Decision Conflict, Power Logit, and the Deferral Outside Option

1. Problem Statement

The paper addresses a gap in discrete choice modeling regarding choice deferral (opting out or delaying a decision).

• The Phenomenon: Decision-makers often avoid making an active choice not because the available alternatives are undesirable, but because the decision is difficult (high "decision conflict"). This occurs when alternatives are similarly attractive, creating hesitation.
• Limitations of Existing Models: Standard Logit models with an outside option (e.g., Anderson, Palma, and Thisse, 1992) treat the outside option as having a fixed, exogenous value. In these models, deferral occurs only if the outside option's fixed utility exceeds the utility of the active alternatives. This fails to capture situations where deferral is driven by the difficulty of comparing active alternatives (decision conflict) rather than their absolute unattractiveness.
• The Gap: There is a need for a model where the value of the outside option is endogenous and menu-dependent, specifically increasing with the difficulty of comparing the active alternatives.

2. Methodology and Theoretical Framework

A. General Decision-Conflict Logit Model

The author extends the foundational Luce (1959) logit model and McFadden's (1973) econometric specification.

• Structure: The model defines a random choice function $\rho(a, A)$ for an alternative $a$ in a menu $A$, and an outside option $o$.
• Key Innovation: The probability of choosing the outside option is determined by a function $D(A)$, which represents the menu-dependent value of deferral.
  $$\rho(a, A) = \frac{u(a)}{\sum_{b \in A} u(b) + D(A)}$$
  Here, $u(a)$ is the menu-independent value of active alternatives, while $D(A)$ captures the "decision cost" or complexity.
• Axioms: The model satisfies the Active-Choice Luce Axiom (Independence of Irrelevant Alternatives holds only between active alternatives) and a new Desirability & Complexity Axiom (A3): An active choice is made with certainty only if the menu contains a single alternative ($|A|=1$). If $|A| > 1$, there is a positive probability of deferral.

B. The Power Logit Special Case

The paper focuses on a specific class of these models called the Power Logit, where decision difficulty is driven by the agent's "noisy resampling" of alternatives.

• Mechanism: The agent samples the menu multiple times (or evaluates alternatives with noise) and chooses an alternative only if it dominates others in all rounds of sampling.
• Formulation: Let $b_u(a)$ be the stimulus intensity of alternative $a$. The choice probability is:
  $$\rho(a, A) = \left( \frac{b_u(a)}{\sum_{b \in A} b_u(b)} \right)^p$$
  where $p \geq 1$ is the power parameter.
  • If $p=1$, the model reduces to the standard Logit (no deferral).
  • If $p > 1$, the model generates a deferral probability: $\rho(o, A) = 1 - \sum_{a \in A} \rho(a, A)$.
• Interpretation of $p$:
  • $p$ reflects the propensity to engage in multiple rounds of sampling.
  • In the Quadratic Logit ($p=2$), the agent effectively samples the menu twice. An alternative is chosen only if it is the "winner" in both independent sampling rounds.
  • The value of the outside option $D(A)$ is endogenously determined by the values of the active alternatives: $D(A) = (\sum b_u(b))^p - \sum b_u(b)^p$.

3. Key Contributions and Theoretical Results

The paper derives three empirically supported behavioral predictions that standard models cannot explain:

1. Similarity-Driven Deferral Effect:

   • As alternatives become more similar in value, the probability of deferral increases.
   • Result: In the Power Logit, the probability of deferral is maximized when all active alternatives have equal value. The maximum deferral probability is $1 - |A|^{1-p}$.

2. The "Roller-Coaster" Choice Overload Effect:

   • Menu expansion does not monotonically increase deferral.
   • Result: Deferral increases when new options are added unless the new options introduce a dominant alternative. If a new option is significantly better than existing ones, it reduces decision difficulty, thereby decreasing the probability of deferral despite the larger menu size. This explains non-monotonic "roller-coaster" patterns observed in consumer psychology.

3. Relative-Desirability Effect:

   • Holding the total value of a menu constant, increasing the difference in value between alternatives reduces deferral.
   • Result: Deferral is higher when options are equally desirable (hard to choose) and lower when one option is clearly superior to the other, even if the aggregate "appeal" of the menu remains the same.

4. Applications and Empirical Analysis

A. Theoretical Application: Duopolistic Competition

The author applies the Power Logit to a duopoly where firms compete on Quality ($q$) and Price ($p$).

• Setup: Consumers have value $u = q/p$. Firms maximize profit $\pi = (p-q) \times \text{Market Share}$.
• Equilibrium Findings:
  • As the decision difficulty parameter ($s$, equivalent to $p$) increases, firms increase quality to make their product stand out and reduce consumer hesitation.
  • However, this leads to lower profits for firms due to reduced profit margins and a lower overall share of consumers making an active purchase (many defer).
  • Consumer welfare decreases as hesitation increases, despite the higher quality offered, because the probability of making any purchase drops significantly.

B. Empirical Application: Film Choice Data

The model is estimated using data from Bhatia and Mullett (2016), where subjects chose between pairs of films or deferred.

• Methodology: The author compares four models:
  1. Standard Multinomial Logit (MNL) with a fixed outside option.
  2. MNL with a random outside option.
  3. Quadratic Logit ($p=2$).
  4. General Power Logit (estimating $p$).
• Estimation: The Power Logit is formulated as a discrete-choice model where the likelihood function accounts for the endogenous deferral probability derived from the power transformation of the utility.
• Results:
  • The Power Logit and Quadratic Logit generally provide a better fit (higher log-likelihood, lower AIC/BIC in many cases) and higher prediction accuracy (41.6% correct predictions) compared to baseline models.
  • Parameter Insights: The estimated power parameter $p$ averages around 1.51, confirming the presence of decision conflict.
  • Correlation: There is a negative correlation between the estimated power parameter $p$ and the absolute difference in film ratings. When ratings are similar (hard to choose), $p$ is higher (more deferral), validating the theory that deferral is driven by similarity-induced difficulty.

5. Significance and Conclusion

• Theoretical Advancement: The paper provides the first random-choice model where the outside option is feasible, has an endogenous, menu-dependent value, and is driven specifically by decision conflict rather than just the unattractiveness of options.
• Policy and Business Implications:
  • Choice Architecture: It suggests that reducing choice overload isn't just about limiting the number of options, but about creating clear dominance (e.g., a "best buy" label) to reduce hesitation.
  • Market Strategy: Firms should invest in quality differentiation to reduce consumer hesitation, even if it lowers margins, to prevent customers from opting out entirely.
• Empirical Utility: The framework offers a robust tool for econometricians to distinguish between deferral caused by "bad options" (standard logit) and deferral caused by "hard choices" (power logit), leading to more accurate welfare and demand estimates.

In summary, Gerasimou successfully bridges behavioral psychology (decision conflict) and econometric modeling (logit), offering a tractable framework that explains complex choice-deferral phenomena and provides actionable insights for market design and policy.