Effort without Evidence

Technical Summary: Effort without Evidence

Problem and Motivation
This paper investigates the feedback loop between strategic communication and evidence production in a sequence of short-lived decision-makers. Standard models of communication and learning typically separate information transmission from information production: agents receive signals and then decide what to disclose. However, in many contexts (e.g., research, production, organizational routines), evidence exists only because an action was taken. Hidden effort governs public outcomes, and a sender's message to a successor influences the successor's action, thereby determining what evidence can be generated in the future.

The core conflict arises from "motivational bias." The sender values the successor's success but underweights the successor's effort cost (parameterized by $\lambda \in (0, 1]$). Consequently, the sender has an incentive to induce actions that are physically costly, even if those actions are not the most informative. The paper asks: How does this bias affect the credibility of causal narratives, the production of evidence, and long-run learning?

Methodology and Framework
The model features a discrete-time sequence of agents ($t=0, 1, \dots$) facing a persistent binary state $\theta \in \{0, 1\}$.

• Timing: Each agent observes a public outcome archive and a message from the predecessor, forms a posterior, privately observes an effort cost $k$, chooses a hidden action $a$, and observes a public outcome $y$. The agent then sends a costless message $m$ (causal assessment) to the next generation.
• Preferences: The agent's utility includes current material payoff and a weighted component of the successor's payoff: $u(y, a, k) + \alpha [y' - \lambda c(a', k')]$. The parameter $\lambda$ captures the degree of motivational bias (underweighting of effort cost).
• Equilibrium Concept: The analysis focuses on Perfect Bayesian Equilibria (PBE). To refine the set of equilibria and address the issue of multiple causal narratives, the paper employs a "credible action-changing neologism" criterion (based on Farrell, 1993). An equilibrium is immune if no group of sender types can credibly propose a new label that induces a strictly preferred action profile for themselves without being imitated by excluded types.
• Key Definitions:
  • Identifying Action ($a_I$): An action profile that generates state-dependent outcomes (informational).
  • Nonidentifying Action ($a_N$): An action profile that generates state-independent outcomes (uninformative).
  • Motivational Collapse: A situation where common sender rankings force every message into a nonidentifying action region, making identification impossible.

Key Results

1. The Cost-Bias Identity and Motivational Collapse
The paper establishes a fundamental identity (Lemma 2) linking the sender's incentive to the receiver's incentive. Let $C$ be the physical cost difference between the identifying and nonidentifying actions. The sender's differential utility is $G_\lambda(s) = D(s) + (1-\lambda)C$, where $D(s)$ is the receiver's material differential.

• Result: Stronger motivation (lower $\lambda$) shifts the sender's preference toward the physically costlier action.
• Collapse: If the identifying action is costlier ($C > 0$), strong motivation makes every sender want to encourage it. If the nonidentifying action is costlier ($C < 0$), strong motivation makes every sender want to suppress the identifying action.
• Bayes-Unavoidability: If the mean posterior lies in a region where the nonidentifying action is robustly optimal for the receiver, and all senders share a strict ranking over actions, Bayes plausibility forces every message to induce the nonidentifying action. Once this "collapse" occurs, strengthening motivation further cannot restore the identifying action.

2. Dynamic Evidence Traps and Absorption
The static credibility failure leads to dynamic traps.

• Absorption: If communication induces a nonidentifying action profile, the public outcome becomes state-independent. No new state information is produced. The posterior support remains within a bounded set, and the common ranking argument repeats indefinitely.
• Result: In every PBE, the system enters an absorbing state where learning fails. If the initial information is insufficient to distinguish the states, complete learning becomes impossible with positive probability.
• Near-Traps: Even if the nonidentifying region is only weakly informative (not perfectly state-independent), the duration of the nonidentifying phase can be arbitrarily long, diverging as the signal quality approaches zero.

3. Application I: Inactivity (The "Do Too Little" Trap)
In this application, effort is binary (experiment vs. no experiment).

• Mechanism: A low-cost agent experiments; a high-cost agent does not. A public failure is "confounded": it could be a failed experiment (low return) or a lack of experimentation.
• Moderate Motivation: Senders with different posteriors want different actions (pessimists want inactivity; optimists want experimentation). This disagreement preserves causal separation. Pessimists credibly signal that the technology is bad, while optimists signal it is good. Learning occurs almost surely.
• Strong Motivation: Even pessimistic senders want to encourage experimentation to help the successor. All senders imitate the "encouraging" narrative. Rational skepticism renders all such narratives incredible.
• Outcome: After a confounded failure, the system shuts down permanently. No experimentation occurs, no new evidence is generated, and the true state (if it is the productive one) is never learned.

4. Application II: Excessive Effort (The "Do Too Much" Trap)
In this application, success requires crossing an unknown effort threshold. Effort above a second, higher threshold yields diminishing returns but is technologically saturated.

• Mechanism: "Restraint as discovery." A high-cost agent who succeeds with lower effort is the informative pioneer because her success reveals the threshold is low. A low-cost agent choosing maximal effort produces state-independent outcomes.
• Moderate Motivation: Senders correctly attribute outcomes. High-cost agents who succeed with low effort induce the receiver to choose low effort (identifying).
• Strong Motivation: All senders, regardless of their causal type, advocate for maximal effort (the physically costlier action).
• Outcome: After a bad pooled history, everyone exerts maximal effort. Outcomes become state-independent. The informative "restraint" experiment disappears. The society appears more industrious and successful (higher success rate per incumbent) but fails to learn that less effort would suffice, resulting in lower long-run welfare.

5. Certification and Policy Interventions
The paper analyzes whether verifying actions (certification) can repair these failures.

• Decomposition: A Kullback–Leibler decomposition shows that certification adds information only if it reveals hidden causal variation. If the system is already in a recursively nonidentifying action region (where actions are state-independent), certification adds no information.
• Inactivity Model: Certifying a failure (distinguishing "failed experiment" from "no experiment") restores the belief at which experimentation is optimal. Failure-only certification is the unique, least-costly policy to guarantee learning.
• Threshold Model: Certifying an intermediate-effort failure is genuine bad news; it correctly updates the belief to a point where maximal effort is chosen. Thus, even perfect attribution can select an uninformative continuation. Furthermore, anticipated success certification can crowd out the identifying action itself.
• Pricing Mechanisms:
  • Inactivity: Subsidize effort to move the receiver across the identifying boundary.
  • Threshold: Subsidize restraint (or tax maximal effort) to preserve the identifying margin.
  • Forced Variation: A public mandate to take the identifying action with some probability can generate information directly when communication fails.

Significance and Claims
The paper claims to unify two seemingly opposite failures of social learning:

1. Shutdown: Society stops learning because nobody works (due to the inability to credibly encourage experimentation).
2. Overexertion: Society stops learning because everyone works too hard (due to the inability to credibly encourage restraint).

The central organizing message is that motivation selects the costly action, not the informative action. The paper demonstrates that the "costly" action is not inherently informative; its informativeness depends on the technology. When motivational bias aligns all senders behind a single action (either effort or restraint), it collapses the informational content of communication, leading to absorbing states where no new evidence is produced.

The paper does not claim to provide a general theory of narratives but isolates a specific mechanism where strategic causal communication recursively selects the next experiment. It highlights that verification (certification) is not a panacea; its value depends entirely on whether it prevents entry into a nonidentifying region or restores the incentives to produce identifying evidence. The corrective policy depends on the direction of the bias: subsidizing effort when agents do too little, and subsidizing restraint when agents do too much.