There is vast heterogeneity in the human willingness to weigh others' interests in decision making. This heterogeneity concerns the motivational intricacies as well as the strength of other-regarding behaviors, and raises the question how one can parsimoniously model and characterize heterogeneity across several dimensions of social preferences while still being able to predict behavior over time and across situations. We tackle this task with an experiment and a structural model of preferences that allows us to simultaneously estimate outcome-based and reciprocity-based social preferences. We find that non-selfish preferences are the rule rather than the exception. Neither at the level of the representative agent nor when we allow for several preference types do purely selfish types emerge in our sample. Instead, three temporally stable and qualitatively different other-regarding types emerge endogenously, i.e., without pre-specifying assumptions about the characteristics of types. When ahead, all three types value others' payoffs significantly more than when behind. The first type, which we denote as strongly altruistic type, is characterized by a relatively large weight on others' payoffs – even when behind – and moderate levels of reciprocity. The second type, denoted as moderately altruistic type, also puts positive weight on others' payoff, yet at a considerable lower level, and displays no positive reciprocity, while the third type is behindness averse, i.e., puts a large negative weight on others' payoffs when behind and behaves selfishly otherwise. We also find that there is an unambiguous and temporally stable assignment of individuals to types. In addition, we show that individual-specific estimates of preferences offer only very modest improvements in out-of-sample predictions compared to our three-type model. Thus, a parsimonious model with only three types captures the bulk of the information about subjects' social preferences.