A Bottleneck Model of E-voting: Why Technology Fails to Boost Turnout

EUROPEAN UNIVERSITY INSTITUTE Department of Political and Social Sciences

A Bottleneck Model of E-voting. Why Technology Fails to Boost Turnout

Kristjan Vassil

Till Weber

Abstract: Recent years have seen an increasing interest in internet voting in theory and practice. According to its proponents, e-voting modernizes the electoral process and boosts turnout. Less optimistic scholars object that citizens remain largely unaffected by the new technology. This study aims to fill the gap between these two claims. We argue that e-voting has a high impact on those citizens who are unlikely to use it in the first place; conversely, the impact is low on the bulk of typical e-voters. We test this hypothesis with new survey data from the 2007 general election in Estonia, the first country to have nationwide and legally binding elections on the internet. In a two-step model of individual behavior, we predict both the usage of e-voting and its impact on electoral participation. Our findings identify variables that increase the impact of e-voting but simultaneously decrease the initial likelihood of usage. In particular, e-voting affects ‘peripheral’ citizens (in a demographic and political sense), but only few of these citizens vote on the internet. This bottleneck effect explains why e-voting has failed to boost aggregate turnout but also points to a role in reducing political inequality.

Paper presented at the Midwest Political Science Association Annual National Conference, Chicago April 2-5, 2009

Contact information:

[email protected], [email protected]

2

Introduction

During the long decline of voter turnout in modern democracies, the question of how to

motivate citizens to participate in elections has remained on the agenda of politics and

political science. One rather recent attempt to address the issue is internet voting, i.e. the

option to cast one’s vote over the internet in (otherwise) normal elections1. When internet

voting was developed, hopes for a boost in turnout were great. However, the first

experiences from Switzerland, the United Kingdom, the Netherlands and the U.S. did not

confirm the expected effect.

Scholars put the blame on the failure of internet voting applications to overcome

social divisions in conventional political participation. Usage of the new technology is

not equally distributed across the population. It seems that instead of mobilizing

disaffected or “peripheral” citizens, internet voting merely constitutes yet another channel

of influence for the politically engaged. Traditional patterns of inequality in political

participation seem to be reinforced, not transformed.

In this paper we further scrutinize the mobilization potential of internet voting

applications. We highlight an analytical distinction that seems crucial to us but has as yet

not attracted much attention in the literature: the mere usage of internet voting is different

from its impact in terms of mobilization. Whereas we do not doubt that internet voting is

mostly used by the politically engaged, we claim that the impact on participation is

highest among peripheral citizens. Thus, internet voting does possess transformative

potential; at the same time, however, this potential remains largely inactive due to a

typical “bottleneck” effect.

The basic mechanism of the bottleneck effect is simple: peripheral citizens are

unlikely to use internet voting, but those few who happen to do so are then exposed to

strong mobilizing forces. More subtly, the effect is based on different motivations that

lead to usage in the first place. Politically engaged citizens are generally experienced with

computers and the internet. They use internet voting as a means to cast their ballot, but

they are not overly interested in the application itself. Peripheral citizens are less

computer literate. For them the application itself has a fascinating aspect, whereas the act

1 Throughout the text we use the terms internet voting and e-voting as synonyms.

3

of voting is barely attractive in its own right. But by using the technology, peripheral

citizens are brought in contact with politics and may then experience a more radical

impact than engaged citizens who merely strive for convenience.

We test these propositions in the case of the 2007 general election in Estonia, the

first time that internet voting has been used in a nationwide parliamentary election. The

analysis follows a two-step model whose components are linked by the bottleneck

mechanism. In the first step we predict individual usage of internet voting (as distinct

from conventional voting and abstention) on the basis of demographic and attitudinal

variables. This model is tested by multinomial probit regression with data from a general

election survey of the Estonian population. In the second step we predict the impact of

internet voting from similar variables by interval regression. To test this model we

introduce a new survey of Estonian internet voters. Before we turn to empirical analysis,

however, we will explicate our theory in more detail.

The Bottleneck Model

In recent years the majority of scholars have become less optimistic about the internet’s

ability to promote political participation in general and voter turnout in particular.

Although the last U.S. primaries demonstrated major novelties in web-campaigning,

possibly contributing to differences in election outcomes, European e-democratic

experiments have remained rather modest. Opposing the excessive cyber-optimism from

the mid-nineties, the contemporary literature admits that in theory the internet may lower

the costs of electoral participation, strengthen democratic practices and include the

disengaged into civic life, but there seems to be little empirical support for these claims.

Internet applications have only weak impact on political participation and civic

engagement.

The standard explanation for this finding is offered by theories of digital divide in

general and political divide in particular: Online politics mirrors the patterns of inequality

experienced in conventional politics and even increases the gap between the engaged and

the disengaged (Alvarez & Nagler 2000; Wilhelm 2000; Putnam 2000; Margolis &

Resnick 2000; van Dijk 2000; van Dijk 2005). Disparities in access to the internet based

on income and education are still widespread. Online politics therefore tends to empower

4

the wealthy and well-educated and to further marginalize the underprivileged

(Mossberger, Tolbert & Stansbury 2003). The prime beneficiaries are elites with the

resources and motivation to take advantage of internet applications, whereas the costs

remain too high for less skilled citizens. The internet provides new opportunity structures

for the elite rather than mobilizing the disengaged periphery. In this sense, promoting

politics on the internet means preaching to the faithful.

Far from mobilizing the general public, the Internet may thereby function to increase

division between the actives and apathetics within societies. /---/ But as the media of

choice par excellence it is difficult to know how the Internet per se can ever reach the

civically disengaged (Norris 2001, 231).

Recently, however, scholars have raised some doubts about the internet’s inability

to reach the disengaged and bring them closer to politics. Based on studies of internet

voting and Voting Aid Applications (VAA) – the two most tangible forms of online

political participation – small but significant mobilization effects have been found. In

particular, the results reported by Alvarez, Hall and Trechsel (2008) show that roughly

one tenth of the internet voters in Estonia would not have turned out without the

possibility to vote online. A mobilization effect of about the same magnitude was found

by Boogers (2006): One tenth of the users of Stemwijzer (the Dutch VAA) reported an

increased motivation to cast their vote after obtaining the advice from the VAA.

Kleinnijenhuis and van Hoof (2008) in their study of the usage of several Dutch VAAs

observed that more people made a choice for a particular party after consulting the VAA.

Although limited in cross-sectional and longitudinal terms, this evidence points toward

some mobilization effects caused by VAA-usage and internet voting. An apparent

question follows from here: Who is being mobilized and for what reason?

If online politics has any effect on participation at all, it is likely to occur among

the converted citizens with particular attitudes and demographic characteristics: Young

individuals with higher income, educational attainment, sense of political efficacy and

positive attitudes toward politics are more likely to participate in online politics in

general (Mossberger, Tolbert & Stansbury 2003; Norris 2001) and in e-voting in

particular (Kersting & Baldersheim 2004; Alvarez & Nagler 2000; Solop 2001).

5

A number of studies have established that the usage of internet voting is

significantly skewed toward younger citizens. After all, it is the young who are exposed

to the new media to a far greater extent than the elderly, and it is self-evident that internet

voting is most conveniently accessible to those already familiar with new technologies.

These preconditions, combined with the fact that turnout has been generally low among

young citizens (Franklin 2004; Wattenberg 2008), raise expectations that precisely the

young will be mostly affected by internet voting (Kersting & Baldersheim 2004; Norris

2003; Alvarez, Hall & Trechsel 2008).

Considering voting behavior by age category, it becomes clear that above all younger

people participated by voting over the Internet. Based on this finding, one can

conclude that the introduction of voting by Internet seems to have significant impact

on the participation of younger voters in elections. The use of internet voting mobilizes

the generally underrepresented young persons, while it is more seldom used by older

voters (Trechsel et al. 2007).

Building on these findings, we would like to highlight a distinction that has as yet

not attracted much attention among scholars. The usage of internet voting often seems to

be insufficiently differentiated from its impact. We argue that without making a

conceptual distinction between the two, the analysis of internet voting may suffer from

some degree of logical imprecision with implications for empirical analysis and

theoretical interpretation. Namely, the act of using internet voting per se does not

necessarily imply an effect on an individual’s propensity to turn out and may therefore

not be an ideal indicator to measure mobilization. The proposition that the young are

more likely to engage in e-voting due to their digital affinity may well hold, but we see

no compelling reason for concomitant mobilization effects. It does not necessarily follow

from the literature on usage that internet voting mobilizes particularly the young and

affluent. Quite the contrary, we expect mobilization effects – if any – among the apathetic

periphery. In particular, the greatest impact on the propensity to turn out should appear

among those who are unlikely to use internet voting in the first place. Conversely, the

impact on individual turnout should be low among typical internet voters.

6

It is important to be aware of some ambiguity about the meaning of “turnout” in

this respect. Previous studies tended to equate turnout with usage and impact, both in

theory and measurement. We suggest that more precision is needed: Of course usage of

e-voting implies turnout, but the mere act of usage (or turnout for that matter) does not

imply impact. Impact may result from the mere availability of the option of e-voting

and/or from the experience of using the application. Although the typical e-voter may be

affected in both dimensions (availability and experience), we claim that impact decreases

with the likelihood of usage.

Why should we expect such a pattern? The following thought experiment is meant

to illustrate the difference between the usage of internet voting and its impact. Imagine

internet voter “A” who is fluent with computers, politically engaged, interested in

political news, discusses politics with his friends and family, and usually participates in

elections. In terms of technology he is an active user of the internet and related

applications. However, technology is so deeply rooted in his everyday life that he pays

minimum attention to it. Technology for him is a means rather than a goal.

Also imagine internet voter “B”. He is much less computer literate, politically

disengaged, rarely shows any interest in politics, and usually abstains in elections. In

terms of technology he is no active internet user. Moreover, by default he rarely thinks of

technology as an intrinsic part of his everyday life. However, when he happens to use it

he finds technology somehow fascinating. For him, the usage of technology per se

appears to be stimulating. For the same reason he finds the idea of casting his vote over

the internet attractive, but he is attracted by the technology and not by the desire to vote.

By using internet voting, both ideal-type voters – “A” and “B” – may be

positively affected in their propensity to turn out. If voter “A” finds that internet voting

works smoothly and is indeed a comfortable alternative to the polling booth, he may be

even more likely to turn out in the future. In this respect internet voting indeed reduces

electoral costs (cf. Norris 2003). And if voter “B”’s fascination with technology brings

him in contact with politics in the first place, he may develop some political interest and

turn out with a higher probability as well. The effect, however, is rather superficial for

Voter “A”, whereas Voter “B” may experience a more radical and potentially much

stronger impact. For voters of type “B” e-voting is a major innovation, but for voters of

7

type “A” it is a mere extension of a technology that they are long used to. The impact of

e-voting then depends on the motive an individual had to use the application. However,

this link serves to differentiate impact from usage, not to equate the two.

In summary, the peripheral citizens of type “B” are unlikely to use internet voting,

but they are strongly affected by it once they manage to clear the first hurdle. Conversely,

voters of type “A” use internet voting more frequently, but the impact on their propensity

to turn out is limited. Similar “bottleneck” effects have been described previously by

Lazarsfeld, Gaudet and Berelson (1944) and Zaller (1991) in the domain of political

communication and its impact on individual preferences.

/---/ the people who did most of the reading and listening not only read and heard

most of their own partisan propaganda but were also most resistant to conversion

because of their strong predispositions. And the people who were most open to

conversion - the ones the campaign managers most wanted to reach - read and

listened least. Those inter-related facts represent the bottleneck of conversion

(Lazarsfeld, Gaudet, Berelson 1944: 95)

We employ the bottleneck metaphor in a similar fashion: the mobilization effect

of internet voting would be strongest among disengaged citizens, but not many of these

citizens manage to use it in the first place. And usage of internet voting is most common

among active citizens, but these citizens do not experience high impact. The interplay of

these two effects constitutes the bottleneck mechanism of internet voting.

If this line of reasoning holds, then usage is both conceptually and empirically

decoupled from impact. By making a distinction between usage and impact we gain the

conceptual clarity required for testing our core hypothesis: characteristics distinguishing

the political periphery from the elite should decrease the probability of usage but increase

impact. Figure 1 represents this hypothetical relationship.

[Figure 1 about here]

8

Context and data

We now turn to empirical evidence from the Estonian parliamentary elections in 2007.

This section briefly describes the main features of Estonian internet voting and introduces

our datasets. On the basis of these data we then proceed to test the two-step model of

usage and impact in the following sections.

Internet voting in Estonia

In October 2005 Estonia became the first country to have statewide local elections where

people could cast binding votes over the internet. This world premiere was followed by

the national parliamentary elections in 2007 where the number of internet voters reached

5.4% of the total turnout2.

[Table 1 about here]

The general feasibility of e-voting in Estonia is based on the widespread use of

electronic identification cards. Since 2002 more than one million of these credit-card size

personal identification documents have been issued. For internet voters they allow to cast

legally binding digital votes at a high security level. Participation in the electronic ballot

requires a computer with an internet connection and a “smart-card reader”. For less than

ten Euro these card readers are easily available at computer shops, supermarkets and bank

offices. For users without personal computer or internet access, internet voting is

accessible through a wide number of free internet access points in public libraries,

community centers, etc.

The process of internet voting is an interaction with the website of the National

Electoral Committee, www.valimised.ee (www.voting.ee). The user first inserts the ID-

card into a card reader and opens the website. Then the user is required to identify

himself/herself through a PIN-code associated to her/his ID-card. If the user is eligible to

vote, the system displays the list of candidates by party in the user’s electoral district. The

user chooses a candidate by clicking on the name and confirms the choice by using a

2 For further details see the reports of the Estonian National Electoral Committee (2005; 2007a; 2007b).

9

second PIN-code. At the end of the process, the voter receives a confirmation that the

vote has been cast3.

Internet voting is available during three days of advance voting (6-4 days before

Election Day). To prevent coercion and fraud, internet voters are allowed to recast their

electronic vote with their previous vote being deleted. For similar reasons, internet voters

can dismiss their electronic vote altogether by casting a paper ballot on Election Day.

Surveys

Our study employs data from two Estonian surveys4. First, we use a representative

population survey to explain individual usage of internet voting. Second, a new online

survey of e-voters will shed light on the impact of e-voting on turnout.

The general population survey was carried out before the 2007 parliamentary

elections between the 10th and 21st of February 2007 on a random sample of 803 adult

Estonians (18 years and older). Data were gathered through interviewer-assisted

questionnaires. The survey performed well is terms of demographic representativeness;

minor deviations were adjusted by weighting.5

The e-voter survey was conducted online within one week’s time after Election

Day between the 5th and 11th of March 2007. It reached a sample of 1206 respondents

what is about 4% of the total e-voter population. The sample was recruited through a two-

stage snowball strategy. First a direct e-mail invitation was sent to more than 50 public

mailing lists and about 100 individuals from academia and the public and private sector.

3 For further details see the Election Assessment Mission Report of the OSCE/ODIHR (2007) and the

Overview of the Estonian E-voting System by the Estonian National Electoral Committee (2005).

4 Both surveys were financed by the University of Tartu, Institute of Journalism and Communication,

supported by grants from the Estonian Ministry of Education and Science (grant nr. 0180017s07) and the

Estonian Science Foundation (grant nr. 6526).

5 The sample was weighted according to age, gender and place of residence. Reference values were

obtained from Statistics Estonia (2008). In both surveys missing values were handled by multiple

imputation. The Amelia II program (King et al. 2001; Honaker, King & Blackwell 2006) was used to

produce five imputed datasets, and all calculations were carried out for each of them. As proposed by

Rubin (1987), the final point estimates simply represent the mean across the five datasets, and the final

standard errors are based on the mean variance within the five datasets plus the variance across the five

datasets (multiplied by a correction factor).

10

Second, online advertisements appeared in the two largest national newspapers, Eesti

Päevaleht and Postimees, for three days right after the election.

This method of data collection allowed us to reach a large sample of the e-voter

population by efficient means what would be hard to achieve with other survey modes.

When it comes to impact, we need to analyze variation within the group of e-voters. To

include a number of e-voters that is sufficient for this approach, conventional random

sampling would require an enormous overall N. Our online strategy has clear advantages

in this respect. However, it also entails two drawbacks: demographic non-

representativeness and potential self-selection bias.

Demographic non-representativeness results from the tendency of variables that

predict usage of e-voting also to predict participation in the online survey. Table 2

compares the sample distribution of internet voters on four criteria – age, gender, place of

residence, and vote choice – to the distribution of the whole e-voter population as

available from official statistics (Statistics Estonia 2008; Estonian National Electoral

Committee 2007a). We find that the younger, male and urban population is

overrepresented in the survey. On the basis of the information in Table 2 we constructed

post-stratification weights to bring the marginal distribution of our sample in line with

that of the population. Estimation results remained robust throughout this procedure,

indicating that although our sampling design affected average levels of some variables,

their relationships (which are of interest here) are adequately represented.


Post-stratification also limits the role of potential self-selection bias: Table 2

shows some discrepancies between the population and the sample, but it also shows that

no group is systematically excluded from the sample. There is sufficient variance on all

criteria to generate a representative image of the e-voter population, especially given the

large N of 1206. However, bias may go beyond these observable criteria if self-selection

operates on the basis of some unobserved criterion. We suspect that two such criteria may

play a role: citizens’ generalized attitudes toward technology and their experience of

using technology. The survey may mostly attract people who are optimistic about e-

11

voting as technological innovation in general (“euphoria” bias) or who have just had a

pleasant experience using e-voting (“conversion” bias). These citizens would be likely to

report relatively high impact. We thus have to be cautious about taking overall levels of

impact at face value. However, our primary interest is not in estimating impact per se but

in predicting its distribution across the electorate. The implications of self-selection for

our hypothesis of high impact on peripheral citizens depend on which form of bias is

present in which part of the population. Table 3 presents the possible scenarios.


Both forms of bias would lead to conservative estimates if they occur among the

elite. We expect citizens who are likely to e-vote to experience low impact. If there is

bias in this group of citizens in favor of higher impact, it becomes harder to confirm our

hypothesis. Bias among the periphery is more complicated. Euphoria bias would mean

we have sampled rather atypical “peripheral” citizens. The role of political and

demographic characteristics that identify this group would be superseded by some

unobserved criterion that unites all respondents in the sample. Then, the data would not

possess structure and our variables would display only null findings. The only scenario

that implies a liberal test of our hypothesis is conversion bias among the periphery. If we

have primarily sampled peripheral citizens with a positive experience of e-voting, we

may overestimate impact in this group. We cannot ultimately exclude this possibility, but

it should be noted that even if liberal bias is present, conservative bias is likely to offset

its effect (the question of overall levels mentioned above). And finally, the scenario of

exact estimation with no bias in any of the groups is not ruled out by any of these

considerations.

Explaining the usage of e-voting

We now turn to the first component of our two-step model, namely the explanation of the

usage of e-voting. Our aim is to go beyond the descriptive information presented above in

three respects. We provide analytical leverage by using multivariate analysis; we consider

12

attitudinal variables in addition to demographics; and we model the choice between

abstention, conventional voting and e-voting.

Variables

The dependent variable is the type of intended behavior in the parliamentary elections

(abstention, conventional voting or e-voting) as derived from two survey questions6. The

choice of independent variables follows from the theoretical discussion of reinforcement

and mobilization effects above. Two main types – demographics and attitudes – will be

treated separately in our analysis. This takes into account that attitudes can be partly

derived from demographics; merging the two models would therefore lead to problems of

collinearity and suppress certain effects that are of particular interest in our analysis.

The demographic variables are age (five groups), income (banded), place of

residence (logged population), gender, and education (elementary, secondary and higher).

Attitudinal variables include political activity, trust, and media consumption. We

aggregated these variables from multiple items, allowing us to capture a concept with

adequate breadth and to reduce random error variance at the same time7.

Political activity is a scale of participation in political meetings, signing public

petitions, contacting the media, and opinion leadership in politics8. Political trust is

operationalized as two scales: one for trust in politics (government, politicians), and one

for trust in the polity (the State, the President and the courts). Media consumption

incorporates the perceived importance of print media, radio and television for obtaining

campaign information. We also include self-reported computer literacy as a baseline

6 Question 1: Are you planning to cast your vote in the coming Parliamentary elections (Answers provided:

Yes; Probably yes; Probably no; No. Collapsed to two categories.). Question 2: Do you, or does someone

of your close friends, plan to cast your vote over the internet in the coming elections? (Relevant answer:

Yes, I intend to).

7 In general the items used here do not only differ in the particular aspect of a concept they measure, but

also in their position on the latent dimension (their “difficulty”). In such a case techniques like factor

analysis often fail to identify the true latent structure. To account for varying difficulty between items, we

applied polytomous Mokken scaling to derive the dimension of interest (Mokken 1971; Hemker, Sijtsma &

Molenaar 1995; Hardouin 2007). All items were rescaled to the same range for this procedure.

8 Question: Do other people ask frequently your opinion about the following subjects? (A battery where

“Politics” is one element).

13

effect that allows us to evaluate the performance of the attitudinal variables over and

above digital affinity.

These variables characterize citizens who should be likely to use internet voting

according to the theories of digital and political divide discussed above. We expect e-

voters to be young, educated and wealthy males from urban areas. As regards attitudes,

we expect a positive influence of political activity and media consumption. The impact of

trust as conceptualized here should depend on the particular aspect: Trust in politics

reflects a passive conception of the individual citizen’s role in democratic affairs; in

comparison, trust in the polity attests to the belief that the individual citizen may play an

active role in a functioning democratic order. We expect e-voters to have a strong sense

of political efficacy and thus to trust the polity but not necessarily politics9.

Method

The aim of our statistical approach is to model the choice between three nominal

outcomes – non-voting, conventional voting, and e-voting. We estimate this model by

multinomial probit regression. Probit is preferred over logit because the latter imposes the

assumption of independence of irrelevant alternatives, i.e. the odds between each pair of

alternatives do not depend on the inclusion of other alternatives (Maddala 1983: 61ff.). In

our case this assumption is likely to be violated in two ways. First, voters may get used to

the convenience of e-voting so that the odds of conventional turnout over abstention

decrease. Second, e-voting may mobilize non-voters for whom the odds of conventional

turnout over abstention increase. In both cases the alternatives are not independent and a

model that does not impose this assumption (such as probit) is required10.

9 The role of trust is based on the distinction between institutional trust and trust toward political actors

(Citrin 1974). We expect e-voters to be “critical citizens” (cf. Norris 1998) in that they trust the system

(polity), but not necessarily the actors within the system (politics). It is this pattern of trust that should go

along with political efficacy. While clearly defined in the case of “critical citizens”, however, one has to be

cautious about the relation of trust and efficacy more generally (cf. Craig, Niemi & Silver 1990).

10 Notwithstanding these theoretical considerations, multinomial logit regression produced similar results

for both models. Moreover, there was reason to test two other estimation strategies: First, the choice

between abstention, conventional voting and e-voting could be modeled in a sequential manner: Citizens

would first decide whether to turn out and then which method to use to cast their vote. Of course such a

model would suppress potential mobilization effects of the availability of e-voting. Conversely, however,

one might wonder whether multinomial probit as a simultaneous-choice model overly favors such effects.

14

Multinomial probit coefficients are generally hard to interpret. To facilitate this

task, we rescaled all independent variables to a range from 0 to 1 and calculated first

differences in probabilities11. These estimates can then be interpreted as the effect of

moving a variable from its minimum to its maximum value on the probability of a certain

outcome (non-voting, voting, or e-voting). The results are presented in Table 4 and 5.

Results: usage


The results of the demographic model in Table 4 confirm the basic expectations

discussed in the literature. First, we find that the probability to turn out increases with

age. However, this effect applies only to conventional voting whereas the probability to

e-vote is similar for the first four age categories. In the oldest category we find a sharp

decline of e-voting of more than 7%. This effect is quite substantial given that the sample

contains only 9% of e-voters. With regard to age, e-voting ranks between abstention and

conventional voting. Age plays an important role in explaining individual turnout

regardless of which method is employed to cast one’s vote. In this sense, e-voters display

the same characteristics common to all voters. However, among those who decide to

vote, e-voters are more likely to be younger than conventional voters.

Among the other variables, urban residence increases the likelihood to e-vote.

Females are more likely to participate in voting, but not in internet voting. Income does

not seem to play a significant role. Education initially increases the likelihood of e-voting

and even more so of conventional voting. Interestingly, however, this order is reversed

for higher education where e-voting prevails. Citizens with higher (as compared to

secondary) education are not notably more likely to turn out in general, but they are more

likely to prefer e-voting to conventional voting.

Therefore we also tested a Heckman selection model (Heckman 1976) with similar results. Second, a

potential problem is the relatively low number of e-voters in the sample. We replicated our analysis using

rare events logistic regression, a technique designed for dependent variables with a rare positive outcome

(Tomz, King & Zeng 1999; King & Zeng 2001). Also this estimator confirmed the results reported here.

11 This was done by averaging over 1,000 simulations drawn from the multivariate normal distribution

using an adapted version of Beber’s (2008) -qi- procedure in Stata.

15


The results of the attitudinal model in Table 5 also confirm theoretical

expectations. The general likelihood of turnout increases with political activity where the

effect on e-voting is particularly strong. Trust in politics increases the likelihood to vote

but decreases the likelihood to e-vote (although not significant). Trust in the polity

increases the probability of both voting and e-voting. Thus, the impact of trust indeed

depends on the particular aspect: Non-voters display low trust toward politics and even

lower trust toward the polity; conventional voters display high levels in both dimensions;

and e-voters tend to trust only the polity. With regard to the media, e-voters exhibit a

level of consumption between the high value of conventional voters and the low value of

non-voters. Finally, computer literate citizens are more likely to substitute e-voting for

conventional voting.

In summary we find that most demographic and attitudinal variables perform in

line with theoretical expectations. Age, education, urban residence, gender, political

activity and computer literacy seem to play their expected role. Moreover, the impact of

trust suggests that e-voters have a strong sense of political efficacy. Interestingly, media

consumption does not increase the likelihood of e-voting. Additional analyses of items

measuring information seeking behavior on the internet showed that e-voters generally

replace traditional media with online sources.12 This seems to add to the image of the

typical e-voter as an independent and demanding citizen.

Explaining the impact of e-voting

We now turn to the second component of our two-step model, namely the explanation of

the impact of e-voting. In particular, we want to know whether e-voting affects individual

turnout and whether any such impact is distributed asymmetrically across the population.

12 In order to avoid endogenous explanations of e-voting we have not included variables related to online

activity in our final models.

16

We first report descriptive statistics of general mobilization effects and then proceed with

multivariate analysis to predict these effects.

Table 6 and 7 present evidence from two survey questions on the impact of e-

voting. The first asks for impact on the motivation to turn out; the second concerns

impact on behavior in the recent election. Each of the two items leans toward one of the

dimensions of impact we have discussed above: past behavior primarily reacts to the

availability of the option of e-voting, and present motivations primarily reflect the

experience of using the application. Responses to both items document general

mobilization effects, and motivations seem to be more strongly affected than behavior.

The size of the behavioral effect corresponds to others found in earlier studies (Trechsel

et al. 2007; Boogers 2006). The size of the motivational effect indicates that future

elections may see an even stronger impact on behavior.



When interpreting these results one should be aware that we are dealing with

subjective evaluations. Given that citizens are not always the best judges of their own

motivations and behavior, we would prefer to also trace the impact of e-voting on the

basis of long-term panel studies. Obviously, this option is not available shortly after the

introduction of the technique in a single country. But of course the lack of alternatives

does not eliminate the shortcomings of subjective evaluations. Below we will propose an

estimation strategy to account for potential bias.

The descriptive findings indicate that the option of e-voting did indeed affect the

propensity to turn out of a good part of our sample. We do not claim that this impact is

representative in its entirety, but an effect of this magnitude is highly unlikely to be a

mere artifact of subjective evaluation. Our next aim is to explain variation in impact: why

are some e-voters affected by the new technology and others are not? Our bottleneck

model suggests that usage and impact of e-voting are negatively related. Thus, we expect

the impact to decrease with those variables predicting the probability that an individual

used e-voting in the first place. Testing this hypothesis requires multivariate models that

17

resemble the models we used to explain usage as closely as possible. Again, we expect

variation by demographics and attitudes.

Variables

To operationalize the dependent variable, we investigated how the two aspects of impact

(motivations and behavior) are related to each other. As is evident from Table 6 and 7,

the two items differ considerably in terms of difficulty (motivational impact is achieved

more easily than behavioral impact). To assess scalability, we therefore subjected the two

items to Mokken analysis (as explained above; with motivation reduced to a binary item)

and achieved a strong Loevinger’s H of 0.67. This means that behavioral change does

generally not occur without motivational change, but it indicates even stronger impact on

top of changing motivations. The two aspects of impact represent the same latent

phenomenon, and their scale will serve as our dependent variable.

Concerning independent variables, the demographics are the same as in the model

of usage (i.e. age, income, residence, gender, and education)13. Attitudinal variables

include political activity, sense of political efficacy, and perceived user friendliness of the

e-voting system. Activity and efficacy are captured through self-assessment questions14.

The question on activity contains very similar elements as the scale used to predict usage.

Efficacy takes the role of trust in the model of usage. The e-voter survey does not contain

a trust item, but the logic of the variable can be adequately represented. We considered

trust in two distinct dimensions: trust in politics reflecting a passive role in democratic

affairs, and trust in the polity reflecting an attitude of confidence in the functioning of

democracy that is closely related to efficacy. This logic allows us to compare the effect of

efficacy on impact with the effect of trust on usage. Finally, user friendliness is a scale of

three items including an assessment of the website, difficulties in installing the smart-

13 Place of residence is limited to a dummy for the two biggest Estonian cities, Tallinn and Tartu.

14 Question on activity: Could you please describe the level of your political activity - how often do you

participate in political events, talk about politics with your friends and family and follow political

developments?

Question on efficacy: Do you think that your vote influences who is in power and how the country is

governed?

18

card reader, and the number of attempts needed to successfully cast an electronic vote.

PC literacy is again included as a baseline effect.

Method

As mentioned above, predicting self-reported impact requires an adequate estimation

strategy. Certain respondents will interpret the items constituting our dependent variable

as if they were limited by an upper bound. Research on turnout shows that a part of the

electorate votes as a matter of mere habit (Franklin 2004; Gerber, Green & Shachar 2003;

Plutzer 2002). These voters have arguably reached some maximum level of electoral

participation and it is unlikely that they will report an impact of e-voting on their past or

future propensity to turn out. However, this effect is not owing to a “natural” upper limit

in the propensity to turn out itself, but owing to the limited number of elections that

citizens can participate in. This could only be avoided in a polity where dozens of

elections are held every day. But with only one general election every four years (and just

a few second-order elections in between) our dependent variable is effectively censored.

In fact, this form of bias is a symptom of the problem of subjective evaluation

discussed above: people may be bad judges of their own behavior. We therefore apply an

estimation technique that accounts for the constraints imposed on people by the limited

number of elections. The impact of e-voting on turnout is treated as a latent variable that

is not fully observable due to upper censoring. Censoring is expected for those

respondents who report having turned out in all six previous elections (national, local and

European). OLS regression would be inconsistent in this case because it takes censored

point data at face value (Wooldridge 2002: 524f.). Instead, we estimate the model by

interval regression. Censored responses are defined as elements of an interval having as

lower bound the measured value of the dependent variable and as upper bound the

maximum value applicable to all respondents. The model parameters can then be

obtained by maximum likelihood estimation.

Again all independent variables were rescaled to a range from 0 to 1 so that the

estimated coefficients can be interpreted as the effect of moving a variable from its

minimum to its maximum value. Table 8 and 9 present the results.

19

Results: impact


The results of the demographic model confirm our core hypothesis of high impact

on peripheral citizens. This is clearly the case with regard to older and rural voters. Less

educated citizens also experience higher impact, but these effects are insignificant and

relatively weak. This null finding, however, should be interpreted in light of our core

hypothesis. We have argued that usage and impact of e-voting cannot be equated.

Moreover, we have seen above that education entails higher usage. The null hypothesis

against which we test our model thus posits equality of usage and impact: if education

leads to usage, it should also lead to impact. This is not what we find. Therefore, the

effect of education supports the distinction of usage and impact. The insignificant effect

of gender may be interpreted in a similar way.


The findings of the attitudinal model are coherent with the demographic model

and our core hypothesis. Political activity decreases the impact of e-voting, but it

increased the likelihood of usage. Political efficacy also reduces impact, an effect that

should be related to the influence of trust on usage. Positive user experience leads to

higher impact even while controlling for computer literacy. Again, the null effect of PC

literacy should be contrasted with the expectation one would derive from an equation of

usage and impact, namely a strong positive effect in this case.

In sum, we find that the impact of e-voting is indeed higher among peripheral

citizens, but usage of e-voting is more likely among the well-educated elite. This implies

that usage and impact should be treated separately, both conceptually and empirically.

The distinction can be demonstrated most clearly by contrasting the predictions from the

two models. Figure 2 achieves this graphically. The vertical axis represents impact as

predicted for the respondents of our online survey from the demographic and attitudinal

models in Table 8 and 9. The horizontal axis represents the likelihood that these

20

respondents voted online. Note that of course all of them did so, but for some this was

more likely than for others. These differences in the likelihood of usage are due to

differences in demographics and attitudes. Each respondent features a combination of

these characteristics that can be expressed as a propensity score: using the parameters of

the two models of usage in Table 4 and 5, we predicted for each respondent of the online

survey the propensity to be included in the e-voter population in the first place.15 Figure 2

plots the expected level of impact for the full range of propensity scores (where both

values were averaged over the predictions of the demographic and the attitudinal model).

[Figure 2 about here]

According to our bottleneck model internet voting fails to increase turnout because its

impact is highest among those citizens who are unlikely to use it. Figure 1 presented this

logic in hypothetical form. The regression line in Figure 2 captures the same effect

empirically: the higher the likelihood of e-voting, the lower the expected impact on

turnout. The variables that gave rise to this prediction identify peripheral citizens as high-

impact but low-probability users. These citizens seem to face many barriers in accessing

e-voting, but once they manage to clear the first hurdle the impact on their propensity to

turn out is high. Conversely (and perhaps somewhat counter-intuitively), it is not the

young and educated who are being mobilized into political life by the new technology.

Frequent usage in this group does not lead to high impact.

Conclusion

The image of a bottleneck is usually evoked to describe a process that is constrained by

one single element while other elements are idling. We have modeled such a process to

explain usage and impact of internet voting applications. Some scholars argued that the

option to vote on the internet should lead to an increase in voter turnout. Others replied

that such effects are unlikely because internet voting merely replicates existing patterns

15 Out-of-sample prediction requires comparable variables across the models of usage and impact. Two

variables required special attention. Efficacy serves to represent the diametrical effect of trust in politics

and trust in the polity. Media consumption does not have a correspondent, so the variable was imputed by

the mean value of the e-voters from the general population survey.

21

of political participation. Our bottleneck model unites these claims in one framework.

Usage of internet voting is mostly restricted to the politically engaged, but the impact of

this technology on the propensity to turn out is highest among peripheral citizens.

To all appearances, then, internet voting does not only increase turnout, but it also

counteracts inequality in political participation. We do not doubt that as of today the new

technology mostly benefits the political elite. However, in the long run the very reasons

of this disparity may undermine their own short-term effects. The elite may well benefit

from e-voting, but these benefits concern matters of mere convenience. More radical

effects are expected mostly among peripheral citizens for whom e-voting may serve as a

stepping stone toward political activity in general. In the long run, this mobilization effect

should offset the pro-elite bias inherent to online politics. However, the pace at which this

is possible critically depends on the narrow bottleneck of usage that restricts the impact

of e-voting.

Our expectations of the future role of e-voting are mixed. Once the new

technology achieves a critical amount of users, some laggards get carried along (Rogers

2003). The bottleneck will widen and let a higher impact pass. In the long run, then,

overall turnout should increase. However, there is an upper limit for the rising tide to lift

all the boats. The bottleneck model assumes that peripheral citizens become e-voters out

of interest and curiosity with regard to technology. Once the technology loses its

innovative character and becomes “domesticated” in everyday practices (Silverstone and

Hirsh 1994), the impact of internet voting may disappear. The development of e-voting in

Estonia, the case we have drawn on here, will shed light on these dynamics.

22

Bibliography

Alvarez, R.M., Hall, T.E. & Trechsel, A.H., 2009. Internet Voting in Estonia. PS: Political Science & Politics, forthcoming

Alvarez, R.M. & Nagler, J., 2000. Likely Consequences of Internet Voting for Political

Representation, The Loyola of Los Angeles Law Review, 34, 1115. Beber, B. 2008. '-qi-: First Differences for Multinomial Probit Models in Stata',

http://www.columbia.edu/~bhb2102/software.htm [Accessed January 31, 2009]. Boogers, M., 2006. Enquete bezoekers StemWijzer, Available at: http://www.publiek-

politiek.nl/Bestanden/Enquete-bezoekers-StemWijzer [Accessed May 20, 2008]. Citrin, J. 1974. The Political Relevance of Trust in Government. American Political

Science Review. 68(3) pp. 973-988. Craig, S., Niemi, R., Silver, G. 1990. Political Efficacy and Trust: A Report on the NES

Pilot Study Items. Political Behavior, Vol. 12, No. 3, pp. 289-314 van Dijk, J., 2005. The Deepening Divide: Inequality In The Information Society, Sage. van Dijk, J., 2000. Widening information gaps and policies of prevention. Digital

Democracy: Issues of Theory and Practice, 166-183. Estonian National Electoral Committee, 2007a. Parliamentary elections 2007. Statistics

of e-voting, Available at: http://www.vvk.ee/english/Ivoting_stat_eng.pdf [Accessed November 28, 2008].

Estonian National Electoral Committee, 2007b. Statistical overview and comparison of

the results of I-voting in 2007 parliamentary elections and 2005 local elections, Available at: http://www.vvk.ee/english/Ivoting comparison 2005_2007.pdf [Accessed November 28, 2008].

Estonian National Electoral Committee, 2005. Estonian I-Voting System: General

Description, Available at: http://www.vvk.ee/elektr/docs/Yldkirjeldus-eng.pdf [Accessed May 15, 2008].

Franklin, M.N., 2004. Voter Turnout and the Dynamics of Electoral Competition in

Established Democracies Since 1945, New York: Cambridge University Press. Gerber, A. S., Green, D.P., and Shachar, R., (2003) 'Voting May Be Habit-Forming:

Evidence from a Randomized Field Experiment', American Journal of Political Science 47(3), 540-50.

23

Hardouin, J., 2007. Mokken Scale Procedures for Stata. Available at: http://www.freeirt.org/ [Accessed October 25, 2007].

Heckman, J., 1976. The Common Structure of Statistical Models of Truncation, Sample

Selection, and Limited Dependent Variables and a Simple Estimator for such Models. Annals of Economic and Social Measurement, 5(4), 475-92.

Hemker, B.T., Sijtsma, K. & Molenaar, I.W., 1995. Selection of Unidimensional Scales

From a Multidimensional Item Bank in the Polytomous Mokken IRT Model. Applied Psychological Measurement, 19(4), 337-52.

Honaker, J., King, G. & Blackwell, M., 2006. Amelia II: A Program for Missing Data.

Available at: http://gking.harvard.edu/amelia/ [Accessed July 22, 2007]. Kersting, N. & Baldersheim, H., 2004. Electronic Voting and Democracy: A

Comparative Analysis, Palgrave Macmillan. King, G. et al., 2001. Analyzing Incomplete Political Science Data: An Alternative

Algorithm for Multiple Imputation. The American Political Science Review, 95(1), 49-69.

King, G. & Zeng, L., 2001. Logistic Regression in Rare Events Data. Political Analysis,

9(2), 137-163. Kleinnijenhuis, J. & van Hoof, A., 2008. The Influence of Internet Consultants.

University of Antwerp. Lazarsfeld, P. F., Berelson, B., and Gaudet, H., 1944. The People's Choice. Columbia

University Press. Maddala, G.S., 1983. Limited-Dependent and Qualitative Variables in Econometrics,

Cambridge University Press. Margolis, M. & Resnick, D., 2000. Politics as Usual: The Cyberspace" revolution", Sage

Publications. Mokken, R.J., 1971. A theory and procedure of scale analysis with applications in

political research, Mouton The Hague. Mossberger, K., Tolbert, C.J. & Stansbury, M., 2003. Virtual Inequality: Beyond the

Digital Divide, Georgetown University Press. Norris, P.(ed) 1999. Critical Citizens: Global Support for Democratic Governance.

Oxford University Press.

24

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. Available at: http://www.pippanorris.com/.

OSCE/ODIHR, 2007. OSCE/ODIHR Election Assessment Mission Report in the 2007

parliamentary elections in Estonia, Available at: http://www.vvk.ee/english/OSCE report_EST_2007.pdf [Accessed November 28, 2008].

Plutzer, E. (2002) 'Becoming a Habitual Voter: Inertia, Resources, and Growth in Young

Adulthood', American Political Science Review 96(1), 41-56. Putnam, R., 2000. Bowling alone, Simon & Schuster New York. Rogers, E. M., 2003. Diffusion of Innovations, 5th ed. New York, NY: Free Press. Rubin, D.B., 1987. Multiple Imputation for Nonresponse in Surveys. New York: J. New

York: Wiley. Silverstone, R. & Hirsch, E. (eds.) (1994). Consuming Technologies: Media and

Information Domestic Spaces. London: Routledge. Solop, F.I., 2001. Digital Democracy Comes of Age: Internet Voting and the 2000

Arizona Democratic Primary Election. PS: Political Science and Politics, 34(02), 289-293.

Statistics Estonia, 2008. Demographic indicators. Available at: http://www.stat.ee/

[Accessed September 11, 2008]. Tomz, M., King, G. & Zeng, L., 1999. ReLogit: Rare Events Logistic Regression.

Available at: http://gking.harvard.edu/stats.shtml [Accessed October 1, 2008]. Trechsel, A.H., Schwerdt, G., Breuer, F., Alvarez, M., Hall, T., 2007. Council of Europe

and Estonian National Electoral Committee Report on Internet Voting in the 2007 parliamentary elections in Estonia, Available at: http://www.vvk.ee/english/CoE and NEC_Report E-Voting 2007.pdf [Accessed November 28, 2008].

Wattenberg, M. P. 2008. Is Voting for Young People? New York: Pearson Longman. Wilhelm, A.G., 2000. Democracy in the Digital Age: Challenges to Political Life in

Cyberspace, Routledge.

25

Wooldridge, J. M. (2002) Econometric Analysis of Cross Section and Panel Data, Cambridge: MIT Press.

Zaller, J., 1992. The Nature and Origins of Mass Opinion.

26

Table 1. Main statistics of internet voting

2005 2007

Eligible voters 1 059 292 897 243

Voter turnout 47.4 % 61.9 %

E-votes counted 9 287 30 243

E-votes among all votes 1.9 % 5.4 %

Source: Estonian National Electoral Committee.

27

Table 2. Marginal distributions of internet voters

Percentages

Population Sample Difference

Gender

Female 48.2 41.5 -6.7 Male 51.8 58.5 6.7

Age (i)

18-24 11.2 16.4 5.2 25-34 32.0 46.4 14.4 35-49 32.7 24.3 -8.4 50-64 18.2 10.9 -7.3 65+ 5.9 2.0 -3.9

Residence

Tallinn and Tartu 47.6 74.3 26.7 Remaining country 52.4 25.7 -26.7

Vote choice

Eesti Reformierakond 34.5 30.2 -4.3 Isamaa ja Res Publica Liit 26.7 33.8 7.1 Sotsiaaldemokraatlik Erakond 13.3 18.1 4.8 Eestimaa Rohelised 10.7 13.1 2.4 Eesti Keskerakond 9.1 2.2 -6.9 Eestimaa Rahvaliit 3.6 1.8 -1.8 Others 2.1 0.8 -1.3

Source of the population data: Estonian National Electoral Committee. (i) The last age category in the original population data starts at 60. Congruence with the survey categories was established by estimating density as a linear function of age and adjusting percentages accordingly.

Table 3. Potential forms of self-selection and their implications

Periphery Elite Euphoria bias Null effects Conservative estimation Conversion bias Liberal estimation Conservative estimation No bias Exact estimation Exact estimation

28

Table 4. Probability of e-voting predicted from demographics

Abstention Voting E-voting

Age (basis: 18-24)

- 25-34 .010 .020 -.030

(.060) (.063) (.036)

- 35-49 -.075 .077 -.002

(.052) (.057) (.037)

- 50-64 -.115** .139** -.024

(.057) (.063) (.038)

- 65+ -.205*** .276*** -.071*

(.055) (.063) (.041)

Income -.093 .094 -.001

(.067) (.072) (.037)

Urban residence .011 -.057 .046*

(.044) (.048) (.025)

Gender (f) -.044 .090** -.046**

(.036) (.038) (.021)

Education (basis: elementary)

- Secondary -.212*** .145** .067*

(.048) (.057) (.039)

- Higher -.227*** .103 .124*

(.041) (.069) (.064)

N 803

Log pseudo-likelihood -626

% correctly predicted 69

Wald test 60***

First differences from multinomial probit regression with robust standard errors in parentheses.

* significant at .1 ** significant at .05 *** significant at .01

Table 5. Probability of e-voting predicted from attitudes

Abstention Voting E-voting

Political activity -.268*** .113 .155**

(.059) (.093) (.079)

Trust (politics) -.139 .213** -.074

(.098) (.106) (.058)

Trust (polity) -.337*** .218* .119*

(.117) (.129) (.063)

Media consumption -.416*** .367*** .049

(.059) (.066) (.032)

PC literacy .056 -.214*** .158***

(.053) (.058) (.038)

N 803


% correctly predicted 69

Wald test 117***

First differences from multinomial probit regression with robust standard errors in parentheses.


29

Table 6. Did the option of e-voting affect your motivation to participate in elections?

Negative effect 0.03%

No effect 53.06%

Positive effect 46.91%

Table 7. Would you have turned out without the option of e-voting?

I would have voted anyway 62.48%

I would rather have voted 21.84%

I would rather have abstained 8.41%

I would have abstained 7.27%

30

Table 8. Impact of e-voting on turnout predicted from demographics

Age (basis: 18-24)

- 25-34 .088** (.043)

- 35-49 .167*** (.044)

- 50-64 .169*** (.053)

- 65+ .209*** (.066)

Income .029 (.057)

Urban residence -.055** (.028)

Gender (f) -.044 (.027)

Education (basis: elementary)

- Secondary -.021 (.049)

- Higher -.029 (.044)

Constant .495*** (.056)

N 1206


Wald test 41***

Interval regression coefficients with robust standard errors in parentheses.


Table 9. Impact of e-voting on turnout predicted from attitudes

Political activity -.081* (.047)

Political efficacy -.152*** (.050)

User friendliness .175** (.085)

PC literacy -.003 (.068)

Constant .571*** (.086)

N 1206


Wald test 15***

Interval regression coefficients with robust standard errors in parentheses.


31

Figure 1. The hypothetical relation of usage and impact

Pro

bab

ility

Political periphery

Usage

Impact

32

Figure 2. Impact of e-voting declining with likelihood of usage

A Bottleneck Model of E-voting: Why Technology Fails to Boost Turnout

Documents

mere usage of internet

impact of e

usage of e

peripheral citizens

political inequality

political sense

bottleneck model of

high impact