Swedish exceptionalism? Investigating the effect of associational involvement on generalized trust with panel data

ABSTRACT For some time been it has been hypothesized that involvement in civic associations creates generalized social trust. Yet, prior panel data studies, based mainly on data collected in Australia, the Netherlands, Switzerland, the United Kingdom, and the United States, have found little support for the existence of such an effect. This article adds further empirical knowledge, focusing on Sweden. The evidence presented here is the first to provide support for the hypothesis using a survey that allows panel data models. In the conclusions, it is discussed whether the differing findings may depend on Sweden being a particularly favourable environment, considering its comparatively democratic and prosperous associational life; or if the reason is that the data at hand do not allow using exactly the same panel models as in some of the prior studies.


Previous Research
'Civic engagement is not an important breeding ground for trust ', Van Ingen and Bekkers (2015, p. 291) wrote in a relatively recent article.Their conclusionbased on rather few but convincing panel data-based studies conducted in five different Western democracies is astonishing for any scholar who has followed this research field since Putnam's seminal study (2000), and also taking into account that empirical findings have long seemed to support the hypothesis that civic engagement creates trust.But is the conclusion true for all kinds of countries and contexts?Before we return to that question, a short overview will follow of the main theoretical arguments within this research field, as well as the main objections.
Social trust is a necessary bond that keeps societies from falling apart, and its presence is a necessary condition for democracy to function well (e.g., Paxton, 2002).In the search for factors and conditions that enhance trusthere defined as the perceived trustworthiness of an average individual, not friend or acquaintance (Paxton & Ressler, 2018, p. 149)particular attention has been devoted to civic associations.This social arena has been argued to be particularly beneficial for the creation of generalized social trust (see, e.g., Putnam, 2000).One argument supporting this belief is that in order to survive, voluntary organizations have to demand cooperation and trust; if the members do not seem trustworthy people will leave and the organization will die, as exit is rather costless (Kramer, 1999, p. 579).Therefore are norms of trust and cooperation especially likely to flourish in this context.For trust to be generalized, it has been suggested that bridging (as opposed to bonding) associations are especially beneficial, i.e., associations characterized by diversity and by its members being socially heterogeneous (Putnam, 2000).It has also been suggested that connected associationsi.e., where members belong to other organizations as wellare particularly favourable, by creating network ties and trust across social cleavages (see, e.g., Paxton, 2002).
However, all scholars do not agree on civic associations being so important for social trust.In fact, opponents argue that trust is a stable personality trait, established mainly during childhood, and that it therefore is unlikely affected by experiences later in life, such as involvement in civic associations (Uslaner, 2002).Moreover, a correlation between civic participation and trust is assumed to be an effect of trusting individuals self-selecting into associational activity rather than the former factor casually affecting the latter (for a recent overview of the critique, see Paxton & Ressler, 2018).
The effects of civic involvement have been extensively tested in numerous studies, generally receiving strikingly strong empirical support (for an overview, see, e.g., Pichler & Wallace, 2007). 1 Empirical research in general is almost exclusively based on cross-sectional data and, to my knowledge, only four studies use panel data when explicitly addressing this topic.The general conclusions of the latter studies, based on data collected in Australia, the Netherlands, Switzerland, the United Kingdom, and the United States, differ markedly from those based on cross-sectional data (Bekkers, 2012;Claibourn & Martin, 2000;Van Ingen & Bekkers, 2015; but see Botzen, 2015).Overall, in panel data-based investigations, the reported effects are nonexistent, substantially small, and/or non-lasting. 2According to these findings, trusting individuals seem to self-select into participation and there are no apparent indications that trust arises from civic involvement.
The types of countries so far studied using panel data are, however, somewhat limited.In this article, a panel data approach is utilized with data collected in Sweden, a country with a reputation of being an exceptionally well-functioning and egalitarian democracy and where a comparatively extensive and stable organizational life already existed in the early 1900s.Today, Sweden's civic associations are known for their relatively democratic and equal internal structure (Van Ingen & van der Meer, 2011) and particularly high levels of involvement (e.g., Pichler & Wallace, 2007).From a country-comparative perspective these associations are general likely to be rather diverse and interconnected (as individuals tend to have many memberships), making Sweden a likely context for civic engagement to enhance trust. 3On the other hand, a ceiling effect may hinder the mechanisms to come about as citizens in general already are rather trusting. 4In sum, it is not completely obvious what to expect regarding Sweden, but it is an interesting case worthy of further investigation.
Previous studies of this country have supported civic engagement increasing trust, but are exclusively based on cross-sectional data (see, e.g., Pichler & Wallace, 2007;Stolle, 1998).The present article provides the first panel data test in this country, paying more attention to the just-mentioned problems of causality than has previously been possible.

Data and Measures
The data analysed here is a nationally representative Swedish panel survey, the so-called Swedish Citizen Study. 5To the author's knowledge, this is the only dataset collected in Sweden that allows panel data analyses of the hypothesis under scrutiny.The first wave was undertaken in 1997 on a random sample of 1964 citizens (16-80 years old); the response rate was 74.3% and the interviews were carried out mainly face-to-face.The second wave used the same sample and was carried out in 1999 as a short mail questionnaire, with a response rate of 61.9%. 6Compared with prior panel data studies, in which the time between waves varies from one to nine years, the corresponding time here was comparatively short; approximately 1.5 years (cf.Bekkers, 2012;Botzen, 2015;Claibourn & Martin, 2000;Van Ingen & Bekkers, 2015).
Generalized trust is measured using an index variable based on three items also used in the European Value Survey (see, e.g., Sønderskov & Dinesen, 2016).The first one is identical to the traditionally employed item, i.e., 'Generally speaking, do you think that most people can be trusted, or that you cannot be too careful in dealing with other people?' Answers are given on a scale of 0-10 where '10' equals 'most people can be trusted' and '0' equals 'you cannot be too careful'.The second item is 'Do you think that most people would try to take advantage of you if they got the chance, or would they try to be fair?' and the third is 'Would you say that most of the time people try to be helpful, or that they are mostly looking out for themselves?' (both measured on 0-10 scales).Analyses support treating these three measures as indicators of a single dimension. 7One additive index variable was created, based on the three items, and standardized to run from 0 (minimum level of trust) to 1 (maximum level of trust).
Organizational involvement is an additive index variable measuring membership during the previous year in 30 specified types of voluntary associations, ranging from recreational to special-interest and ideological associations (presented in Table A1 in the Appendix).

Results
The findings are presented in Table 1.The results of a cross-sectional analysis are shown in the first column (all variables measured in the first wave).In line with prior cross-sectional studies, organizational involvement has a positive and significant effect on generalized trust.Additionally, it should be reassuring to find that the signs and levels are generally as expected also for the control variables, such as education, foreign born, living with partner, and income.The second column displays the results of a similar model, but with the dependent variable (i.e., generalized trust) measured in the second wave approximately 1.5 years later.With the exceptions of age and gender, which now display significant effects, the coefficients generally resemble those in the previous model.Most noticeably, the effect of associational engagement still substantial and statistically significant, even being somewhat greater than in the previous model.Intuitively, this seems reasonable, as it should take some time for an increase in organizational involvement to affect trust.
Finally, the third model adds one additional control variable, i.e., generalized trust measured in the first wave.Utilizing a lagged dependent variable in this way is assumed to deal with two-way causation more satisfactorily than does using a cross-sectional model.The lagged dependent variable can, in addition, be viewed as a surrogate or proxy for omitted variables (e.g., family background and self-selection; Finkel, 1995).As expected, when trust measured in wave 1 is included as a control, Table 1 shows that the effects of several independent variables decrease, leaving one of them substantially weaker, though still statistically significant (education), and three of them no longer statistically significant (income, living with partner, and foreign born).Most importantly, although somewhat weakened, organizational involvement still has a statistically significant and positive effect.The size of the coefficient means that a change from complete passivity (not uncommon) to being noticeably involved, such as membership in six organizational types (also not uncommon), is expected to increase trust by 0.10 units on the 0-1 scale.This is arguably a noticeable effect, considering that more than half of the respondents' trust levels are found between 0.4 and 0.7 on the trust scale. 8 In additional analyses, shown in Tables A3 and A4 in the Appendix, the effects of each one of the 30 different types of organizations are investigated in separate regression analyses (c.f., Coffé & Geys, 2007;Stolle & Rochon, 1998).The coefficients were in most cases positive but not statistically significant (or more or less zero), with the exception of four types of organizations where the positive effect was statistically significant: Associations for disabled people or medical patients, sport clubs, trade unions, and professional (although the latter two were significant only on the 0.1 level).It is hard to convey a pattern of some general types of organizationssuch as interest groups or recreational organizationsbeing more useful for the development of generalized trust than others.However, it should be noted that many of the 30 organization variables are heavily skewed, limiting the possibilities of reaching statistically significant effects (significant effects appear mainly among the least skewed ones, cf., Table A1). 9

Concluding Discussion
The results presented here are clear.No signs of self-selection removing the effects of civic engagement on generalized trust appeared and neither did any significant ceiling effects seem to occur (suspected due to Swedish citizens already being comparatively trusting).On the contrary, a significant and positive impact of organizational involvement on generalized trust was unearthed; in fact, the effect is substantial even when controlling for the lagged dependent variable.
However, three noticeable limitations of the data used should be mentioned.First, only information on organizational membership was available, while previous studies mostly analysed both membership and volunteering, often distinguishing between the two.However, it seems rather unlikely that the effects would have been weaker if active associational involvement could have been analysed separately in this article. 10 A second limitation is the lack of measures of organizational involvement in the second wave (unlike generalized trust, which was measured in both waves).Methods such as fixedeffects and first-difference regressions, which permit stronger controls for omitted variable bias, could therefore not be employed (cf.Van Ingen & Bekkers, 2015).We cannot be sure what the results would be with these methods.Nonetheless, the lagged model used here is preferable to a cross-sectional approach when handling self-selection and two-way causality (Finkel, 1995).Furthermore, all three panel methods have been used in previous studiesundertaken in other countrieswith the findings being highly similar (no effect).
Finally, the panel waves were conducted with an interval of only approximately 1.5 years, so we do not know whether the observed effect would last for a longer period.What we do know is that it survives controlling for relevant variables, including a lagged dependent variable, and in this sense seems persistent.
The data limitations mentioned should be taken into consideration when evaluating this study.Still, unlike previous studies undertaken in Australia, the Netherlands, Switzerland, the United Kingdom, and the United States, the findings presented here do support the hypothesis, at least that an effect of a more short-term character exists.Is it possible that the diverging results in fact are related to the different countries under study?As discussed above, Sweden is a rather deviant case, with its comparatively democratic and vivid organizational life seemingly facilitating for the hypothesis to gain support.However, before final conclusions are drawn as regards Sweden, more panel data should be collected in this country based on more than two waves and/or longer time periods.This would allow for investigation of the persistence of the effect as well as the use of even more robust methods for taking omitted-variable bias into account.If the findings presented here still hold, further theorizing about the importance of the context where organizations operateincluding the national contextindeed seems called for.Notes 1.A limited of these studies investigate the impact of different types of organizations (see, e.g., Coffé & Geys, 2007;Stolle & Rochon, 1998; see also, Glanville, 2004).The evidence is somewhat mixed and does not reveal any consistent pattern.2. See also see Brehm and Rahn (1997), who find support for the hypothesis using cross-sectional data and a latent variable approach, and Glanville, Andersson, and Paxton (2013), who find a causal effect on social trust of informal social tiesi.e.social interaction with friends, relatives and neighboursusing American longitudinal data.However, the conditions for informal social ties to affect trust are rather different from the conditions regarding formal associational activity.3. The impact of contextual variation within Sweden, on the link between civic engagement and trust, would also have been interesting to investigate.However, the data at hand do not permit such analyses, as there are too few respondents representing the less populated categories of the relevant geographical units (municipalities and regions).4. Participation in civil society even seems to be higher in Sweden than in other countries known for high social capital, such as Denmark and the Netherlands (Pichler & Wallace, 2007). 5. Its panel component has been used in several research articles but for other purposes than the present one (see, e.g., Adman, 2008;Bäck, Teorell, & Westholm, 2011).6.Of the original sample, 52.7% took part in both waves.Professors Jan Teorell and Anders Westholm, at Lund University and Uppsala University, respectively, were principal investigators.Analyses of missing cases in the 1997 survey identified the elderly, immigrants, and people living in big cities as less often taking part, but the differences were found to be too small to have any probable significant effect on the results (Petersson, Hermansson, Micheletti, Teorell, & Westholm, 1998, pp. 22-24).For the sake of comparison, all models in Table 1 are based on the same respondents (i.e.individuals for whom there are values for all relevant variables).However, comparing cross-sectional analyses of the entire 1997 sample with corresponding analyses of only those who took part in both surveys achieves similar results (the coefficient for organizational involvement is substantially somewhat weaker than in Table 1 but still substantially and statistically significant).Hence, the results do not seem to be significantly affected by the potentially lower participation of 'low-trusters' in the second wave.7. Using principal component analysis, only one factor survives the Kaiser criterion (Eigenvalue >1.0), explaining 68% of the variance in the first wave and 71% in the second wave (Cronbach's alpha is 0.76 and 0.79, respectively).The main models presented below have also been rerun using the original trust indicators analyzed separately, one by one.The results are, both when it comes to substantial and statistical significance, very similar to the ones presented below.8.In further analyses (not shown) additional controls were included, but the effect of organizational involvement remained unaffected (included were number of children, residing in a city or the countryside, subjective health, media consumption, political participation, internal political efficacy, and party identification).Several interaction models have also been investigated, e.g., whether the effects of organizational involvement differ depending on contexts such as if living in a city or in the countryside, without any significant effects occurring.9. Dimensional analyses (not shown) did not reveal any interpretable pattern, as comes to clusters of different types of organizations (e.g., in general did recreational organizations not load on the same dimension).Therefore I chose to investigate the 30 different organizational types separately.10.Cross-sectional studies in fact tend to indicate stronger effects among active than passive members (see, e.g., Bekkers & Schuyt, 2008;Welch, Sikkink, & Loveland, 2007).

Table 1 .
Effect of organizational involvement on generalized trust.