Technical Summary: Reputational Conservatism in Expert Advice

1. Problem Statement

The paper investigates the behavior of experts who possess private information about the success of a risky action versus a safe action, operating under "career concerns." The central empirical puzzle addressed is the observation that senior, high-status experts often propose risky options less frequently than junior experts, yet achieve higher success rates when they do take risks.

The model formalizes a one-shot advisory problem where:

• The Expert: Observes a continuous private signal regarding the state of the world (success/failure of a risky option) and recommends either a risky ($R$) or safe ($S$) action. The expert's utility depends on their posterior reputation (a belief about their ability) and potentially a success-contingent bonus.
• The Principal: Always implements the safe option. However, the probability of implementing the risky option, denoted $\lambda(\rho)$, is a weakly increasing function of the expert's current reputation $\rho$. This captures institutional frictions such as committee approvals, client uptake, or regulatory gatekeeping where high-status advice is more likely to be acted upon.
• Information Flow: If $R$ is implemented, the outcome is observed, updating the expert's reputation. If $S$ is implemented (or $R$ is vetoed), no new information is generated, and reputation remains unchanged.

2. Methodology

The authors employ a game-theoretic framework with Bayesian updating and continuous signals.

• Signal Structure: The expert observes a signal $x$ with a Monotone Likelihood Ratio Property (MLRP). The posterior probability of success, $p(x)$, is strictly increasing in $x$.
• Equilibrium Concept: The analysis focuses on cutoff equilibria, where the expert recommends $R$ if and only if their signal exceeds a threshold $c(\rho; b)$ (or equivalently, if the posterior success probability $p(x)$ exceeds a threshold $\bar{p}(\rho; b)$).
• Reputational Payoffs: The expert maximizes a reduced-form utility $W(\rho^+)$, where $W' > 0$. The model distinguishes between the reputational gain from a successful risky implementation ($\Delta_s(\rho) > 0$) and the reputational loss from a failure ($\Delta_f(\rho) < 0$).
• Key Condition: The paper introduces a Relative-Diagnosticity (RD) condition. This condition posits that as reputation increases, the marginal reputational penalty from a visible failure grows (or declines less) relative to the marginal gain from a visible success. Formally, at the cutoff, $p \Delta'_s(\rho) + (1-p)\Delta'_f(\rho) \leq 0$.

3. Key Results

A. Existence and Uniqueness of Cutoff Equilibrium

The paper establishes that for any given reputation $\rho$ and bonus $b$, there exists a unique cutoff threshold $\bar{p}(\rho; b)$ in the posterior success probability space.

• The expert recommends the risky option if $p(x) \geq \bar{p}(\rho; b)$.
• This cutoff is continuous with respect to reputation and the bonus.
• The result holds robustly for any signal distribution satisfying MLRP.

B. Reputational Conservatism

The central theoretical finding is that under the Relative-Diagnosticity condition, the experimentation threshold is weakly increasing in reputation.

• Mechanism: High-reputation experts are more likely to have their risky advice implemented ($\lambda(\rho)$ is higher). Consequently, failures are more visible and damaging to their reputation. Conversely, successes are not proportionally more rewarding in terms of reputation gains. To "break even" on the increased reputational risk, a high-reputation expert requires a stronger signal (higher $p$) to recommend risk.
• Outcome: Senior experts propose risky options less frequently but have a higher conditional success rate when they do, matching the empirical pattern described in the introduction.

C. Comparative Statics

The paper derives clean, closed-form comparative statics:

1. Signal Informativeness: Higher signal precision (informativeness) lowers the signal cutoff required for risk, increasing the frequency of risky recommendations.
2. Prior Success Probability: A higher prior probability of success lowers the signal cutoff, increasing risk-taking.
3. Career Concern Intensity:
   • If a success-only bonus exists ($b > 0$), increasing the intensity of career concerns raises the cutoff (more conservatism).
   • If there is no success-only bonus ($b = 0$), the intensity of career concerns alone does not shift the cutoff; the ratio of reputational gains to losses remains constant.

D. Institutional Design Levers

The authors identify two distinct levers a principal can use to tune expert behavior:

1. Success-Contingent Bonuses ($b$): There is a one-to-one mapping between the bonus size and the equilibrium experimentation rate (frequency of risky recommendations) at a given reputation. This allows principals to target specific exploration rates without redesigning the entire contract.
2. Gatekeeping ($\lambda(\rho)$): Adjusting the implementation probability (e.g., via committee approval or monitoring) scales the realized experimentation and the speed of learning but does not change the expert's recommendation threshold. Stricter gatekeeping reduces the frequency of implemented risks and slows reputational learning without altering the advice itself.

4. Significance and Contributions

The paper claims to offer a simple, tractable explanation for the "playing it safe at the top" phenomenon observed in organizational settings. Its contributions are threefold:

1. Mechanism Identification: It isolates a specific mechanism where the visibility of failure (driven by higher implementation probabilities for high-status experts) combined with asymmetric reputational updates drives conservatism. This distinguishes the model from "political correctness" models where advisors shade messages to avoid bias, or standard cheap-talk models where communication is coarse.
2. Policy Tools: It provides actionable design tools for organizations. By decoupling the proposal of risk (controlled by bonuses) from the implementation of risk (controlled by gatekeeping), institutions can balance prudence and learning across experts of different seniority.
3. Robustness and Extensions: The paper discusses how these static results interact with dynamic career concerns (where long-term incentives might encourage early experimentation) and endogenous implementation rules, suggesting that the conservative selection effect is robust and potentially strengthened in more complex settings.

The authors emphasize that their framework delivers "clean comparative statics" and "implementable levers," allowing for direct empirical testing (e.g., observing changes in proposal rates and hit rates following policy changes in bonus structures or gatekeeping intensity).