We compare three approaches to test for guilt aversion in two economic experiments. The first approach elicits second-order beliefs using self-reports. The second approach discloses first-order beliefs of matched players to decision makers, which are taken as exogenous second-order beliefs of decision makers. The third approach lets decision makers make choices conditional on a sequence of possible first-order beliefs of matched players. We find that the first and third approach generate similar results, both qualitatively and quantitatively. The second approach, however, generates significantly higher levels of ‘kindness’ for low levels of beliefs: at a second-order belief of zero, the probability of choosing the ‘kind’ action is between 43 and 65 percentage points higher than with the other approaches.