It’s job market season again, and I’m always curious to see what type of work by the Young Turks is popular in any given year. The present paper, by Berkeley’s Mitch Hoffman, strikes me as a nice example in the genre “Behavioral Second Best,” a genre best exemplified by Ben Handel’s medical insurance paper, which I see is now R&R at AER (and which surely will be accepted in the end, right?). The classic theory of the Second Best says that, in the presence of two market distortions, fixing one may decrease social welfare. For instance, reducing the tax burden of a firm that pollutes leads to less distortionary investment across market sectors, but more investment into a sector with large negative externalities. The Behavioral Second Best (BSB) is similar: the presence of a market distortion plus a behavioral bias can mean that correcting one of these decreases welfare. In Ben’s medical insurance paper, the reluctance of workers to switch their health plans over time had the side benefit of letting the insurance market offer multiple price/feature bundles without the usual unraveling of a separating equilibrium.
Unsurprisingly, this is a popular genre, so what does it take to write a good paper along these lines? I think there’s no way to do it without a first-rate dataset plus some empirical analysis that convincingly shows an important, real-world effect. Merely pointing out a theoretical curiosity related to BSB is simply uninteresting at this stage. So surely Hoffman’s bevy of top-flight flyouts is related to some careful data work. The question he addresses is an old one: why would firms pay for general training when workers can just leave the firm once they’ve been trained, or else why won’t they ask for higher pay after being trained as in a classic hold-up problem? Workers can’t always pay for training themselves due to credit constraints. Labor market frictions surely explain some of the puzzle – it is not always so easy to take one’s talents to a new employer – but it is tough to imagine these frictions accounting for training valued in the tens of thousands, which is not unusual in many industries. And certainly this is something firms worry about: witness the recent scandal where Apple, Google and other tech companies had a secret do-not-compete-for-labor pact based precisely on the worry that workers they train will flee after training is finished. (As an aside, how is it that some of these executives, particularly Eric Schmidt, are not facing criminal charges here? The quotes documented look like bald-faced admission of criminal activity to me!)
Hoffman proposes a reason why workers may be trainable in piece-rate industries even if they could leave whenever they wished: overconfidence. As a trucker, say, I am paid by how many miles I drive. The trucking firm agrees to hire and train me, but if I leave before date X, I have to pay back the cost of my training. After my training, I am overconfident about the number of miles I will be able to drive. Overconfidence that is not attenuated (or only slowly attenuated) by learning will make me less likely to quit, since the (perceived) value of keeping my trucking job is higher, at any given piece-rate, compared to my best outside option. This makes firms more willing to train me in the first place. In theory, this might improve outcomes for everyone: the firm gets more profit from better-trained workers, and the workers perhaps can extract a higher wage. Teaching workers to be less confident might make things worse.
Hoffman has a fantastic data set of payroll data from a large trucking company, indicating actual miles driven per week, plus the relevant training contract of each employee, plus weekly subjective reports on how many miles the employee expected to drive. Those reports were secret as far as the employer was concerned, so there is little reason for workers to lie, but Hoffman also runs a side experiment where he pays workers small bonuses for guessing their miles driven correctly; such incentives do not change in a significant way the reported expectations. The average worker is terribly overconfident, and his overconfidence attenuates as he gains experience, but only slowly. Overconfidence is linked to lower quit probabilities at any stage in the training contract, as you would expect Running a counterfactual structural model, Hoffman examines how various contracts will affect quit probabilities, and therefore firm and worker welfare (though for all the usual reasons, you should be skeptical of welfare estimates involving behaviorally biased agents). Eliminating overconfidence massively harms the trucking firm’s profits, as quit rates will increase and training becomes less viable. A government ban on penalties for quitting after you train may actually improve worker welfare for BSB reasons: though these penalties allow for more cases where worker training is possible, overconfidence also means workers are willing to accept huge penalties for quitting in exchange for tiny increases in post-training wages.
My only real quibble here is how the outside option is defined, theoretically. The usual worry with worker training is not hold-up, but the increase in the outside option. The model here is really quite specific to firm-specific training in piece-rate industries. My prior is that such industries are a rather small part of all industries where worker training occurs: examples in the paper involving MBA education surely don’t qualify. That said, I imagine you can tell a Behavioral Second Best story in the more general case, as long as the overconfidence retains some level of firm-specific nature, or as long as there is some divergence between the impact of the overconfidence of MP within the firm and outside the firm. There are many ways to do this: labor search frictions are one.
http://econgrads.berkeley.edu/hoffman/files/2011/12/Training-Contracts_JMP.pdf (November 2011 working paper)