Page 177 - ISC PROCEEDINGS 21.4

primary regressor is internet penetration (IT.NET.USER.ZS), the share of the population
using the internet in the past three months. Two alternative ICT measures are used for
robustness: fixed broadband subscriptions per 100 people (IT.NET.BBND.P2) and mobile
cellular subscriptions per 100 people (IT.CEL.SETS.P2). Controls include log GDP per capita
(constant 2015 USD), tertiary school enrolment, trade openness (% of GDP), and inflation.
Income groups follow World Bank Atlas thresholds. All continuous variables are
winsorised at the 1st and 99th percentiles.
Table 1. Variable definitions and sources

Variable Description WDI Code Source
World Bank /
gini Gini coefficient SI.POV.GINI
PIP
internet Internet users (% of population) IT.NET.USER.ZS ITU / WDI
Fixed broadband subscriptions per
broadband IT.NET.BBND.P2 ITU / WDI
100
mobile Mobile cellular subscriptions per 100 IT.CEL.SETS.P2 ITU / WDI
ln(gdppc) Log GDP per capita (const. 2015 USD) NY.GDP.PCAP.KD World Bank WDI
edu Tertiary school enrolment (%) SE.TER.ENRR UNESCO / WDI
trade Trade (% of GDP) NE.TRD.GNFS.ZS World Bank WDI
inflation Inflation, consumer prices (annual %) FP.CPI.TOTL.ZG IMF / WDI
Source: Sample period 2000–2022. Retrieved via wbopendata.
All variables winsorised at 1st–99th percentiles.
Table 2. Descriptive statistics (N = 1,346)
Variable Obs Mean SD Min Max
Gini coefficient 1,346 37.81 9.13 19.20 64.80
Internet (%) 1,346 43.72 28.94 0.09 99.00
Broadband (per 100) 1,275 10.18 13.52 0.00 46.10
Mobile (per 100) 1,346 84.52 43.61 0.43 176.80
ln(GDP per capita) 1,343 8.44 1.30 5.49 11.36
Tertiary enrolment (%) 1,054 42.38 27.43 1.10 119.80
Trade (% GDP) 1,316 78.29 43.05 14.22 220.40
Inflation (%) 1,316 5.14 11.22 −4.48 89.10
Source: Sample period 2000–2022. Retrieved via wbopendata.
3.2. Model specification
The baseline model is a quadratic two-way fixed effects regression:
Ginᵢₜ = αᵢ + γₜ + β₁ Internetᵢₜ + β₂ Internet²ᵢₜ + X′ᵢₜδ + εᵢₜ (1)
where αᵢ captures time-invariant country heterogeneity, γₜ absorbs common global
shocks, and Xᵢₜ is the control vector. A finding of β₁ < 0 and β₂ > 0 confirms the U-shaped
DKC. The turning point is recovered as:
̂
τ = −β̂₁ / (2β̂₂) (2)
and tested via the nlcom delta-method in Stata. Income-group heterogeneity is
tested by augmenting Equation (1) with full interactions between the internet terms and
World Bank income group dummies (High income as reference), yielding group-specific
turning points.
3.3. Estimation strategy
A Hausman test confirms FE consistency over random effects. Given the global
panel structure and high probability of cross-sectional dependence, the preferred

176