The Determinants of Broadband Availability: Economics, Demographics,

1
The Determinants of Broadband Availability
Economics, Demographics, State Policy

• Kenneth Flamm
• University of Texas at Austin
• kflamm_at_mail.utexas.edu

2
Motivation

• Very preliminary work presented today
• FCC data on broadband entry now offers
  opportunity for longitudinal analysis relevant to
  major telecomm policy issues
  
• Linking to multiple other data sets, have
  constructed rich data set, sophisticated models
  with greater range of explanatory variables now
  possible
• Extends and improves on early work of others,
  some new approaches to be outlined below
  
• New results, relevant to policy

3
Overview of FCC Data

• FCC identifies zip codes where at least 1 high
  speed line installed
  
• Estimates zip codes where no high speed lines, to
  track penetration
  
• FCC maps point zip codes to geographic zip
  codes
  
• Result remote areas with no regular mail service
  absorbed into zips with mail delivery
  
• Census maps remote areas with no regular mail
  service to post office of boxes/general delivery
  for remote residents
  
• Maps geographic areas to point zips (actually
  ZCTAs)
  
• 3245 areas with P.O. Box-only delivery zip codes,
  no conventional mail delivery in 2000
  
• Census the only organization mapping zip codes to
  people

4
Implications for FCC-Census Match-up

• Implications
• FCC BB numbers probably overestimate
  providers/zip in zips to which point zips are
  mapped
  
• FCC BB numbers for zips with ANY service probably
  about right
  
• Probably very few remote areas without mail
  service but with broadband, that adjoin more
  populated areas with mail service but without BB
  
• Point Census zips not showing up on FCC list do
  NOT necessarily not have broadband service
  
• Confining analysis to geographic zips only
  probably best fix
  
• But understand that remote, sparsely populated
  rural zips underrepresented in resulting sample
  
• Issue important for geographic BB coverage, but
  no longer important for population BB coverage
  

5
Rapid Change in US Broadband Penetration,
Competition over 4 Years
Note Census zips not showing up on FCC list
credited with 0 BB providers overestimates
true zeros to unknown extent
6
99 Population now has at least 1 provider in
their zip code
Note providers may be overestimated in
geographic zips to which point zips have
been assigned by FCC
7
Economic models of broadband penetration

• L-R Approach Firms enter markets to make
  profits
  
• Market characteristics
• Demand side consumer socioeconomics,
  demographics
  
• Supply/cost sides technology, geography,
  regional cost factors
  
• Approach estimate reduced form
• solve for number of firms that fit into
  market as function of characteristics
  
• Price and quantity solved for as functions of
  exogenous variables, given N players in market
  and all above characteristics
  
• Simplest decision for anyone to enter
  marketrequires few assumptionsjust ask whether
  a hypothetical monopolist would make a profit
  
• Much more complex decision if we ask how many
  enter
  
• Need to assume oligopoly model
• Need to deal with asymmetries among players

8
Ordered Choice Models

• The natural way to think about this decision
• Hypothetical monopoly profit 0, enter,
  otherwise dont
  
• An unobserved latent variable a function of
  market characteristics
  
• Logit or probit a natural solution
• For number of entrants
• Profit of least profitable potential entrant 0,
  enter
  
• Next least profitable entrant ends up with profit
  
  
• Construct function N giving number of entrants
  that just makes marginal entrant profit 0
  
• Since N is integer, largest integer N defines number of entrants in equilibrium
  
• N is a latent variable that gives number of
  firms, falls below integer N cut point
  
• Ordered logit or probit marginal the natural
  choice

9
Data Issues

• Have constructed zip code level longitudinal
  (2000-2003) panel from 7 sources
  
• FCC high-speed survey
• FCC CLEC competiton survey
• 1997 Economic Census
• 2000 Population Housing Census
• Universal Service Fund School and Library
  (eRate) and Rural Health Care funding
  Commitments
  
• Hydrographic, topological, land cover geophysical
  databases
  
• Plus, various zip code data bases
• The bad newsA lot of tedious work
• Still not done!
• Still fixing small issues in data
• The good newsA very rich data set

10
Current Research Road Map
Simple logit/probit (single years)
Correlated Data model (panel data)
Any Entrant at all (fewest assumptions)
Bivariate logit/probit (Use info on CLEC competit
ion)
Todays Talk
X
Ordered logit/probit
Fails proportional odds/ lines test
Number of entrants (more assumptions)
Non-proportional ordered models
Partial proportional odds Continuation ratio
Generalized ordered logit
11
Initial observations

• Functional form
• Preliminary work showed logs for selected
  continuous variables worked marginally better
  
• Little difference in coefficient signs,
  significance
  
• Years covered ?? terrain variables
• Results for 2000 led to investigation of
  geophysical/terrain variables
  
• 2000 known to have data collection problems, FCC
  revised
  
• 2000 results qualitatively different from later
  years
  
• 2000 dropped
• CLEC competition data
• In principle, could be used to separate telephone
  competition from other elements of state BB
  policy
  
• Simultaneity, identification issues
• Stuck with completely reduced form

12
Econometric Approach

• Start with standard binary logit models for
  12/00, 12/01, 12/02, 12/03
  
• Any statistically significant variable (10
  level) in any year goes into interesting pool
  
• Relax statistical assumptions
• Error term generalized to entire exponential
  family
  
• Calculate robust standard errors
• Exploit information in repeated observations on
  zips over time
  
• Longitudinal panel data structure
• Allow coefficients to vary over time
• Allow for correlations in observations over time
• Generalize estimating equation (GEE) estimator
• A cousin of generalized method of moments (GMM)
  estimator

13
Whats Statistically Significant, Inclusive GEE
Model (all significant variables in any standard
2001-03 logit included, 10 level), 2001-02

• Effect on BBand Penetration
• Population density
• Geographic size of zip
• Mean CTI Wetness index (actually marginal at
  11 level)
  
• MODIS land cover classification (1, 4, 11,
  baseline urban)
  
• NAICs 31, 44, 54, 72, 81 Estabs
• Pct Pop on Farms
• Pct Pop 55-74
• Afro American (2001 only), Native American,
  Native Hawaiian Pop
  
• English is 2nd language
• Higher educational attainment
• Share Over 16 in Armed Forces, Civ Labor Force or
  Not in Labor Force
  
• Share pop working in education
• Per capita income
• Occupied housing density
• Share of houses occupied
• Share of housing indoor plumbing
• Share Living in new building
• Share living in 50yr old building

• Sign
• -
• -
• -
• -
• -
• -

14
Whats Statistically Significant, Parsimonious
GEE Model (only significant variables from
inclusive models, 10 level), 2001-02

• Effect on BBand Penetration
• Population density
• Geographic size of zip
• Mean CTI Wetness index (actually marginal at
  11 level)
  
• MODIS land cover classification (1, 11, baseline
  urban)
  
• NAICs 31, 44, 54, 72, 81 Estabs
• Pct Pop on Farms
• Pct Pop 55-74
• Afro American (2001 only), Native American,
  Native Hawaiian Pop
  
• English is 2nd language
• Higher educational attainment
• Share Over 16 in Armed Forces, Civ Labor Force or
  Not in Labor Force
  
• Share pop working in education
• Per capita income
• Percent pop female
• Occupied housing density
• Share of houses occupied
• Share of housing indoor plumbing
• Share living in 50yr old building

• Sign
• -
• -
• -
• -
• -
• -
• -

15
Dogs that did not bark Small point estimates,
not significant

• Numbers of households
• for given pop density, zip size is scale
  variable
  
• Similarity of coefficients?pop is scale
• Pop/housing unit
• Household size
• Age variables (except 55-74, over 16)
• eRate, rural health care grants

16
Role of State Policy/Effects

• Baseline was Texas, impact on BB
• May not be best baseline, but many zip codes,
  active state subsidy program (TIF, 1996-2004,
  1.5B), relatively competitive market
  
• Statistically significant effects
• CA,CO,FL, MD, MA, NY, NC, OR, TN,
• MD TN increasing in 2002 relative to 2001
• Statistically significant effects
• IL, IN, IA, KS, MN, MO, NE, NV, ND, PA, SD, UT,
  VA, WI
  
• IL, IA, KS increasing in 2002 relative to 2001
• Greater in 2001, parity in 2002
• CT, ME
• Less in 2001, parity in 2002
• HI, MI, WV

17
First Pass at an Ordered Model
4 levels 0, 1-3, 4-7, 8
Score Test for the Pr
oportional Odds Assumption
Chi-Square DF Pr ChiSq
2001 2406 218 More flexible model needed!
18
Possible endogeneity issues

• Industry ? ?broadband availability
• But pre-bb industrial mix (97 econ census)
• e-Rate ? ?broadband availability
• But long lags for e-Rate apps-approvals-commitment
  s-disbursements
  
• Similar issues for car ownership, home quality?

19
Conclusions

• Estimated state effects correlate with accounts
• Terrain effects significant in some parts of
  country
  
• Terrain effects exciting!
• Instrumental variables for demand studies
• Income, pop density expected effects
• eRate irrelevant
• Industrial activity very significant (prof
  technical services largest)
  
• Gender, education, farm location as expected
• Age effects generally not supported
• Digital divide ethnicity/gender show up, but
  small effects and decreasing
  
• Tests for proportional odds/parallel lines
  hypothesis critical in ordered logit/probit
  models