• No results found

Income changes and voting in Sweden during the 1990’s: : Is voting an inferior good?

N/A
N/A
Protected

Academic year: 2021

Share "Income changes and voting in Sweden during the 1990’s: : Is voting an inferior good?"

Copied!
39
0
0

Loading.... (view fulltext now)

Full text

(1)

Income changes and voting in Sweden

during the 1990’s:

Is voting an inferior good?

Author: Erik Wätterbjörk

Spring 2019

Degree project: Thesis, Second Cycle, 30 Credits Subject: Economics

Örebro University School of Business

Supervisor: Dan Johansson, professor, Örebro University School of Business Examiner: Henrik Jordahl, professor, Örebro University School of Business


(2)

Abstract

There is a large literature in political science on the individual level variables that are associated with voting in elections at a high rate. Income is typically one variable that is positively connected to high turnout. Mainstream economic theory also tend to see voting as a normal good that increases with income, if it is rational at all. Using an individual panel from the 1991-2002 Swedish National Election Studies surveys that follows voters across two elections the

reliability of earlier cross sectional estimates of the effect of voting on income is tested by using the difference in income between elections. The result shows that it is possible that voting is an inferior good that decreases with rising individual income.


(3)

Table of contents

1. Introduction 2. Institutional framework 3. Theoretical background 4. Previous studies 5. Data 6. Empirical model 7. Results 8. Discussion 9. Conclusions References

(4)

1 Introduction

How many people choose to take part in the political process can have large effects on economic outcomes. For example as seen in the Larcinese (2007) cross country study on social spending and using more robust methods in the specific cases of Sweden and Finland on taxes (Aggeborn, 2016). Why citizens of democracies vote is to themselves and others around them perhaps not a mystery. Voting is maybe something that you do or something that you do not do but to most people who are not economists or political scientists describing using one of the most central aspects of political participation as a paradox might seem strange. But seeing the widespread act of voting as a paradox is precisely what many economists did in works beginning with Downs (1957) introduction of the rational voter hypothesis. In the Downsian framework later elaborated on by Tullock (1967) the voter approaches the voting decision rationally and will only cast a vote if the expected utility of doing so outweigh the costs. Since the institutional framework assumed by Downs and his followers is a single member district majority vote the precise probability of each voter’s influence on the election can be estimated. When this was done it was quickly found that most votes are wasted (Beck, 1975) since the probability of an individual voter affecting the election outcome was very very small.

Economists responded to this failure of the original model to explained observed behavior by three ways 1) redefine the voters calculation in some way to allow voting 2) relax the rationality assumption 3) relax the self-interest assumption (Muller, 2003). This essay uses the case of Sweden in the 1990s and 2000s to test alternate hypothesis for the first strategy of redefining the rational voters calculation to allow voting. The alternative hypothesis is the original model described by Larcinese (2005) where voting occurs because political infor-mation has private value and acquiring political inforinfor-mation creates a political interest. In Larcineses specification high income or net worth individuals will acquire this information and thus vote more often. A contrary interpretation of this is that it is equally likely that low income individuals has an incentive to acquire this information and then start to vote. The literature on the effects of income on likelihood to vote is large and there is also a literature on the effects of exogenous lottery winning income on what party to vote for but there is no direct investigation into the effects of changing income on the probability to vote itself. The purpose of this essay is to test if a change in individual income between two elections leads to a change in turnout, and if yes, in what direction. The contribution of this essay is both to conduct a more credible test of the role of income in the voting decision along

(5)

the lines of what Persson (2014) did for education and to check if there is a role for rational voting as proposed by Larcinese (2005).

A first difference staggered panel set-up with individual survey and register data from the 1991-2002 Swedish National Election Survey’s (SNES) are used to investigate the effect of changing individual income on voting behaviour. The results shows a possible small negative effect of changing income on propensity to vote using OLS but this result is not robust to a logistic specification.

The outline of the essay is as follows: a primer on the Swedish electoral system is presented in section two, section three presents an overview of voting theory, the fourth section describes some previous work, the data is presented in section five, the methods in section six, the results in section seven and a discussion is found in section eight. Section nine concludes.

(6)

2 Institutional framework

In Sweden national elections are to the one chamber 349 member Riksdag parliament. Re-gular elections takes place on the 3rd Sunday of September every fourth year since 1994, and at the same time elections to the municipal and county councils takes place. Prior to 1994 the electoral cycle was three years and thus the preceding elections took place in 1991. Votes are cast by filling an envelope with the ballot. All eligible voters are automatically registered on the voting rolls and voting is also possible in a period prior to election days, early or mail voting. Votes are cast for parties. Prior to stepping into the voting booth there is an opportunity to take pre-filled party ballots or you can use a neutral ballot and write the party name on it. Preference for a particular individual can be noted on the party list ballot but there is no possibility of voting for a candidate without at the same time voting for a party.

The objective of the seat distributing mechanism is proportionality - i.e. a 20 percent vote share should lead to a 20 percent seat share. The exception to this is the 4 percent party vote share threshold for entering the Riksdag. An exception to the exception is that a 12 percent vote share in an electoral district also allows seat representation even if the national vote share is below 4 percent. Seat proportionality is attempted to be achieved by having only 310 of the 349 seats represent the voters of the 29 electoral districts and having 39 of the districts as leveling seats (utjämningsmandat). After counting the votes after an election the fixed seats are first distributed by a modified Sainte-Laguë method (divisors are 1.4,3,5,7,. . . ) then the 39 leveling seats are used to achieve further proportionality. Each electoral district (valkrets) is further divided up in local districts (valdistrikt) totaling over 6000 in the most recent election. Each of these most local districts has its own polling station so that there is roughly one polling station per 1000 voters.

Since the current system with a single parliamentary chamber was introduced in 1971 and in the preceding half century of universal suffrage the party system has consisted of at least five parties in parliament. During this time it has often been the case that one party has been near the 4 percent threshold. Opinion polling is now regularly done and published in the newspapers in the period leading up to the election so voters have a the possibility of having a good idea of the party support strength when making up their own minds. However exceedingly few if any opinion polls published are done at the local level so there is little possibility of making precise seat estimates for individual parties prior to voting.

(7)

is based on parties and that the system is highly proportional. For the individual voter facing a voting decision in this set-up the physical aspects of voting are not large but the information requirement to make a decision in a multi-party system might be comparatively large compared to a two party system.

During the period studied Sweden was the first industrialized country to suffer a major financial crisis since the 1930’s. The acute phase of the banking crisis took place in the autumn of 1992 when the Riksbank famously set it’s policy interest rate at 500 percent in an unsuccessful effort to try to save the recently implemented currency-peg (Englund, 2015). Unemployment was at 2.8 percent during the first election included in the sample in 1991, by the next election in 1994 the unemployment rate had risen to 8 percent. Only by the next election in 1998 was the unemployment rate below 7 percent for the first time in over 5 year (SCB, 2019a). Incomes followed a similar, if less dramatic, path as can be seen in Figure 1. Aggregate turnout remained high during 1991 and 1994 but by 1998 it had decreased by over 5 percentage points to the low 80’s. There were two changes of government during the period. The Social democratic party lost power in 1991 to a center right Moderate party led government but then the Social democrats with minor party support regained power in 1998. In sum the political-economic situation of the case studied was in something of a period of flux with the situation leading to decreasing turnout and incomes generally falling and then rising.

(8)

80 82 84 86 88 R iksd a g t u rn o u t % 140 160 180 200 220 Me d ia n i n co me 1 0 0 0 s 2 0 1 7 SEK 1990 1995 2000 2005 2010 Year

Median income 1000s 2017 SEK Riksdag turnout %

Disposable median income per consumption unit. Official turnout figures. Source: SCB.

(9)

3 Theoretical background

In 1929 after more than 40 years of intense formal theoretical development in economics since Walras (1883) Harold Hotelling presented a speech before the American Mathematical Society in New York (Hotelling, 1929). Hotelling (1929) noted that after the work of Francis Ysidro Edgeworth in formalising the Bertrand model of competition (Edgeworth, 1899) there could seem to be little need for further work in competition theory for the case when there are a few sellers of a good. Citing work by Piero Sraffa (1926) that a market is often geo-graphically divided into quasi-monopolistic sellers Hotelling went on to introduce a location model of competition. In the Hotelling Location Model there are many otherwise identical consumers of a good that are uniformly spread out over a geographical line of fixed length where the consumer prefers close sellers. Firms enter the market and compete not with qua-lity characteristics but with prices and placement along the line. If we assume two producers who each enter the market at the far ends of the line with prices p1 and p2 respectively then

Hotelling’s Law says that in equilibrium the firms will “meet” at the halfway point of the line and p1 = p2. After presenting his formal analysis and his own intuitive interpretation of this

result Hotelling (1929) notes in passing that this result reminds him of how party politics in the United States during this time played out.

In 1948 Duncan Black (1948) set out to model small group decision making by voting. Black’s primary interest was to understand economic decisions taken by for example labour unions and corporate boards. Congleton (2002) notes that prior to Black (1948) in his work, which advocated for majority-rule, Condorcet (1785) had mentioned that there were pivotal voters but Condorcet had no formal model for predicting who would take on that role. In the Black model of majoritarian voting on a single issue domain with single peaked preferences (each agent has one ideal point and choices further away from that ideal point are less preferred) the median voter along the issue domain emerges as the pivotal voter and the policy preferred by the median voter will be adopted by the group after voting.

Duncan Black had mentioned electoral politics as a possible application of his model but before his book on the topic was published in 1958 (Black et al., 1958) Anthony Downs (1957a) had made significant further expansions to this line of models. The Downsian model of political competition as it would be come to be known was described in detail in An Economic Theory of Democracy (Downs, 1957b). Öhrvall (2018) calls the publication of Downs seminal book"(Öhrvall, 2018 p.6) of having had a fundamental importance for the theoretical study of voter turnout. The model developed by Downs is a full economic model

(10)

of electoral politics with rational optimizing behaviour by three different types of agents, voters (who cares about policy outcomes), incumbent politicians or political parties (who wants to be re-elected), and a challenger politician or political party (who wishes to become elected). In Downs’ framework the opposing political candidates, c 2 C, both choose among the pure strategies sc 2 SC that are the same and available to all candidates. These strategies

are referred to by Downs as a “party ideology”(Downs, 1957b p.98) and are assumed to be agreed by all to be one dimensional, left to right, showing its direct descent from the Hotelling Location Model (Downs, 1957b p.115). Voters have a fixed preferred point on the ideological dimension and in the basic set up of the model they will always vote for the candidate that offers them the in their eyes best policy. However the more interesting case for the purposes of this essay is the model adaption featured in chapter eight when voters face a trade off between voting and abstaining.The decision of whether or not to vote in an election is taken after a cost benefit analysis by the voter. A voter will take into account the benefits to her/him of voting through the electoral process changing its policies to the interest’s of the voter as a result of the voter participating in the process. However since the voter is a strict rational agent it will vote only if pB C > 0 where p is the probability of your vote affecting the outcome, B is the difference in policy that a political change would bring about, and C is the cost of voting. Downs points out that the cost of voting for the general voter may be low, mainly consisting of time costs in performing the physical act and the time cost in acquiring information beforehand. At first glance the benefits to voting might seem obviously to be higher than the modest time investment needed to cast a vote because changes in government policy can greatly affect a persons welfare. However it is not enough for the electoral decision to have good consequences for you and for you to have voted for it for it to be worth it to vote.

The key problem in voting theory is that the vote must be or at least have a reasonable chance of, being decisive. A decisive vote is the marginal vote that decides whether candidate A or candidate B gets elected. Strictly speaking any vote cast for the winner that is above and beyond a one vote margin is in some sense “wasted”. This is the paradox of voting that Downs explicitly mentions, elections should be decided by a much smaller percentage of voters because the outcome does not change above and beyond this level. To solve this dilemma Downs introduces another benefit of voting, DDowns, that is unrelated to the policy outcome

but instead captures a perhaps more long term benefit of voting. This Downs argues makes it rational for some voters to take the relatively small cost of keeping democracy running. This solves the apparent paradox that voting should not be as widespread as it is if voting is rational but at the cost of introducing into the model a variable that is not as theoretically

(11)

precise. Riker and Ordeshook (1968) built on Downs and Tullock (1967) rational voter theory but criticised the Downsian rational voting model as relying on a set up that would require overcoming a large collective action problem (Olson, 2009). Instead of assuming the kind of long term rationality advocated by Downs Riker and Ordeshook stressed the importance of subjective experiences and allowed for voting to give potential voters a mental reward, for example by fulfilling his or her civic duty to vote. This could also better explain behaviour but again at the cost of theoretical imprecision since this kind of rational theory could explain any behaviour (Verba, Schlozman and Brady, 1995).

Fiorina (1976) describes another possible solution to the voting paradox puzzle. In this model voters are assumed to have a “taste” for voting by using their vote as a signal about their preferences regarding different policies. Fiorina describes this as something resembling partisanship, that voters have a preference for one party winning over the other irrespective of the policy difference between them. Voting is done over and above the standard, instrumental, rate to do this. I.e. will vote if pB C+DFiorina > 0. The problem from a strictly theoretical

point of view is again the same for the other psychological and sociological models, including them might be plausible but there risk of creating a model that it is not possible to falsify (Blais, 2000).

In To Vote or Not to Vote? Blais (2000) hypothesizes that the voter performs a mental calculation of the efficacy of voting but that the voter often performs non-rational cognitive mistakes in estimating the probability of her or his own vote affecting the outcome. Voting is thus not a rational affair but an outcome of arithmetic mistakes that could have positive social effects. In a sense the Blais model is that the collective good is provided for by individual mistakes whereas the full Downsian model where voting takes places sees it as enough voters realise that without their effort the collective good will not be provided for unless they “take the hit”.

In “An Economic Theory of Democracy” Downs spends six out of sixteen chapters directly defining and explaining the role of uncertainty and information costs in his model. Downs seems to both see uncertainty as something of an Achilles heel of democracy, for instance on p. 79 he writes “imperfect knowledge allows the unequal distribution of income, position, and influence -which are all inevitable in any economy marked by an extensive division of labour-to share sovereignty in a realm where only the equal distribution of votes is supposed labour-to reign” and something that makes democracy possible in the first place because with certainty the government would simply loot as much as it could before it would be surely defeated due to the so called Arrow problem (Arrow 1950). The Arrow problem Downs refers to in this

(12)

case is that no non-dictatorial electoral rule can stably satisfy complex preferences so each election will be a cycle of wins and losses from equally non-preferred candidates.

Larcinese (2005) picks up this thread from Downs about the importance of information costs and formulates a model of voting that is, if not completely economical as opposed to psycho-logical, at least more falsifiable. Larcinese argues that the model developed by Meltzer and Richard (1981) that says that the size of government is a function of the income difference between the median voter and the median income in the economy fails to predict empirical realities. The levels of government redistribution that Meltzer and Richard predicts for the high levels of turnout that occur in the developed world simply does not occur (Lindert, 1996). Political information as such, i.e. of the type “who will win the next election?”, Larci-nese argues could be an important factor driving turnout behaviour. In the LarciLarci-nese model political information of that kind, positive predictive statements and detailed information about for instance the technical aspects of the tax regime, could be valuable to activities ot-her than voting in elections. For example Larcinese mentions the benefit of knowing whetot-her Forza Italia would win the 2001 election or not because the election manifesto contained va-rious tax reductions on investment and donations. If you were someone who faced a decision whether or not to invest or give a gift some knowledge of the probability that Forza Italia would be able to form a govenment coul have private returns. What lets Larcinese tie this to the Meltzer and Richard debate and to the Downsian voting paradox debate is his assump-tion that acquiring this kind of informaassump-tion is more valuable to well-off individuals. Higher income (or wealth, formally “full income” [Larcinese, 2005, p. 6] or the endowment) indivi-duals will then acquire more of this information and be more aware of policy suggestions by the political parties. Once aware of the policy difference between parties or candidates it is argued the step to acting on this acquired information in the voting booth is relatively small as the mental cost of voting has already been borne. An increased dynamic effect of this would be that the parties will be aware that this particular group of citizens is especially cognizant of the parties’ policy pronouncements and that the parties would then tailor their policy proposals more to the interests of this population. This would give an even larger incentive to well-off individuals to acquire political information and vote.

However Larcinese’s model of voting behaviour through information acquisition and income changes could still be true even if the effect is not driven by well-off individuals. Instead of thinking about the possibility of a party forming the government and implementing a tax reduction policy that affects investment decisions it is easy to imagine the possibility that a government would institute changes to the unemployment, sickness, or social insurance schemes which could also call for private decisions such as the purchasing of private insurance.

(13)

In those cases as well it could be argued that political information has private returns but it would be driven by a response at the opposite end of the income distribution. In this case as well information gathering for private reasons could potentially cause political interest and then voting. Larcinese argues that the private returns on political information for well-off individuals are absolutely higher because of their larger investment portfolios and tax optimisation possibilities but if utilities are convex the returns could be relatively more important for the less well-off because a larger share of disposable income is affected by government transfer policies.

The imagined causal chain in the general form that leaves open the possibility of both a positive and a negative effect then looks like something like this:

Income! Information ! Knowledge ! Interest ! V oting

The perspective of all of these theories is that voting takes place in a single member majority electoral district. The case studied in this essay is as described in section two a proportional representation (PR) electoral system. The difference from a theoretical point of view with a PR system is that on the surface there are many more possible parties to vote for due to lessening the constraints of Duverger’s Law (Durverger, 1959) allowing for more parties to “survive” competitive pressures over the long-term. However parties are still constrained by coalition formation such that they either have to plan coalitions beforehand and lessen their ideological diversity ahead of the election or compromise on promises after it. In either case the need to achieve stable coalitions will significantly decrease the real “political real estate” that is available to parties and thus there is even in PR a mechanism that acts to reduce the potential choices of voters. This in turn will make the voters potential benefits of either party being elected is less than a simple PR world without coalition formation would give them. There is though still probably more ideological diversity in a PR system and this would in itself lend a helping hand to rational voting because a larger difference in policy between the possible the exact parties that make up the government leads to a larger potential benefit of voting. On the other hand there is in PR systems with two relatively stable party groups a linear pattern between votes earned and probability of winning. In single member district systems the geographical concentration of votes between seats matters and because of that the number of seats earned is not a stable linear function of votes earned. Sweden during the period studied was such a PR case with two relatively stable party groups vying for power and it is not clear that from the voters point of view the increased options should lead to a less complicated voting calculus even if precisely determining the probability of being decisive is hard in a PR system.

(14)

4 Previous studies

The study of elections and voting is a large academic field. A search on the Web of Science Core collection database for the topics ”voting” + ”elections” yields over 12,000 results of which over 10,800 are published articles. Adding the topic “individual*” still returns over 1600 articles. In the short review of the voting literature in their 2016 monograph Svenska Väljare, Swedish Voters, Oscarsson and Holmberg (2016, p. 48) states that few areas within political science are so thoroughly studied as political participation"(my translation). They go on to explain that voting has been studied as a key form of political participation for now over 80 years ever with one of the first important works being Swedish political scientist Herbert Tingsten’s 1937 monograph Political Behaviour. In that work Tingsten (1937) develops a formal law, the law of dispersion, which states that as voting turnout increases the level of social inequality in turnout necessarily must decrease.

The voting literature took a large leap with the 1960 publication of The American Voter which was the first large work based on empirical findings in the American National Election Studies (ANES). In this work the authors argued that votes are cast primarily for, often inherited, partisan affiliations and that unaffiliated voters are often some of the least informed (Campbell et al., 1960), i.e. voting is not a rational act because voters are not even trying to be rational. Other political scientists using the same or similar data source argued otherwise and one who responded, V. O. Key, Jr., said: “Voter’s are not fools” (Key 1966 p.150). Since then the political science literature on explaining voter turnout has expanded in many directions considering the effects of electoral institutions (Grofman and Lijphart, 1986), an individuals social environment and context (Huckfeldt and Sprague, 1993) and the national political polarization (Rosenstone and Hansen 1993). Oscarsson and Holmberg (2016 p.10) sees three main schools of though for explaining voting behaviour: the historical-sociological explanations of Lipset and Rokkan (1967) emphasising class, the more psychological Michigan school following Campbell et al (1960), and the rational economic Downsian perspective used by for example Fiorina (1981) to explain voting as a retroactive “check” on bad politicians. For a review of the macro level literature of turnout predictors see Geys (2006) and also Cancela and Bays (2016). When it comes to individual specific explanations the rich data gathered in projects such as the pioneering ANES, and the second such large national project the one in Sweden SNES, has enabled researchers to compare results of cross sectional regressions in many contexts to get a more consistent understanding.

(15)

90 studies on that topic published in the top 10 journals in political science between 2000 and 2010 (Smets and van Ham 2013). The articles included in the meta-analysis were majority, 82 percent, self reported turnout versus the more reliable validated turnout from voter rolls (Smets and van Ham, 2013, p.5) and were generally of the cross sectional kind. The results from the meta-analysis were that many factors influence turnout at the individual level but the number of different variables under the category rational choice (58 percent) were more successful than the number of variables under the categories mobilization (41 percent), soci-alization (45 percent), psychological (47 percent), and resources (39 percent). The resource category was the largest category with 22 variables and the most successful in that category were age, generation, education, socio-economic status, income and marital status. Smets and van Ham notes that age, education and income are some of the most studied turnout predictors that are consistently found to have an effect whereas the most studied variable, gender, does not at the general level have an effect.

Income is thus systematically connected to voting across many contexts. However the pro-blem with interpreting this is that all these studies, as far as I understand, use only cross sectional variation to explain voting behaviour by income. As Persson (2013a) and Persson (2013b) has showed for another very important predictor of voting, education, it is not clear that variables found to have statistical predictive power in a cross sectional analysis actually represents a true causal effect. Persson (2014) shows that absolute education level does not matter for voting but that an increased relative education level is a strong proxy for relative social status and that it is this which determines voting. For other turnout predictors such as mobilization drives and information campaigns there exists a large literature of experi-mental studies (Gerber and Green, 2000, Lassen, 2005 and Green et al., 2013). Powdthavee and Oswald (2014) studies a natural experiment in the form of UK lottery winners in a large household sample to find that winning the lottery, an exogenous shock to income, tend to make people to start voting for the political right. Bagues and Esteve-Volart (2016) instead looks at the macro level outcomes of exogenous lottery winnings being geographically corre-lated and thereby increase incumbent vote shares in winning regions. Other studies have used other methods that are more credible than a simple cross sectional analysis to investigate the effects on political preferences and ideology from specific individual economic conditions such as job loss (Margalit, 2013 and Martén, 2019). After a database search using different keywords has not succeeded in finding a similar study to Persson (2014) but for income.

(16)

5 Data

Data source is the Swedish National Election Studies (SNES)1 panel data where individual

voters are in a staggered way followed across two elections from the elections of 1991 to 2002. SNES has since 1956 surveyed a nationally representative sample of the electorate about the respondents political preferences and voting in current and previous elections. In addition to survey responses to these questions about substantive materiel there are responses to questions about background information. In further addition to this Statistics Sweden has added register data on on (pre-tax) income, if voted in the election, age, gender and marital status.

The response rate for the surveys was 73 percent in 1991, 80 percent in 1994, 82 percent in 1998 and 70 percent in 2002. Since 1979 SNES allows respondents to complete a shortened form to help increase response rates. Some of the individual characteristics of interest are only found as answers in the survey and some of those questions were not posed to those respondents who only answered a shortened form so the response rate for individual questions can lower than for the survey as a whole. The lack of full cover of the survey responses is a serious issue and all results come with a caveat that this is something that simply can’t be assumed away. However it should also be noted that a large majority of respondents does in fact answer the survey and since the questionnaire is quite substantive the selection bias in response should be less than if it was a very brief survey since it’s reasonable that many more (and different types of) people don’t find the time to answering it. The variables that come from the register that is matched to the survey has a near perfect “response” rate to the survey.

The panel approach that is of interest to this study starts from the 1991 election where the rotating panel is composed of individuals who are interviewed in two successive elections and includes register data on individual pre-tax income. The data on pre-tax income comes from the tax registry via Statistics Sweden and refers to the income reported to the tax authority the previous year. The individuals participation in the election is available both as a survey response (not used) and from the Statistics Sweden registry where the source is the electoral roll (used). Income is in 1000’s SEK and deflated using the Consumer Price Index (CPI)

1Data and tables in this essay has partially been made available to the author by SND, Svensk Nationell

Datatjänst. The materiel in SVENSK VALUNDERSÖKNING 1991, 1994, 1998 and 2002 where initially collected for research projects at the Department of Political Science, University of Gothenburg under the direction of Mikael Giljam (1994), Sören Holmberg and Mikael Giljam (1994), Sören Holmberg (1998), and Sören Holmberg and Henrik Oscarsson (2002). Neither SND nor the primary researchers bear any responsibility for any of the analyses or interpretations carried out in this essay.

(17)

since the Labour Cost Index was not available for the period. The sample includes the SNES surveys from 1991, 1994, 1998 and 2002 since no raw income data was available for 2006 and 2010 that were also available. Each election around 3000 individuals are surveyed but since half of those were interviewed during the previous run from the five different surveys you can assemble a rotating panel that includes around 1200-1800 individuals from each "election wave". I.e. the first 1991 election panel is made up from individuals re-interviewed since their first interview in 1988 and a new sample of the electorate that is also again in 1994. But since income data is only available from 1991 the first panel begins there. The complete merged dataset has 13564 observations with 6335 unique individuals.

The education variable follows Martén (2019) in a three category setup for highest completed education. There is one category for those with less than high school (gymnasium), another category for those with any of the different kinds of high school and lastly one category for those with university or comparable education. Marital status (civilstånd) is transformed into a binary currently married or in a domestic partnership alternatively widowed, divorced or single. This variable also from the register. In addition to actual age, age squared a dummy for age 75 and above is included in some specifications to better capture non-linear age related behaviour. The labor market group variable is used with no altered coding in the nine possibilities. The categories are employed, doing relief work, in a labour market program, unemployed, old age pensioner, early retirement, doing housework and being a student. But to get a more detailed look at the labor market situation a student dummy and a employed versus unemployed binary is created from this question. To study a possible mechanism survey responses to questions about news consumption and political interest are coded to binaries. For news consumption “quite often” and “every day” is coded as 1 and “some times” (“nån gång”) and “never” as 0. For political interest “very interested” and “pretty interested” (“ganska intresserad”) are coded as 1 and “not particularly interested” (“inte särskilt intresserad”) and “not at all interested” (“inte alls intresserad”) coded as 0. Gender presented for summary statistics, male gender coded as 1. Covariate balance on the independent the outcome variable is presented in Table 1.

(18)

Tabell 1: Covariate balance

(1) (2)

Did not vote Voted

mean max min sd mean max min sd

Gender .5400956 1 0 .4985221 .5005993 1 0 .500021

Income CPI deflated 287.3882 26974 0 937.6467 404.8484 29901 0 813.2549

Education 3 cat 1.734082 3 1 .7536242 1.81365 3 1 .7901893

Unemployed .071934 1 0 .258531 .039874 1 0 .1956749

Part/full time worker .1704958 1 0 .3762953 .2023442 1 0 .4017723

Student .3426573 1 0 .4754299 .3100033 1 0 .462571

Labor market group 3.021088 9 1 2.705262 2.807925 9 1 2.622778

age 43.67074 83 18 17.86107 46.36661 84 18 17.01967

In relationship .2278279 1 0 .4195424 .5083469 1 0 .4999517 Over 74 years old .0727562 1 0 .259805 .0557363 1 0 .2294216

Reads news .2691808 1 0 .4438227 .4779954 1 0 .4995473

Political interest .3308458 1 0 .4708105 .5360256 1 0 .4987324

(19)

6 Empirical model

The most straightforward way to study the influence of income on voting would be to perform a cross sectional ordinary least squares, OLS, regression in a single year of individual income on individual turnout. This approach would be feasible and easy to interpret but it would suffer from omitted variable bias, OVB, if there is a factor that both influences an individuals’ income and her likelihood of voting. As discussed earlier there are numerous such factors. Another, ideal, way to study this relationship would be to conduct an experiment with a truly exogenous change in income and then observe the turnout consequences. However performing such an experiment is not feasible in this setting and would possibly raise ethical concerns. The approached used in this essay is a second best attempt at capturing the effect that would be present in such an experiment by using a panel set-up following individuals across time.

The main concern with a cross sectional OLS for a given election year is the individual level OVB, the presence of a confounder. One can imagine the biasing factors as being of two kinds, fixed and time varying. Fixed factors are confounders that do not change over the life of a person. In this case a fixed factor could for instance be general personality type which possibly influences an individual’s income level and voting propensity but is relatively stable over time. Time varying OVB factors are those determinants of both income and voting behaviour that does vary over time. An example of this would be age since the very young typically have both lower incomes and are less likely to vote. The individual fixed effect set up is a second best attempt eliminating these fixed factors. The econometric model used is based on the approach used by Martén (2019) and is a first difference set up instead of fixed effects, both similar.

Consider if Yit is a binary outcome variable for voting for individual i in election year t, Xit

is the income of individual i in election year t, ci are fixed individual characteristics that

biases the estimate of 1, Zt is individual year effects and uit is the residual predictors of Y

not correlated with X.

Yit = 1Xit+ ci+ Zt+ uit, t = 1, ..., T, i = 1, ..., I

(20)

Differencing both equations, gives:

Yit = Yit Yit 1 = 1Xit+ uit

If there were only two elections in the sample the first difference (FD) and the fixed effects (FE) estimator would be completely the same but since the data is a rotating panel a first difference approach gives more leeway to include also non-differenced variables as covariates. Martén uses a FD model to study how changes in employment status affect individuals preferences for economic redistribution. That study uses the same data source, panels from the Swedish national election study, which is why this study follows the coding procedure of that study to limit the degrees of freedom of the investigator. A linear time trend is used to capture long term trends in voting and incomes that can bias the estimate.

Yit = 1Xit+ Zt+ uit

Even if FD and a time trend can eliminate individual fixed and macro trend biases that still leaves the room for time varying individual effects to affect our estimate. To eliminate this bias we in principle need to include as covariates in the model all of these time varying factors that have an effect on both income and voting. The most important covariates to include are those that affect the bias and those are the ones that are time varying, are associated with income and have an effect on voting turnout. An example of an important covariate like this is age. Voting has a strong habit formation process to it and every time you do vote there is a strong tendency for you to continue voting the next time. So increasing age will definitely change over time and most likely affect voting at least in some age groups. Since Martén (2019) studies political attitudes and not the voting decision, and because her independent variable is job loss and not income, the individual covariate mix should in theory differ between those set ups. However since the two outcome variables are conceptually related and a tying of the investigator’s hand is seen as a positive the same individual covariate mix will be used in this study. Additionally Marténs focus on labour market variables to isolate the effect of job loss is probably a good idea in the light of Persson (2013a) who showed that it is social economic status and not education per se that causes voting. In view of this it is a good idea to not lean too exclusively on education along with age and relationship status as a core predictive set, especially in a panel set up. In sum a Martén differenced

(21)

covariate individual covariate mix, V, is used in the standard specification both because it seem reasonable also for this study but also to limit analyst degrees of freedom.

Yit= 1Xit+ Zt+ V + uit

Where V includes age, age squared, education, unemployment status, part time worker status, student status, general labour market group status, and relationship status. All of these except the age covariates are differenced (since the age difference is fixed). The effect of age on voting and income is as previously mentioned the archetypical time varying confounder and since there is good reason to believe age has a parabolic relationship with both income and voting age squared is a necessary covariate. Education is included in the model despite the discussion previously over whether that effect is causal or not. For our purposes if it has a direct causal effect or if it is just the proximate agent of social status does not matter in our case because both an absolute causal effect and a relative education non causal effect is capturing a factor that biases our estimate. Unemployment, part time work and labour market group are also important covariates, since the status is time varying and has an effect on voting according to Öhrvall (2012) and Smets and van Ham (2013) for the international literature, and is correlated with income in many cases including this sample. Student status is strongly negatively correlated to income in this sample but the evidence for it’s effect on voting is mixed, partly due to the difficulty in isolating it’s effect from that of the interaction effect of age and relatively high socio-economic status along (Niemi and Hanmer, 2010) but inclusion is warranted because of the same reasons as for education. Even if not directly causal inclusion of differenced student status is has the possibility to pick up time varying bias in regressing income on voting. Relationship status is an important variable to include both for it’s own strong effect on voting (Oscarsson and Holmberg, 2016) but also as a proxy for other social characteristics of individuals that can vary across time and are likely to be correlated with voting, income and relationship status. In particular an individuals social network has an effect on voting (Öhrvall, 2012) and a change in relationship status is likely to be correlated with changes in an individuals social network.

Another note on comparisons with Martén (2019) is that since she studies preferences which are given in an interview conducted over a period over several months it introduces problems with using income data that are from tax data that cover a fixed time period every year. However with voting this is not a problem since every election in the sample was conducted

(22)

during the same time each election year compared to the tax year.

The distribution of voters and non voter for all individuals in all the panels together is 11681 (86.12 percent) voters and 1883 (13.88 percent) non voters. This puts it close to the region where the linear and logistic models no longer correspond strongly (and thus use the linear model for ease of interpretation) and where you possibly should switch to a logistic model to better fit the underlying distribution (Von Hippel, 2015) and Hellsvik, 2009). For this reason both models will be used and presented. This has drawbacks in terms of introducing another layer of difficulty in interpretation but hopefully with gains for transparency. Standard errors are clustered at the individual level in all models to handle within subject correlation of errors across elections.

In order to both handle problems arising from the right tail skew of the income distribution and help in interpretation the income measure is transformed using the natural log. To further test reliability of results additional models were run with models also including a dummy for persons over the age of 75 and using the non differenced income as predictor as well as testing dropping some of the labour market covariates. Just to go slightly further than just the standard reduced form estimate of income change on voting behaviour change a primitive investigation into the income change to information acquisition, and from information to political interest. and finally from interest to voting is done.

A robustness check with absolute income included as a control is also performed. This is done to protect from possible bias from the fact that though we are analysing changes in income the effect of the four year difference in income is potentially overwhelmed by the stronger influence of an individuals actual absolute income level. Additional robustness checks are done with altering the socio-economic covariate inclusion strategy.

In addition to the main analysis a small tentative investigation into a possible mechanism is performed. This is done by using the base linear model but first regressing income change on news consumption change and then regressing changing news consumption on changing political interest. Finally changing political interest is regressed on voting behaviour.

(23)

7 Results

Table 2 and Table 3 presents the main results, the linear regressions in Table 2 and the logistic regressions in Table 3. Independent variable in all models is the individual differenced turnout. In the first column of Table 2 the results for the model with only age covariates is presented along with the time trend. Here you can see that an income change has a statistically significant negative effect at the 1 percent level. When the education change variable is introduced in the second model in the second column the income change effect is still significant but only at the 5 percent level and the effect is reduced. Again when further (differenced) social and economic characteristics are added in the third model the income change variable is still significant at the 5 percent level but the explanatory power of the model is lower. When in model four the differenced relationship status dummy is added to create the base model the income change variable is statistically significant with a negative point estimate at the 5 percent level and the explanatory power of the model is increased. The effect is of such size that if you were to increase income by 100 percentage points the probability of voting would be reduced by about 6 percentage points.

The fifth and final model specification adds the 75 and above age dummy which does not alter the estimate. The same covariate progression is used for the logistic regression models present in Table 3. In none of the logit model specifications is the income change variable statistically different from zero. In addition to this difference the most interesting finding is that using the logit set up the estimates are not only less precise they are also of the opposite sign. All logit model estimates of the effect of differenced income on differenced voting are positive compared to all negative using linear models. In all of the logit models where they are included both age and differenced education are significant at the 5 percent level. I.e. the difference between a linear and a logit model leads to different variables “capturing” the predictive power. But it is also noteworthy that the pseudo R2 for the fourth model, the

base model, in logit specification is roughly 3.5 times the adjusted R2 of the linear model

with the same covariates.

The results of the preliminary investigation into the possible mechanism behind the the result is showed in Table 4. In this analysis no step of the mechanistic progression from an income change to changing news consumption and on from that to a change in political interest and voting is significant at any level.

In the appendix a robustness check for the main income-voting analysis is presented with alternate specifications and with absolute income as a control. The first model in Table 5

(24)

Tabell 2: OLS models

(1) (2) (3) (4) (5)

Dvoted Dvoted Dvoted Dvoted Dvoted

Dlnincome -0.0079⇤⇤⇤ -0.0061-0.0061-0.0064-0.0063⇤ (0.0016) (0.0025) (0.0027) (0.0027) (0.0027) age 0.0013 -0.0032 -0.0039 -0.0028 -0.0030 (0.0022) (0.0031) (0.0033) (0.0033) (0.0036) age2 -0.000022 0.000027 0.000034 0.000025 0.000027 (0.000022) (0.000029) (0.000031) (0.000032) (0.000035) Time trend -0.0043 -0.013 -0.0071 -0.0079 -0.0078 (0.0065) (0.011) (0.012) (0.012) (0.012) Deducation 0.0041 0.0043 0.0048 0.0048 (0.016) (0.016) (0.016) (0.016) Dunemployed -0.023 -0.024 -0.024 (0.039) (0.039) (0.039) Dpart 0.032 0.031 0.031 (0.020) (0.020) (0.020) Dstudent 0.00071 -0.00068 -0.00068 (0.0076) (0.0074) (0.0074) Dlabgroup 0.00066 0.00059 0.00059 (0.0021) (0.0020) (0.0020) Drel 0.069⇤ 0.069⇤ (0.028) (0.028) Dover74 -0.0066 (0.031) Constant -0.023 0.10 0.10 0.075 0.078 (0.056) (0.079) (0.089) (0.090) (0.093) Observations 4007 1850 1733 1733 1733 Adjusted R2 0.0073 0.0048 0.0033 0.0087 0.0081

Standard errors in parentheses

(25)

Tabell 3: Logit models

(1) (2) (3) (4) (5)

Dvoted Dvoted Dvoted Dvoted Dvoted Dvoted Dlnincome 0.0085 0.018 0.020 0.020 0.021 (0.015) (0.025) (0.028) (0.028) (0.028) age -0.11⇤⇤⇤ -0.075-0.076-0.081-0.092⇤ (0.019) (0.032) (0.036) (0.037) (0.039) age2 0.00097⇤⇤⇤ 0.00050 0.00050 0.00053 0.00067 (0.00019) (0.00033) (0.00037) (0.00038) (0.00040) Time trend 0.046 0.30⇤⇤ 0.300.300.31⇤ (0.065) (0.11) (0.13) (0.13) (0.13) Deducation -0.42⇤⇤ -0.35-0.36-0.36⇤ (0.16) (0.17) (0.17) (0.17) Dunemployed -0.043 -0.048 -0.042 (0.37) (0.37) (0.36) Dpart 0.0084 0.014 0.017 (0.22) (0.22) (0.22) Dstudent 0.032 0.037 0.036 (0.075) (0.075) (0.075) Dlabgroup 0.043 0.044 0.042 (0.044) (0.044) (0.044) Drel -0.23 -0.24 (0.26) (0.26) Dover74 -0.62 (0.68) Constant 0.40 -0.76 -0.76 -0.66 -0.46 (0.45) (0.74) (0.85) (0.86) (0.89) Observations 4007 1850 1733 1733 1733 Pseudo R2 0.015 0.033 0.031 0.032 0.033

Standard errors in parentheses

(26)

in the appendix is for comparison the same as model four in Table 2. In the second model with absolute income included as a variable the differenced income variable effect size is increased and the estimate is still statistically significant. In the third row in Table 5 showing a model without labour market characteristics but with full age covariates the estimate is very similar to the base model. Since the base logit model was not statistically significant it is to be suspected that the alternate specifications presented in Table 6 are not statistically significant either. It can be noted that the point estimate for differenced income with the absolute income variable included is also higher as in the linear models however the standard error of this estimate is very large.

(27)

Tabell 4: Possible mechanism, OLS

(1) (2) (3)

Dnews Dpolinterest Dvoted Dlnincome 0.0034 (0.0045) Deducation -0.028 0.043 0.0046 (0.028) (0.043) (0.016) Dunemployed 0.065 0.078 -0.071 (0.043) (0.057) (0.044) Dpart -0.061 -0.035 -0.0021 (0.032) (0.046) (0.020) Dstudent 0.021 0.075⇤⇤⇤ 0.0094 (0.013) (0.015) (0.0081) Dlabgroup -0.00043 0.013⇤⇤⇤ 0.0017 (0.0029) (0.0031) (0.0023) Drel 0.065 0.019 0.063⇤ (0.036) (0.049) (0.028) age -0.0044 0.00038 -0.0037 (0.0049) (0.0070) (0.0039) age2 0.000039 0.0000027 0.000029 (0.000049) (0.000071) (0.000036) Time trend 0.031 0.12⇤⇤⇤ -0.014 (0.021) (0.034) (0.015) Dnews 0.11⇤⇤ (0.038) Dpolinterest 0.0062 (0.016) Constant 0.018 -0.43⇤ 0.13 (0.13) (0.18) (0.10) Observations 1466 1111 1236 Adjusted R2 0.0029 0.030 0.011

Standard errors in parentheses

(28)

8 Discussion

The results show the necessity of including time varying covariates such education to not bias the estimate of the effect of income change on voting change. A naive first difference estimator without any covariates produces a more precise and a 24 percent stronger negative estimate of the income change effect. This most likely shows the necessity of conditioning on time varying covariates.

The main threats to identification of the main effect studied is in the author’s opinion still bias arising from unobserved individual heterogeneity in the income, education, social network, and political behaviour nexus as well as bias from selective non-response to the survey. These problems are compounded in that education is measured through the survey. However since education is included in the shortened survey used to achieve maximum survey compliance the problem is not as bad as it could be. The sampling from the electorate to the survey is where the majority of survey non-response takes place as opposed to incomplete survey forms. Still, it is not known whether this is an unbiased sample of the reference population and it is possible that differences between survey responders and non-responders account for the effect found. The role of education itself is also a potential issue even if there would be full coverage from the survey. First of all the education measure used is only of three categories. For the period studied this does not seem like a large problem but there are of course large human capital and sociological differences within these categories. If the income data was available with a higher frequency the problem of separating out the effects of education and income on voting would be easier and a more credible estimate could be produced. Unless you turn to pure exogenous variation such as a lottery it is hard to find a “pure” treatment in the income domain. The problem with using that approach however is that for many countries it is unlikely that lottery ticket consumers are fully representative of the general population. This study by using differenced income thus trades some problems of internal validity to achieve good external validity, since the income measured is the full income experienced by voters at least for tax purposes.

When it comes to how to interpret the results the main issue is how much weight should be put on the linear results, small statistically significant negative effect of income change on voting change, versus the logit results, not statistically distinguished from zero. In addition to this the respective R2’s of the different model types were different favouring the logit

null result. Since including both model types in the analysis was a definite choice, plainly disregarding either once the results are known does not make sense. However this does not

(29)

mean that just because the different model results disagree that one can’t say anything at all about how one should judge the evidence. Since the main purpose of this essay is to provide a more confident and a more causal interpretation to the income-voting relationship total model predictive power is not as important as correct specification and statistical significance. This is because it would be very unlikely if all the cross sectional predictors of voting such as age, education and relationship status would due to unobserved heterogeneity and disappear in a panel set up with a differenced specification. I.e. some predictive power must still exist with the covariates and the question of how much explained variance each model produces is secondary to the estimation of economically and statistically significant effects.

The main difference between the model used by Martén (2019) and this essay except the main outcome variable of interest and the main independent variable is that Marten in-cluded geographical covariates in her analysis. This is done because there are systematic differences in attitudes towards redistribution between rural and non-rural communities and that this difference is correlated with local unemployment level and therefore an individu-als unemployment risk (her main independent variable under). And superficially the same could be true for this study. There is evidence that there are effects on voting by the local environment of an individual (Oscarsson and Holmberg 2016, p.50). There is also a clear relationship between income and for instance municipality size (SCB 2018). But the diffe-rence is that income is in some ways a broader phenomenon than unemployment risk since there is no reason to exclude causal paths from relocation to income change to voting change (perhaps through the private information acquisition mechanism). The difference with the unemployment status binary is that it because it is measured at the same time as the main outcome variable it is much more important to control for it. Since the income measure in this study is from the tax register, i.e. not measured at the time of the survey, and because the main outcome variable comes from the voting rolls there is much less possibility of a simultaneity problem.

Having said that why include any time varying covariates at all? The reason why the “social variables” education and relationship status are so important to include even in a differenced model is because it is much more likely that their effect on voting is completely separate from any of their potential effect on changing a person’s income. As an example imagine a change in an individual’s relationship status, like a divorce. The divorce has the potential to quite quickly change many of this person’s interactions with others, from seeing more of their friends and acquaintances to joining a new social grouping. All these changes has the possibility of shifting the individual’s mental framework on voting. If this shift also produces a change in income the estimate of the income change effect on voting will be biased.

(30)

As for the question of the relevance of the time and country sample studied here it’s clear that even if one accepts that individual income changes could have a negative effect on turnout in this sample it is difficult to say how general you think this result could apply. On the one hand Sweden during the period studied had an internationally very high voting turnout level (International IDEA, 2019) which could make Sweden an outlier in all comparisons. On the other hand the fact that voter turnout in Sweden is so high makes finding such a negative result of an income change for Sweden all the more surprising as you would expect a high overall turnout to be realised through many different positive effects. The time period studied is also both a positive and negative for ease of generalizability. The positive is that there is large amount of real variation in income due to macroeconomic fluctuations that enables us to more credible study it’s effects on voting. The negative is that this level of fluctuation is unusual and may not represent a mature industrialized democracy working as usual.

A final interesting finding is that in the second model of Table 5 where absolute income is included in the regression the effect size of income change roughly doubles but stays negative. The absolute income variable is not statistically significant from zero but the point estimate is positive. In the same model the unemployment dummy is still not significant suggesting the effect is not simply driven by an event such as this. Together this could suggest that higher income individuals vote at at higher rate than lower income individuals but that for those individuals that have not formed a stable voting habit, the marginal voters in the electorate, a decreasing income is more likely to bring you to the polls than an increase. If voting was done for purely irrational reasons and was simply a product of erratic psycho-logical misconceptions it would be hard to understand why there would be an association with voting behaviour and an income change. On the other hand if individuals truly followed a Downsian voting calculation for solely individually rational cost-benefit reasons it is also hard to see why changing incomes should lead to people beginning to start voting or quit doing the same. It is perhaps reasonable to see the voting decision as some kind of rational-irrational chimera, such as the miscalculations suggested by Blais (2000), with at least some marginal individuals whose political behaviours are driven by economic circumstances.

(31)

9 Conclusions

Through a first difference panel data model individual income increase between elections was found to possibly have a negative effect on voting that is small but empirically significant. The result is far from certain but plausible and interesting. That voting could in a sense be an inferior good would be a surprising result given the large cross sectional literature finding that higher incomes are associated with higher turnout. But even though the mechanism of political information for private reasons leading to voting failed to gain support in the simple analysis performed it is one possible way to explain the result. This line of inquiry could be expanded further using longer panel structure data. To get a more credible estimate of the main effect studied a larger sample period, more frequent income data and a set of socio-economic covariates with more coverage is needed.

(32)

10 References

Aggeborn, L. (2016). Voter turnout and the size of government. European Journal of Political Economy, 43, 29-40.

Arrow, K. J. (1950). A difficulty in the concept of social welfare. Journal of political economy, 58(4), 328-346.

Bagues, M., & Esteve-Volart, B. (2016). Politicians’ luck of the draw: Evidence from the Spanish Christmas lottery. Journal of Political Economy, 124(5), 1269-1294.

Beck, N. (1975). A note on the probability of a tied election. Public Choice, 75-79.

Black, D. (1948). On the rationale of group decision-making. Journal of political economy, 56(1), 23-34.

Black, D., Newing, R. A., McLean, I., McMillan, A., & Monroe, B. L. (1958). The theory of committees and elections.

Blais, A. (2000). To vote or not to vote?: The merits and limits of rational choice theory. University of Pittsburgh Press.

Campbell, A., Converse, P., Miller, W., & Stokes, D. (1960). The American Voter. University Chicago Press.

Cancela, J., & Geys, B. (2016). Explaining voter turnout: A meta-analysis of national and subnational elections. Electoral Studies, 42, 264-275.

Condorcet, M. D. (1785). Essay on the Application of Analysis to the Probability of Majority Decisions. Paris: Imprimerie Royale.

Congleton, R. D. (2004). The median voter model. In The encyclopedia of public choice (pp. 707-712). Springer, Boston, MA.

Downs, A. (1957). An economic theory of political action in a democracy. Journal of political economy, 65(2), 135-150.

Downs, A. (1957). An economic theory of democracy. Harper, New York.

(33)

Methuen.

Edgeworth, F. (1889). The pure theory of monopoly"., reprinted in Collected Papers relating to Political Economy. 1. Macmillan. 1925.

Englund, P. (2015). The Swedish 1990s banking crisis-A revisit in the light of recent experi-ence. In Riksbank Macroprudential Conference, Stockholm (pp. 23-24).

Fiorina, M. P. (1976). The voting decision: instrumental and expressive aspects. The Journal of Politics, 38(2), 390-413.

Fiorina, M. P. (1981). Retrospective voting in American national elections. Yale University Press

Gerber, A. S., & Green, D. P. (2000). The effects of canvassing, telephone calls, and direct mail on voter turnout: A field experiment. American political science review, 94(3), 653-663. Geys, B. (2006). Explaining voter turnout: A review of aggregate-level research. Electoral studies, 25(4), 637-663.

Green, D. P., McGrath, M. C., & Aronow, P. M. (2013). Field experiments and the study of voter turnout. Journal of Elections, Public Opinion and Parties, 23(1), 27-48.

Grofman, B., & Lijphart, A. (Eds.). (1986). Electoral laws and their political consequences (Vol. 1). Algora Publishing.

Harold, H. (1929). Stability in competition. Economic Journal, 39(153), 41-57.

Hellevik, O. (2009). Linear versus logistic regression when the dependent variable is a dicho-tomy. Quality & Quantity, 43(1), 59-74.

Holmberg, S., & Gilljam, M., Statistiska centralbyrån (SCB). (1995). Svensk valundersökning 1991. Version 1.0. Svensk Nationell Datatjänst. http://dx.doi.org/10.5878/002513

Holmberg, S., & Gilljam, M., Statistiska centralbyrån (SCB). (1997). Svensk valundersökning 1994. Version 1.0. Svensk Nationell Datatjänst. http://dx.doi.org/10.5878/002514

Holmberg, S., Statistiska centralbyrån (SCB). (2002). Svensk valundersökning 1998. Version 1.0. Svensk Nationell Datatjänst. http://dx.doi.org/10.5878/002515

(34)

institutio-nen (2014). Svensk valundersökning panel 1998-2002. Svensk Nationell Datatjänst. Version 1.0. https://doi.org/10.5878/002208

Holmberg, S., & Oscarsson, H., Statistiska centralbyrån (SCB). (2006). Svensk valundersök-ning 2002. Version 1.0. Svensk Nationell Datatjänst. http://dx.doi.org/10.5878/002643 Huckfeldt, R., Plutzer, E., & Sprague, J. (1993). Alternative contexts of political behavior: Churches, neighborhoods, and individuals. The Journal of Politics, 55(2), 365-381.

Key, V. O. (1966). The Responsible Electorate. Belknap Press

Lindert, P. H. (1996). What limits social spending?. Explorations in Economic History, 33(1), 1-34.

Lipset, S. M., & Rokkan, S. (Eds.). (1967). Party systems and voter alignments: Cross-national perspectives (Vol. 7). Free press.

Margalit, Y. (2013). Explaining social policy preferences: Evidence from the Great Recession. American Political Science Review, 107(1), 80-103.

Martén, L. (2019). Demand for Redistribution: Individuals’ Responses to Economic Setbacks. The Scandinavian Journal of Economics, 121(1), 225-242.

Meltzer, A. H., & Richard, S. F. (1981). A rational theory of the size of government. Journal of political Economy, 89(5), 914-927.

Mueller, D. C. (2003). Public choice III. Cambridge University Press.

Niemi, R. G., & Hanmer, M. J. (2010). Voter turnout among college students: New data and a rethinking of traditional theories. Social Science Quarterly, 91(2), 301-323.

Larcinese, V. (2005). Electoral competition and redistribution with rationally informed voters. Contributions in Economic Analysis & Policy, 4(1).

Larcinese, V. (2007). Voting over redistribution and the size of the welfare state: the role of turnout. Political Studies, 55(3), 568-585.

Lassen, D. D. (2005). The effect of information on voter turnout: Evidence from a natural experiment. American Journal of political science, 49(1), 103-118.

(35)

Oscarsson, H., & Holmberg, S. (2016). Svenska väljare. Wolters Kluwer

Persson, M. (2013). Does education cause participation in politics?. Göteborg Studies in Politics 134

Persson, M. (2013). Is the effect of education on voter turnout absolute or relative? A multi-level analysis of 37 countries. Journal of Elections, Public Opinion & Parties, 23(2), 111-133. Persson, M. (2014). Testing the relationship between education and political participation using the 1970 British cohort study. Political Behavior, 36(4), 877-897.

Powdthavee, N. & Oswald, A. (2014). Does Money Make People Right-Wing and Inegalita-rian? A Longitudinal Study of Lottery Winners, IZA Discussion Papers, No. 7934, Institute for the Study of Labor (IZA), Bonn

Tingsten, H. (1937). Political behavior: Studies in election statistics (No. 7). PS King. Tullock, G. (1967). Toward a mathematics of politics. University of Michigan Press.

Riker, W. H., & Ordeshook, P. C. (1968). A Theory of the Calculus of Voting. American political science review, 62(1), 25-42.

Rosenstone, S. J., & Hansen, J. (1993). Mobilization, participation, and democracy in Ame-rica. Macmillan Publishing Company,.

Smets, K., & Van Ham, C. (2013). The embarrassment of riches? A meta-analysis of individual-level research on voter turnout. Electoral studies, 32(2), 344-359.

Sraffa, P. (1926). The laws of returns under competitive conditions. Economic Journal, (36) 535-550. Reprinted in Sraffa, P. (2005). The laws of returns under competitive conditions. In Readings In The Economics Of The Division Of Labor: The Classical Tradition (pp. 216-233).

Statistiska Centralbyrån. (2018). Skilda världar? Det demografiskt delade Sverige. Demogra-fiska rapporter 2018:2

Statistiska Centralbyrån. (2019). Arbetskraftsundersökningarna (AKU)- Länkade data från 1987 och framåt. Viewed 5 June 2019

</urlhttps://www.scb.se/hitta-statistik/statistik- efter-amne/arbetsmarknad/arbetskraftsundersokningar/arbetskraftsundersokningarna-aku/pong/tabell-och-diagram/lankade-data-justerad-for-tidsseriebrott/lankade-data-fran-1987-och-framat/>

(36)

Statistiska Centralbyrån. (2019). Indikatorer inkomstfördelning 1975–2017. Viewed 5 June 2019 </urlhttps://www.scb.se/hitta-statistik/statistik-efter-amne/hushallens-ekonomi/inkomster- och-inkomstfordelning/inkomster-och-skatter/pong/tabell-och-diagram/inkomster–ekonomisk-standard-riket/indikatorer-inkomstfordelning-19752017/>

Statistiska Centralbyrån. (2019). Valdeltagande i svenska val efter valår 1973-2018. Viewed 5

June 2019 <https://www.scb.se/hitta-statistik/sverige-i-siffror/manniskorna-i-sverige/ valdeltagande-i-sverige/>

The International Institute for Democracy and Electoral Assistance (International IDEA). (2019). VOTER TURNOUT - REGIONAL COMPARISON. viewed May 23 2019 <https: //www.idea.int/data-tools/regional-comparison-view/521/40/>

Von Hippel, P. Linear vs. Logistic Probability Models: Which is Better, and When?, Statiscal Horizons, viewed 29 April 2019,< https://statisticalhorizons.com/linear-vs-logistic> Verba, S., Schlozman, K. L., & Brady, H. E. (1995). Voice and equality: Civic voluntarism in American politics. Harvard University Press.

Walras, L. (1883). Théorie mathématique de la richesse sociale. Corbaz.

Öhrvall, R. (2012). Svenskt valdeltagande under 100 år. Statistiska Centralbyrån.

Öhrvall, R. (2018). Growing into Voting: Election Turnout among Young People and Habit Formation (Vol. 752). Linköping University Electronic Press.

(37)
(38)

Tabell 5: OLS models

(1) (2) (3) (4) (5)

Dvoted Dvoted Dvoted Dvoted Dvoted

Dlnincome -0.0063⇤ -0.011-0.0061-0.0070⇤⇤ -0.0079⇤⇤⇤ (0.0027) (0.0047) (0.0025) (0.0025) (0.0016) Deducation 0.0048 0.0036 0.0050 0.00025 (0.016) (0.016) (0.016) (0.016) Dunemployed -0.024 -0.026 -0.032 (0.039) (0.039) (0.037) Dpart 0.031 0.032 (0.020) (0.020) Dstudent -0.00068 -0.00027 (0.0074) (0.0074) Dlabgroup 0.00063 0.0010 (0.0020) (0.0021) Drel 0.069⇤ 0.0680.0660.068⇤ (0.028) (0.028) (0.027) (0.028) age -0.0023 -0.0039 -0.0012 -0.0016 0.0019 (0.0040) (0.0037) (0.0037) (0.0038) (0.0026) age2 0.000019 0.000036 0.0000081 0.000011 -0.000029 (0.000040) (0.000036) (0.000037) (0.000038) (0.000027)

Over 74 years old 0.011 0.017 0.016 0.013

(0.033) (0.033) (0.033) (0.029)

Time trend -0.0081 -0.0050 -0.012 -0.0091 -0.0045

(0.012) (0.012) (0.010) (0.010) (0.0065)

Dover74 -0.0070

(0.031) Income CPI deflated 0.00019 (0.00016)

Constant 0.065 0.073 0.054 0.053 -0.035

(0.100) (0.093) (0.089) (0.091) (0.063)

Observations 1733 1733 1850 1807 4007

Adjusted R2 0.0082 0.0083 0.0092 0.010 0.0071

References

Related documents

46 Konkreta exempel skulle kunna vara främjandeinsatser för affärsänglar/affärsängelnätverk, skapa arenor där aktörer från utbuds- och efterfrågesidan kan mötas eller

Coad (2007) presenterar resultat som indikerar att små företag inom tillverkningsindustrin i Frankrike generellt kännetecknas av att tillväxten är negativt korrelerad över

Närmare 90 procent av de statliga medlen (intäkter och utgifter) för näringslivets klimatomställning går till generella styrmedel, det vill säga styrmedel som påverkar

I dag uppgår denna del av befolkningen till knappt 4 200 personer och år 2030 beräknas det finnas drygt 4 800 personer i Gällivare kommun som är 65 år eller äldre i

Hibbs (2000) obtained nearly identical estimates of the e¤ect of real income growth rates (sustained over the entire term, rather than just the election year) on aggregate

Intellectual ability Oral expression Written expression Emotional maturity Imagination and probable creativity Potential as teacher Motivation for proposed program of study

(v) Even in the absence of tax evasion, complementarity between labor supply and consumption of real cash balances does not guarantee the optimality of the Friedman rule as a

rather undecided in regard to the beet manner of getting the ioe out of the shaft, but said in case he found it necessary to use the boiler to thaW the ioe, he planned on getting