The canonical model of life-cycle labor supply predicts a positive response of labor supplied to transitory wage changes. We tested this prediction by conducting a randomized field experiment with bicycle messengers. In contrast to previous studies we can observe in which way working hours as well as effort respond to a wage increase and we have full control regarding the workers? anticipation of the wage increase. The evidence indicates that workers increase monthly working time and decrease their daily effort but since the working time effect dominates the effort effect overall labor supply increases. The decrease in daily effort contradicts the canonical model of intertemporal labor supply with time separable preferences, since the wage in our experiment directly rewarded effort. We show that a simple model of loss averse, reference dependent, preferences can account for both the increase in working time and the decrease in daily effort. Moreover, we elicit independent individual measures of loss aversion and show that workers who are more prone to loss aversion are more likely to reduce effort in response to higher wages. Our model and our results also reconcile the seemingly contradictory evidence reported in previous studies (Camerer et al. 1997, Oettinger 1999) of high frequency labor supply.