Human Speed-Accuracy Tradeoffs in Search

When foraging for information, users face a tradeoff between the accuracy and value of the acquired information and the time spent collecting it, a problem which also surfaces when seeking answers to a question posed to a large community. We empirically study how people behave when facing these conflicting objectives using data from Yahoo Answers, a community driven question-and-answer site. We first study how users behave when trying to maximize the amount of acquired information while minimizing the waiting time. We find that users are willing to wait longer for an additional answer if they have received a small number of answers. We then assume that users make a sequence of decisions, deciding to wait for an additional answer as long as the quality of the current answer exceeds some threshold. The resulting probability distribution for the number of answers that a question gets is an inverse Gaussian, a fact that is validated by our data.

💡 Research Summary

The paper investigates how users of a large‑scale question‑and‑answer community balance the competing objectives of acquiring high‑quality information and minimizing the time spent waiting for that information. Using an extensive dataset from Yahoo Answers (over two million questions posted between 2005 and 2009), the authors first examine empirical patterns in users’ “information‑foraging” behavior. They find a clear “scarcity effect”: when a question has received only a few answers, askers are willing to wait considerably longer for an additional response; as the number of existing answers grows, the marginal willingness to wait drops sharply. This pattern mirrors the classic diminishing marginal utility of information and suggests that users implicitly compare the expected benefit of an extra answer against the time cost of waiting.

To formalize this intuition, the authors propose a sequential decision‑making model. Each incoming answer i is assigned a quality score q_i based on observable proxies such as user rating, answer length, and keyword relevance. The asker maintains the highest quality observed so far (q_max) and continues to wait for another answer as long as q_max exceeds a pre‑determined threshold θ. The threshold is assumed to be drawn from a multivariate normal distribution that captures variations across topics, difficulty levels, and asker characteristics. Under these assumptions, the stopping time—the number of answers a question ultimately receives—follows an inverse Gaussian (Wald) distribution characterized by a mean μ and a shape parameter λ. The inverse Gaussian is asymmetric with a heavy right tail, reflecting the empirical fact that a small fraction of questions endure long waiting periods while most terminate quickly.

The authors validate the model by fitting μ and λ to the observed distribution of answer counts and then comparing the theoretical inverse Gaussian density to the empirical histogram. Goodness‑of‑fit tests (Kolmogorov‑Smirnov) and QQ‑plots show no statistically significant deviation, confirming that the inverse Gaussian accurately captures the shape of the data, including the long tail associated with “high‑cost waiting” for scarce answers.

From a practical standpoint, the findings have two immediate implications. First, platform designers can improve user satisfaction by accelerating the delivery of early answers—e.g., by incentivizing initial responders or routing new questions to experts—thereby reducing the overall waiting time and preventing the steep decline in willingness to wait that occurs after a few answers are posted. Second, the inverse Gaussian model can be embedded into automated moderation tools: if a question’s current answer quality remains below the estimated threshold after a certain elapsed time, the system could suggest closing the question, prompting the asker to rephrase, or sending targeted reminders to potential answerers.

The paper also acknowledges limitations. The quality metric is a simplified aggregate of observable signals and may not fully capture semantic usefulness. Moreover, the model treats the waiting threshold as static across a user’s session, ignoring individual differences in urgency or patience. Future work could extend the framework by employing Bayesian updating to personalize θ for each asker, or by integrating deep‑learning‑based text embeddings to obtain richer quality estimates. Cross‑platform validation (e.g., on Stack Overflow or Reddit) would further test the generality of the inverse Gaussian description.

In summary, the study provides a rigorous empirical and theoretical treatment of the speed‑accuracy tradeoff in online information seeking. By demonstrating that the distribution of answer counts conforms to an inverse Gaussian law derived from a simple sequential stopping rule, the authors bridge behavioral observations with a tractable statistical model. This contribution offers both a deeper understanding of human foraging behavior in digital environments and actionable insights for designing more efficient, user‑friendly Q&A systems.