Data

Our analysis centers on Sweden. Associations in Sweden are numerous; the average adult is associated with nine associations (Lundström and Wijkström Reference Lundström and Wijkström1997).Footnote ³ However, political participation in Sweden is already high, which could mean that rates of membership in civil associations are due to self-selection.

To assess membership in civil society associations, we rely on Statistics Sweden's yearly Swedish Living Conditions Survey (ULF/SILC). The survey is administered through personal interviews, and a subset of respondents belong to a panel survey conducted every seven years. We focus our empirical analyses on this panel of individuals.

Our outcome variables capture different forms of political participation. First, we study voter turnout in European Parliament elections rather than Swedish national elections, because these have more variation in turnout. Secondly, we study other acts of political participation, such as contacting political officials and participating in demonstrations.Footnote ⁴ We conduct a factor analysis for all dependent variables except for voter turnout. Our prior is that variation in the variables may be captured in one factor, which is confirmed in Figure A1. We thus have two main dependent variables in our main empirical analysis: voter turnout and the first factor score for these other measures.Footnote ⁵

Our right-hand side variables come from questions on civil society associations in the ULF surveys. We include information on passive and active membership in an association. We also account for the type of association. In the main analysis, we include binary indicators for reported passive or active membership in any organization and a count variable for the number of passive and active memberships a respondent reports. We only focus on associations that were included twice in the survey so that we can run panel regressions.Footnote ⁶

Our final dataset consists of a pooled panel dataset administered in three different waves: (1) 1992 and 1993, (2) 2000 and 2001 and (3) 2008 and 2009. For voter turnout, we only use the second and third waves, because Sweden did not elect representatives to the European Parliament before becoming an EU member in 1995.

The empirical analysis is conducted in two parts. First, we pool the data from all years together and estimate the associations between membership (separating passive and active membership) and our two outcome variables. Next, we conduct a panel data analysis. We estimate the two regression equations below.

(1)$$Y_i = \beta _0 + \beta _1Association_i + \beta _2W_i + u_i$$

(2)$$Y_{i, t} = \beta _0 + \beta _1Association_{i, t} + \beta _2W_{i, t} + f_i + \tau _t + e_{i, t}$$

(3)$$t\in \{ 1, \;2\} $$

Y is voter turnout or the factor score for the other political participation measures, while ui and e_i,t are error terms. Association is an individual's reported membership in the survey, which is either a binary variable for any type of passive or active membership or a count variable for the number of passive or active memberships. W is a vector of covariates including years of education, employment status, age, gender, immigration status, standardized income for each cohort and year, and whether the person's family receives social welfare. f_i refers to individual fixed effects, and τ_t represents period fixed effects.Footnote ⁷

To assess the effects of membership in associations on political participation, it is helpful to think about the ideal randomized experiment. In such an experiment, we would randomize individuals into passive, active and no membership in different types of civil society organizations. This design would keep all other factors constant and would isolate the causal effect of interest. Given that we are unable to run such an experiment, we attempt to partially overcome some of these concerns in Equation 2.

By including fixed effects, we account for all characteristics that fluctuate between individuals but that remain constant over periods, and for characteristics that vary over time periods but are constant between individuals. In the present case, we only use the variation in political participation and civil association membership within the individual unit over the two time periods. We demonstrate the degree of between- and within-unit variation in our dependent and independent variables in Appendix Tables A2–A5.

We are careful to not argue that these estimated coefficients in Equation 2 amount to causal estimates, however. Such an analysis requires a more elaborate identification strategy with exogenous variation in membership.Footnote ⁸ Notwithstanding these limitations, the estimates in Equation 2 are less plagued by self-selection concerns, and can provide empirical results that may nuance earlier empirical findings on the relationship between membership in organizations and political participation.

Main results

Table 1 presents the results from our first model (Equation 1), where we regress our measures of political participation on civil association membership and our set of socioeconomic control variables.Footnote ⁹ We distinguish between passive and active membership where an individual cannot be an active member without also being a passive member. The estimated coefficients for active membership should thus be interpreted as the additional impact of active membership.Footnote ¹⁰ With one exception, the estimated coefficients for our variables of interest are statistically significant at the 5 per cent level. The first column in Table 1 reveals that being a passive member of any organization is associated with a 7.2-percentage-point increase in the probability of voting in the European election. For someone who is also an active member of the organization, their voting propensity increases by an additional 3.3 percentage points. A similar pattern is found for other forms of participation. Being a passive member of an organization is associated with an increase in the factor score of political participation by 0.12 standard deviations. For active members, the factor score is estimated to increase by approximately 0.4 (0.12 + 0.26) standard deviations.

Table 1.

Pooled data with covariates

Note: robust standard errors in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01

The results from Table 1 confirm prior scholarship and reveal a positive relationship between membership and political participation, but the estimated coefficients are plagued by self-selection. Members of civil associations are likely very different from non-members, and it is unlikely that we have controlled away these differences and obtained estimates where all else is equal except for civil association membership.

We therefore turn to our panel analysis (Equation 2) next and present the results in Table 2. Using this empirical strategy, the estimated coefficients decrease substantially in magnitude, but remain positive despite rarely manifesting as statistically significant. Voter turnout data are displayed in the first column: being a passive member is no longer associated with a large increase in the probability of voting in European Parliament elections. The estimated coefficient for passive membership, which is no longer statistically significant, corresponds to a 1.3-percentage-point increase in the likelihood of voting for passive members of any civil association. Active membership in an organization continues to have a statistically significant impact on the factor score, but the estimated impact is smaller than in Table 1. For the factor score, we do still estimate one positive and statistically significant coefficient for the number of passive memberships, but the coefficient is again smaller in size.

Table 2.

Panel data analysis

Note: clustered robust standard errors at the individual level in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01

These results demonstrate that the relationship between membership in civil associations and political participation is rather nuanced, and that the previously observed strong “effects” of civil association membership are likely due to a high degree of self-selection behavior. For robustness, we provide results from a propensity score matching analysis in Appendix Figures A2–A4 and Table A13. We have also run analyses for different age groups, presented in Tables A14–A15, as well as analyses where we use an additive index instead of the factor score as the dependent variable (Table A7) and analyze each participation indicator separately (Table A8).

Distinguishing between different types of associations

Given that the theoretical causal chain runs from membership to social capital and finally to political participation, it is likely that different associations yield various degrees of social capital, and thus have differential impacts on political participation. The degree of self-selection is also likely larger for some types of associations compared to others. The panel data analysis in the main results section rests on the premise that the difference in characteristics between individuals is captured by the inclusion of individual fixed effects, time fixed effects, and covariates where the identifying variation stems from within-subject variation between time periods 1 and 2. However, it is unlikely that the estimated coefficients from the previous section are causal estimates, given that we have not accounted for all types of self-selection behavior in our analyses.

To take these analyses one step further, we introduce an approach of categorizing associations into two groups and run the panel data analysis for each organization separately. Our purpose here is to investigate whether the positive associations found in Table 2 are primarily driven by membership in organizations where we suspect a high degree of self-selection may exist. If so, the claim that membership in civil associations leads to higher political participation is further weakened.Footnote ¹¹

We distinguish between high and low self-selection to address two different concerns. First, we worry about reverse causality, which is likely present for membership in organizations more closely connected to the political sphere. As previously discussed, a person who is participating in demonstrations may be more likely to join a political organization or protest group. We have therefore put peace organizations, environmental organizations and women's organizations in the high self-selection group. Secondly, some associations (for example, parental and pension associations) are only open to people who fall within a certain age range. For these associations, our fixed effects are not able to capture latent confounding factors that may be shaping their membership. Thus, we expect a higher degree of selection bias to still be present in the panel regression with membership in these organizations. The remaining associations (for example, sports, housing, culture associations (theatre and dance), and organizations for immigrants and persons with disabilities) are categorized into the low self-selection group.

The analysis is presented in Table 3, and estimates the same model as in Table 2 but with separate effects for associations where we expect high self-selection and for those where we expect smaller selection bias. Beginning with voter turnout, membership in low-self-selection associations (passive and active membership) has no positive impact on voter turnout. The point estimates are sometimes positive and sometimes negative, but always small in magnitude and statistically insignificant. For high-self-selection associations, we estimate consistently positive and sometimes statistically significant coefficients for both passive membership and active membership. Overall, these results suggest that the positive ‘effect’ of membership on voter turnout is mainly driven by high-self-selection associations, which could mean that the remaining impact is simply a result of self-selection due to political interest and not due to a causal effect of membership in civil society associations. For the factor score analysis, we estimate both negative and positive coefficients depending on the specification. We find two statistically significant results related to the number of associations a person belongs to. Here, we find a positive impact for both the high- and low-selection groups on the factor score. Nonetheless, these results should be interpreted with caution, given that the other estimates vary greatly.

Table 3.

Organizations grouped with different level of selection bias (high and low)

Note: clustered robust standard errors at the individual level in parentheses. * p < 0.1, ** p < 0.05, *** p < 0.01

For completeness, we have also run the entire analysis for all types of association separately (not grouped together). Appendix Tables A9–A12 present these results, which substantiate our main findings.