A Switch or a Process? Disentangling the Effects of Union Membership on Political Attitudes in Switzerland and the UK

The paper examines the effects of union membership on individual political attitudes using panel data for Swiss and British workers. Considering union membership as an on/off switch (member vs. non&#8208;member), as it is often done, it is only possible to distinguish between a selection effect (unions attract like&#8208;minded individuals) and a molding effect (the experience of membership has a transformational impact on the individual). Exploiting the longitudinal structure of the data reveals that union membership is best characterized not as a switch, but as a dynamic process involving anticipation effects (which start well before becoming affiliated) and maturation effects (which become noticeable only after a certain duration of membership and may not dissipate after leaving the union). Empirically, the selection effect appears the most important in the two countries we focus on, while the molding effect is less pronounced. Anticipation and maturation effects are also non&#8208;negligible and hitherto unexplored.


Introduction
A time-old literature argues that membership in labor unions increases the individual propensity to be politically involved, participate in elections, and vote for pro-labor parties. An even broader social science literature maintains that associational membership in general, including union membership, has a transformative effect on participating individuals. In reality, the empirical validation of these plausible claims is less straightforward than most literature assumes. In particular, as we discuss in this article, the modal type of analysis -based on cross-sectional data and lacking a longitudinal dimension-generally fails to adequately control for self-selection, and consequently too hastily interprets cross-sectional correlations in causal terms.
An emerging literature improves on the previous generation of studies by making an effort to distinguish between selection and causal effect of union membership through appropriate research designs. However, the focus of this new literature is either on particular types of unions-representing longshoremen (Ahlquist, Clayton, and Levi 2014;Ahlquist and Levi 2013)-or on workers in industries strongly exposed to the pressures of globalization (Kim and Margalit 2017), and this makes it difficult to understand the effect of unions in general. Furthermore, the new literature shares with the old a view of union membership as an "on/off" switch: when the variable is "on" certain effects are expected and vice versa when it is "off." In this article, we make two contributions: First, we provide a rigorous assessment of the average treatment effect of union membership on political attitudes in Switzerland and the UK as a whole (as opposed to membership in particular unions), carefully distinguishing the effect of attitudinal transformation-which we refer to as molding-from the selection effect of unions, which derives from unions having a tendency to attract workers who are systematically more interested in politics, more likely to participate in elections, and more willing to vote for labor parties than non-members.
Second, we show that union membership is a process that unfolds in time before, during, and sometimes after the experience of union membership, and not just an on/off switch. Using this approach, we uncover an anticipation effect, which means that in some cases workers begin modifying their attitudes before joining unions, and a maturation effect, i.e., an attitudinal change which becomes noticeable only after a certain duration of membership (and thus may fail to materialize if workers do not remain members for a sufficiently long time). Appreciation of the dynamic effects of union membership requires longitudinal as opposed to cross-sectional data, and this is another implication of the article.
Our empirical focus on Swiss and British workers is primarily data driven: Focusing on Switzerland and the UK allows us to analyze two high-quality household panel surveys (the Swiss Household Panel [SHP;1999-2014 [UKHLS;1991-2014), which are uniquely suited to the type of analysis we conduct in this article. In particular, they allow us to analyze long union membership spells. In addition, these two countries provide interesting variation in both labor and political institutions. With a centralized industrial relations system at the industry level and a consensual political system (Lehmbruch 1993;Lijphart 1998), Switzerland is usually considered a coordinated market economy (CME; Hall and Soskice 2001), while the UK is widely regarded as a prototypical liberal A Switch or a Process? / 467 market economy (LME). Thus, the choice of these two countries provides variation of institutional conditions and boosts the external validity of results. 1 Nonetheless, as we argue in the Discussion section, the generalization of our findings to national contexts characterized by predominantly non-voluntary forms of union membership (e.g., the United States, Canada) may not be warranted.

and the British Household Panel [BHPS]/UK Household Longitudinal Study
The remainder of the article is organized as follows. After discussing the literature and presenting the theoretical framework in the next section, and the characteristics of the data in the third, we deliver the empirical analysis in two parts. In the fourth section, we treat union membership as an on/off switch, and examine the average treatment effects associated with it. We then shift to union membership as a process, and analyze how its attitudinal effects vary dynamically. After providing a discussion of the findings, the final section concludes with limitations of the current research and avenues for future research.

Theoretical Framework
The existing literature on the individual-level impact of trade unions on political outcomes maintains, almost unanimously, that union membership has important effects on political attitudes: It increases, on average, the political involvement of individuals (interest in politics and participation in elections) and their closeness to left-wing political parties. 2 The bulk of the literature is based on cross-sectional regressions correlating union membership with political outcomes, while controlling for some individual and (sometimes) contextual characteristics. The estimated coefficients are generally interpreted in causal terms, with (sometimes) a discussion of why endogeneity should not be considered a problem. Similar results have been found for several countries, and thus appear robust to cross-national variation (Bryson et al. 2013;Bryson 2014;Budd, Lamare, and Timming 2018;D'Art and Turner 2007;Freeman 2003;Geissbühler 2000;Gray and Caul 2000;Kerrissey and Schofer 2013;Lamare 2010Lamare , 2016Leigh 2006;Leighley and Nagler 2007;Rosenfeld 2010Rosenfeld , 2014. In Table A1, Appendix A in supporting information, we provide a summary of recent studies. In an extension of the classic work on union effects in the United States (Freeman andMedoff 1984), Freeman (2003) reports that union members are 4 percent more likely to vote and to prefer Democratic candidates than non-members with comparable characteristics.
As for the reasons why trade unions produce the effects that are attributed to them, two types of explanations are generally invoked: the "school of democracy" and the "class consciousness" explanations. The former explains why unions increase political interest and participation (Putnam, Leonardi, and Nanetti 1994;Tocqueville [1840] 2016); the latter accounts for the effect on partisan affiliation and voter choice.
With regard to the first explanation, unions, similar to other voluntary associations, are argued to exert a "civilizing" effect on members. Interacting with other individuals inside organizations augments the "civicness" of members; teaches them the virtues of reciprocity and the importance of doing one's share for the common good; and broadens the political horizons of individuals (Lazer et al. 2010;Putnam 2001). Furthermore, taking part in meetings and being involved in organizational activities enhances the propensity to be interested in and to participate in politics (Baggetta 2009;Hanks 1981;Hooghe 2003;Minkoff 2016;Quintelier 2008;Sobieraj and White 2004;Terriquez 2011;Verba, Schlozman, and Brady 1995;Wollebaek and Selle 2002). Unions can also directly mobilize their members to vote (Lamare 2010(Lamare , 2016. In particular, for less educated individuals, unions act as information providers, thus reducing the costs of voting and clarifying potential gains from participation (e.g., Kerrissey and Schofer 2013).
Some literature has also tried to identify the organizational features that are most conducive to interest and participation. For example, the literature on civic culture and social capital has argued that organizations without strong status differences between leaders and followers and without an explicit political mission exert a greater "civilizing" effect than politically oriented and vertical organizations (Almond and Verba 1989;Baggetta 2009;Putnam 2001;Quintelier 2008). However, some studies challenge these views by emphasizing the persuasive ability of union leaders as an important determinant of union members' political involvement and orientation (Cregan, Bartram, and Stanton 2009;Foerster 2004), and by pointing out that, even in apolitical organizations, the exposure to political dialogue and information is a necessary condition for organizational membership to have an impact on political involvement (Hanks 1981;Sobieraj and White 2004). The dichotomy between apolitical and political organizations is relevant for our topic because unions are both economic and political organizations. On the one hand, it may be argued that the core of union activity, the improvement of members' working conditions, is of eminently economic nature and not strictly political. On the other hand, possibly as a consequence of the union density decline threatening their legitimacy as bargaining partners, unions are increasingly active in the political arena, they have a higher visibility in the public sphere, and rely on political alliances (Baccaro, Hamann, and Turner 2003;Streeck and Hassel 2003). Therefore, the discussion of political matters is also a key feature of union meetings and of union affairs (Kerrissey and Schofer 2013;Verba, Schlozman, and Brady 1995).
The effects on partisan affiliation and voting choice in Europe are generally linked to the emergence of a sense of solidarity among workers or class consciousness (Brooks 1994;Hyman 1978;Wright 1996: 373-518), or alternatively to the information-provision function of unions that makes members more aware of their interests (Kim and Margalit 2017). By becoming members of trade unions, it is argued, individual workers develop a clearer sense of the structural similarities between their life conditions and the life conditions of other workers; develop a greater sense of efficacy; overcome passivity; and understand the importance of electorally supporting parties committed to improving workers' conditions. Through union membership, the individual worker may come to identify with a "community of fate" and experience a widening of horizons (Ahlquist and Levi 2013). Furthermore, union membership may promote the development of a more encompassing view of individual interests (Mosimann and Pontusson 2017). There is a classic literature on American "exceptionalism," starting with Sombart ([1906] 1976) and Perlman ([1928] 1979), which argues that American workers are (or were) less class conscious than their European counterparts. Nonetheless, the mechanisms evoked above, e.g., information transmission and the building of solidarity ties, are sufficiently general to be broadly applicable.
All explanations for why union membership would have an impact on individual attitudes invoke a "molding" effect of sorts. Attitudes are reshaped by the experience of union membership; individuals are made more civic-minded and willing to participate politically, or more aware of their interests and willing to support certain types of parties. There is, however, other literature casting doubt on the presence of such molding effect of unions.
Some literature suggests that political preferences are highly inertial (Campbell et al. 1980;Denny and Doyle 2009;Gerber, Green, and Shachar 2003;Hooghe and Wilkenfeld 2008;Prior 2010;Sears and Funk 1999), and that the political outlook of an individual is essentially shaped during the childhood and adolescent years by the family background and by the educational path (Avdeenko and Siedler 2017;Eckstein, Noack, and Gniewosz 2012). Thus, the impact of associational membership may be noticeable only during adolescence (McFarland and Thomas 2006;Quintelier and Hooghe 2012). Some longitudinal studies have found that the political attitudes of students are already stable before high school, and that the differences between students in different education paths even predate the high-school period (Hooghe, Dassonneville, and Marien 2015;Persson 2012). These findings cast doubt on the ability of unions to modify the members' political views. At the same time, the increasing volatility of voting behavior (e.g., Drummond 2006;Gomez 2018;Pedersen 1979) suggests that political preferences are probably less engrained than the above literature suggests and possibly more susceptible to associational influence.
In addition, the literature on voluntary organizations distinguishes between active and passive membership and contends that active membership is more likely to be conducive to attitudinal change than passive membership (Almond and Verba 1989). Actual involvement in associational dynamics, as opposed to nominal membership, is seen as necessary in order to observe an attitudinal change (Howard and Gilbert 2008;Putnam 2001). Thus, one may expect union molding effects not to be a generalized phenomenon but to be limited to, or at least stronger for, "active" members. However, several authors highlight (e.g., Minkoff 2016;Putnam 2001;Skocpol 1999) the rise of "checkbook membership" in which financial contributions are the only sign of attachment of members to associations increasingly managed by professionals. Nonetheless, if information provision is the central mechanism through which union membership operates, the newsletters sent to members may have attitudinal effects (Muskett 2012). Based on a different perspective, Wollebaek and Selle (2002) suppose that face-to-face interactions are not necessary for associations to reinforce and transform the values shared by their members. Scandinavian countries in particular (Dekker and van den Broek 1998) show a high proportion of passive members feeling a sense of belonging to "imagined communities" (Anderson 1991). Because they are convinced that their financial contributions make a difference, even passive membership can increase generalized trust and reduce political alienation (Wollebaek and Selle 2002). From this perspective, multiple affiliations matter more than active involvement (Alexander et al. 2012;Wollebaek and Selle 2002) for the appearance of molding effects.
From a methodological viewpoint, we would argue that most existing literature does not carefully distinguish between selection and treatment effect of union membership. Although some studies acknowledge the possibility of a selection bias (Freeman 2003;Geissbühler 2000;Kerrissey and Schofer 2013;Kim and Margalit 2017), the inclusion of control variables is considered sufficient to eliminate the selection effect. In other words, it is implicitly assumed that individual selection into unions is based on observable characteristics (Winship and Morgan 1999). Alternatively, it is argued that distinguishing between union effects in right-to-work and non-right-to-work states in the United States-in the former membership is considered voluntary; in the latter quasi-compulsory-is sufficient to allay concerns about endogeneity (Kerrissey and Schofer 2013;Kim and Margalit 2017;Rosenfeld 2010Rosenfeld , 2014. 3 An incipient new literature on union effects seeks to tease out treatment effects from selection effects through appropriate research designs. For example, Kim and Margalit (2017) have argued that unions alter members' opinions on free trade. The bulk of their evidence relies on a matching estimator, which in turn is based on a set of observed covariates. This approach is in principle not dissimilar from a regression approach and open to the same threat of unobserved omitted variables. 4 However, their strongest evidence comes from the analysis of the United Auto Workers union, which shifts its view on free trade during the period of the survey. This is a more convincing test, but it is based on a union that is particularly active and whose members, predominantly low educated individuals, may be more malleable than the average member. In other words, it is unclear to what extent these results can be extended to unions as a whole. Ahlquist, Clayton, and Levi (2014) tackled the puzzle of International Longshore and Warehouse Union (ILWU) members being more likely to oppose trade liberalization than non-members, even though their rational self-interest should lead them to favor it. Their statistical evidence is again based on a matching technique, which, however, they complement with in-depth historical and sociological information about the union environment. Most importantly, they are explicit about the specificities of the ILWU case with regard to its market power, democratic internal governance, and external activism.
Another approach to the problem of disentangling selection and union effect has been the use of an instrumental variable estimation (Leigh 2006;Radcliff and Davis 2000). We discuss this modeling choice in the second subsection in the Average Treatment Effects section and in Appendix E in supporting information.
Here it will suffice to say that instrumental variable estimation is not a panacea: finding valid instruments is difficult; it reduces the efficiency of the estimates; and, if the population of "compliers" is not large or representative enough, it creates a problem of external validity of the results (Angrist and Pischke 2009: 150-72). Furthermore, just like the other studies summarized above, an instrumental variable approach treats the experience of union membership as a switch (on/off) variable, and therefore it is unable to capture effects that unfold in time.
Summing up, most existing literature argues for the existence of causal effects of union membership on political attitudes, and invokes a molding effect; that is, an attitudinal transformation as a result of organizational membership. However, most studies discard too quickly the selection effect. Working with longitudinal data, in this article we are able not just to distinguish between selection and molding, but to go a step further and explore additional, dynamic effects.
We draw on the idea that union membership should be treated as an experience good, which exerts its effects only gradually and with repeated exposure (Gomez and Gunderson 2004). This suggests that a newly affiliated individual may need some time to become aware of the benefits of unions and to be influenced by union dynamics. We expand the experience-good approach by hypothesizing that the gradual attitudinal changes provoked by union membership may precede joining. In this respect, a useful theoretical tool is the concept of anticipatory socialization developed by Merton and Rossi (1968). According to Merton and Rossi (1968), individuals aspiring to become members of a reference group develop attitudes congruent with other group members even before formal access to membership. This anticipatory socialization process serves a double function: it increases the likelihood of becoming part of the reference group, and it makes the adaptation to the new group easier. Crucially, for anticipatory socialization to play a role, the choice to join a union has to be a matter of choice, as opposed to imposition.
Drawing on these insights, we hypothesize the presence of anticipation effects of union membership, which would lead individuals to start altering their political attitudes even before joining unions. For example, workers may begin interacting with union members or being exposed to the arguments of union leaders even before joining the union. These interactions may lead non-union workers to modify their political attitudes. In due time, these workers may also decide to become union members. In line with the experience-good perspective, the membership experience may take time to provoke an attitudinal change (maturation effect), such that only individuals who are affiliated for a sufficiently long period see their attitudes re-molded. Finally, the political effects of union membership may require some time to be learned but, once they are assimilated, they are unlikely to be easily forgotten after leaving a union (Artz 2010).
The remainder of the article examines these issues empirically by analyzing panel data for Switzerland and the UK.
A Switch or a Process? / 473 Data The Swiss and British cases on which the analysis is focused provide interesting variation on the institutional context of union membership. 5 For the Swiss case, we use all waves of the Swiss Household Panel (SHP) in which the question about union membership is asked : 1999-2009, 2011, and 2014. Regarding the British case, we use both the British Household Panel (BHPS) and the UK Household Longitudinal Study (UKHLS). 6 The waves in which the membership questions are asked are : 1991-2008, 2010, 2012, and 2014. We focus on wage-earners-that is, on the target group for unions. 7 This leaves us with 51,299 observations on 14,312 individuals for the Swiss case and 108,786 observations on 12,388 individuals for the British case.
We examine the impact of union membership on interest in politics, political participation, closeness to any political party, partisan choice, and attitudes toward high incomes and taxation. Specifically, we consider the following dependent variables: interest in politics; participation in federals polls for Switzerland and, in the absence of a similar measure in the British panel, a measure of closeness to any party for Britain; electoral preference for the Socialist Party and the right-wing populist Swiss People's Party for Switzerland, and two variables capturing closeness to the Labour Party and the Conservative Party for Britain; attitudes about high-income taxation in Switzerland and about high-income ceiling in Britain. The latter variables are aimed to capture left-wing attitudinal shifts independent of shifts in partisan choice. For the British case, we also consider two variables about the strength of support for the Labour or Conservative Party. All non-binary variables are expressed on a 0-10 scale. The wording of survey questions and operationalization of the main independent variable and of the dependent variables are reported in Table 1.
In the models described below, we use a standard set of control variables appearing in previous literature: sex, age class, level of education, region of residence, economic sector, type of occupation, and a set of time dummies for each year taken into account. We select control variables which are clearly exogenous, because this is a very important aspect when estimating causal  • Non-member (0) (either because a union does not exist at the workplace or because the individual does not belong to the workplace union) • Member (1)  • Non-member (0) • Passive member (1) • Active member (2) 1991-1995, 1997, 1999, 2001, 2003, 2005, 2007 UK Interest in politics How interested would you say you are in politics? Would you say you are (10) • (obtained by multiplying by 10/3 the original values) 1991-1996, 2001-2008, 2010, 2012, 2014 476 / HADZIABDIC AND BACCARO • Strongly disagree (0) • Disagree (2.5) • Not agree/disagree (5) • Agree (7.5) • Strongly agree (10) • (obtained by multiplying by 2.5 the original values) 1992,1994,1996,1998,2001,2003,2006 effects (Wooldridge 2010: 53-57). 8 Besides estimating a main effect, for each dependent variable we also re-estimate the effect of union membership on the type of membership of respondents (active or passive). More information and descriptive statistics on the main independent variable, on the dependent variables, and on the controls are available in Tables C1 and C2, Appendix C in supporting information. 9 Average Treatment Effects In this section, we analyze union membership as a switch from the "nonmember" to the "member" status and estimate the average treatment effects associated with such event on individual political attitudes.
Model specification. We estimate three sets of models: (1) pooled ordinary least squares (OLS) without controls, (2) pooled OLS with controls, and (3) fixed effects with controls. Model 2 is similar to the existing cross-sectional literature but is vulnerable to the selection bias related to unobserved individual heterogeneity. Model 3 addresses this problem by including fixed effects. In algebraic terms: where the subscripts i and t represent individuals and time periods, respectively; Y it is an attitude; α the intercept; β is the coefficient of the union membership status; M it a dummy variable coded as 0 if the individual is a not a union member in a given year and 1 otherwise; 10 C it is the set of observed control variables described in Tables C1 and C2, Appendix C in supporting information, or the null vector for model 1; γ the coefficients of the control variables; ν i (excluded from models 1 and 2) captures the fixed effects; μ it is the error term assumed to be white noise. 11 8 The set of controls we consider is hence parsimonious, only including variables that are certainly exogenous. However, below we also discuss the inclusion of potentially endogenous control variables (e.g., job security, job change). 9 All statistical analyses have been produced using (and cross-checking) R and Stata 16. 10 In the analysis, we exclude individuals who become non-members after a previous spell of membership. An individual that has already experienced a membership spell is arguably not a true non-member if union membership has a durable effect. 11 A potential problem of measurement error has been cited in previous studies examining the impact of union membership on wages (Card 1996). We consider this problem negligible for our analyses. The most important reason is that we are interested in the attitudinal effects of union membership. If an individual declares herself a non-member while being in reality a member, it makes sense for us to consider her as a non-member because she obviously does not see herself as being part of the union (and vice versa for members). The problem would be completely different if we were examining the link between union membership and an objective outcome such as a wage premium, which does not depend on any subjective awareness of membership.
The coefficient of union membership when controlling for individual-level heterogeneity (model 3) should provide an estimate of the molding effect of union membership conceived as on/off switch, after adequately controlling for the selection effect. Furthermore, by comparing models 1, 2, and 3 we are able to assess to what extent the selection bias can be explained by observed control variables or by the unobserved heterogeneity between union members and non-members.
We also analyze, to our knowledge for the first time, the attitudinal impact of leaving a union, using the same three types of models as above. The only difference concerns the definition of the treatment variable: M it is coded as 0 if an individual is a union member and 1 if she is a non-member having been a member in at least one previous year.
To correct for heteroscedasticity and serial correlation, we use cluster-robust standard errors, with the individual as the cluster unit.
Are time-varying omitted variables a problem?. Before proceeding further, we need to discuss a potential problem with our empirical approach: In principle, controlling for fixed effects leaves the estimates open to a problem of time-varying endogeneity; that is, unobserved time-varying factors that are correlated with both the change in the union membership variable and the change in the outcome variables.
We try to address this problem in three ways. First, our specifications include time dummies, and these should be able to control for time-varying shocks occurring at the macro level, such as broad economic trends, which may be correlated with broad trends in unionization and political attitudes.
Second, we explicitly model one possible time-varying shock at the individual level: job security. A decrease in perceived job security (a time-varying shock) might simultaneously increase the individual willingness to join a union and change the individual's political attitudes. These models (available in Appendix H in supporting information), show that controlling for job security leaves our estimates of the union effect unchanged. 12 Third, we also experiment with instrumental variable models. For reasons discussed in Appendix E in supporting information, when instrumenting the union membership variable, we replace the fixed effects estimator with a generalization of the first differences estimator, which we refer to as "adjusted differencing procedure." Differently from the fixed effects estimator, this approach allows us to use instruments in levels, as opposed to differences 12 In addition, one may argue that changing workplace and/or job may overlap with membership transitions and hence represent a confounding factor. Additional analyses available in Appendix I in supporting information show that the estimates of average treatment effects remain unchanged if we restrict the analyses only to those individuals who change neither workplace nor job during their participation in the survey. (Wooldridge 2010: 354). We are able to find convincing instruments only for the Swiss case due to the lack of sufficiently nuanced geographical information in the British panel. The results of the adjusted differencing procedure applied to the Swiss case (reported in Appendix H in supporting information) suggest that two-stage least squares (2SLS) and OLS estimates are not significantly different from one another. Thus, we have some evidence (at least for the Swiss case) to suggest that time-varying omitted variables are unlikely to bias our estimates of the union effect.
One further methodological aspect deserving to be discussed is our use of a linear estimator (OLS) throughout, and in particular of a linear probability model (LPM) for voting behavior and partisan identification, which are binary variables. Estimates issuing from non-linear models depend on the amount of residual variance and this generates two types of problems (Mood 2010): it is not possible to compare nested models or models estimated on different subgroups, and it is not possible to directly interpret the estimates as marginal effects without additional transformations. In a panel data setting like ours, the last issue becomes crucial. In fact, the only non-linear model that permits us to control for time-invariant unobserved heterogeneity is the conditional logit model (Wooldridge 2010: 619-22). Wooldridge (2010: 622) shows that the estimates of such a model cannot be converted into average treatment effects, which is the main goal of our analysis.
Results. Because we estimate a large number of models and are only interested in the impact of the union membership status, we only report, in graphical form, the average treatment effects (ATEs) of union membership for the Swiss and British cases. Tables with the union effects (Tables D1-D36) are provided in Appendix D in supporting information. Full models including control variables are available in Appendix G in supporting information.
For each dependent variable, we run the three models described above (OLS with no controls, OLS with controls, and fixed effects with controls) on the overall sample for both the joining and leaving phase. As the presence of a selection effect makes the first two models clearly inappropriate, we then provide only the fixed effects estimates for the impact of union membership by type of membership (passive/active) and consider only the joining phase because we do not have enough within variation to produce reliable estimates for the leaving phase. In Figures 1-3 we plot the union effects by distinguishing those significant at least at the 5 percent level from the other ones.
The estimates plotted in Figures 1 and 2 reveal clearly that most effects usually attributed to joining or leaving a union are largely the result of self-selection. The first set of models ("OLS with no controls") gives a purely descriptive account of the relationship between union membership and political 480 / HADZIABDIC AND BACCARO attitudes: union members are on average more politically involved, closer to left-wing parties than right-wing ones, and more favorable to taxing/putting a ceiling on the income of the rich than non-members in both countries. The effects are reversed when individuals leave the union membership status even though the differences with those who remain are less pronounced than for the joining phase.
Including a set of observed controls in the second group of regressions ("OLS with controls") reduces the magnitude of all coefficients, but the coefficients remain highly significant almost everywhere. Regarding the leaving phase, there is a reversed pattern with smaller magnitudes (in absolute value) and less significant estimates. However, for some outcome variables (e.g., "Opinion on high income taxes" for the Swiss case or "Feeling close to no party vs any party" for the British case) we observe an increase in the magnitude and significance of the estimates when controlling for observable covariates.
The magnitude of all coefficients declines dramatically with fixed effects and almost all coefficients become insignificantly different from zero for both transitions. For example, if we consider the effect of joining a union on "Participation in federal polls" for Swiss workers, the first pooled OLS model gives a highly significant estimate (0.95, p < 0.1%), which decreases but remains highly significant when observable characteristics are controlled for in the second model (0.61, p < 0.1%), and finally becomes a slightly negative but insignificant coefficient in the fixed effects model (−0.051, p > 5%).
Controlling for fixed effects, joining or leaving a union has no significant effect on: political participation and feeling close to any party (with the exception of small negative and barely significant effects for the UK [−0.081, p < 5%] on interest in politics and on feeling close to at least a Party when leaving a union [−0.016, p < 5%]); partisan preference (with the exception of a small negative impact [−0.017, p < 1%] regarding feeling close to the Labour party when leaving a union); and attitudes regarding redistribution from high incomes. The effect on joining a union for "Interest in politics" in Switzerland is among the few that remain highly significant even in the fixed effects model, although the magnitude is small (0.14, p < 1%). Interestingly, among British workers there is a rather strong and highly significant negative impact of joining a union on the "Strength of support for the Conservative Party" (−0.50, p < 0.1%).
Turning to how the effect varies between passive and active membership, even though the estimates are in most cases higher for active members, they (1) as expected, joining a union increases interest in politics only for active members in Switzerland (0.24, p < 0.1%) and in the UK (0.23, p < 1%); (2) becoming an active member leads to an increase in the likelihood to have a preferred party (0.032, p < 5%) in the UK; and (3) surprisingly, union membership decreases the strength of support for the Conservative Party more for passive members (−0.60, p < 0.1%) than for active members in the UK. Figure 3 reports these results.

Dynamic Analysis
So far, we have treated union membership as a switch event in the life of a worker: an on/off switch that may, or may not, set in motion a set of attitudinal changes. In this section, we adopt a different viewpoint: we treat membership as a continuous process whose effects unfold dynamically before, during, and after the membership experience. In particular, we are interested in Model specification. To detect anticipation effects, we perform (to our knowledge for the first time) a dynamic leads and lags analysis (Powdthavee 2011). In these specifications, we use ten dummy variables, each one capturing a specific moment in the union membership trajectory. 13 In equation form: We use again a fixed effects estimator to estimate at the same time the effect of all time dummies while controlling for time-invariant unobserved heterogeneity.
For the effect of leaving a union, we proceed analogously and create ten dummies, each one identifying a specific year in the period before leaving a union and afterward. We choose as reference point the most distant time dummy (5 years before) from the transition we consider (joining or leaving a union). This reference point makes sense because it is the furthest from the two transitions, thus representing a good approximation to the steady state. The leads and lags analysis introduces multicollinearity, because consecutive leads and lags are highly correlated with each other. Standard errors are likely to be larger than usual, thus any trend in the evolution of the estimates is more important than statistical significance of single estimates. Furthermore, the analyses are restricted to the first spell of membership. 14 The way we deal with data gaps is discussed in Appendix F in supporting information. 15 Results. The estimates of the leads and lags analysis are plotted in two separate graphs for the joining and for the leaving transition, respectively. Regression results are synthetically reported in Appendix D, Tables D37-D48 in supporting information. In plotting the graphs, we include 95% confidence intervals and provide explicit tests of the significance of key differences in estimates. In commenting, we focus on the clearest trends only.
Beginning with interest in politics in Switzerland (Figure 4), there is an increasing trend in the dependent variable, which starts 2 years before the transition to union membership. The increase continues during the membership phase until the third year (difference in coefficients 3 years after membership versus 2 years before: 0.55, p < 0.1%). Interestingly, the increased interest in politics seems durable because no significant decrease is observed in the leaving phase.
Participation in federal polls in Switzerland displays a similar anticipation effect ( Figure 5). The increase begins before joining a union and continues gradually during the membership phase (difference in coefficients 5 years after membership versus 5 years before: 0.92, p < 0.1%). As with interest in politics, there is no decline in participation after leaving unions. A Switch or a Process? / 485 With regard to the British case, the most interesting findings of the leads and lags analysis pertain to party support. In Figure 6, we see that the propensity to support the Labour party increases 2 years before joining and continues at least until the first year of membership (difference in coefficients 1 year after membership versus 2 years before: 0.027, p < 1%). The opposite trend is observed in the leaving phase (difference in coefficients 1 year after leaving versus 2 years before leaving: −0.031, p < 5%).
When looking at the propensity to support the Conservative Party (Figure 7), we observe a continuous decrease both in the joining (difference in coefficients 4 years after membership versus 2 years before: −0.027, p < 5%) and in the leaving phase (difference in coefficients 5 years after leaving versus 5 years before leaving: −0.041, p < 5%). Interestingly, union members reduce their propensity to support the Conservative Party not only during the membership spell but also after leaving unions.
A similar pattern is found with regard to the strength of support for the Conservative Party in Britain (Figure 8): there is a continuously decreasing trend before joining the union and during the union membership phase (difference in coefficients 3 years after membership versus 5 years before: −1.17, p < 1%). Leaving a union is also associated with a general decline in the strength of support for the Conservative Party (difference in coefficients 5 years after leaving versus 5 years before: −1.62, p < 1%).
The remaining variables for the Swiss (voting intentions and opinion on taxes on high income) and British (interest in politics, propensity to support a party, strength of support for the Labour Party, and opinion on income ceiling) cases do not show important significant trends. The relative graphs are included in Appendix D ( Figures D1-D7 in supporting information).

Discussion
A large literature on union effects, and an even larger literature on the attitudinal effects of secondary associations, make broad-ranging claims about the transformational impact of membership on political attitudes. The analysis presented above has shown that, while such claims are not unfounded, union effects are more nuanced than it has so far been assumed. A Switch or a Process? / 487 Due to the heavy reliance on cross-sectional data and associated approaches, the existing research has largely lumped together two very different effects of unions: the selection effect-implying that unions attract workers who are systematically more interested in politics, willing to participate, and likely to vote for certain parties than non-joiners-and the molding effect, i.e., the ability of union membership to alter individual attitudes toward politics. Differences between members and non-members, which are really a feature of self-selection, have often been attributed to the causal impact of union membership.
Virtually all of the union effects estimated through pooled OLS with controls (the equivalent in our setting of the dominant strategy of estimating cross-sectional regressions with controls) clear the threshold of statistical significance. Yet very few of these effects survive when individual-level heterogeneity is controlled for. Our estimates of the overall treatment effect of unions (summarized in Table 2 together with other main results) suggest that in only two cases union membership has a significant causal impact on wage earners as a whole: joining a union produces a small increase in interest in politics in Switzerland (0.14 on a 0-10 scale); becoming a union member

/ HADZIABDIC AND BACCARO
decreases the strength of the support for the Conservative Party (-0.50 on a 0-10 scale). Furthermore, the distinction between active and passive membership-a dimension highlighted by the literature on civil culture (Almond and Verba 1989)-is much less important than expected. Almost all effects that are insignificant in the overall sample are also insignificant for both passive and active members separately. The only exceptions are interest in politics in Switzerland and the UK, for which the union effect is significant only for active members; the propensity to declare a preferred party; which increases for active members in the UK; and the strength of support for the Conservative Party, which decreases only for passive members in the UK.
It is also interesting to note that the selection bias linked to unobserved heterogeneity is much larger than the one accounted for by the observed covariates, as revealed by the comparison of pooled OLS, pooled OLS with controls, and fixed effects in Figures 1-3. This suggests that the strategy of controlling for a rich set of covariates under the assumption that individual selection into unions is "on observables" is unlikely to address the problem of self-selection.
Yet, the fact that treatment effects of unions are few and sparse for Swiss and British workers as a whole does not mean that the experience of union   membership has no impact on the political attitudes of workers. Exploiting the longitudinal structure of the data, the dynamic analysis has revealed that the transformational process associated with union membership begins, in some cases, before workers join unions. We have found strong anticipation effects for interest in politics and participation in federal polls in Switzerland and for supporting a given party, supporting the Labour Party, supporting the Conservative Party, and the strength of support for the Conservative Party in the UK. The variables showing an anticipation effect present also clear maturation effects, with the impact of union membership increasing and becoming noticeable only after a certain number of years of membership. Finally, the analysis of the leaving transition shows that some effects last even after leaving unions (those on interest in politics and participation in federal polls in Switzerland and the effect on the support for any party in the UK). Other effects tend to increase even more (support and strength of the support for the Conservative Party), while more rarely the effect tends to fade out after leaving a union (support for the Labour Party).
Taken together, these results resonate with the experience-good model of union membership, which suggests that the union experience has gradual effects (Gomez and Gunderson 2004), but also broaden it by showing that attitudinal changes can precede formal membership and in general do not dissipate or even continue after leaving the union. The coexistence and combination of anticipation and molding effects suggests that, rather than thinking of union membership as an "on/off" event, it is more appropriate to treat it as a process unfolding in time: before, during, and after the change in union status. Some workers may start modifying their attitudes due to contacts with unionized coworkers before changing membership status. Social exchanges may lead both to attitudinal change and to a change in membership status. Alternatively, non-unionized workers may be influenced by the message and communication strategies of external unions. It may be that workers experiencing an anticipation effect are individuals that join unions at least in part because of political reasons (either because they have had political discussions with individuals who are already members or because of the role of unions as political actors in the public sphere). These individuals experience an anticipatory socialization process (Merton and Rossi 1968) that makes their views similar to those of group members seen as reference group. Conversely, individuals that become members only because of economic reasons (e.g., looking for income or employment protection) are much less likely to show changes in their political attitudes before joining. We leave it to future research to explore the micro-mechanisms that may be responsible for the anticipation effects.
For the time being, it suffices to say that if the process of anticipatory attitudinal change is important enough, the potential impact of formal union membership may be limited. In other words, the newcomer may come to resemble the existing members so much that there is only a very small leeway, or no leeway at all, to move closer. This may explain the absence of important average treatment effects. Furthermore, in an ongoing paper project we show that the effect is highly heterogeneous depending on pre-membership attitudes. Workers whose pre-membership attitudes are neither too far nor too close to those of existing union members are those most likely to be affected by the union experience.
Thinking about the external validity of such dynamic patterns, we cannot exclude that the anticipation effects may be less relevant in industrial relations systems in which the act of joining a union is less voluntary than in the UK and in Switzerland. In fact, if the worker does not choose, but is rather compelled to become a member, anticipatory socialization is less likely to play a role (Merton and Rossi 1968). Thus, anticipation effects may be less central in North American industrial relations systems (the United States and Canada), in which the joining act is not an individual choice, but the result of a majoritarian decision at the workplace. This would imply the existence of two distinct types of union membership regimes: the one predominantly based on voluntary membership would be associated with attitudinal changes throughout the membership trajectory (before, during, and after joining) and a predominance of selection effects over molding effects; while the one predominantly based on a constrained form of membership would produce political effects only during and potentially after joining a union, and selection effects would be less predominant.
The distribution of the membership duration we observe in Switzerland (Table C3, Appendix C in supporting information) and in the UK (Table C4, Appendix C in supporting information) provides some evidence about the voluntary nature of union membership in these systems. The most striking pattern is that in both countries most individuals that join unions remain members only for short spells. The biggest leak of members takes place between the first and the second year of membership (after the first year, 72 percent leave a union in Switzerland, while 50 percent exit a union in the UK; both decreases are only marginally affected by attrition issues). The decrease continues for longer membership durations, but at lower rates. If North American unions are indeed characterized by a more constrained form of membership, we should expect the average membership duration to be longer. Longer membership spells, in turn, would provide unions with greater leeway to provoke important molding effects. We leave it to future research to determine whether these expectations about two fundamentally different union membership regimes are empirically supported.
A Switch or a Process? / 493

Concluding Remarks
With this article we hope to have shown that a longitudinal analysis of union effects is vastly preferable to even the most methodologically sophisticated cross-sectional analysis. The reason is that the attitudinal effect of union membership is better approached as a continuous process than as an on/off switch. While the switch approach only allows us to distinguish (in the best of circumstances) between two types of effects on individual attitudes-selection and molding-the process approach also detects anticipation and maturation effects. By embracing this approach, we were able to show that in Switzerland and the UK unions mostly attract like-minded workers, while the ability of the membership experience to modify the political attitudes of workers is on average more limited. Nonetheless, there are interesting anticipatory effects, which have gone hitherto unnoticed. For example, workers start increasing their interest in politics before they formally join. We have attributed this finding to the mechanism of anticipatory socialization. At this stage, however, we cannot exclude that in countries such as the United States and Canada, where union membership is determined in large part by institutional arrangements forcing workers to join unions, the presence of anticipatory socialization patterns could be less relevant than in Switzerland and the UK.
Unions effects are more than likely to vary by type of unions. Large, encompassing organizations behave differently from small, sectional ones (Olson 1971). This is likely to affect the ability of unions to shape worker attitudes and the direction of their influence. Ideally, we would like to know which union the worker belongs to, the union's position on key policy issues, and its organizational practices. In addition, the distinction between passive and active membership is probably too coarse to capture the level of workers' engagement with unions. We would want to have information about union meetings, assemblies, worker referenda, frequency, content, etc. Yet this information is not available in any of the large panel surveys of which we are aware.
Going forward, it would also be interesting to extend this type of research to other associations. Our results suggest that the molding effect of secondary associations on members' attitudes is more nuanced than previously thought. Per se, this does not necessarily invalidate the argument about the "civilizing effect" of secondary associations (Lazer et al. 2010;Putnam 2001), but invites a critical reexamination of that literature, the bulk of which is based on crosssectional studies that are potentially subject to the same type of problems we have identified for unions. More generally, future research should try to determine which types of associational settings have the greatest ability to alter the way an individual interprets and interacts with the surrounding world.