Application of multilevel models to the study of the individual digital

divide in Russia

Introduction and prior research

The Digital Divide has emerged as an important research and policy issue during the past thirty
years. It can be defined as an inequality in access to Information and Communication
Technology (ICT), such as personal computers, Internet and mobile phones (Norris, 2001). The
digital divide on the individual level is defined as inequality in access to and ability to use ICTs
among individuals (Dewan & Riggins, 2005); it causes an inequality in the ability to obtain
important information, which in turn leads to inequality in the ability to gain employment,
participate in online communities, e-government, receive important health information etc.
While extensive research already exists on this subject, most existing quantitative digital divide
studies are limited to descriptive statistics or simple linear models.

Most of existing digital divide studies examine the difference in the access to ICTs by different
demographic groups. Such studies in the USA demonstrate the existence of income,
geographical and racial divides and some studies find a gender divide (Chakraborty & Bosman,
2005; Fairlie, 2004; Martin, 2003). Demunter (2005) in his study of the digital divide across
countries included in the European Union concludes that age and education level predict the
level of use of ICTs – individuals with higher income and higher levels of education have better
access to ICTs. Gender is not a significant factor in the divide and the gender gap is widest in
rural areas. Sciadas (2002) in a study of the digital divide in Canada also finds that income is a
significant predictor of access to ICTs. The growth rate of Internet use is higher at lower income
levels than at higher income levels. Willis and Tranter (2006) find income, age and the level of
education to be predictors of the individual digital divide in Australia, whereas gender is found
not to be a significant predictor.

Research on the digital divide in Russia, similarly to the overall individual divide research, is
mostly qualitative in nature or includes basic descriptive statistics only. The same predictors that
were found significant in studies of the digital divide in the USA, such as gender, age, and
income, are identified as predictors of the digital divide in Russia (Beketov, 2009; Delitsin,
2006; Lihobabin, 2006). Individuals with lower income, older people and women have lower
access to ICTs. In addition to these common problems, a language barrier affects the ability to
use Internet content for most Russians – most Internet content exists in English and is
inaccessible to most Russians (Beketov, 2009). Another aspect specific to the situation in Russia
is a very significant regional divide (Beketov, 2009; Delitsin, 2006). A gap in the rate of use of
ICTs arises between regional capitals and areas outside the capitals, with the highest rates of the
ICT use concentrating in Moscow and St. Petersburg. Lack of competition in the
telecommunication market and low purchasing power of the population are found to be reasons
for the regional divide.

While studies described above shed some light on the problem of the individual digital divide in
Russia, they are mostly qualitative and descriptive in nature. To the best of our knowledge, no
detailed quantitative study addressing the problem of the individual digital divide in Russia
exists. In this study we employ multilevel models to provide a comprehensive analysis of the
state of the individual digital divide in Russia and assess differences among individuals in their
use of Internet.

Data

We use data from the Russian Longitudinal Monitoring Survey (RLMS), a comprehensive living
standards survey covering every year from 1992 to 2010. Data collection is conducted at the
individual, household and communal level. The advantage of using this data source is the fact
that it is nationally representative and the fact that the data provide variety of information.

The RLMS dataset includes several variables representing ownership and use of technology by
individuasl. We chose to build models for the measures of the use vs. ownership of ICTs since
ownership does not guarantee use and individuals gain advantage only when ICTs are actually
used. Based on the high level of correlation between measures of the Internet and PC use (.85),
we chose to select one of these variables - Used Internet in the last 12 months – as a target
variable.
The choice of independent variables is guided by previous literature. Variables include gender,
age, measures of education, marital status, nationality and the knowledge of a foreign language
at the individual level; household income per capita at the household level, and the rural/urban
indicator, availability of an Internet Café, steady mobile communications and several other
measures at the population center (site) level. The complete list of variables included in the
model is presented in Table 1. We also included several interactions in the model to determine
the gender effect is different at different age levels. Originally the model included all of the
independent variables listed in Table 1, but the final model includes only variables that turned
out to be significant.

Table 1: Indicators used in the multilevel model

Variable name Question Level
Dependent Variables
Internet Used Internet in the last 12 Individual Yes = 41% No = 59%
months?
Independent Variables
ISGENDER - Male Gender Individual Male = 42% Female=58%
ISBIRTHY - Age Year of birth Individual Range: 18 - 104 Mean = 44
ISSPFLAN Do you speak any foreign Individual Yes = 19% No = 81%
For_language language other than the
languages of the USSR’s former
republics?
ISNATION - Nationality Individual Russian – 86% Non-Russian =
Russian 14%
ISHIEDUL Highest educational level Individual Incomplete secondary
confirmed by certificate or education = 26% Complete
diploma? secondary education = 27%
Completed professional
school = 23% Bachelor ,
Masters or post graduate
degree = 24%
ISMARIST Marital status Individual Single = 21% Married = 53%
Divorced = 13% Widowed =
13%
tincm_rs Real household income Individual Range: 0 - 391 Mean = 5.2
Income_per_cap (000)/number of household
members
sett_typ 1=Urban 2=Small Urban 3=Rural Site Urban = 68% Small Urban =
6% Rural = 26%
Elec_inter Frequent electricity interruptions Site Yes = 22% No = 78%
High_sp_itr High speed Internet available Site Yes = 93% No = 7%
Inter_cafe Internet Café Site Yes = 76% No = 24%

Steady_mob_comm Steady Mobile connection Site Yes = 95% No = 5%

1=Yes 0 = No
Trip_to_mos_000 Cost of trip to Moscow (000) Site Mean = 3 Range: 0 - 20
sites Site
Region Region
Interactions
Age*Gender
Income*Gender
Foreign
Language*Gender
Educational
variables*Income
Urban*Gender
Urban*Income

Methodology – Multilevel models

Multilevel models are an extension of a regression model which takes into account the complex
structure of the data. For example, in education, when analyzing the difference in students’ tests
scores, one needs to account for the fact that the data are hierarchical – students belong to
schools, schools belong to districts etc. Students in the same school tend to have some
similarities due to the effect of the school; households in the same area tend to be more similar
than households in different areas. In these cases the assumption of independence of observations
is violated. Multilevel models take into consideration the hierarchical structure of the data,
simultaneously analyze variables from different levels and correct the estimates of statistical
significance, overstated by conventional regression models (Goldstein, 2002).

Multilevel models are used in many disciplines, such as education, psychology, medical
research, economics etc. (Goldstein, 2007).

We build a 4-level model for the year 2010. The 4 levels include the individual level i (lowest
level), the household level h, the site level s and the regional level r (highest level). The term
“site” refers to a population center.

In general, the model takes the form:

Internet use=β 0 hsr + ∑ β whsr x wihsr +e ihsr
w

β 0 hsr =γ 00 +γ 01 Ζ hsr + f 0 r + v 0 sr +u0 hsr

β whsr=γ w 0+ γ w1 Ζ hsr + f wr + v wsr +uwhsr

where the dependent variable is the indicator of Internet use by the ith person from the hth
household from the sth site from the rth region. The x wihsr represent values of the wth explanatory
variable measured for the ith person from the hth household, from the sth site, from the rth region.
The β 0 hsr represent intercepts for each household h from site s and region r. The β whsr represent
coefficients for each household h from the site s from the region r for each independent variable.
The term e ihsr represents the level 1 random error that has mean 0 and variance σ 2ihsr. The term
Ζ hsr represents one independent variable, measured at the household level h at the site s from
region r. The coefficients γ 00 , γ 01 , γ w0 , γ w1 represent fixed regression coefficients, the u0 hsr and
uwhsr are random residual terms at the household level (they have mean 0 and are independent
from the terme ihsr ) with constant variance. The terms v 0 sr and v wsr represent random residual
terms at the site level which have mean zero and are independent from e ihsr . The terms f 0 r and f wr
represent random residual error terms at the regional level. This model allows the linear
combination for each individual to shift from the overall linear combination by an amount of
u0 hsr + v 0 sr +e ihsr + f 0 r . Also we allow the coefficients of some of the covariates to shift by a
random amount.

Results

A final model was selected based on the lowest value of the Bayesian Deviance Information
Criterion (DIC). Results of the Internet use model are presented in Table 2.

Results indicate that holding all other variables constant, married people have estimated odds of
Internet use approximately 63% higher than those of single people; and those who are divorced
have estimated odds of Internet use approximately 20% higher than single people. Ability to
understand a foreign language, being younger, having Russian nationality, having higher income,
living in an urban area and having education all positively affect the level of Internet use. Being
Russian, compared to other nationalities, increases the estimated odds of Internet use by
approximately 46%. The higher the level of education, the higher the level of Internet use – the
magnitude of coefficients increases gradually with each educational category, compared to the
level of Internet use by individuals with incomplete secondary education. The level of education
has the highest influence on the Internet use among all of the predictors used in the model.
Holding all other predictors constant, having at least a bachelor degree or completed professional
education increases the odds of using the Internet by 9 and 2 times respectively, compared to
those who have not completed secondary education.

Several population center level variables are significant in the model. The odds of using the
Internet almost double for those who live in an area with an Internet café available. Those who
live in an area with steady mobile communications have higher odds by 61% then those who do
not. One possible explanation of this result is that steady mobile communication in the area
coincides with other infrastructure developments, which in turn increase the odds of using the
Internet. Living in an urban area increases the level of Internet use, compared to living in a rural
area) by 57%. We used a cost of a trip to Moscow as a proxy for remoteness and not
surprisingly, it is significant. The higher the cost of a trip to Moscow is, the lower the odds are of
using the Internet.

Interestingly, the effect of gender is only significant in interaction with the urban effect. This
means that holding all other variables constant, a positive effect of an urban area on the odds of
using the Internet is higher for males than for females. This effect is true even outside the model:
the rate of the Internet use for both men and women in rural and small urban areas is very low –
only approximately 25% of both men and women have used the Internet in the past 12 months.
Whereas in urban areas, the rate of the Internet use is higher for males than for females: 45% of
women and 53% of men have used the Internet in the past 12 months.
Table 2 Results*
Fixed effects
Variable Coefficient St Error t value Log (Odds)
Foreign language 0.738 0.063 11.71 2.09
Age -0.109 0.002 54.5 0.9
Russian 0.376 0.079 4.76 1.46
Higher Education 1 2.275 0.072 31.6 9.73
1 0.353 0.064 5.52
Completed Secondary Education 1.42
Completed Professional education1 1.027 0.068 15.10 2.79
Urban 2 0.448 0.173 2.6 1.57
Divorced3 0.184 0.074 2.5 1.20
3 0.486 0.055 8.84
Married 1.63
Male*Urban 0.291 0.057 5.11 1.34
Log income per capita 0.687 0.030 22.9 1.99
Steady mobile communication 0.482 0.153 3.15 1.62
Cost of trip to Moscow (000) -0.50 0.017 29.4 0.61
Internet Café 0.687 0.141 4.87 1.99
Random effects
Site 0.357 0.068 5.25 1.43
Intercept -4.478 0.257 17.42 0.01
1
Reference category – Incomplete Secondary School
2
Reference category – Rural
3
Reference category – Single
*DIC=12327, n=16922, only site random effects were significant

Site effects on Internet Use

Interestingly, even after accounting for all above mentioned factors, regional effects became
insignificant, but site effects are still significant, which indicates that location itself affects the
odds of Internet use. Figure 1 shows the magnitude of this effect and the ranks of residuals for all
the sites included in the dataset.
Figure 1 Site residuals ranking

Site effects holding all predictors constant are listed in Appendix 1, sorted from the highest
negative to the highest positive value. The RLMS excludes site names in order to protect the
privacy of the respondents, so sites are identified by region names to which they belong and site
numbers.

Not surprisingly, Moscow, Moscow Oblast, St. Petersburg city are among the regions with the
highest positive effect on the level of Internet use, even after accounting for all variables
included in the regression model. Three Moscow Oblast sites and three sites from St. Petersburg:
Leningrad Oblast: Volosovskij Rajon are in the top 15 sites with the highest positive effect on
the level of the Internet use. Other sites with the highest positive effects are Vladivostok,
Krasnodarskij Kraj: Kushchevskij Rajon, Khanty-Mansiiskij AO: Surgut CR and Krasnojarskij
Kraj: Nazarovo CR. Interestingly, one site from Komi ASSR: Usinsk CR region has one of the
highest positive effects and one site has one of the highest negative effects. This indicates a high
level of diversity in terms of the Internet use within the region.
Multiple sites with the highest negative effects belong to the same regions: Volgograd Oblast:
Rudnjanskij Rajon, Kalinin Oblast: Rzhev CR, Amurskaja Oblast: Arkharinhskij Rajon and
Altaiskij Kraj: Biisk CR. This indicates that these regions have consistently negative effect on
the level of the Internet use. Some of these sites have some of the highest cost of a trip to
Moscow, which is a measure of site remoteness. However, even after accounting for this effect in
the model these sites still have a high negative effect.

Discussion

We find similarities and differences between significant predictors of the use of the Internet in
Russia and other countries. Similarly to studies in the US, EU, Australia and Canada we find an
income divide in the rate of use of the Internet. The geographical divide we find in our study is
also significant in the US. The level of education is also found to be significant in studies in the
EU, Australia and Netherlands. Gender was found to be insignificant in most studies and the
greatest gender differences were found in rural areas. We find the opposite result in our study:
the gender divide is the greatest in urban areas with men more likely to use Internet than women.
However, the effect of gender on the use of mobile phones is reversed: women are more than
men likely to use mobile phones.

Our results support some of the findings in previous studies in Russia. We confirm that age,
income, ability to speak a foreign language, identified as important in previous studies, are
significant predictors of the Internet use. We also find a significant regional divide, identified as
important in previous studies. We confirm the existence of a gender divide, but find that the
effect is greatest in urban areas. While the gender effect in rural areas is smaller, the rate of the
Internet use by both genders in rural areas is very low. We find other significant predictors of the
Internet use not used in previous studies, such as Russian nationality, marital status and
population center characteristics, such as availability of an Internet Café, steady mobile
communication and remoteness. We are able to account for these differences while holding other
predictors constant, extending past studies which used descriptive statistics. We are also able to
identify and quantify effects of different population centers on the rate of the Internet and Mobile
phone use.
This study has a number of limitations. The most significant limitation is data availability. The
RLMS dataset includes a sample of 38 Russian regions that were selected out of total of 1,850
regions. This limits our ability to explore the spillover effects of population center on one
another. Another data limitation is the limited availability of data related to technology use by
individuals. For instance, we have very limited information on the extent of the use of the
Internet by individuals.

Appendix 1: Site effects on Internet use

*Includes 10 sites with the highest positive and highest negative results

Region name Site # Site Rank

residual
Lipetskaya Obl: Lipetsk CR 72 -1.403 1
Kalinin Obl: Rzhev CR 68 -0.988 2
Volgograd Obl: Rudnjanskij Rajon 41 -0.984 3
Kalinin Obl: Rzhev CR 67 -0.899 4
Penzenskaya Obl: Zemetchinskij Rajon 121 -0.823 5
Amurskaja Obl: Arkharinhskij Rajon 96 -0.800 6
Kabardino-Balkarija, Zolskij Rajon 77 -0.773 7
Komi ASSR: Usinsk CR 89 -0.764 8
Volgograd Obl: Rudnjanskij Rajon 42 -0.738 9
Amurskaja Obl: Arkharinhskij Rajon 99 -0.713 10

Moscow City 138 0.557 138

Khanty-Mansiiskij AO: Surgut CR 162 0.686 142
Moscow Obl 152 0.726 143
Komi ASSR: Usinsk CR 91 0.774 144
St. Petersburg: Leningrad Oblast: Volosovskij Rajon 8 0.813 145
Krasnojarskij Kraj: Nazarovo CR 76 0.821 146
St. Petersburg 141 0.878 147
St. Petersburg: Leningrad Obl: Volosovskij Rajon 6 1.014 148
Moscow Obl 178 1.169 149
St. Petersburg: Leningrad Obl: Volosovskij Rajon 4 1.169 150
Moscow Obl 143 1.177 151
References

Beketov, H. B. 2009. The role of the digital diversity and the digital divide in Russian
development.
Chakraborty, J., & Bosman, M. M. 2005. Measuring the Digital Divide in the United States:
Race, Income, and Personal Computer Ownership. The Professional Geographer, 57(3):
395-410.
Delitsin, L. L. 2006. The problem of digital divide and development potential of the Internet in
Russia.
Demunter, C. 2005. The Digital Divide in Europe. Statistics in Focus, 38: 1-7.
Dewan, S., & Riggins, F. 2005. The Digital Divide: Current and Future research Directions.
Journal of the Association for Information Systems, 6(12): 298-337.
Fairlie, R. W. 2004. Race and the Digital Divide. Contributions to Economic Analysis & Policy,
3(1): 1-38.
Goldstein, H. 2002. Multilevel Statistical Models: A Hodder Arnold Publication.
Goldstein, H. 2007. Becoming familiar with multilevel modeling. Significance, 4(3): 133-135.
Lihobabin, M. Y. 2006. Gender Determinants of Information Societies.
Martin, S. P. 2003. Is the Digital Divide Really Closing? A Critique of Inequality Measurement
in a Nation Online. IT & Society, 1(4): 1-13.
Norris, P. 2001. Digital Divide? Civic Engagement, Information Poverty and the Internet in
the Democratic Societies. New York: Cambridge University Press.
Willis, S., & Tranter, B. 2006. Beyond the 'Digital Divide'. Journal of Sociology, 42(1): 43-59.