Experimental Design Details
In the experiment, we want to establish whether the revenue ranking of different auction formats (second-price vs. Japanese English auctions) predicted by standard theory (Milgrom & Weber, 1982) hold in a laboratory experiment. In the theory, the additional revealed information embedded in drop-out prices of a Japanese English auction is expected to increase revenue. However, an alternative hypothesis is that the information revealed might help bidders overcome strategic misconceptions, such as the winner's curse. If this is the case, revenue may decrease when more information is revealed. In the experiment we want to test these competing hypothesis.
In the oral outcry auction, we mimic auction formats commonly used in practice, which might also allow for some information revelation. Additionally, in this treatment we do not restrict bidders to incremental bidding. We can thus study the effect of allowing for non-incremental bidding in auctions.
Our primary measures in this experiment are the revenues of different auction formats and bidding strategies, as measured by dropout prices in second-price and Japanese English auctions, or bids submitted in the oral outcry auctions.
Our goal is to distinguish between the rational Nash benchmark (Milgrom & Weber, 1982), some classical behavioral models, such as incorporating winner's cursed bidders, and four potential naïve models, where bidders use available information in different ways.
The naïve models we will test are:
- Bidding equal to signal
- Bidding the expected value conditional on own signal and unconditional on winning (Goeree and Offerman, 2003)
- Bidding according to the signal-averaging rule (Levin et al. 1996)
- Naïve signal-averaging rule: Bidding average signals while assuming all other bidders bid their signal
We will focus on individual bidding strategies. We will classify bidders into the aforementioned models, based on the distance of the observed bidding strategies from the predictions of different models.
We will also compare models based on the extent to which they correctly predict the revenue ranking across auction formats. Revenue differences across formats will be assessed with the Mann–Whitney U test.
In addition, we are interested in potential learning effects, and perform the above analysis separately for the first and second half of the periods.
We will classify bankrupted bidders as bidders whose average earnings across all rounds, including their starting capital, is below zero. Then, we will discard all periods in which a bankrupted bidder won the auction. We will present the analysis both with and without the discarded data.