Summary: You randomize here?/ You will randomize there, too./ It’s a stable trait.

Summary: People use either AIC/BIC or cross-validation to pick which behavioral model fits data best. But how often do they actually pick the "right" model? With cross validation it can fail pretty badly! AIC/BIC perform better, but are too infelxible.

Summary: If you ask someone their favorite things from k menus, you're eliciting only partial information about their entire preference. However, it is incentive compatible if you just pay one randomly. What other kinds of partial information could you elicit? Depends, but pretty much that's it.

Summary: In our other paper on experiment incentives we model gambles as Savage acts, which are very general. Here we repeat the exercise when gambles are viewed as objective lotteries. In practice the conclusions are the same, but the characterization is more beautiful here.

Summary: A risk averse agent will misreport their beliefs in a proper scoring rule. We use duality techniques to characterize what beliefs they will report. Useful if the indirect utility function is easier to work with. We also show how to bound the degree of misreport with CARA preferences.

Summary: In some markets where pollution permits are auctioned off, a few players get consigned an initial allocation of permits and receive the revenue from their sale. How does this alter the auction outcome? We test this setting in the lab.

Summary: If we assume only monotonicity (dominated gambles are never chosen), paying for one random period (the RPS mechanism) is typically the only incentive compatible way to pay subjects. Paying every period is similarly justified if we assume a 'no complementarities at the top' (NCaT) condition. Which should you use? That's your decision.

Summary: In a multiple price list experiment (eg, Holt-Laury), we find that paying one randomly is not incentive compatible when all decisions are shown in a single table or list. When we show each decision on a separate screen, then it becomes incentive compatible. Yeah, subjects exhibit more multiple switches, but those are their true preferences!

Summary: Nobody knows this, but Groves & Ledyard's (1977) mechanism is actually just a VCG mechanism where people with non-quadratic preferences announce quadratic approximations. We generalize this procedure to construct some really messy-looking mechanisms.

Summary: We test cross-game stability of the Level-k/CogHi model. Ann's estimated level in one set of games fails to predict her level in another. In fact, Ann being a higher level than Bob in the first set of games doesn't predict who's higher in the other.

Summary: Let C be the set of preference profiles that admit a Condorcet winner. Any social choice correspondence that picks the Condorcet winner and whose domain is a subset of C can be implemented. But if you allow for more preferences, it cannot be implemented.

Summary: Ann and Bob play a finitely-repeated prisoners dilemma (FRPD). Ann can see how Bob played a previous FRPD against Charlie. If Bob was good to Charlie, then Ann trusts Bob more than she would've if she hadn't seen his history. But if Bob was a jerk to Charlie, Ann doesn't trust him less! Standard reputation models can't explain this.

Summary: Our experiments suggest that, when perceiving inflation, people tend to overweight those goods whose prices they see frequently. For example, if you see gas prices shooting up, you think the entire economy is inflating faster than it really is.

Summary: We characterize all mechanisms that implement Walrasian or Lindahl allocations in Nash equilibrium. Then we prove that they can't be dynamically stable if their message space is one-dimensional, but can be stable if messages are two-dimensional. Example mechanisms are given.

Summary: For normal distributions, the posterior mean always lies between the prior mean and the signal. That's not true in general. We characterize distributions for which that is true. Symmetry and unimodality of the error distribution are important.

Summary: Fix any bounded prior. We can construct a signal distribution and two possible signals where the higher signal is actually worse news, in terms of stochastic dominance of the posterior. (A failure of MLRP says this must be true for some prior; we say here it can be true for any.)

Summary: A prediction market works great if you have lots of traders to help equilibrate prices. When you don't, we find that incentivized, iterated polls perform better.

Summary: Suppose you use an incentive compatible mechanism to determine an ideal public goods allocation. But then some people don't mail in their check to the social planner. If you can't punish them, then no allocation better than the endowment is achievable when the economy grows large.

Summary: First, we distinguish between overplacement, overestimation, and overprecision. Then we show negative correlation between overplacement and overestimation as task difficulty varies, and rationalize this observation via a simple Bayesian argument.

Summary: I show how rational cooperation between groups can be sustained when sufficient stereotyping (modeled as a belief in type correlation among group members) is present. Behavior in experimental labor markets ('gift exchange games') is consistent with this theory.

Summary: Using simple theory and lab experiments, we show that the current method of funding space science projects within NASA is inferior to a proposed alternative with a cost sharing feature. (Thronson, Ulrich & Varsi are/were NASA employees.)

Summary: We know dynamic stability (supermodularity?) is important for public goods mechanisms. Here, I test 5 mechanisms, and provide a simple best-reply dynamic that predicts play (and convergence) reasonably well. (This was my job market paper.)

Summary: Non-equilibrium behavior in the 4-player `contestants row' game on the Price is Right is replicated in the lab. But when the game is made simpler (requiring less backwards induction), players play equilibrium more frequently.

Summary: This is a 6-page version of "The Trouble With Overconfidence", Psychological Review 115(2): 502–517. April 2008.

Summary: I agree with Rothkopf's shortcomings of the VCG mechanism, but urge computer scientists to look beyond dominant strategy implementation.

Summary: Take a standard G.E. model. As the economy grows large, the Pareto optimal public good level is infinitely larger than the private good levels. This seems unrealistic. We fix this by assuming attention constraints limit agents' consumption of public goods.

Summary: This was supposed to be a test of Yoon's (2015) prediction that truthful-but-noisy ratings can actually increase certain types of information cascades. But the crazy subjects just loved to purchase regardless of the history! Appears to be some sort of activity bias going on.

Summary: In belief-based learning models, agents form predictions based on past play and then best respond to them. To uncover the model players actually use, I either track their calculator usage or have them make incentivized predictions as they play.

Summary: Pratt gave a risk premium for gambles in which a random variable gets added to your wealth. We give several notions of a risk premium for gambles in which your wealth gets multiplied by a random variable.

Summary: This is a working paper version of "The Trouble With Overconfidence", writen for an economics audience and with more focus on the theory. This paper won the 2007 Roman Weil prize at CMU's Tepper School of Business.

Summary: In 1999 I was an undergraduate student in Tim Cason's Experimental Economics (ECON 690) course at Purdue. I ran a pilot experiment testing behavior in Final Jeapordy against the predictions of Metrick (1995, AER).

Summary: I wrote this while studying for qualifier exams in graduate school. It's incomplete and messy. And I failed the exam.

Summary: This is a master record of all BibTeX entries I've entered over the years. Last updated: 4/27/17 10:50:08pm.