Getting Meta About Uncertainty

There's uncertainty in how we reduce uncertainty, and that's okay.

Posted Aug 16, 2019

The world is complicated. Lots of stuff happens, mainly because lots of other stuff happened previously. Making sense of all of it—and predicting what will happen next—is a challenge. Humans spend a lot of time grappling with this challenge.

The good news is that we’ve developed some tools to help us cut through the noise. On an individual level, people observe the world, look for information and patterns that might be useful for understanding that world, and then form beliefs. In short, beliefs are things we think are true. These mental shortcuts help us explain what has happened and figure out what will happen next.

From Helena Lopes on Pexels
There's a 60% chance of rain?
Source: From Helena Lopes on Pexels

For instance, you might believe that if you see cows lying in a field, a storm is on the way. This belief gives a useful guide for predicting future events: Worried about rain? Consult the cows.

Scientists are like most humans (but maybe slightly more neurotic). We are similarly trying to form beliefs about the world and figure out what’s most likely to be true. We use fancy equipment, fancy statistics, and, most importantly, fancy terminology, but we’re fundamentally playing the same game. We’re observing what’s out there, trying to figure out why it happened, and then using that knowledge to try to predict what will happen next.

These efforts, by both scientists and people generally, are aimed at reducing uncertainty in the world. This is a noble pursuit. However, it’s not a perfect one. There is uncertainty in how we try to reduce uncertainty. 

An example of meta-uncertainty: the many ways scientists can analyze data

There’s a lot of information in the world. There are many patterns that can be used for uncertainty-reduction. Picking the right information to pay attention to can be a challenge. 

As an example, let’s say a certain Midwestern guy (purely hypothetical) wanted to test his belief that cows can predict storms. He wants his belief to be empirically supported, so he goes out and collects a bunch of data on the weather and the position of cows. Now that he’s got his data, he should have the answer, right?

Well, it depends on how he wants to analyze his data. For instance, what counts as a storm? Any rain? Does there have to be lightning? Also, maybe there are features of his bovine subjects that he should be taking into account. Maybe he should consider how long the cows in his study were standing before they laid down (they should get tired sometimes, after all). 

Each of these choices is defensible. From that perspective, his data don’t provide him with the answer to his question, but rather several answers to his question. He has to pick which answer makes the most sense.

Scientists face this same problem. In most scientific studies, there are many ways a researcher could analyze her data. Many of those ways probably make sense. However, they will each give a slightly different answer to the same research question. In one great example of this, several teams of researchers were given the same dataset and the same question and asked to answer it. In this case, they were asked to answer the question “are dark-skinned soccer players more likely to receive red cards than lighter-skinned players?” They were then told to come up with what they considered to be the best statistical test of that question.

Twenty-nine teams of researchers produced an answer to the research question. They collectively produced twenty-three different answers. Like our Midwestern meteorologist, these researchers had many choices to make when analyzing the data. This led to many different choices. Overall, answers tended to point to an unfortunate conclusion — about two-thirds of the researchers found support for the conclusion that dark-skinned players are more likely to be red-carded. However, how strong this bias is depends on how you calculate it.

How to think about meta-uncertainty

There’s a pessimistic way to view studies like these. If scientific findings are dependent on the choices of researchers, and there are many possible choices, how can we trust any one answer? That’s a fair concern. 

I take a more optimistic approach. Studies like this help to reveal the uncertainty that is baked into the uncertainty reduction process. If we know about this uncertainty, we can take it into account when adjusting our beliefs about the world. Better yet, we can do things to help combat that uncertainty (for instance, by having researchers make analysis choices before they play around in the data, or by running lots of different analyses and looking at the distribution of results).

Meta-uncertainty is fine when we know where to find it. It’s the things that we’re certain about when we shouldn’t be that are the problem.