Information, entropy and the paradox of choice: A model for understanding human choice behavior

Technical Summary: Information, Entropy, and the Paradox of Choice

Problem Statement
The paper addresses the "paradox of choice" or "choice overload," a phenomenon where an abundance of alternatives leads to decreased decision-making satisfaction rather than increased utility. While behavioral evidence consistently demonstrates an inverted U-shaped relationship between choice set size and satisfaction (where satisfaction peaks at an intermediate number of options and declines as options increase), quantitative frameworks linking information processing to this subjective experience remain underdeveloped. Existing models often rely on qualitative trade-offs between benefits and costs but lack a rigorous information-theoretic basis to explain the generic non-monotonic dependence of satisfaction on set size.

Methodology
The authors propose a theoretical framework grounded in information theory, specifically utilizing Shannon entropy and relative entropy (Kullback-Leibler divergence).

1. Theoretical Model:

   • Preference Representation: Individual preferences within a choice set of size $N$ are modeled as a probability distribution $P = \{p_1, p_2, ..., p_N\}$, derived from subjective ratings via a softmax function.
   • Baseline State: A state of "indifference" is defined where all options are equiprobable ($p_i = 1/N$), corresponding to maximum Shannon entropy ($H_{max} = \log_2 N$) and minimal satisfaction.
   • Effective Information ($I$): The authors define choice satisfaction as effective information, quantified as the relative entropy between the observed choice distribution ($P'$) and the baseline indifference distribution ($P$). Mathematically, $I = H_{max} - H_{obs}$, where $H_{obs}$ is the observed entropy of the individual's preferences.
   • Optimal Set Size ($m$): The model introduces a cognitive constraint parameter $m$, representing the maximum number of alternatives an individual can effectively evaluate.
     • For $N \leq m$: The individual can discriminate among options, leading to a non-uniform distribution where $H_{obs} < H_{max}$. As $N$ increases, $H_{max}$ grows faster than $H_{obs}$, causing $I$ (satisfaction) to rise.
     • For $N > m$: Cognitive limitations prevent full evaluation. The model assumes the unprocessed options ($s = N - m$) are assigned equal probabilities (missed choices), causing the distribution to become more uniform. Consequently, $H_{obs}$ approaches $H_{max}$, and effective information $I$ declines.
   • Simulation: The model was implemented using C++11 to simulate probability distributions and calculate effective information across varying set sizes and parameters ($m$, and parameters $a, b$ governing the probability mass of missed choices).

2. Behavioral Experiment:

   • Participants: 240 students were divided into six groups ($N=4, 8, 12, 16, 20, 24$).
   • Task: Participants rated ceramic mugs (varying in design/color) on a 7-point Likert scale.
   • Measurements:
     • Survey 1: Ratings transformed into choice probabilities to compute model inputs.
     • Survey 2: Four Likert-scale items measured perceived abundance, overwhelm, confusion, and choice satisfaction (specifically: "If I were to buy one mug, I would have chosen my first priority").
   • Analysis: Statistical comparisons (ANOVA, t-tests) were performed on survey responses. Model predictions of effective information were fitted to a quadratic function to compare with experimental satisfaction data.

Key Contributions

• Quantitative Metric for Satisfaction: The paper introduces effective information (the reduction in uncertainty relative to a baseline of indifference) as a robust, quantitative metric for choice satisfaction.
• Mechanistic Explanation: It provides a principled derivation of the inverted U-shaped relationship. The model demonstrates that satisfaction is non-monotonic because it is a balance between the informational benefit of having more options (increasing $H_{max}$) and the cognitive cost of processing them (increasing $H_{obs}$ toward $H_{max}$ due to missed choices).
• Empirical Validation: The study validates the theoretical framework against behavioral data, showing that the effective information derived from participant ratings closely matches the observed inverted U-shaped pattern of subjective satisfaction.

Results

• Experimental Findings:
  • Perceived abundance, overwhelm, and confusion increased monotonically with choice set size.
  • Choice satisfaction exhibited a significant inverted U-shaped dependence on set size ($F(5, 234) = 7.13, p < 0.001$). Satisfaction was significantly higher at $N=12$ compared to both $N=4$ and $N=24$.
• Model Performance:
  • The model successfully reproduced the inverted U-shaped curve.
  • A quadratic fit to the model's effective information predictions yielded a coefficient of determination $R^2 = 0.947$.
  • The model predicted a peak satisfaction at $N^* \approx 13.39$, which closely aligns with the experimental peak at $N=12$.
  • The analysis confirmed that the structure of choice probabilities (specifically the tendency toward uniformity in large sets due to cognitive limits) is the key determinant of the satisfaction curve.

Significance and Claims
The paper claims to offer a "principled theoretical account" of the paradox of choice by linking subjective satisfaction directly to the informational structure of preferences.

• Theoretical Value: It moves beyond qualitative trade-off models to provide a mathematical framework where satisfaction is defined by the ability to differentiate options from a state of indifference.
• Generalizability: The framework is presented as a general mechanism applicable across domains (from consumer goods to life decisions), suggesting that the inverted U-shape is a population-level tendency arising from heterogeneous cognitive constraints.
• Practical Implication: The authors suggest that the findings imply decision environments should be optimized not by maximizing variety, but by structuring information to facilitate effective comparison within an individual's processing capacity.
• Modesty: The authors explicitly acknowledge limitations, noting that the model treats the optimal set size ($m$) as a fixed parameter rather than dynamically emergent, assumes a simplified representation of alternatives (ignoring multi-dimensional attributes), and does not currently distinguish between physical and online environments. They frame the model as a parsimonious starting point for understanding decision-making in environments with abundant alternatives.