«NETWORKING IN EVERYDAY LIFE by Bernard J. Hogan A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy ...»
Raudenbush and Bryk, 2002). As this is a relatively simple model with no random slopes, cross-level effects or interaction terms, I will focus on the inclusion of selected variables based on the above hypotheses. I examine three random intercept models to assess the proportional reduction in variation, and to understand the behaviour of tie strength across models. The ﬁrst model merely includes the variable ‘ring’. If one recalls from Chapter 4, participants arranged alters on a large sheet with concentric rings. The inner ring included the individuals who are the most strongly tied, while the outer rings included weaker ties. This measure strongly correlates with the traditional ‘somewhat’ and ‘very’ close measures, however it adds additional granularity that should facilitate better estimation.6 For this variable, the innermost ring is given a value of 4 while the outermost ring is given a value of 1. Thus, a positive value means a positive association with increased tie strength. This value has been group-mean centered.7 This is important since not everyone used all four rings. This way, I can assess deviations from an average tie strength value for each respondent.
The next model includes qualities of alter, ego, and their relationship. Speciﬁcally:
Age (ego’s age and alter’s age [group-mean centered]): I include a continuous measure of ego’s age as well as a continuous measure of alter’s age. Given that adoption rates of mobile phones and Internet vary substantially by age (Ling, 2004), this is an important control. Raw age scores performed much better than relative age scores (i.e., difference between ego and alter’s age), suggesting that the relationship between media use and age is more related to social trends regarding media uptake and other age-related factors, such as personal mobility, than to differences between ego and alter. To note, this analysis uses a mean centered approach. The mean age for this Additional details about the correlation between somewhat/very close and these rings are found in Section 4.7.6.
There is an unfortunate aspect to the terminology here. Group mean refers to the mean for a speciﬁc collection of alters. So indeed, at least for the purposes of properly centering variables in these models, a group is a network.
CHAPTER 7. MEDIA USE WITH SPECIFIC NETWORK MEMBERS 195sample is 43. So a coefﬁcient represents a 1 year increase or decrease from this age on the number of media used.
Kin/non-kin: Despite having a rich granularity of possible roles, as mentioned in Section 4.7.4, the optimal distinction in these models was merely the kin/non-kin distinction. Kin in this case includes extended family members as well as immediate family. It is a dichotomous variable.
Have a computer: This is a dichotomous ego-level variable about whether or not ego has a computer. Since two of the possible media are computer based, I control for this signiﬁcant difference. To note, even those who do not have a computer at home may access others via email at work or elsewhere, although this is uncommon in the interviews.
Distance to ego (group-mean centered): This is a measure of the distance in kilometers between ego and alter. It has been log transformed to better represent differences in orders of magnitude (where 1 and 100 kilometers makes a far bigger difference than 2000 and 2100 kilometers).
Structural metrics (degree and density): Despite the bevy of possible structural metrics available, only the most basic metrics (degree and density) had any effect on the overall models. This is fortunate as these two measures are easily interpreted.
Degree represents the number of ties shared between ego and alter, whereas density represents the overall number of connections in the network. Theoretically, this means that alter is more embedded in ego’s personal network. Degree ranges from 0 (for an isolate) to 23 (the highest observed value). Density ranges from zero to one, where zero means no alters are tied to each other and one means all alters are tied together.
The third model includes communication frequency between ego and alter measured in terms of days per year. These variables include frequency of in-person contact, socializing, telephone, email and instant messaging. Since there were great asym
like distance, this allows for media use to more effectively represent differences in the order of magnitude between frequent and infrequent contact.8 As is common in multilevel models, most of the variables have been group-mean centered. This is a common practice in HLM work, so much so that some books advocate an almost ritualistic adherence to this particular transformation (Wheaton, 2004;
Wellman and Frank, 2001). Group mean centering usually decreases the standard errors while having little effect on the overall parameter estimates (Gelman and Hill, 2007). However, its use is not always ideal. When one is interested in the raw count for a variable, and interpreting this same raw count across respondents (alters in this
case), a raw variable can make more sense. Also, as Raudenbush and Bryk point out:
In some applications, of course, an X value of zero will in fact be meaningful. For example, if X is the dosage of an experimental drug, Xij = 0
sult, the intercept β0j is the expected outcome for such a subject. That is β0 j = E(Yij |Xij = 0). We wish to emphasize that it is always important to consider the meaning of Xij = 0 because it determines the interpretation
In this case, a degree of zero is an important and meaningful number—these are the 18 percent of isolates in the data set. Moreover, in this particular model, I am interested in the number of mutual ties, not deviation from an average number of mutual ties.
In all other continuous level-1 variables the parameter estimates are meaningful and indeed more signiﬁcant using a group mean centered approach.
Telephone frequency is a composite measure that includes both cell and landline contact. This is due to the nature of the question asked. The data permits me to know whether or not ego and alter used cell, landlines or both as dichotomous values (which is why I could produce the distribution plots of cell phone use above). However, it only asked about frequency of all combined telephone contact.
CHAPTER 7. MEDIA USE WITH SPECIFIC NETWORK MEMBERS 197
The multilevel models show that tie strength is indeed related to the number of media chosen but that its relationship is unstable and dependent on the other variables in the model. The ﬁrst non-null model simply includes tie strength. The standard error is 1.56 and there is a tiny (less than one percent) reduction in variance from the null model. This value is found at the bottom of Table 7.2. Superﬁcially, this does not bode well for the media multiplexity hypothesis, at least once you get beyond the threshold of ‘strong enough to be included in the personal network’. To note, a similar model using the dichotomous variable for somewhat close versus very close was similarly non-signiﬁcant.
Interestingly, tie strength performs much better with the inclusion of controls.
First, this suggests that the variable is somewhat unstable, but also that there are numerous reasons other than tie strength as to why an individual might want to increase their accessibility. For example, merely being more proximate to ego means that ego will be more likely to use a cell phone (due to exorbitant charges). Thus controlling for this factor seems to be an effective way to accentuate the effect of tie strength. This model has an overall R2 of.14. As one can see from R2 values, most of the variance explained is at the second (network) level. That is to say, most of the variation from these controls explains differences between networks, rather than differences within networks. This is understandable given the important level-2 controls such as having a computer and age.
The effects of age are particularly expected. As noted above and found in the literature, older people are more likely to be late adopters of technology. Since it ‘takes two to tango’, the model accounts not only to the respondent’s age, but the age of alter. The interaction term was non-signiﬁcant, suggesting that these are independent
Table 7.2: Nested random-intercept models predicting number of media (including face-to-face and socializing)
MEDIA USE WITH SPECIFIC NETWORK MEMBERS 199ference in age between ego and alter. It is simply an effect of the fact that there is a negative relationship between diversity of technology and age, and one that is worth controlling for. Again, this helps to increase the signiﬁcance of tie strength. Once controlling for the fact that older individuals are less likely to use many media, for whom will these people make exceptions? Their most strongly tied alters.
The variable “is kin” is a signiﬁcant contributor to the model. The coefﬁcient is negative meaning that all else equal, individuals will use fewer media with family members than with other individuals in the personal network. I believe this speaks to the logic of how individuals are included in the personal network. Namely, individuals who are family are a part of the network partially because they are family, not because ego actively maintains as solid a relationship with these individuals as with other alters. As Fischer has noted, urbanites are rather selective when considering kin in their network. He notes that “the decline in kin involvement with greater urbanism is less a sign of family disintegration and more a sign of selective family integration” (Fischer, 1982, 84). Fischer was referring to the inclusion or exclusion of family members in the network. However, it can also be extended to the inclusion or exclusion of family members into the ‘inner circle’ of accessibility via many media. This is to say there are different thresholds for inclusion. While one threshold might be ‘named in the network’, another would be access by media. Family members might have an easier time passing through the ﬁrst threshold but not the second. This is corroborated by Marin’s work on recall in name generators (2004). She ﬁnds that individuals often recall clusters of individuals. Since family have a discernible kin structure, people are likely to list off a large set of family members simply because they come to mind when people think about a few salient family members.
Curiously, the effect of being a family member is reversed once I include the frequency of contact. However, this value is insigniﬁcant suggesting that once I include
to make a clear claim about the number of media used.
The effects of network structure show a pattern that some might consider obvious, while others consider surprising. Under a logic that individuals will want to optimize their accessibility with their alters, I initially thought individuals of lower degree would have greater media use. This is in keeping with Wellman’s networked individualism. Implicit in this theory is the idea that individuals compensate for a fragmentation of personal networks with the use of more media. This is in keeping with the idea of the ‘person-as-portal’. For example, he suggests that [t]he shift to a personalized, wireless world affords truly personal communities that supply support, sociability, information, and a sense of belonging separately to each individual. It is the individual, and neither the household nor the group, that is the primary unit of connectivity (2001a, 238).
Yet, here I ﬁnd that individuals who are of higher degree are also those with whom ego uses more media. That is, there is a structural basis to connectivity, rather than a solely individualistic one. Moreover, this result persists when removing individuals of particularly high degree (the top 20 percent, who have a degree of 15 more). Also there is no curvilinear effect of degree. That means this ﬁnding refers to a linear increase in social connectivity rather than being about a few very highly connected individuals.
From an accessibility perspective as well as an efﬁciency perspective, this makes sense.
Individuals will seek to make contact with those who share the most mutual ties. One would want to give greater access to those with whom one is the most embedded in a network of relations. Also, if one needs to plan a future event or distribute information to one’s network, being available to someone who is tied to the most members of one’s network makes sense. The effect of density is surprisingly strong, as well as robust. It suggests that all else equal, individuals in a fully connected network will be contacted
knows each other). Part of maintaining a network as well as forging links across this network may be in linking network members together. However, this is a multi-causal relationship. Individuals may link each other via media, thus making the network more dense, or dense networks may make it easier to reinforce the sharing of contact information.