Last updated on March 3, 2020
I will be teaching the last quarter (i.e., half-semester) of our first-year graduate econometrics sequence this year. This of course means that I will be teaching causal inference.
To do so, I am using the second edition of Morgan and Winship’s wonderful Counterfactuals and Causal Inference, which features a wonderful discussion of the stable unit treatment value assumption (SUTVA).
Many people who were trained in econometrics prior to the Credibility Revolution are not familiar with the acronym SUTVA, and even the full name “stable unit treatment value assumption” can sound more confusing than not. In economics, people sometimes refer to it as the “no-macro-effect” or “partial equilibrium” assumption.
What SUTVA says, basically, is that for a treatment D and and outcome Y, the value of D for individual i in time period t should should not have any effect on the value of Y for individuals who are not i in any time period t or for individual i in any time period that is not t.
Put more simply: That individual i gets treated in period t should have no effect on any other individual’s outcome at any given time, nor should it have any effect on that individual’s outcome in other time periods.
Put yet more simply: There should not be any spillovers.
The SUTVA can be extremely difficult to satisfy, and as with many other assumptions, though it might be feasible to rule out certain types of SUTVA violations, it may be difficult if not impossible to rule them all out.
In Bellemare and Nguyen (2018), for instance, we were interested in the relationship between farmers markets and food-borne illness in a given state in a given year. In an attempt to rule out contemporaneous spillovers from neighboring states, we controlled for the average number of farmers markets in neighboring states, but this did not help with any potential spillovers from year to year, or across states from year to year, no matter how unlikely they are.
In preparing my other graduate class–microeconomics of agricultural development–this semester I read an article which does a wonderful job of testing for SUTVA violations. In their 2019 article investigating the puzzle of “sell-low, buy-high” behavior (i.e., the phenomenon whereby smallholders sell their crops at low prices around harvest time, only to buy the same commodities later in the year at high prices), Burke et al. test for SUTVA violations by randomly varying the intensity of a randomly assigned treatment.
This double randomization allows first to estimate the impact of their treatment, which consists of a loan at harvest time, and then to estimate the impact of treatment spillovers. The idea behind the latter is that if SUTVA holds, the estimate of the treatment effect should be invariant to how many people receive a loan within a given community.
Burke et al.’s findings are telling: When few people are treated in a given community, receiving a loan at harvest reduces the extent of “sell-low, buy-high” behavior and increases the welfare of smallholders via an increased use of storage. But when many people are treated in a given community, smallholders are not significantly better off, since the use of storage is not more profitable.
As I said above, testing whether SUTVA holds can be extremely difficult, if not impossible. Burke et al. randomly varied treatment intensity to get at whether the SUTVA held, but not everyone can do so. Testing whether SUTVA holds can be particularly difficult with observational data. But this need not doom one’s findings. One way out of this is to admit that one cannot test for SUTVA, and that one’s treatment effect estimate should hold for “similar situations,” which ultimately limits external validity.
In Burke et al.’s case, had they not varied treatment intensity and only offered loans to small proportions of smallholders in each community, this would have meant saying that the treatment effect should hold in other situations where only a small proportion of smallholders are treated in each community.