Political Ignorance and the Internet

3.1 Data and Descriptives

The data set used for our empirical analysis consists of 108,594 observations and is a merge of six yearly cross sections of surveys conducted by “Forschungsgruppe Wahlen” (a pollster institute) in Germany. We use the data for the period 2001–2004, 2010 and 2014.^[1] Hence, we have data referring to the period of the rise of the Internet, as well as data referring to later years, when alternative possibilities of using the Internet entered the scene, for example linked to the invention of the smartphone or to social media.

Regarding the survey mode,^[2] note that the survey is conducted on the phone and answers are reported electronically. Moreover, using administrative data, the institute randomly selects households in Germany that have a telephone. Within a selected household, the person whose birthday was last is selected for the survey. One yearly survey wave consists of several smaller survey waves. There is usually one survey wave each month with approximately 1500 respondents. For each of these small survey waves, a new random sample of individuals with the described procedure is selected. The main aim of the pollster institute is to obtain a representative sample of the whole German electorate for each survey wave. Hence, we are quite confident that the cross-sectional surveys are comparable across years as they aim at being representative for the German electorate. Any change in the cross-sectional results should therefore be attributable to changes in the German electorate and not in the sample.

Each observation describes one single individual. One group of variables refers to the characteristics of the individuals,^[3] like age, religious affiliation, economic position, union membership, interest in politics, strength of the affinity to individual’s favorite political party (given by affinity to party), the number of persons in the household in which the individuals live, gender, origin, formal education as well as the variable profession, which is a measure of the socio-economic status regarding the autonomy of action and the position in the job constructed with a method from Hoffmeyer-Zlotnik (2003). Moreover, we constructed a variable measuring the Internet usage of an individual (see the Online Appendix for further details). Table 1 shows descriptive statistics.

A further set of 602 variables contains individuals’ answers to questions about their opinions on recent political topics. Each variable documents the answer to one question.^[4] For example, one question was how the individual assesses the introduction of minimum wages in Germany, or the introduction of a toll for passenger cars. The corresponding variable contains the answer of the individual, when she gave one. To be clear, not all 602 questions were asked to all individuals – each participant received only a small set of approximately 20 questions on average for two reasons: first, asking all questions to all individuals would take too much time and asking so many questions would come with the thread of very low response rates. Second, the questions partly change over time in order to refer to recent political topics and issues that do change over time. Hence, asking always the same questions is not possible, as important political topics do change and some questions are solely relevant during a specific period of time. When an individual did not answer a question (i.e. refused to give an opinion), we call this individual indifferent about the issue in question. In this case, the individual had to state explicitly that she did not have an opinion about the specific question.

3.2 Constructing an Indifference Index

As a measure of individual political ignorance (or political indifference) we construct an indifference index. High (low) values of the indifference index indicate high (low) levels of political ignorance. The indifference index is constructed by aggregating and weighting the 602 variables referring to the questions asked in the yearly surveys and assigning a single index value between zero and 100 to every individual. In the following, we will first explain how we compute these weights. Afterwards, we will argue why using these weights leads to a useful indifference index.

The weights of the questions asked to individuals are constructed with hand-collected publication data. Each question refers to a particular topic. For instance, if the question refers to the introduction of a minimum wage, the underlying topic is “minimum wage”. To construct the weight for a certain question, we count the number of publications in a representative sample of German media on the topic(s) covered in this question in the year this question was asked as well as the previous years. These numbers of publications are hand-collected from the search engine FACTIVA. For each question, we searched for all terms related to the topic that is referred to with this question on FACTIVA and collected the number of German media publications that FACTIVA reports that cover this topic.

To construct the indifference index, we define dummy variables O k for k = 1 , ··· , 602 , where O k i is zero if individual i states an opinion for question k , or if i was not asked to answer question k , and one otherwise. Hence, O k i is one if the individual was asked to give her opinion on this question but was not willing and/or not able to share her opinion on this topic. Moreover, we define the dummy variables θ k for k = 1 , ··· , 602 . Here, θ k i is equal to one if individual i was asked to answer question k during the survey and zero otherwise. Intuitively, the index equals the weighted share of questions, which the individual was indifferent about (i.e. had no opinion/answer to this question). The index takes values between 0 and 100 and for individual i surveyed in year j , the index I i , j is constructed by:

The weight ω k j of the question k asked in year j ∈ { 2001 , 2002 , 2003 , 2004 , 2010 , 2014 } is computed by considering the representative sample of German media publications about the underlying topic of question k in year j relative to the publications about the same topic in the five years before year j . The number of publications in year j − t , related to the topic covered in question k , is defined by p k , j − t . For each question, there is publication data available for the year in question as well as for the previous five years, such that t can be any integer between 0 and 5. Furthermore, we define the number of questions asked in year j as K j , where ∑ j K j = 602 . To avoid negative weights, we define:

Note that the weight is simply the number of standard deviations over the mean value of publications in the last five years corrected by the minimum value. For instance, suppose the weight of the minimum wage question is 2. Ignoring the addend of the minimum weight in the weight construction for a second, this means that in the year the question was asked, the number of German media publications referring to the minimum wage were two standard deviations above the mean of publications referring to the minimum wage in the five previous years.

The weighting is conducted in order for the indifference index to fulfill two conditions: First, the index values are comparable across weeks and years, even though different numbers and types of questions are asked from week to week. Second, in the process of aggregation, the single questions are weighted according to the recent relative relevance of a topic in comparison to the relevance of the same topic in the last five years. This second property of the index means that when the underlying topic is totally new, or when the topic is not new but receives some new relevance through new media publications, we assign a high weight. In contrast, topics which are already being discussed for years receive a low weight. To understand why we weight the questions before we aggregate them, consider the topics ‘climate change’ and the ‘Ukraine crisis’ in 2014. The former topic was already being discussed for years, the latter is a totally new topic in 2014. With our analysis, we want to measure the link between the current Internet usage of an individual and her current political ignorance. Considering, for example, climate change, it is very likely that the individual has already formed an opinion about this topic in the past and her current Internet usage does not substantially influence her opinion and decision about climate change. However, the current Internet usage still can influence it, which demands a low, but non-zero weight for the climate change question. By contrast, the crisis in the Ukraine was a late-breaking topic in 2014 and it is very likely that the current usage of the Internet highly influences the decision-making process regarding this topic calling for a high weight according to our scheme. Thus, by weighting the questions, the differences of the relevance of the Internet usage for the decision making process are taken into account.

Figure 1 indicates a relatively stable average indifference value (black curve) across the six years of the sample for the aggregate of all individuals. Also the 25- and 75-percentile index value among individuals in each year does not change substantially. Therefore, we do not observe that there is a general increase/decrease in political indifference over time.

Figure 1:

The graph shows for each year: the average indifference index value of all individuals in that year (black curve), the 75-percentile value of indifference index values for all individuals (grey curve above the black curve) and the 25-percentile value of indifference index values for all individuals (grey curve below the black curve, which is always equal to zero across all years).

3.3 Econometric Analysis

In order to estimate the link between Internet usage and the indifference index, we use the following econometric specification:

The coefficient of interest is β 1 ; x i = ( x i 1 , ··· , x i j ) is the vector of j control variables and γ the corresponding vector of coefficients; ϵ i is the independent and identically distributed error term. We conduct OLS^[5] estimations of different specifications with a varying set of controls and with time dummies. Our main focus will be on the estimates of coefficient β 1 . The estimation results shoud be interpreted in a descriptive way due to potential endogeneity (see Section 3.4 for further details).

For the estimations, we first use the whole sample including all years of observation and year dummies. We also run regressions separately for the year-wise samples in order to illustrate directly how coefficients change over time.

In the full sample, the estimated coefficient β ˆ 1 is negative and significant in all specifications, implying a negative link between Internet usage and the indifference index (see Table 2). This result suggests that individuals using the Internet are better informed and thus less likely to be politically ignorant. As the interaction with time dummies shows, the negative link between the indifference index and Internet usage is reinforced in 2002, but alleviated in the following years. In 2014, the negative coefficient of Internet usage is even outweighed by the coefficient of the interaction term resulting in a significant positive link between political indifference and Internet usage in that year. This pattern is similar for several specifications and holds if we add further controls (see specifications 3, 4, and 6 in Table 2). By conducting year-wise estimations (see Table 4 in Appendix A.3) the same results for the Internet usage coefficient can be shown more clearly (see Figure 2): we obtain negative estimated coefficients β ˆ 1 for the early years of observation that are decreasing in size over time, and a positive estimated coefficient for the last year of observation. Thus, in the early years of Internet usage, individuals using the Internet seem to be less indifferent with respect to the political debate while, in later years, those who use the Internet are more likely to be politically ignorant.

Figure 2:

Coefficient of Internet usage, estimated separately across years with OLS including all control variables and heteroscedasticity-robust standard errors. The grey area above and below the coefficients represents the 95%-confidence interval.

This evidence supports the hypothesis of an increasing information overload of the Internet over time, making it more costly to take decisions. More specifically, as the Internet gets more and more overloaded with information, the information overload effect should be stronger in later years of Internet diffusion. In fact, what we see in Figure 2 suggests that in early years of Internet diffusion, additional availability of information helped users to be better informed and less politically ignorant. By contrast, in later years, the increasing amount of information has led to an overload letting Internet users ignore information and making them politically ignorant.

There are indeed further alternative explanations for our findings. First, individuals’ purposes for using the Internet might have changed over time. While individuals were primarily seeking for information in the early years of Internet diffusion, in 2014, they might have had higher preferences for entertainment because the Internet offered this entertainment more than it did in 2001–2004 (e.g. Persson 2017; Prior 2005). Hence, the group of Internet users in later years might simply use the Internet for different purposes than those in earlier years, changing the link between being a user and political ignorance. Relatedly, political misinformation and hate speech in social media content might cause individuals to withdraw from these media and turn to more entertaining Internet content (see for example the overviews by Blesse et al. (2021) and Zhuravskaya et al. (2020) showing inconclusive evidence on the link between social media and knowledge or political participation). Second, the set of Internet users might have changed over time in a way we cannot capture with our control variables. In earlier years of Internet diffusion, its usage was restricted rather to high-skilled individuals using the Internet predominantly for information search and professional communication. In later years, due to the increasing broadband diffusion and the invention of the smartphone in 2007, more and more individuals with different characteristics and different preferences have been connected to the Internet. With respect to the latter explanation, Figure 3 shows the development of the average probability that an individual uses the Internet distinguishing between individuals with above- and below-average formal education. There is a clear upward trend in the Internet usage for all individuals. Furthermore, individuals with below-average formal education use the Internet much less on average in the earlier years of the sample while they fully catch up with the individuals who have above-average formal education later on.

Figure 3:

Average probability of using the Internet for all individuals (black curve with dots), for individuals with above-average formal education (grey curve with squares) and below-average education (grey curve with triangles) for each year in the sample.

To get some more insights into the determinants of the indifference index, the estimated coefficients of the 12 control variables are examined in more detail in the following. The estimations refer to the full specification (6) in Table 2. The coefficient estimates are quite stable over the different specifications and are all significant. The coefficients of the control variables show the expected sign. Political indifference is negatively related to individuals’ general interest in politics or belonging to a political party or a religious group, living with further individuals in one household or belonging to a union. There is also a negative and significant link for individuals who have a higher level of formal education, a higher economic or professional status. West Germans are politically less indifferent than East Germans whereas older people and women are significantly more indifferent to political issues.

A one-unit increase in the age of a person (measured in categories 1–11) is related to an increase in the indifference index by approximately 0.46 points. This might be explained through preferences becoming more stabilized with increasing age (“preference stabilization”). An older decision-maker already formed her political preferences, inducing the voting decision to be mainly driven through experiences and decisions made in the past. Therefore the information on the most recent political topics have less of an impact on her preference seeking, which means she becomes politically more ignorant.

A further variable reflecting the stability of preferences is the affinity to a political party. A higher degree of affinity is linked to a 0.91 decrease in the indifference index, i.e. the more a person is committed to a party, the less she will be politically indifferent.

A one-unit higher (categories 1–4) self-assessed economic status is related to a 0.64 points lower indifference index, i.e. an individual who evaluates her economic position as good might have more resources to get informed about political topics than those, whose main interest is to get out of the personal “recession”.^[6]

The membership in a union is connected with a 0.60 lower indifference index indicating that the membership is associated with a person’s generally higher political awareness and thus a lower indifference index.

A one-unit increase in the interest in politics as a general measure (categories 1 to 5) is associated with a decrease of the indifference index by 3.71 points indicating that the more one is interested in politics, the more one is informed about it and the less one is indifferent.^[7]

Each additional person in the household of the individual is related to a 0.41-point decrease in the index indicating that living together with other people might foster political discussions^[8] or just make the division of tasks attractive. The division of tasks might include acquiring and processing information and thus yielding a higher degree of information to those living together with others in their household.^[9]

There are two coefficients measuring the role of education and the socio-economic status^[10] with respect to the indifference index, formal education and profession. With a higher level of profession (categories 1–5) the index decreases by 0.48 points whereas a higher level of formal education (categories 1–5) has a coefficient only half as large as that of profession, i.e. it is related with a decrease in the index of 0.20 points. Thus, higher education and higher socio-economic status are related to a lower level of political indifference. This result is in line with the existing literature.^[11]

The coefficient for West/East indicates that individuals living in the western part of Germany compared to the eastern part (post-soviet state) have a 1.10 lower indifferent index on average. The differences in political participation for post-soviet states and other western states are well studied in the literature.^[12]

The estimated coefficient for religion indicates that an individual with no religious affiliation has an indifference index which is 1.02 points lower than the average. One possible explanation for this effect is that individuals who left the church (those are individuals without any religious affiliation) might be those with a higher political awareness in the sense that those people might be more sensitized and critical to specific societal issues and are thus also better informed about political topics in general. There is evidence in the literature supporting this view.^[13]

The regression results for the dummy variable female yield a significant and positive coefficient of 2.91 for women. Since this effect is very large, we further analysed by which group of women this effect is driven and thus included an interaction term between female and age.^[14] By including the interaction with age in the year-wise specifications, the coefficient for female becomes insignificant, indicating that the measured effect of “female indifference” in the first regression is solely driven by older women. These findings are in line with the literature, as they are well-established in political science.^[15]

3.4 Limitations and Robustness Checks

The empirical analysis we have conducted so far has several limitations. First, the econometric results might be prone to endogeneity due to omitted variables or due to simultaneity and therefore can be interpreted as descriptive evidence only. As regards omitted variable bias, for instance, we do not observe individuals’ ability to process information. This characteristic might affect the use of the Internet, however, it is likely to be correlated also to education and profession and thereby to affect political indifference. With respect to simultaneity, using the Internet as a source of information is likely to have an effect on political indifference. However, the level of political indifference might also determine individuals’ purpose for which they use the Internet and thus also determine the Internet usage per se.

Second, since we do not have a panel data set but a series of cross-sections, the estimate of the Internet coefficient might be prone to sample selection. More specifically, it is possible that in later years a sample was chosen which is substantially different in its characteristics than the sample which was chosen in earlier years, which might explain the change in the estimated coefficient over time. However, we argue that it is very unlikely that our analysis suffers from such a sample selection problem. The reasoning of this claim lies in the representativeness of the sample for the whole German electorate. The pollster institute, which is renowned for its election forecasts, claims that the sample is representative for the whole German electorate in each year. Thus, a substantial change in the sample, which might explain the change in the coefficient is unlikely. Moreover, the stability of the estimated control-variable coefficients over time, as well as the reasonable signs of these estimations underline the credibility of the evidence we found with respect to the Internet usage coefficient. Still, with panel data, we could take into account unobserved heterogeneity by fixed effects estimations.

Third, the analysis conducted so far might suffer from the sample of questions changing over time. Most of the questions refer to recent topics and do solely appear in one year. Hence, the change in the Internet coefficient might simply be attributable to the change in the set of questions. To gather some evidence against this concern, we conduct further estimations. Of all questions asked, we could identify eight questions which are asked in all six years without any changes. We construct a weighted indifference index in the exact same way as before using solely those eight questions. OLS estimations with this reduced indifference index as the dependent variable are shown in Appendix A.2. They show qualitatively similar results to the ones above: we see a clear upward trend in the Internet coefficient over time. The coefficient is significantly negative in the early years and it increases substantially over time. Hence, this finding provides some evidence against the concern that the observed change in the Internet coefficient in the main analysis might stem from a change in questions.

Finally, considering the weighting of the questions in the aggregation process of the indifference index, one might ask how the results change when the aggregation of the questions is done without the weighting scheme. More specifically, in such a simple aggregation, an individual’s value of the unweighted indifference index is equal to the share of questions about which this individual was indifferent. Table 2 contains the results of the OLS estimations with the unweighted indifference index as dependent variable in specification (7). The results show that there is no qualitative difference compared to the estimation results obtained by using the weighted index as dependent variable. To understand the meaning of this result, consider the relationship between the weight of a question and the share of people who were indifferent about it, which is analyzed in the Online Appendix. The results of this analysis show that there is no statistically significant relationship between the share of indifferent people and the weight of a question. Assume that the weighting is indeed a good measure of the recent relative relevance, i.e. the novelty of a question. Then, for the individuals, it follows that, on average, being politically indifferent does not depend on the novelty of a question and a topic. Hence, the link between Internet usage and political indifference does not only change over time for those questions for which we would assume that individuals have a higher need for information, but for any question which is asked to individuals.