«NETWORKING IN EVERYDAY LIFE by Bernard J. Hogan A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy ...»
7.3 Which alters are accessed (or accessible) by media Not all media are available to all individuals. Haythornthwaite’s original work was done in a context where virtually all students or collaborators had access to similar media. Yet, individuals who have a cell phone do not necessarily give their number to everyone. Similarly, individuals who email their networks do not email everyone who has an address. These issues tell us a lot about the relative role of media in sustaining ties to alters. This section examines the relationship between those alters that are accessible by media and in person versus those who are regularly contacted by these media. The purpose of this analysis is twofold. First, this will help familiarize the reader with the measures and distributions which are later used in the multilevel models of media multiplexity. The second is to promote the idea that new media are primarily used to maintain contact with very close network members rather than one’s weaker ties. Or at least, they are used with a small slice of the network rather than most of it. This latter claim is somewhat controversial as several scholars have related media use with network size and the ability to harness weak ties (Boase et al., 2006; Zhao, 2006). While I accept that there are times when many individuals are emailed, this is not the same as suggesting that email is one of the key ways in which full personal networks are sustained.
Figure 7.1 shows the distribution of media use with ego’s personal network.
ters contacted monthly for 84 personal networks.3 Each chart has a discernible shape.
The charts for in-person, socializing, and telephone contact display a logarithmic distribution while the charts for mobile phones, email, and instant messaging show a linear distribution among the network members who used these media. The trend lines in each chart indicate how faithfully the sample matches the proposed distribution. In all cases, it can be considered a very good ﬁt, with the R2 values for every distribution being above 0.95.
The biggest differences in this chart are clearly between new media and old media.
These differences accentuate the problems inherent in a tidy model of media multiplexity in everyday life. Firstly, the new media are not uniformly adopted by the population. All respondents telephoned, socialized or saw in person at least one other network member in the last month. Yet, many of the respondents did not email, call by mobile phone or instant message any of their network members. Many of these zeroes are obviously due to those who do not have access to the media, but that should not be taken as code for the respondents not being able to gain access to these new media. Virtually everyone in the study had the economic means to afford a computer.
Many respondents were simply disinterested in adopting this form of connectivity.
This especially marks the difference between instant message use and email. If they are emailing these people, they almost certainly also have the equipment ready for instant messaging, even if they do not use it.
The difference between new and old media also show in the distributions of proportion contacted monthly. In-person, socializing, and telephone proportions show a logarithmic distribution, whereas email, mobile, and instant messaging show a linear distribution. These distributions are non-trivial. A log distribution means that start
0.6 0.6 0.4 0.4
0.6 0.6 0.4 0.4
0.6 0.6 0.4 0.4
off. A linear distribution means that starting from zero, there is a steady increasing distribution of the percentage of ties contacted, with no particularly steep climb nor a particularly level area at the top. These two different distributions highlight two different logics at play. For in-person, socializing, and telephone, most people contact most of their alters at least monthly. One can expect to see alters monthly (or semimonthly) as well as expect to be seen by them. Also, most people expect to call or be called by most of their alters at least monthly. As such, it makes it to see this as a habitual part of networking in everyday life. There is less consensus, however, about how much of the network is maintained by ICTs. A linear distribution means that the probability of ego emailing an alter is quite ambiguous. People who email their alters are just as likely to email a small share of them as they are to email most of them. Same for instant messaging and cell phones. Differences in network size notwithstanding, people are just as likely to call a few of their alters via cell phone as they are to call most of their alters.4 Thus, the use of ICTs appears to be at least partially a personal decision, based on one’s media tastes and ability to access network members rather than a cultural convention. This harkens back to Swidler’s discussion of the cultural toolbox during settled and unsettled times (1986). At least in everyday life in 2005, telephone and in-person contact appears to be a settled part of the conventions for networking; email, mobile phones and instant messenger do not.
This is important for an analysis of media multiplexity in personal networks—it means that one cannot rightly perform an analysis of media multiplexity via dyads alone, as individuals operate on different logics in their network about how many individuals they are going to contact via any medium.
It is important to note that this claim does not take into account differences in the size and structure of the network. If this was done, one can see that those with larger networks email a disproportionate number of these individuals. However, I believe this says more about the speciﬁc individuals than the media, especially considering the ﬁndings in Chapter 5 about how heavy communicators use all media at their disposal.
CHAPTER 7. MEDIA USE WITH SPECIFIC NETWORK MEMBERS 189
7.4 Dyadic reports of media use Two dependent variables were posited for this analysis. The ﬁrst includes seven possible “points-of-access”: face-to-face, socializing (which is considered a subset of faceto-face interaction, where the former can include chance meetings and propinquity), landline telephone contact, calling by cell phone, calling to a cellphone, email, and instant messaging. Admittedly, the ﬁrst two are not media. However, they have been included in analyses up to this point since they are of making contact with other individuals. Moreover, as mentioned in Chapter 2, in-person interaction has certain affordances for social activity, just like media. By taking an affordances-oriented view, there are plausible reasons for considering in person and social activity, and including them as separate variables. Some individuals have network members who they never see anymore, but still keep in contact occasionally via telephone and via email. Also, many respondents would know people they consider at least somewhat close alters but do not socialize with them. This may lead to different ordering principles in the network—if you see your neighbour frequently but do not socialize with him, how would that play into your willingness to be accessible to him via many media?
Fortunately, the parameter estimates were relatively close whether in-person and socializing were included, or whether they were excluded. Given that these terms give the variables a larger range and a more normal distribution, thereby increasing the potential for stronger estimates, this expanded variable is preferred.
Figure 7.2 shows the distribution of the number of media-contexts used, characterized in two ways.
The lighter bars in back show the percent of dyads that use the speciﬁed number of media-contexts. The darker bars in front show the distribution of the mean values per network. As one can see, unsurprisingly, the mean values per network cluster together far more tightly. Individuals, in general, use three-to-four
Figure 7.2: Distribution of number of media used by alter and network (average) even the most ardent networker uses more than ﬁve media on average with all of their alters.
This ‘regression to the mean’ helps to illustrate that even though individuals may reciprocally reinforce or condition each other in terms of the number or intensity of media use, this is not the case for all network members. Ultimately, ego will want to include some people that he neither wishes to access via every medium, nor is it even possible as not everyone included uses every medium.
7.5 A multilevel model of media multiplexity This model seeks to test the theory that tie strength leads to increased numbers of media used with alters. In theoretical terms, this suggests that individuals are actively regulating their social accessibility by maintaining higher levels of access to those with whom they feel there is a stronger relationship. Yet, there are practical reasons why people use certain media with their alters. They may be proximate, well-embedded, and frequently in contact. Yet, if after controlling for these factors, tie strength is still
theory, and a relevant one for understanding how people decide which media to use with which alters. By implication, this will also point to media multiplexity as a useful theory for understanding everyday networking.
This analysis proceeds as a series of nested models demonstrating the overall story that the relationship between tie strength and the number of media used is underspeciﬁed. That is not to say that the correlation is spurious. As I am dealing with cross-sectional data, it is difﬁcult to say which covariates cause others. For example, one of the highly signiﬁcant variables in the model is degree (that is the number of links that alter has to others in the network). A theory of triadic closure suggests that individuals link two alters together, because both alters are strongly tied to ego (Granovetter, 1973; Simmel, 1950). So, in some sense I can suggest that degree is an intermediary between tie strength and number of media-contexts used. Thus, my assertion is not that tie strength is a spurious causal factor, but that the processes leading to media multiplexity are under-speciﬁed.
7.5.1 Within network variance (the variance components model) I begin with the simple model that assesses whether there is sufﬁcient variance between networks in order to consider using a multilevel model. If the error terms are correlated within networks (meaning that there are network-speciﬁc biases), then standard regression tests will give insufﬁcient and biased estimates. Most personal network studies assume that a multilevel model is warranted since alters nested within network behave very differently across networks (or reportedly behave very differently according to the respondent, van Duijin, van Busschbach, and Snijders, 1999;
Wellman and Frank, 2001). This is especially the case when considering media use since some egos do not have either a mobile phone or a computer. Media do not cleanly substitute for one another; people who buy a computer do not give up their
to add on to other forms of communication rather than substitute for them (Carrasco and Miller, 2006; Madell and Muncer, 2005; DiMaggio, Hargittai, Neuman, and Robinson, 2001; Quan-Hasse and Wellman, 2006). As such, it is important to take this variation among networks into account.
The variance components model was performed using xtmixed in Stata with Maximum Likelihood Estimation.5 The results of this model are shown in table 7.1.
variance between networks and within networks, respectively. The likelihood ratio ˆ χ2 test of ψ shows that it signiﬁcantly varies from zero, and thus I can reject the null
these two scores. Lower values in subsequent models will indicate the percentage of unexplained variance reduced. The second fact is that I can calculate the percentage of unexplained variance in this model that is due to level-1 factors (differences in dyads)
To assess the proportional reduction in variance at level-1 (the alter level), the equation Stata offers several packages for the estimation of multilevel models. xtmixed is a fast native package, although it has limitations for more complex models with cross-level effects and random coefﬁcient models. In such a case one can use the GLLAMM package. Since GLLAMM models did not signiﬁcantly alter the results but did make the presentation signiﬁcantly more complex, I will only be presenting the random-intercept models via xtmixed herein. Moreover, the authors of GLLAMM encourage individuals to use the xtmixed mixed package unless necessary (Rabe-Hesketh and Skrondal, 2008).
CHAPTER 7. MEDIA USE WITH SPECIFIC NETWORK MEMBERS 193
as given by Raudenbush and Bryk (2002) in this notation is:
level-2 (the network level). There is more variation within networks than between networks. This makes sense from a media multiplexity perspective, as individuals within all networks are afforded fewer media the weaker the tie. However, it also points to the need for a multilevel analysis of this question. The subsequent random-intercept ˆ ˆ models will use the ψ and θ values from Table 7.1 when calculating the proportional reduction in error (i.e., the R2 ).
Table 7.1: Variance components model of the number of media used with alter by network 7.
5.2 Random-intercept model with covariates Multilevel models are expensive in terms of both the degrees of freedom as well as computational power. This means that standard errors are going to be much larger in a two-level model than in an ordinary regression. For this reason, only a small purposefully selected set of variables are included, rather than all possible covariates. This forward-selection style of model-building based on the theoretical reasons
because spurious variables can mask more signiﬁcant ones (van Duijin et al., 1999;