Despite being discussed often in both practice and academic circles, the sunk cost effect remains empirically elusive. We develop an experimental design that overcomes the deficiencies of the previous research in identifying the effect. Additionally, we propose a model based on reference-point dependence, and test a novel prediction that when a sunk cost intervention constitutes a gain the effect is smaller (or non-existent) than when it is coded as a loss. We randomize the price of entry into a counting game and observe its effect on playing time. The price (low, medium, or high) is realized after the participant’s decision whether to enter or not. The distribution is fully disclosed prior to the decision, with the low price placing the participant in the gain region and the high price – in the loss region (with placements robust to different reference points).