Discussion about this post

User's avatar
cagnew's avatar

Hi Noah,

I have long enjoyed your blog . This is a topic near to my heart -- I wrote a thesis on the subject back in 2015.

I think you correctly summarize the upshot of Deaton and Nancy Cartwright's position. I would just like to clarify some terminology, that may make the critique easier to understand. No one, including Deaton and Cartwright as far as I'm aware, thinks that RCTs are bad tool. In fact, almost everyone thinks they are very good tools for one, specific, task: making causal inferences.

The critique is really focused on what you do with causal inferences -- sometimes known in the literature as so-called 'evidence-for-use.'

To see the problem, it's helpful to think what an ideal RCT tells you: An ideal RCT gives you an extremely strong evidence that the intervention (whatever form it takes) is the cause of the effect in the model population. What an RCT, however well designed, can never tell you is whether the same intervention will have the same effect in some other population. In order to jump from the inference in the model population to some other target population, you need to extrapolate. (I would emphasize, in passing, here that a target population is *always* distinct from the model population -- the same intervention may have different effects based solely on the time the intervention is administered!)

The difficulty is that an RCT -- by design -- does not explain *why* an intervention worked in the model population. All it tells you is that it did work. RCTs, to the extent we want them to do anything more than generate a true causal inference, have to be accompanied by some theory of mechanisms -- a theory that may need only be intuitive, as I understand Deaton to suggest. And this theory must explain why the *reason* the intervention had its effect in the model population can be expected to obtain with respect to some other population. In other words, to turn a causal inference from an RCT into evidence for, e.g., a policy's efficacy in some other (later, more widespread, geographically distinct, whatever) setting, you have to discharge the burden of showing the reason the effect occurred will hold in the target population.

To tie this back to reality, it is helpful to think about medicine -- aspirin, to use you example. Consider an RCT showing aspirin is effective in population A. We have reason to think that asprin will be effective in population B (say, all mankind more or less) because we know that the mechanism by which aspirin has its effect in population A will be unaffected by any differences in population B. People are the same in the relevant respects, across space and time. The casual pathway is essentially invariant. (Causal pathway is a term in medicine that development economists should pay greater attention to, in my view.) This assumption is more or less a fair one for the majority of medical interventions and the associated RCTs.

The same cannot be said for many RCTs in development economics. Consider deworming children. It may be that a mass deworming program has the effect of better educational outcomes (and associated human capital development) in model population A. But how do we extrapolate that result to other settings? We need to assume that causal pathway, i.e. the mechanism by which the intervention has its effect -- the drug works (okay), school children have fewer parasitic infections (okay), so they have more energy (maybe), are able to attend school earlier (are you sure?) and are more attentive in class (maybe), leading to better educational outcomes -- also obtains in the target population, often in some very different social context. In other words, the causal inference is only useful when it is accompanied by all kinds of other evidence as to whether the causal inference can survive extrapolation. (Please don't take me to be saying deworming is bad. I'm in favour of it, but not because of its affects on human capital accumulation!)

So what's the problem? Well in some sense there isn't one. RCTs are great! But they have to be accompanied by careful empirical research, as you say. But ideal (or close to ideal) RCTs are extremely expensive and time consuming. Furthermore, from a policy design perspective, placing RCTs on a pedestal may come at the cost of the other types of research necessary for good usable policies. Institutional demands for RCTs may also restrict funding for plausible evidence-based (but not RCT based!) interventions. There is nothing wrong with experimenting.

In my view, extrapolation is the real challenge for RCTs in development economics. It is a problem that medicine doesn't really need to grapple with to the same extent -- but they do. And economists should too.

Expand full comment
Rory Hester's avatar

I used to be an educational blogger for a few years. I am pretty familiar with educational research, and how poor it is. Even when its good... people ignore the results.

I suggest looking at "Project Follow Through" https://en.wikipedia.org/wiki/Follow_Through_(project)

It was the largest most rigorous education experiment ever conducted, and it clearly showed that Direct Instruction was the most effective teaching method for young children, yet here we are with project based learning taking over our schools.

RCTs are only good if we actually design them well, and then actually pay attention to their results even if they give answers that we disagree with.

Expand full comment
20 more comments...

No posts