Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
The diffusion of internet voting. Usage patterns of internet voting in
Estonia between 2005 and 2015☆
Kristjan Vassil
a,
⁎, Mihkel Solvak
a
, Priit Vinkel
b
, Alexander H. Trechsel
c
, R. Michael Alvarez
d
a
University of Tartu, Estonia
b
Tallinn Technical University, Estonia
c
European University Institute, Italy
d
California Institute of Technology, United States
abstractarticle info
Article history:
Received 28 December 2015
Received in revised form 21 June 2016
Accepted 22 June 2016
Available online xxxx
E-voting has the potential to lower participation thresholds and increase turnout, but its technical complexity
may produce other barriers to participation. Using Rogers' theory of the diffusion of innovations, we examined
how the use of e-voting has changed over time. Data from eight e-enabled elections between 2005 and 2015
in Estonia, were used to investigate changes to the profile of e-voters and contrast them to those voting by con-
ventionalmeans. Owing to the aggregate shareof e-voters increasing with each election, with one third of voters
now castingtheir vote remotely over the internet, there wasa lack of conclusive evidence regarding whether the
new voting technology had diffused homogenously among the voting population, or remained a channel for the
resourceful and privileged. Our findings show that diffusion has taken place, but not until after the first three e-
enabled elections. Thus, internet voting has the potential to be used by a wide range of voter types,bridge societal
divisions, and emerge as an inclusive innovative voting technology.
© 2016 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Keywords:
Internet voting
E-voting
Diffusion
Voting behavior
1. Introduction
Remote internet voting
1
has long been discussed as a means of in-
creasing voter turnout in developed democracies, especially among
younger people (Alvarez & Hall, 2004; Alvarez, Hall, & Llewellyn,
2008; Norris, 2001, 2003). However, such technology can only have a
significant impact on political participation when its usage becomes
widely diffused. Voting technologies can empower people who have
faced participation hurdles (Vassil & Weber, 2011). Socially excluded
groups or people with reduced mobility should especially benefitfrom
modes that make it easier to cast a vote (Alvarez & Hall, 2004; Gibson,
2001). Such increased empowerment might also increase voter confi-
dence and their willingness to participate in elections (Alvarez & Hall,
2006; Alvarez et al., 2008). As participation is required for effective rep-
resentation, easily usable voting modes should, in theory, ensure a bet-
ter overlap between the elected representatives and society. However,
technology can also present additional barriers to thealready disadvan-
taged, in effect nullifying its theoretical promise (Berinsky, 2005; Norris,
2003). It also needs to be acknowledged that e-voting will not address
underlining reasons for abstention, such as political disillusionment or
a lack of political interest. This does not mean that internet voting is a
“technological fix”to an issue that cannot be fixed using technology.
E-voting can impact turnout among those who have accessibility prob-
lems, such as the disabled and elderly. Moreover, it can also mobilize
those who do not have clear mobility problems, but who simply do
not vote due to inconveniences related to conventional voting. Thus,
e-voting is first and foremost a convenient voting method and therefore
should appeal to those parts of theelectorate who have abstained due to
paper voting being too cumbersome.
The actual practice of remote e-voting has been implemented in a
limited number of countries. Exactly how remote e-voting influences
voting behavior and parties' strategies is unknown. Studies on technol-
ogy usage show that the most likely users and beneficiaries are young,
technology savvy, well-resourced, and connected people (Schlozman,
Verba, & Brady, 2010; van Dijk, 2000, 2005). There is clear evidence
that the same applies to the early adopters of e-voting (Alvarez, Hall,
& Trechsel, 2009; Trechsel & Vassil, 2011). However, what we do not
know is whether e-voting has the potential to diffuse beyond this sub-
population to a broader and less homogenous group of voters, or
whether it remains a tool for those with skills and resources. As diffu-
sion is the prerequisite of e-voting having a large impact upon turnout,
discussions about how and why new modes of voting might improve
participation or representation, require empirical evidence of the
Government Information Quarterly xxx (2016) xxx–xxx
☆This research was supported by Estonian Research Council grant nr. PUT523.
⁎Corresponding author at: Institute of Government and Politics, University of Tartu,
Lossi 36, 51003 Tartu, Estonia.
E-mail address: kristjan.vassil@ut.ee (K. Vassil).
1
We use the terms e-voting, remote internet voting, and internet voting interchange-
ably throughout this paper to describe online voting using a remote computer and digital
identification, i.e. voting without visiting a polling station.
GOVINF-01181; No. of pages: 7; 4C:
http://dx.doi.org/10.1016/j.giq.2016.06.007
0740-624X/© 2016 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Contents lists available at ScienceDirect
Government Information Quarterly
journal homepage: www.elsevier.com/locate/govinf
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
conditions and patterns by which new technologies are adopted over
time. If the rate of adoption of a new voting technology is slow and its
diffusion limited to specific subpopulations of the electorate, it is unlike-
ly that e-voting will have a positive impact upon voter turnoutand qual-
ity of representation.
This paper addresses precisely the question: Who are the e-voters
and has their profile changed over time? We used unique cross-
sectional survey data from all eight of the legally binding e-enabled
electionsin Estonia between 2005 and 2015. Our goal was to determine
whether the technology has diffused among the voter population, or
whether it remains a convenient technical solution for a group of people
already engaged in politics and who face limited barriers to participa-
tion in the first place.
1.1. E-voting in Estonia
Since 2005, Estonia has had a total of eight e-enabled elections
where eligible voters could cast binding ballots over the internet. Inter-
net voting has been used for local, national and European elections. The
number of e-voters in the first e-enabled election was only 9317 (Fig. 1).
However, the number increased in each succeeding election, reaching
176,491 in the 2015 national elections. In relative terms, the share of in-
ternet votes of total votes grew from a mere 2% in 2005 to N30% in 2014
and 2015.
A prerequisite for casting an electronic vote is a credit card sized
electronic ID-card,
2
which are compulsory for all Estonian residents.
Using digital identification, voters can use their personal computers
when connected to the internet and equipped with a smart card reader,
to cast an electronic vote (Alvarez et al., 2009). E-voting is available dur-
ing the advanced voting period via a website hosted by the EstonianNa-
tional Electoral Committee (2005–2011). E-voting itself involves three
steps; first, the user opens the website and with their ID-card and first
PIN-code to identify themselves, enters the system; second, after the
system has verified the identity of the voter, it displays the list of candi-
dates by party in the voter's respective district; third, by clicking on a
candidate's name and then entering their second PIN-code, the voter
casts their vote.
3
The first five elections were reasonably similar for the user-end, with
the only marked difference being the length of period during which e-
voting was available: three days in 2005 and 2007; and 7 days in
2009, 2011 and 2013. From 2009, e-voters needed to download a voting
program instead of voting via the web-embedded application. In 2013, a
vote verification feature was introduced to the e-voting system that
allowed voters to verify—using a smartphone or tablet—whether their
electronic vote was received as cast. Other than these differences, the
eight e-enabled elections were reasonably similar, providing a valid
point of comparison of the related dynamics in user behavior.
On the technical side, e-voting requires internet access and a mini-
mum level of computer literacy, both of which are not universal in
Estonia. However, the act of e-voting is no more difficult than other on-
line activities, such as banking or shopping.
2. Measuring diffusion
Theories on the diffusion of technological innovations provide a
foundation for measuring and explaining the potential spread of e-
voting in a society. The classical accounts of the diffusion of innovations
provided by Ryan and Gross (1943) and Rogers (2003[1962]) have
stood the test of time, being used over the years to explain a wide vari-
ety of phenomena, rangingfrom the spread of agricultural practices (e.g.
Fliegel, 1993) to political reforms and policies (e.g. Starr, 1991; Jahn,
2006), medical practices (e.g. Greenhalgh, Robert, Macfarlane, Bate, &
Kyriakidou, 2004), management (e.g. Abrahamson, 1991), and most
crucially, technological applications in very different fields (e.g.
MacVaugh & Schiavone, 2010). Rogers' (2003) account sees the diffu-
sion of technology as a sequence of steps in an innovation decision pro-
cess. This process includes gaining knowledge of the technology, being
convinced of its usefulness, and ultimately, deciding to implement it.
Adoption occurs if expectations are positively confirmed by experience.
Once a distinct subgroup has reached the adoption stage and built up a
critical mass of users, subsequent diffusion is reminiscent of a bank-run,
where the number of people adoptingit is partly a function of the num-
ber of prior adopters (Rogers, 2003: 206). This sequence has been dem-
onstrated to apply to both collective and individual actors (see Wejnert,
2002).
The crucial aspect of using Rogers' account to explain e-voting
regards the changing profile of adopters of technology at different
stages of the process. The first adopters tend to be a small number of
well-informed, innovative risk-takers (Rogers, 2003: 263). The second-
ary and tertiary adopters should more closely resemble thegeneral pop-
ulation,and the unique characteristics associated with thefirst adopters
should continually become less prominent. Eventually, even technolog-
ical laggards might be motivated to adopt the technology, as the relative
gains outweigh the costs of adopting (Rogers, 2003:263–265).
As with every new internet technology, adoption requires a certain
level of digital literacy, which is not always evenly distributed across so-
cial groups. This suggests that internet voting is most likely to appealto
Fig. 1. Dynamics of e-voting in Estonia, 2005–2015.
2
Since 2011voters can alsouse a smartphone-based mobileID (using a special SIM card
and PIN-codes) to authenticate themselves to the e-voting system. The ID card, however,
is the more widely used identification method.
3
For further details on the process of e-voting, see: http://vvk.ee/voting-methods-in-
estonia/engindex/; Estonian National Electoral Committee (2005);OSCE/ODIHR (2007,
2011);Vassil and Weber (2011);Trechsel and Vassil (2011).
2K. Vassil et al. / Government Information Quarterly xxx (2016) xxx–xxx
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
those with a good command of modern technologies. It is precisely this
mechanism that has fueled accounts claiming that e-democracy in gen-
eral reflects underlining societal divisions and even augments these by
further marginalizing the marginalized and connecting the connected
(Alvarez & Nagler, 2000; Van Dijk, 2000, 2005; Margolis and Resnick,
2000; Putnam, 2001; Wilhelm, 2000). New voting technologies might
therefore also only diffuse amonga distinct non-random subpopulation
of voters distinguished by higher socio-economic status, but not
beyond.
Eventually it remains an empirical question that can be tested using
Roger's theoretical framework. The literature on diffusion suggested
two different expectations for our subsequent analyses. First, we ex-
pected a gradual dispersion of e-voting usage, driven by early adopters,
who would be distinguishable by their socio-demographic profile:
younger, better-educated, and comparatively economically well-off in-
dividuals. Presumably, they should also be more technology savvy and
trust the technology more. As the use of technology widens with each
additional e-enabled election, a gradual increase in the number of e-
voters should be observable. Over time, these new e-voters should re-
sult in the total group of e-voters becoming less distinct from paper bal-
lot voters, both in terms of socio-demographics, as well as behavioral
and attitudinal characteristics. If diffusion of a technology is actually tak-
ing place, the population of e-voters will become more heterogeneous
over time. Thus, we expected that the characteristics associated with
the likelihood of e-voting during the first e-enabled elections would
subsequently become less pronounced and lose their explanatory
power. Based on this, we formulated a first hypothesis to identify the
presence of diffusion:
H1. Characteristics that explained e-voting during the first e-enabled
elections will lose their predictive power over time, suggesting that
usage of internet voting has diffused among the electorate.
A contrary expectation, however, follows from the fact that a new
voting technology can present a barrier to potential users. Thus, al-
though a technology mightspread according to the initial sequence sug-
gested by Rogers and user numbers initially rise, a barrier may prevent
larger segments of potential users from adopting the new technology.
We posited that the likely barriers would be an insufficient level of dig-
ital literacy, a lack of trust in thee-voting system, andage related factors.
If such barriers do indeed exist e-voters will remain a distinct subgroup
of total voters, because the growth of e-voters will plateau owing to the
technology failing to bridge the gap between more and lesstechnology-
savvy voters. It would also mean that first-time e-voters should remain
clearly distinguishable frompaper ballot voters, irrespective of time or a
growth in e-voters. Thus, our competing hypothesis was:
H2. Characteristics that explained e-voting during the first e-enabled
elections would retain their predictive power over time, suggesting
that usage of internet voting has not diffused among the electorate.
3. Data, variables and model specification
In order to investigate whether diffusion of e-voting occurred, we
used a unique series of individual-level surveys, whereby data were col-
lected after each of the eight e-enabled elections in Estonia. We chose to
work with different types of elections in one temporal sequence. Thus,
we deliberately ignored the possibility that different election types
may mobilize different voter types, vary in saliency, and influence over-
all turnout. However, as we were interested in measuring diffusion as a
function of recurring experiences with e-voting, we compared elections
over time. We argue that diffusion should be observable irrespective of
whether elections are treated as one temporal sequence or grouped ac-
cording to type or cycle, as time is a proxy for cycle and vice versa. We
return to the empirical implications of this analytical choice in the dis-
cussion section.
The first five surveys consisted of quota sampling (with the sample
containing an almost equal share of internet voters, ballot-paper voters,
and non-voters), to ensure a sufficient number of e-voters for analysis;
all surveys had a sample size of 1000 respondents and used the CATI
method. The three subsequent surveys consisted of stratified random
sampling, because the number of e-voters in the overall voting popula-
tion had become sufficiently large for analysis (Fig. 1). They also had a
sample size of 1000 respondents and used the CAPI method. The sur-
veys had response rates of 62.3%, 61.7% and 60.0% in 2013, 2014 and
2015 respectively.
4
All were post-election surveys conducted during
the three week period post-election day and are representative of the
voting eligible population. Table 1 shows the sample composition ac-
cordingtovotingmode.
3.1. Variable selection
Our dependent variable consisted of a dichotomy of factors that dis-
tinguished e-voters from ballot-paper voters. However, we must be ex-
plicit about our choice of response category of interest, because
comparing e-voters to regular voters over time across elections inevita-
bly introduces noise. Such noise was owing to the fact that in later elec-
tions, the population of e-voters contained first-time e-voters (early
adopters in the first elections) and those who had voted online in mul-
tiple elections. As the motivations and characteristics of these groups
may differ considerably (Rogers, 2003), we preferred—in-line with our
theoretical argument—to decompose the response category to distin-
guish between first time and recurring e-voters in each election. We
chose to compare ballot-paper voters (coded as 0) only to first time e-
voters (coded 1) in each e-enabled election. Effectively, our
operationalization yielded a response category that captured first time
e-voters that were: early adopters in the first elections; the early major-
ity and majority voters in subsequent elections; and late majority and
laggards in the most recent elections. Any change in the profile of first
time e-voters would thus reveal whether diffusion of the new voting
technology occurred.
Prior studies in diverse settings on diffusion patterns have demon-
strated effects of socio-demographic and economic factors, including
ethnicity, in predicting diffusion among actors (e.g. Berry & Berry,
1990; Hedström, 1994; Tolnay & Glynn, 1994). More importantly,
prior studies on e-voting in Estonia have shown that age, education,
trust in the e-voting system, and first-language, should be particularly
strong predictors of whether someone e-votes (Trechsel & Vassil,
2011). The latter is true due to the fact that Estonia is a multilingual so-
ciety with about one third of the population Russian-speaking. As the
system of e-voting is only offered in the Estonian language, it could
limit its use among the Russian-speaking minority (Trechsel & Vassil,
2010). In addition, economic well-beingand literacy with new technol-
ogies, have been shown to be systematically correlated with the likeli-
hood of internet voting (Alvarez et al., 2009; Trechsel & Vassil, 2011;
Vassil & Weber, 2011). Finally, people's political self-positioning (e.g.
left or right) may provide a useful control for our models, because ag-
gregate election results consistently show that liberal parties gain
more e-votes than those on the ideological left.
5
Followingthese empirical accounts, we used the following indepen-
dent variables for our analysis: both age in years and age squared (to
allow for non-linear effects); education (using a dummy variable for
higher education, with secondary and elementary education as the ref-
erence); gender (male = 1; female = 0); ethnicity (Estonian as home
language = 1; Russian as home language = 0); income (measured by
decile); computer literacy (using a dummy for good skills level, with av-
erage and basic levels as reference); trust toward the e-voting system
4
Comparable responserates for earlier surveys cannotbe computed because they used
quota-sampling and CATI methods.
5
Further information is available on the National Electoral Committee website: www.
vvk.ee
3K. Vassil et al. / Government Information Quarterly xxx (2016) xxx–xxx
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
(trust = 1; no trust= 0)
6
; and political left-right self-positioning (using
a10pointscale).
3.2. Model specification
In order to evaluate whether a diffusion mechanism was at play, we
estimated a separate logistic regression model for each of the eight elec-
tions, and compared the coefficients for each independent variable and
model fit over time. If the diffusion process happened as stated in H1,
we should observe a gradual reduction in model fitand disappearing co-
variate effects in terms of magnitude and statistical significance over
time, which would confirm that the sample of first time e-voters had in-
deed become more heterogeneous. By contrast, if the effects retained
their power and significance, it would suggest that the sample of first
time e-voters in later elections were as distinct in terms of their charac-
teristics as during the first few elections, thus confirming H2. The model
took the following generic form (Eq. (1)):
ln Pr evote ¼1ðÞ
1−Pr evote ¼1ðÞ
¼β0þβiXið1Þ
where X
i
is the vector of the listed independent variables for individual
i; the model was estimated separately for each of the eight elections.
Our particular interest was in how β
i
changes over time. We were not
interested in any particular covariate effects per se, but whether and
how the effects of independent variables changed over time.
For improved interpretation, we converted the logistic regression
coefficients into average marginal effects, which show the average of
the variation induced in the probability of interest by a marginal change
in an independent variable for each individual in the sample (Baum,
2006). An average marginal effect is interpreted as an effect of one-
unit change of the independent variable on the change of probability
of interest. The appeal of average marginal effects in our analytical set-
ting, was that they are less affected by unobserved heterogeneity that
is unrelated to the independent variables in the model, and can thus
be compared across models, groups, samples, or years (Mood, 2010:
78). Missing values were multiply imputed for all datasets.
7
4. Findings
The findings from all eight regression models, with relevant fitdiag-
nostics, are presented in Table 2. First and foremost, we saw that associ-
ations between e-voting and age, ethnicity, computer literacy, and trust
toward e-voting, weakened substantially with time. More specifically,
Fig. 2 displays the non-linear impact of age-squared on the likelihood
of internet voting. The findings clearly show that the effect of age
flattens gradually over time. Only for the first three elections did age
have the expected inverted U-shape effect on the probability of voting
online. The likelihood of internet voting was initially highest among
40 to 50 year olds, and lowest for the younger and older. However,
this once strong relationship started to gradually disappear after the
third e-enabled election in 2009, flattening and losing its predictive
power entirely by the fourth election. We assumed from this that over
time, the likelihood of e-voting becomes almost equally probable for
all age groups.
Similarly, and in accordance with previous studies, we found that
the Estonian language was an important predictor of e-voting during
the first e-enabled elections. Between the 2005 and 2009 elections, eth-
nic Estonians were approximately 26–38 percentage points more likely
to vote online compared to non-Estonians. However, this difference was
lower by more than half by 2011, and had completely vanished by the
2014 elections. We infer from this that ethnicity has lost its explanatory
power over internet voting, and aswith the effect of age, most of thisex-
planatory power seemed to disappeared after the third election, render-
ing the once significant disparity between Estonian and Russian-
speakers, with respect to internet voting, negligible (Fig. 3).
Computer literacy followed the same trail: it shows strong associa-
tion with the likelihood of internet voting, with those with high PC-
skills approximately 17 percentage points more likely to vote online
than those with average and poor skills. This is not surprising, as the
general setup of the e-voting system requires several interactions with
a computer, relevantperipherals, andthe ID-card. The effect was consis-
tent over the first two elections, after which it became insignificant, pro-
viding evidence that voters may have become more familiar with the
system and learned to useit (Fig. 3). However, we did find it surprising-
ly that the effect of PC-literacy reappeared during the last election of
2015, with a small effect at the 0.01 level. We believe this was partly
due to differences in the electorates between different election types.
Trust toward the system of e-voting has been shown to be one of the
strongest predictors of e-voting (Trechsel & Vassil, 2011). We observed
the same, but only for the first e-enabled elections. In particular, we saw
that thosewho trusted the system of e-voting were about 49 percentage
points more likely to vote online than those who found it less trustwor-
thy. The effect hovered between 35 and 70 percentage points in the first
four e-enabled elections and then significantly lost its explanatory
power. Unlike previous variables, trust decreased substantially in effect
size, but retained its statistical significance (Fig. 3).
Regarding education, gender, income, and left-right self-position, we
found no substantially strong relationships (Table 2). A higher educa-
tion appeared to be weakly but positively associated with internet vot-
ing, though its effect was not consistent. The same was the case for
gender and income. As for left-right self-positioning, we found it partic-
ularly reassuring that nowhere in the data did we find a statistically sig-
nificant and sizable effect to provide evidence that internet voting is
unequally likely for those on the left compared to the right of the polit-
ical spectrum.
Taken together, we found that multiple socio-demographic and atti-
tudinal variables were strongly associated with the likelihood of
Table 1
Voter types in the sample.
Voter type 2005 local 2007 national 2009 EP 2009 local 2011 national 2013 local 2014 EP 2015 national Total
Normal voter 318 450 448 403 480 560 477 613 3749
(%) (33.9) (45.8) (44.93) (40.3) (47.7) (53.8) (48.0) (61.4) (47.1)
1st time e-voter 315 309 264 108 139 65 48 62 1310
(%) (33.6) (31.5) (26.5) (10.8) (13.8) (6.2) (4.8) (6.2) (16.5)
Recurring e-voter 0 60 85 142 72 106 90 134 689
(%) (0.0) (6.1) (8.5) (14.2) (7.2) (10.2) (9.1) (13.4) (8.7)
Non-voter 306 163 200 347 316 311 379 190 2212
(%) (32.6) (16.6) (20.1) (34.7) (31.4) (29.9) (38.1) (19.0) (27.8)
Total 939 982 993 1000 1007 1042 994 999 7956
(%) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0) (100.0)
Column percentages might not sum to 100% due to rounding.
6
The dummy was created from a four item Likert scale (2005–2011) or a 10 category
ordinal item scale (201 3–2015). Both variables were split at the middle, with people
who trusted or tended to trust coded as 1, and those with no trust or a tendency not to
trust coded as 0.
7
We used STATA's margins package to compute averagemarginal effects, and their mi-
impute package to multiply impute missing values.
4K. Vassil et al. / Government Information Quarterly xxx (2016) xxx–xxx
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
internet voting in the first three e-enabled elections, but they started to
gradually lose their explanatory power over time, becoming less and
less relevant. Thus, we took this as thefirst evidence of diffusion having
taken place and that the data supports H1.
However, as the diffusion process should render the first-time e-
voter population more heterogeneous overtime, not only the individual
effects of covariates should weaken, but also the fit of the model. Con-
versely, ifno diffusion has taken place, model fit should remain relative-
ly immune to change over time. Fig. 4 presents various model fit
parameters over time. First, there was a significant drop in model per-
formance, as measured by the drop in pseudo-R
2
,reducingfroma
healthy 0.52 in 2005 to 0.18 in 2015. This clearly points to increasing
heterogeneity among first time e-voters, providing evidence that diffu-
sion was indeed taking place. More importantly, there was a radical
drop in the model's sensitivity, i.e. its ability to correctly classify first
time e-voters (the true positives). Fig. 4 shows that it substantially
dropped after the third election. Moreover, from the 2013 election on-
ward the model failed to classify e-voters. In sum, the diminishing effect
of covariates in explaining first time e-voters, as well as a lower overall
model performance, with time, allowed us to refute H2 and support the
thesis of ongoing diffusion based upon the increasing heterogeneity of
e-voters. Notice that the general classification accuracy of the model
remained high throughout the years; this was because it initially cor-
rectly predicted e-voters and ballot-paper voters, and continued to
Table 2
Predicting first time e-voting (base: only paper ballot voters).
2005 local 2007 national 2009 EP 2009 local 2011 national 2013 local 2014 EP 2015 national
Age 1.72⁎⁎ 1.77⁎⁎ 1.97⁎⁎⁎ 1.28 0.54 0.24 0.23 0.57
(0.58) (0.57) (0.58) (0.69) (0.52) (0.42) (0.37) (0.39)
Age
2
−0.02⁎⁎
−0.02⁎⁎
−0.02⁎⁎⁎
−0.02⁎
−0.01 0.00 0.00 −0.01
(0.01) (0.01) (0.01) (0.01) (0.01) (0.00) (0.00) (0.00)
Estonian language 26.05⁎⁎⁎ 38.40⁎⁎⁎ 29.86⁎⁎⁎ 11.86⁎14.66⁎⁎ 2.31⁎1.68 5.92
(5.23) (7.08) (7.50) (4.92) (5.57) (3.43) (2.61) (3.60)
PC literacy: good 17.35⁎⁎⁎ 16.60⁎⁎⁎ 3.75 0.86 5.40 5.10 3.57 9.88⁎⁎
(3.56) (3.41) (3.68) (3.75) (3.34) (2.82) (2.72) (3.06)
Trust e-voting 49.25⁎⁎⁎ 36.23⁎⁎⁎ 69.56⁎⁎⁎ 35.30⁎⁎⁎ 25.25⁎⁎⁎ 14.80⁎⁎⁎ 9.25⁎⁎ 13.50⁎⁎⁎
(4.00) (5.35) (11.51) (7.67) (3.73) (3.63) (3.28) (3.38)
Left-right self-position −0.17 0.13 −0.04 −0.26 0.54 0.42 0.32 0.19
(0.72) (0.78) (0.66) (0.90) (0.70) (0.53) (0.48) (0.63)
Education: higher 5.24 8.81⁎⁎ 9.01⁎⁎ 5.06 8.98⁎⁎ 5.50⁎1.50 4.34
(3.33) (3.29) (3.26) (3.52) (3.03) (2.48) (2.28) (2.63)
Male −1.04 2.05 0.47 0.09 6.38⁎1.81 −1.56 5.11⁎
(3.13) (3.12) (3.18) (3.35) (2.91) (2.36) (2.10) (2.41)
Income decile 0.31 1.79⁎⁎ 0.85 0.97 0.97⁎
−0.10 0.23 −0.58
(0.58) (0.59) (0.60) (0.63) (0.48) (0.46) (0.42) (0.53)
Constant −7.49⁎⁎⁎
−7.54⁎⁎⁎
−9.09⁎⁎⁎
−5.90⁎⁎⁎
−5.45⁎⁎⁎
−4.70⁎⁎⁎
−4.71⁎⁎⁎
−6.54⁎⁎⁎
(1.12) (0.96) (1.25) (1.32) (1.11) (1.38) (1.50) (1.39)
Observations 633 759 712 511 619 625 646 617
Nagelkerke Pseudo R
2
0.52 0.36 0.41 0.34 0.37 0.24 0.16 0.18
Correctly classified 0.78 0.74 0.72 0.81 0.81 0.90 0.93 0.90
Sensitivity 0.90 0.67 0.68 0.24 0.40 0 0 0
Specificity 0.66 0.79 0.75 0.96 0.76 1 1 1
Average marginal effects as percentages. Standard errors in parentheses.
⁎⁎⁎ pb0.001.
⁎⁎ pb0.01.
⁎pb0.05.
Fig. 2. Effect of age on the likelihood of e-voting.
5K. Vassil et al. / Government Information Quarterly xxx (2016) xxx–xxx
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
accurately predict (only) on-paper voters, this further supporting the
evidence on diffusion.
5. Discussion
E-voting has become a widely used voting mode in Estonia. Howev-
er, the aggregate number of e-voters might disguise a situation where
the technology has not diffused across societal boundaries, but instead
is only being increasingly used by a distinct subpopulation of well-
resourced, technologically savvy voters. Real diffusion over time
would mean that voters from a broad cross-section of the population,
regardless of their social status or level of resources, use e-voting.
Followingthe expectations derived from Rogers' diffusion of innova-
tion, we examined the profile of first time e-voters over the course of
eight e-enable elections (a period of ten years, 2005–2015), to deter-
mine to what degree usage of this new voting technology has been
adopted by the wider voter population. The aggregate number of e-
voters increased sizably over time, with every third vote being cast on-
line during the past two elections. At the level of the individual, we
found that the characteristics of first time e-voters became more similar
to the characteristics of traditional paper ballot voters over time. E-
voters used to be Estonian-speakers from a distinct age group, who
have good computer literacy and trust in the system of internet voting.
However, this was only the case for the first three elections in which e-
voting was used. From the fourth election onward, we consistently saw
that these characteristics were only weakly, if at all, associated with the
choice to vote online. As a result, our model's ability to predict and cor-
rectly classify first-time e-voters based only on socio-demographic and
attitudinal data, becomes increasingly limited. Our results show that e-
voting has diffused among the overall voter population, and not just
remained an activity of the privileged few. Importantly, we found that
the process of diffusion did not occur immediately, but was shown via
a plateau effect, by which diffusion became visible only after the first
three elections.
We focused on elections in one temporal sequence, irrespective of
their type. Alternatively, one could focus on electoral cycles by separat-
ing them by type. We chose not to do so, because we modelled the dif-
fusion of e-voting in one country, where eligible voters substantially
overlap from one election to another. However, if one did focus on elec-
toral cycles, the evidence provided in this paper should also point to-
ward the diffusion. Namely, some of the effects already started to
disappear after the second, and others after the third, electoral cycle.
As a result of our findings, we draw two main conclusions. First,
technology has the potential to bridge societal divisions and ease polit-
ical participation, not only for the already connected and resourceful,
but also for the less privileged, who have fewer resources and remain
at the periphery of using modern technologies. Similarly, internet vot-
ing may appeal to those finding conventional voting to cumbersome.
As a more convenient mode of participation e-voting may also have
the potential to ease participation for those who are connected and en-
gaged but may still abstain due to the inconveniences related to on-
Fig. 3. Impact of computer literacy, ethnicity, trust and ideological auto-position on the
likelihood of internet voting. Whiskers represent 95% confidence intervals. (L –local, N –
national, EP –European Parliament election.).
Fig. 4. Model fitdiagnostics(L–local, N –national, EP –European Parliament election).
6K. Vassil et al. / Government Information Quarterly xxx (2016) xxx–xxx
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
paper voting. The experience of e-voting usage in Estonia shows that
technology should not be considered as a hurdle, but as an enabler for
political participation. The caveat is that technology only provides an ef-
ficient mode for participation; structural hurdles that inhibit participa-
tion in general, regardless of the mode voting, will most likely stay
unaffected. However, what we have demonstrated in this paper is that
technology itself does not seem to exclude anybody, as the skeptics
have suggested.
The second conclusion is that the potential enabling effects did not
surface immediately in the electoral realm after the introduction of
the new voting technology, but required a period of at least three elec-
tions to appear. Adoption amonga select subgroup can happen immedi-
ately, but this is limited to people who already have the resources and
skills to use new technologies. A wider public benefit can only be real-
ized once usage has diffused and this does take time. Policymakers are
well advised not to expect immediate results following the introduction
of new voting technologies, but should recognize that different sub-
groups of the electorate adopt and use new technologies at different
rates. From a positive perspective, our evidence showed the process to
be fairly quick; characteristics that used to predict internet voting
started to lose their predictive power after only three separate elections
within four years. What seemed to matter most was not time as such,
but the frequency of being exposed to the possibility of casting their
vote over the internet.
Regarding the generalizability of the three-election argument to
other contexts, we point out that Estonia was an early adopter of inter-
net voting. Ten years ago internet penetration, broadband communica-
tions, and the use of social media, was markedly lower than today. In
countries where such factors are higher, therate of diffusion of internet
voting may be substantially accelerated following its introduced.
Taken together, we found evidence that e-voting has diffused among
a wide and heterogeneous group of Estonian voters, and has not just be-
come an exclusive form of participation for a privileged few. That
Estonian e-voters are a widespread and heterogeneous group was con-
vincingly shown by the model fit, which went from excellent to ex-
tremely poor in just over the course of eight elections. Therefore, we
are confident that new voting technologies are not necessarily exclu-
sive, as early studies on e-voting have suggested, but are inclusive for
a wide range of voter types.
References
Abrahamson, E. (1991).Managerial fads and fashions:The diffusion and rejection of inno-
vations. Academy of Management Review,16(3), 586–612.
Alvarez, R. M., & Hall, T. E. (2004). Point, click, and cote: The futur e of internet voting.
Brookings Institution Press.
Alvarez, R. M., & Hall, T. E. (2006). Controlling democracy: The principal–agent problems
in election administration. Policy Studies Journal,34(4), 491–510.
Alvarez, R. M., & Nagler, J. (2000). The likely consequences of internet voting for political
representation. Loyola of Los Angeles Law Review,34,1115–1154.
Alvarez, R. M., Hall, T. E., & Llewellyn, M. H. (2008). Are Americans confidenttheir ballots
are counted? The Journal of Politics,70(3), 754–766.
Alvarez, R. M., Hall, T. E.,& Trechsel, A. H. (2009). Internet voting in comparative perspec-
tive: The case of Estonia. Political Science & Politics,42(3), 497–505.
Baum, C. F. (2006). An Introduction to Modern Econometrics Using Stata. Stata Press.
Berinsky,A. J. (2005). The perverse consequences of electoral reform in the United States.
American Politics Research,33(4), 471–491.
Berry, Frances S]–>F. S., & Berry, W . D. (1990). State lotte ry adoptions as policy i nnova-
tions: An event history an alysis. The American Political Science Review,84(2),
395–415.
van Dijk, J. (2005). The deepening divide: Inequality inthe information society. Sage Publica-
tions, Inc.
van Dijk, J. (2000). Widening information gapsand policies of prevention. In K. L. Hacker,
& J. van Dijk (Eds.), Digital democracy: Issues of theory and practice (pp. 166–183).
London: Sage.
Estonian National Electoral Committee (2005). Estonian I-voting system: General de-
scription. http://www.vvk.ee/elektr/docs/Yldkirjeldus-eng.pdf (February 12, 2014)
Fliegel, F. C. (1993). Diffusion research in rural sociology: The record for the future. (West-
port, CT).
Gibson, R. (2001). Electionsonline: Assessing internet voting in light of the Arizona Dem-
ocratic Primary.Political Science Quarterly,116(4), 561–583.
Greenhalgh, T., Robert, G., Macfarlane, F., Bate, P., & Kyriakidou,O. (2004). Diffusion of in-
novations in ser vice organizations: Systematic review an d recommendations.
Milbank Quarterly,82(4), 581–629.
Hedström, P. (1994). Contagious collectivities: On the spatial diffusion of Swedish trade
unions, 1890–1940. American Journal of Sociology,99(5), 1157–1179.
Jahn, D. (2006). Globalization as 'Galton's Problem':The missing link in the analysis of dif-
fusion patterns in welfare state devel opment. International Organization,60(2),
401–431.
MacVaugh,J., & Schiavone, Francesco]–>F. (2010). Limitsto the diffusion of innovation: A
literature review and integrative model. European Journal of Innovation Management,
13(2), 197–221.
Margolis, M., & Resnick, D. (2000). Politics as usual: The cyberspace revolution. Inc: Sage
Publications.
Mood, C. (2010). Logistic regression: Why we cannot do what we think we can do, and
what we can do about it. European Sociological Review,26(1), 67–82.
Norris, P. (2001). Digital divide: Civic engagement, information poverty, and the internet
worldwide. Cambridge University Press.
Norris, P. (2003). Will new technology boost turnout? Evaluating experiments in e-voting
v. All-postal voting facilities in UK local elections. Paper for the British Study Group
Seminar Friday 31st October 2-4 pm, Minda de Gunzberg Center for European studies.
Harvard University.
OSCE/ODIHR (2007). Election assessment mission report in the 2007 parliamentary elec-
tions in Estonia . http://www.osce.org/odihr/elections/estonia/25925 (December 20,
2013)
OSCE/ODIHR (2011). Election assessment mission report in the 2011 parliamentary elec-
tions in Estonia. http://www.osce.org/odihr/77557 (December 20, 2013)
Putnam,R.(2001).Bowling alone: The collapse and revival of American community. Simon
and Schuster.
Rogers, M. E. (2003). Diffusion of innovations (5th ed.). New York: Free Press (1962).
Ryan, B., & Gross, N. C. (1943). The diffusion of hybrid seed corn in two Iowa communi-
ties. Rural Sociology,8,15–24.
Schlozman, K. L., Verba, S., & Brady, H. E. (2010). Weapon of the strong? Participatory in-
equality and the internet. Perspectives on Politics,8(2), 487–509.
Starr, H. (1991). Democratic dominoes. Diffusion approaches to the spread of democracy
in the international system. Journal of Conflict Resolution,35(2), 356–381.
Tolnay, S. E., & Glynn, P. J. (1994). The persistence of high fertility in the American south
on the Eve of the baby boom. Demography,31(4), 615–631.
Trechsel, A. H., & Vassil, K. (2010). Internet voting in Estonia - a comparative analysis of
four elections since 2005. Report for the Council of Europe. Council of Europe.
Trechsel, H. A., & Vassil, K. (2011). Internet voting in Estonia. A comparative analysis of
five elections since 2005. Report to the Estonian Na tional Electoral Comm ittee.
http://www.vvk.ee/valijale/e-haaletamine/raportid/ (January 12, 2014)
Vassil, K., & Weber, T. (2011). A bottleneck model of e-voting: Why technology fails to
boost turnout. New Media & Society,13(8), 1336–1354.
Wejnert, B. (2002). Integrating models of diffusion of innovations: A conceptual frame-
work. Annual Review of Sociology,28,297–326.
Wilhelm, A. (2000). Democracy in the digital age: Challenges to political life in cyberspace.
Routledge.
Kristjan Vassil is a senior research fellow at the Johan Skytte Institute of Political Studies,
University of Tartu. He defended his doctoral degree in the European University Institute
in January 2012 and his thesis focused on the effect of information and communication
technologies on voting behavior.
Mihkel Solvak is a senior research fellow at the Johan Skytte Institute of Political Studies,
University of Tar tu. He defended his doct oral dissertation in the University o f Tartu,
Estonia.His dissertationfocused on parliamentary roll-callbehavior. Mihkelhas published
a wide rangeof academic peer-reviewedjournal focusingon voting behavior,politicalpar-
ticipation and technology.
Alexander H. Trechsel (PhD), a professor at the European University Institute, Florence,
Italy; faculty fellow at Berkman Center for Internet & Society, Harvard University, Cam-
bridge, MA, USA.
R. MichaelAlvarez (PhD), a professor of political scienceat California Institute of Technol-
ogy, Pasadena, California, US. For more than a decade, R. Michael Alvarez has been one of
the most prominent academic researchers studying electiontechnologies, especially elec-
tronic and Internet voting.
7K. Vassil et al. / Government Information Quarterly xxx (2016) xxx–xxx
Please cite this article as: Vassil, K., et al., The diffusion of internet voting. Usage patterns of internet voting in Estonia between 2005 and 2015,
Government Information Quarterly (2016), http://dx.doi.org/10.1016/j.giq.2016.06.007
