Experimental Design
Game Structure
The experiment is based on a discrete cheap talk game played over 60 rounds between one Sender (S) and one Receiver (R).
- State of the World (s): Uniformly drawn from S={1,2,3,4,5} in each round, privately observed by S.
- Message (m): S sends a cheap talk message m from M={1,2,3,4,5}, of the form "The state is m."
- Action (a): R observes m (but not s) and chooses an action a from A={1,2,3,4,5}.
- Sender payoff: U(a,s,b) = 110 − 20 |s+b−a| ^1.4.
- Receiver payoff: U(a,s) = 110 − 20 |s−a| ^1.4
- Incentives: R's optimal action is a=s. S's optimal action is a=s+b. The parameter b captures the degree of preference misalignment.
- Perturbations: A small, continuous perturbation is applied to the bias parameter b in every round to reduce potential demand effects, though the overall incentive structure remains unchanged. Participants are informed that payoffs vary from round to round.
Treatments
The experiment uses a 2×2 between-participants design varying the bias parameter b across two unannounced parts of the 60 rounds.
- Part One (Rounds 1–30): Aligned: b=0.2 (Low Conflict), Conflict: b=2 (High Conflict).
- Part Two (Rounds 31–60): Frequent: All 30 rounds use b=1, Rare: 10 rounds use b=1, and 20 rounds repeat the bias from Part One (b=0.2 or b=2). The 10 rounds with b=1 are randomly pre-selected and fixed across sessions.
- Treatments: Aligned-Rare (AR), Aligned-Frequent (AF), Conflict-Rare (CR), and Conflict-Frequent (CF).
Primary outcome: Correlations between states and actions.
Prediction: ρ_{AR} > ρ_{AF} > 0.65 > ρ_{CF} > ρ_{CR}.
Primary analysis (aggregate level)
- Kruskal-Wallis for testing hypotheses of equal correlations across all four treatments. If Kruskal-Wallis rejects hypothesis, Dunn’s post-hoc tests for pairwise comparisons with a Bonferroni correction.
- Wilcoxon exact one-sided signrank tests to test separately whether ρ_{AR} > 0.65, ρ_{AF} > 0.65, ρ_{CF} <0.65, ρ_{CR} < 0.65
Robustness checks
- Ordered logistic regressions of action on state, with errors clustered at the matching group level (for first test)
- Regression method suggested by Cai & Wang, 2006 (for second test)
Secondary analysis (individual level)
- Defining habitual participants: We classify a participant as habitual if their decisions satisfy two requirements: (i) high automaticity, and (ii) reduced dependence on goals. For high automaticity, we require participants to converge to a stable strategy in part one. Since the habit formation process takes time, we ignore the first ten rounds where participants could potentially still be using trial and error. For reduced dependence on goals, we require participants to use the same stable strategy in part two as they did in part one, despite the change in the underlying bias. A participant is classified as habitual if their decisions satisfy both requirements.
- Defining possible strategies: For each of the five states observed, senders can choose among five messages, resulting in 3,125 possible strategies. Symmetrically, for each of the five messages received, the receivers can choose among five actions, also resulting in 3,125 possible strategies. For each strategy, we compute the percentage of decisions consistent with it. The consideration set consists of strategies which are consistent with at least 60% of participant decisions. If the set consists of more than one strategies, we select the one which matches the highest percentage of decisions. The threshold of 60\% is used for both part one and part two.
- Comparisons: Following prior studies, we present qualitative evidence rather than formal statistical tests. We compare: (i) habitual participants between Rare and Frequent treatments, (ii) habitual participants between Aligned and Conflict treatments, (iii) decision time between habitual and non-habitual participants, (iv) CRT scores between habitual and non-habitual participants, and (v) change in decision time (Part One - Part Two) between habitual and non-habitual participants.