Extended Data Fig. 4: Simulation design.
From: Resampling reduces bias amplification in experimental social networks
For the network simulations and power analyses, we alternated between simulating participants’ judgements and the effects of our resampling procedure. We first fit the parameters \({\tilde{\Phi }}\) of the oracle models to Experiment 1 data Xe1 using Markov Chain Monte Carlo. One model was fit to Asocial/Motivated participants, and one to Social/Motivated participants. For the power analyses, at each wave t, we used either the asocial (wave 1) or social (waves 2-8) oracle to sample participant biases θt and simulate judgements θt (this process is illustrated here). By contrast, for the network simulations, we sampled a set of participant biases θ using the social oracle model and used these parameters to simulate judgements Xt at each iteration (there was no population turnover in our network simulations). To simulate our resampling procedure, we then fit IRT parameters θt to the simulated judgements at each wave. As in Experiments 1 and 2, these IRT models did not have access to the ground truth or participants’ true biases. We used our fitted IRT model to determine the importance weights wt for each judgement and resample a set of judgements \({\tilde{X}}_{{{{\rm{t}}}}}\) to propagate to the next wave or iteration. For each simulation, we repeated this process for 8 waves (the same fitted oracle model was used in all simulations).
