The diffusion of internet use across China: Spatiotemporal models with policy implication

Although spatiotemporal models have been developed to capture the diffusion of new offerings in space and time, these efforts can hardly be used to assess the potential impact of policy decisions due to (i) their relatively large number of parameters (when data series are often limited) and (ii) their lack of time varying covariates. Besides, most of these models are intrinsically nonlinear in the parameters which can raise challenging issues when one deals with large systems of equations that describe diffusion in several areas. This study introduces relatively new spatiotemporal models to the diffusion literature; these models include a small number of parameters as well as time varying covariates. They can be applied when the number of areas becomes large. The study operationalizes three kinds of spatial autocorrelation models to uncover the underlying spillover effects which describe the influence of one region over the others: (i) The endogenous spatial lag model (SAR), (ii) The spatially autocorrelated error model (SEM), and (iii) A model that includes both the endogenous spatial lag and spatially autocorrelated errors: The spatial Durbin model. To the best of my knowledge, these models have not yet been used in diffusion studies despite their potential appeal. In contrast with the diffusion models which have been rooted for the most part in the Bass model, an intrinsically non-linear model in its parameters, these spatiotemporal models are extensions of the classical linear model. They offer a parsimonious and flexible manner for the analysis of the interdependence between observations in space and time.;The study deals with the diffusion of Internet use across Mainland China. Tested on 12 years of annual data (1998-2009) drawn from the China Internet Network Information Centre (CNNIC), the spatiotemporal models exhibit larger log-likelihoods than the non-spatial models despite a smaller number of parameters (except for the SAR model). The spatial Durbin model appears more consistent with the data than SAR and SEM. The empirical results support the existence of a strong spatial autoregressive coefficient (p = .444). The dependence arises from the endogenous interaction as well as exogenous interaction effects: The number of Internet users in a province can be driven (i) by the number of Internet users in the neighboring provinces and the more distant ones, and (ii) by the exogenous determinants of Internet use in both the focus province and the neighboring provinces. The impact of these accelerators (or hindrances) can be decomposed into two parts: the direct (i.e., within province) effects and the indirect effects (i.e., those from the neighboring and the more distant provinces). For example, a decrease in illiteracy rate results in an increase in the Internet use rate both in the focus province, the neighboring provinces and the more distant provinces. A simulation shows that decreasing the illiteracy rate in Tibet does not only benefit Tibet (with an increase in the proportion of Internet users) but also other provinces: the ripple effect. Decreasing illiteracy rate in Tibet by 80% over an eight-year period increases Internet use by at least 26.2% in Tibet, 5.8% in the first-order neighboring provinces, 3.5% in the second order neighboring provinces and 2.1% in the third-order neighboring provinces (compared to the predictions without accelerated change) over the same time span. These results provide guidance on the potential impact of policy decisions.