Fig. 3: Results of skin conductance level analyses.

Differences in the temporal dynamics of the police interaction across experimental conditions depicted as (A) overall skin conductance level (SCL) and (B) conditional average treatment effects on SCL (N = 177). Note: SCL Skin conductance level. Results in panels A and B were generated by a multilevel random intercept model that used Bayesian Additive Regression Trees (BART), a machine-learning algorithm with conservative prior distributions, to estimate the non-parametric effect of time on SCL and on the treatment effects. Panel A presents loess-smoothed means of the posterior distribution of SCL reactivity estimated by the BART model. In panel B, dark lines correspond to the median of the posterior distribution of effects, boxes correspond to the inter-quartile range, and whiskers correspond to the 2.5^th to 97.5^th intervals. SCL was indexed by the average second derivative of EDA with respect to time during each 10-second window. A positive second derivative signals more rapid increases in EDA, which signals an impending peak and thus an SCL event, while a negative second derivative signals that an SCL event is ending and therefore the participant’s SNS arousal was in the process of recovering to homeostasis. The second derivative was used to follow recommendations from papers validating SCL data which have argued^43,44 that a better way to use information from a continuously monitored SCL is to calculate rates of change in slopes rather than discretely coding SCL events, which discards observations, especially when there is a low sampling rate, as in the present case. In our separate pilot study that validated our SCL facets, the second derivative metric was three times more effective at distinguishing challenge and threat levels to a lab stressor than a frequently-used skin conductance response event coding algorithm (see the SI).