Kevin Bryan, the author of one of my favorite blogs, writes summaries of recent research in his field. Before I get my own blogging voice, I will start as copycat and write some summaries of my own. My first summary is of a recent experiment in network cognition.

Social network theory makes assumptions about how much people know about the structure of their social network. There is some very interesting research that studies how strategic behavior in network games may change with different levels of information. This research shows that knowledge of the network structure is a critical assumption — the results change depending on how much you assume people know. Experiments are a great way to validate critical assumptions, so Dessí, Gallo and Goyal (2016) use an experiment to gauge how much people know about the structure of their social networks, investigate if this knowledge differs amongst people and identify any biases in perception.

The authors get students into a lab, show them a network structure for one minute (see figure below) and then ask them questions about the structure they have just seen. I was pretty skeptical about this design until I read about the one-two punch to internal and external validity. The biases they observe in the lab are also present in survey data from real-life social networks! People make the same mistakes whether they look at some network structure on a screen or they think about the structure of their own social networks. This is great for internal validity because showing a picture proves to have the same behavior as the real-life process that the authors wanted to mimic in the lab and it is great for external validity because the lab (at Cambridge) and the two surveys (of employees of an American company and an Italian research center) are situated in extremely different settings.

Figure 11 from Dessí, Gallo and Goyal (2016)

Even though the authors find the same biases in the lab and in existing surveys, I am still skeptical. The experiment is just a memory game using network diagrams. The three biases they identify are underestimation of the mean number of friends, overestimation of the “rare types” and underestimation of the number of frequent types. A “type” is an individual with a certain number of friends. James has three friends so he is a Type 3. The trick of the diagrams was there was only Type 3, 4 and 7 in the network. If the respondent is asked how many people have two friends in the network, the correct answer is zero. Surprise, surprise, some students filled in positive numbers for Type 1, 2, 5, 6 and 8, even though these types were not present in the network. This error generates the last two biases. The underestimation of the mean number of friends is probably the most interesting to compare to the survey data.

In the surveys, respondents must complete an adjacency matrix of their own social networks. Names of all individuals are in the rows and columns and the respondent fills a one if the person i and person j are friends and zero otherwise. If respondents do not perceive that certain friendships exists, you generate the bias of overestimation of rare types and underestimation of frequent types. The actual and perceived degree distributions (Figure 6) from the surveys show that the main bias was to overestimate the number of people who have zero friends. This sounds more like the survey respondents did not complete the full matrix and just left zeros as the default. The overestimation of Type-zero in the surveys brings about the underestimation of degree.

I am going to pause this blog post here and read the paper another couple of times. There are many other aspects that I didn’t mention (such as the public goods game on networks and individual heterogeneity). When considered in totality, Dessí, Gallo and Goyal (2016) make a great first contribution to a new strand of research in social network economics. The stage is set for further research.