Aidan Coville and I have a new working paper in which we look at how policymakers, development practitioners (such as international organization staff), and researchers update in response to evidence from impact evaluations. In particular, we test whether they are subject to two biases: updating more on “good news” than “bad news” and not taking the variance of the estimates adequately into consideration (here, we mean confidence intervals, but you could imagine testing for neglect of other kinds of variance, such as inter-study variance).

The first bias is sometimes called overconfidence (not to be confused with the kind of overconfidence in which one simply is too certain in one’s beliefs), and in our experiment we are able to distinguish it from confirmation bias. The second bias is novel to my knowledge, and we call it “variance neglect”. In our context, it refers to ignoring confidence intervals, but you can also imagine other types of variance being ignored, such as inter-study variance. Variance neglect is related to “extension neglect”, in which people may, for example, neglect sample size when considering how much to trust a new finding. However, I think it is distinct, since sample size is not all that matters when considering variance. It is also reminiscent of Kahneman and Tversky’s prospect theory, in which people overweight low probability events and underweight high probability events. However, if someone overweights low probability events and underweights high probability events, that should have the effect of increasing the dispersion of their beliefs. What we find, in supplemental tests, is more consistent with a fundamental misunderstanding of confidence intervals. We give respondents one interval and ask them to provide a different interval (e.g. giving a 95% interval and asking for the interquartile range). For small ranges, they have overly disperse distributions, but for large ranges they report overly narrow distributions. Variance neglect is perhaps more closely related to the hot hand fallacy and the gambler’s fallacy, which also have to do with incorrect treatment or perception of variance. I would also distinguish variance neglect from the hot hand fallacy and the gambler’s fallacy, though, both because variance neglect does not require as restrictive a functional form and because the latter two biases conceptually have to do with seeing repeated streaks, while I am more concerned with seeing noisy data once. (Especially given how I know from past work how rare it is to have multiple studies on the same intervention covering the same outcome – instead, studies often “run away from each other” as authors seek to make them unique.)

We show how these biases can be easily fit into a quasi-Bayesian model. Testing for these biases is also straightforward. First, we elicit respondents’ priors. Then we randomly vary whether we show them high or low point estimates (relative to their priors) and large or small confidence intervals. Finally, we elicit their posteriors. That allows us to cleanly estimate whether they update more on the “good news” than the “bad news”. Testing for variance neglect is a little more complicated: for someone to suffer from variance neglect, they don’t need to update equally on large confidence intervals and small confidence intervals, nor do they need to perversely update more on large confidence intervals. Instead, all we need to show is that they do not update *as differently* when they see small confidence intervals as opposed to large confidence intervals as they would if they were Bayesian. Since we know their priors and we know what data we showed them, we can determine how a Bayesian would have updated and compare their responses.

The setting is also pretty cool. We can’t bring policymakers to a lab, but we can bring the lab to policymakers. We leverage a series of World Bank and Inter-American Development Bank impact evaluation workshops in various countries. These workshops tend to be one week long, and over the course of the week policymakers working on a particular program learn about impact evaluation and try to design one for their program with the help of a researcher. Several different types of participants attend these meetings: “policymakers”, or officials working in developing country government agencies (both those in charge of the particular programs and monitoring and evaluation specialists); “practitioners”, which are mostly international organization staff (like World Bank operational staff) and NGO partners; and researchers. Apart from the workshops, we also conducted some surveys at the headquarters of the World Bank and the Inter-American Development Bank. Finally, we also ran the experiment on Amazon’s Mechanical Turk (where workers take surveys for cash) to obtain another comparison group.

Quick summary of results: we found significant evidence of overconfidence and variance neglect, and no respondent type (policymakers, practitioners, researchers, and MTurk workers) updated significantly better or worse than any other. We also disaggregated results by gender; a few results indicated that women suffered more from these biases but it is difficult to interpret given that a) other results did not indicate that and b) the distribution of women in the sample varied by respondent type. Finally, we found that respondents updated more on more granular data, even when the data should have theoretically provided the same informational content (e.g. providing more/fewer quantiles of the same distribution). This suggests that when you have bad news, you should come bearing a lot of data.

The working paper has more details. There is still some work to do (e.g. non-parametric tests, robustness checks), but we are pretty happy with the initial results. We also ran several follow-up experiments on MTurk that try to parse the ultimate cause of the variance neglect we observe. We are running these experiments on a pre-specified larger sample and will add them to the next draft. Preliminary results suggest that misunderstanding confidence intervals (as opposed to inattentiveness or lack of trust in the experiment) is the crux of the issue.

This is not the only experiment that we have run in this space. We also ran a set of experiments on a similar sample that tried to determine how people weight different aspects of a study’s context or research design. For example, would a policymaker, practitioner or researcher prefer to see results from an RCT done in a different region or results from a quasi-experimental study done in their setting? This was a simple discrete choice experiment in which respondents were asked to repeatedly pick one of two studies with different attributes. Other attributes tested (in various permutations) include program implementer (government or NGO), sample size, point estimates, confidence intervals, and whether the program was recommended by a local expert. You would need far stronger assumptions than we would be comfortable making in order to say that anyone was making a “correct” or “incorrect” choice, but their preferences were interesting nonetheless. Tune in next time to find out how respondents answered.