Institutional Setting and Data

Voter Turnout

In Norway, elections take place every four years at the national and local (that is, county and municipal) levels of government, separated in time by two years. All Norwegian nationals aged eighteen years or older are eligible to vote (with few exceptions), and they become automatically included in the Election Roll upon reaching this age threshold.Footnote ² Municipal governments are responsible for all organizational aspects related to the elections. They also compile the election results and report them to the Election Directorate via an electronic system.

Using the information from the electronic voting systems, Statistics Norway collects individual-level data on electoral participation since 2013. We obtained access to these data for the 2015 and 2019 local elections as well as the 2017 and 2021 national elections. The data cover the complete Election Roll in 27 Norwegian municipalities in 2015, 255 municipalities in 2017, and all municipalities starting with the 2019 elections. As can be seen from Online Appendix Table A.1, this implies that our dataset covers register-based voter turnout for 49–100 per cent of the Norwegian population over the time period of our analysis.

Voter turnout in recent Norwegian elections hovers around 78 per cent for national elections and 59–65 per cent for local elections. Consistent with a vast academic literature (see above), Fig. 1 shows that turnout in Norway is considerably higher among affluent citizens compared to less affluent citizens in both local and national elections (Online Appendix Fig. A.6 confirms a similar pattern for distinct education levels). Clearly, however, these substantial differences in turnout along the income distribution may reflect several confounding factors (Doherty et al. Reference Doherty, Gerber and Green2006; Peterson Reference Peterson2016). Hence, our analysis exploits random, exogenous income windfalls arising from lottery wins to identify the causal relationship of interest.

Figure 1. Income and voter turnout.

Note: The diagrams show the average voter turnout in the election years 2015–2021 for local and national elections conditional on gross personal income. Income is calculated in annual deciles for the entire eligible population.

State-Owned Lottery Monopoly

Two state-owned companies – Norsk Tipping and Norsk Rikstoto – hold exclusive rights to provide lottery and gambling services in Norway. Prizes won in any of the games they offer (henceforth ‘lotteries’) are tax exempt, even though wins exceeding 10,000 NOK up to 2006 and exceeding 100,000 NOK since 2007 are automatically reported to the tax authorities.Footnote ³ Our register data provided by Statistics Norway offer full coverage of all lottery wins exceeding the applicable legal threshold for the period 1993–2021. Panel A in Table 1 presents descriptive statistics for the 2009–2021 period, which covers up to six years before the first election where we have turnout data (that is, the 2015 local elections). Although the vast majority of Norwegians engage in lottery games at least sometimes (estimated at up to 70 per cent in our dataset), Panel A reveals some notable differences between winners and non-winners. Those who won typically have lower educational attainment and are more likely to be older and male.Footnote ⁴ This suggests that there may exist systematic differences between lottery participants and non-participants (including, potentially, unobservable factors such as preferences for gambling).

Table 1. Summary statistics

Note: The table displays summary statistics for person-years observed in the years 2009–2021. Non-winners and winners differ in (the lack of) a lottery win in the period under analysis. Panel A shows descriptive statistics for the entire population aged eighteen years or more. Panel B displays statistics for individuals winning a prize of NOK 10,000 or more at least once in the period 1993–2006 and those winning one NOK 100,000 prize in the period 2009–2021. Observe that the number of observations in the top row of each panel is lower due to the four-year election cycle and the incomplete coverage for the 2015 and 2017 elections.

Our data unfortunately do not include a direct way to distinguish lottery participants from non-participants since we only observe that individuals win in any given year. Still, winning must imply participation. Hence, we can exploit the time dimension of our dataset to develop a proxy for current participants (that is, in the 2009–2021 period of our main analysis) based on who won at least once also in an earlier period (that is, in the 1993–2006 period, where we observe any win above 10,000 NOK). As we have full coverage of the Norwegian adult population, most of these historical winners remain observable in the 2009–2021 period of our main analysis. Under the assumption that they continued to participate in lotteries during the 2009–2021 period, this offers ‘counterfactual’ outcomes from individuals who won in one or more years during the 1993–2006 period but did not register a large win in the 2009–2021 period. Similarly, assuming that all individuals winning in 2009–2021 also participated in the years they did not win, these non-winning years further help identify counterfactual outcomes. Panel B of Table 1 highlights that zooming in on these ‘participants’ brings us closer to a most-similar research design (Doherty et al. Reference Doherty, Gerber and Green2006; Peterson Reference Peterson2016; Geys and Sørensen Reference Geys and Sørensen2024). That is, the characteristics of winners and non-winners in the period 2009–2021 are now much more evenly matched in terms of income, gender, and education level. People experiencing a lottery windfall also look no different on average from non-winners in terms of turnout levels (even when the income windfall is very large; see Online Appendix Fig. A.5). The subsample in Panel B will therefore be at the heart of our main analysis.

Finally, it is important to note that the average income windfall when winning a prize of at least 100,000 NOK in the 2009–2021 period equals 1,206,540 NOK. As this is pulled up by several extremely large wins (see the Online Appendix Fig. A.8), it appears more meaningful to look at the median win size. Conditional on winning, this equals 250,000 NOK (that is, roughly 60 per cent of the median income in our sample period). As this involves a jump of three to four deciles in the income distribution, it reflects a very considerable boost in individuals’ affluence.

Model Specification

To identify the causal impact of unearned income windfalls on voter turnout, we estimate the following linear probability regression model (with subscript i for individuals and t for time):

(1) $${V_{i,t}} = \mathop \sum \nolimits_{l = 0}^6 {\beta _l}{W_{i,t - l}}\; + {\theta _t} + \left[ {{\varphi _i}} \right] + {\varepsilon _{i,t}}$$

where the dependent variable – ${V_{i,t}}\;$ – is a dummy variable indicating the electoral turnout of individual i in year t (1 for voting, 0 otherwise). Since our dataset includes all lottery wins exceeding 100,000 NOK for the 2009–2021 period, the main independent variable ${W_{i,t - l}}$ equals 1 if individual i had a lottery income windfall exceeding 100,000 NOK in year t-l (0 otherwise).Footnote ⁵ With l=0,…,6, we capture both the immediate impact (l = 0) and the longer-term effects of income windfalls up to six years before an election takes place (for example, lottery wins dating back to 2009 for the 2015 election).

Our baseline model is rooted in a robust tradition of social science research, which posits that lottery wins are random events for lottery participants (Imbens et al. Reference Imbens, Rubin and Sacerdote2001) – both across different individuals and over time for the same person. In our main analysis, we, therefore, estimate equation (1) without individual fixed effects ${\varphi _i}$ using two subsamples. The main estimation sample zooms in on lottery participants identified using winners also from the 1993–2006 period (cf. panel B in Table 1). This effectively assumes that a lottery prize is a random event within the subsample of lottery participants (Doherty et al. Reference Doherty, Gerber and Green2006; Dunning Reference Dunning2012). While this assumption cannot be tested directly, it is reassuring to see that this sample is fairly well balanced across winners and non-winners in the 2009–2021 period (Table 1, Panel B). Our second estimation sample restricts the analysis to individuals winning exactly once in the 2009–2021 period (cf. top-right panel in Table 1).

As a robust check, we also estimate equation (1) extended with individual (and year) fixed effects. Such an event-study model directly controls for any (un)observed individual-specific characteristics that do not change over time (such as gender, education, and gambling preferences). By assuming that the timing of lottery wins is random for each individual, the parallel trends assumption will be satisfied (Schmidheiny and Siegloch Reference Schmidheiny and Siegloch2023).

Before we turn to our main results, two methodological issues require attention. First, one might worry that winners of a prize exceeding 100,000 NOK may also win a smaller prize in addition. This would bias our estimates downwards (upwards) if the smaller prize is won first (last). Crucially, the timing of any such smaller wins is random. Hence, on average, they are unlikely to bias our main estimates. Second, one might worry that some individuals identified in our subsample as non-winners may nonetheless collect lottery wins below 100,000 NOK. If such small wins affect turnout, this could generate a bias toward zero when estimating equation (1). Still, we observe only 90,287 lottery wins between 10,000 and 100,000 NOK in 48,256,701 person-year observations over the 1993–2006 period. Even when (conservatively) assuming that 60 per cent of the population are lottery participants, this leads to a win probability among participants of about 0.31 per cent. Considering that the average win between 10,000 and 100,000 NOK is just over 38,200 NOK (median ≈ 31,300 NOK), any presence of small-prize winners among our ‘non-winners’ is unlikely to have a major impact on our estimates.

Empirical Findings

Main Findings

Figure 2 summarizes our main results. The horizontal axis shows the timing of the lottery win relative to the election (with 0 reflecting a win in the election year, –1 the year before the election, and so on). The vertical axis shows the point estimate from equation (1). The left-hand diagram analyses the subsample that includes winners in the 1993–2006 period (Panel B in Table 1) while the right-hand diagram restricts the sample to those winning one prize exceeding 100,000 NOK in the 2009–2021 period (top-right panel in Table 1). In both cases, the point estimates in Fig. 2 reflect the effect of a lottery-based income windfall at a given moment before an election on the probability of turning out to vote in that election (relative to individuals not treated by an income shock).

Figure 2. Lottery windfalls and voter turnout for lottery participants.

Note: The diagrams show the estimated effects of lottery income windfalls on voter participation. The right diagram shows effects when only including individuals winning in the period 2009–2021 (cf. top-right panel in Table 1), while the left diagram displays effects for the sample extended with all lottery winners in the years 1993–2006 (cf. Panel B in Table 1). The diagrams illustrate 95 per cent confidence intervals, calculated using standard errors that are clustered at the individual level. For additional details, refer to Appendix Table B.1.

The results in Fig. 2 indicate that witnessing a random income windfall only has a statistically significant impact on voter turnout when it arises in the near vicinity of an election. The estimated effect size in the two years prior to an election ranges from 1.6 to 1.9 percentage points.Footnote ⁶ Online Appendix Fig. A.3 highlights that our point estimates are somewhat larger for women (2.2 to 3.5 percentage points) compared to men (always below 1.3 percentage points), and only change marginally when adding individual-level controls for age, gender, income, and education (1.4 to 2.6 percentage points). Likewise, our small effect size persists when using data from the entire Norwegian population, and adding individual fixed effects to control for any (un)observed individual-specific characteristics (Online Appendix Fig. A.4). Remember that the median income windfall exceeding 100,000 NOK in the 2009–2021 period is 250,000 NOK (see above), which would move people three to four deciles up the income distribution. In Fig. 1, such an income shock would be associated with an increase in voter turnout of 9.2 to 13.4 percentage points. Our causal estimate remains well below that, strongly suggesting that at best a small portion of the positive income-turnout relationship depicted in Fig. 1 – and observed throughout much of the academic literature – is causal.Footnote ⁷

Our positive point estimates may indicate that winners increase their political engagement more than individuals untreated by an income shock, or it may be that non-winners decrease their political engagement more than winners. Since electoral participation is sticky and tends to increase over most of the life cycle, we present age-specific estimates in Online Appendix Fig. A.7 to differentiate both interpretations. The results indicate that our findings in Fig. 2 are largely driven by the 66-and-above age group (who document effect sizes between 2.0 and 4.7 percentage points). Assuming that these age-specific estimates represent life-cycle effects, a natural interpretation would be that a positive income shock may make it easier for (some) elderly to sustain their electoral participation.

Heterogeneous Effects by Income Levels

One might argue that gaining additional income is more important when one is (very) poor, but less relevant when one is already affluent (Schaub Reference Schaub2021; Markovich and White Reference Markovich and White2022; Schafer et al. Reference Schafer, Cantoni, Bellettini and Ceroni2022). This section therefore extends our analysis by estimating the effects of lottery wins conditional on income levels. Let $Inc_{i,t}^d$ (with d=1,…,10) represent a set of indicator variables equal to 1 if individual i is in income decile d in year t, and 0 otherwise. The extended model then takes the following form:

(2) $${V_{i,t}} = \;\mathop \sum \nolimits_{d = 1}^{10} {\gamma _{ - x}}{W_{i,t - x}}Inc_{i,t - x}^d + \mathop \sum \nolimits_{d = 1}^{10} {\gamma _{ - x}}Inc_{i,t - x}^d + {\theta _t} + {\varepsilon _{i,t}}\;\;\;\;with\;x = 0,1,2$$

where x refers to the number of years since the lottery win materialized. Given that the strongest effects in Fig. 2 are observed near-contemporaneously, we focus on lottery wins in years immediately prior to an election (that is, x=0,1,2).

The results in Fig. 3 fail to generate a clear pattern. As such, our analysis offers no clear evidence of a meaningfully distinct impact of income windfalls on voter turnout across the income distribution.

Figure 3. The effect of lottery wins on voter turnout, by income levels.

Note: The diagrams display the effect of lotto wins conditional on income, defined by annual deciles. The analyses employ a sample of lottery participants that also includes winners in the years 1993–2006 (cf. Panel B in Table 1). We estimate effects with zero, one- and two-year lags for income and the lottery win variables. The diagrams illustrate 95 per cent confidence intervals, calculated using standard errors that are clustered at the individual level. For additional details, refer to Appendix Table B.2.

External Validity

Panel B of Table 1 indicates that lottery winners are quite representative of the population of ‘lottery players’. Even so, as mentioned, lottery windfalls are an unusual form of income, and one might wonder whether unearned income shocks from lottery windfalls yield different effects on electoral participation compared to changes in earned income (Doherty et al. Reference Doherty, Gerber and Green2006; Peterson Reference Peterson2016). We address this external validity concern by exploiting population-wide data across four election years (2015, 2017, 2019, and 2021) on individual-level variation in total annual earned income (measured in income deciles). We estimate the effect of individuals moving up from decile 1 (the reference group) to any of the other nine income deciles between two election years. The results are summarized in Online Appendix Table C.5. Model (A) estimates the key parameters of interest controlling for election years alongside individual fixed effects. Model (B) furthermore includes municipality fixed effects as well as linear municipality-specific time trends to capture regional (income) growth differences.

Remember that the median lottery windfall reflects moving three to four deciles up the income distribution. Both models in Appendix Table C.5 indicate that an equivalent four-decile jump in earned income would raise turnout by 0.6 to 4.1 percentage points depending on the individual’s initial position in income distribution. This again constitutes a small part of the turnout differences across the income distribution observed in Fig. 1. Overall, and keeping in mind potential omitted-variable bias in these models, we conclude that the turnout effects from earned income changes are broadly consistent with our main results using lottery wins.