Thursday, November 1, 2018

Gender representation on UACES 2018 panels

I recently read a blog post on unequal gender representation on academic panels (especially all-male panels, so-called 'manels'). In it, the authors linked to another post that estimated that all-male panels are "getting more rare each year" at the International Studies Association conference.

I realized that after my social network analysis of panels at the 2018 UACES conference and the follow up on the role of gender in the conference's network structure, I already had the data to create a detailed estimate of gender representation on UACES panels.

The caveats

This analysis comes with similar caveats to my previous post on the topic.

First, I coded gender myself, so the results are based on my own interpretation, and not self-identification.

Second, the data is based on the UACES public program, and so does not measure the people who actually participated in person. For example, my panel (Panel 205) is listed as having a 40%/60% female/male split, but was an all-male panel if measured by the people in the room.

Finally, the network is held together by panels that share formal participants, and does not address more relevant - but difficult to measure - aspects such as who went to panels as an audience member.

The findings

As a reminder, for the UACES conference as a whole, the gender split was almost exactly 50-50: 258 participants were women (49.4%) and 264 were men (50.6%).

The headline findings on panel composition mirror this 50-50 split.

Across the 127 panels analyzed, the average (mean) representation overall was 51% women, 49% men. The most common percentage was an exact 50-50 split (16 panels).

Unsurprisingly (for me) these averages masked a lot of variation.

In Figure 1, I show all UACES 2018 panels categorized by the percentage of women participating. There were two all-male panels. In a little more than half of the panels, 50% or more of the participants were women (65 panels, 51%). Around a fifth of panels had more than 80% female participants (24 panels, 19%); of these, 14 featured only women.

Figure 1: UACES 2018 panels by percentage of participants that are women

This information is visualized differently in the network map in Figure 2, where darker green panels have a higher percentage of female participants.

Figure 2: UACES panels connected by shared participants. Greener panels have a higher percentage of participants that are women.

Panels, gender, and network importance: Four angles

How is the gender balance of a panel related to its importance within the network?

In this section I will explore this question using four network statistics: betweenness centrality, closeness centrality, eigenvector centrality, and weighted degree. These three values measure the importance of nodes in different ways, which allows us to take a multifaceted approach. In Figures 3-6, I take the network presented in Figure 2 and re-size the nodes based on their ranking on these statistics (larger nodes have higher scores).

In Figure 3, the size of the nodes depends on their betweenness centrality, a measure of "how often a node appears on shortest paths between nodes in the network." Nodes will a high betweenness centrality are more important to the overall network's connectivity, and removing them is more likely to lead to network fragmentation.

Figure 3: UACES panels re-sized according to their betweenness centrality, a measure of "how often a node appears on shortest paths between nodes in the network (larger nodes have higher centrality). Greener panels have higher percentages of participants that are women.

In Figure 4, the size of the nodes depends on their closeness centrality, "the average distance from a given node to all other nodes in the network."

Figure 4: UACES panels re-sized according to their closeness centrality, which measures average distance from all other nodes (larger nodes have a higher centrality). 

In Figure 5, the size of the nodes depends on their eigenvector centrality, which measures a node's importance based on the importance of the nodes it is connected to; "a node is important if it is linked to by other important nodes."

Figure 5: UACES panels re-sized according to their eigenvector centrality, which measures which measures a node's importance based on the importance of the nodes it is connected to (larger nodes have a higher centrality). 

Finally, in Figure 6, panels are re-sized according to their weighted degree, the number of connections they have with other panels.


Figure 6: UACES panels re-sized according to their weighted degree a measure of how many connections they have with other panels (larger nodes have a higher number of connections). 

Table 1 compares the scores on these four statistics of a number of sub-groups. First, I divided panels into those that had 50% or more female participants, and those that had 49% or less. I then compared the 14 all-female panels with the 14 panels with the lowest percentages of female participants (including the two all-male panels). Finally, for each statistic, I took the top  3, 5, and 10 panels and compared their average percentage of female participants.

Panels with more women had lower betweenness centrality, but were close to even on closeness, eigenvector, and degree. The all-female panels had much lower betweenness than the bottom 14, somewhat lower on closeness, even on degree and significantly higher on eigenvector. The top scoring panels for betweenness and closeness centrality had lower than average female participation, but higher than average for eigenvector centrality and weighted degree.

In summary, different measures of network importance come to different conclusions in this case.

Table 1: Comparing sub-groups on the four network statistics and percentage of participants that are women.


Conclusions

I'd like to emphasize again that these findings should be interpreted with caution: as noted above, the information used to build the network (shared participants) is far from ideal and there could be a number of explanations for the findings (which are not analyzed here).

Hopefully the analysis above can help spark productive conversations on these issues.