Do #nzpol contributors exist in echo chambers?

As part of my Masters in Environmental Policy I did a course on Social Network Analysis—essentially analysing social structures by looking at the connections between people (or other things). For our final projects we had to something with Twitter data, so I looked at accounts tweeting with the #nzpol hashtag. The following is an adaptation of the full report (written in R Markdown).

Aotearoa/New Zealand (NZ) is a developed nation of around 5 million people, governed as a parliamentary democracy. Since 1996, national elections have been run on a ‘mixed member proportional’ basis, and have always resulted in representation from at least five parties. Accordingly, political debate encompasses a wide range of views and political figures, and having a relatively small population means that discussion is mostly conducted at a national level. Twitter is home to many of these debates, under the hashtag ‘#nzpol’. The community of people who tweet with this hashtag has been described in a media article by ‘The Spinoff’ as “small and attentive”, and “insular”—arguably a reflection of the wider political culture¹. The same article suggests that the same names “[come] up over and over again”, and that a significant proportion of active NZ political Twitter users are journalists and politicians themselves.

The real world representation of #nzpol

Research Questions

This research aims to understand more about the contributors to #nzpol discussion on Twitter by doing network analysis on the links between users. Two research questions have been defined to guide the enquiry:

1. Do ‘echo chambers’ exist (or not)? [RQ1]
2. Are there key actors (‘brokers’) that link between these clusters (if observed)? [RQ2]

‘Echo chambers’ refers to the “formation of groups of like-minded users framing and reinforcing a shared narrative”, and is something that can be observed across different social media platforms². In this context RQ1 seeks to understand if there are communities of users within the network whose grouping could potentially explained by political views.

Assuming that echo chambers are observed, RQ2 seeks to understand whether there are key accounts that join these clusters/communities—that is, determining accounts that connect otherwise unconnected accounts and/or fill gaps in the network ³. My hypothesis is that these are the accounts of journalists, news/media organisations or other reputable sources of information—by international standards, New Zealanders’ trust in news is still relatively high ⁴.

Methods and Results

Data processing

A three-week period (14 March 2022 until 3 April 2022) of #nzpol tweets were collected, a time frame chosen because of practical constraints. The dataset consisted of 6031 tweets in total from 813 distinct accounts, reduced to 776 accounts after verifying that they were neither suspended nor protected. The friends (links) of these accounts were then collected and filtered to only show links between accounts in the dataset.

Adding political leanings

To understand more about the presence (or absence) of echo chambers, an account classification system was determined. Accounts were labelled either as ‘news’ accounts, representing journalists or nominally neutral institutions (such as parliament); or they were labelled as ‘left’ or ‘right’, representing political leanings.

This classification was completed manually by reading Twitter account biographies, looking for key terms—an approach used elsewhere for classifying political leanings ⁵. In this instance a certain amount of subjectivity was necessary, but accounts were classified only if they demonstrated a clear leaning, or were political commentators with well-known views. Biographies had to be carefully read to make sure the account was not expressing opposition to an ideology (e.g. “[I] like arguing with lefties”).

Example of classification terms used:

• Left: ‘leftie’, ‘ardern supporter’, ‘labour’
• Right: ‘conservative’, ‘individual’, ‘freedom’, ‘liberty’, ‘national’
• News: organisations (‘TVNZ’, ‘RNZ’, ‘Discovery’, ‘Stuff’, etc.), official emails, ‘journalist’

An equal number of accounts were classified as left/right (25) along with 15 news accounts.

Generating the graph

From here a network ‘graph’ can be generated, showing the links between accounts:

Figure 1: A graph showing connectivity between accounts with greater than two degrees of connectivity.
Legend: ‘left’ (medium red circles); ‘right’ (medium blue squares); ‘news’ (large black circles).

Visually it is immediately apparent that the ‘left’ accounts and ‘right’ accounts are generally clustered on opposing sides of the graph, with ‘news’ accounts positioned reasonably centrally. This provides evidence in support of echo chambers [RQ1] and ‘news’ accounts acting as brokers [RQ2], however this can be confirmed by algorithmic analysis.

Network analyses

Community identification

Using a ‘walktrap’ algorithm, it is possible to identify communities within the graph, that is, parts of the graph that are “highly connected between them but with few links to other vertices” ⁶. If echo chambers exist, at least two communities would be expected and the account membership (accounts that make up each community) would be expected to be divded along political leanings.

Figure 2: The results of the walktrap algorithm, showing communities identified within the graph.
Legend: ‘left’ (medium circles); ‘right’ (medium squares); ‘news’ (large circles).

Indeed, two communities are present and all 25 of the ‘left’ accounts are members of one community, and almost all (21) of the ‘right’ accounts are members of the other community, providing strong evidence for the ‘echo chamber’ effect [RQ1]. ‘News’ accounts all appear to be a part of one community.

Identifying key actors (brokers)

To verify the presence (or not) of brokers [RQ2], brokerage scores were calculated using ‘Gould-Fernandez brokerage analysis’, with higher scoring accounts considered to be brokers ³. As expected, the vast majority of accounts had low brokerage values, but there were a few accounts with comparatively high values. To test the theory that ‘news’ accounts are key brokers, they were differentiated from ‘non-news’ accounts and then brokerage values graphed on a box plot. If ‘news’ accounts were key brokers, then it would be expected to see higher values in the box-plot for the ‘news’ accounts.

Figure 3: The brokerage scores between ‘non-news’ and ‘news’ accounts.

Indeed, the box plots provide evidence that ‘news’ accounts act as key brokers in the network—the median value for ‘news’ accounts is greater than that of ‘non-news’ accounts. Whilst there are a few ‘non-news’ accounts that have high brokerage scores, these are statistical outliers.

Discussion

The analyses suggest that there are echo chambers, divided along political leanings [RQ1] and key brokers, which are often ‘news’ accounts [RQ2]. Nonetheless, there are a number of limitations to this research.

Regarding data collection, the #nzpol hashtag only collects a subset of Twitter discussion and it was quite time-limited. A more complete analysis might include all accounts linked to NZ politicians/political actors, and be conducted over a longer period. In addition, ‘political leaning’ identification was simplistic and somewhat subjective, limitations which could be overcome by conducting algorithmic sentiment analysis of accounts, or introducing more classification categories^{5 7}.

In the network analyses, the walktrap function was used with the default steps number (4). If this value is altered, it outputs a different number of detected communities, and a more detailed analysis should consider the implications of changing these values. For example, the potential reasons why the ‘news’ accounts were grouped in the same community as the ‘left’ accounts was not analysed, but further exploration with the walktrap function might provide some clues.

Similarly, the brokerage function only tells us that certain accounts are key to ‘filling gaps’ in the network, but more detailed analysis might provide information on the nature of these links, as in, whether they are actually linking the echo chambers, or they are linking disparate accounts within them. In addition, not all news organisations can be considered as ‘neutral’, so the potential biases of these accounts could be included.

Despite these limitations, this research is still a useful insight into NZ political debate. Whilst it is hard to say how much Twitter discussion has implications for wider society, politicians and journalists do actively use it as an avenue for sharing and engage with the public–and there have been examples of social media pressure causing government action ^{1 8}. In addition, an increasing number of people get their news from social media, thus understanding more about these networks can help with tracking the spread of information, and combatting disinformation ⁴.

Conclusion

To conclude, in order to understand more about the dynamics of NZ Twitter political discussion, an analysis of accounts using the #nzpol hashtag was conducted. This was done using visual plots, community detection and brokerage analysis, following the classification of accounts by political leanings or ‘news’ status. The analyses provided evidence of echo chambers along political lines, and also the role of ‘news’ accounts acting as key brokers in the network. Future research could address some of the limitations of this study, for example by looking at a wider group of Twitter accounts or introducing different processes of classification. Whilst it is difficult to fully understand the wider societal implications, network analysis like this still provides useful insight into political debate, and understanding more about social networks can help with tracking information and disinformation.

Reference

Moon, G. (2022). Do contributors to NZ political discussion (#nzpol) on Twitter exist in echo chambers?.