Associative Memory, Beliefs and Investments: A Replication and Extension

We study whether associative memory shapes beliefs, recall and investment decisions through five laboratory experiments with Chinese university students. Each subject evaluates 14 hypothetical companies, with the cue condition (Cue, NoCue, or SimilarCue) varied within subject. We elicit beliefs, memory reports and investment decisions in Beliefs, Recall and Invest experiments with independent subject pools. Beliefs and Recall experiments are each run with two context pairs (Cue+NoCue and SimilarCue+NoCue), while investment is run only with Cue+NoCue. Within each pair, the 14 companies are split 7-7 between the two conditions, with NoCue as the common within-subject baseline.

2026-08-31

2027-06-30

The study uses the following outcome variables, as explained under Experimental Design:

Beliefs (Cue+NoCue and SimilarCue+NoCue experiments)
• Subjects' second belief, i.e., their stated probability that the company is good in Part 2.

Recall (Cue+NoCue and SimilarCue+NoCue experiments)
• Subjects' second belief, i.e., their stated probability that the company is good in Part 2.
• Subjects' recall, i.e., the number of Part 1 positive and negative signals that they recall having seen when asked in Part 2.

Invest (Cue+NoCue experiment)
• Subjects' second belief, i.e., their stated probability that the company is good in Part 2.
• Subjects' investment share, i.e., the proportion of a fixed endowment that they allocate to the company's stock in Part 2.

The study uses the following secondary outcome variables, as explained under Experimental Design:
• Subjects' belief in Part 1, i.e., their stated probability that the company is good, elicited after the Part 1 signals and before the distraction task (Beliefs and Recall experiments).
• Subjects' belief in Part 2 in the Invest experiment, i.e., their stated probability that the company is good, used for the belief-investment consistency test.
• Comprehension-check pass rates.
• Self-reported demographic and risk-preference variables from the post-experiment questionnaire.

Following the original paper (Enke et al., 2024), we manipulate the presence of contextual cues to test whether associative memory drives belief overreaction and affects subsequent recall and investment decisions.

I. PROCEDURES AND LOGISTICS

At the Experimental Economics Laboratory of Dongbei University of Finance and Economics (Dalian, China), we conduct in-person laboratory experiments. We conduct five independent experiments with separate subject samples: two replicate the original Beliefs and Recall experiments (the Beliefs Cue+NoCue and Recall Cue+NoCue experiments); three are new extensions (the Beliefs SimilarCue+NoCue, Recall SimilarCue+NoCue experiments, and Invest Cue+NoCue).

The basic structure of the experimental setup is as follows.
• Subjects are asked to estimate the probability that each of 14 hypothetical companies is of a good type.
• Each company is either a good or bad type, determined by the computer independently for each company and with 50-50 probability.
• Participants do not learn about the actual realization of the computer draw, but receive multiple signals. Signals, presented in the form of news, are symmetric and correct with 65%, so Pr(news=good|company=good) = Pr(news=bad|company=bad) = 65%.
• News are independently drawn conditional on the true type of the company (good or bad).

Comprehension checks and exclusion: After instructions, subjects answer a short comprehension quiz. Subjects who answer more than one comprehension question incorrectly are excluded from analysis and replaced by newly recruited subjects until the target sample size is reached (this follows the original Enke et al. 2024 protocol; reported exclusion rate approx. 5%). In the Invest experiment, because the investment task and the return distribution are introduced only at the start of Part 2, a brief additional instruction screen and comprehension question on the investment payoff structure are administered at that point; the same exclusion criterion applies across all comprehension questions.

The experiment consists of the following parts:

Part 1:

• For each of the 14 companies, a subject first observes a sequence of news (the first-period signals). The news appears sequentially on a subject's computer screen. The number of signals per company, denoted by k, is 0, 1, 2, 3, or 4. The distribution across the 14 companies per subject (identical to Enke et al. 2024) is k=0 for 2 companies, k=1 for 2, k=2 for 4, k=3 for 2, k=4 for 4.
• The subject is then asked to state the probability (0-100%) that the company is good (we will henceforth refer to this as "first belief"). This belief is incentivized through a binarized scoring rule.
• This procedure is repeated for all 14 companies.
• Then the subject completes real effort tasks for 8 minutes as a distraction task.

Part 2:
• A subject observes a final piece of news about the company (the second-period signal).
• The subject then states the probability (0-100%) that the company is good (we will henceforth refer to this as "second belief", denoted by b_2). This belief is financially incentivized through a binarized scoring rule. This is the outcome elicited in the Belief experiment.
• In the Recall experiment, the subject reports the number of positive and the number of negative signals they recall having seen for that company. This recall is elicited after observing the final piece of news and before stating the second belief. This recall task is a surprise task.
• In the Invest experiment, subjects additionally make an investment decision: they allocate a per-company budget (1,000 ECUs) between the company's stock and cash. s in [0,1] denotes the investment share, i.e., the proportion of the budget allocated to the stock. The investment task and the stock's return distribution are introduced only at the start of Part 2, immediately before the first investment decision, so that the Part 1 belief task remains identical to the original Beliefs experiment. Subjects are told that they invest for a single period and therefore earn a profit based only on the company's next-period realized return, drawn from the following state-contingent distribution: in a good company, +20% with probability 0.65 and -10% with probability 0.35; in a bad company, -20% with probability 0.65 and +10% with probability 0.35. The order of the second belief and the investment decision is randomized at the subject level (counterbalanced across subjects). To avoid hedging between the incentivized belief and the investment decision, each subject is paid for either the belief task or the investment task for one randomly chosen company, determined at random at the end of the experiment.
• This procedure is repeated for all 14 companies.

Part 3: Questionnaire.

II. TREATMENTS

The design has two dimensions. The first is the outcome measure, i.e., what is elicited in Part 2. We distinguish three experiments, Beliefs, Recall, and Invest, that differ only in the Part 2 outcome: Beliefs elicits the second belief only; Recall additionally elicits recall of the first-period signals (after the second-period signal and before the second belief); and Invest additionally elicits an investment decision (with the order of the second belief and the investment decision randomized at the subject level). The second dimension is the context condition, of which there are three, Cue, NoCue, and SimilarCue, defined as in the original paper. The three conditions are:

• Cue. Each piece of news is communicated together with a context consisting of a story and an image. There is a one-to-one mapping between the type of news for a given company and its context: every piece of positive news for a company is shown with the same context, and every piece of negative news for a given company is shown with the same (but different) context. This replicates the original Main Individual treatment.
• NoCue. The setup is identical to Cue, except that each piece of news is communicated with a different context; a given story and image never appears twice, even for the same company and news type. This replicates the original No Cue Individual treatment.
• SimilarCue [NEW]. The setup is identical to Cue, except that for each (company, news direction) the contexts share the same theme while each occurrence draws a different instance of that theme. For example, every piece of positive news for a company relates to the same theme (e.g. celebrity endorsement) but with a different specific story and image each time; every piece of negative news relates to a different theme (e.g. environmental pollution) again with a different specific story and image each time. SimilarCue introduces theme-level association without exact repetition.

The context condition is manipulated within subject: for each subject the 14 companies are split 7-7 between two conditions, always with NoCue as the common baseline.
• For the Beliefs and Recall outcome measures we run both a Cue+NoCue experiment (the replication) and a SimilarCue+NoCue experiment (the extension);
• For Investment we run only the Cue+NoCue experiment.
This gives five experiments in total (there is no Investment SimilarCue+NoCue experiment). Cue and SimilarCue, conducted on different samples, are compared between subjects, each against its own NoCue baseline.

The Beliefs Cue+NoCue and Recall Cue+NoCue experiments constitute a loyal replication of the original Beliefs and Recall experiments. On top of it we add two extensions: the SimilarCue condition (the Beliefs SimilarCue+NoCue and Recall SimilarCue+NoCue experiments), testing whether theme-level association alone triggers associative recall; and the new Investment Cue+NoCue experiment, testing whether associative memory affects individual asset allocation in a non-strategic setting.

We commit to following all pre-registered hypotheses and the analyses specified above. Any deviation will be explicitly reported in the resulting paper, with the rationale stated and the corresponding pre-registered analysis also presented.

Not available

Randomization is conducted by computer through the oTree software when subjects begin the experiment at their laboratory terminal; all sessions are administered in person.

Randomization operates at two levels. First, each laboratory session runs one of the five experiments, and subjects do not know which one before they arrive. Second, for each subject in a given experiment, the following are independently randomized:
• The 14 companies are split evenly 7-7 between the session's two conditions (Cue vs. NoCue; SimilarCue vs. NoCue). The assignment is randomized across subjects so that every company appears in every condition across the relevant subject pool.
• The order of the 14 companies is randomized separately in Part 1 and Part 2.
• Each company is good or bad with equal probability of 0.5; given the type, all signal realizations are drawn independently with 65% accuracy.
• The number of first-period signals per company, k, is randomized subject to the constraint that the original distribution is preserved per subject (k=0: 2 companies, k=1: 2, k=2: 4, k=3: 2, k=4: 4).
• In the Invest experiment, the order of the second belief elicitation and the investment decision is randomized and counterbalanced across subjects.

There is no cluster randomization.

The subject. Assignment to one of the five experiments is between subjects, at the level of the laboratory session. Within an experiment, the context condition is manipulated within subject: the 14 companies are split 7-7 between the session's two conditions, randomized per subject.

No

n/a

494 subjects

The sample size will be given by:
• Treatments Beliefs Cue and Beliefs NoCue: 127 subjects total
• Treatments Recall Cue and Recall NoCue: 70 subjects total
• Treatments Beliefs SimilarCue and Beliefs NoCue: 127 subjects total
• Treatments Recall SimilarCue and Recall NoCue: 70 subjects total
• Treatments Invest Cue and Invest NoCue: 100 subjects total

Pre-Analysis Plan for Associative Memory, Beliefs and Investments: A Replication and Extension