Experiments: Always LATE
07 Oct 2016
This is the second of a series of posts on Angus Deaton and Nancy Cartwright’s working paper, “Understanding and Misunderstanding Randomized Controlled Trials.” The first post is available here.
Deaton and Cartwright’s paper makes an excellent point about the distinction between an average treatment effect (ATE) and the local average treatment effect (LATE). While commonly distinguished in instrumental variables analyses of observational data, the distinction is not regularly made in experiments (except those where non-compliance is rampant and easily observable). They raise this point on (p.24) in a discussion of blinding:
Blinding is rarely possible in economics or social science trials, and this is one of the major differences from most (although not all) RCTs in medicine, where blinding is standard, both for those receiving the treatment and those administering it. Indeed, the ability to blind has been one of the key arguments in favor of randomization, from Bradford-Hill in the 1950s, see Chalmers (2003), to welfare trials today, Gueron and Rolston (2013). Consider first the blinding of subjects. Subjects in social RCTs usually know whether they are receiving the treatment or not and so can react to their assignment in ways that can affect the outcome other than through the operation of the treatment; in econometric language, this is akin to a violation of exclusion restrictions, or a failure of exogeneity. In terms of (1), there is a pathway from the treatment assignment to another unobserved cause, which will result in a biased ATE. This is not to argue in favor of instrumental variables over RCTs, or vice versa, but simply to note that, without blinding, RCTs do not automatically solve the selection problem any more than IV estimation automatically solves the selection problem. In both cases, the exogeneity (exclusion restriction) argument needs to be explicitly made and justified. Yet the literature in economics gives great attention to the validity of exclusion restrictions in IV estimation, while tending to shrug off the essentially identical problems with lack of blinding in RCTs.
This is a really important point and one that we tend to gloss over. In field experiments and in medical trials where compliance is often directly observable (answer the phone; take the pill; etc.), it is easy to estimate the LATE and thus obtain an estimate of the average treatment effect local to the population of those who comply with treatment assignment. In many other contexts, however, compliance is undefined or unobservable and thus we estimate ATEs but are in practice estimating “intention to treat” (ITT) effects (in essence, the LATE averaged with apparent zero effects for those who are unwilling to accept treatment assignment).
What we forget in this exercise is that there is always noncompliance, it just cannot always be seen. Consider for example, a “framing” experiment in which I randomly assign participants to read either a “free speech” message about a hate rally or a “public safety” message about a hate rally. There are two ways to look at this design.
The first interpretation is that this is a political communication experiment. In this case, we are estimating an ATE. One group has the free speech treatment and the other has the public safety treatment. Everyone is treated. The estimate derived from comparing mean outcomes in the two groups is the (S)ATE. There is no compliance concern per se because the treatment we care about is, in essence, a policy: we are studying the effect of making something available to some and withholding it from others.
The second interpretation, however, is that this is a political psychology experiment. In this case, we are estimating an ITT or - depending on how we conceptualize our treatment - a LATE. Let me explain. From this perspective, we are not so much interested in the difference in response to exposure but rather we are interested in differences in outcomes between those that think about the hate rally in different ways. Thus the effect we want to know is about the frames in thought, not the frames in communication. As such, our treatment is actually just an instrument. We think it will change how people think. When we compare the two groups then, we are estimating an ITT where we as the researchers are blind to compliance (i.e., thinking the way we induced the subjects to think). If we can conceptualize and measure frames in thought, then we can instead estimate a LATE using treatment assignment as an instrument for reasoning
That, however, assumes that the exclusion restriction is not violated (i.e., that there is no other path by which our treatment affects the outcome but through the specific reasoning process we are interested in). An observational IV paper would spend considerable time discussing the plausibility of the exclusion restriction. Deaton and Cartwright’s point above is that experimental papers rarely do the same. (Though see my research trying to tease out what part of “framing” treatments are actually changing opinions.) I agree we should do more to defend the plausibility of our treatments as instruments, when we are using them as such.
In short, for this second lens on the experiment, we are only able to estimate an ITT or a LATE, never an ATE. To achieve that we would have to envision a way to actually set the value of frames in thought perfectly (in the way that we can set the value of media content perfectly in the political communication interpretation of the study). Because we randomize, we are inclined to think that the thing we are randomizing is the independent variable that we are care about. But often we randomize something because it can be manipulated and ignore the causal process that we are actually trying to instrument for. We pretend when we can estimate the ATE we care about but find ourselves always LATE.
Except where noted, this website is licensed under a Creative Commons Attribution 4.0 International License.