Experimental Design Details
For each candidate reference point the test asks whether the probability of selling depends on the current value and the candidate only through their difference. If it does, raising both by the same amount leaves the probability unchanged. Each candidate is tested in both arms. There are twenty-three candidates per arm, in four groups. Every quantity built from the history uses the values of the current stock only; the history restarts at the first value of each stock.
Single reference points (twelve): the situation benchmark, which the Framed arm displays and the Plain arm never shows; the values one, two, three, four and five rounds earlier; the first, second and third values of the stock; and the minimum, maximum and mean of the stock’s earlier values.
Functional rules (six): adaptive expectations; a moving average of the last M values; a peak-end blend of the highest earlier value and the most recent one; the Baucells–Weber–Welfens weighting of the value sequence; extrapolative expectations formed from the most recent change; and experience weights that favour recent or distant values. Each rule has one parameter, two for the Baucells–Weber–Welfens weighting. The parameter is estimated within each arm by an auxiliary logit fit of the sell decision on the gap between the current value and the rule’s reference point, and is re-estimated in every bootstrap draw of the test.
Mixture (one): six reference points entering jointly: the previous value, the first value, the minimum, maximum and mean of the earlier values, and the situation benchmark.
Lagged distribution (four): the last K values, for K = 2, 3, 4 and 5, treated as the support of one stochastic reference distribution in the sense of Kőszegi and Rabin. The test asks whether raising the current value and every point of that distribution together leaves the probability of selling unchanged.
The pre-analysis plan specifies the estimator, averaging weights, tuning constants, bootstrap and test statistics for every candidate.