IMPACTS OF SOCIAL NETWORKS AND SPACE ON OBESITY The Rights and Wrongs of Social Network Analysis
THE SPREAD OF OBESITY IN A LARGE SOCIAL NETWORK OVER 32 YEARS Nicholas A. Christakis James D. Fowler Published: New England Journal of Medicine, 2007
Introduction Based on the spread of other activity through social networks, researchers examined whether obesity spreads from person to person. Does weight gain in one person influence weight gain in others with whom he/she has social contact? Authors have received considerable acclaim for studies. Additionally applied to loneliness, happiness, and smoking cessation. Increasing interest in this type of application of social networks due to rising availability of micro data.
Methods Data from: Framingham Heart Study Over 12,000 adults evaluated over a 32 year period. Defined social relationships Family, friend (directional), spouse, neighbor Three explanations for clustering of obese individuals Homophily: people chose to associate with those similar to them Confounding: friends may share attributes or experience unobserved concurrent events which cause their weights to vary at the same time Induction: one person may exert social influence on another
Methods Longitudinal logistic-regression models Y it+1 = α 0 + α 1 Y it + α 2 Age it+1 + α 3 Sex it+1 + α 4 Education it+1 + α 5 Y jt + α 6 Y jt+1 + ε Y it+1 = obesity status of the ego at time of interest Y it = obesity status of ego at previous time, controls for homophily and correlation Y jt+1 = obesity status of the alter at time of interest Y jt = obesity status of the alter at previous time
Methods Define BMI as binary variable. If obese ( > 30) designated with 1 Additional regressions included: Geographic component to measure distance effects Smoking status to measure impact of smoking behavior
Social Network Subcomponent, 2000
Animation Changes, 1975-2000
Results Related increases in the probability of obesity persists up to three degrees of separation. Spread of obesity depends on type and direction of relationship. Geographic distance does not play the same role as social distance.
Results Directional results: A mutual friendship exerted the strongest effect on obesity probability increase. Gender results: Same-sex friends (and siblings) have a larger impact than opposite-sex friends. Female to female friends spread was not significant! Spouse results: Husbands and wives effect their spouses similarly.
THE SPREAD OF EVIDENCE- POOR MEDICINE VIA FLAWED SOCIAL-NETWORK ANALYSIS Russell Lyons Published: Statistics, Politics, and Policy, 2011
Introduction Claims made by Christakis and Fowler are not sound, due to two types of errors: 1. Use statistical models that contradict their data, as well as conclusions. 2. Even if the models and tests are accepted, the results are interpreted incorrectly.
Major Problems 1. Directionality 2. Random Networks 3. Modeling and Model Estimation
Directionality Obesity spread based on perceived directional differences in friendships. Causality is the best explanation for: 171% difference for egos alters 57% difference from egos alters 13% difference from alters egos
Directionality 1. Differences are not statistically significant. Use 0 as a value when in fact data is too imprecise to distinguish values from 0. Compare coefficients from different models; cannot estimate the differences between the coefficients. 2. Idea that the differences are net of homophily is incorrect. Use time lag to control homophily, if lagged obesity were used for effect, conclude that social networks inhibit the spread of obesity. Conclude that homophily affects alters and egos in different ways. 3. The differences are consistent with all three possible explanations: induction, homophily, and environment. Directional differences found are what would be expected for these types of explanations; differences do not distinguish among them.
Random Networks Use both an actual network and several randomly generated networks. Friendship data used is incomplete. Creates friendships that do not exist: 45% of 5124 egos named a friend, but there were 3604 unique friendships in the actual network. Distinctions between the artificial and real networks are blurred. Observations in generated networks are compared with actual ones, despite lack of similar observations.
Modeling and Model Estimation Experiment was not done and not enough information for multi-dimensional cross-tabulation, so statistical models are used. Make assumptions of what the data would look like if they had an experiment or much more data. Do not pay enough attention to assumption, make errors which ultimately contradict other assumptions and conclusions made about directionality.
Modeling and Model Estimation Estimate coefficients using: general estimating equations (GEE) method. Method designed for repeated measures and dependencies, but requires independence between groups. All measurements on an alter were a single cluster or group, but these groups are not independent. Model requires coefficient β 1 = 0 but estimated that β 1 0.
Conclusions Major problems of studies: 1. Data not made available to others. 2. The unavailable data is sparse for friendships. 3. The models used to analyze the data ultimately contradict it and the conclusions. 4. The method used to estimate the models does not apply. 5. The statistical significance tests from the estimates do not show the proposed differences. 6. The proposed differences do not distinguish among homophily, environment, and induction. 7. Associations at a distance are better explained by homophily than induction.