Phoenix Center Policy Paper Number 33: The Broadband Efficiency Index: What Really Drives Broadband Adoption Across the OECD?

Transcription

1 PHOENIX CENTER POLICY PAPER SERIES Phoenx Center Polcy Paper Number 33: The Broadband Effcency Index: What Really Drves Broadband Adopton Across the OECD? George S. Ford, PhD Thomas M. Koutsky, Esq. Lawrence J. Spwak, Esq. May 2008), George S. Ford, Thomas M. Koutsky and Lawrence J. Spwak 2008).

2 Phoenx Center Polcy Paper No. 33 The Broadband Effcency Index: What Really Drves Broadband Adopton Across the OECD? George S. Ford, PhD Thomas M. Koutsky, Esq. Lawrence J. Spwak, Esq. Phoenx Center for Advanced Legal & Economc Publc Polcy Studes, George S. Ford, Thomas M. Koutsky and Lawrence J. Spwak 2008).) Abstract: In ths PAPER, we assess the performance and effcency of OECD countres wth respect to broadband Internet subscrpton. Usng the econometrc technque of Stochastc Fronter Analyss, we estmate scores ndcatng the effcency wth whch a country converts ts economc and demographc endowments nto broadband subscrptons. Wth very few exceptons, we fnd that broadband subscrpton n OECD countres s consstent wth those endowments about two thrds of OECD countres have an effcency rate of 95% or better. Sgnfcantly, the Unted States has an effcency ndex of 96.7%, whch s slghtly hgher than Japan 96.3%) and Korea 95.8%). Consstent wth earler research, we fnd that economc and demographc endowments explan nearly all of the varaton n broadband subscrptons 91%). Ths fndng suggests that publc polcy s role for broadband adopton may be more effectve f t s targeted at mprovng or mtgatng the adverse effects of those underlyng demographc and economc condtons, such as computer ownershp and educaton programs. Fnally, because countres have dfferent demographc and economc condtons, the most effectve mx of polces wll vary from country-to-country. As such, our fndngs ndcate that blndly followng the polces of countres ranked hgher n the OECD raw rankngs s not lkely to result n optmal success. Chef Economst, Phoenx Center for Advanced Legal & Economc Publc Polcy Studes. Resdent Scholar, Phoenx Center for Advanced Legal & Economc Publc Polcy Studes. Presdent, Phoenx Center for Advanced Legal & Economc Publc Polcy Studes. The vews expressed n ths paper are the authors alone and do not represent the vews of the Phoenx Center, ts Adjunct Fellows, or any of ts ndvdual Edtoral Advsory Board members. We are ndebted to Randy Beard, Adjunct Fellow, for hs assstance n formulatng the economc model presented n ths paper.

3 2 PHOENIX CENTER POLICY PAPER [Number 33 TABLE OF CONTENTS: I. Introducton... 2 II. Emprcal Framework... 5 III. Emprcal Detals... 9 A. Regressors and Expectatons... 9 B. Data Sources C. Estmaton Specfcs IV. Results A. Margnal Effects and Influence B. Measures of Performance C. Calculatng the Fronter D. Effcency Improvements V. Concluson I. Introducton The Internet has become an essental component of communcatons and a growng body of lterature has establshed the sgnfcance of ts role n economc growth, productvty and compettveness. A recent study by Gllett, Lehr & Srbu 2006) found that communtes where broadband was avalable experenced more rapd growth n employment, the number of busnesses overall, and busnesses n IT-ntensve sectors, relatve to comparable communtes wthout broadband. 1 As a result, many countres have ntated natonal polces for broadband Internet access and subscrptons and many more are consderng such polces. It s not surprsng that countres oftentmes feel that they are n the mddle of a Broadband Arms Race n whch the adopton and dffuson of broadband nfrastructure and technology s seen as a key to a country s economc future. 2 In large part, comparsons of broadband nfrastructure and adopton among more developed countres are based on the wdely-reported fgures regardng 1 S.E. Gllett, W.H. Lehr & M. Srbu, Measurng Broadband s Economc Impact, Fnal Report, p. 3 Feb. 28, 2006) avalable at: _2epdf/v1/mtcmubbmpactreport.pdf); see also G. Ford & T. Koutsky, Broadband and Economc Development: A Muncpal Case Study from Florda, 17 REVIEW OF URBAN & REGIONAL DEVELOPMENT STUDIES ). 2 Gllett et al., supra n. 1, at 3, specfcally note that n order for a communty to realze many economc gans from broadband, broadband had to be used, not just avalable.

4 Sprng 2008] BROADBAND EFFICIENCY INDEX 3 broadband adopton across member countres of the Organsaton for Economc Co-Operaton and Development OECD ). These raw broadband subscrptons per capta data and the rankngs thereof are reported by the OECD on a bannual bass. 3 To make the data more sensbly comparable across countres wth wdely dsparate populatons, the OECD normalzes or condtons) the data on populaton, expressng subscrpton counts n per-capta terms. However, as the sgnfcant dfferences across OECD countres are not lmted to populaton, ctng to raw OECD data wthout further analyss presents a msleadng pcture of broadband adopton and provdes a poor bass upon whch responsble publc polcy can be developed. 4 A more relevant comparson of broadband success takes nto account a wde range of economc and demographc endowments, lookng not to raw subscrptons as an effcency measure but rather at a falure to perform up to expectatons. Put smply, a country wth low GDP can be a more effcent adopter of broadband than a rch country even f ts raw, broadband subscrptons per capta rate s lower. In PHOENIX CENTER POLICY PAPER NO. 29, we frst proposed a tool to make such comparsons. 5 Usng broadband subscrptons and country-specfc data on ncome, ncome nequalty, educaton attanment, age, and so forth, we used regresson analyss to calculate a predcted broadband subscrpton rate and then compared ths to the actual subscrpton rate. Ths dfference, whch equals the dsturbance of the regresson, was then normalzed to become the Broadband Performance Index BPI ). 6 Countres that fell well short of expectatons were deemed poor performers, and those that were above expectatons were good performers. Most countres performed n lne wth expectatons, but there were a few obvous specal cases of over- and underperformance. In ths PAPER, we buld upon our pror work by usng a dfferent approach to assess and compare the adopton of broadband n the thrty OECD countres. 3 The latest OECD broadband data to December 2007) s avalable at The OECD updates ths data every sx months. 4 See, e.g., C. Holahan, The Sad State of U.S. Broadband, BUSINESS WEEK May 22, 2008). 5 G.S. Ford, T.M. Koutsky & L.J. Spwak, The Broadband Performance Index: A Polcy-Relevant Method of Comparng Broadband Adopton Among Countres, Phoenx Center Polcy Paper No. 29 June 2007)avalable at: For an analyss of the Unted States, see G.S. Ford, T.M. Koutsky & L.J. Spwak, The Demographc and Economc Drvers of Broadband Adopton n the Unted States, Phoenx Center Polcy Paper No. 31 Nov. 2007) avalable at: 6 See Secton II.A, nfra, for the calculaton.

5 4 PHOENIX CENTER POLICY PAPER [Number 33 Here, we calculate a Broadband Effcency Index BEI ), whch s derved drectly) usng the regresson technque of Stochastc Fronter Analyss SFA ). Wth ths alternatve, the techncal effcency wth whch endowments are converted nto the subscrpton can be separated from the typcal econometrc dsturbance. The BEI, whch s computed wthn the estmaton technque, measures how far a country s from the fronter of broadband subscrpton that s, the subscrpton rate observed under optmal effcency). The further a country s from the fronter, the lower ts effcency. For comparson purposes, we estmate and compute n ths paper both the BPI and the BEI usng the most-recent OECD data. Despte the theoretcal and practcal dfferences n computaton, the two measures of performance yeld very smlar results. Greece, Ireland, Slovaka, the Czech Republc, New Zealand and Luxembourg are all relatvely poor performers. In contrast, Iceland, Belgum, and Portugal are exceptonal performers, wth broadband adopton rates well above expectatons. The other OECD countres are all good performers, meanng that they are currently convertng ther demographc and economc endowments nto subscrptons at a very hgh rate of effcency. Sgnfcantly, the Unted States has an effcency ndex of 96.7%, whch s slghtly hgher than the purported broadband mracles of Japan and Korea 96.3%, 95.8%), 7 consstent wth the results of our earler paper. As a result, our new fndngs agan reveal that the Unted States s mproperly crtczed n for laggng behnd ts peers n broadband adopton. 8 In addton, our results suggest that the common legal and polcy framework for telecom n the European Unon EU ) has not led to common results. There s a wde varety n effcency among EU membershp, wth some EU countres exhbtng low effcency. 7 See T. Ebhara, Understandng the Japanese Broadband Mracle, Apr. 7, 2007) avalable at: C.P. Larsen, Experences from Korea, p. 8 Jan , 2006) avalable at: I. Tuom, EU Jont Research Center, The Korean Broadband Mracle 2004) avalable at: see generally T. Bleha, Down to the Wre, 84 FOREIGN AFFAIRS ) notng that Japan and South Korea wll lead the charge n hgh-speed broadband over the next several years ). 8 For example, see Letter from Consumer Federaton of Amerca, et al. to Charman Kevn J. Martn, Charman, Federal Communcatons Commsson U.S.) Nov. 13, 2007) avalable at: whch clams that [t]he U.S. s clearly tralng most of our major economc rvals n broadband speed transmsson speed, nvestment, subscrbershp and compettveness.

6 Sprng 2008] BROADBAND EFFICIENCY INDEX 5 We beleve that our fndngs are relevant to telecom polcymakers worldwde. It s mportant for polcymakers to understand the condtons that drve the rate of broadband adopton and the extent to whch telecom polcy may have a role. It s not uncommon for advocates to clam that a country s behnd a peer n broadband subscrptons per capta to support a desred polcy outcome, but our analyss demonstrates that such arguments are lkely msguded. Raw subscrpton rates are exceedngly poor ndcators of relatve effcency. Whle we beleve that telecom polcy can play an mportant role n the dffuson and adopton of broadband technology, ts nfluence and motvaton s complex. What we attempt to do n ths PAPER s provde a bass for understandng the factors that drve broadband adopton and provde some methods for polcymakers to examne targeted responses to the partcular condtons that may be holdng back the pace of broadband adopton n ther country. For example, we show below that ncome nequalty s a major factor that explans the rate of broadband adopton n the Unted States, so programs targeted at mtgatng that effect such as computer tranng and ownershp programs) may be an effectve means of drvng up the overall rate of adopton. Ths PAPER s organzed as follows. In Secton II, we provde the basc theoretcal underpnnngs of the estmaton approaches and calculatons of the effcency measures.e., BPI and BEI). Next, n Secton III, we summarze the detals of the estmaton. Results are summarzed n Secton IV, wth conclusons n Secton V. II. Emprcal Framework Our statstcal approach s straghtforward. Usng data on broadband subscrpton rates and demographcs across the OECD, we frst employ our technque set forth n POLICY PAPER NO. 29 and use a regresson analyss to quantfy the relatonshp between economc and demographc endowments and subscrpton. In partcular, we examne GDP per capta, ncome nequalty, educaton, populaton age, populaton densty, relatve sze of the country s largest cty, household sze, busness sze, telephone penetraton, and the prce of broadband servces. 9 We fnd that each one of these demographc and economc condtons s a statstcally sgnfcant determnant of broadband subscrpton. In 9 Throughout ths paper we called these demographc and economc condtons endowments. We regard them as such because telecom polcymakers generally have very lttle control over these condtons. As a result, broadband polcy at best can only affect the effcency wth whch a country s economy converts these endowments nto broadband subscrptons.

7 6 PHOENIX CENTER POLICY PAPER [Number 33 fact, taken together, our regresson shows that these factors explan 91% of the dfferences n the broadband subscrpton rate of the 30 OECD countres. Usng the results of that regresson, we then compare the actual and predcted subscrpton rates to see whether a country meets, exceeds, or falls below what would be reasonably expected gven ts demographc and economc endowments. In ths PAPER we use two methods of makng ths comparson. In our frst effort at developng a polcy-relevant means of comparng broadband adopton across countres, we computed a Broadband Performance Index BPI ), derved from the dfference n actual and predcted subscrpton, for each OECD country as a smple ndex wth whch to compare how that country performs relatve to expectatons. 10 In contrast, the BPI calculaton was relatve. Stated smply, the BPI benchmarked the broadband penetraton of one country relatve to the poorest performer n the OECD and generated an ndex from that comparson. In ths PAPER, we employ Stochastc Fronter Analyss to generate a measure of performance drectly. Unlke the least squares method of our earler PAPER, the fronter analyss computes a measure of effcency as part of the estmaton process tself that s separate and ndependent from statstcal nose. We call ths measure of performance the Broadband Effcency Index BEI ). Stochastc Fronter Analyss SFA ) s deally suted to comparng the performance of OECD countres wth regard to broadband subscrpton rate. In essence, SFA s a lnear regresson technque estmated by maxmum lkelhood) wth a dsturbance that has two components a standard two-sded dsturbance and an addtonal strctly non-negatve dsturbance. 11 In the present context, ths latter part of the dsturbance s a drectly estmated measure of performance n that t captures neffcency n the converson of endowments nto broadband subscrpton. Wth SFA, statstcal nose and effcency are separated, so that an arguably cleaner measure of performance s rendered at least relatve to the BPI). 12 Whle the fronter approach seems suted to ths challenge, t s not wthout ts problems. Least squares estmaton s exceedngly robust, but fronter models 10 Supra n For dscussons of SFA, see, e.g., S. C. Kumbhakar & C. A. K. Lovell, STOCHASTIC FRONTIER ANALYSIS 2000); W. Greene, ECONOMETRIC ANALYSIS 2000), ; T. J. Coell, D. S. Rao, C. O Donnell & G. Battese, AN INTRODUCTION TO EFFICIENCY AND PRODUCTIVITY ANALYSIS 2005). 12 The addtonal dsturbance term must be estmated, however, and n some cases no neffcency component can be extracted from the data.

8 Sprng 2008] BROADBAND EFFICIENCY INDEX 7 are estmated by maxmum lkelhood and convergence can be fckle to model specfcaton. In some cases, the effcency measure cannot be adequately estmated and the fronter approach devolves nto least squares. In other cases, convergence s not acheved. However, usng the general emprcal format of POLICY PAPER NO. 29, we are able to get sensble results for both least squares and fronter specfcatons. In ths PAPER we calculate both the BPI and the BEI, so we can compare the results of the two approaches. Dong so provdes smlar, but not exact, predctons regardng effcency. The strongest evdence of good and bad performance s generally consstent across both procedures. Consequently, f the fronter analyss fals to converge or poorly estmates the effcency measures, the least squares approach of the BPI remans a vable opton. A formal descrpton of our econometrc approach s as follows. Let the relatonshp between per capta) broadband subscrpton and a country s endowments be fx,β), so that B f X, β). 1) = Gven the standard dsturbance term of least squares regresson, we have B = f X, β)exp v ). 2) If there are k endowments, then Equaton 3) can be expressed n econometrc form as: k ln B = β 0 + β j ln X j + v, 3) j= 1 whch s the model estmated by POLICY PAPER NO. 29. Usng the estmates from Equaton 3), n POLICY PAPER NO. 29 we then compute the Broadband Performance Index BPI ) as BPI = vˆ /max vˆ ), 4) where the BPI s scaled so that t les between -1 and 1, wth larger more postve) values ndcatng better performance than endowments would suggest and negatve values poor performance). A value near zero ndcates performance s consstent wth the country s demographc and economc endowments.

9 8 PHOENIX CENTER POLICY PAPER [Number 33 A problem wth ths approach s that the dsturbance v measures not only varatons n the effcency of endowment converson but also statstcal nose. Ideally, ths nose could be separated from the measure of effcency, thereby provdng more useful polcy gudance. Further, effcency should be measured from an deal performance, so the effcency measure needs to be one-sded. To ncorporate effcency nto the estmaton process tself, let θ capture the varyng degrees of effcency n the converson of endowments to subscrpton. Now, Equaton 2) becomes B = f X, β) θ exp v ) 5) where θ s the level of effcency for country [where 0 < θ 1]. When θ = 1, the country s optmally effcent n ts converson of endowments nto subscrpton; a value of θ < 1 ndcates some degree of relatve neffcency. Of course, effcency can only be establshed relatve to the countres n the sample.) Effcency cannot exceed the deal, so θ 1. The econometrc specfcaton of the regresson model becomes: k ln B = β 0 + β ln X + v u. 6) j= 1 j j where u = -lnθ) f we restrct u > 0 then we have 0 < θ 1. Equaton 6) s the standard specfcaton for Stochastc Fronter Analyss SFA ) where the term u s the estmated neffcency term, whch s strctly postve one-sded). If there s no effcency varaton ether t s zero or cannot be estmated), then Equaton 6) s dentcal to Equaton 2) and we have least squares estmaton. Practcally, the term u measures the percentage by whch a country fals to acheve the deal converson rate of endowments nto broadband subscrptons. The error term v, alternately, s a standard two-sded dsturbance and captures the statstcal nose. In estmaton, the dstrbuton of u must be specfed and avalable dstrbutons are the exponental, truncated normal, half normal, among others. We assume the dstrbuton of u s exponental. 13 From Equaton 6), the Broadband Effcency Index BEI ) s computed usng 13 The rankng of effcency s typcally robust to dstrbutonal assumptons, though the effcency measure tself may vary. Coell et al., supra n. 11, at 252.

10 Sprng 2008] BROADBAND EFFICIENCY INDEX 9 BEI = exp u ), 7) where the BEI s smply the techncal effcency parameter θ from SFA. The BEI has a maxmum value of 1.0 and a mnmum value of 0.0. Values closer to 1.0 mply greater effcency, wth a value of 1.0 ndcatng optmal effcency. III. Emprcal Detals For estmatng Equatons 3) and 6), we use the broadband subscrpton data publshed bannually by the OECD. The last three perods of avalable data are used to render a suffcent sample sze of recent data. Gven the short perod covered 18 months), we do not treat the data as a panel. 14 The explanatory varables are dentcal to those used n POLICY PAPER NO. 29 wth two exceptons both the subscrpton and prce data are updated. 15 Further detals are as follows. A. Regressors and Expectatons The vector X contans 14 varables ncludng a constant term. PRICE s an ndex of broadband prce n country, GDP s gross domestc product per household n country, GINI s the naton s Gn Coeffcent a measure of ncome nequalty) n country, EDUC s the percent of persons wth postsecondary or tertary educaton n country, AGE65 s the percent of the labor force age sxty-fve or older as a percentage of the labor force n country, DENSITY s the number of households per square klometer n country, BIGCITY s the percent of the populaton lvng n the country s largest cty n country, PHONE s the number of telephones landlne and moble) per 100 persons n country and PHONE 2 s ts square to allow for non-lnearty, 16 HHSIZE measures persons per household n country, BUSSIZE measures persons per busness establshment n country, and JUNE07 and DEC06 are dummy varables that equals 1 for the relevant perod of the data 0 otherwse). All contnuous varables are n natural log form. 14 Panel estmaton was attempted, but we were unable to estmate non-zero techncal effcency parameters. 15 We also corrected a codng error for Australa s household sze. 16 Specfcaton tests ndcated a non-lnear relatonshp wth respect to PHONE, so we nclude the square of the regressor.

11 10 PHOENIX CENTER POLICY PAPER [Number 33 Based on earler research, we have the followng expectatons regardng the regressors. The followng varables are expected to have negatve sgns: PRICE though we cannot clam that Equaton 1 s a demand curve), AGE65, GINI, BUSSIZE and PHONE Postve sgns are expected on these varables: GDPCAP, EDUC, DENSITY and PHONE. We do not have an a pror expectaton for the BIGCITY snce DENSITY s held constant. Furthermore, a large urban populaton s not smply a measure of populaton densty but may reflect other factors. We make no a pror predctons on the sgn of HHSIZE due to conflctng effects. HHSIZE mght be negatve snce t corrects for the per-capta nature of the dependent varable only one connecton s needed per household), but larger households may have larger demands for broadband servces. The tme dummy varables are expected to have negatve sgns, snce broadband subscrpton grows over tme. B. Data Sources The bulk of the data s provded by the OECD FACTBOOK 2006 and the World Bank s WORLD DEVELOPMENT INDICATORS Subscrpton rate data, rankng, and populaton are provded by the OECD. We estmate the model usng the latest three perods of data for a total of 90 observatons December-06, June-07 and December-07). Most of the regressors are at least three-year lags wth the excepton of PRICE), due to data avalablty. 19 Usng lagged values has some advantages, snce t commonly asserted that broadband mpacts economc development and other economc and demographc factors. Thus, the lagged data helps attenuate the potental for smultanety bas. 20 We use the last year of data avalable for all perods. All values of the regressors are constant over the sample. 17 The varable BUSSIZE s expected to be negatvely sgned snce the subscrpton data s n per capta terms. In other words, larger values of BUSSIZE ndcate fewer busnesses thereby ndcatng fewer busness subscrptons on a per capta bass. 18 Organsaton of Economc Co-operaton and Development, OECD FACTBOOK ECONOMIC, ENVIRONMENTAL AND SOCIAL STATISTICS 2006 avalable at: World Bank, WORLD DEVELOPMENT INDICATORS Wth few exceptons, the varables used n the regresson change very slowly, f at all, over tme. 20 D. Gujarat, BASIC ECONOMETRICS 1995), 654. Ths choce of lag was also motvated by the avalable data.

12 Sprng 2008] BROADBAND EFFICIENCY INDEX 11 Prce data s provded by an OECD report provdng detaled prce data for broadband servces. 21 Of all the varables, prce s the most dffcult to measure snce there are many prces pad for broadband servces n a populaton. Further, most of the subscrptons were not ntated under the current prce, so smply usng current average advertsed rates may lead to ncorrect results. Any sngle measure of prce cannot be exactly ndcatve of prces pad. Nevertheless, we expect prce to be an mportant determnant of subscrpton, so we nclude the varable as a regressor. Whle the prce varable has the expected sgn and s statstcally sgnfcant n the regressons, we nevertheless caveat our fndngs by observng that a sngle ndex of prce for broadband servce suffers from numerous shortcomngs. Others have done the same. 22 Data on GDPCAP, EDUC, AGE65, DENSITY, TAXES, and PHONE are all provded by the OECD FACTBOOK WORLD DEVELOPMENT INDICATORS also provded data for BIGCITY. Mssng observatons on some varables were flled usng other data sources. Notably, the BUSSIZE varable s computed usng populaton and busness establshment data, the latter of whch was unavalable for four countres. 23 We estmate usng least squares regresson the number of busness establshments for the four countres from the data avalable. 24 C. Estmaton Specfcs Snce the subscrpton rate s akn to a penetraton rate, we estmate Equaton 3) by weghted least squares WLS ) to account for the non-constant varance of the dependent varable. 25 In the natural log form, the varance of B s 1 - B )/N B where N s populaton, so the least squares and fronter regressons are weghted wth N B /1 - B ) 0.5. We use the same weght for the fronter estmaton, and the neffcency component of the error s assumed to follow the 21 Organsaton of Economc Co-operaton and Development, OECD COMMUNICATIONS OUTLOOK 2007 avalable at: Tbl S. Wallsten, Broadband and Unbundlng Regulatons n OECD Countres, AEI-Brookngs Jont Center Workng Paper No June 2006) avalable at: 23 These countres nclude Australa, Canada, Greece and Japan. 24 See Greene, supra n. 11, at ; R. Pndyck and D. Rubnfeld, ECONOMETRIC MODELS & ECONOMIC FORECASTS 1991), The same approach was adopted n Ford et al. June 2007), supra n G. S. Maddala, LIMITED DEPENDENT AND QUALITATIVE VARIABLES IN ECONOMETRICS 1983), 29. Ths specfcaton s the mnmum ch-square method for the lnear and log-lnear model.

13 12 PHOENIX CENTER POLICY PAPER [Number 33 exponental dstrbuton. Convergence of the SFA was fckle, but was acheved n ths specfc format. IV. Results The econometrc results and descrptve statstcs are summarzed n Table 1. Model 1 s estmated by WLS and Model 2 by SFA. The coeffcent estmates across technques are smlar. Table 2 provdes the BPI and BEI estmates for each country, and the countres are sorted by relatve effcency scores. Each country s rankng for the latest avalable broadband raw subscrpton data s also provded n the table. Both models exhbt good statstcal sgnfcance, wth all regressors sgnfcant at the 5% level or better. 26 All sgns are as expected. The unweghted) R 2 of Model 1 s 0.91, so 91% of the varaton n sample broadband subscrpton rates s explaned by the model. Consequently, non-polcy varables explan nearly all varatons n subscrpton rates. For the least squares model, we cannot reject the null hypothess of RESET no specfcaton error ) at even the 10% level, but the null hypothess of Whte s test homoscedastc dsturbances ) s rejected. 27 So, robust standard errors are used. For Model 2, the rato of the standard devatons of the neffcency and nose components s 0.79 λ 0.79), so the neffcency components s nearly as varable as the statstcal nose. A. Margnal Effects and Influence Margnal effects for Model 1 are nterpreted as those from an average effects model. The coeffcents from Models 2 and 4, alternately, descrbe the fronter. Note that n Model 2, the estmated coeffcents are unbased and consstent except for the constant term. Typcally, fronter analyss focuses on effcency estmates, rather than the coeffcents, but we also summarze the margnal effects from ths model. 26 Note that robust standard errors are used to compute the t-statstcs for Models 1 and Gven the large number of regressors, Whte s test for heteroscedastcty s based on regressng the squared resduals on the ftted and square of the ftted value from the regresson. As detaled by Wooldrdge, ths test s useful n that the test statstc χ 2 ) has only two degrees of freedom yet remans asymptotcally vald. It s a specal case of Whte s test for heteroscedastcty. J. Wooldrdge, ECONOMETRIC ANALYSIS OF CROSS SECTION AND PANEL DATA 2002), ,

14 Sprng 2008] BROADBAND EFFICIENCY INDEX 13 For Model 1, the margnal effects are as follows. Other than the tme dummes, the largest effect s PHONE, wth an elastcty of 2.0 computed at the sample mean). Second s GINI wth an elastcty of -1.2, and then GDPCAP 0.58), AGE ), PRICE -0.39), HHSIZE 0.35), BUSSIZE -0.23), EDUC 0.20), BIGCITY -0.20) and DENSITY 0.03). The partal R 2 values for the regressors are AGE ), GINI 0.54), PRICE 0.42), GDPCAP 0.38), BUSSIZE 0.33), BIGCITY 0.31), PHONE 0.27), PHONE ), EDUC 0.20), HHSIZE 0.12) and DENSITY 0.09). For both measures of nfluence, t appears that ncome, prce, populaton age, and hstorcal telephone demand are the key determnants of broadband subscrpton. The nfluence of BUSSIZE ndcates that rankng countres on the bass of per-capta broadband connectons as the OECD does) may be an nherently flawed method of normalzng data. 28 The margnal effects from Model 2 are very smlar to those from Model 1 as expected). For Model 2, the largest effect s agan PHONE, wth an elastcty of 2.84 a very large effect. Second s GINI wth an elastcty of -1.14, and then GDPCAP 0.55), AGE ), PRICE -0.31), HHSIZE 0.24), BUSSIZE -0.19), EDUC 0.16), BIGCITY -0.15) and DENSITY 0.03). The partal R 2 values for the regressors are AGE ), GINI 0.54), PRICE 0.42), GDPCAP 0.38), BUSSIZE 0.33), BIGCITY 0.31), PHONE 0.27), PHONE ), EDUC 0.20), HHSIZE 0.12) and DENSITY 0.09). Agan, ncome, prce, populaton age, and hstorcal telephone demand are the key determnants of broadband subscrpton. Table 3 utlzes the results of our regressons to analyze the margnal effects that each regressor has upon broadband penetraton n each of the thrty OECD countres. We use the estmates from Model 1. The values n the table are constructed by frst computng the contrbuton of each regressor to the departure of a country s subscrpton rate from the OECD mean subscrpton rate. 29 We then express ths contrbuton as a percentage of all the contrbutons of the regressors. 30 Table 3 demonstrates the magntude, n percentage terms for comparablty, whch a partcular factor plays n explanng each partcular country s expected 28 On ths pont, see also S. Wallsten, Understandng Internatonal Broadband Comparsons, Technology Polcy Insttute May 2008). 29 If β s the coeffcent, X the country s regressor value, and X the mean, the contrbuton s βx - X). 30 Each regressor s contrbuton s βx - X), so the values n the table for Country and regressor j are β j X j - X j )/Σ β j X j - X j ).

15 14 PHOENIX CENTER POLICY PAPER [Number 33 rate of broadband adopton n reference to the OECD mean. For example, the PRICE varable has a negatve sgn, so a country wth a relatvely hgh prce wll have a lower expectaton of broadband subscrpton. The hgh prce n the Czech Republc, whch s 16% above the sample average, explans 17.1% of the country s negatve departure from the sample mean. Conversely, the Czech Republc s smaller than average busness sze BUSSIZE) drves up ts expected broadband subscrpton level by 20%. Table 3 helps explan why there can be dspute between commentators over the mportance of certan factors because not every factor plays a role n every country. 31 For example, consder populaton densty DENSITY), whch our analyss shows to be sgnfcant n many countres but whch some commentators dspute. 32 Table 3 shows that DENSITY plays a major role n explanng the broadband adopton rate n countres lke Australa -20.2%), Belgum +10.4%), Canada -21.7%), Fnland -10.1%), Japan +11.8%), the Netherlands +10.1%), and New Zealand -11.6%), whch all have populaton denstes substantally departng from the OECD average. Populaton densty also plays a smaller but not nsgnfcant role n the Unted States -3.3%) but other factors notably ncome nequalty, GINI) play a more mportant role. In other words, populaton densty does matter but to a dfferent degree n the thrty OECD countres. Ths observaton accentuates our fndng that approprate broadband polces need to be nuanced and talored for a country s partcular demographc and economc condton. In nterpretng Table 3, one should not regard large percentage contrbutons to mply that there are large departures from the mean of the regressors, however. If a country s average n all respects except for, say, DENSITY, then 100% of ts departure from the sample mean subscrpton rate wll be explaned by DENSITY, even f the departure s small. For ths reason, we have ncluded the predcted and sample average values of lnb and the percentage dfference of the two. An examnaton of the data suggests that n most cases, however, large 31 The prmary defect n the densty does not matter argument s that those makng such clams operate n a unvarate settng. For example, densty may strongly related to subscrpton, but f hgh densty countres have relatvely hgher prces or low ncomes or dfferences n a varety of other factors), the varaton across countres n these other factors can mask the role of densty n a unvarate framework. Unvarate analyss s rarely useful, and the true nfluence of densty can only be determned n a multvarate settng lke that used here. 32 See Ebhara, supra n. 7, at 5; see Danel K. Correa, Assessng Broadband n Amerca: OECD and ITIF Rankngs Apr. 2007), 5-6 avalable at:

16 Sprng 2008] BROADBAND EFFICIENCY INDEX 15 values n Table 3 are assocated wth large departures from the sample mean of the regressors and ths s true for DENSITY). But, large departures from the mean are not always assocated wth large values n the table, snce n some cases there are more nfluental varables. In fact, Korea s densty s further from the mean than Japan s, but the nfluence of AGE65 n Korea overwhelms the effect of DENSITY. B. Measures of Performance Our measures of performance, the BPI and BEI, for the four models are provded n Table 2. An llustraton of the BEI s provded n Fgure 1. The effcency measures are averaged over the three perods of data for presentaton, and the countres are sorted by relatve performance. For about two-thrds of the OECD countres, the BEI s 0.95 or better, ndcatng that most countres are performng very well n terms of convertng endowments nto broadband subscrpton. As already mentoned, the endowments explan almost all of the varaton n subscrpton 91%), so there s lttle room for neffcency explanatons. Surprses among the top performers are Portugal, Belgum and Turkey. Portugal, arguably a poor performer wth a raw subscrpton rank of 23 rd, s actually the 3 rd best performer wth a BEI of Whle Portugal s overall subscrpton rate s low, the country s expected subscrpton s reduced by ts unfavorable endowments for GDPCAP, GINI, and EDUC, as shown n Table 3. Belgum s ranked 12 th n the raw OECD data, but ts effcency score s Turkey lkewse s ndcated as havng respectable effcency BEI 0.963) despte beng ranked 29 th n the raw subscrpton data. These surprses are mportant because they can provde polcymakers a more nuanced vew of broadband adopton trends. If you were smply to look at the OECD rankngs, then you would not lkely gve Portugal, Belgum and Turkey a second thought. Our tools nstead ndcate that they are models worthy of study and even possbly emulaton. The opposte s true havng a hgh rankng by the OECD does not mean that a country s dong a partcularly good job. For example, Luxembourg s a very weak performer on our effcency ndex BEI 0.769) gven ts subscrpton rank 9 th ). The poorest performers are unquestonably Greece BEI 0.619), the Slovak Republc BEI 0.651), Ireland BEI 0.696), the Czech Republc BEI 0.755), and Luxembourg BEI 0.769). Ths same set of countres s also labeled poor performers by the BPI. It seems these countres have the most to be concerned about regardng broadband subscrpton.

17 16 PHOENIX CENTER POLICY PAPER [Number 33 It s also nterestng to observe that Japan BEI 0.958) and South Korea BEI 0.963), sometmes touted as broadband polcy mracles, 33 are mddle-ofthe-pack performers n the OECD, accordng to the BPI. The U.S. s slghtly more effcent than both wth a pont estmate) BEI of whch s close enough to conclude they are statstcally equal). In addton, the BEI reveals a sgnfcant dfference between EU member states, wth some rankng near the top and others near the bottom. Ths ndcates that whle there s a common legal framework for telecom polcy n the EU, to date there s nothng close to common results. C. Calculatng the Fronter In Table 4, we provde the December 2007 subscrpton rates for the OECD countres n two forms. In the frst column, the actual subscrpton rate as publshed by the OECD s provded, and the second column ranks these rates. These are the data commonly cted n dscussons of relatve performance across OECD countres wth respect to broadband subscrpton. In the thrd column of the table, the fronter subscrpton rate s lsted. The fronter s computed by settng the measure of techncal effcency the BEI) for all countres equal to 1.0. Remanng dfferences between the actual and fronter subscrpton rates are accounted for now only by the statstcal nose normally ncluded n the random dsturbance term of a regresson. Note that f the twosded error term v from Equaton 6) s large enough, the fronter subscrpton rate may be below the actual subscrpton rate. Table 4 demonstrates that the broadband rankngs released by the OECD and used contnually by advocates, n fact, have very lttle to say about the effcency of broadband adopton. Table 4 shows that wth a few exceptons, the rankngs of actual and fronter subscrpton rates are smlar. Ths means that even f all OECD countres were perfectly and equally effcent n convertng endowments to broadband penetraton, that acton often has very lttle effect on the fnal raw rankngs that the OECD reports. Demographc and economc condtons so pervasvely drve the broadband subscrpton per capta number that utlzng the rankng of OECD countres, condtoned only on populaton, to advocate for or aganst broadband polcy changes s nonsenscal. 33 See supra n. 7.

18 Sprng 2008] BROADBAND EFFICIENCY INDEX 17 Our method of analyss strongly ndcates that rankng countres by broadband subscrptons per capta makes lttle sense. One can use our tools to determne where any partcular country would rank f t was optmally effcent n convertng ts endowments nto broadband subscrptons BEI = 1.0) and then derve ts rank. Ths what f approach n essence provdes a celng for each country s broadband rankng among the OECD. For nstance, f Luxembourg became optmally effcent, then t would lead the OECD n broadband. On the other hand, f the Unted States were optmally effcent, then ts December 2007 rank would be 14 th only one place hgher than ts actual rank of 15 th. Ths analyss demonstrates that usng the per capta rankngs as a bass for makng polcy comparsons and decsons s a fool s errand. D. Effcency Improvements For the analyss of performance just presented, we compute the BPI and BEI as an average over the three tme perods of data n the sample. By comparng the SFA effcency ndex from last and frst perod of data, we can determne whch countres are most sgnfcantly ncreasng ther effcency over tme. We suspect that the least effcent countres wll mprove the most, and ths result would suggest that the OECD countres are convergng to a more effcent outcome. Table 5 presents the rato of the BEI from December-07 to the BEI from December-06 BEI D07 /BEI D06 ) from Model 2. All the BEIs are summarzed n the table. The larger the rato, the larger the ncrease n the effcency over the twelve month perod. In Table 5 we see that Greece and to a lesser extent the Slovak Republc, both poor performers, are sgnfcantly mprovng n effcency over tme. In the eghteen month perod, the effcency of Greece countres rose by about 67%. Ireland s mprovng as well, wth about a 22% mprovement n effcency. We observe modest mprovements n effcency for other poor performers ncludng the Czech Republc 1.105) and Luxembourg 1.117). Most of the better performers have lttle room for relatve mprovement, and ths s confrmed n Table 5. V. Concluson The dffuson of broadband technology s perhaps the most sgnfcant telecom polcy challenge of the last thrty years. Polcymakers need to have useful tools that help them determne whether ther polces are havng an mpact on broadband subscrpton. Unfortunately, whle countres are routnely ranked by organzatons lke the OECD on ther broadband subscrpton rates, those postngs and surroundng rhetorc have very lttle analytcal foundaton for showng that there are polcy-relevant dfferences between countres that explan

19 18 PHOENIX CENTER POLICY PAPER [Number 33 those rankngs. Our analyss suggests that broadband adopton s ntmately ted to demand-sde factors lke ncome nequalty and educaton, and polces drected at those factors may be more cost effectve than supply-sde subsdes and regulaton. Whle not partcularly useful rhetorcally, the fact s that demography, geography, and economc condtons affect the rate of broadband adopton and those condtons cannot necessarly be affected drectly or ndrectly by communcatons polcy. We do not mean to suggest that polcymakers should be content wth the current level of performance, or that broadband polcy s rrelevant. Indeed, our results should encourage polcymakers to focus ther attenton on polces that wll cultvate or enhance the endowments that ncrease broadband adopton or that wll counterbalance the adverse effects of endowments that suppress broadband adopton. For example, programs focused on overcomng the effect of ncome and ncome nequalty mght sgnfcantly spur broadband adopton. ConnectKentucky s No Chld Left Offlne program s an example of such a program. 34 Broadband polcy s a serous ssue and polcymakers deserve serous tools of analyss. We fnd heren that much of the rhetorc regardng broadband rankngs and broadband mracles s suspect at best. In fact, the better performers n a rhetorcal sense often fal to lve up to expectatons. In partcular, both Japan and Korea rank below the Unted States, Canada and France n broadband effcency, whch suggests that ther relatvely hgh rates of broadband adopton have less to do wth dfferent telecom polcy approaches and more to do wth demographc and economc condtons such as populaton densty and age. Even though Japan and Korea have fber optc networks servng large portons of the populaton, these advanced networks appear not to have nfluenced substantally broadband adopton. Our analyss shows that broadband adopton n Iceland, Portugal and Belgum are substantally more effcent n convertng ther demographc and economc condtons nto broadband subscrptons than Japan, Korea, Germany and the Unted States. 34 B. Marshall, No Chld Left Offlne, RICHMOND REGISTER Sep. 25, 2007) avalable at: Obvously, the subscrpton to broadband servce s dependent on computer ownershp, and many poorer households cannot afford a computer. See, e.g., The Kentucky program ams to provde low cost or free computers to low ncome households.

20 Sprng 2008] BROADBAND EFFICIENCY INDEX 19 As always, ths analyss should be consdered one part of the portfolo of evdence needed to drve publc polcy. Further research s warranted and encouraged.

21 20 PHOENIX CENTER POLICY PAPER [Number 33 Table 1. Summary of Econometrc Results Model 1 WLS Model 2 SFA Coef. t-stat) Coef. t-stat) Mean [St. Dev.] Constant )* -6.01)* lnprice )* )* [0.29] lngdpcap )* )* [0.42] lngini )* )* [0.18] lneduc )* )* [0.43] lnage )* )* [0.34] lndensity )* )* [1.41] lnbigcity )* )* [0.50] lnhhsize )* )* [0.22] lnbussize )* )* [0.56] lnphone )* )* [0.29] lnphone )* )* [2.68] DEC )* )* [0.47] JUNE )* )* [0.47] Unw. R RESET F 2.00 Whte χ * Obs lnσ 2 v) -5.81* lnσ 2 u) -6.28* λ 0.79 * Statstcally sgnfcant at the 5% level.

22 Sprng 2008] BROADBAND EFFICIENCY INDEX 21 Table 2. Broadband Performance Index and Broadband Effcency Index Rank of Raw Subscrpton Rate for Dec-07 Data n Parenthess) Model 1 BPI Model 2 BEI Iceland 3) Iceland 3) Belgum 12) Belgum 12) Portugal 24) Portugal 24) Swtzerland 5) Swtzerland 5) Turkey 29) Denmark 1) Denmark 1) Fnland 6) Fnland 6) France 13) Norway 4) Norway 4) France 13) UK 11) Hungary 25) Canada 10) UK 11) Netherlands 2) Netherlands 2) Sweden 8) Sweden 8) US 15) Canada 10) Turkey 29) US 15) S. Korea 7) S. Korea 7) Hungary 25) Span 21) Japan 17) Japan 17) Italy 22) Italy 22) Span 21) Poland 27) Poland 27) Austra 18) Australa 16) Mexco 30) Mexco 30) Australa 16) Germany 14) Germany 14) Austra 18) New Zealand 19) New Zealand 19) Czech Rep. 23) Luxembourg 9) Luxembourg 9) Czech Rep. 23) Ireland 20) Ireland 20) Slovak Rep. 28) Slovak Rep. 28) Greece 26) Greece 26) 0.619

23 222 PHOENIX CEN NTER POLICY Y PAPER [N Number 33 Fguree 1. Broadban nd Effcency y Index 2006//7) Rank of Raw w Subscrpto on Rate for Deec-07 Data n Parenthess) Iceland d 3) Belgum 12) Portugal 24) Swtzerland d 5) Denmarkk 1) Fnland d 6) France 13) Norwayy 4) UK 11) Canada 10) Netherlandss 2) Sweden n 8) US 15) Turkey 29) S. Korea 7) Hungary 25) Japan 17) Italy 22) Span 21) Poland 27) Australa 16) Mexco 30) Germany 14) Austra 18) New Zealand 19) Luxembourgg 9) Czech Rep. 23) Ireland 20) Slovak Rep. 28) Greece 26) Phoenx Center C for Advancced Legal and Econ nomc Publc Polccy Studes ww ww.phoenx-center..org

24 Sprng 2008] BROADBAND EFFICIENCY INDEX 23 Table 3. Explanng Departures from the Mean Broadband Subscrpton Rate Contnued on next page) Country Predcted Dff. lnb from Mean PRICE GDPCAP GINI EDUC AGE65 Sgn > Australa % -3.9% 18.7% 0.2% 8.4% 8.8% Austra % -2.1% 20.8% 26.3% -9.4% -14.6% Belgum % 1.4% 17.8% -10.7% 6.9% -33.0% Canada % -3.2% 20.6% 2.2% 18.2% 17.1% Czech Republc % -17.1% -19.2% 14.2% -9.6% -2.4% Denmark % 11.8% 16.7% 33.6% 7.4% -5.3% Fnland % 18.3% 17.2% 22.0% 10.2% -0.8% France % 13.5% 16.8% 18.9% 1.4% -32.4% Germany % 13.6% 8.8% 11.0% 1.8% -16.1% Greece % 5.7% -13.2% -14.7% -4.3% -37.3% Hungary % -6.6% -36.6% 4.8% -7.9% -18.5% Iceland % -4.0% 10.9% -21.0% 2.0% 15.7% Ireland % 7.3% 33.1% 0.7% 4.3% 9.4% Italy % 4.4% 5.5% -12.4% -13.7% -27.4% Japan % 16.2% 16.2% -4.4% 14.7% -4.3% Korea % 3.7% -13.3% -3.4% 5.0% 43.8% Luxembourg % -1.8% 49.8% 14.2% -6.2% -11.3% Mexco % -5.0% -24.6% -16.7% -2.7% 21.3% Netherlands % 7.6% 16.8% 24.3% 1.9% 5.0% New Zealand % -1.1% -4.3% -15.8% 9.9% 18.8% Norway % -6.1% 35.1% 19.5% 7.6% -6.0% Poland % -4.8% -42.9% -16.7% -7.3% -0.5% Portugal % -4.3% -25.9% -21.5% -17.4% -0.8% Slovak Republc % -11.3% -27.6% 12.0% -7.7% 2.9% Span % -20.1% 0.5% 2.7% 4.2% -35.5% Sweden % 10.9% 12.4% 24.0% 7.6% -15.0% Swtzerland % 13.4% 21.7% 15.5% 4.2% -0.7% Turkey % -6.4% -33.0% -14.8% -5.8% 14.3% Unted Kngdom % 16.6% 21.4% -9.8% 6.5% -10.3% Unted States % -2.9% 25.6% -13.3% 8.4% 10.0% Average % 0.0% 0.0% 0.0% 0.0% 0.0%

25 24 PHOENIX CENTER POLICY PAPER [Number 33 Table 3. Explanng Departures from the Mean Broadband Subscrpton Rate Contnued from prevous page) Country DENSITY BIGCITY HHSIZE BUSSIZE PHONE Australa -20.2% -7.1% 2.1% 23.8% 5.5% Austra 2.2% -10.4% 2.2% -5.2% 5.6% Belgum 10.4% 10.2% 2.2% 0.6% 5.5% Canada -21.7% 1.8% 2.3% 10.8% 1.6% Czech Republc 2.6% 4.1% -6.4% 20.0% 5.6% Denmark 3.1% -7.7% -8.0% -1.5% 4.8% Fnland -10.1% -0.9% -10.2% -3.9% 5.0% France 3.6% -1.1% 2.6% -2.8% 4.6% Germany 6.3% 12.0% -8.2% -16.6% 5.5% Greece 1.1% -11.7% 1.9% 4.5% 5.2% Hungary 2.4% -3.6% 1.8% 14.5% 3.1% Iceland -9.8% -12.2% -4.8% 17.1% 4.1% Ireland -1.3% -11.8% 2.3% -22.6% 5.5% Italy 4.6% 10.4% 1.5% 15.2% 5.5% Japan 11.8% -12.9% 2.5% -10.5% 4.6% Korea 9.1% -7.6% 7.8% 1.7% 5.1% Luxembourg 3.3% -1.3% 1.4% 5.5% -6.3% Mexco -0.5% -0.9% 3.1% -2.3% -64.5% Netherlands 10.1% 13.2% -9.0% -5.8% 5.6% New Zealand -11.6% -11.3% 2.5% 19.0% 3.9% Norway -9.4% -1.9% 2.0% 7.8% 4.2% Poland 2.4% 8.1% 1.4% 3.4% -15.2% Portugal 3.3% -6.3% 2.2% 11.7% 5.3% Slovak Republc 1.5% 7.6% 1.1% -26.2% -3.5% Span 1.7% 8.5% 2.9% 14.7% 5.6% Sweden -6.2% -0.7% -7.6% 12.7% 3.2% Swtzerland 5.2% 5.3% -8.4% -20.5% 4.9% Turkey 0.5% 1.4% 5.2% -7.8% -28.9% Unted Kngdom 8.9% 5.1% 2.4% -11.5% 5.4% Unted States -3.3% 10.9% 1.3% -21.4% 3.7% Average 0.0% 0.0% 0.0% 0.0% 0.0%

26 Sprng 2008] BROADBAND EFFICIENCY INDEX 25 Table 4. Actual and Fronter Subscrpton Rates and Ranks December 2007) Country Dec Subscrpton Rate Rank Dec Fronter Subscrpton Rate* Rank Australa Austra Belgum Canada Czech Republc Denmark Fnland France Germany Greece Hungary Iceland Ireland Italy Japan Korea Luxembourg Mexco Netherlands New Zealand Norway Poland Portugal Slovak Republc Span Sweden Swtzerland Turkey Unted Kngdom Unted States * Fronter subscrpton rates may be below actual subscrpton rates due to statstcal nose. Raw rate data from OECD