers using legacy (cable and DSL, other copper) wireline technologies in serving a block, (ii) number of providers using fiber but not legacy technology, and (iii) number of providers using fixed wireless but not fiber or legacy technology. The total number of broadband providers is simply the sum of these three numbers.27 We calculate analogous market-structure variables for mobile wireless providers.28
Based on this classification of service provider technology, we identify the largest subpopulation of U.S. census block markets: so-called classic legacy (cable/DSL) duopolies, duopoly blocks in which there were precisely two terrestrial “legacy” providers in December 2014—ISPs who made use of cable, or DSL, or both technologies in offering service. Recall our previous observation that in urban areas, almost 60% of urban census blocks had two or fewer providers in 2014: about two-thirds of this group of census blocks were “legacy duopolies” by this definition.
We classify blocks in this manner in order to better accommodate the institutional diversity of broadband service in U.S. block level, micro markets. In many local markets (e.g., in Texas), analysis of the detailed FCC data suggested that new entrants often consisted of small and typically regional firms deploying fixed wireless technology to offer lower quality/speed service at a discount from the price points used by the incumbent legacy duopolists.
The “pure” fixed wireless and fiber providers entering into local, residential block markets in recent years may potentially have had a disruptive impact on market outcomes in blocks that previously were relatively stable legacy wireline duopolies. Our classification of census blocks accommodates simple model specifications that seek to measure how entry may affect service quality in the most dominant (and relatively homogeneous) single category of competitive structures within a complex U.S. national ecosystem of locally defined markets, the legacy duopoly.
We assign to each census block the value corresponding to the TRI values for the grid “cell” containing the official Census “interior point”29 latitude and longitude coordinates of a census block.30
Data Panel
Our analytical dataset is a panel dataset where the basic unit is a service availability record at the provider/census block/technology/time period level. The panel we utilized has five periods: December 2014, December 2015, December 2016, December 2017, and December 2018. Our data set contains approximately 67 million observations and contains information for approximately six million U.S. census blocks, in 3,142 U.S. counties, with over 2000 different broadband providers. This article uses the subset of the data consisting of observations for urban census blocks during endpoint years 2014 and 2018.
We further trim our population of interest to just those urban census blocks that had a legacy duopoly market structure in 2014, and study how the maximum quality of legacy technology broadband service available in 2018 for this same subpopulation of (approximately 1.4 million) census blocks varied in response to entry or exit of ISPs using any terrestrial broadband technology. Our rationale for trimming the sample in this way and focusing on maximum service quality offered in urban duopoly census blocks by ISPs using legacy network technology is related to our understanding of measurement errors in the FCC 477 data, which we discuss in the next section.
Identification Strategy Overview
As just remarked, the population being studied is the almost 40% of U.S. urban census blocks in 2014 that could be described as legacy (cable/wireline) duopolies. We think of entry by new ISPs as a “treatment” applied to local census block markets. This approach is motivated in part by recent changes in FCC subsidy policies for high-cost areas (intended to improve their service quality) that can have the effect of subsidizing new entrants into local markets. Viewed in this light, numbers of entering or exiting firms are the “treatment dose,” with three types of treatments, based on ISP technology type: entry/exit of “legacy” ISPs, entry of new, “pure fiber” ISPs,31 and entry of fixed wireless ISPs.
This taxonomy of entry “treatments” is partially motivated by FCC data showing qualitatively different distributions of maximum download speeds among ISPs using different technologies, with fixed wireless ISPs characterized by a distribution of maximum download speeds that is centered an order of magnitude below median speeds for legacy and fiber ISPs in 2018.32 There was a huge amount of variation in maximum speeds offered by legacy wireline ISPs in 2018; in contrast, variation in maximum speeds offered among fiber and fixed wireless ISPs was very limited in comparison. Best available fiber ISP service quality was extremely concentrated at the high end of the quality spectrum (near 1000 Mbps or faster download speeds), while fixed wireless ISPs mainly offered maximum speeds in the 20–50 Mbps range, with a median speed near the bottom of that range.
By contrast, the distribution of maximum legacy technology download speeds among treated and reference group urban duopoly census blocks prior to any changes in market structure looks very similar. Figure 1 shows the distribution of maximum download speeds in treatment and reference group urban legacy duopoly census blocks in 2014.
Box Plot for Legacy Duopoly Block Maximum Download Speeds, 2014.
Source: Authors’ calculations based on FCC Form 477 data.
Box Plot for Legacy Duopoly Block Maximum Download Speeds, 2014.
Source: Authors’ calculations based on FCC Form 477 data.
Threats to Identification
Our goal is to estimate a medium-run reduced form model that will allow us to analyze how maximum quality (measured by maximum download speed) available in a census block market varies as market structure changes. We recognize that market structure is likely to be endogenous for two reasons. First, there is likely to be some measurement error in our measures of market structure (numbers of providers). Second, unobserved local factors or local shocks that affect service provider quality offer choices may also affect decisions to enter or exit a local market. In both cases, endogeneity (correlation with the model disturbance term) can bias coefficient estimates.
We believe we have two significant sources of measurement error in our market structure variables—numbers of broadband ISPs offering service to households in a local census block market, disaggregated by technology used as reported to the FCC on its Form 477—that must be addressed. The first of these measurement problems is that in rural areas (with census blocks substantially larger than the single physical block bounded by streets typically found in urban areas) census blocks may span hundreds of square miles. The fact that an ISP can serve at least one household in such a large “block” may be a very misleading indicator of its competitive role as a potential service provider to other homes in that far flung rural block. Previous studies and direct observation both suggest that this measurement issue is of great empirical significance primarily in rural areas.33 We address this measurement problem by simply dropping rural census blocks from our sample and studying only urban census blocks (also limiting the relevance of our empirical results to urban areas only).
The second significant set of measurement errors is for fixed wireless providers where, as previously remarked, we expect to observe the vast bulk of ISP mismeasurement within urban areas. We do not address this problem in this article but are currently developing an approach that utilizes instrumental variables, that is, identifying plausibly exogenous, relevant variables associated with variation in wireless ISP entry and exit that do not also directly affect download speeds offered by ISPs using legacy wireline technologies.
To deal with the threat to identification from any unobservable local factors and shocks that affect both quality choice and market structure, we will take two distinct approaches. First, we have included a very large set of control covariates, in order to make it more plausible that conditional on these covariates, the model disturbance term is approximately a mean zero random variable. Equivalently, including this very extensive set of quality demand and cost-shifting controls means that the selection of a given census block into a given market structure “treatment” is likely to be determined by a subset of these same covariates, so that market structure conditional on the included covariates is approximately as good as randomly assigned.
Second, to the extent that any remaining unobservables affecting maximum offered (and advertised) download speeds are approximately time-invariant at the census block level, we can purge potential bias by using census block fixed effects. Use of a large set of time-varying quality demand and cost shifters, along with fixed effects for census blocks, will be our primary identification strategy in this analysis.
Possible time-variation in long run entry cost related to local spatial and economic factors for wireline networks (both legacy and fiber) can be proxied by interacting our 2014 Costquest census block-level greenfield network cost estimates with a time dummy variable. These cost categories seem to be largely based on local geospatial characteristics of census blocks, including terrain, soil characteristics, road and household density, and can be interpreted as primarily geographically determined. Because legacy (cable and DSL) network speed upgrade costs may in part depend on some of the same physical network infrastructure building costs that are the basis for the Costquest network construction measures, we include them as explanatory covariates out of an abundance of caution. For the same reasons, we interact the Terrain Roughness Index (TRI) with time as a proxy for other local, time-varying cost changes linked to terrain roughness that may affect legacy network speed upgrades. Speed upgrade costs in legacy networks are also likely to be tied to changes in costs for network electronics not dependent on the local geography and the natural environment, and a simple time dummy variable should capture these cost shifts, as well as national shifts over time in demand for the highest possible broadband speeds.
Wireline network construction costs should only affect network speeds directly to the extent they also have significant impact on ISP speed upgrade costs. Network construction costs will always affect speed indirectly via market structure, through their role in shaping entry and exit by wireline ISPs, and the resulting degree of competition between ISPs.
Finally, we think of the CAF subsidy vintage variables described earlier in our discussion of the FCC CAF as controls for the average effects of accepted CAF broadband subsidy offers on advertised maximum download speeds in subsidized census blocks. Conditioning on these binary indicator variables is intended to remove a potential threat to identification (of the effects of market structure on download speed) coming from otherwise unobserved subsidies to legacy speed upgrades in high-cost census blocks. It may also be reasonable to suppose that the selection mechanisms for offer and acceptance of CAF subsidies by wireline ISPs are based on observables already included as explanatory covariates in our main statistical model.34
Sample Population Dynamics
Table 1 displays 2014 legacy duopoly urban block population dynamics over time. A small number of singleton blocks not present in both 2014 and 2018 were dropped, creating a balanced sample of about 1.6 million urban blocks. About 22% of blocks in the sample were no longer duopolies by 2018; blocks in the “treatment group” constituted a slightly larger 23.6% of the sample. Some of the 2014 duopoly blocks still classified as duopolies in 2018 were nonetheless in the treatment group because their technology composition changed—for example, a transition from a block with two legacy tech ISPs to a block with one legacy tech and 1 fixed wireless ISP (coded as a net legacy tech exit [legacy wireline ISPs changing from 2 to 1] combined with a fixed wireless entrant [zero to 1] treatment).
ISPs per Urban Block, 2014 Legacy Duopoly Population
ISPs/block: . | 2014 . | 2018 . |
---|---|---|
1 | 0 | 28,782 |
2 | 1,565,177 | 1,223,170 |
3 | 0 | 254,418 |
4 | 0 | 50,623 |
5 | 0 | 8,055 |
6 | 0 | 122 |
7 | 0 | 7 |
All Blocks | 1,565,177 | 1,565,177 |
ISPs/block: . | 2014 . | 2018 . |
---|---|---|
1 | 0 | 28,782 |
2 | 1,565,177 | 1,223,170 |
3 | 0 | 254,418 |
4 | 0 | 50,623 |
5 | 0 | 8,055 |
6 | 0 | 122 |
7 | 0 | 7 |
All Blocks | 1,565,177 | 1,565,177 |
Note: Authors’ calculations based on FCC Form 477 2014 and 2018 Data. The table shows the urban block distribution by the number of ISPs in the block for years 2014 and 2018.
The remaining 76.4% of the sample constituted the comparison group. The treatment group was about 370 thousand census blocks; the reference group about 1.2 million blocks.35
Because the upgrade path to gigabit speeds for DSL networks (as well as some cable networks with old coaxial cable infrastructure) may require replacing or augmenting copper twisted pair (or older coax) wirelines to households with fiber wireline residential connections, we need to distinguish between “pure fiber” entrant ISPs, and legacy technology providers who offer gigabit speed upgrades using optical fiber for residential customer hookups within a given block. (The ISP would then show up in FCC data as using a combination of both fiber and legacy technology to serve that block.) If an ISP uses both fiber and a legacy technology to connect to customers in a census block, we classify that ISP as a legacy network provider to the block and use its maximum speed using only legacy technologies in calculating maximum legacy speed for the block. If the ISP only uses fiber to provide broadband to a block, we classify that provider as a non-legacy, “pure fiber” entrant competitor, unless there were, in 2018, less than two legacy ISPs serving that 2014 duopoly block.
In that latter case, some ISPs offering fiber-only service to a block in 2018 may have previously served that block with legacy technology (and therefore would have been classified as a “legacy network” provider in 2014). Accordingly, we adjust the count of net nonincumbent “pure fiber” entrants over 2014–2018 to be the number of ISPs serving a block using only fiber in 2018, less any reduction in numbers of legacy ISPs (any ISP using legacy cable or DSL technology in 2014) below two (since the entire subpopulation of 2014 “legacy duopoly” census blocks we are studying had exactly two legacy providers and no other ISPs in 2014). We define a new variable for 2018 legacy “upgrade fiber” providers in these cases as equal to the decline in reallocated fiber ISPs.36
Finally, after constructing this additional “treatment” type, we have four sets of market structure “treatment”-related variables for 2018:
Count of number of legacy ISPs. All census blocks (in the 2014 legacy duopoly subpopulation we are studying) have a value of two pretreatment, in 2014, for legacy providers. The comparison group of census blocks (continuing 2018 legacy duopolies) also has a value of 2 in 2018 for legacy ISP count (as well as counts of zero for fiber, upgrade fiber, and fixed wireless “treatment dose”).
Count of “pure fiber” ISPs in 2018. This equals number “pure” fiber ISPs in 2018 less any measured reduction in number of legacy ISPs less than or equal to the fiber ISP count for a census block. This variable takes on value zero in 2014 in all census blocks; zero in 2018 for all blocks without “pure fiber” entrants, count of pure fiber ISPs when a census block is “treated” with “pure fiber.”
Count of legacy “upgrade to fiber” ISPs, as described above. This variable is only nonzero in census blocks with the “legacy upgrade to fiber” treatment in 2018 imputed, for 2018 only, and equal to the count of legacy ISP fiber upgrades. In our sample, the only fiber upgrade “dose” ever observed was 1, a single upgrade to fiber in a census block that previously was a legacy duopoly, but now has only 1 legacy ISP.
Count of fixed wireless ISPs in 2018.
We think of a zero value in 2018 for any of these categories other than legacy (cable or DSL) as a “no treatment” indicator, and 2 as the reference no-treatment “dose” for legacy broadband duopolies. Number of providers can then be viewed as “dose” for four qualitatively different treatments37: legacy entrance/exit, “pure fiber” entry, “upgrade fiber” entry, and fixed wireless entry.38Table 2 shows the distribution of entry/exit types for “treated” 2014 urban legacy duopoly census blocks in 2018.39 Almost 240,000 blocks in our urban legacy duopoly sample experienced some form of wireless entry, about 45,000 some level of fiber entry, and close to 50,000 some form of legacy entry by 2018.
Number of Urban Census Blocks by Entry and Exit Category
Entry/Exit Type . | Number Blocks . |
---|---|
Legacy Exit | 60,141 |
1 or More Legacy Entrant | 48,116 |
2 or More Legacy Entrant | 392 |
1 or More Wireless Entrant | 239,846 |
2 or More Wireless Entrant | 45,479 |
3 or More Wireless Entrant | 6,102 |
1 or More Fiber Entrant | 44,898 |
2 or more Fiber Entrant | 696 |
Fiber Upgrade | 26,117 |
Entry/Exit Type . | Number Blocks . |
---|---|
Legacy Exit | 60,141 |
1 or More Legacy Entrant | 48,116 |
2 or More Legacy Entrant | 392 |
1 or More Wireless Entrant | 239,846 |
2 or More Wireless Entrant | 45,479 |
3 or More Wireless Entrant | 6,102 |
1 or More Fiber Entrant | 44,898 |
2 or more Fiber Entrant | 696 |
Fiber Upgrade | 26,117 |
Note: Authors’ calculations based on FCC Form 477 2014 and 2018 Data. The table shows the urban block distribution by entry/exit type. Census blocks with missing information are excluded from this table. Individual census blocks may be counted within multiple categories.
Urban census blocks in our treatment and comparison groups look very similar in terms of demographic characteristics. Table 3 shows the means of a subset of demographic variables. Census blocks in each group have a similar share of population receiving public assistance income (PAI), percentage of population in poverty, share of non-Hispanic-Latino population, share of occupied housing units, and share of household with a college degree or graduate educational attainment.
Summary Statistics. Mean of Variables Subset for Control and Treatment Groups
. | Control . | Treatment . |
---|---|---|
% PAI | 0.03 | 0.03 |
% Population in Poverty | 0.14 | 0.15 |
% Non-Hispanic-Latino Population | 0.86 | 0.87 |
% Occupied Housing Units | 0.88 | 0.88 |
% College Degree or Higher | 0.29 | 0.27 |
Population per Block | 74.74 | 67.10 |
Housing Units per Block | 31.25 | 28.72 |
Number Observations | 2,865,922 | 962,958 |
. | Control . | Treatment . |
---|---|---|
% PAI | 0.03 | 0.03 |
% Population in Poverty | 0.14 | 0.15 |
% Non-Hispanic-Latino Population | 0.86 | 0.87 |
% Occupied Housing Units | 0.88 | 0.88 |
% College Degree or Higher | 0.29 | 0.27 |
Population per Block | 74.74 | 67.10 |
Housing Units per Block | 31.25 | 28.72 |
Number Observations | 2,865,922 | 962,958 |
Note: PAI stands for Public Assistance Income. An observation is a census block-year.
Econometric Model
We next describe a simple difference-in-differences model. In a fixed effect version of this model, we capture all census-block-specific, time-invariant quality cost or demand shifters as components of a fixed effect. We will also include time-varying explanatory covariates for census blocks, including U.S. census variables relevant to broadband demand and cost (like population and housing units, and density), common time trend variables (to capture time-varying changes in the costs of offering higher legacy network speeds along with shifts in market demands for higher speed), and detailed time-varying measures of census block household and job characteristics relevant to demand for higher broadband speeds. Implicit in this formulation is a “parallel slopes” assumption: that absent treatment, maximum download speeds in treatment group census blocks that were treated with entry/exit in numbers of competitors would have increased, on average, by the same amount as in the comparison census block group.
The simplest “naïve” difference-in-differences model we estimate is
- 1
Sit = constant + btreat e_trti + bN_leg N_legit* t_e + bN_F N_Fit + bN_UF N_UFit + bNw N_Wit + bt_e t_e + uit
where subscript i indexes census block and subscript t indexes time period:
Sit: maximum download speed available in block I at time t, the dependent variable (in mbps)
e_trti: binary variable = 1 if census block is in the treatment group, =0 if comparison group, captures any pretreatment differences between treatment and comparison groups
N_legit: number of ISPs using legacy wireline technology in block i at time t (2 (in 2014 for all blocks, different from two in 2018 for blocks receiving legacy ISP entry or exit “treatment,” > 2 if additional legacy entry, possibly 1 with legacy ISP exit or fiber upgrade)
N_Fit: number of “pure fiber” entrant ISPs in a block at time t (always equals 0 in 2014, possibly > 0 in 2018 if in pure fiber treatment group)
N_UFit: number of “upgrade fiber” ISPs in a block at time t (0 in 2014, possibly 1 in 2018)
N_Wit: number of fixed wireless ISPs in block i at time t, (always =0 in 2014, > 0 in 2018 if in wireless treatment group)
t_e: binary time variable=0 in 2014, 1 in 2018
uit: disturbance term for census block i, time period t
A second version of the model adds on a complete set of control covariates, giving
- 2
Sit = constant + btreat e_trti + bN_leg N_legit* t_e + bN_F N_Fit + bN_UF N_UFit + bNw N_Wit + bt_e t_e
+ b3g tmw_2git + b4g tmw_4git
+ bcaf15 caf15it + bcaf16 caf16it + bcaf17 caf17it + bcaf18 caf18it
+ bdshifters x [demand shifters] + bcost_groupi x [block cost categoryi]
+ bcshiftersi x [block cost categoryi] * t_e
+ bTRI x TRI + bTRIshifters x TRI * t_e + uit
Where
tmw_2git: number of mobile 2G providers serving block i at time t;
tmw_4git: number of 4G providers serving block i at time t;
caf15it: dummy variable taking on value 1 for block i in year t, for t > 2014 if that census block claimed high-cost CAF subsidies for the first time in 2015, 0 otherwise;
caf16it,caf17it,caf18it: dummy variables defined like caf15it but for 2016, 2017, and 2018 respectively;
[block cost categoryi]: set of categorical census block cost dummy variables based on Costquest estimated census block cost group category in 2014;
[demand shifters]: 16 block group level demographic/economic variables from the Census ACS for block i in year t, three FCC staff estimates of annual population, households, and housing units at the block level derived using block group level annual ACS data, and 34 annual RAC for employees with residences within a census block40;
TRI: time-invariant TRI, and TRI interacted with time period, to capture changes in quality/speed upgrade cost related to terrain roughness.
A third version is a fixed effects version of this diff-in-diff specification, where every block has its own fixed effect/intercept, but maintains the assumption of a common change (parallel slopes) over time absent treatment:
- 3
Sit = ai + bN_leg N_legit* t_e + bN_F N_Fit + bN_UF N_UFit + bNw N_Wit + bt_e t_e
+ b3g tmw_2git + b4g tmw_4git
+ bcaf15 caf15it + bcaf16 caf16it + bcaf17 caf17it + bcaf18 caf18it
+ bdshifters x [demand shifters]
+ bTRIshifters x TRI * t_e
+ bcshiftersi x [block cost categoryi] * t_e + uit
Note that the time invariant 2014 block cost category and TRI terms disappear in the fixed effects model, as does the treatment group indicator, since they are now all included in the unobserved fixed effect ai. Inclusion of the fixed effects effectively removes any time-constant, block-level, unobserved heterogeneity in the disturbance term that might be correlated with the treatment-related variables. Each of the above linear model specifications assumes a constant marginal effect on best available quality (maximum download rate) from an additional competitor.
Our preferred specifications are a more flexible, nonparametric and nonlinear version of model specifications (1)-(3) that divides block counts for legacy, fiber, and fixed wireless ISPs into categorical treatment groups: number of legacy ISPs equal to 1, equal to 3 or more, equal to 4 or more (with no treatment baseline equal to 2); number of pure fiber ISPs equal to 1 or more, 2 or more (baseline zero); and fixed wireless ISPs (1 or more, 2 or more, 3 or more, with baseline zero).41 We report both specifications (linear in continuous entry/exit counts, vs. nonparametric categorical) in results shown below.
We do not attempt to analyze the timing of ISP entry and exit in this article. Additionally, limitations in the collection and reporting of FCC Form 477 data t create practical barriers to such an analysis.42
Difference-in-Difference Results
The difference-in-differences models (Equations (1)-(3)) as written above were estimated with maximum legacy ISP speed offered within a census block as the dependent variable, and market structure variables as the “treatment” variables of interest (i.e., N_legit, N_Fit, N_UFit, and N_Wit). In addition, we estimated the same set of model specifications with discretized, categorical indicator variables for different kinds of treatment, with an “I” indicator prefix substituted for the “N” continuous treatment variable name.
The alternative discretized treatment categories created from N_leg were (i) I_L1 (exit leaves only a single legacy ISP in the block), (ii) I_L3 (3 or more legacy ISPs in 2018), and (iii) I_L4 (4 or more legacy ISPs in 2018).43 For fiber ISPs, the discretized market structure variables were (i) I_f1 (1 or more fiber ISPs in 2018), and (ii) I_f2 (2 or more fiber ISPs).
For fixed wireless ISPs (N_W in the continuous linear specification), the corresponding categorical variables created were (i) I_W1 (1 or more fixed wireless ISPs), (ii) I_W2 (2 or more wireless ISPs), and (iii) I_W3 (3 or more). Note that nonlegacy treatment variables all take value zero in 2014 in both treatment and comparison groups of census blocks, and nonzero values only in the treatment group in 2018.
Table 4 summarizes estimation results for each of these models. Since the fiber upgrade to legacy treatment took on only a binary 0/1 value in this sample, the same variable appears untransformed in both the continuous linear and nonparametric categorical model specifications.
Difference-in-Difference and Fixed Effects Model Results
. | Naïve Diff-in-Diff . | Diff-in-Diff w/full Controls . | DiD w/ FE and full controls . | |||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
. | (1) . | (2) . | (3) . | (4) . | (5) . | (6) . | (7) . | (8) . | (9) . | (10) . | (11) . | (12) . |
e_trt | -32.46*** | -10.35 | -5.605 | -5.525 | -23.61*** | -4.507 | 0.243 | 0.277 | ||||
(6.681) | (6.655) | (5.619) | (5.627) | (6.134) | (6.193) | (4.874) | (4.865) | |||||
t_e | 735.3*** | 745.8*** | 748.0*** | 748.0*** | 753.2*** | 762.8*** | 765.0*** | 765.1*** | 756.9*** | 764.5*** | 765.3*** | 765.3*** |
(10.61) | (11.14) | (10.95) | (10.94) | (11.90) | (12.39) | (12.15) | (12.14) | (11.56) | (11.85) | (11.76) | (11.76) | |
N_leg | 172.6*** | 159.9*** | 160.4*** | |||||||||
(15.74) | (15.23) | (17.24) | ||||||||||
N_F | -12.68 | -10.39 | -3.097 | |||||||||
(23.30) | (21.01) | (23.57) | ||||||||||
N_UF | 165.2*** | 290.7*** | 290.0*** | 289.8*** | 143.0*** | 265.4*** | 264.7*** | 264.5*** | 112.6*** | 270.2*** | 269.8*** | 269.8*** |
(29.21) | (34.95) | (34.93) | (34.93) | (28.71) | (34.57) | (34.56) | (34.55) | (28.33) | (33.33) | (33.32) | (33.32) | |
N_W | -1.470 | -16.47 | -5.060 | -17.85 | -23.63 | -24.29 | ||||||
(16.84) | (16.52) | (16.07) | (15.77) | (17.83) | (17.83) | |||||||
I_L1 | -323.8*** | -327.6*** | -327.7*** | -304.7*** | -308.5*** | -308.6*** | -323.5*** | -323.4*** | -323.3*** | |||
(27.91) | (27.61) | (27.60) | (27.85) | (27.52) | (27.51) | (27.16) | (27.20) | (27.20) | ||||
I_L3 | 62.38** | 58.48** | 58.30** | 60.22** | 56.34** | 56.22** | 58.97** | 58.97** | 58.99** | |||
(19.85) | (19.25) | (19.23) | (18.56) | (17.89) | (17.87) | (20.81) | (20.78) | (20.76) | ||||
I_L4 | 25.71 | 24.96 | 24.91 | 31.06 | 30.36 | 30.16 | 43.93 | 43.69 | 43.56 | |||
(22.29) | (22.13) | (22.15) | (22.60) | (22.35) | (22.39) | (34.49) | (34.36) | (34.37) | ||||
I_F1 | -32.44 | -35.43 | -35.88 | -27.17 | -30.05 | -30.49 | -5.610 | -5.150 | -5.211 | |||
(23.52) | (23.16) | (23.17) | (21.52) | (21.18) | (21.19) | (22.98) | (22.96) | (22.94) | ||||
I_F2 | 94.54** | 92.68* | 93.50** | 64.95+ | 63.09+ | 63.90+ | 99.11* | 98.00* | 98.36* | |||
(36.15) | (36.10) | (35.95) | (34.70) | (34.74) | (34.57) | (46.42) | (46.52) | (46.38) | ||||
I_W1 | -38.44 | -32.35 | -39.90 | -34.07 | -37.26 | -35.32 | ||||||
(24.64) | (23.05) | (24.68) | (22.94) | (26.39) | (24.64) | |||||||
I_W2 | 39.91 | 38.99 | 18.40 | |||||||||
(32.69) | (30.63) | (34.74) | ||||||||||
I_W3 | -55.10 | -57.56 | -61.17 | |||||||||
(74.15) | (63.47) | (76.17) | ||||||||||
caf15 | -184.2*** | -184.2*** | -183.6*** | -184.2*** | -201.3*** | -198.2*** | -197.9*** | -198.1*** | ||||
(36.42) | (36.81) | (36.86) | (36.84) | (36.33) | (37.16) | (37.20) | (37.20) | |||||
caf16 | -248.5*** | -245.6*** | -244.4*** | -245.1*** | -226.5*** | -222.2*** | -221.2*** | -221.4*** | ||||
(33.11) | (32.76) | (32.80) | (32.78) | (31.73) | (31.51) | (31.43) | (31.45) | |||||
caf17 | -399.2*** | -390.4*** | -387.7*** | -388.6*** | -326.7*** | -316.0*** | -313.9*** | -314.2*** | ||||
(34.65) | (35.70) | (35.94) | (35.87) | (32.02) | (33.42) | (33.58) | (33.52) | |||||
caf18 | -428.8*** | -423.0*** | -418.5*** | -419.7*** | -403.0*** | -392.0*** | -389.0*** | -389.1*** | ||||
(47.61) | (47.02) | (46.96) | (46.90) | (48.95) | (47.81) | (48.01) | (47.90) | |||||
Census-Block FE | N | N | N | N | N | N | N | N | Y | Y | Y | Y |
Full set of Controls | N | N | N | N | Y | Y | Y | Y | Y | Y | Y | Y |
N | 3130354 | 3130354 | 3130354 | 3130354 | 2928901 | 2928901 | 2928901 | 2928901 | 2798076 | 2798076 | 2798076 | 2798076 |
r2 | 0.777 | 0.779 | 0.779 | 0.779 | 0.788 | 0.789 | 0.789 | 0.789 | 0.900 | 0.902 | 0.902 | 0.902 |
r2_within | 0.777 | 0.779 | 0.779 | 0.779 | 0.788 | 0.789 | 0.789 | 0.789 | 0.887 | 0.888 | 0.888 | 0.888 |
. | Naïve Diff-in-Diff . | Diff-in-Diff w/full Controls . | DiD w/ FE and full controls . | |||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|
. | (1) . | (2) . | (3) . | (4) . | (5) . | (6) . | (7) . | (8) . | (9) . | (10) . | (11) . | (12) . |
e_trt | -32.46*** | -10.35 | -5.605 | -5.525 | -23.61*** | -4.507 | 0.243 | 0.277 | ||||
(6.681) | (6.655) | (5.619) | (5.627) | (6.134) | (6.193) | (4.874) | (4.865) | |||||
t_e | 735.3*** | 745.8*** | 748.0*** | 748.0*** | 753.2*** | 762.8*** | 765.0*** | 765.1*** | 756.9*** | 764.5*** | 765.3*** | 765.3*** |
(10.61) | (11.14) | (10.95) | (10.94) | (11.90) | (12.39) | (12.15) | (12.14) | (11.56) | (11.85) | (11.76) | (11.76) | |
N_leg | 172.6*** | 159.9*** | 160.4*** | |||||||||
(15.74) | (15.23) | (17.24) | ||||||||||
N_F | -12.68 | -10.39 | -3.097 | |||||||||
(23.30) | (21.01) | (23.57) | ||||||||||
N_UF | 165.2*** | 290.7*** | 290.0*** | 289.8*** | 143.0*** | 265.4*** | 264.7*** | 264.5*** | 112.6*** | 270.2*** | 269.8*** | 269.8*** |
(29.21) | (34.95) | (34.93) | (34.93) | (28.71) | (34.57) | (34.56) | (34.55) | (28.33) | (33.33) | (33.32) | (33.32) | |
N_W | -1.470 | -16.47 | -5.060 | -17.85 | -23.63 | -24.29 | ||||||
(16.84) | (16.52) | (16.07) | (15.77) | (17.83) | (17.83) | |||||||
I_L1 | -323.8*** | -327.6*** | -327.7*** | -304.7*** | -308.5*** | -308.6*** | -323.5*** | -323.4*** | -323.3*** | |||
(27.91) | (27.61) | (27.60) | (27.85) | (27.52) | (27.51) | (27.16) | (27.20) | (27.20) | ||||
I_L3 | 62.38** | 58.48** | 58.30** | 60.22** | 56.34** | 56.22** | 58.97** | 58.97** | 58.99** | |||
(19.85) | (19.25) | (19.23) | (18.56) | (17.89) | (17.87) | (20.81) | (20.78) | (20.76) | ||||
I_L4 | 25.71 | 24.96 | 24.91 | 31.06 | 30.36 | 30.16 | 43.93 | 43.69 | 43.56 | |||
(22.29) | (22.13) | (22.15) | (22.60) | (22.35) | (22.39) | (34.49) | (34.36) | (34.37) | ||||
I_F1 | -32.44 | -35.43 | -35.88 | -27.17 | -30.05 | -30.49 | -5.610 | -5.150 | -5.211 | |||
(23.52) | (23.16) | (23.17) | (21.52) | (21.18) | (21.19) | (22.98) | (22.96) | (22.94) | ||||
I_F2 | 94.54** | 92.68* | 93.50** | 64.95+ | 63.09+ | 63.90+ | 99.11* | 98.00* | 98.36* | |||
(36.15) | (36.10) | (35.95) | (34.70) | (34.74) | (34.57) | (46.42) | (46.52) | (46.38) | ||||
I_W1 | -38.44 | -32.35 | -39.90 | -34.07 | -37.26 | -35.32 | ||||||
(24.64) | (23.05) | (24.68) | (22.94) | (26.39) | (24.64) | |||||||
I_W2 | 39.91 | 38.99 | 18.40 | |||||||||
(32.69) | (30.63) | (34.74) | ||||||||||
I_W3 | -55.10 | -57.56 | -61.17 | |||||||||
(74.15) | (63.47) | (76.17) | ||||||||||
caf15 | -184.2*** | -184.2*** | -183.6*** | -184.2*** | -201.3*** | -198.2*** | -197.9*** | -198.1*** | ||||
(36.42) | (36.81) | (36.86) | (36.84) | (36.33) | (37.16) | (37.20) | (37.20) | |||||
caf16 | -248.5*** | -245.6*** | -244.4*** | -245.1*** | -226.5*** | -222.2*** | -221.2*** | -221.4*** | ||||
(33.11) | (32.76) | (32.80) | (32.78) | (31.73) | (31.51) | (31.43) | (31.45) | |||||
caf17 | -399.2*** | -390.4*** | -387.7*** | -388.6*** | -326.7*** | -316.0*** | -313.9*** | -314.2*** | ||||
(34.65) | (35.70) | (35.94) | (35.87) | (32.02) | (33.42) | (33.58) | (33.52) | |||||
caf18 | -428.8*** | -423.0*** | -418.5*** | -419.7*** | -403.0*** | -392.0*** | -389.0*** | -389.1*** | ||||
(47.61) | (47.02) | (46.96) | (46.90) | (48.95) | (47.81) | (48.01) | (47.90) | |||||
Census-Block FE | N | N | N | N | N | N | N | N | Y | Y | Y | Y |
Full set of Controls | N | N | N | N | Y | Y | Y | Y | Y | Y | Y | Y |
N | 3130354 | 3130354 | 3130354 | 3130354 | 2928901 | 2928901 | 2928901 | 2928901 | 2798076 | 2798076 | 2798076 | 2798076 |
r2 | 0.777 | 0.779 | 0.779 | 0.779 | 0.788 | 0.789 | 0.789 | 0.789 | 0.900 | 0.902 | 0.902 | 0.902 |
r2_within | 0.777 | 0.779 | 0.779 | 0.779 | 0.788 | 0.789 | 0.789 | 0.789 | 0.887 | 0.888 | 0.888 | 0.888 |
Notes: Standard errors in parentheses.
p < 0.1
p < 0.05
p < 0.01
p < 0.001
The table shows estimate for three models: Naïve Diff-in-Diff, Diff-in-Diif with full controls, and FE with full controls. The dependent variable is the maximum download speed available in the block (in levels).
Three sets of specifications for the market structure variables are shown: naïve, full controls, and fixed effects with full controls. Within each of these three sets of DiD models, market structure treatment measures for ISPs using different technologies (which bound their quality choices technologically) are parametrized in three different ways: linear in continuous number of firms; nonparametric (and nonlinear) discretized treatment category indicators used for legacy and fiber, but retaining a linear, continuous measure for number of fixed wireless ISPs; and with nonparametric (and nonlinear) discretized treatment effect categories for all technologies. Within this last, third set of model specifications we further differentiate with two variations in how we code fixed wireless entry: a single discrete treatment indicator for any number of fixed wireless entrants, versus a suite of indicator variables for 1 or more, 2 or more, or 3 or more fixed wireless entrants.
Discussion of DiD Results
One striking aspect of these results is the broad similarity in the estimated coefficients of greatest interest across vastly different sets of controls and model specifications. With the nonparametric specification, in particular, estimated coefficients with fixed effects and a full set of controls are in most cases not very different from those in the simplest, naïve DiD model. Figure 2 visualizes point estimates for effects of entry on maximum speed available in a census block in 2018, by entrant technology type, for fixed effects, full controls models. Both linear continuous and nonparametric (and nonlinear) categorical specifications generate nearly identical predictions of speed improvement for the legacy duopoly comparison reference group over 2014–2018 and diverge only in assessing the impact of different market structure “treatments” in census blocks moving away from legacy duopoly in 2018. This is driven by the sample composition, since the legacy duopoly comparison group accounts for three-quarters of the observations used in estimation. Note that in the nonparametric categorical model specification, market structure categories were truncated at two fiber entrants and three wireless entrants (so two means two or more for fiber; three means three or more for fixed wireless), to avoid creating categories with tiny numbers of nonzero indicator observations in an otherwise large sample.44 Since we have quite a large sample size (about 1.4 million census blocks observed at two points in time in the models with a full set of controls), there is little reason to favor the more parsimonious continuous linear specification, and we focus our discussion on the more flexible (and nonlinear) nonparametric, categorical specifications of market structure change “treatments.”
Entry/Exit Vs. Change in Maximum Download Speed, 2014–2018.
Source: Authors’ calculations based on FCC Form 477 data.
Entry/Exit Vs. Change in Maximum Download Speed, 2014–2018.
Source: Authors’ calculations based on FCC Form 477 data.
In the comparison group (unchanged legacy duopoly blocks), maximum available speeds in 2018 grew by over 750 megabits per second relative to their 2014 level in all models. Losing a legacy competitor over this time span (going from a duopoly to a monopoly) in the nonlinear specification was associated with a reduction of about 320 Mbps in 2018, that is, so maximum available speed would have risen by only 440 Mbps after a block’s conversion from duopoly to a legacy tech monopoly.45 In contrast, gaining a legacy competitor, that is, going from legacy duopoly to legacy triopoly would be expected to raise maximum available download speed by almost 60 Mbps, for a net gain of about ~810 Mbps in 2018 (with fixed effects).46 Going from triopoly to four or more legacy ISPs produces a statistically insignificant additional gain of about 44 Mbps, for an estimated increment of roughly 850 Mbps in 2018 relative to 2014. This asymmetric pattern—going from duopoly to monopoly having large effects, going from duopoly to triopoly have smaller but significant impacts, and with larger numbers of competitors bringing only even smaller additional changes has been observed previously in empirical studies of oligopoly price competition.47
A surprising feature of these results is the high explanatory power (R2) of the market structure “treatment” variables, with or without additional covariates added to the model. In the naïve DiD model with no additional controls, market structure variables alone explain about 78% of variance in the outcome (speed) variable. Adding many dozens of covariates raises that to 79%, and even adding an additional ~1.6 million census block fixed effects elevates explained variance by only an additional 10 percentage points, to roughly 90%! Differences in 2018 market structure seem to be the most important predictor of variation in improvements to legacy broadband service quality registered across U.S. urban duopoly census blocks over the 2014–2018 period.
We note that there is not a huge amount of variation in the outcome variable in 2018. In that year, almost 85% of urban census blocks had maximum available download speeds above 900 Mbps, and 55% speeds above 940 Mbps (see Flamm and Varas, 2021). There was much more speed variation back in 2014, however, and the fixed effect models we are using explain changes in download speed over time within census blocks. This variation seems to be very well predicted by changes in market structure—numbers of competitors—alone.
Consistent with previous empirical studies, more competition in practice seems to translate into higher quality. In the linear model, an additional legacy competitor “overbuilding” in a census block added on an additional 160 to 170 Mbps to maximum available legacy speed in 2018 (or, with the unrealistic symmetry imposed by the linear specification, subtracted this same increment if an incumbent duopolist exited). Allowing for nonparametric asymmetry (and nonlinearity) in these effects lowers this to a 60 Mbps boost from an additional legacy competitor, and a much greater 320 Mbps drop with a transition to legacy monopoly. An upgrade to fiber by an incumbent legacy ISP actually seemed to reduce maximum available legacy speed slightly relative to what an undisturbed legacy duopoly would offer in a census block in 2018, by about −50 Mbps in a model with block fixed effects.48
Unexpectedly, adding a single “pure fiber” competitor to a block seemed to have no significant effect on legacy ISPs’ service quality. In contrast, adding two or more fiber competitors in a census block did increase maximum available broadband speed by a statistically significant (at 5%) 90 Mbps. Limited or no impact on legacy broadband speed after entry by a fiber ISP has also been observed in other empirical studies.49
Finally, wireless competitors apparently played little role at all in influencing the maximum speeds offered in urban census blocks, based on these models. Point estimates are generally negative, often near zero, and statistically insignificant.50 This is consistent with our earlier observation that the distribution of maximum wireless ISP speeds is centered well below that for wireline ISPs, including both legacy and fiber connections to the home. Despite the theoretical ability of wireless ISPs to offer very high speeds over short distances, this apparently is not often an economically practical network design option. There was little difference between using a single wireless variable (either a count, or an indicator for any wireless competitor) and multiple indicators. However, as previously noted, measurement error for wireless competitors may be affecting these estimates, a problem we have not addressed in this study.
The CAF control coefficients are included in this table because of their strikingly large, negative size, and their statistical significance, even though the CAF program is not the subject of this article and the number of census blocks receiving CAF subsidies was relatively small. The CAF subsidies over this time period were a subsidy to incumbent legacy ISPs intended to induce them to upgrade their service speeds. These coefficient estimates suggest they may have perversely had the opposite effect—perhaps by signaling to both DSL incumbent subsidy recipients and their potential competitors that a challenger was unlikely to win an unsubsidized battle for census block markets based on service quality. The issue is interesting and important, and merits further study.51
Interpreting Our Results: Summary
We have adopted identification assumptions that seem reasonable (common trends in outcome—maximum download speed—over time, on average, within all 2014 urban legacy duopoly blocks absent entry/treatment, after conditioning on a linear index of controls that includes census block-specific fixed effects and a very large and extensive set of time-varying cost and demand shifters). As is well understood, in a world where heterogenous entry/treatment effects vary across census blocks, the difference-in-difference model we are using estimates an average treatment effect on the treated census blocks.52 If we were also comfortable assuming that entry effects were homogeneous and constant across all census blocks, then the average effect of entry on all 2014 legacy duopoly census blocks would be identified; if not and effects are heterogenous, then we are measuring the average effect of entry on the subpopulation of treated blocks (roughly 24% of 2014 legacy duopoly blocks).
For fiber upgrades by a legacy ISP, we see a slowdown in maximum speed—about −50 Mbps. The 2018 difference in speed between a legacy monopoly and a legacy duopoly is a whopping 320 Mbps. Going from two to three legacy ISPs adds about 60 Mbps to maximum available legacy speed in the block. Additional entry beyond that has a small but statistically insignificant additional impact.
The effect of increasing competition on speed is relatively small compared to the time trend observed across the industry. Maximum available legacy speed in duopoly markets which did not experience changes in market structure changes increased by around 760 Mbps from 2014 to 2018, on average. We believe this primarily reflects ongoing technological improvements to the performance of legacy networks that were implemented widely across the residential broadband industry.53
We also observe that one fiber entrant produces no statistically significant effect, but two fiber entrants are associated with an almost 90 Mbps speed improvement in legacy ISP speeds. Adding fixed wireless competitors to a census block seems to have had a slightly negative but statistically insignificant effect on legacy ISP speeds.
As remarked above, we believe that measures of numbers of fixed wireless competitors are flawed by relatively large measurement errors. Since this measurement error would be expected to bias estimated coefficients toward zero, the possibility remains that the true impact of fixed wireless entry might actually be even more negative.
Controls for the operation of the Connect America subsidy program in a census block have large, statistically significant, and an intriguingly negative influence on legacy ISP service quality in the small (~1200) number of recipient blocks in our sample.
Conclusion and Policy Implications
We found strong evidence that market structure (competition) is very important in explaining the evolution of maximum available speeds available from legacy technology ISPs serving U.S. urban census blocks over 2014–2018. Differences in 2018 market structure seem to be by far the most important predictor of variation in improvements to legacy broadband service quality registered across U.S. duopoly census blocks over this period. Maximum offered legacy ISP speeds with a single fixed wireless ISP entrant has a negative, but statistically insignificant effect, which we speculate is due to the lower quality service entrant shifting the competitive focus toward price, and away from higher speed.
With exit, and a duopoly census block devolving into a legacy monopoly, maximum legacy ISP speed offered dropped by about 320 Mbps. Going from two to three legacy ISPs adds about 60 Mbps. A second fiber entrant resulted in about 90 Mbps added to maximum speeds available from a legacy ISP. These preliminary results argue strongly for the proposition that increased competition in local broadband markets associated with entry by new ISPs using high-speed wireline technologies (fiber, coax) resulted in economically and statistically significant quality improvements.
Conversely, a continuation of reductions in competition in local markets—the apparent trend since 2017, per the official FCC data—are likely to diminish future quality improvements. For this reason, the prepandemic surge in mergers and acquisition among U.S. ISPs raises policy concerns. One effect of the COVID-19 pandemic has been to make clear just how important adequate and robust broadband service quality can be for local communities struggling to adjust to costly economic and social shocks.
Our results are in some respects welcome, but not surprising, given a previous empirical literature that shows mixed, but generally positive effects of increased competition on product quality. More striking is the relatively small impact of increased competition on legacy ISP service quality when compared to the impacts of underlying time trends, which likely reflect dramatic quality-adjusted cost declines in commercially available high-speed broadband electronics over the 2014–2018 period, as well as continued gains in adoption of high-speed Internet-based services among residential consumers.54 The continuing impact of technological innovation and changes in market structure on future broadband service quality are something of an unknown, given that the most significant technological changes currently underway seem to be in wireless technology.
Fixed wireless broadband ISP entry in this study was found to have little effect on quality competition in this population of legacy duopoly census blocks. Based on this we speculate that price competition, in lieu of quality competition, is likely to become a more salient focus for broadband rivalry if and when new (5G) fixed wireless entrants using lower cost technology to provide lower quality service come into the market in force. For the highest quality broadband service we study in this article, by contrast, which primarily utilized advanced wireline (optical fiber and coaxial cable) technologies to connect to individual residences, higher available quality seems to have been the primary outcome of greater competition among peer ISPs offering services more similar in quality. Research currently in progress will address measurement error for fixed wireless ISPs, determine the merits of our speculation about the direction of resulting bias, and provide a better estimate of the impact of entry by lower quality entrants on legacy ISP quality improvement.
Because of a general lack of spatially disaggregated, granular public data on broadband prices, we are silent on another interesting and important question: How increased competition affected pricing of broadband services to residential consumers? Though not the focus of this article, we also uncovered intriguing statistical evidence that subsidies to legacy wireline ISPs through earlier vintages of the CAF subsidy program may have worked to dampen legacy ISPs’ maximum available download speeds in recipient census blocks, possibly by reducing entry threats from nonincumbents. Both questions merit further research.
APPENDIX A
The FCC Form 477 Data
Since 2000, the Federal Communications Commission (FCC) has been collecting data on the availability of broadband service across the United States. The original May 2000 Form 477 required “facilities based” broadband service providers to provide a list of all five-digit ZIP codes in which they provided service to at least one end user, and defined high-speed service connections as those with data transfer rates exceeding 200 kilobits per second (kbps).55 The data was collected twice a year, in June and December.
In 2008, the program was revised significantly: the geographic unit for data collection became the census tract. Fixed wireline and wireless ISPs, as well as satellite providers, now had to report total subscribers by census tract, broken down by technology and speed tier, and the share of those subscribers that were residential customers. Mobile wireless broadband providers were asked to list the census tracts that “best represent” their broadband service footprint, by service tier quality.56 Beginning around this time, public data released by the FCC, derived from this form, was expanded to include the numbers of mobile wireless providers serving a census tract (as with fixed broadband data, results for the range of 1–3 providers were censored, as were numbers of mobile providers exceeding a cap of 7), and a qualitative coding of the share of households in a census tract with broadband service was added to the public version of the Form 477 dataset.
Shortly afterward, in 2009, in association with passage by Congress of the American Recovery and Reinvestment Act of 2009 (which contained approximately $7.2 billion for broadband-related investment projects), the FCC was directed to create a National Broadband Plan (NBP), released in March 2010. In parallel, when the FCC released 477 data for June 2009, the public release of the 477 data also began reporting numbers of broadband providers exceeding a faster “NBP” threshold of 3 Mbps download, 768 Kbps upload speeds, by census tract.57 In August 2013, the FCC revised its 477 program once again. Most importantly, it required fixed service ISPs (including satellite as well as fixed wireline and wireless providers) to submit lists of all census blocks (vs. tracts) “where they offer service” (vs. where they have actual subscribers!). Finally, in late 2014, the FCC revised rules for its Connect America Fund (subsidies supporting fixed rural broadband) to require minimum 10/1 Mbps connection speeds. Form 477 data submitted from 2014 on report maximum speeds (by technology) of fixed broadband service ISPs, by census block.
APPENDIX B
Connect America Fund Transitional Subsidy Mechanisms, 2014–2018
Connect America Fund (CAF) broadband subsidies were offered by the FCC to a subset of “grandfathered” census blocks that had previously received legacy high-cost voice subsidies. The FCC began this process by analyzing the new and more detailed Form 477 data that became available at the census block level after 2014. (See Appendix A.) Information on broadband availability tabulated from prior year Form 477 data was available to determine which census blocks within a “high-cost” incumbent local exchange carrier (ILEC) local voice service territory (or “study area”) lacked adequate broadband service, and therefore potentially justified a subsidy offer. The first claims by price-cap ILECs for CAM-based subsidies at the census block level begin the following year, in 2015.
Though simplified somewhat through this process, the distribution of Universal Service subsidy funds remained quite complex. In 2011, there had been no less than seven different support mechanisms distributing “high-cost” subsidies to three different categories of recipient telephone voice service providers. There were the ILECs in rural areas that were subject to “price cap” regulation of their interstate service charges (generally larger national or regional providers, the “price cap” ILECs); so-called “rate-of-return” ILECs (primarily small local telephone companies serving rural areas); and finally, so-called “competitive eligible telecommunications carriers” (CETCs), in rural areas served by all other firms (other than the historical ILECs present at the time of the breakup of the U.S. Bell System telecommunications monopoly in the 1980s).
The cost models used by the FCC to formulate subsidy offers combined engineering cost models with cross-sectional data on census block characteristics. The resulting cost model allowed for variation across blocks in population, housing counts, and spatial density. Based on census block characteristics, incumbent voice service providers were offered model-driven funding levels in exchange for provision of specified broadband speed levels to existing network customers upon request, and to additionally accept an obligation to serve new locations in subsidized blocks if the improved service level could be provided “at reasonable cost.”58 Simple economic logic predicts that profit-maximizing incumbent providers would first accept these subsidy offers for those particular census blocks where net profit after subsidy would have been highest, and decline these offers when the expected net return was too low, or even negative.
From 2012 through 2014, a set of “Phase I” transitional funding mechanisms began shifting both old and new Universal Service subsidy funds to new mechanisms supporting broadband deployment to high-cost areas serviced by price cap ILECs. Initially, approximately $486 million in funding was allocated to Phase I subsidies to broadband deployment (initially defined by a 4 Mbps download/1 Mbps upload standard, relaxed to a 3 /.768 standard after the initial phase of funding) to about half a million high-cost locations lacking “broadband” (defined by speeds at or exceeding .768/.200). Acceptance of the funding (and service obligations) by the ILEC was voluntary, with frozen “legacy” high-cost support available as an alternative.
From 2015 on, the “price cap” ILECs were offered “Phase II” support based on an FCC cost model for broadband provision (the “Connect America Model,” or CAM) on a voluntary basis. The Phase II support obligated price cap ILECs to deploy a minimum 10/1 broadband standard to census blocks lacking an unsubsidized competitor offering broadband service. Frozen legacy voice subsidies to price cap providers in high-cost areas were cut sharply during Phase II.59
Most recently, after 2018, the FCC added significant funding ($1.5 billion over 10 years) to support eligible providers with winning bids in a new reverse auction mechanism. Providers bid the subsidy value at which they would accept broadband service obligations in specified high-cost “price cap” census blocks in 45 states. These were an FCC-defined list of census blocks where either the Connect America Model support had been declined by the incumbent price cap provider, or where costs were deemed “extremely high,” or blocks had been removed for other reasons from the FCC’s previous CAM offers to price cap carriers. This auction, completed in late 2018, was known as the Connect America Fund (CAF) Phase II auction.60
The new reverse auction mechanism is a significant qualitative break from previous subsidy mechanisms. It subsidizes new entrants into high-cost census blocks where an incumbent ILEC had declined a previous FCC subsidy offer. Previously, only incumbent voice services providers had been offered subsidies in exchange for broadband service provision commitments in high-cost areas.
The second large group of incumbent voice service providers serving rural high-cost areas, the rate-of-return carriers, continued to have access to subsidies at frozen legacy levels through 2016. Beginning in 2017, rate of return carriers in high-cost areas were offered substantial new subsidies for broadband deployment, as a voluntary option, based on another FCC cost model (the “Alternative Connect America Model,” or ACAM). Service speed obligations ranged from under 4/1, to 4/1, to 10/1, and up to 25/3, and were expected to bring broadband to 714,000 new locations by 2026.61 Legacy universal service support claims from rate-of-return carriers began to decline after the ACAM broadband offers started in 2017.62
FOOTNOTES
We think of a higher “quality” product or service in the context of models of vertically differentiated products, with a set of vertically ordered characteristics such that all consumers prefer higher quality to lower quality at a given price, with at least some consumers willing to pay more than others for a higher quality product.
We note, however, that it was plausibly argued that the typical U.S. household in pre-COVID markets would make relatively little use of broadband bandwidth over about 100 megabits per second. See for example S. Ramachandran et al.
The canonical model of vertical product differentiation is “a three stage game in which a number of firms choose firstly, whether to enter an industry; secondly, the quality of their respective products, and thirdly, their prices” (Shaked and Sutton, 12). In the abstract, we think of the sequential stages of this game as the long, medium, and short runs.
Appendix A briefly reviews the history and limitations of the FCC Form 477 data we are using to measure both market structure and service quality; Appendix B summarizes the Connect America Fund program.
Published data on the relationship between advertised speeds and delivered speed, by ISP, based on a small national sample of households, has been reported publicly by the Federal Communications Commission since 2011. See https://www.fcc.gov/general/measuring-broadband-america.
In the Communication Marketplace Report, the FCC analyzes services of delivering voice, video, audio, and data services, among others.
A limitation in Whitacre and Gallardo is that they use county-level data for their analyses.
Wilson. Also studies multimarket contact effect on download speeds offered by ISPs.
Molnar and Savage. Use variables related to the sunk cost of entry as excluded instruments. In particular, they use the number of roads, intersections, houses, and bedrock and wetland terrain measures. These variables could also directly affect the cost of offering or upgrading to higher speeds for both incumbent and entrant ISPs. In addition, they assume joint normality with zero covariances between unobservables affecting broadband ISP entry and quality choice.
Roughly 1,200 urban census blocks out of the 1.4–1.6 million used in our analysis. (Numbers of census blocks used in the statistical analysis vary slightly with availability of covariates included in the model.) For a brief history of CAF funding, see FCC, “In the Matter of Connect America Fund,” Report and Order, FCC 14-190, W.C. Dockets No. 10-90, 14-58, 14-192, December 2014. In 2011, the FCC issued a report and order transforming the process by which Universal Service Fund subsidies of communications services to “high-cost” areas would be undertaken. While previously voice telephone service to (mainly) rural households had been supported by subsidies from Universal Service funds (generated by fees paid by all U.S. voice service purchasers), the Commission now expanded subsidies to include both voice and broadband service in so-called “high-cost” areas. Accompanied by extensive public comment—and litigation—a subsequent 2014 report and order set out the specifics of how this reformed system was to operate. A summary of the transitional mechanisms for this new set of subsidy mechanisms—the Connect America Fund (CAF)—is given in Appendix B.
“Each year, the FCC conducts a survey of the fixed voice and broadband service rates offered to consumers in urban areas. The FCC uses the survey data to determine the local voice rate floor and reasonable comparability benchmarks for fixed voice and broadband rates for universal service purposes.” https://www.fcc.gov/economics-analytics/industry-analysis-division/urban-rate-survey-data-resources.
“See April 2014 Connect America Order, 29 FCC Rcd at 7070-75, paras. 59-72. C.f. 47 CFR § 54.202 (requiring any carrier petitioning to be federally-designated ETCs [Eligible Telecommunications Carriers] to “[c]ommit to provide service throughout its proposed designated service area to all customers making a reasonable request for service” and to certify that it will provide service `on a timely basis’ to customers within its existing network coverage and `within a reasonable time’ to customers outside of its existing network coverage if service can be provided at reasonable cost).” (FCC, 68)
The currently novel, more competitive CAF Phase II reverse auction mechanism mentioned in Appendix B did not come into use until after the end of our sample period.
https://www.fcc.gov/general/broadband-deployment-data-fcc-form-477. Note that prior to 2014, service providers only reported their data for areas in which they actually had subscribing customers. See Flamm and Varas. For a detailed discussion of the 477 data definitions and how they have changed over time, along with an analysis of historical trends in competition in local markets.
The problem is particularly noticeable in Texas, where wireless providers seem to pop up in urban areas for a year or two, then exit from the FCC dataset. The metropolitan Dallas area has one such “phantom” wireless provider popping up in the FCC Form 477 records as serving most of the Dallas area for a short period, before vanishing with no evident trace. One of the authors’ homes is in a census block in the Austin suburbs (literally a single suburban city block) which has long been a good example of the classic cable-DSL duopoly. For a couple of years in the middle of our study period, a wireless provider apparently reported serving this block on its Form 477 but appears to have made no active attempt to win any customers or even advertise availability of its service to residents, to the best of our knowledge. A likely but problematic scenario leading to such misleading reporting would be a wireless service provider setting up an antenna on high ground in a rural area on the periphery of a city in order to serve primarily rural customers, then reporting all city blocks within line-of-sight of this antenna, as determined by some automated computer calculation, as “offered service” on FCC forms.
There will be two separate records if the service is available for both residential and business consumers.
The FCC’s Mobile Deployment Data is also collected twice per year by the FCC through the Form 477. For each data collection period, each mobile broadband provider must report all geographical areas where they can provide service. A major difference from the Fixed Deployment Data is that the mobile data is reported using shapefiles, so covered areas can be smaller (or larger) than census blocks. In this data, each provider is identified by its “Doing Business As” (DBA) name. Providers have to report a shapefile for each technology they use to offer service (e.g., WiMAX, LTE, etc,). The Mobile deployment data is also available every six months from December 2014 to December 2018, except for the period June 2015.
Available at https://opendata.usac.org/High-Cost/High-Cost-Connect-America-Fund-Broadband-Map-CAF-M/r59r-rpip.
The 2014 vintage baseline value is embedded in a census block fixed effect.
Costquest is an FCC contractor that constructed estimates of census block level costs per household for building out a greenfield fiber network in telephone operating company territories using a highly detailed engineering model combined with demographic, geospatial locational, and other data. These cost estimates were designed to be used by the FCC in formulating its CAF subsidy offers. The bulk of these network costs is the building of physical optical network infrastructure plant (installation of conduits and fiber cables in trenches, or fiber cables on poles).
This data was briefly made available to the public for download on the Costquest web site from roughly mid-2019 through mid-2020. These data are, unfortunately, no longer available on the Costquest web site.
This index was originally developed by Riley et al. (1999) and measures the amount of elevation difference between adjacent cells of a geographical grid.
The geographical terrain ruggedness is described and available for download at https://diegopuga.org/data/rugged/.
Conversation with FCC data administrator, Washington, D.C., September 2018.
Because of the change in Form 477 definition of areas served (from actual customers pre-2014, to “could serve”), in most urban markets and even in many rural census blocks there was a noticeable uptick in numbers of providers after 2013 (typically, by two or more providers). Much of this increase was related to satellite‐based ISPs now being included in the ISP counts, even for urban census blocks in which they rarely if ever sold a competitive service offering to paying customers. (The theoretical service footprint of the major satellite‐based ISPs covers most of the continental United States) The FCC clearly took note of this, since beginning in 2014, publicly released data on ISPs by census block distinguish between counts including and excluding satellite‐based service providers.
Also, while satellite ISPs can provide upload and download speeds for digital content comparable to fixed terrestrial broadband service, latency (the round‐trip time to send, then receive a single digital packet) is about 20 times greater with satellite service. This significantly affects interactive applications, like gaming, or point and click applications (like moving or zooming dynamically on a map or menu). Based on stated preference survey data, one study estimates that a representative consumer would be willing to pay $8.66 monthly to avoid the increased latency associated with moving from terrestrial to satellite broadband service, at a given download/upload speed. Liu and Wallsten.
If a provider does not offer service with a given technology set (either wireline or wireless), then its maximum advertised speed is defined as missing, not zero.
Although we do not observe the share of households served by various individual providers within a census block, the inverse of our market structure measure (1/N)—with N the sum of the provider counts across technology types—defines a lower bound on the Herfindahl–Hirschman Index (HHI) of concentration (with 1 defined as the upper bound HHI were all households in a block served by a single provider).
Using the FCC’s mobile deployment data, we can also calculate analogous mobile broadband market structure measures for each successive mobile technology standard (which is also associated with significant service speed improvements). In particular, we have defined market structure measures for 2G or better and 4G or better mobile broadband data service providers. These measures are cumulative: our 2G count includes providers offering 2G or better service, while our 4G count covers 4G or better service. So a 4G provider, for example, would be included in both 2G, and 4G provider counts, while a 2G-only provider would be included in the 2G count, but not in the 4G count.
The U.S. Census latitude and longitude coordinates for a census block interior point are effectively the centroid, modified to lie within the census block boundaries for unusual block shapes where it would otherwise not.
Because the raw data is in shapefile format, we use the user-written Stata command geoinpoly to count how many providers provide mobile broadband service in each census block (Picard).
“Pure fiber” because in urban areas today, most ISPs use some amount of fiber in their local distribution networks. We use “pure fiber” to describe an ISP connecting to homes in a census block only using fiber, that is, not also using legacy cable/DSL technology to connect to other homes in the same census block, which would indicate a legacy network that is being upgraded to fiber-to-the-home (FTTH).
For empirical support for this assumption, see Trostle et al., Ford (2019), US General Accountability Office (2018), and US General Accountability Office (2020).
If true, this would make it reasonable (after conditioning on observed covariates) to assume that remaining unobservables affecting an ISP’s choice of maximum service speed (as embedded within a statistical disturbance term) are uncorrelated with the CAF subsidy variables, allowing us to think of the CAF subsidy coefficients as measures of another type of causal “treatment” effect.
Availability of nonmissing values for other covariates used in different model specifications resulted in from 1.4 to 1.6 million census blocks being used in statistical models.
Underlying this correction is the default assumption that “old legacy wireline ISP networks exiting just as fiber enters” is generally a sign of a speed upgrade to an incumbent legacy network. So, for example, if unadjusted number of legacy ISPs declined from 2 to 1 and fiber ISPs increased by 1 (the most common situation with apparent legacy ISP exit in the actual FCC data) from 2014 to 2018, number of “upgrade fiber” ISPs was coded as 1, and “pure fiber” ISPs as 0. Alternatively, if the number of legacy ISPs declined from 2 to 1 and fiber ISPs increased by 2, the number of “upgrade fiber” ISPs in 2018 would be coded as 1, and “pure fiber “entrant” ISPs as 1. Maximum speeds offered by either “pure fiber” or “upgrade fiber” ISPs were not considered when calculating maximum legacy speeds for a census block.
Note that because of a very considerable level of merger and acquisition activity over the 2014–2018 period, frequent changes from year-to-year in holding company names, and the existence of a large number of small, privately held ISPs, it is quite challenging to link ISP Form 477 ISP holding company names from one year to the next. Using aggregate statistics on numbers of ISPs, by network technology, to indirectly measure entry and exit is a next best alternative.
This simple specification implies that the effect of any treatment is approximately proportional to dose size (count), with no treatment equivalent to a zero dose.
The total effect of an upgrade to fiber by a single legacy ISP in a census block, in this framework, adds the effect of +1 dose of the legacy ISP “upgrade fiber” treatment to the effect of a −1 dose of legacy ISP count change treatment.
Table 2 is restricted to census blocks for which all variables used in the regression analysis are available.
The 16 included block group level demographic/economic demand shifter variables from the ACS were: total block group population, percent male population, non-Hispanic/Latino, non-Hispanic Afro-American, American Indian, Asian American, other non-Hispanic, below the poverty line, receiving public assistance income, has a high school degree education, has a college degree or better education; median age, median family income, median housing value, percent of housing units that are occupied, percent of population with no telephone service. The RAC variables used include job shares by categorical age range, by race category, by educational attainment category, by ethnicity category, and by gender category of employees resident in a census block.
Absent this truncation, we would have had very small numbers of census blocks in high entry count categories.
In our current data set, we only measure year of entry or exit, and are unable to specify month of entry within the year. In addition, particularly for smaller, nonpublicly listed ISPs, we would be unable to distinguish mere ownership name changes from bona fide entry/exit.
Effects of entry are additive in this specification—the difference in effect between 2 (the omitted baseline level) and 4 or more ISPs in a census block would sum the coefficients of I_L3 and I_L4. I_L3 is the difference in effect between 2 (the omitted baseline level) versus 3 legacy ISPs; I_L4 is the difference between 3 vs. 4 or more ISPs.
The fourth column under each model in Table 3 also portrays a specification with the wireless entry market structure variable even more truncated, to a single entry category: one or more fixed wireless entrants.
We see about 30,000 census blocks in our sample (the entire treatment group numbers about 370,000 blocks) go from two legacy providers to a single legacy provider. Many of these appear to be AT&T, Windstream, Centurylink, Verizon, and other legacy DSL ISPs abandoning any attempt to sell their oldest and slowest DSL service to new customers in selected census blocks they do not wish to make further investments in, or owners of older cable infrastructure (like Wehco Video, Mediacom, Cogeco, Vyve, Wavedivision, West Alabama TV Cable) either ceasing operation or abandoning sales of broadband service to new customers in selected urban census blocks (as uneconomic, presumably).
We see almost 50,000 census blocks in our sample with new legacy entry. In some cases these appear to be resellers connecting to existing legacy infrastructure (in particular, DSL circuits connecting to local incumbent telephone company networks), while in other cases spatially adjacent competitors seem to be “overbuilding” into another ISPs territory. For example, in about 4,000 urban census blocks, WOW! and Verizon were replaced by WOW!, Charter, and Frontier. In roughly 8,000 urban Tennessee census blocks, AT&T and either Comcast, Charter, or Infostructure faced new competition from Ecsis using DSL or a combination of DSL and wireless. In roughly 3,300 urban census blocks, Comcast and AT&T were joined by Telephone Electronics Corporation. In California, resellers like Sonic and Raw Bandwidth connected customer premise equipment to leased AT&T local network infrastructure and competed successfully with both AT&T and a local cable ISP.
See Bresnahan and Reiss. Four or more competitors brings a smaller speed increment and larger p-value.
A reduction of −52.6 Mbps is our point estimate, with a p-value of .015. One scenario as to why a fiber upgrade by a legacy ISP might reduce observed 2018 maximum available speed in an urban block is that an upgrade to fiber by one of the legacy ISPs reduces the probability of a nonincumbent fiber-based ISP entering a census block (supposing that upgrades to maximum offered speeds in “traditional” legacy duopoly blocks are in part undertaken to preempt possible entry by fiber-based nonincumbent ISPs). Already having fiber available may make a market less attractive for a nonincumbent fiber ISP considering entry, somewhat reducing the preemptive motive to upgrade maximum speeds on the part of incumbent legacy ISPs.
Kotrous and Bailey. One scenario that may explain this result is that many of the “single fiber entrant” blocks could be incumbent DSL providers switching the entirety of their internal network within a census block to fiber to the home, thus looking like a “pure fiber” entrant in 2018, while simultaneously reselling use of their obsolete copper telephone lines to third party DSL ISPs. This could lead to an average effect that is close to zero in “single fiber entrant” blocks—since fiber upgrades by legacy ISPs would not be detected by our legacy-to-fiber-upgrade coding algorithm in this case, and we would be averaging negative (DSL upgrade to fiber) and positive (bona fide pure fiber) effects.
As previously noted, measurement error in fixed wireless ISP counts is likely an issue in our data, and the usual classical measurement error arguments suggest that this would bias this coefficient toward zero, that is, the coefficient of a properly measured wireless ISP count variable would be more negative.
CAF money was allocated using “study areas” derived from telephone exchange boundaries. We are not surprised to find that roughly 1200 urban duopoly census blocks received some CAF funding as a consequence. The definition of these blocks as urban is the 2010 Census Bureau classification. We have estimated our model after dropping the roughly 1,200 CAF-recipient blocks, and find no significant or substantive change to coefficients or their estimated standard errors. We prefer to report the model with the blocks included, because with roughly 300 observations per estimated CAF parameter, we find the coefficients’ signs, size, and significance interesting, provoking useful thought as to what may be going on here.
Lechner, 182–184.
Many readers will be familiar with the adoption of new DOCSIS 3.0 and 3.1 standards that raised legacy cable network speeds over this period. U.S. readers will be less familiar with new, higher speed DSL standards that were introduced in 2006 (VDSL2, up to 200 mbps), 2015 (VDSL2 Vplus35b, up to 300 mbps), and 2014 (G.fast, with speed of 500-1000 mbps over distances <100 meters, 1000 Mbps over distances of 50 meters). Commercial equipment incorporating these newer ITU standards was offered by multiple vendors. Because speed drops rapidly with distance, the main economic use case for this technology is in short wire runs from fiber-connected pedestals or cabinets to individual homes or multi-unit dwellings with embedded interior twisted-pair copper telephone wiring. It seems to have been adopted in urban areas in Europe on a reasonable scale. It was also adopted in the US, particularly in dense urban use cases, as announced publicly by Centurylink, AT&T, and Frontier (see, for example https://www.fiercetelecom.com/telecom/gfast-passes-over-3m-premises-33-providers-advance-roll-outs-trials-says-analyst). AT&T marketed its use of VDSL2 and G.fast under the umbrella term “IPBB” technology, though it limited use to guaranteed speeds far lower than those available with shorter cable runs than AT&T apparently has standardized on it on US service territory. (“IPBB includes ADSL2+, VDSL2, G.Fast and Ethernet technologies delivered over a hybrid of fiber optic and copper facilities which provides subscribers with significantly faster download speeds compared to traditional DSL connections.” See https://about.att.com/sites/broadband/performance.) While deployment in the US has been much more limited than in Europe and Asia, use of these faster DSL technologies nonetheless shows up in reported speeds on the FCC Form 477 data for “pure” DSL ISPs. Generally, in less dense US urban markets, the economics seem to have favored former DSL ISPs generally increasing speeds by deploying fiber connections to homes.
For statistical evidence on dramatic variations over time in historical quality-adjusted price decline rates for communications hardware, see Flamm (1989), Gordon, Flamm (1999), Byrne and Corrado.
The history of FCC’s Form 477 program is summarized in FCC, Modernizing the FCC Form 477 Data Program, Final Rule, published in the Federal Register on August 13, 2013. The FCC originally published this data in the form of a list of zip codes in which broadband service provider reported any customers, and the number of providers reporting customers in the zip code (with numbers of providers in the 1 to 3 range per zip code censored, and reported as an ‘*’. Unfortunately, zip codes with no end users reported were not shown in the public list, which made these data of limited utility to researchers and policymakers, since zip codes were created and withdrawn by the postal service frequently, and zip codes were never actually used to define stable spatially defined areas. Researchers attempting to use these data were forced to devise ad hoc schemes to associate postal zip codes with Zip Code Tabulation Areas (ZCTAs) defined by the Census during decennial census years, in order to try to figure out what areas had no service at all. After 2004, the FCC began reporting “zero provider” zip codes, which enabled researchers to at least enumerate the universe of zip codes being considered by the FCC for its public reports.
FCC (2013). The speed tiers reported by mobile wireless providers appear to have corresponded to what generation (2G, 3G, 4G non-LTE, 4G LTE) of mobile wireless technology was available to serve customers in a census tract.
The National Broadband Plan actually set a 4/1 Mbps benchmark, but the 3/.768 Mbps tier in the National Broadband Map that was being constructed at the time was the closest speed tier to this benchmark, and ended up becoming the original “NBP broadband speed.”
“See April 2014 Connect America Order, 29 FCC Rcd at 7070-75, paras. 59-72. C.f. 47 CFR § 54.202 (requiring any carrier petitioning to be federally-designated ETCs [Eligible Telecommunications Carriers] to “[c]ommit to provide service throughout its proposed designated service area to all customers making a reasonable request for service” and to certify that it will provide service “on a timely basis” to customers within its existing network coverage and “within a reasonable time” to customers outside of its existing network coverage if service can be provided at reasonable cost).” FCC, REPORT AND ORDER, ORDER AND ORDER ON RECONSIDERATION, AND FURTHER NOTICE OF PROPOSED RULEMAKING, FCC 16-33, WC Dockets No. 10-90, 14-58, 01-92, March 2016, p. 68.
See FCC, Universal Service Monitoring Report 2018, Table 3.3. The CAM support to price cap carriers was reasonably large— set at $1.7 billion annually over six years and expected to result in deployment of 10/1 broadband to 3.5 million locations by 2020, See V. Gaither, p. 7.
See FCC, “Connect America Fund Phase II Auction (Auction 903),” available at https://www.fcc.gov/auction/903.
The broadband service obligations were complex. “Carriers who elected this option will have the certainty of receiving specific and predictable monthly support amounts over the 10 year support term (2017–2026). Those that elected model support must maintain voice and existing broadband service and offer at least 10/1 Mbps to all locations fully funded by the model. They must also offer at least 25/3 Mbps to a certain percentage of those locations by the end of the support term. In addition, carriers must also offer at least 4/1 Mbps to a certain percentage of capped locations (where caps on ACAM subsidy levels were in effect) by the end of the support term, and provide broadband upon reasonable request to the remainder.” https://www.usac.org/hc/funds/acam.aspx. See also Gaither, p. 8.
See FCC, Federal-State Joint Board on Universal Service, Universal Service Monitoring Report 2018, Table 3.2.