Spatial Search: Experimental Evidence

How to cite

Morozov, Ilya; Malladi, Suraj; Marquina, Alejandro Martinez. 2024. "Spatial Search: Experimental Evidence." AEA RCT Registry. AEARCTR-0013777. https://doi.org/10.1257/rct.13777-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Ilya Morozov* , Suraj Malladi , Alejandro Martinez Marquina

* Primary investigator

Registry ID
AEARCTR-0013777
DOI
https://doi.org/10.1257/rct.13777-1.0
First registered
Last updated
Status
In development
Trial dates
2024-06-10 – 2024-06-30
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/13777

Abstract

In various contexts, such as shopping, experimentation, and problem-solving, agents engage in dynamic trial-and-error search processes. They decide both where to search and when to stop before choosing from a set of available options. We investigate how behavior changes with the complexity of the search problem. To this end, we design a controlled lab experiment where participants uncover positions with unknown prizes, paying a search cost for each revealed position. We vary the complexity by restricting the possible prizes at different positions and telling participants about these restrictions.

Intervention

Experimental Conditions:
1. [“Unrestricted”]. Prizes are not restricted.
2. [“High Variability”]. Prizes in adjacent positions cannot differ by more than 10 cents.
3. [“Low Variability”]. Prizes in adjacent positions cannot differ by more than 5 cents.
4. [“Quasi-Concave”]. Prizes in adjacent positions cannot differ by more than 10 cents AND the function that maps positions to prizes is quasi-concave (i.e., there is a unique maximum, and prizes decrease the farther one gets from this maximum).
5. [“Known Maximum”]. Prizes in adjacent positions cannot differ by more than 10 cents AND the participant knows that at least one position contains a prize of $1.

Experimental design

We built an online app where participants can perform search tasks. In each task, a participant is presented with 100 positions, each containing an unknown prize between $0 and $1. The participant must choose which and how many positions to uncover. To reveal a prize in a given position, the participant must pay a search cost. Prizes are uncovered sequentially, which allows the participant to learn from past discoveries.

Design details

Randomization method
Randomization done at the back-end of our online platform hosted on kelloggsearchdistribution.shinyapps.io.
Randomization unit
Participant
Clusters
2,500 participants.
Observations
2,500 participants.
Arms
We will aim to recruit 500 participants per condition, which amounts to 2,500 participants in total.
Primary outcomes
Hypotheses and outcome variables: As we vary complexity, we look for evidence that consumers change their search behavior in response to the information provided about the variability of prizes. The following hypotheses are based on the stylized predictions of rational models of spatial search in Malladi (2022) and Banchio and Malladi (2024): H1: Participants respond to the similarity of prizes in nearby positions. - Test H1A: Compare average net payoffs (maximum discovered prize minus total search costs) in conditions 1 and 2. Better performance in condition 2 would indicate that consumers make use of spatial correlation to select better search locations. - Test H1B: Compare average net payoffs in condition 3, 4, or 5 against those in condition 2. We expect participants to earn lower net payoffs in condition 2 due to higher search complexity. - Test H1C: Test whether time spent on search per round increases with complexity. Condition 1 is the most complex and should be the slowest, condition 2 should lead to faster search, and conditions 3, 4, and 5 should lead to the fastest search due to the lowest complexity. H2: Uncovering a low prize leads the participant to make a large step away from this position, whereas uncovering a high prize leads them to make smaller steps, staying near this position. - Test H2: In condition 2, regress step size on (a) first discovered prize, (b) latest discovered prize, (c) difference between the latest discovered prize and the maximum prize uncovered by that point, or (d) difference between the last and second to last discovered prizes. We expect a negative coefficient. Repeat the same analysis in condition 3 where we expect a larger coefficient (stronger correlation leads to larger steps). H3: Knowing there is a sweet spot leads participants to engage in “funneling”, taking progressively smaller steps and gradually zooming in on the maximum. Funneling is also more likely to arise in the low variability condition than in the high variability condition. - Test H3. We will test whether participants in conditions 3 and 4 engage in funneling search more frequently than those in condition 2. Funnelers take steps in the same direction after discovering positive news (uncovered prize is better than the previous prize) and change direction after discovering bad news. H4: Uncovering low prizes discourages participants from further searching in high-variability and low-variability conditions but not in unrestricted and known-maximum conditions. - Test H4A: We will compute the share of search tasks with recall (returning to a previously uncovered position) by condition. Seeing recall would suggest the presence of discouragement effects. - Test H4B: We will estimate the stopping regions nonparametrically as a function of (a) last uncovered prize, and (b) maximum prize the participant had before uncovering the last prize. We will use classification trees for estimation but may try other semi- and nonparametric estimation methods. We expect discouragement effects in conditions 2 and 3 but not in conditions 1 and 5.

Cited by

No citation data available for this registration yet.

Keywords