Productivity impact of broadband and ICT use

1
Productivity impact of broadband and ICT use
George van Leeuwen Section Sience and
Technology Statistics Statististics Netherlands
SCB conference Stockholm, October 2008
2
Outline of presentation

• Some results of core analysis EUROSTAT ICT
• project for the Netherlands, w.r.t.
• readiness
• intensity of ICT use.

• Econometric analysis of ICT impact on
  productivity
• using linked data.
• Joint research with Shikeb Farooqui (ONS),
• chapter 10 final report EUROSTAT Project.

• Current research on ICT and innovation modes,
• follow up of chapter 12 final report.

3
ICT readiness

• macro core analysis EUROSTAT ICT project
• panel linked data EC, SBS, and Investment
  surveys

4
Intensity of ICT use

• How to assess productivity impacts ?
• Focus on broadband use (DSLpct DSLPCpct)
• and E-commerce

5
ICT use and ICT investment3 questions, 2
countries

• Using Netherlands and UK data we aim to answer
• Is IT a general purpose technology?
• If so, then IT use variables should not contain
  any extra explanatory power once IT investment
  and capital has been accounted for in the
  production regression
• Can we use the e-commerce IT use variables to
  predict IT capital stocks?
• Can we identify a broadband productivity impact
  if we do not use (cumulated) IT investment?
• (DSLpct DSLPCpct)

6
A graphical explanation

• Y/L value added productivity
• K/L capital intensity
• F frontier technology

7
Construction of IT capital Stock
Survey Investment in Fixed Assets
ARD Business Survey Quarterly Inv Survey Annual
Inv Survey
Initial Conditions Deflators PIMS
IT Capital Hardware Stocks Purchased Software
IT Capital Hardware Stocks
8
Data
9
Productivity regressions should control for skill
differences
10
Basic Production Function Specification

• LP f(kNIT , kIT) g(DSLPCT , epurch ,
  esales, ITM) h(skills) z(Firm controls)
• Firm controls Region, Sector, Year Dummies where
  available
• Skills Firm level wages
• If IT GPT find insignificant coefficients on g()
• Significant coefficients on DSLPCT, esales IT
  maturity
• Differences in IT use explored with broad sector
  breakdown
• UK insights Benefits of broadband enabling
  highest in Differentiated services
• Netherlands insights Weak evidence at sectoral
  level

11
Base Regressions

• Negative in manufacturing, positive in
  differentiated services

12
Potential Problems Concerns

• Sample Selection
• Are the results representative?
• Overlap sample displays characteristics that
  differ from Business Register. Problem more acute
  for UK correct using Heckman Selection
  procedure
• Endogeneity
• Skilled biased coefficients on IT capital and
  DSLPCT may crowd out impact of automated business
  processes
• Use wages as proxy for skills but wages highly
  correlated with productivity need to control
  for possible endogeneity
• Firms with better IT infrastructure e.g. high
  DSLPCT more productive
• But more productive firms likely to adopt
  broadband earlier need to control for possible
  reverse causality

13
IT use and IT capital stocks

• Two Step process
• The first equation is the prediction equation and
  gives us the relationship between ICT use and IT
  capital levels
• kIT f(PC, web, DSL, DSLPCT, epurch, esales,
  BI)
• h(skills) z(Firm controls)
• But IT capital levels are only observable for a
  sub-sample of all e-commerce firms, in order to
  correct for this we need to define a selection
  equation
• (Seln 1) f(DSLPCT, epurch, epurch, esales,
• esales) g(K, L) z(Firm controls)

14
IT Prediction equation
Range of e-commerce variables tested In both
countries we find significance for the same core
variables
15
Systems Approach to ICT use impact

• Augment the productivity equation
• LP f(kNIT , kIT) g(DSLPCT , epurch ,
  esales)
• h(skills) z(Firm C)
• With the IT capital prediction equation
• kIT f(PC, web, DSL, DSLPCT, epurch, esales,
  BI)
• z(Firm C)
• A wage equation
• w f(wt-1, DSLPCT) z(Firm C)
• And a (prediction) equation for NIT capital
• kNIT f(Proxy Capital Inputs) z(Firm C)

16
Implementation

• Estimation Procedure
• Estimate a Heckman selection model on each
  equation separately
• Evaluate the equation specific selection bias and
  capture in a new mills bias variable
• Re-estimate all four equations jointly, in a
  system, adding the mills bias variable to the
  core specifications
• What do we achieve by doing this?
• Estimating equations jointly controls for
  endogeneity
• Adding the mills bias variable controls for
  selection
• Procedure very demanding on data We can no
  longer split sample into broad sectors

17
SEM - Netherlands
18
SEM - UK
19
Productivity with predicted ICT stocks

• Use of predictions Heckman models in PF
• Additional firms NLD 4001, UK 3261

20
Discussion

• Useful approach for exploiting correlation
  structures in case of missing data resulting from
  survey merges
• Evidence that selected firms (marginally) more
  productive in their use of ICT technologies
  coefficients on ICT variables decrease when
  predicted firms also included in estimation
• Is e-sales a special case of innovation
  implemented by firms? (see evidence of this in
  NLD work on innovation
• in chapter 12 of final report EUROSTAT project)
• Is DSLPCT primarily capturing IT capital
  deepening or is it also a proxy for knowledge
  management (see evidence of this in UK work in
  chapter 12)

21
ICT and innovation (current research, (1))

• Furthers research of
• Hagen et al. (Sweden, 2007)
• Chapter 12 of final report EUROSTAT (NLD, UK)
• Robin and Mairesse (France, 2008)
• Looks at different innovation modes (Product -,
  Process -and Organisational innovation
• Identifies contribution of ICT (use) and RD to
  three modes of innovation
• Test for complementarities of innovation modes
  for productivity contributions
• Joint research of UNU/MERIT (Mohnen, Raymond) and
  Statistics Netherlands (van Leeuwen and Polder)

22
ICT and innovation (2)

• A Basic structure of CDM innovation model
  (Crépon, Duguet, Mairesse, 1998)
• 1. RD input is latent variable observed only
  for firms that
• reported RD
• 2. RD is input for producing new/improved
  products and
• production processes (with other
  factors measured by CIS)
• 3. Productivity depends on successful RD
  not on RD
• expenditure

B Potential weaknesses of CDM innovation
model 1. Not all innovation is RD related
(role of ICT neglected) 2. Too much emphasis on
product innovation (given problems to
measure innovation output)? 3. Process
innovation is exogenous.
23
ICT and innovation (3)

• Extension of Robin and Mairesse, 2008 (and
  before)
• 1. Down-grade role of innovative sales and
  up-grade role of ICT
• investment and ICT use for explaining
  product , process and
• organisational innovation, by
• 2. Using the same model for predicting per
  employee ICT and
• RD investment (two types of innovation
  expenditure) stage 1
• 3. Predict propensity to be involved in the three
  modes of innovation,
• conditional on (predicted) innovation
  expenditure (including ICT
• investment) from stage 1 and other ICT use
  variables stage 2
• 4. Use predictions of stage 2 in productivity
  regression and test for
• complementarity between innovation modes
  stage 3.

24
ICT and innovation (4) First results
25
Some conclusions

• ICT use indicators are good predictors for
  missing data on ICT capital stocks
• Broadband connectivity has no direct impact
• on productivity (but indirect via ICT capital
  deepening)
• ICT use indicators important for explaining
  differences in innovativeness
• Main productivity impact of ICT use is through
  enabling innovation (and thus TFP).

26
Thanks for your attention