The mobilization effect of social media use: an instrumental variable approach

There is ongoing disagreement in the literature on OPP, with some researchers arguing that OPP reinforces (i.e., normalizes) existing patterns of offline political participation, while others emphasize that social media can facilitate (i.e., mobilize) political engagement. Theories put forward by the former group are called normalization theories, while those proposed by the latter are referred to as mobilization theories. Normalization theories focus on the existing network structure, which reflects the status quo. According to normalization theorists, while it may appear that SMU brings about social change, these changes are largely driven by offline political dynamics in traditional political processes, and the impact of SMU is limited (Gerl et al., 2018; Mejias, 2010). On the other hand, mobilization theories focus on the horizontal and decentralized networks that are created and reinforced through the use of social media. Mobilization theorists generally posit that as the communication network becomes more connected and participatory through SMU, the population using the platform gains increased access to information, more opportunities to participate in public discussion, and greater capacity for collective action (Lilleker & Koc-Michalska, 2018; Shirky, 2011). They also believe that widespread SMU will promote the sharing of opinions in the public sphere, gradually building a framework for citizen participation over time (Boulianne, 2018; Shirky, 2013). Consequently, one’s perceived impact of SMU on OPP may depend on his/her acceptance of certain assumptions regarding network properties.

In addition to the divide between mobilization and normalization theories previously mentioned, there are two other notable strands of research within the field of OPP. One strand focuses on identifying the characteristics of individuals who participate online and primarily employs the resources approach (Best & Krueger, 2005; Boulianne, 2018; Ruess et al., 2023; Smith, 2013; Vasilescu et al., 2014; Wang et al., 2018). This strand, which generally assumes that SMU leads to OPP and that the network structure is decentralized and distributed, examines differences between individuals (i.e., between variations). The other strand of research on OPP focuses on examining the ways in which individuals engage in political activity. In doing so, it refers to social learning mechanisms and information exchange theories to understand the impact of SMU on OPP (Velasquez, 2012; Yamamoto et al., 2020). This area of research examines changes within an individual over time (i.e., within variations), taking into consideration that traditional resources related to connection intensity within a network are less likely to change over time. The latter strand of research on OPP implicitly assumes the status quo and may be more suitable for measuring the impact of an individual’s SMU on OPP, regardless of income, gender, age, and educational attainment. However, there is a significant risk of endogeneity in studies that adopt this assumption.

Normalization theories posit that the political potential of social media is primarily utilized by individuals who are already politically active and engaged. In other words, these theories propose that social media usage for political purposes does not significantly enhance overall participation (Chadwick, 2006; Gerl et al., 2018). This is attributed to preexisting inequalities in material, social, and political resources that also lead to disparities in online participation. This phenomenon, referred to as the ‘digital divide,’ closely mirrors disparities in offline participation among socio-demographic groups (Castells, 2008; Hirzalla et al., 2011; O’Neil, 2017; Pariser, 2011; Strauß et al., 2020; van Dijk, 2020).

Research on the digital divide, sometimes used interchangeably with normalization, has broadly examined two aspects of SMU: intensity and forms (Hirzalla et al., 2011). Studies have shown that differences in motivational, material, and skills access contribute to three socio-demographic gaps in most developed and democratic countries regarding the intensity of SMU: a ‘socioeconomic gap,’ a ‘gender gap,’ and an ‘age gap’ (di Gennaro & Dutton, 2006; Quintelier & Vissers, 2008). In terms of forms of SMU, research has focused on how access differences create or maintain a ‘democratic divide’ between those who can and cannot use the Internet and social media for political purposes (Matthews, 2021; Mossberger et al., 2003; Norris, 2001). These trends have led normalization theorists to view social media as reinforcing the status quo (e.g., Dalton, 2017), and they generally suggest that OPP is primarily influenced by political and social resources or a reflection of offline political participation (Bonfadelli, 2002; Matthews, 2021; Oates et al., 2006; van Dijk, 2005, 2006). However, it is important to note that digital divides are caused by differences in resources between individuals or groups, and socio-demographics alone do not fully explain differences in SMU (Hirzalla et al., 2011).¹

Mobilization theories emphasize how social media can facilitate politically-motivated activities, serving as a political forum where individuals can develop civic skills and acquire the knowledge necessary for political engagement (e.g., Kann et al., 2007; Theocharis & van Deth, 2017; Valenzuela, 2013). According to Hirzalla et al. (2011), mobilization claims involve four types of OPP. First, there are studies on ‘digital activism,’ which refers to confrontational, sometimes highly radical, forms of interaction between citizens and political or economic elites (e.g., Dahlberg & Siapera, 2007; George & Leidner, 2019). Second, studies examine democratic conduct in discussions on web forums and social networking sites, with some scholars considering whether social media is realizing a new iteration of the public sphere concept (e.g., Bennett & Pfetsch, 2018; Dahlgren, 2005). Third, a body of research examines how individuals utilize the informative or educational potential of websites with politically or civically oriented content. These studies often aim to demonstrate the wealth of information available online to the general public, and how social media grants new freedoms to non-governmental organizations (NGOs), governments, and individual political entrepreneurs to disseminate their messages (Feezell et al., 2016; Montgomery et al., 2004). Fourth, much research oriented towards mobilization focuses on the concept of ‘e-democracy,’ which refers to an interactive online environment where citizens can educate themselves about and communicate with political representatives (Aichholzer & Rose, 2020; Chadwick, 2006).

Informed by the existing literature, we can summarize that normalization theories focus on citizens’ general abilities and goals, while mobilization theories examine the potential and actual instances of political activity on social media (Hirzalla et al., 2011). Normalization theories concentrate on identifying who participates and the socio-demographic differences among participants. According to normalization theory, OPP is seen as one of the socio-demographic characteristics of an individual. Therefore, it does not significantly increase or decrease depending on SMU, especially in estimates using a model with fixed individual characteristics. In contrast, mobilization theories are more interested in how participants engage in the political forum, how their behavior changes over time, and the features and characteristics of networks that facilitate and enable OPP. According to mobilization theory, as the average individual’s use of social media increases, so does the level of OPP. While these two theories present antithetical expectations, which one has more explanatory power in the current situation has not yet been empirically tested.

Considering the debate on whether SMU leads to normalization (where online political engagement mirrors existing offline patterns) or mobilization (acting as a catalyst for increased and diverse political participation), our first hypothesis is:

If the use of social media has a mobilization effect, what becomes important are the consequences of this mobilization. While most studies assume that it leads to an increase in OPP, this outcome can depend on the characteristics of the network, such as homophily or heterophily.

Homophily, the tendency of individuals to form relationships with others similar to themselves, is based on individual traits such as age, gender, race, educational attainment, or personality, or on group membership or shared experiences (Bond & Sweitzer, 2022; Boutyline & Willer, 2016; Granovetter, 1983; McPherson et al., 2001). This tendency can lead to the formation of homophilous clusters within networks, where individuals with similar characteristics or interests are more likely to connect. It is widely accepted that homophilous clusters play a significant role in promoting political participation (Christakis & Fowler, 2009; Kim & Chen, 2016; McPherson et al., 2001).

Meanwhile, heterophily refers to the tendency of individuals to associate with those who are dissimilar, based on individual characteristics or group membership (Boutyline & Willer, 2016; Halberstam & Knight, 2016; McPherson et al., 2001). This concept is often discussed regarding whether interactions with different individuals within the network will reinforce existing consistencies or lead to more accurate perceptions through mutual understanding (e.g., Song et al., 2020). Communication with heterogeneous individuals does not always reduce the consistency of beliefs; at least, such reduction may lead to a decrease in partisan attitudes and, further, a decrease in actual participation in the context of mobilization theory. However, to the best of our knowledge, there is a lack of prior research on the types of networks formed by individuals on social media platforms and their potential impact on facilitating OPP.

In terms of the role of networks, the impact of heterophily, and by extension homophily, on political participation, including OPP, remains inconclusive according to Valenzuela et al. (2012). Some research has found that high levels of heterophily can lead to political ambivalence and fear of isolation, resulting in decreased participation (e.g., Chu & Yeo, 2020; Mutz, 2006), while other studies have indicated that heterophily is positively related to participation (e.g., Scheufele et al., 2006; Strauß et al., 2020). Therefore, it can be hypothesized that the formation of networks characterized by homophily through SMU may enhance OPP, whereas the absence of such networks may decrease OPP. Furthermore, if the remaining variation is insignificant, then this supports the normalization hypothesis.

Regarding the nature of social interactions on social media, particularly among mobilized users, and whether these interactions exhibit homophily (engagement with like-minded groups) or heterophily (engagement with a diverse range of opinions), (Fig. 1) the study proposes the following second hypothesis:

Fig. 1. Summary of hypothesis.

Download Original Figure

Combining the above theories and approaches, factors influencing online participation predominantly focus on the demographic characteristics of individuals (Boulianne, 2018), such as gender, age, education level, income level, regional affiliation, marital status, occupation, and religion (e.g., Anduiza et al., 2010; Best & Krueger, 2005; Dutton & Blank, 2011; Park, 2018; Saglie & Vabo, 2009; Shin et al., 2017; Smith, 2013; Vicente & Novo, 2014). Recent research has also directed attention to situational factors affecting online participation. Among these noteworthy research areas is privacy-related research (Hoffmann & Lutz, 2023; Lutz & Hoffmann, 2019). Privacy concerns have sparked debates on whether they prompt individuals to engage in more active online activities, including online participation, or conversely, deter them, turning them into passive consumers. This phenomenon is sometimes referred to as the ‘privacy paradox’ (Blank, 2013; Hoffmann & Lutz, 2023). In particular, where concerns about privacy invasion are high, there is a possibility of negative effects on online participation (Jang & Sung, 2021). Furthermore, overall privacy concerns associated with the online platform itself may lead individuals not only to refrain from participating but also to avoid using the platform altogether.

Since this study employs a fixed-effects model that controls for individual time-invariant characteristics, our interpretation of the descriptive statistics focuses on the important time-varying variables (Table 2, 3, 4). Both OPP and SMU incrementally increased during the analysis period. However, the average of the OPP measurements is close to 1, and the distribution of OPP shows positive skewness. This suggests that most people do not engage in political activities online, and only a small number are highly active. Similarly, the sub-measurement variable of SMU, akin to that of OPP, indicates that SMU is not significantly high either. On the other hand, privacy concerns, a control variable, did not change significantly until 2017 but then rapidly increased by more than 1 point on average in that year, with an average of 3–4 points, indicating high levels of privacy concerns. Finally, telecommunication costs, the instrumental variable, is measured on a scale of 1 to 11 points. The average telecommunication costs did not change significantly during the analysis period, showing a slight overall decline.

However, when we decompose the variation, it reveals dynamics that were challenging to discern through simple trend analysis. Firstly, the within-variation for OPP is more significant than the between-variation. This suggests that there is more variability in OPP within the same individuals over time than across different individuals. Consequently, our analysis focuses on within-variation in OPP, which is why we test the mobilization hypothesis. Secondly, within-variations are larger than between-variations in SMU. This indicates that the variation is more influenced by changes within the same individuals over time. Thirdly, the between-variation in telecommunication cost is significantly higher than the within-variation. Among the time-varying variables, the telecommunication cost variable is the only one with higher between-variation than within-variation. This can be attributed to the randomness in subsidies. Lastly, within-variation is considerably higher than between-variation in the case of privacy concerns. This suggests that an individual’s level of concern about privacy is subject to more significant changes over time than the differences in privacy concerns among different individuals.

Using a fixed-effect model is preferable when there are not many observable variables available, and the risk of omitted variable bias is high. However, it is insufficient to handle all endogeneity issues, especially the problem of reverse causality (Schaffer, 2020; Wooldridge, 2010). If a reverse causal relationship exists between SMU and OPP, the effect of the independent variable can be underestimated when the instrumental variable is not used (e.g., Cawley & Meyerhoefer, 2012; Kwon, 2016). To address the remaining problem, we conduct FE2SLS (Model 1) alongside FE (Model 3), as well as RE2SLS (Model 2) and RE (Model 4) for efficient estimations additionally.

In the following, we examine whether the telecommunication costs variable is appropriate as the instrumental variable. Instrumental variables must meet the criteria of exogeneity, relevance, exclusivity, and identifiability. However, as the exogeneity of instrumental variables is challenging to identify statistically, the necessity of instrumental variables can only be confirmed by examining the endogeneity of explanatory variables (Schaffer, 2020; Wooldridge, 2010). The problem of endogeneity of explanatory variables relates to the correlation between the variation of the explanatory variables and the residual variation. The endogeneity of explanatory variables can be confirmed by calculating the F-statistic in multiple regression analysis, where the residual of the first-order regression equation is added as a variable to the second-order regression equation. In other words, the null hypothesis $COV\left({\overline{SMU}}_{i,}{\overline{u}}_i\right)\ne 0$ (Wooldridge, 2010). The analysis results reveal that the F-value, with degrees of freedom 1 and 9282, is 32.29, leading to the rejection of the null hypothesis that SMU has no endogeneity (p<0.0001)(Table 5).

Next, we examine the relevance of the instrumental variable. Assessing the relevance of an instrumental variable involves determining whether there is a correlation between the instrumental variable and the explanatory variable that has been identified as endogenous. In other words, the null hypothesis is $COV\left({\overline{SMU}}_{i,}\overline{Phonebil{l}_i}\right)\ne 0$ (Wooldridge, 2010). If the partial F-statistic value of the first-stage regression equation exceeds 10, then the instrumental variable is considered strong (Stock & Yogo, 2005). The analysis results reveal a t-value of 5.68 and a partial F-value of 32.31, indicating that the relevance condition is satisfied. Therefore, the instrumental variable in this study is not weak.⁴

Additionally, we test for underidentification of the simultaneous equations model. The analysis results show that the Kleibergen-Paap rk LM statistic, which follows a chi-square distribution with 1 degree of freedom, is 32.29. Consequently, the null hypothesis is rejected, confirming satisfaction of the ranking condition and indicating that there are no issues with the identification of the model. In summary, the instrumental variable is appropriately identified, and there are no concerns related to a weak instrumental variable. Therefore, we can confidently assert that the specification of the FE2SLS model is free from major concerns.

The analysis results indicate that there is a significant difference between the model using the instrumental variable and the model without the instrumental variable. Furthermore, the effect of SMU on OPP is underestimated in the models that exclude instrumental variables, consistent with previous findings. The Hausman test results show that the random effects model is not suitable because the null hypothesis is rejected in both cases, whether the instrumental variable is used or not. Therefore, we select Model 1 as the final model.

According to Model 1, a higher level of SMU is associated with a higher level of OPP. In other words, if the level of SMU increases by one unit between different time points within an individual, OPP increases by approximately 0.48. However, since this result is based on estimations using within variations, it cannot be interpreted as implying that individuals with high levels of SMU have higher OPP than those with low levels of SMU. This result suggests that the mobilization hypothesis is likely to be more acceptable than the normalization hypothesis and that the network features enough homophily to facilitate user participation (Table 6).

The result indicating that SMU has a significantly positive impact on OPP can be understood as an increase in opportunities for OPP resulting from SMU (Shirky, 2011). This is primarily because participation opportunities, such as online surveys, are often promoted on the same platform or across easily accessible platforms. Furthermore, the frequent and intensive interactions and exchanges of opinions among individuals through SMU may have inspired them to participate (Shirky, 2013).