R and Inquisit scripts for ensemble coding of trustworthiness in faces:

This repository contains data, stimuli, & code from experiments (Inquisit scripts) and analyses (R scripts) on the ensemble coding of trustworthiness in faces. Six studies were conducted online with Inquisit Web, using participants from Amazon Mechanical Turk. Details for these studies are as follows:

• Study 1 (ensemble1; n = 50 subjects): Subjects did a task with 100 trials, where they estimated the mean trustworthiness of a group of 4 faces (grid faces) vs. a single face (compare face). We used computer-generated faces from Alex Todorov's database of 25 maximally distinct faces on the trustworthiness dimension. On each trial, a fixation appeared (500ms), followed by a grid of 4 faces (2000ms), followed by the compare face (until response). Subjects were asked "Is THIS face more (M) or less (L) trustworthy than the group?" (no time limit to respond). The experiment program (.iqx) constrained each trial such that faces from the same identity were never shown during the same trial, and the average of trust ratings for the grid faces cannot equal the trust rating of the compare face (if so, the program flags the trial and runs another one). Subjects were given the following instructions:

"You are going to be doing a timed perception task on pictures of different faces. On each trial, you will first see a grid of 4 faces. Next, you will see a single face, and we will ask you to judge whether that one face is MORE or LESS TRUSTWORTHY than the group of 4 faces that appeared before it. If you think the single face looks MORE trustworthy than the group, press the M button on your keyboard. If you think the single face looks LESS trustworthy than the group, press the L button on your keyboard. Note that the grid of 4 faces will only appear for a couple seconds, so make sure to pay close attention at all times."

• Study 2 (ensemble2; n = 50 subjects): Same as Study 1, except that subjects were tasked with comparing the VARIANCE of trustworthiness between two grids of 4 faces each (rather than the mean trustworthiness for 4 faces vs. a single compare face). The experiment program (.iqx) constrained each trial such that faces from the same identity were never shown during the same trial, and the variance of trust ratings for the first grid of faces could not equal the variance of trust ratings for the second grid of faces (if so, the program flags the trial and runs another one).

• Study 3 (ensemble3; n = 50 subjects): Same as Study 1, except that we used images of White male faces from the Chicago Face Database. We calculated a trust score for each face by z-scoring the normed trustworthiness ratings provided for the stimuli in the Chicago Face Database codebook (see "stimuli" folder). All other aspects of the experiment were the same as Study 1.

• Study 4 (ensemble4; n = 411 subjects): This was a norming study to get z-scored trustworthiness ratings for other faces aside from White males in the Chicago Face Database. We did this by having subjects give trust ratings for all Black/White male/female faces in the Chicago Face Database, using clickable 9-point scales ("How trustworthy do you find this person?"). All subjects rated 190 faces each, after being randomly assigned to 1 of 4 conditions (the R script for this study uses ratings from condition 1 to create trust z-scores, which are then outputted to a new .csv file and used in Studies 5-6):

  1. Mixed race and gender - all faces randomized by race/gender in one block
  2. Separate race and gender - all faces blocked by race & gender (i.e., 4 blocks of White male, White female, Black male, & Black female; block order randomized)
  3. Mixed race and separate gender - faces blocked by gender, with races mixed (i.e., 2 blocks of all Black/White female faces and all Black/White male faces; block order randomized)
  4. Mixed gender and separate race - faces blocked by race, with genders mixed (i.e., 2 blocks of all male/female Black faces and all male/female White faces; block order randomized)

• Study 5 (ensemble5; n = 124 subjects): Similar to Study 1, with 3 main changes:

  1. We used the faces and trust z-scores from Study 4. As with Study 1, on each trial, a fixation would appear (500ms), then the 4 grid faces (2000ms), followed by the compare face (until response). Subjects were asked "Is THIS face more (M) or less (L) trustworthy than the group?" (no time limit to respond). The experiment program (.iqx) constrained each trial such that faces from the same identity were never shown during the same trial, and the average of trust ratings for the grid faces cannot equal the trust rating of the compare face (if so, the program flags the trial and runs another one).

  2. Subjects completed 120 grid trials, where the races of the faces varied: 40 of the trials contained all White faces, 40 of the trials contained all Black faces, and the other 40 trials were mixed-race (i.e., there was at least 1 White face and 1 Black face in the grid/compare set of 5 faces).

  3. Subjects filled out additional scale measures after completing the 120 grid trials. These included the Attitudes Towards Blacks (ATB) scale, Internal and External Motivation to respond without prejudice (IMS/EMS), SDO7 scale, and race feeling thermometers (i.e., on scales of 0-100, "How cold or warm do you feel towards White-Americans?" and "How cold or warm do you feel towards African-Americans?).

• Study 6 (ensemble6; n = 123 subjects): Just a replication of Study 5, without the ATB and IMS/EMS scales (SDO7 and feeling thermometers were still included in Study 6).