Social contagion theory: examining dynamic social networks and human behavior

Abstract

Here, we review the research we have conducted on social contagion. We describe the methods we have employed (and the assumptions they have entailed) to examine several datasets with complementary strengths and weaknesses, including the Framingham Heart Study, the National Longitudinal Study of Adolescent Health, and other observational and experimental datasets that we and others have collected. We describe the regularities that led us to propose that human social networks may exhibit a 'three degrees of influence' property, and we review statistical approaches we have used to characterize interpersonal influence with respect to phenomena as diverse as obesity, smoking, cooperation, and happiness. We do not claim that this work is the final word, but we do believe that it provides some novel, informative, and stimulating evidence regarding social contagion in longitudinally followed networks. Along with other scholars, we are working to develop new methods for identifying causal effects using social network data, and we believe that this area is ripe for statistical development as current methods have known and often unavoidable limitations.

Figure 1

Results from network permutation tests, using five different observational and experimental datasets, show significant associations up to between 2 and 4 degrees of separation for a variety of 15 different behaviors and affective states. The Y axis represents the percentage increase in probability that an ego has the trait of interest given that an alter has it, compared with the probability that an ego has the trait given that the alter does not have it. Vertical black lines indicate 95% confidence intervals. For more details, see the related manuscripts cited in the text. Colors indicate data source: yellow: Framingham Heart Study Social Network [14]; blue: AddHealth [1]; green: lab experiment [6]; red: Facebook strong ties [2]; orange: Hadza hunter gatherers [5].

Figure 2

Network visualizations showing clustering in obesity (top) and happiness (bottom) in the Framingham Heart Study Social Network in 2000. The top graph shows the largest connected component of friends, spouses, and siblings for whom information about body mass was available. Node border indicates gender (red=female subject, blue=male subject), node color indicates obesity (yellow=BMI>30), node size is proportional to BMI, and tie colors indicate relationship (purple=friend or spouse, orange=family). The bottom graph shows a portion of the largest component of friends, spouses, and siblings for whom information about happiness was available. Each node represents a subject and its shape denotes gender (circles are female, squares are male). Lines between nodes indicate relationship (black for siblings, red for friends and spouses). Node color denotes the mean happiness of the ego and all directly connected (distance 1) alters, with blue shades indicating the least happy, and yellow shades indicating the most happy (shades of green are intermediate). The bottom image involves both ‘geodesic smoothing’ and sampling, as noted in the text.

Figure 3

Schematic of a network infection and sampling process. (A) The full (unobserved) network with the initially infected node colored green (upper left corner of the network). (B) The shortest path from the source node to the target node colored red (lower right corner of the network) corresponds to the most likely infection path in the fully observed network and has a length of 2. (C) The (unobservable) spreading process unfolds in the (unobserved) network. The actual path taken by the infection is shown with wavy edges. The target node is reached in three steps giving a length of 3. (D) The partially observed network has some nodes and links missing depending on the sampling. The shortest path from source to target has a length of 3 (shown in the dotted lines), corresponding to the length of the most likely path taken by the infection. In this case, using the shortest path length in the fully observed network to estimate the actual path length would result in an underestimate of path length, whereas using the path in the partially observed network correctly yields a path length of 3.

Figure 4

Illustrative results from longitudinal regression models for various relationship types and outcomes. Horizontal bars show 95% confidence intervals derived from GEE models by simulating the first difference in alter contemporaneous outcome (changing from 0 to 1) using 1000 randomly drawn sets of estimates from the coefficient covariance matrix and assuming all other variables were held at their means.