CountryLevel Variation in Open Source Software Policy and Environment

1
Country-Level Variation in Open Source Software
Policy and Environment
ICT Roundtable February 19, 2009
2
Outline

• OSS background/theory/empirics
• Interviews
• Data collection
• Index construction
• Maps
• SLOSI

3
OSS Background

• Free/Libre Open Source Software (FLOSS)
• Passion, noise, potential, and accomplishments
• Relatively new area for academic inquiry
• Lots of interesting questions about
• Decisions to contribute / coordinate / adopt
• Quality / Cost advantages over closed-source
• Policy levers to affect OSS development

4
Understanding OSS

• Economics and business of OSS
• altruism, signaling, reputation
• alumni effects, lock-in, standards
• business models (complementarity), hierarchy
• Law and economics of OSS
• licensing and IP
• Cultural studies of OSS
• hacker, international norms

5
Overview of OSS knowledge

• Beliefs, hype, arguments and prophecies
• Lots of anecdotal evidence
• Firm- or project-specific stories
• Sampling on the dependent variable
• Snapshots of participants, adopter groups
• Paucity of systematic data collection and robust
  hypothesis-testing

6
OSS empirics

• Empirical evidence on OSS activity limited
• Open nature limits data collection
• Informal participation, org. structure
• difficult to monitor / efforts proprietary
• No transactions
• Dynamic, disaggregated, dispersed activity
• Predictors difficult to observe
• culture, training, altruism, reputation, etc.
• IT costs for OSS vs. other (IT) costs, output

7
Interviews

• Red Hat requests development of OSS
• OSS activity? OSS potential?
• Background/Context
• Interviews with OSS professionals
• Red Hat executives/developers
• International OSS experts
• - (Brazil/Latin America/India/Singapore/Germany/Fr
  ance)

8
The goal?

• Develop an OSS Potential Index (OSPI)
• Global, country-by-country index
• akin to HDI, Index of Economic Freedom, etc.
• let policymakers and advocates (e.g., Red Hat?)
  point to a particular countrys rank
• Compare to neighbors, link to policies

9
Index Design

• Develop OSS Potential Index (OSPI)
• theoretically relevant
• consistent data available
• direct and indirect measures
• Compile the index
• Test for sensitivity to construction
• Keep OSPI construction/composition open

10
Data collection

• 750 variables collected
• All publicly available data
• Issues of coverage
• Across time, across countries
• Priority is establishing the framework
• open index to entice data provision

11
Index construction

• Conceptual approaches
• Activity vs. Potential
• OSPI , OSAI
• Index composed of
• Dimensions government business
  community/education
• Indicators transformations of variables
• Variables

12
INDEX f(Dimension 1, ..., Dimension i, .,
Dimension I) Dimension i g(Indicator 1,
..., Indicator j, , Indicator
J) Indicator j h(Variable j)
13
INDEX f(Dimension 1, ..., Dimension i, .,
Dimension I) Dimension i g(Indicator 1,
..., Indicator j, , Indicator
J) Indicator j h(Variable j)
Active INDEX f(GA, FA, CA) Potential INDEX
f(GP, FP, CP) G government F firms or
commercial enterprises C community and
educational system
14
Index construction

• Conceptual approaches
• Activity vs. Potential
• OSPI , OSAI
• Index composed of
• Dimensions government business
  education/community
• Indicators transformations of variables
• Variables
• direct (related to or impacting OSS) or indirect
  (contextual)
• (Arbitrary) weights

15
Index construction

• Index theory
• Weights/transformations/aggregations affect
  rankings
• - E.g., the HDI rankings shuffle if ln(GDP) is
  used
• - No theory for appropriate transformation,
  weights, in OSS index
• Every numeric variable is ratio or interval
• - ratio has natural zero (e.g., pop., Firefox
  installs)
• - interval does not (e.g., F, Linux language
  support)
• Geometric means of ratio vars preserves rank
  ordering (preferred as g function)

16
proposed index structures
Index A
Index B direct indirect
Index C
OSAI
G
C
GA
FA
CA
OSPI
GP
FP
CP
17
Index construction

• Variables classified as
• Active / Potential
• Direct / Indirect
• Long / Short
• Ratio / Interval
• Missing values create problems
• Transformations of variables to remove scale
• Z-scores used here (as h function)
• Aggregations (g function)
• Weights arbitrary? endogenous?

18
Index construction

• Lots going on
• 2 primary indices (active potential)
• 5 aggregation rules for f
• arithmetic mean, maximin, minimean, geometric
  mean, R2 weights
• 2 sets of variables (long short)
• 3 dimensions (govt, firms, community/edu)
• A total of 60 combinations

19

• 11 out of 57 cells lack suitable and available
  variables
• 2 out of 23 indicators had no available variables
• All Potential indicators have an available
  variable

20
Summary of results

• Correlations among rank-orderings were high
  across different aggregation rules
• Rankings are fairly stable
• Geometric mean rankings were least correlated
  with other rules
• Recommended indices (a.m. g.m., su.am.) are
  correlated in value but do differ in ranks
• Correlated at 0.79 for OSAI, OSPI
• a.m. and g.m. correlated at 0.87 (P) and 0.67 (A)
• Coverage a.m. has most (Ngt132), g.m. has least
  (N51), su.am. in middle L has double Ss
  coverage

21
Frontier analysis

• Stochastic frontier analysis uses background
  attributes or endowment to predict OSPI / OSAI
• Lets us inductively see if factors (e.g., income,
  education) predicts index score
• Lets us see which countries under- and
  over-achieve based on their endowments
• Another index thus becomes available

22
Maps
OSAI (arithmetic mean, long) Green indicates
low, Red indicates high
23
Maps
OSPI (arithmetic mean, long) Green indicates
low, Red indicates high
24
Maps
OSAI (geometric mean, long) Green indicates low,
Red indicates high
25
Maps
OSPI (geometric mean, long) Green indicates low,
Red indicates high
26
Maps
OSAI (weighted mean, long) Green indicates low,
Red indicates high
27
Maps
OSPI (weighted mean, long) Green indicates low,
Red indicates high
28
Maps
OSAIeff (geometric mean, long) Green indicates
low, Red indicates high
29
Maps
OSPIeff (geometric mean, long) Green indicates
low, Red indicates high
30
State-level index

• Experimented with constructing a within-US index
  (SLOSI)
• Take 5 indicators (per capita)
• Rank each 1 lowest value
• Sum the states rank across all 5 indicators
• Rank the sum the SLOSI

31
State-level index

• 5 indicators
• open source hits in state website,
• firefox hits in state website,
• Linux jobs _at_ Monster,
• open source jobs _at_ Monster,
• Linux user groups in state
• Also collected a statepolicy variable
• State-level policy activity concerning OSS
  binary (0/1)

32
(No Transcript)
33
State maps
34
State maps
35
Predictions

• logit model predicts statepolicy using things
  like
• SLOSI
• Number of software companies per capita
• Not so successful.
• At best, only OSjobs significant and positively
  related
• Measurement error? Endogeneity?
• OSS policy may arise when OSS is strong or
  because it is weak and needs policy help

36
Acknowledgements

• The authors wish to acknowledge the generous
  support of Red Hat, Inc. in funding this
  research.
• The authors also wish to recognize research
  assistance provided by the following individuals
  Art Seavey, Nathan W. Moon, Ankit Kharadi and
  Saswat Anand.