Extended Data Fig. 3: Tie determinants of respondent accuracy.
From: Cognitive representations of social networks in isolated villages
We find that a range of properties of ties have statistically significant associations with their tendency to be accurately conceived. In each panel, LHS, marginal effect on accuracy in ROC-space. Grey shading represents the 95% bootstrapped confidence ellipse of the predictions from the two models. RHS, marginal effect of each individual accuracy measure: the true positive and false positive rates and the summary measure, Youden’s J. Intervals represent 95% confidence levels, calculated via normal approximation for the two rates, and bootstrapped for J, around the mean estimates. Estimates are stratified by whether they are of a tie among kin or not. (a) Relationship type; we include a covariate for the two relationships considered, free-time or personal-private. (b) Gender combination of tie members, for example, both women or both men. (c) Average age of tie members. (d) Difference in age between tie members. (e) Average degree of tie members. (f) Difference in degree between tie members. (g) Cognizer-to-tie geodesic distance. Individuals may or may not have a defined path between them in the reference network; when there is a path, individuals exist at a geodesic distance defined as the minimum number of steps between them; note that individuals who do not have a path between them necessarily have a path in at least one of other networks considered in this study, by design. (h) Distance between tie members. When a tie does not exist between two individuals, a specific geodesic distance may separate them (or they may have no path between them in the network). The TPR is set to the population average; but it does not have a meaningful interpretation in assessments of ties that do not exist. Parameters are fit from separate models of each rate, conditional on tie verity in the reference network. See Methods for details of model specification. Results are from n = 9,998 survey respondents in both panels, corresponding to 177,928 individual responses for the TPR estimates, and 477,393 responses for the FPR estimates.
