Ethnic digital divide? Evidence on mobile phone adoption

ABSTRACT

Between 2006 and 2014, Vietnam experienced a true information technology revolution. In rural areas, mobile phone ownership increased from 18% to 89%. I use a panel dataset from this period to study the ‘digital divide’ between the ethnic majority group and ethnic minorities. I find that underprivileged ethnic minorities are lag behind in phone adoption. These differences are fully explained by standard factors of demand and by education level, that is, I do not find unobservable barriers in adoption. Analysis using spatial data on digital mobile network coverage reveals that infrastructure is also not a barrier to adoption, even in remote regions of Vietnam. Results are robust to alternative model specifications.

KEYWORDS:

• Information technology
• digital divide
• ethnicity
• Blinder–Oaxaca decomposition

JEL CLASSIFICATION:

• J15
• L96
• O12
• O33

Acknowledgments

I thank Professor Finn Tarp and colleagues at the Development Economics Research Group at the University of Copenhagen, the participants at #DIYCSAE Oxford 2020, and seminar participants at SSDEV workshop in Prato Italy for useful comments and suggestions. I take sole responsibility for any remaining errors.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ In particular, Vu, Hanafizadeh, and Bohlin (2020) summarize evidence from 208 studies on the topic.

² GSM is a standard developed by the European Telecommunications Standards Institute to describe the protocols for second-generation (2G) digital cellular networks used by mobile devices such as mobile phones and tablets. More information can be found in: https://www.gsma.com/

³ VARHS cannot differentiate between a fixed-line phone and a mobile phone, but the vast majority of the increase in phone ownership is attributable to the increase in mobile phones as illustrated in Figure 1. The mean resale value of a phone in VARHS was 707,000 VND, that is around 31 USD (Using Google exchange rate 15 April 2017).

⁴ Only the 2014 data is shown, as the GSM coverage is the same in 2012 and 2014.

⁵ Operators have submitted strong ($\ge$ −92dBm) and variable ($<$ −92dBm and $\ge$ −100dBm) signal strengths. The data includes both types but does not make a distinction between the two (Collins Bartholomew 2014).

⁶ In this kind of static setting the durable good and consumption do not differ, but the distinction would matter in a dynamic setting.

⁷ Multiplicative network externalities $U(c,d,D)=c^{\alpha }(df(D){)}^{1-\alpha }$ would yield similar implications for demand.

⁸ Plugging this formulation in the demand function yields an individual’s demand for phones (now for additive formulation): $d^{\star }=\left(\left(1-\alpha \right)\frac{y}{p_d}+f(D)\right)h$.

⁹ This is similar to the approach by Middleton and Chambers (2010); Tien and Fu (2008). We also conduct robustness checks using logit and probit for our binary outcome variable of interest in Table B4, similar to approach taken by Middleton and Chambers (2010).

¹⁰ Specifically, the empirical counterpart of $D$ is $D_{ik}=\frac{{\sum }_{j=1}^{n_k-1}{d}_{j\ne i}}{n_k-1}$ where $n_k$ is the number of households in district $k$, and $d$ is a dummy denoting phone ownership. This is the closest empirical approximation possible for network benefits given the data available. The measure now could however be subject to reflection problem as described by Manski (1993), such that the network benefits may be indicative either on how $n_k-1$ mean group behaviour affects individual behaviour, or, indicate an aggregation of individual behaviours in the $n_k-1$ group. The empirical network benefits are considered a proxy, due to the absence of exact information on the respondents’ phone networks.

¹¹ Following theoretical possibilities presented in Appendix C Metcalfe’s law is closest to the first option. Zipf’s law (Zipf 1935,1949) would best be approximated by the first option with small network sizes, and by the second option with large network sizes.

¹² The minimum price can be either a bundled good with both airtime and the phone for a specified time-period, or those two goods separately, depending on which option in each year was the cheapest. Alternatively, a quality-adjusted price index would capture the evolution of the average prices (Yun, Kim, and Kim 2019; Kim and Kim 2018). The construction of such price index would require detailed information on phone prices for models sold in Vietnam, which, in the absence of such data, is not feasible to construct.

¹³ The survey is a collaboration between the Development Economics Research Group (DERG) at the Department of Economics at the University of Copenhagen, the Central Institute of Economic Management (CIEM), the Institute for Labour Studies and Social Affairs (ILSSA), and the Institute of Policy and Strategy for Agriculture and Rural Development (IPSARD) in Hanoi, Vietnam.

¹⁴ These are calculated at the level of the district of which there are 136. In rural communities people interact with neighbouring and nearby communes in their district for trade, non-farm labour, education etc.

¹⁵ The Mobile Coverage Explorer provided by Collins Bartholomew (Collins Bartholomew 2014).

¹⁶ Mobifone’s signal is centred in urban areas and province capitals, which are also covered by the other operators (According to OpenSignal.com maps (accessed 14 August 2016).

¹⁷ This is in line with government reports of full access and also the relatively high ICT infrastructure readiness in the country observed prior to the study period (Hanafizadeh, Saghaei, and Hanafizadeh 2009).

¹⁸ In merging the GSM coverage data with VARHS, not all VARHS communes could be merged with the GSM data: I only have information from 1,681 households.

¹⁹ The quadratic formulation is used due to the possibly non-linear nature of adoption patterns, as discussed in Section 2.

²⁰ The variable takes the form: $\frac{\mathrm{p}\mathrm{h}\mathrm{o}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{a}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a}-mean(\mathrm{p}\mathrm{h}\mathrm{o}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{a}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a})}{sd(\mathrm{p}\mathrm{h}\mathrm{o}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{a}\mathrm{p}\mathrm{i}\mathrm{t}\mathrm{a})}$

²¹ The coefficients not displayed are the number of children below 15 years, a dummy for whether the household speaks Vietnamese, log food expenditures, total area owned, toilet, good water, and dummies for shocks.

²² The theoretical frameworks of Metcalfe’s and Zipf’s laws do not provide a straightforward explanation for this, as they only describe a relationship between the utility of belonging to a network as a function of the network size.

²³ While the key variables of interest also have high correlations with phone adoption (Table A2), also other variables such as food expenditures and wealth have high correlations. Yet, the gap is explained away without adding all highly correlated variables to the model.

Additional information

Funding

This work was supported by the Samfund og Erhverv, Det Frie Forskningsråd [7024-00030B]; University of Copenhagen [DERG – Coordinated by Professor Finn Tarp].