Eric J Ma's Website

Simplifying Uncertainty Responsibly

written by Eric J. Ma on 2019-12-19 | tags: bayesian data science uncertainty decision making

I read an article from Brandon Rohrer titled "Oversimplify". The article provoked some thoughts.

As someone who strives to model uncertainty in a day-to-day setting, it’s easy to misread Brandon’s article as saying, "discard your modelling of uncertainty". It took me a few reads and a bit of thinking to realize that my misunderstanding would have been wrong. The gist of Brandon’s article is to communicate the fact that most people don’t like uncertainty, so we must simplify the communication of uncertainty.

In a Bayesian setting, I think this corresponds mostly to "minimize regret" when making decisions.

Here’s a classic example: weather forecast says "30% probability of precipitation" (assume this is rain). How does this fit into Brandon’s conception of simplifying the communication of uncertainty?

Simply telling someone the probability of precipitation doesn’t help. It’s like stating an unhelpful fact - unhelpful if the audience doesn’t have a thought framework for acting on that uncertainty.

Any good card-carrying Bayesian who also knows how to minimize regret would say, "there’s a 30% probability of precipitation. Since getting wet is more miserable than carrying an umbrella on a cloudy day, take that umbrella, and throw in your waterproof jacket and boots while you’re at it."

By contrast, someone other consultant might take the same probabilities, and instead recommend, "Oh, there’s basically greater odds of not raining than raining, so wear your cotton jacket since it’s cold, but don’t bother about an umbrella." This consultant, we might say, is a definite risk-taker, but the consultant is also simplifying the communication of uncertainty by providing an actionable recommendation.

The key question most people are seeking an answer to is not "what do the data say", but rather "how should I act?" Though the latter question is normative, it can be informed by quantitative reasoning.

So to summarize: _most people don’t like uncertainty; we can simplify the communication of uncertainty by providing an actionable recommendation based on that uncertainty.

I send out a newsletter with tips and tools for data scientists. Come check it out at Substack.

If you would like to sponsor the coffee that goes into making my posts, please consider GitHub Sponsors!

Finally, I do free 30-minute GenAI strategy calls for organizations who are seeking guidance on how to best leverage this technology. Consider booking a call on Calendly if you're interested!