Telecommunications Policy 38 (2014) 1011–1023 Contents lists available at ScienceDirect Telecommunications Policy URL: www.elsevier.com/locate/telpol Broadband's contribution to economic growth in rural areas: Moving towards a causal relationship$ Brian Whitacre a,n, Roberto Gallardo b, Sharon Strover c a b c Department of Agricultural Economics Oklahoma State University Oklahoma State University, Stillwater, OK 74075, USA Mississippi State University, USA University of Texas, USA a r t i c l e i n f o abstract Available online 11 July 2014 This paper uses recent data on both broadband availability and adoption to empirically gauge the contribution of broadband to the economic growth of rural areas of the United States over the past decade. Availability data from the National Broadband Map aggregated to county level is used in conjunction with county-level adoption data from Federal Communication Commission. Economic variables of interest include median household income, number of firms with paid employees, total employed, percentage in poverty, and the percentage of employees classified as either creative class or non-farm proprietors. A propensity score matching technique (between a “treated” group associated with various broadband thresholds and a control group) is used to make preliminary causal statements regarding broadband and economic health. Growth rates between 2001 and 2010 for different economic measures are tested for statistical differences between the treated and non-treated groups, restricting the analysis to non-metropolitan counties. Results suggest that high levels of broadband adoption in rural areas positively (and potentially causally) impacted income growth between 2001 and 2010, and negatively influenced unemployment growth. Similarly, low levels of broadband adoption in rural areas lead to declines in the number of firms and total employment numbers in the county. Broadband availability measures (as opposed to adoption) demonstrate only limited impacts, suggesting that future broadband policies should be more demand-oriented. & 2014 Elsevier Ltd. All rights reserved. Keywords: Broadband Rural Average treatment effects Economic growth 1. Introduction The diffusion of broadband Internet access across America during the 2000s brought with it a significant amount of concern that rural areas might be left behind in terms of the availability, adoption, and benefits of this technology (Malecki, 2003; Parker, 2000; Strover, 2001).1 The presence of a rural–urban broadband “digital divide” is well documented in the economic literature (Dickes, Lamie, & Whitacre, 2010; Strover, 2011; Whitacre & Mills, 2007). Several studies have suggested that broadband access positively impacts employment as well as other quality of life issues (health care, education, social $ This study was supported by the National Agricultural and Rural Development Policy (NARDeP) Center under USDA/NIFA Grant no. 2012-70002-19385. Corresponding author. þ1 405 744 9825. E-mail address: brian.whitacre@okstate.edu (B. Whitacre). 1 The FCC's definition of broadband has changed over time. Historically, the definition has been 200 kilobits of data transfer per second (kbps) in at least 1 direction. The most recent (2010) definition is 4 megabits (Mbps) download and 1 Mbps upload. This report incorporates various thresholds, depending on the data used for analysis. n http://dx.doi.org/10.1016/j.telpol.2014.05.005 0308-5961/& 2014 Elsevier Ltd. All rights reserved. 1012 B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 linkages) in rural areas (Katz & Suter, 2009; LaRose, Strover, Gregg, & Straubhaar, 2011; Stenberg et al., 2009); however, many analyses were based on hypothetical assumptions or case studies. Until recently, very little reliable and useable broadband infrastructure data has been available, and assessments of programs designed to improve broadband access and adoption are quite limited. Contemporary empirical evaluations of the economic impacts of broadband in rural areas are generally lacking. State and federal policies related to broadband investment and deployment remain contested. As the funds flowing to additional advanced broadband infrastructure under the American Reinvestment and Recovery Act end and as the Federal Communication Commission (FCC) initiates changes to universal service, questions regarding where American broadband infrastructure stands and how high speed capability contributes to economic development are receiving new attention. Better data from the National Telecommunications and Information Administration (NTIA)/Census Current Population Supplements – and the more detailed data that NTIA's investments in statewide mapping efforts provide – can be used to deliver better responses to questions regarding broadband's contributions to economic growth. This study uses recent data from the National Broadband Map (NBM) and the FCC to assess the potentially causal relationship between broadband availability/adoption and measures of economic growth, specifically focusing on rural areas. Propensity score matching techniques are used to match non-metropolitan counties with high/low levels of broadband availability and adoption as of 2010 with otherwise similar counties. The growth rates of a variety of economic development variables between 2001 and 2010 are then compared across groups. The results demonstrate that “high adoption” non-metro counties had faster growth in median household income and lower growth in unemployment rates, while “low adoption” non-metro counties saw lower growth in both the number of firms and total employment. In general, thresholds for broadband adoption are shown to be more important to economic growth measures than those for availability. While this methodology moves the discussion of causality forward, we are careful to note that we cannot make strict claims of a causal relationship, due to (1) the potential influence of unobservables and (2) limitations of the NBM data. 2. Recent literature on broadband and economic growth Many recent studies in the economics literature have attempted to demonstrate broadband's influence on productivity or economic gains, although most have been focused on aggregate or urban economies. In one of the earliest (and most widely cited) studies, Lehr, Osorio, and Gillett (2005) concluded that relative to communities without broadband, communities with consumer broadband between 1998 and 2002 experienced larger levels of growth in employment, numbers of businesses, and businesses in IT-intensive sectors. Gillett, Lehr, Osorio, and Sirbu (2006) similarly demonstrated a positive relationship between broadband availability and employment/business growth – especially growth in IT-related businesses. They did not, however, find any relationship with wage levels. Each of these studies noted that the data available at that time were primarily supply-side, and that better data on demand (i.e. adoption) were sorely needed. Holt and Jamison (2009) provide a useful summary of multiple studies between 2000 and 2009 on this topic, generally finding support for the position that broadband has been positively associated with economic outcomes. No studies discussed by Holt and Jamison looked extensively at the relationship in rural areas, however, and few delved into causality. In particular, the Lehr and Gillet studies have often been critiqued for potential endogeneity problems. As Kolko (2012) notes, these early studies were careful to caution that their results do not imply causality – but warns that without an explicit focus on this relationship, suggesting policy implications is difficult. Several current studies have attempted to more rigorously assess the causal relationship between broadband and economic development. Koutroumpis (2009) incorporates a simultaneous equations model to disentangle the relationship between broadband infrastructure and GDP (country-level) growth. Using data from 22 OECD countries over the period from 2002 to 2007, Koutroumpis found a significant causal positive link between broadband infrastructure and growth. This link became stronger as thresholds for broadband penetration increased. However, OECD data are notoriously weak, and country-level relationships are prone to measurement problems. Kolko (2012) focused instead on broadband's causal relationship with local economic growth, including its effect on employment and wage levels. Recognizing that broadband could have heterogeneous impacts on employment (due to the degree an industry is reliant on information technology), Kolko meshed FCC broadband provider data with North American Industrial Classification System (NAICS) employment data at the zip-code tabulation area level from 1999 through 2006. Nearly all industries demonstrated a positive relationship between increasing broadband providers and employment growth over this time; the strongest relationships were found in IT-intensive industries such as information and finance. However, this study found no evidence that areas with faster broadband expansion had higher household incomes or better employment rates. Similarly, Jayakar and Park (2013) found no evidence that broadband deployment significantly impacted changes in unemployment rates between 2008 and 2011. None of the Lehr et al., Gillett et al., Koutroumpis, Kolko, or Jayakar and Park studies focused explicitly on rural parts of the country. The most thorough examination of broadband's role in rural America is likely Stenberg et al. (2009), who focused on broadband use by consumers, communities, and businesses. One finding, using FCC data, is particularly noteworthy. Comparing rural counties with relatively high levels of broadband by the year 2000 with otherwise similar rural counties, they found that the counties with early access to broadband had higher levels of growth in wage and salary jobs, non-farm proprietors, and private earnings between 2002 and 2006. They were careful to caution, however, that their research does not necessarily imply causality. Stenberg et al. also summarized ways that rural communities and businesses can potentially benefit from broadband, including research on distance education, telehealth, and telework. Kuttner (2012) concurs with B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 1013 the notion that broadband is important for rural areas, but focuses instead on attempting to quantify the opportunity costs of not having broadband in these communities by using hypothetical examples. Similarly, Whitacre (2011) demonstrates the economic importance of telemedicine in rural areas (which requires broadband), but his analysis is based on simple case studies. Further, while most of the literature suggests that broadband is important for sustaining and improving rural economies, some work has questioned its relative value. In particular, Galloway (2007) implies that rural businesses do not tend to seek growth opportunities, which is more important for job growth than broadband availability. In general, the studies related to broadband and rural economic development are limited. Most analysis related to broadband and economic outcomes are focused on aggregate or urban measures, with investigations specific to rural typically being more anecdotal or hypothetical. Further, nearly all of the papers attempting to quantify broadband's impact on the economy use measures of broadband availability, such as broadband penetration or number of lines per capita. The relative importance of adoption (versus availability) is left unanswered. This paper seeks to address these shortfalls and provide empirical evidence of a potentially causal relationship between broadband availability/adoption and economic growth in rural areas. 3. Data This paper meshes recent broadband availability data with recent adoption data. The specific data sources used are 2010 versions of FCC County-level broadband adoption data (county-level broadband adoption rates) and National Broadband Map infrastructure availability data (aggregated to county level). Each of these sources is detailed below. 3.1. FCC County-level adoption data Since 2008, the FCC has provided categorical data on county-level household broadband adoption rates, along with measures of the number of broadband providers for residential fixed (wired) connections. One of the most useful features of these data is that they can be easily meshed with other county-level sources, such as demographic data provided by the Census or economic measures provided by the Bureau of Economic Analysis (BEA) or elsewhere. Counties are also readily classified as non-metro, micro, and non-core, allowing for a lower level of analysis for more rural parts of the country.2 Counties are considered metropolitan if they have a core community of at least 50,000 people or 25% of their workforce commutes to a neighboring core; micropolitan if they have an urban core of at least 10,000 up to 49,999 people or 25% of their workforce commute to a neighboring core; and noncore if they do not have a core of at least 10,000 people. There are 3073 counties in each year of the FCC data, of which 2037 are non-metropolitan (671 micropolitan and 1366 non-core).3 The FCC data also include information on two distinct speed thresholds for “broadband” – one defined under the traditional measure of at least one direction with 200 kbps, and another under a faster definition of 768 kbps download, 200 kbps upload.4 The spatial nature of the data allows for informative maps to be drawn. The FCC broadband adoption data are split into 5 categories based on the proportion of households that connect to the Internet with a high-speed connection: 0–19.9% adoption, 20–39.9% adoption, 40–59.9% adoption, 60–79.9% adoption, and 80–100% adoption. While this results in a loss of fidelity regarding the actual percentage of households that adopt, it does provide useful data in assessing the extent of the digital divide. It is worth noting that this primary variable of interest deals with residential fixed (wireline) broadband connections – therefore, wireless or phone connections are not included. Although the FCC does provide data on the number of residential fixed providers, for confidentiality reasons counties with between 1 and 3 providers all have the same designation. Since the National Broadband Map data discussed below do not have this limitation, we defer to their provider counts. We use adoption data from 2010 to model broadband's impacts on economic growth measures that can be captured at the county level. 3.2. National Broadband Map (NBM) data Fall 2010 data5 from the National Broadband Map were utilized to obtain average values for the maximum advertised download/upload speeds and unique number of providers at the county level. The NBM is an online database that allows users to assess broadband availability at the neighborhood (census block) level. The NBM data has been critiqued on several points, namely that it is provided by infrastructure carriers who have an incentive to overstate their service areas, that a 2 We recognize the difference between rural/ urban (defined at the community level) and metro / non-metro (defined at the county level). Our data is county-oriented, so we generally speak in terms of metro/micro/non-core; however we still use the term “rural” to connote a lack of population density. 3 We mesh Virginia independent cities with the counties where they reside. 4 This speed (768 kbps down, 200 kbps up) was adopted by the FCC at one point as a definition for broadband, and BTOP likewise used it for reporting purposes. The most current broadband speed definition the FCC uses is 4 Mbps for download and 1 Mbps upload. 5 This dataset does not include census tracts larger than 2 square miles. 1014 B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 census block is considered served if even one customer in that area has access to broadband, and the manner in which empty census blocks are handled (Grubesic, 2012). These caveats may inflate the availability rates for some rural areas since a small portion of those areas may receive the same level of broadband service as a neighboring urban community. Ford (2011) also illustrates the impact that problems in this dataset, like measurement errors and sample selection bias, could potentially cause. He concludes that use of the NBM data do not permit any strong causal conclusions. Nevertheless, this data represents a marked improvement from previous data collection efforts related to broadband infrastructure provision. This study focuses on several variables from the NBM: maximum advertised6 upload/download speeds and number of providers. However, since data are available at the block group level, aggregation to the county-level was necessary. In order to achieve this, Microsoft Access and Excel software were utilized. First, due to the size of the datasets, Microsoft Access was utilized to “break up” the dataset into smaller sub-datasets so these in turn could be analyzed in Excel. Second, a unique identifier was assigned after combining the holding company unique number and the county-level FIPS code. Third, pivot tables in Microsoft Excel were used to obtain the unique number of providers as well as the average maximum advertised upload/download speeds at the county level. Data provided for the NBM resulted in another useful measure: the percentage of the population for which no wired broadband infrastructure was available. These data, referred to in the text that follows as “no broadband,” are only available for 2010 and are based on an alternative definition of broadband (3 Mbps download, 768 kbps upload) than is typically used elsewhere. However, this measure is quite useful in providing information about broadband availability; such a measure cannot be gleaned from county-level numbers of providers. Only wireline technologies were used for this measure due to concerns about the accuracy of the mobile wireless broadband data (FCC, 2012).7 Again, for the purposes of this paper, these data were aggregated to the county level. 3.3. Demographic and economic data This paper seeks to answer whether broadband availability and/or adoption contribute to (and potentially cause) economic health outcomes in rural areas. Recent studies (Goetz, Fleming, & Rupasingha, 2012; McGranahan, Wojan, & Lambert, 2011; Stephens & Partridge, 2011) point to the percentage of self-employed (i.e. nonfarm proprietors), the percentage of employees classified as creative class, income levels, and employment levels as useful measures of rural economic health. Therefore, we use these indicators as well as standard ones (such as poverty rates and number of firms) to better understand the relationship between economic growth in rural areas and broadband adoption/availability. We also include a wide array of demographic data that may be related to broadband adoption such as age category, race, natural amenity ranking, and county dependencies such as farming or manufacturing. These data come mostly from the Census and the Bureau of Economic Analysis. 3.4. Maps and descriptive statistics Fig. 1 displays the 2010 FCC adoption data from a geographic perspective. Several states exhibit low levels of adoption, notably those in the South (Georgia, Mississippi, and parts of Louisiana, Texas, and Oklahoma). Very high levels of broadband adoption exist in the Northeast, and also near Denver in Colorado. Interestingly, most states have pockets of counties with high levels of adoption, and there is a general spatial trend among the data (Global Moran's I value of 0.15, p-score 0.00). Many of the counties with low levels of adoption are lightly populated and have lower income levels. In fact, the average county population in 2010 for counties with the lowest adoption levels (o20%) is 11,760 (compared to the national average of 25,020 for all non-metro counties). Similarly, the average median household income level in these counties is $34,500 compared to $39,500 for all non-metro counties. Fig. 2 displays the NBM availability data related to the number of unique fixed residential providers per county. Again, some states demonstrate relatively low numbers of providers (West Virginia, Kentucky, North and South Dakota, Wyoming) while others generally have high numbers (Ohio, Indiana, California, Michigan). Overall, there is not an overly strong correlation between household adoption rates and the number of residential providers (correlation coefficient ¼0.31) and the relationship is even weaker for non-metropolitan counties (0.17). We also look at the percentage of the population in each county without access to broadband capability, and how these trends differ by metro/micro/noncore status (Fig. 3). Here, broadband is defined as 3 Mbps download speed and 768 kbps upload, a higher threshold than that used in the previous maps. As expected, most metropolitan counties have high levels of broadband availability (only about 3% of the metropolitan population lacks it), while the non-core areas have the worst (26% of the non-core population lacks availability). There are large pockets of micro and non-core counties with very poor levels of broadband availability in several states, notably the mountainous terrain of West Virginia and portions of many Midwest states. 6 Advertised maximum speeds were utilized rather than typical speeds for two main reasons. First, data availability is higher when using advertised maximum speeds and second, according to the FCC's “Eighth Broadband Progress Report”, there is no significant difference between advertised maximum speeds and typical speeds. The report finds that, “most of the broadband providers studied deliver actual speeds that are generally 80 to 90 percent of advertised speeds or better.” (Federal Communications Commission (FCC), 2012, p. 56) 7 According to the FCC report, “…we have concerns that providers are reporting services as meeting the broadband speed benchmark when they likely do not. … although mobile networks deployed as of June 30, 2010 may be capable of delivering peak speeds of 3 Mbps/768 kpbs or more in some circumstances, the conditions under which these peak speeds could actually occur are rare” (Federal Communications Commission (FCC), 2012, pp. 25–26). B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 1015 Fig. 1. County-level household broadband adoption rates, 2010. Source: FCC Form 477 Data, 2010. As another way of looking at this data, Fig. 4 displays the spectrum of broadband availability categories across metro, micro, and non-core counties. This figure shows that there are wide discrepancies in broadband availability across these designations. While more than 40% of metropolitan counties have the highest level of availability (with less than 2% of the population lacking access), only 5% of non-core counties are in this category. And perhaps more strikingly, nearly 30% of all non-core counties have more than 40% of their population lacking access to wired broadband infrastructure. 3.5. Defining broadband thresholds In order to assess the relationship between broadband and economic growth, we develop several thresholds related to broadband adoption and availability. Our hypothesis is that high or low levels of broadband adoption/availability may promote or deter economic activity within the county. In particular, we define high and low thresholds for (1) broadband adoption, (2) broadband availability (% of population with access), (3) number of broadband providers, and (4) average advertised download speed.8 Table 1 summarizes these thresholds and the percentage of non-metropolitan counties in each category. 23% of non-metro counties achieve the “high-adoption” threshold (household adoption rates of over 60%), while 27.8% are classified as “low-adopting” (household rates of less than 40%). In terms of availability, 41% of non-metro counties have more than 85% of their population with broadband available to them, and are classified as “high-availability.” 14.5% have more than 50% of their population without access to wired broadband, and are classified as “low-availability.” Similar cutoff points are applied to the number of residential broadband providers (high threshold: 5; low threshold: 2) and average advertised download speeds (high threshold: 10 Mbps; low threshold, 3 Mbps). 8 Although these thresholds are arbitrary, they were defined to ensure that a significant amount of observations existed across low and high categories; and that the differences in categories had significance in the real world. For example, a difference between a 3 Mbps and a 5 Mbps download speed may not be noticeable to most users, but the difference between 3 and 10 Mbps likely would be. 1016 B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 Fig. 2. Number of residential broadband (wired) providers, 2010. Source: National Broadband Map, 2010. As a preliminary way of investigating how achieving these thresholds has affected economic prosperity, Table 2 summarizes how the percentage change between 2001 and 2010 for eight economic measures has varied across the high and low categories. These statistics demonstrate that while higher levels of economic improvement are often found in counties meeting the “high” broadband thresholds, this is not always the case. For example, non-metro counties meeting the high threshold for broadband adoption experienced higher growth in median household income levels, number of firms, and total employment when compared to non-metro counties that did not meet either the high or low threshold. However, these same counties also experienced lower growth rates in the percentage of non-farm proprietors. Similarly, under the broadband availability category, non-metro counties with high levels of availability experienced more growth in the percentage of employees in the creative class and in the overall number of firms. However, they also had lower growth rates in median household income levels and the percentage of non-farm proprietors, and had higher growth rates of poverty and unemployment. Comparable trends can be found with the high threshold for the number of residential providers (higher growth in the creative class, but lower levels of income and higher rates of poverty/unemployment). Non-metro counties with high levels of average advertised download speeds demonstrate mostly positive comparisons, with higher levels of creative class employees, lower losses in the number of total employees, and lower increases in poverty rates. Similar stories can be told regarding the “low” broadband thresholds: low adoption is associated with lower growth in the number of firms and the number employed, but higher rates of non-farm proprietors. Counties with lower levels of broadband availability surprisingly demonstrate higher growth in income levels and the number of firms, but also show higher growth in poverty. The statistics provided in Table 2 represent a first step in assessing whether achieving specific broadband thresholds might influence economic outcomes in a non-metropolitan county. There is, however, a significant amount of self-selection occurring: counties with high levels of income or employment growth may see their broadband adoption rates also rise. To begin discussing causal relationships, more rigorous statistical techniques are required. The following section discusses the use of a propensity score matching technique, which, along with the FCC and National Broadband Map data, is used to assess the impact of various broadband thresholds on the eight economic indicators discussed above. B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 1017 Fig. 3. Percent of population with no broadband availability by metropolitan status, 2010. Source: National Broadband Map, 2010. 4. Propensity score matching (average treatment effects) The technique implemented to estimate broadband's (arguably) causal contribution to economic growth is known as propensity score matching, used in the context of estimating average treatment effects (ATE). The ATE measures the average causal difference in outcomes between a treated group and a control group (Rosenbaum and Rubin, 1982). The previously defined thresholds of broadband availability or adoption serve as “treatment” indicators – for instance, having a high level of broadband adoption (460%) or having a high level of average download speed available (410 Mbps) in your county. The “controls” would be otherwise similar counties that do not meet this criterion. In general, if ΔY1 and ΔY0 represent the changes to economic indicators to areas that have and have not met the broadband criterion, respectively, then the ATE (or more precisely, the ATE on the Treated or ATET) is written as ATET ¼ EðΔY t1 j BBt ¼ 1Þ EðΔY t0 j BBt ¼ 1Þ ð1Þ where BBt equals 1 for areas that meet the broadband availability criteria (treated) and 0 for areas that do not (non-treated). The second term in Eq. (1) is non-observable: it is the expected value of the economic indicator change for control group if in fact they had achieved the broadband threshold of interest. Our modeling strategy seeks to produce a control group that is measurably similar to the treated group, but did not achieve the broadband threshold. Put another way, we can observe either ΔYt1 or ΔYt0 for a particular place, but not both, since each county will have either met or not met the broadband threshold of interest. This implies that there is self-selection into the treatment group, which would typically cause biased estimates of the treatment's impact (Wooldridge, 2002).9 To obtain unbiased estimates, an assumption of conditional independence is applied (Imbens, 2004), which means that there are no unobservable differences between areas that meet the broadband threshold and those that do not. This is not a trivial assumption, since there are many unmeasurable factors 9 Another way of interpreting the self-selection issue is that there is some unmeasured variable (including the presence of a broadband ‘champion’ in the community) that influences whether two otherwise similar counties end up in different treated versus non-treated groups. 1018 B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 Fig. 4. Percentage of residents with no broadband availability by metropolitan status, 2010. Source: National Broadband Map, 2010. Table 1 Broadband Threshold Definitions and Percentage of Non-metro Counties in Each. Threshold Non-metro average Broadband adoption category (1–5 category) High (Category 4,5: 460% adoption) Low (Categories 1,2: o 40% adoption) 2.95 Broadband availablity (% with access) High ( 485% with access) Low ( o50% with access) 0.74 Number of residential broadband providers High (5 þ) Low (1–2) 4.30 Average advertised download speed (1–11 Category) High (Category 7þ: 410 Mbps) Low (Categories 2–5: o 3 Mbps) 5.72 % of non-metro counties in category 0.230 0.278 0.410 0.145 0.402 0.223 0.376 0.310 (active community leaders, providers with/without local ties, intensity of provider advertising) that could potentially influence levels of broadband availability and adoption. There are also other unmeasured factors (such as local development policies or programs) that could influence the outcome (i.e. economic) variables of interest. In fact, the prospective influence of these unobservables is a primary reason why this methodology cannot make strict claims of causality, although it does make advances over simple regression. However, the existing literature has successfully identified the most important determinants of both broadband access and adoption – and there is general consensus about these factors. To estimate the average treatment effect, each ‘treated’ county needs a comparable, nontreated counterpart. To accomplish this, the ATET technique seeks to “match” counties that met the broadband criterion with otherwise similar communities that did not. The first step in doing so is to estimate the propensity score – that is, the likelihood of meeting the broadband adoption/availability criteria. Most applications in the statistics literature use a logit model to estimate this propensity score, where the conditional probability of meeting the broadband threshold is modeled on observable predictors such as the socioeconomic variables that might influence adoption or availability. A wide array of literature has already assessed this topic; we draw from many of these studies to develop the socio-economic characteristics to be included in our propensity score model. These characteristics include income and education levels, age, race/ethnicity, and level of urbanization (Aron & Burnstein, 2003; Flamm & Chaudhuri, 2007; Hitt & Tambe 2007; Horrigan, 2007, 2009; Whitacre & Mills, 2007; Whitacre, 2010).10 The outcomes of these models are then used to match treated and non-treated counties by creating blocks of counties with similar propensity scores. That is, we are looking for groups of non-metro counties with similar likelihoods of achieving the broadband threshold as predicted by 2001 values of income, education, number of firms, levels of non-farm proprietors, and other variables (including growth rates during the 1990s). The primary difference between the two groups is that one actually achieved the threshold while the other did not. A test developed by Becker and Ichino (2002) is used to determine whether the treated and non-treated counties in each block have the same distribution of covariates – essentially ensuring that the matches are in fact ‘good.’ The literature suggests various methods for matching, the simplest of which is nearest-neighbor. This technique matches treated and non-treated units by searching 10 Shadish (2013) demonstrates that including “effective” variables (widely known to affect the treatment) in propensity score specification is crucial for reducing bias. B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 1019 Table 2 Summary statistics on changes in economic growth (2001–2010) across broadband threshold categories. Non-metropolitan categories High Broadband adoption %Δ Median HH income %Δ Non-farm proprietors %Δ Non-farm proprietor income %Δ Creative class %Δ Number firms %Δ Total employed %Δ Poverty %Δ Unemployment rate Broadband availability (% of pop with access) %Δ Median HH income %Δ Non-farm proprietors %Δ Non-farm proprietor income %Δ Creative class %Δ Number firms %Δ Total employed %Δ Poverty %Δ Unemployment rate Number of residential providers %Δ Median HH income %Δ Non-farm proprietors %Δ Non-farm proprietor income %Δ Creative class %Δ Number firms %Δ Total employed %Δ Poverty %Δ Unemployment rate Average advertised download speed %Δ Median HH income %Δ Non-farm proprietors %Δ Non-farm proprietor income %Δ Creative class %Δ Number firms %Δ Total employed %Δ Poverty %Δ Unemployment rate ( 460%) 0.231 0.207 0.090 0.041 0.027 0.001 0.221 0.763 ( 485%) 0.209 0.231 0.089 0.048 0.002 0.040 0.256 0.855 (5 þ ) 0.202 0.221 0.050 0.050 0.016 0.056 0.284 0.853 ( 410 Mbps) 0.219 0.243 0.075 0.052 0.012 0.037 0.202 0.814 n Neither n nn nnn nnn nnn nnn n nnn nn nnn nnn nnn nnn nnn nnn nnn nn n nnn 0.221 0.234 0.069 0.040 0.013 0.048 0.216 0.784 0.226 0.262 0.056 0.036 0.026 0.052 0.125 0.746 0.231 0.254 0.057 0.037 0.002 0.041 0.172 0.759 0.212 0.237 0.079 0.041 0.001 0.061 0.244 0.794 Low (o 40%) 0.226 0.300 0.066 0.035 0.036 0.081 0.176 0.772 (o 50%) 0.264 0.244 0.080 0.015 0.001 0.047 0.185 0.641 (1–2) 0.254 0.280 0.142 0.016 0.015 0.038 0.121 0.666 (o 3 Mbps) 0.244 0.260 0.065 0.018 0.017 0.043 0.171 0.711 nnn nn nn nnn nnn nnn nn nnn nnn nnn nn nnn nnn nnn nnn nnn nn nnn nnn nnn Statistically significant difference from the “Neither” category at p ¼ 0.01 level. Statistically significant difference from the “Neither” category at p ¼0.05 level. Statistically significant difference from the “Neither” category at p ¼0.10 level. nn nnn for the closest propensity score between the two groups, and can be altered to include nearest ‘groups’ of neighbors to prevent outliers from skewing the results. The results presented in the following section use the five closest neighbors from the comparison group. As a robustness check, we also include kernel matching estimators. These estimators offer a solution in the case of potentially large discrepancies between neighbors by calculating the distance between a treatment case and various control cases, and giving higher weight to the control cases that are closer. To ensure that the logit model capturing the likelihood of meeting a particular broadband threshold is not influenced by broadband investments already made, relatively deep lags are used in each specification. In particular, the probability of meeting each of the broadband thresholds is modeled on 2001 county characteristics, which is before most private cable or phone companies began aggressively investing in broadband – particularly in non-metro areas (Bright, 2001).11 For each model run, the logit specification included 2001 levels of variables that could potentially influence future broadband availability and adoption: population, income, and education; age group composition, race/ethnicity, and percent of the population residing in an urban area. 12 In some models, growth rates over the 1990s for specific variables (population, income, percentage with bachelor's degree) are also included to control for historical trends in those variables. Measures from 2001 (and earlier) are chosen in order to generate matches that are similar before broadband became an important part of American life. The logit models are restricted to non-metro counties, and the resulting probabilities of meeting the chosen broadband threshold are used to develop the propensity scores. The ATE technique then tests whether the growth 11 Note that additionally, the USDA's Rural Utilities Service pilot broadband loan program focused on providing infrastructure to rural areas did not begin until FY2001 (Kruger, 2012). 12 The final specification for each propensity score logit model will vary based on whether or not the covariates selected satisfy the balancing property developed by Becker and Ichino (2002). 1020 B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 rates for different economic measures (in this case, between 2001 and 2010) are statistically different for the treated and non-treated groups. In an ideal situation, we would have data available for the post-treatment period (i.e. after 2010) so that we could assess the impact of meeting the broadband threshold in 2010. However, our broadband availability data are limited (2010 is the only time we have specific measures such as the percentage with no access) and as such we are limited to assessing growth over the 2001–2010 time period, based on meeting a threshold in 2010. This is not ideal; however, the implication is that movement towards these thresholds began to occur after 2001 and before 2010. Our modeling approach, then, looks at the likelihood of obtaining a 2010 threshold based on 2001 demographics (including accounting for growth rates during the 1990s); and then suggests that economic growth between 2001 and 2010 may be influenced by the progress made towards those broadband thresholds over the same time period.13 Some non-metro counties achieved the threshold (and are placed in the treatment group); others with similar 2001 characteristics did not (and are placed in the control group). The ATET methodology allows for a comparison between the groups. 5. Results The results of the propensity score matching/ATET technique are displayed in Table 3. Only variables that demonstrated statistically significant differences between treated and non-treated groups of economic measures are displayed.14 More stringent broadband/availability thresholds were also tested to assess robustness; however these results generally lacked any statistical significance. All economic outcomes are changes over the 2001–2010 time period. The pseudo R2 of the logit model specification is also displayed. The results shown are those for the nearest neighbor matching method; however all significant outcomes were verified with the kernel matching technique. All outcomes reported in Table 3 demonstrated significant results under both methods. The results suggest that, generally, broadband adoption, availability, and download speed do have meaningful impacts on growth rates of economic health measures in non-metropolitan counties. In particular, non-metro counties that demonstrated high levels of broadband adoption (defined as county-level adoption rates 460%) had significantly higher levels of growth in median household income and significantly reduced growth in unemployment when compared with otherwise similar counties that did not meet this threshold. Alternatively, low levels of broadband adoption (o40%) imply detrimental impacts for rural businesses, with low-adoption counties having firm and employment growth rates approximately 3 percentage points lower when compared to their matched counterparts. The logit models underlying these two results had relatively high goodness of fit measures, with pseudo R2 values of around 0.25. Broadband availability thresholds also demonstrate some (potentially causal) relationships with economic growth, although the results are somewhat counterintuitive. Non-metro counties with high levels of broadband availability (485%) had growth rates of non-farm proprietors income that were over 5 percentage points lower than comparable counties with lower levels of availability. This suggests that high levels of availability are actually causing non-farm proprietor income to decline – perhaps implying that non-farm proprietors in these areas are not responding to customer expectations related to broadband. If residents have high levels of availability, they may expect most businesses they deal with (including the self-employed) to offer online payment options or have a viable web presence. Entrepreneurs who do not offer these elements might find their incomes declining. Broadband adoption is in fact higher in nonmetro counties with high levels of availability as defined in this study (categorical mean 3.2 versus 2.7); and recent evidence suggests that less productive workers may have been pushed into rural entrepreneurship during the recent recession (Goetz et al., 2012; Rupasingha & Goetz, 2011). This provides some limited evidence in support of this argument. Table 2 also demonstrates that non-metro counties with low levels of broadband availability (o50%) had growth rates of median household income that were marginally higher than otherwise similar counties. Although this result is not expected, recall that the treated and non-treated groups are matched based on their probabilities of reaching the broadband threshold – in this case, having very poor broadband availability. Counties with high likelihoods of having such poor levels of infrastructure likely have low population densities, and relatively low income and education levels. Changes to median household income over a 10-year period can be driven by any number of factors, including returns to those residents that do have access to (and productively use) broadband. Alternatively, this counterintuitive finding may indicate problems with the controls used for the matching procedure, or could simply be a false positive that would be reported nearly 10% of the time for this specification. In terms of broadband download speeds, attaining a high threshold (4 10 Mpbs) appears to be causally linked to higher increases in the percentage of employees classified as creative class workers. Although the difference between treated and control groups was less than one percentage point in this case, many rural communities are actively seeking to attract creative class workers, and these results indicate that having high download speeds available plays a role. Poverty levels were also roughly 2.6 percentage points lower in non-metro counties with high download speeds compared to otherwise similar counties, suggesting that broadband speed can potentially contribute to general community well-being. Finally, having only low levels of average download speeds (o3 Mbps) is associated with marginally higher growth rates in median household income. This counterintuitive result may suggest that many productive uses of broadband can still be accomplished at lower 13 We acknowledge that there may be some degree of endogeneity between the outcome variables (different measures of economic growth between 2001 and 2010) and the treatment variable (levels of broadband in 2010). This is another limitation against any strict claims of causality. 14 All variables displayed in Table 2 were tested for statistically significant differences using the ATET methodology for each broadband threshold. Any variable not displayed in Table 3 demonstrated no statistical differences between treated and control groups. B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 1021 Table 3 Propensity score matching results. Propensity scores with significant results: Adoption 1 High levels of BB adoption in 2010 ( 460%) NM Only (þ) ( ) 2 Low levels of BB Adoption in 2010( o 40%) NM Only ( ) ( ) Availability 3 High levels of 2010 BB availability ( 485%) NM Only ( ) 4 Low levels of BB availability ( o 50%) – NM only (þ) Treated Control Diff 0.254 %Δ MHHI (2001–2010) %Δ Unemp (2001–2010) 0.234 0.751 0.221 0.847 0.013 1.84 0.096 2.74 n nnn 0.244 %Δ Number Firms (2001–2010) 0.033 0.005 0.028 2.42 %Δ Total Employment 0.078 0.044 0.034 1.88 (2001–2010) nnn n 0.166 %Δ NFP Income (2001–2010) 0.087 0.034 0.054 2.33 nn %Δ MHHI (2001–2010) 0.265 0.248 0.017 n 0.051 0.201 0.043 0.227 0.008 1.83 0.026 2.41 n 0.245 0.235 0.010 n 0.222 Download Speed 5 High levels of 2010 Avg D/L Speeds ( 410 Mbps) – NM Only (þ) %Δ Creative Class (2001–2010) ( ) %Δ Poverty (2001–2010) 6 Low levels of 2010 Avg D/L Speeds (o 3 Mbps) – NM Only (þ) %Δ MHHI (2001–2010) Propensity Scores with no significant results: High Levels of 2010 BB Providers (Z 5) Low levels of 2010 BB Providers ( r 2) n Psuedo R2 T-stat 1.94 0.041 nnn 0.075 1.84 0.229 0.218 Statistically significant difference at the p ¼0.10 level. Statistically significant difference at the p ¼ 0.05 level. Statistically significant difference at the p¼ 0.01 level. nn nnn speeds. Note, however, that the fit of these models is significantly lower than previous efforts (pseudo R2 of only 0.04–0.07). Thus, relatively poor model specifications may be contributing to these findings. Finally, having a high or low number of broadband providers does not provide any significant results for the economic growth measures in question. It is important to note that because the only difference between the treated and control group is meeting the broadband criterion, use of the ATE technique allows for preliminary statements about causality (Dehejia and Wahba, 2002; Rubin, 2006). As previously noted, however, issues with both the NBM dataset and the potential influence of unobservables deter any claims of strict causality. In general, then, the results of the propensity score specifications suggest that broadband does contribute to the economic health of non-metro counties, and that meaningful relationships are found for broadband adoption, availability, and download speeds. In particular, high levels of broadband adoption are (arguably) causally linked to increases in median household income and lower unemployment growth; low levels of broadband adoption negatively impact non-metro counties in the form of lower numbers of firms and employment. The relationships with availability and download speeds are less intuitive – the impacts of achieving some “high” or “low” thresholds are not always in the hypothesized direction. From a policy standpoint, this suggests a need to focus on increasing adoption rates in order to spur economic growth, and that simply improving levels of infrastructure availability will not necessarily achieve that goal. 6. Conclusion and policy implications This research yields important findings on the effect of broadband on economic gains, namely on household income and employment levels in rural portions of the country. Although concerns about the NBM data itself (Ford, 2011; Grubesic, 2012) and the potential influence of unobservables prevent any strict claims of causality, the methodology used represents a step forward in the search for a causal relationship between broadband and economic development.15 The ability to do matched county comparisons, specifically in non-metro counties, demonstrates the influence of adoption (as opposed to availability) in producing these positive outcomes, and constitutes another indication that development efforts should focus on mobilizing populations to subscribe to and use broadband capabilities. Again, cultivating local leadership, mobilizing the services of cooperative extension educators nationwide, and working more closely with each State Broadband Initiative could be fruitful avenues for targeting adoption.16 Our matching results do produce some puzzling results, including a positive relationship between low levels of both availability/download speeds and median household income growth. 15 Strict claims of causality might be made if it could be demonstrated that the NBM data used was, in fact, reliable, and that there are no unobservable variable differences (that impact either broadband adoption/availability or economic outcomes) between treated and control groups. Additionally, potential endogeneity between the outcome and treatment variables could be addressed if post-treatment outcome data were available. 16 Under NTIA, the State Broadband Initiative launched in 2009 awarded funds to an entity in each state to undertake mapping, data gathering, and planning for broadband. 1022 B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 We can hypothesize that median income can grow on the basis of those that do have access to at least some level of broadband, but these results still warrant further investigation – including a deeper understanding of the pathways through which broadband might impact economic outcomes. A particularly interesting result is that achieving high levels of download speed in non-metropolitan counties does seem to have the desired result of attracting creative class workers and lowering poverty levels. This suggests that while promoting adoption should be the first and foremost goal, achieving higher levels of speed in rural areas is also a worthy policy premise. In the future, case studies on very high speed networks such as those in Chattanooga and Kansas City may also be warranted given that most (76%) economic development professionals in a recent survey felt that speeds of 100 Mbps or greater were needed to effectively attract new businesses (Settles, 2012). The recent growth of fiber networks in rural portions of the country (many funded by the American Recovery and Reinvestment Act) will also provide an opportunity for studies related to economic growth. Inasmuch as adopting (and using) broadband must be a focus of digital divide policy, future options must consider the means to encourage people to subscribe to broadband services once they are present. Hauge and Prieger (2010) provide an excellent overview of issues associated with encouraging broadband adoption, including the need for systematic evaluation of programs focused on demand. The FCC's attempt to experiment with the Lifeline programs through the Broadband Adoption Pilot Program, in which providers are expected to help address “other challenges” to broadband adoption such as the cost of devices and digital literacy, represents an interesting behavioral economics approach to this issue (FCC, 2013; National Telecommunications and Information Association (NTIA), 2013). As well, the endeavors of municipalities and other groups to provide broadband services, particularly when local privately-owned options are deemed insufficient, should be carefully examined and supported when community needs warrant this option. References Aron, D., & Burnstein, D. (2003). Broadband adoption in the United States: an empirical analysis. In: A. Shampine (Ed.), Down to the wire: Studies in the diffusion and regulation of telecommunication technologies. Bright, M. (2001). Federal strategies encouraging rural broadband access: Intelligent options to minimize the digital divide. Report submitted to the Institute of Electrical and Electronics Engineers, Inc. Retrieved from 〈http://www.wise-intern.org/journal/2001/matthewbright2001.pdf〉. Dehejia, R., & Wahba, S. (2002). Propensity score-matching methods for nonexperimental causal studies. Review of Economics and Statistics, 84(1), 151–161. Dickes, L., Lamie, D., & Whitacre, B. (2010). The struggle for broadband in rural America. Choices, 25(4). Retrieved from 〈http://www.choicesmagazine.org/ magazine/article.php?article=156〉. Becker, S. O., & Ichino, A. (2002). Estimation of average treatment effects based on propensity scores. The Stata Journal, 2, 358–377. Federal Communications Commission (FCC) (2012). Eighth broadband progress report. Retrieved from 〈http://www.fcc.gov/reports/eighth-broadband-progressreport〉. Federal Communications Commission (FCC) (2013). Lifeline broadband pilot program. Retrieved from /http://www.fcc.gov/encyclopedia/ low-income-broadband-pilot-programS. Flamm, K., & Chaudhuri, A. (2007). An analysis of the determinants of broadband access. Telecommunications Policy, 31, 312–326. Ford, G. (2011). Challenges in using the National Broadband Map's data. Phoenix Center Policy Bulletin, 27, 1–27. Galloway, L. (2007). Can broadband access rescue the rural economy? Journal of Small Business and Enterprise Development, 14(4), 641–653. Gillett, S., Lehr, W., Osorio, C., & Sirbu M. (2006). Measuring the economic impact of broadband deployment. Economic Development Administration, Department of Commerce. Retrieved from 〈http://cfp.mit.edu/publications/CFP_Papers/Measuring_bb_econ_impact-final.pdf〉. Goetz, S., Fleming, D., & Rupasingha, A. (2012). The economic impacts of self-employment. Journal of Agricultural and Applied Economics, 44(3), 315–321. Grubesic, T. (2012). The U.S. National Broadband Map: Data limitations and implications. Telecommunications Policy, 36, 113–126. Hauge, J., & Prieger, J. (2010). Demand-side programs to stimulate adoption of broadband: What works?. Review of Network Economics, 9(3). (Article 4). /http://www.degruyter.com/view/j/rne.2010.9.3/issue-files/rne.2010.9.issue-3.xmlS. Hitt, L., & Tambe, P. (2007). Broadband adoption and content consumption. Information Economics and Policy, 19(3), 362–378. Holt, L., & Jamison, M. (2009). Broadband and contributions to economic growth: Lessons from the US experience. Telecommunications Policy, 33(10), 575–581. Horrigan, J. (2007). Home broadband adoption 2007. Report for PEW Internet and American Life Project. Retrieved from 〈http://www.pewinternet.org/ /media//Files/Reports/2007/PIP_Broadband%202007.pdf〉. Horrigan, J. (2009). Home broadband adoption 2009. Report for PEW Internet and American Life Project. Retrieved from 〈http://pewinternet.org/Reports/ 2009/10-Home-Broadband-Adoption-2009.aspx〉. Imbens, G. (2004). Nonparametric estimation of average treatment effects under exogeneity: A review. Review of Economics and Statistics, 86, 4–29. Jayakar, K., & Park, E. (2013). Broadband availability and employment: An analysis of county-level data from the National Broadband Map. Journal of Information Policy, 3, 181–2000. Katz, R., & Suter, S. (2009). Estimating the economic impact of the broadband stimulus plan. Working paper. Retrieved from: 〈http://www.elinoam.com/ raulkatz/Dr_Raul_Katz_-_BB_Stimulus_Working_Paper.pdf〉. Kolko, J. (2012). Broadband and local growth. Journal of Urban Economics, 71(1), 100–113. Koutroumpis, P. (2009). The economic impact of broadband on growth: A simultaneous approach. Telecommunications Policy, 33, 471–485. Kuttner, H. (2012). Broadband for rural America: Economic impacts and economic opportunities. Report prepared for the Hudson Institute. Retrieved from: 〈http://www.hudson.org/files/publications/RuralTelecom-Kuttner–1012.pdf〉. Kruger, L. (2012). Broadband loan and grant programs in the USDA's Rural Utility Service. Congressional Research Service report RL33816. Retrieved from 〈http://www.nationalaglawcenter.org/assets/crs/RL33816.pdf〉. LaRose, R., Strover, S., Gregg, J., & Straubhaar, J. (2011). The impact of rural broadband development: Lessons from a natural field experiment. Government Information Quarterly, 28, 91–100. Lehr, W., Osorio, C., & Gillett, S. (2005). Measuring broadband's economic impact. Presented at the 33rd research conference on communication, information, and internet policy (TPRC), Arlington, VA. Retrieved from: 〈http://www.andrew.cmu.edu/user/sirbu/pubs/MeasuringBB_EconImpact.pdf〉. Malecki, E. J. (2003). Digital development in rural areas: Potentials and pitfalls. Journal of Rural Studies, 19(2), 201–214. McGranahan, D., Wojan, T., & Lambert, D. (2011). The rural growth trifecta: Outdoor amenities, creative class and entrepreneurial context. Journal of Economic Geography, 11, 529–557. National Telecommunications and Information Association (NTIA) (2013). NTIA Broadband adoption toolkit. Retrieved from: 〈http://www2.ntia.doc.gov/files/ toolkit_042913.pdf〉. Parker, E. B. (2000). Closing the digital divide in rural America. Telecommunications Policy, 24, 281–290. B. Whitacre et al. / Telecommunications Policy 38 (2014) 1011–1023 1023 Rosenbaum, P., & Rubin, D. (1982). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41–55. Rubin, D. (2006). Matched sampling for causal effects. Cambridge: Cambridge University Press. Rupasingha, A., & Goetz, S. (2011). Self-employment and local economic performance: Evidence from US counties. Papers in Regional Science, 92(1), 141–161. Settles, C. (2012). Moving the needle forward on broadband and economic development. Community Broadband Snapshot Report authored in partnership with the International Economic Development Council. Retrieved from 〈http://cjspeaks.com/msp/IEDC2012.pdf〉. Shadish, W. (2013). Propensity score analysis: Promise, reality and irrational exuberance. Journal of Experimental Criminology, 9, 129–144. Stenberg, P., Morehart, M., Vogel, S., Cromartie, J., Breneman, V., & Brown, D. (2009). Broadband Internet's value for rural America. Economic research report no. 70 (ERR-78). Retrieved from 〈http://www.ers.usda.gov/publications/err-economic-research-report/err78.aspx〉. Stephens, H., & Partridge, M. (2011). Do entrepreneurs enhance economic growth in lagging regions?. Growth and Change, 42, 431–465. Strover, S. (2001). Rural Internet connectivity. Telecommunications Policy, 25, 331–347. Strover, S. (Ed.). (2011). Scholars' Roundtable: The effects of expanding broadband to rural areas. Center for Rural Strategies, USA. Retrieved from 〈http://www. ruralstrategies.org/sites/all/files/Broadband_Investment.pdf〉. Whitacre, B. (2010). The diffusion of Internet technologies to rural communities: A portrait of supply and demand. American Behavioral Scientist, 53(9), 1283–1303. Whitacre, B. (2011). Estimating the economic impact of telemedicine in a rural community. Agricultural and Resource Economics Review, 40(2), 172–183. Whitacre, B, & Mills, B. (2007). Infrastructure and the rural-urban divide in high-speed residential Internet access. International Regional Science Review, 30 (3), 249–273. Wooldridge, J. (2002). Econometric analysis of cross section and panel data. Cambridge, MA: MIT Press.