Fig. 2: Race-based worker rating and income differences eliminated after change to thumbs up/down scale.

a, The average likelihood that a customer rates a completed job using highest rating (five stars or thumbs up) for white (dark grey) versus non-white (light grey) workers; error bars represent 95% confidence intervals (n = 30,540). Observations were grouped into 30-day periods; negative values represent the five-star scale and positive values represent the thumbs up/down scale. Using the five-star scale, non-white workers had a lower likelihood of receiving the highest rating than white workers. Using the thumbs up/down scale, there is no evidence for a difference in the likelihood of receiving the highest rating after the rating-scale change. In the 30-day period before the rating-scale change, non-white workers were less likely to receive the highest rating (five stars) than white workers (88.5% versus 92.4%; −3.9 percentage points, t(1,899) = 2.76, P = 0.006, 95% confidence interval (CI), [−0.07, −0.01]). In the 30-day period after the rating-scale change, we do not find evidence for a race-based difference (93.7% for non-white workers versus 94.1% for white workers; −0.4 percentage points, t(1,717) = 0.308, P = 0.758). b, The marginal effects from Table 3 model 2 (n = 5,485). The sample includes workers who joined within 180 days of the rating-scale change. For workers who joined before the rating-scale change, only jobs completed before the change were analysed. Income is the normalized amount that workers received after completing the job, and error bars represent 95% CI. Using the five-star scale, white workers (crosses and solid line) earned a normalized income of 169.48 and non-white workers (triangles and dashed line) earned 136.67 (−32.81; t(4,939) = −3.27, P = 0.001, 95% CI, [−52.49, −13.14]). After the rating-scale change (to thumbs up/down), white workers earned 169.44 and non-white workers earned 179.00 (+9.56; t(4,939) = 1.15, P = 0.250, 95% CI, [−6.73, 25.85]).