Will Eating Chocolate Make Me More Likely to Win a Nobel Prize?

I'm going to go out on a limb here and say 'No' but stay with me because, as you've probably guessed, there's more to the story than that.

Consider the chart below. It shows chocolate consumption on the x-axis and Nobel Laureates on the y-axis. Both are adjusted for population size. Each datapoint is a country. The evidence is pretty compelling: people in countries that eat more chocolate win more Nobel Prizes. The strength of the correlation is very higher. 0.79 is higher than, say, height and shoe size. And without getting into debate on what a p-value represents, the 0.0001 tells us that this is almost certainly not random.

So, if this is a real correlation (and it is) and it's not just random noise (and it isn't) then what is going on? Does it mean that consuming chocolate causes people to win Nobel Prizes?

Perhaps surprisingly, Dr. Franz H. Messerli did posit something similar in his paper "Chocolate Consumption, Cognitive Function, and Nobel Laureates" in the New England Journal of Medicine (the source of the above chart).

Messerli noted the strong correlation and also that "improved cognitive performance with the administration of a cocoa polyphenolic extract has even been reported in aged Wistar– Unilever rats".

It's worth looking up Dr. Messerli's paper because it's a short and enjoyable read. He also concedes the limitations of the study, "It remains to be determined whether the consumption of chocolate is the underlying mechanism for the observed association with improved cognitive function", and tellingly closes by reporting his own "daily chocolate consumption, mostly but not exclusively in the form of Lindt’s dark varieties".

Of course this isn't really about chocolate and Nobel Prizes but is a cautionary tale about confusing correlation and causation, especially in situations where we'd really like causation to be true. There is a correlation above and it's driven by other (exogenous) factors like income, and social and geographic proximity to Northern Europe. I suspect Dr. Messerli knew this all along and was being a little playful in his writing but, nevertheless, his paper is often cited as an example of the perils of over-interpreting correlations. Most notably in a follow up paper titled: "Does Chocolate Consumption Really Boost Nobel Award Chances? The Peril of Over-Interpreting Correlations in Health Studies".

In business, correlation and causation get confused on a daily basis. In part because sometimes it is very hard to untangle the two, but more likely because we are all incentivised to believe that our efforts are causing impacts. Any finance team will tell you that an observed lift in revenue will find itself with multiple internal causes and any drop in revenue will find itself with none.

If we really want to understand causality we need to go beyond the observed correlation. We could proactively do this in the form of AB testing. In Messerli's case this would be allowing half the population to eat chocolate, preventing the other half, and measuring the difference in Nobel Prize wins between the groups. Analytically this would be the easiest way to determine the causality, but it's fair to say that such a test would be difficult to execute.

A more practical solution would be an expanded overarching model that controls for all those other factors that could be influencing Prize wins (social, geographical, linguistic etc). This would be more retrospective and would not require a chocolate embargo but the analytical process would involve more assumptions.

And so finally we arrive at the point of this blog. Businesses face similar choices when trying to understand the impact of their advertising. AB testing (or lift-testing) is great where available because the results are transparent and intuitive. But it's not always possible to cleanly split groups of people and give them different treatments. The "expanded overarching model" is akin to a Media Mix Model. Less intrusive in terms of treatment but much more complex analytically. Both great options though and both very doable with the right resource.

We have more AB testing here and Media Mix Models here.

Or reach out to us directly if you'd like to talk more info@codarossa.co