Parsimonious Theorising

How to cite

Kazemi, Saeid; Camuffo, Arnaldo; Jannace, Diego; Gambardella, Alfonso. 2025. "Parsimonious Theorising." AEA RCT Registry. AEARCTR-0015867. https://doi.org/10.1257/rct.15867-1.0

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

Saeid Kazemi* , Arnaldo Camuffo , Diego Jannace , Alfonso Gambardella

* Primary investigator

Registry ID
AEARCTR-0015867
DOI
https://doi.org/10.1257/rct.15867-1.0
First registered
Last updated
Status
Completed
Trial dates
2025-04-24 – 2025-04-25
Countries
Italy (Milan)
Registry record
http://www.socialscienceregistry.org/trials/15867

Abstract

We apply the “law of parsimony”, also referred to as “Occam’s razor”, to entrepreneurial theories of value. Our key intuition is that decision-makers should build theories of value that are parsimonious, i.e., the simplest contingent on the level of novelty/complexity of the problem at hand. We explore this research question developing a formal model of how and why decision-makers should select the optimal level of complexity of the conceptual causal structure underlying their theories of value. By theory complexity we mean the number of elements, relationships, and layers of articulation it comprises.

This randomized experiment tests the empirical predictions of a formal model predicting that contingent on the level of novelty/entropy of the problem, the theory’s conceptual causal structure has an optimal level of complexity. Additionally, our interventions will allow to test the causal effect of sharper causal reasoning, reduced cognitive constraints, access to broader information sets, and different levels of problem novelty/entropy on the architectural properties of theories.

To test these predictions, we leverage on the randomized control trial registered on this website with registration identifier AEARCTR-0014579. We use the same experimental design and data, with additional variables and measures to investigate our theoretical model. Participants will be tasked with developing theories of value for assigned entrepreneurial challenges that vary in their level of novelty/entropy. Participants will also be exposed to different technological conditions that expand their information set / knowledge base, reduce cognitive costs, and influence the level of causality of their reasoning. By virtue of these manipulations as well as a data collection survey to gather individual characteristics, we aim to observe how the key factors of our formal model affect the complexity of theories, and more specifically, their level of parsimony.

Intervention

Interventions

In this study, we employ two main types of experimental interventions:

Novelty/Entropy Treatment: this intervention exogenously changes the level of novelty/entropy of the target state (high/low problem novelty/entropy)

AI Treatment: This intervention exogenously affects the cognitive capabilities and the size and understanding of the decision maker’s domain. It does so by assigning participants to one of three following experimental conditions:

Human-only – No additional assistance. Participants will be instructed to develop their theories to address the given decision challenge without any additional assistance or tools.

General AI Assistance - Access to an advanced general AI model (OpenAI O3-mini). Participants will have access to a General AI model as a support to aid in brainstorming ideas, answering questions, and refining their theories.

Specialized Agentic AI Assistance – Access to a Specialized Agentic AI System trained on notions of causality and fit (orthogonality of causes). Participants will have access to a specialized agentic AI system based on OpenAI O3-mini, instructed on notions of causality and fit (orthogonality of causes).

We also leave room for exploring the possible interaction effects between these interventions, although they are not the primary focus of our study.

Participants are randomly assigned to one of the experimental conditions by using blocked minimized randomization, due to the different levels of attrition across arms.

Experimental design

The design of this randomised controlled trial mimics that of the trial “A Design Centric View of AI Enhanced Decision Making” (Camuffo et al., 2025), registered on the AEA registry website with registration identifier AEARCTR-0014579. Whereas we use the same experimental design, we complement data collection with additional variables to test our theoretical model.

Design details

Randomization method
To allocate participants to experimental conditions, we use minimized blocked randomization (batch size = 50) based on three covariates: a) education level and field; b) AI aversion; and c) level of self-confidence. This provides the baseline allocation of participants to the experimental conditions. To ensure a balanced number of participants in each group and to manage the flow of the experiment, we limit the number of participants who can enter the experimental platform at the same time. This helps prevent an imbalance at the end of the experiment caused by participants who start but do not finish the experiment.
Randomization unit
Individuals
Clusters
No clusters
Observations
2500 managers from Prolific following the trial “A Design Centric View of AI Enhanced Decision Making” (Camuffo et al., 2025), registered on the AEA registry website with registration identifier AEARCTR-0014579
Arms
208
Primary outcomes
The outcomes that we measure are the following: Main Outcomes: Complexity of the theories: Number of nodes, edges, layers, and conditional independence of the theory. Expected Probability of Success: Participants' estimated likelihood of successfully achieving the target state of the decision problem based on their developed theories. Expected probability of success of the Theory as evaluated by human experts and/or LLMs Main Explanatory Variables: Novelty/entropy of the target state Domain size and understanding of the decision-makers Cognitive capabilities of the decision-makers Covariates Demographic data about participants: Age Gender Education Level Education Field Industry Job Function Experience Managerial Experience Entrepreneurial Background Salary Country of employment AI Aversion AI Familiarity Confidence Cognitive self-assessment Knowledge depth Knowledge breadth
Secondary outcomes
As a secondary outcome, we are also interested in exploring the following performance hypothesis: More parsimonious theories will be evaluated (by human experts and/or LLMs) as having higher expected probability of success.

Cited by

No citation data available for this registration yet.

Keywords