Feedback matters: Children perform worse after negative feedback in a large online learning platform

Education is increasingly dependent on digital tools. Children now learn skills like mathematics in an online environment, typically involving computerized feedback. To assess how this feedback relates to children’s performance, we here leverage big educational data (N=395 children, T=60,228 trials and N=347 children, T=54,000 trials). Using Bayesian feedback models fitted on dense trial-to-trial data from an adaptive mathematics learning platform, we show that, in 60% of children, performance on a given trial varies depending on the feedback on the previous trial. At the group level, children have a 13% higher probability of making a mistake on trials following negative than positive feedback. However, children showed substantial individual differences: some demonstrated large adverse effects of negative feedback, others showed large favorable effects of positive feedback, and again others more moderate effects, effects in the opposite direction, or no effects. These results emphasize the need to tailor the feedback to the individual.