Informational Content of Auction Prices

Technical Summary: Informational Content of Auction Prices

Problem Statement
This paper investigates the comparative informational content of auction prices in multi-unit, common-value auctions. Specifically, it compares the discriminatory price auction (where each of the $k$ winning bidders pays their own bid) against the uniform-price auction (where all $k$ winners pay the highest losing bid, i.e., the $(k+1)$-st highest bid).

The central question is: which auction format conveys more information about the true, unknown value $V$ of the asset to an outside observer (such as a future investor) who observes only the resulting prices? In a symmetric environment with monotone equilibria, the discriminatory auction reveals the highest signal ($X_{(1:n)}$) via the highest winning bid, while the uniform-price auction reveals the $(k+1)$-st highest signal ($X_{(k+1:n)}$) via the clearing price. The problem thus reduces to a statistical comparison: under what conditions is the highest order statistic of a sample more informative than the $(k+1)$-st highest order statistic?

Methodology
The authors employ a decision-theoretic framework to compare the "experiments" provided by these order statistics.

1. Lehmann Dominance: The paper rejects the Blackwell criterion as too strong for comparing order statistics. Instead, it utilizes the Lehmann (1988) criterion, which requires that every decision maker with single-crossing preferences (or interval dominance order utility) prefers one experiment over the other. This is equivalent to comparing the sensitivity of the cumulative distribution functions (CDFs) with respect to the underlying state $v$.
2. Model Structure: The analysis covers auctions of $k$ identical objects to $n$ bidders ($n > k$) with independent private signals $X_i$ drawn from a distribution $F(\cdot | V=v)$ satisfying the Monotone Likelihood Ratio Property (MLRP).
3. Analytical Approach: The authors derive conditions under which the highest order statistic $X_{(1:n)}$ Lehmann dominates $X_{(k+1:n)}$. They analyze three specific classes of signal distributions:
   • Scale Families: $X = V \times Z$ (multiplicative errors).
   • Location Families: $X = V + Z$ (additive errors).
   • General Families: Arbitrary distributions with support boundaries that may depend on $V$.

Key Contributions and Results

• Sufficient Conditions for Dominance:
  The paper establishes that $X_{(1:n)}$ Lehmann dominates $X_{(k+1:n)}$ if the ratio of objects to bidders ($k/n$) exceeds a threshold determined by the informativeness of the signals at the boundaries of their support.

  • In scale families, the condition is $\frac{k}{n} \ge 1 - \frac{\rho(1)}{\rho(0)}$, where $\rho(z)$ is the elasticity of the error distribution. This implies dominance occurs when high signals are highly informative (large $\rho(1)$) and low signals are not very informative (small $\rho(0)$).
  • In location families, the condition is $\frac{k}{n} \ge 1 - \frac{r(0)}{r(-\infty)}$, where $r(z)$ is the reverse hazard rate.
  • In general families, the condition involves the cumulative score function at the support endpoints: $\frac{k}{n} \ge 1 - \frac{S_v(\omega_v)}{S_v(\alpha_v)}$.

• Qualitative Necessity:
  The authors demonstrate that the sufficient conditions are "qualitatively necessary." For $X_{(1:n)}$ to dominate $X_{(k+1:n)}$, three features are required:

  1. Conclusive High Signals: There must be signals that definitively rule out lower values of $V$ (i.e., the likelihood ratio is unbounded at the upper support boundary). If high signals are not conclusive, the $(k+1)$-st order statistic may be preferred because it aggregates more evidence.
  2. Uninformative Low Signals: Low signals must not be very informative about low states. While the sufficient condition in scale families requires the lower support boundary to be constant (implying no conclusive low signals), the qualitative necessity result specifically highlights that low signals must lack strong informativeness (small elasticity or reverse hazard rate) rather than strictly requiring them to be inconclusive in all distributional contexts.
  3. Non-vanishing Ratio: The ratio $k/n$ must be bounded away from zero. If the number of objects is negligible relative to bidders, the highest order statistic does not dominate.

• Specific Findings:

  • In scale families, a lower-order statistic cannot Lehmann dominate the highest-order statistic.
  • For power distributions ($F(z)=z^\alpha$), the highest order statistic dominates for all $k$ and $n$.
  • For truncated normal, exponential, and Gumbel distributions, specific thresholds for $k/n$ are calculated (e.g., approx. 0.29 for truncated normal).

Significance and Applications
The paper claims significance in three primary areas:

1. Auction Theory: It provides a rigorous answer to which auction format is superior for information revelation to outside observers. The results suggest that discriminatory auctions are more informative than uniform-price auctions when the supply of objects is significant relative to demand ($k/n$ is high) and the signal structure allows for conclusive high values. This contrasts with literature on information aggregation (convergence to true value as $n \to \infty$), showing that an auction can be more informative without necessarily aggregating information in the limit, and vice versa.

2. Statistical Theory: The results extend beyond auctions to the general problem of comparing order statistics. The paper identifies conditions under which the maximum of a sample is a better estimator of a parameter than the median or other lower order statistics. This applies to contexts such as restaurant ratings (where the best review vs. the median review is more informative) or academic admissions (highest test score vs. median score).

3. Voting Theory: By mapping the informativeness of order statistics to voting rules, the authors apply their findings to jury decisions. They show that under the derived sufficient conditions (specifically the presence of conclusive signals indicating innocence), the unanimity rule can be welfare-superior to the majority rule. This challenges earlier findings (e.g., Feddersen and Pesendorfer, 1998) that suggested unanimity rules often fail to aggregate information, by highlighting the role of signal structure (conclusive vs. inconclusive signals) rather than just the number of voters.

The paper concludes that the ranking of auction formats and voting rules depends critically on the nature of the information structure (specifically the presence of conclusive signals and the relative uninformative nature of low signals) and the ratio of winners to participants, rather than solely on the number of participants.