Abstract
The key assumption of spiral of silence theory is that opinion climate perceptions affect political opinion expression. We meta-analyzed the strength of this relationship and clarified the impact of theoretically relevant moderators. Sixty-six studies collectively including more than 27,000 participants were located. We observed a significant positive relationship (r = .10; Zr = .10) between opinion climate and opinion expression. This relationship was not weaker in online as compared with offline opinion expression environments. Also, the relationship did not vary by the number of the targets of opinion expression, the opinion of the targets, the opinion climate characteristics, and the design, measurement, and sample characteristics. The largest silencing effect (r = .34), however, was observed when participants talk to their family, friends, or neighbors about obtrusive issues. Overall, our findings suggest that the relationship between opinion climate perception and political opinion expression is stronger and more robust than previously thought.
The theory of public opinion, usually referred to as the spiral of silence (Noelle-Neumann, 1974, 1993), has become one of the most studied theories of political communication and public opinion research. The theory starts with the premise that individuals fear becoming socially isolated, and as a consequence, they constantly monitor their opinion environment (Hayes, Matthes, & Eveland, 2013), such as their interpersonal networks or the mass media. The observation of the opinion environment helps individuals to determine whether the majority shares their own political views or not. One key hypothesis of the theory is that opinions that are held by the perceived majority are more likely to be expressed publicly compared with opinions shared by a perceived minority (Gonzenbach & Stevenson, 1994; Lasorsa, 1991; Petrič & Pinter, 2002; Scheufele, Shanahan, & Lee, 2001; Shamir, 1997; Willnat, Lee, & Detenber, 2002). The majority of research on the spiral of silence has focused on testing this assumption, although it is a simplification of the theory. Based on that, the theory further posits a spiraling process over time in which perceived majority opinions become dominant and perceived minority views are silenced (Matthes, 2015).
Since its original articulation in the early 1970s, spiral of silence theory has received a tremendous amount of scholarly attention from around the world covering numerous political topics using a diverse array of methodologies. Concurrently, the spiral of silence is a debated and much challenged theory in public opinion research (Hayes & Matthes, 2017; Matthes & Hayes, 2014; Salmon & Kline, 1985; Scheufele & Moy, 2000). In a recent essay about the “retirement of concepts,” E. Katz and Fialkoff (2017) even suggested that the spiral of silence is a theory that communication scholars may abandon in the future. In fact, one could argue that the upsurge in partisan media may undermine a consonant opinion climate, which is one of the key prerequisites for a spiraling process to unfold (see Moy & Hussain, 2014). The theory has been developed in the context of classical mass media such as newspaper and television. With the skyrocketing rise of online media, pressures to suppress minority opinions may decline. This challenges the theory’s key assumption (Schulz & Roessler, 2012). This suggests that it is appropriate to revisit the theory and its core assumption. This article aims to do that by means of meta-analysis.
The purpose of this meta-analysis is to estimate the average effect of climate perceptions on political opinion expression. We follow the footsteps of the only available meta-analysis on the spiral of silence conducted by Glynn, Hayes, and Shanahan (1997) who found a significant but very small silencing effect (see Shanahan, Glynn, & Hayes, 2007). Glynn and Huge (2014) recomputed this single effect with an additional 12 studies without examining the role of moderators. However, we go beyond both studies in several important ways. First, the Glynn et al. (1997) study is based on 17 studies (plus an additional 12 in Glynn & Huge, 2014), and numerous studies have been conducted after this study was published. In contrast to prior research, we also examined experiments in addition to surveys. Second, and more importantly, the Glynn et al. (1997) meta-analysis was conducted prior to the age of online media, and also the update by Glynn and Huge (2014) did not investigate the role of online media. As Perloff (2015) stated with respect to media effects theories, “New media, particularly blogs, social media, and galloping mobile phone technologies, are changing media use patterns and reshaping the arc of effects” (p. 551). In fact, as we will explain in more detail below, there are several theoretical arguments for why the online age could make the spiral of silence obsolete. There is thus a strong need to compare online and offline opinion expression in the context of the spiral of silence. A meta-analysis is the perfect way to do this.
Third, also related to the small number of studies in Glynn et al. (1997), the authors could not find any substantial moderated effects potentially helping to understand the conditions under which climate perceptions and opinion expression are related or not (Glynn & Huge, 2014, did not examine such effects). In addition, there is no meta-analytic knowledge about the role of the type of the opponent when expressing an opinion (i.e., express opinions toward friends or strangers), nor is there information on the role of issue characteristics such as issue obtrusiveness. We also lack an understanding about the reference group related to the opinion climate (i.e., the whole population or close friends). All these variables arguably play a role in shaping the effects of opinion climate on opinion expression. Finally, we lack any knowledge about the role of the study and sample characteristics that may contribute to the detection of effects in line with the theory. This meta-analysis seeks to address these shortcomings. We integrated research published through July 2016 incorporating 66 independent studies with a total of 27,192 participants. We performed moderator analyses to test the impact of key, theoretically relevant variables and then estimated the role of the study, measurement, and sample characteristics. In terms of methodological advancement, we applied multilevel modeling accounting for the dependence of multiple effect sizes, and we estimated publication bias—both of which are important issues in meta-analysis (Borenstein, Hedges, Higgins, & Rothstein, 2009). In doing this, our meta-analysis aims to provide theoretical and methodological explanations for the factors that influence the size of the effects of the opinion climate on opinion expression. We then suggest some directions for future research on the spiral of silence.
The Key Assumption of the Theory
The willingness to express an opinion is the most important concept and the key dependent variable of the theory. In her seminal study, Noelle-Neumann (1974) measured opinion expression with the so-called “train-test”—a procedure that asks respondents to imagine they are on a train journey on which another passenger is voicing an opinion. In this test, respondents are asked if they would be willing to voice their views in front of the passenger. However, the hypothetical nature of this test has been frequently criticized (Glynn et al., 1997; Hayes, Uldall, & Glynn, 2010; Scheufele et al., 2001). In fact, the majority of studies on the spiral of silence have followed a different approach by estimating the effect of the perceived climate of opinion on individual’s willingness to express their views to friends, strangers, or people with opposing views (see Glynn & Park, 1997; Ho & McLeod, 2008; C. Katz & Baldassare, 1994; Matthes, Morrison Rios, & Schemer, 2010; Willnat, 1996). Most of these studies have observed a positive, albeit small correlation between climate perceptions and the willingness to express an opinion (Glynn et al., 1997). The theory, however, does not assume that the relationship between climate perceptions and opinion expression holds for all individuals. Noelle-Neumann (1974, 1993) theorized that the so-called “avant-garde” voice their views even at a time when they are still in the minority. Likewise, so-called “hardcore” individuals are those “who are not prepared to conform, to change their opinions, or even to be silent in the face of public opinion (Noelle-Neumann, 1974, pp. 48-49).
In the only full meta-analysis available to date, Glynn et al. (1997) identified 17 studies based on responses from over 9,500 participants. They examined if people who perceive less (current or future) support for their opinions are less willing to express their views compared with those who perceive more support. The average effect was very small (r = .054) but highly significant. Glynn and Huge (2014) recomputed and confirmed this effect (r = .050) with an update of 12 studies, but without conducting any moderator analyses. In Glynn et al.’s study, whether opinion climate perceptions refer to current or future support did not make a difference for the overall relationship. Furthermore, it did not matter whether respondents expressed an opinion to a member of the media, strangers, community members, and/or friends. It did not make a difference whether or not the target of the opinion expression disagreed. Therefore, the effect seemed to be robust with respect to different measurements. Based on this meta-analysis and findings from previous scholarship, we can formulate our first hypothesis:
Despite its landmark status, the meta-analysis by Glynn et al. (1997) as well as the update by Glynn and Huge (2014) left some questions unanswered. First, Glynn et al. (1997) only included the available empirical evidence until 1995. Based on the comparatively small sample size, theoretically and methodologically relevant moderators could not be examined. Glynn and Huge (2014) did not look at moderators at all. Second, both analyses looked at survey studies only, thus neglecting experiments. One could argue that effects are stronger in experiments compared with surveys (Matthes & Arendt, 2017). Third, and more importantly, neither study analyzed the role of online media and other theoretically important moderators. Clearly, there are strong reasons to revisit the available empirical evidence including the estimation of an overall effect as well as an analysis of key moderating variables.
Moderators
Offline Versus Online Opinion Expression
In its original conception, the spiral of silence is a theory that aims to predict the opinions expressed in face-to-face settings. More recently, however, scholars have begun to explore the relationship between opinion climate and opinion expression in online contexts (e.g., Ho & McLeod, 2008; W. Lee, Detenber, Willnat, Aday, & Graf, 2004) or social media (Gearhart & Zhang, 2014). Also, anonymous and non-anonymous online settings have been compared (Yun & Park, 2011). While most studies concluded that the spiral of silence might still work in the online world, the role of online environments is far from understood. There are two competing theoretical views (Metzger, 2009).
On the one hand, the effect of the opinion climate on individuals’ tendencies to divulge their opinions may be weaker in an online than in an offline setting (Ho & McLeod, 2008). The online environment allows a huge variety of opinions, and those with minority views can more easily find opinion support relative to offline venues. By the same token, individuals may simply self-select content that concurs with their views—thus, the majority pressures may be weak. Especially in anonymous online settings, unpopular opinions can be easily voiced without putting personal relationships at stake as it is the case in interpersonal communication (see Matthes, 2013; Mutz, 2002). Likewise, the majority of online opinions are expressed with text, while offline interpersonal discussions also include nonverbal communication. It can be argued that nonverbal reactions are perceived as an additional cue that particular views are unwanted and therefore ostracized. The online world, by contrast, offers fewer social cues and is thus less intimidating when expressing opinions (Ho & McLeod, 2008). The same is true for so-called “click speech,” that is, people “like” and share content without having to produce their own message content (Pang et al., 2016). Such click speech does not necessitate much cognitive effort and blatant commitment. Thus, it may not depend on the majority opinion climate. Schulz and Roessler (2012) argued that it is less likely to reach consensus on what the majority opinion climate is in an online environment compared with an offline environment because there may be multiple opinion climates online depending on the content that the users select. In line with this, Iyengar and Hahn (2009) presented evidence for ideological selectivity and the preference of media users to expose themselves to media contents based on partisan affinity. Finally, the role of “echo chambers” and “filter bubbles” (Pariser, 2011) has been discussed in relation to online media consumption “in which algorithms inadvertently amplify ideological segregation by automatically recommending content an individual is likely to agree with” (Flaxman, Goel, & Rao, 2016, p. 299). All of this suggests that the spiral of silence is weaker online than offline.
On the other hand, one could argue that the boundaries between online and offline communication are becoming increasingly blurred (Pang et al., 2016). Individuals may go back and forth between talking to their social contacts online and offline. For instance, Facebook contacts often are based on real-world social relationships (Pang et al., 2016). Furthermore, studies within the online ostracism paradigm have repeatedly demonstrated that being ignored or excluded online can pose a serious threat to our need for belonging—similar to our fear of being socially isolated (Williams & Nida, 2011). As Metzger (2009) has stated, political orientations are often not disclosed on social network sites because users may fear losing a friend or professional contact. In fact, politically motivated unfriending on social media is a new kind of political gesture that can be understood as a new form of exerting social pressure. If social media contacts “unfriend” a user, then this may trigger fear of isolation setting the spiral into motion. All of this suggests that the fear of not holding a majority viewpoint is still a likely consideration before expressing opinions online (Pang et al., 2016).
As there are multiple ways of expressing one’s views online and offline, no single study can ever answer the question of whether or not the relationship between perceived opinion climate and opinion expression is stronger offline or online. Meta-analysis, by contrast, is a perfect way to answer that question because the entire available empirical evidence can be considered. Because there are competing views on the likelihood of a spiral of silence to occur in online and offline settings, we formulate a research question rather than a hypothesis:
Issue Characteristics: Controversiality and Obtrusiveness
Noelle-Neumann (1993) theorized that a spiral of silence can only occur when the topic of discussion contains a strong moral component, that is, a topic is controversial and emotionally laden. She argued that without a moral component, there will not be a strong social pressure to suppress one’s views in public situations. In other words, when one voices minority views on morally laden issues, then individuals risk isolating themselves from the morally correct majority. In fact, most prior studies have analyzed the spiral of silence with respect to morally controversial topics such as affirmative action (Moy, Domke, & Stamm, 2001), abortion (Salmon & Neuwirth, 1990), genetically modified food (Scheufele et al., 2001), or naturalization of immigrants (Matthes et al., 2010). And most of these studies found support for the theory. Based on Noelle-Neumann’s (1974, 1993) original theory, we hypothesize,
Obtrusiveness is another issue characteristic discussed in the spiral of silence literature. It is defined as the amount of personal experience with an issue (Winter, 1981). For issues that obtrude in our lives, we have more firsthand knowledge and are less reliant on the news media. This makes media effects less likely compared with unobtrusive issues (Zucker, 1978). In line with this argument, Salmon and Neuwirth (1990) argued that the obtrusiveness of an issue is a key determinant of willingness to express an opinion (see also Moy et al., 2001; Willnat, 1996). This is because salient issues are generally perceived as more important than less salient issues, that is, they matter to one’s life. As Willnat et al. (2002) put it, “those who consider an issue more important or are more interested in public affairs in general, might be more willing to discuss these issues in public, even when faced with an opposing majority” (p. 400). This leads us to suggest that the effects of climate perceptions on opinion expression are stronger for unobtrusive compared with obtrusive issues. For obtrusive issues, people may be inclined to speak their minds independent of the opinion climate.
However, other research suggests quite the opposite. In fact, there are grounds to theorize that individuals are more likely to monitor their opinion environment for obtrusive compared with unobtrusive issues (Willnat et al., 2002). That is, it is quite important to see what others think about an issue when the issue is relevant to my life. As a consequence, obtrusive issues in particular might foster a spiral of silence. In line with this, Willnat et al. (2002) found that “perceptions of majority opinion only affected the outspokenness of those who were concerned about an issue and who thought that their viewpoints were losing ground among the public” (p. 408). Furthermore, given that people care more about obtrusive issues than about unobtrusive issues, one could argue that the risk of social isolation is weaker for issues that most people have no direct experience with. The pressure not to voice dissenting opinions may thus be weaker for unobtrusive issues than for obtrusive ones. This would lead us to suggest the effect of climate perceptions on speaking out are stronger for obtrusive than unobtrusive issues. Based on these conflicting theoretical considerations, we pose the following research question:
Opinion Climate Characteristics: Reference Group and Time Frame
The reference group of the opinion climate refers to the population for which an opinion climate is perceived. Noelle-Neumann (1974) operationalized the reference group as the majority of a country (see also Shamir, 1997; Willnat, 1996; Willnat et al., 2002). Information about this reference group can be conveyed through the mass media (Salwen, Lin, & Matera, 1994) or interpersonal communication. In contrast to the original theory, scholars have frequently operationalized the reference group with respect to strong-tie relationships such as family, friends, and neighbors (Neuwirth & Frederick, 2004; Salmon & Kline, 1985). As Oshagan (1996) put it, “Other possibly important sources of social influence such as family members, close friends, co-workers, or neighbors do not necessarily constitute the overall majority opinion of a system, and may even be opposed to the systemic majority opinion” (p. 336). Based on this literature, we believe it is worthwhile to compare Noelle-Neumann’s (1974) original conceptualization of reference groups as the majority opinion with an operationalization of reference groups as strong-tie relationships. Of course, the psychological situation as well as the societal consequences are totally different when comparing the majority with strong ties.
Interestingly, there is a long-standing body of evidence in the research on majority social influence suggesting that conforming to the majority increases when the discrepant person is a member of the reference group (Cialdini & Goldstein, 2004). That is, conformity pressures are weaker as the reference group becomes less relevant and more distant. This is also in line with the findings of Oshagan (1996). He found that, “when both forces are made salient, reference groups take primacy over societal majorities in influencing individual opinions” (p. 349). Based on these findings, we theorized that strong-tie reference groups especially drive the spiral of silence. Strong ties refer to contacts with trusted others such as family, friends, and neighbors. It can be theorized that the opinion climate with respect to family, friends, and neighbors is more important for shaping opinion expression than the majority opinions voiced in the media, the government, or by unknown others. The reason is simply that fear of isolation is much higher when it comes to strong ties compared with rather weak ties (Oshagan, 1996). That is, it is worse to be isolated by strong ties compared with the general public, even though isolation from strong ties may not be very likely. Again, no study has ever been able to test this explicitly. Thus,
In her seminal article, Noelle-Neumann (1974) explicitly distinguished between current and future opinion climate (i.e., time of the majority opinion climate). More specifically, she clearly recommended operationalizing future opinion trends: If there is a divergence in the assessment of the present and future strengths of a particular view, it is the expectation of the future position which will determine the extent to which the individual is willing to expose himself. . . . If he is convinced that the trend of opinion is moving his way, then the risk of isolation is of little significance. (p. 44)
However, the empirical support for this claim seems inconsistent (see Ho & McLeod, 2008; Salmon & Neuwirth, 1990). In the meta-analysis by Glynn et al. (1997), there was no difference whether studies measured future or current opinion climates. Therefore, we pose a research question rather than a hypothesis:
Target of Opinion Expression
One of the less frequently examined aspects is the target of opinion expression, that is, toward whom a dissenting view is voiced. Glynn et al. (1997) investigated strangers, community, and the media, and found that targets did not matter. The train test by Noelle-Neumann (1974) used strangers as targets, while other studies have used the media (i.e., expression of minority opinions in front of a camera; see C. Katz & Baldassare, 1994). We are not aware of studies that systematically compared the role of several targets of opinion expression for the theory. This lack of research refers to the type of target (i.e., strangers, the media, family, etc.), the number of targets (i.e., single or multiple), and the opinion of the targets (i.e., disagreeing, agreeing).
So why should the target of the opinion expression matter at all? When citizens express minority views, they have to expect negative reactions by the target of the expression. Ultimately, there is a risk that dissenting views can put social relations at stake (Mutz, 2002). It seems reasonable to argue that this risk is especially present in the case of strong-tie targets such as family and friends. Also, one could assume that dissenting opinion suppression is less likely when the target is agreeing rather than disagreeing. Likewise, the risk to social relations may be higher when there are multiple targets compared with speaking one’s mind to only one person. None of these important aspects have been systematically investigated. Thus, we ask,
Design, Measurement, and Sample Characteristics
Findings may not be uniform across designs, measures, and samples. In meta-analysis, design, measurement, and sample characteristics are typically determined by the characteristics of the sampled studies. Based on the studies we collected, we focus on three sets of variables. First, experimental studies allow strong causal inferences but are often criticized with respect to validity. Surveys, by contrast, are often more externally valid but are prone to spuriousness. Second, measurement of opinion expression include diverse concepts such as discussion, political participation, opinion avoidance, or response latencies. Third, sample characteristics typically cover age, gender, and origin as well as the question of student or nonstudent samples. As all these aspects are arguably important for the detection of effects, we ask,
Method
Study Retrieval
Our systematic literature search strategy is visualized in Figure 1. Studies were collected from two major databases: Communication and Mass Media Complete, and PsychINFO. The search was limited to journal articles and conference papers written in English. The databases were searched through July 2016. They were examined searching the term “spiral of silence” in titles, abstracts, keywords, and key concepts following Glynn and colleagues (1997). To identify additional literature, we searched Google Scholar checking who cited Noelle-Neumann’s (1974) original article or the meta-analysis by Glynn et al. (1997). This step was taken to ensure that the latest literature, possibly not yet traceable through databases, was included in the analysis, too. Finally, we referred to the literature of the two existing meta-analyses (Glynn et al., 1997; Glynn & Huge, 2014). This led to an initial list of N = 228 identified papers.

Literature search strategy.
Study Selection
We applied three consecutive steps. In the first step, we excluded all research presenting no quantitative data. We also excluded content analyses, methodological research, research unrelated to the goal of the meta-analysis, and research presenting simulated data. Thus, in this first step, we excluded n = 114 papers. In the second step, we applied two inclusion criteria (Glynn et al., 1997; Glynn & Huge, 2014). The first criterion pertained to the operationalization of the independent variable “perception of opinion support.” Participants had to be clearly assigned to a majority or a minority opinion group, either in form of two groups or as a continuum. The allocation had to be based on participants’ subjective perception of the opinion climate in relation to their own opinion. Three kinds of studies were eligible: (1) studies that asked participants about their own opinions and contrasted these with their perception of the majority opinion, merging both into a single “perception of opinion support” variable (e.g., Chia, 2014); (2) studies that directly asked participants whether their opinion was congruent or incongruent to their perception of the majority (e.g., Matthes, 2015); and (3) experiments that allocated participants to groups where their opinion was clearly in the majority or in the minority (e.g., Hayes, 2007). By contrast, studies were excluded if participants were only “objectively” in the majority or minority (e.g., based on a nationwide poll). “People who are, in fact, members of a minority may misperceive they are the majority and vice versa” (Glynn et al., 1997, p. 455). In addition, studies were excluded if they merely took participants’ perception of the majority opinion as the independent variable without relating it to participants’ own opinion. Furthermore, experiments were excluded if they simply assigned participants to a majority or minority opinion condition without ensuring that the participants were actually in the majority or minority according to their personal opinion (e.g., Moreno-Riaño, 2002). The second criterion concerned the dependent variable. Studies were only included if they measured participants’ expression or their willingness to express themselves. This included various strategies to avoid expression (e.g., Hayes, 2007). Yet, studies were excluded if they operationalized expression by taking attitude measures from anonymous surveys (e.g., McDonald, Glynn, Kim, & Ostman, 2001). Reporting one’s attitude in a survey is not the same as expressing oneself by talking to other people or the media. After the second step, we excluded n = 54 papers. In the third step, we excluded Willnat (1995) as it was based on the same sample as Willnat (1996) and reported the identical results. We included Willnat (1996) because it provided more relevant information. We excluded Kim (2012) as it only reported dependent (i.e., moderated) regression coefficients. However, Kim, Kim, and Oh (2014) presented the same coefficients, stemming from the identical sample, as independent effects. Hence, no information was lost. Lacking the appropriate statistical information to calculate effect sizes with the formulas by Lipsey and Wilson (2001), two additional papers had to be excluded (McDevitt, Kiousis, & Wahl-Jorgensen, 2003; Oshagan, 1996). The authors were contacted but it was not possible to obtain the missing information. Thus, after the third step, we excluded n = 4 papers. Our final sample consisted of 56 papers. They yielded 66 independent studies (i.e., independent samples) coming to 27,192 participants. 1 By comparison, Glynn et al. (1997) included 17 studies with 9,500 participants.
Effect Size Calculation and Integration
Pearson’s r was used as the effect size estimate. A positive r indicates that people who perceive greater support for their opinion are more willing to express themselves than those who perceive less support (see also O’Keefe, 2017). In the case of expression avoidance, a positive r indicates that people who perceive less support for their opinion are more willing to avoid expression than those who perceive greater support. In studies reporting correlation coefficients, r was directly taken from the articles. 2 In studies reporting regression results, standardized regression coefficients were transformed to r according to the formula provided by Peterson and Brown (2005). Their research has shown that standardized regression coefficients can be transformed to r independent of the number of predictors included in the respective regression model. To further support this approach, we conducted a sensitivity analysis. Effect sizes did not differ depending on whether they were originally based on correlation coefficients or standardized regression coefficients (χ2(1) = .73, p = .39). Furthermore, the number of predictors included in the respective regression analysis did not influence effect size (χ2(1) = .87, p = .35). In studies reporting means and standard deviations or frequencies, r was calculated according to the formulas provided by Lipsey and Wilson (2001). Before performing the syntheses, correlation coefficients (r) were converted to the Fisher’s z scale (Zr; Borenstein et al., 2009; Lipsey & Wilson, 2001). In total, 324 effect sizes were obtained.
Meta-analysis was carried out using the R metafor package (Viechtbauer, 2010). Estimates were based on random-effects models assuming differing true effect sizes varying. In addition, random-effects results may be generalized beyond the studies included in the analysis because the investigated studies are treated as a random subset of a larger study population (Hedges & Vevea, 1998). Several studies reported results that enabled obtaining more than one effect size per study. Performing a meta-analysis on these studies would violate the assumption of independence of effect sizes and would assign more weight to the studies producing more than one effect size. Researchers recently suggested treating meta-analysis as a multilevel model to address these issues (e.g., Cheung, 2014; Field, 2015; Konstantopoulos, 2011). The basic idea nests the effect sizes (first level) within the studies (second level; Konstantopoulos, 2011; for more detailed information, see Field, 2015). Effect sizes stemming from the same study receive the same random effect while effect sizes stemming from different studies receive different random effects. Hence, the dependence or independence of effect sizes is explicitly modeled by assigning the correct random effect (Konstantopoulos, 2011; Viechtbauer, 2015). Consequently, all effect sizes can be taken into account without aggregation and loss of information. This is especially valuable when it comes to moderator analysis as multiple effect sizes within studies are usually connected to different levels of a moderator variable. Following this reasoning, the moderator analyses were carried out using the rma.mv() function of the R metafor package that enables the estimation of multilevel mixed-effects models (Viechtbauer, 2010). A maximum likelihood estimator was applied. 3
Moderators
Controversiality was assessed by distinguishing controversial and emotionally laden issues (coded 1) from uncontroversial and neutral issues (coded 0) following Noelle-Neumann (1993). To most people, controversial issues are, for instance, abortion, environment, biotechnology, women’s rights, or affirmative action. In addition, we also took the issue description of the authors (e.g., “morally controversial issue,” “controversial social issue,” “controversial topic”) into account. We were not able to assess the subjective perceptions of controversiality by the studies’ respondents. Issue obtrusiveness was coded as unobtrusive (0) or obtrusive (1). Following Winter (1981), obtrusive issues were defined as issues most people have or had personal experience with and/or issues that have significant consequences for most people’s personal lives. Obtrusive issues were, for instance, unemployment, various local issues, issues related to ongoing elections, and issues related to one’s health or safety. Unobtrusive issues were, for instance, foreign politics, gay rights, or prostitution. It is important to note that we had to employ a macro-categorization of obtrusiveness, neglecting subjective perceptions of obtrusiveness. Subjective assessments cannot be taken into account in a meta-analysis. Reference group related to the opinion climate was coded as population in general (0); as government (1); as media (2); as family, friends, and neighbors (3); as experimental manipulation (i.e., other participants or fictitious stimuli (4); or as colleagues (5). The time frame related to the opinion climate was current (0) or future (1).
When it comes to the target of opinion expression, the number of targets that the participants were asked to express themselves to was coded as single (0) or as multiple (1). The types of target were strangers (0), politicians (1), media (2), or family, friends, and neighbors (3). The opinion of the targets was coded as unknown (0), disagreeing (1), agreeing (2), balanced (3), or disagreeing and agreeing (4). The latter pertained to studies where participants were once exposed to disagreeing opponents and once exposed to agreeing opponents. Finally, it was assessed whether participants expressed their opinion offline (0), non-anonymously online (1), or anonymously online (2). Study design was coded as experimental (0) or survey (1). The type of measurement was coded as opinion expression (0), as conversation (i.e., willingness to talk about or discuss an issue (1), as political participation (i.e., willingness to demonstrate, to campaign, to sign petitions, to donate money to political groups, etc.) (2), as response latency (i.e., time taken to report one’s opinion (3), or as avoidance of opinion expression (i.e., willingness to self-censor, to ignore discussions, to express indifference, to say nothing, etc.) (4). Effect sizes obtained for response latency and avoidance of opinion expression were both reverse coded. Positive effect sizes indicate slower responses or greater avoidance of opinion expression when people perceive less support for their opinion. Sample characteristics were mean age coded in years, the percentage of women, and participants’ origin coded as North America (0), as Europe (1), as Asia (2), or as Oceania (3). In addition, it was coded whether the sample was a student (1) or nonstudent sample (0). All variables were coded by two independent coders (second and third author). Reliability as measured by Krippendorff’s alpha was perfect for all variables. The full data set is available from the authors upon request.
Results
Overall Effect Analysis
In line with Hypothesis 1, the overall effect analysis revealed a small (Cohen, 1988), positive effect of perception of opinion support on expression (r = .10; Zr = .10). The effect was highly significant, 95% confidence interval (CI) = [.06, .13], p < .0001. Individual study results can be seen in the forest plot in Figure 2 (Lewis & Clarke, 2001). It displays aggregated effect sizes for each study as well as corresponding 95% CIs. Larger squares indicate larger samples sizes. Following Rosenthal (1979), a so-called file drawer analysis was calculated. It investigates the number of zero-effect studies needed to nullify the found result (Borenstein et al., 2009). The analysis revealed a fail-safe N of 5,686. That means, there has to exist an additional number of 5,686 zero-effect studies to reduce the overall effect’s significance to just nonsignificant. In other words, the found effect revealed itself as highly robust. Looking at heterogeneity, highly significant variability was found among effect sizes, Q(65) = 487.52, p < .0001. This suggests that effect sizes vary considerably due to between-study differences. The I2 statistic—the amount of total variability (sampling variance + heterogeneity) that can be attributed to the heterogeneity among the true effects (Higgins & Thompson, 2002)—gives further insights. About 83% of the total variability could be attributed to between-study differences (I2 = 83.39). It is likely that some of these differences might be explained by our moderators (Huedo-Medina, Sánchez-Meca, Marín-Martínez, & Botella, 2006).

Forest plot of the studies in the meta-analysis according to authors.
Moderator Analysis
Moderated effects were tested by calculating the aforementioned multilevel mixed-effects models (i.e., multilevel meta-regressions). For each moderator, a separate meta-regression was calculated. Categorical moderators (i.e., study design, type of measurement, origin, sample, issue obtrusiveness, kind of targets, opinion of targets, online vs. offline, opinion climate reference group, and opinion climate time frame) were dummy coded. Results are displayed in Tables 1 and 2. Estimates represent changes in effect size according to the changes in moderator levels. The chi-square test statistic indicates whether a moderator, taken as a whole, significantly affects effect size (Q test; Borenstein et al., 2009). By contrast, the z-test statistic indicates whether or not a certain level of categorical moderator was significantly different from the reference category of this moderator (Z test; Borenstein et al., 2009). The reference categories equal the moderator levels signified as zero in the methods section. Answering RQ1, effect sizes did not differ between offline or online expression.
Meta-Regression Results for Testing the Influence of Study and Sample Characteristics on Effect Size.
Note. k: number of effect sizes; Estimate: meta-regression coefficients for Zr; CI = confidence interval with lower (LL) and upper limit (UL); χ2: test statistic of Q test; z = test statistic of z test.
Meta-Regression Results for Testing the Influence of Characteristics of the Expression Situation and the Opinion Climate on Effect Size.
Note. k: number of effect sizes; Estimate: meta-regression coefficients for Zr; CI = confidence interval with lower (LL) and upper limit (UL); χ2: test statistic of Q test; z = test statistic of z test.
p < .05. **p < .01.
When it comes to Hypothesis 2, we were unable to compute effects of controversiality because we could not code any uncontroversial issues. That is, all examined issues were coded as controversial and were emotionally laden. With respect to RQ2, however, the silencing effect was significantly stronger for obtrusive issues when compared with unobtrusive ones (χ2(1) = 4.64, p < .05). Specifically, obtrusive issues evoked an effect size of r = .15 (95% CI = [.09, .20], p < .0001, k = 168) whereas unobtrusive issues only came to an effect size of r = .06 (95% CI = [−.001, .12], p = .05, k = 136). There were no effects regarding Hypothesis 3 and RQ3. Effect size appeared to be independent of the opinion climate’s reference group and time frame. When it comes to the target of opinion expression (RQ4), the number and the opinion of targets did not significantly matter. However, the kind of targets the participants were asked to express themselves to had a significant moderating effect (χ2(3) = 11.62, p < .01). The silencing effect was significantly stronger and came to an r of .29 (95% CI = [.16, .42], p < .0001, k = 32) when participants expressed themselves to their family, friends, or neighbors as compared with strangers (r = .08, 95% CI = [.04, .12], p < .0001, k = 257). Both effect sizes were significant. There were no significant differences when strangers, the reference category, were compared with politicians or the media. Yet, family, friends, or neighbors evoked significantly greater effect sizes compared with both targets (politicians: z = −2.34, p < .05; media: z = −3.36, p < .001). Effect sizes for politicians and the media were both nonsignificant (politicians: r = .07, 95% CI = [−.06, .20], p = .27, k = 11; media: r = .02, 95% CI = [−.07, .10], p = .68, k = 17). However, the latter results have to be interpreted carefully as both estimations were based on a small number of effect sizes. In an additional analysis, we tested the interaction between opinion climate reference group and target of opinion expression. One could expect that both interact, that is, the silencing effect is strongest when the reference group of the hostile opinion climate is family, friends, or neighbors while the target of the opinion climate is also family, friends, or neighbors. However, we found no such interaction χ2(5) = 7.47, p = .19. That means, when the target group is family and friends, the silencing effect is strongest, no matter what the opinion climate refers to.
Looking at Table 1, there was neither an impact of study nor of sample characteristics. Specifically, effect size showed itself unaffected by the specific study design, the type of measurement, participants’ mean age and gender, and participants’ origin. Effect size was independent of whether participants were students or not (RQ5).
To further illustrate the impact of the two significant moderators, we calculated their combined influence with the predict() function of the R metafor package (Viechtbauer, 2010). Table 3 presents predicted r values and corresponding confidence intervals for various levels of the two moderators. The largest effect size, coming to r = .34 (95% CI = [.19, .50]), can be expected when participants are asked to talk to their family, friends, or neighbors about an obtrusive issue. The size can be considered medium to large (Lipsey & Wilson, 2001). By contrast, the lowest effect sizes can be expected for unobtrusive issues discussed with the media.
Predicted Effect Sizes at Levels of the Moderators Opinion Expression Target and Issue Obtrusiveness (k = 297).
Note. Predicted r: predicted Pearson correlation coefficient; CI = confidence interval with lower (LL) and upper limit (UL).
Publication Bias Analysis
We tested whether studies with small samples and minor effect sizes failed to be published. A funnel plot and Egger’s regression test for funnel plot asymmetry were applied (Egger, Smith, Schneider, & Minder, 1997). Looking at the funnel plot (Figure 3), there was no evidence of publication bias in terms of smaller studies with minor effect sizes missing at the bottom left corner. This was further confirmed by a nonsignificant Egger’s regression test (t(64) = .34, p = .73). There were no indications of publication bias.

Funnel plot of the studies in the meta-analysis.
Additional Analyses
We additionally checked whether the differential effect of offline versus anonymous online or non-anonymous online expression settings might be dependent on issue obtrusiveness, number of targets, type of targets, opinion of targets, the opinion climate reference group, or the opinion climate time frame. However, there was no significant interaction between offline versus online expression settings and one of the other six moderators (Issue Obtrusiveness × Online vs. Offline: χ2(2) = 1.13, p = .57; Number of Targets × Online vs. Offline: χ2(1) = .15, p = .70; Opinion of Targets × Online vs. Offline: χ2(6) = 6.99, p = .32; Opinion Climate Reference Group × Online vs. Offline: χ2(4) = 7.79, p = .10; Opinion Climate Time Frame × Online vs. Offline: χ2(2) = .65, p = .72; we were unable to test the interaction between offline vs. online expression settings and type of targets due to lack of effect sizes).
Discussion
The aim of this study was to revisit the core assumption of spiral of silence theory that opinion climate perceptions are related to the willingness to speak out. The overall size of the effect we observed suggests that the spiral of silence is by no means a “concept about to retire” (E. Katz & Fialkoff, 2017). Furthermore, as one of the most pressing questions in this line of research, our results demonstrate that the silencing effect did not disappear in online environments, nor was it weakened. This suggests that online opinion expression environments may very well trigger fears to become socially isolated when voicing dissenting views.
The relevance of this finding is threefold: First, a key assumption of the spiral of silence in the offline world is the existence of consonance of majority views in the mass media (Noelle-Neumann, 1974). On the one hand, one may argue that online environments do not offer one consonant majority opinion, but many subjective majority climates depending on selective exposure (Schulz & Roessler, 2012). This would make a spiral of silence very unlikely. On the other hand, the most frequently visited online media are often the online versions of offline mainstream mass media (see Van Aelst et al., 2017). As Van Aelst et al. (2017) put it, News media with an ambition to cover politics in a balanced and neutral way still constitute the main source of political information for most people, and selectivity based on political interest seems to be more widespread than selectivity based on political beliefs. (pp. 13-14)
Although more research is needed, the findings of our meta-analysis are in line with this view. Second, compared with classical mass media, the online world offers new and additional forms of monitoring the opinion environment. Click speech such as likes and shares as well as user comments can serve as cues about the majority views (von Sikorski & Hänelt, 2016). People high in fear of isolation may especially monitor such cues, setting the spiral into motion. This also supports the view that the spiral of silence is far from being obsolete in the new media environment. Third, when it comes to opinion expression, of course, the online world does not offer the same kind of possible sanctions when dissent is voiced as do face-to-face situations. Nevertheless, there are several scenarios for opinion suppression when having minority views online. Online contacts may overlap with offline contacts, so even if there is no direct threat of social isolation online, it may exist offline. Also, even in anonymous opinion expression settings, online media allow direct interaction, that is, reactions by others. If, for instance, a user posts a dissenting view in a political forum, she or he may have to fear a so-called “Shitstorm,” rude comments, and personal attacks, which are perceived as highly unpleasant to many users, despite anonymity.
Another key finding is that obtrusiveness moderates the silencing effect. We observed that the effect is stronger for obtrusive rather than unobtrusive issues. This suggests that the fear of being socially isolated due to minority views is stronger for issues that are relevant to the life of most people (Willnat et al., 2002). When an issue matters for most people’ lives, expressing minority views may be more intimidating. That is, if someone holds a political minority view, she or he can expect that others can present arguments based on personal experience to question that view. Unfortunately, we were not able to test the role of controversiality. In fact, most issues in the news are controversial to some extent and have a morally laden component. In her seminal article, Noelle-Neumann (1974) has analyzed the following topics, all of them characterized as controversial: abortion, alcohol threshold for car driving, capital punishment, unmarried couples living together, corporal punishment for children, foreign workers, the “achievement-oriented society,” the treaties of Moscow and Warsaw, the recognition of the German Democratic Republic (GDR), the ban of the Communist Party, or the evaluation of political candidates. We believe that Noelle-Neumann’s list illustrates that most topics are controversial to some extent, and that is exactly why they are debated in the news. Apparently, spiral of silence researchers have followed the advice of Noelle-Neumann (1974) to select issues where some moral dimension is salient. As the degree of controversiality cannot be assessed in a meta-analysis due to the lack of available details about the respective issues, this remains a question for future research.
We also found that the type of the target of opinion expression matters for the silencing effect. Interestingly, expressing dissenting views to family, friends, or neighbors has a stronger effect as compared with strangers, politicians, or the media. This contradicts the original ideas of Noelle-Neumann (1974, 1993) who preferred measuring opinion expression toward strangers. The explanation for this novel finding can be found in the social accountability argument put forth by Mutz (2002). The “need for social accountability creates anxiety because interpersonal disagreement threatens social relationships, and there is no way to please all members of one’s network and thus assure social harmony” (p. 840). When voicing dissent to close others, social accountability is arguably higher compared with voicing disagreement to journalists, politicians, or strangers. Thus, citizens suppress their views to maintain social harmony (Mutz, 2002). In terms of the spiral of silence, this means that disagreement with friends and family may lead to a higher state of fear of social isolation as compared with strangers. Friends and family represent close ties.
The silencing effect did not vary by the number of targets, the opinion of the targets, the opinion climate characteristics, and the study, design, measurement, and sample characteristics, despite the sufficient statistical power. This has a number of key theoretical implications: First, when it comes to the number of targets that the participants were asked to express themselves to, the null-finding suggests that one single opinion target is fully sufficient to trigger the silencing effect. That is, more targets do not increase or decrease the effect. Rephrased, the fear of social isolation becomes salient even when there is only one single person that could isolate someone holding minority views. This seems to contradict Noelle-Neumann’s (1974) original views according to which fear of isolation refers to a community, a group, or one’s entire social environment. Thus, the scope of the silencing effect may be stronger than originally theorized by Noelle-Neumann. This is, in fact, a novel finding deserving more comprehensive research.
Second, surprisingly, the opinion (i.e., unknown, agreeing, or disagreeing) of the target of opinion expression did not matter for the silencing effect. That is, when people feel they are in the minority, they choose to remain silent, independent of the question whether or not their target agrees or disagrees with the minority view. One reason for this may be that it is hard for people to estimate or predict what their opinion targets truly think and how they may act after a minority view is disclosed. The safest option seems to be remaining silent when holding minority opinions. Again, this suggests that the silencing effect is robust; it is even triggered when the opinion target also holds a minority view.
Third, the opinion climate characteristics did not matter either. This means, as long as there are cues about the minority/majority opinion climate—no matter if with respect to the whole population or strong-tie others—the silencing effect occurs. The effect does therefore occur in all cases in which one dissents from a reference group, no matter how big or small this reference group is. Fourth, when it comes to the study characteristics, the origin of the study—pointing to the universality of the theory (Scheufele & Moy, 2000)—also did not matter. That is, effects did not differ between Europe, the United States, and Asia. Based on the assumption that collectivist cultures like those in Asia value the collective work of groups while Western cultures tend to emphasize the individual, it has been assumed that the spiral of silence should be more pronounced in collectivist rather than individualist cultures (Scheufele & Moy, 2002). As a truly novel finding, we can demonstrate that this is not the case. The silencing effect is also robust no matter the design or sample of a study. All of this suggests that the relationship between opinion climate and opinion expression is highly significant and largely unconditional. All in all, the strength of the silencing effect is maximized when combining the effects of the two significant moderators: When dissenting views about obtrusive issues are presented to family, friends, or neighbors, the effect is impressively strong (r = .34).
Limitations and an Agenda for Future Research
As is common in meta-analyses, we were only able to include papers available in English. Also, coding the participant’s origin with Europe, the United States, and Asia ignores cultural differences between single countries. Another limitation refers to the fact that we only examined one aspect of the theory. In fact, most of the analyzed studies did not empirically integrate the silencing hypothesis into the overall context of the theory. Future research should strive to examine the theory more comprehensively, rather than focusing on one single assumption. We want to stress that we did not assess subjective perceptions of obtrusiveness, but categorized issues as obtrusive or unobtrusive at a macro level. Thus, in order to fully understand the role of issue obtrusiveness, we need studies that systematically manipulate and model obtrusiveness.
As in any meta-analysis, we cannot rule out that single studies were missed, especially unpublished ones. Nevertheless, we believe that this does not diminish our findings as we applied a random-effects model. Thus, in our analysis, the investigated studies were treated as a random subset of a larger study population (Hedges & Vevea, 1998). We also found no evidence for a publication bias. Some of the theoretically and methodologically important questions could not be addressed with our meta-analysis and thus can be considered as limitations of spiral of silence research as a whole. That is, as with any meta-analysis, ours may leave some blind spots because relevant moderators cannot be analyzed due to a lack of studies investigating those moderators. First, the role of issue controversiality needs more reflection. It is one of the core assumptions of the theory that has never been systematically examined. Second, when it comes to research on online opinion environments, we need more insights about different types of environments. Voicing dissenting views on Facebook among non-anonymous people may evoke different psychological mechanisms than speaking one’s mind in an anonymous forum (Neubaum & Krämer, 2016; see also Neubaum & Krämer, 2017a, 2017b). In this meta-analysis, we compared offline and online contexts based on the available studies. We need more studies on online opinion expression in order to further differentiate the factors that foster or dampen a spiral of silence (see Neubaum & Krämer, 2016, 2017a, 2017b). Third, our meta-analysis focused on the relationship between opinion climate and opinion expression. This, however, is a simplification of the original theory. Fourth, we suggest to further distinguish family, friends, and neighbors as targets of opinion expression. The effect of dissent on social harmony may be different when looking at each group individually. Due to the lack of relevant studies, this could not be tested here. The same is true for the opinion of the target, especially when the target agrees. This should dampen the silencing effect. We could not observe such a dampening effect, but additional studies are needed to corroborate this claim. Fifth, we lack studies from Latin America, Africa, and other parts of the world. The universal nature of the theory thus remains a pressing question for future research. Sixth, we could also not examine the role of personality traits such as willingness to self-censor and fear of social isolation (see Hayes & Matthes, 2017; Hayes et al., 2013; Matthes et al., 2012) which may serve as additional moderators. Finally, looking at research methods, longitudinal survey designs are clearly missing.
Conclusion
This study aimed at quantifying the effect of opinion climate perception on political opinion expression by using a meta-analytical approach. The results showed a small (Cohen, 1988) but highly robust and highly significant overall effect. We found that the silencing effect is strongest (i.e., medium size; see Cohen, 1988) when dissenting views about obtrusive issues are voiced to family, friends, and neighbors. However, several other theoretically and methodologically relevant moderators did not exert any effect. We can thus conclude that the relationship between opinion climate perception and political opinion expression is stronger and more robust than previously thought.
Footnotes
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
