How to Spot if the Data is Lying

“Old-school bullshit” has been there for ages – it includes lies, rumours and propaganda, write Carl Bergstrom and Jevin West in Calling Bullshit. Their book, though, is about ways in which we can detect a different kind of bullshit:

“New-school bullshit uses the language of math and science and statistics to create the impression of rigour and accuracy.”

Is there a difference though between the two kinds? Yes, unlike old-school benefits where one can Google or apply one’s own knowledge, one often has no idea how to question data and numbers:

“New-school bullshit can be particularly effective because many of us don’t feel qualified to challenge information that is presented in quantitative form.”

The book is about ways to overcome this self-perceived limitation.

 

The first thing to ask, they point out, is if the data is valid. This is increasingly important since we rely and trust AI and Machine Learning algorithms so much these days. We’d do well to remember that if the data used to “teach” the system was flawed, its conclusions will be flawed. For example, take this claim that a Machine Learning algorithm could identify criminals vs the rest of us with remarkable precision based on just a photo. Unfortunately, it turned out that the algorithm was trained via mugshots of criminals and Facebook posts of the rest of us. The problem with that training data? Nobody smiles in a mugshot, and everybody smiles on Facebook; so what the algorithm had really learnt was to use the lack of a smile as a higher sign of being a crook!

 

A related question is easier: does it sound like the data is unbiased? Ask yourself if an election survey cover both rural and urban areas? Was the “global opinion” on Ukraine collected only in the West or in other countries as well?

 

A key question to ask oneself is this: Is the data relevant to the conclusion being drawn? An example will help. Many companies say their caffeine content is less than, say, 1%. The authors point out that the concentration of caffeine in any drink is far less than 1% since they are all over 99% milk/water! So while saying caffeine is less than 1% isn’t wrong, it doesn’t mean that it has less caffeine than any other beverage either. If they had said 1% of the caffeine content in regular products, that would mean something. But an open-ended “1% caffeine” claim is not only meaningless, but even worse, misleading.

 

Sometimes absolute numbers are more important than percentages. At other times, it is the other way around. Inevitably then, presenters use the metric (absolute number, or percent) that makes their case appear stronger. Watch out for this by asking yourself how the data would look if you flipped things.

 

A risk with the usage of percentage is if negative numbers are involved. An example is global smartphone profits – if you add Apple and Samsung’s shares, it will be close to 100%. Told like that, many conclude it means that everyone else is making losses. Wrong! How come? Because the rest could include 5 manufacturers who make, say, 10% of global profits, while all the remaining are at losses (there’s that negative number I was talking of), i.e., their collective share is minus 10%.

 

Then there is the subtler problem we’ve all seen at the workplace. When a (measurable) proxy is used for quantifying an (immeasurable) intangible, everyone maximizes the proxy to a point where there’s no connection between the two anymore. Examples include counting patent filings as a proxy for creativity; and using test marks as a proxy for learning. It’s good to remember this and ask oneself what has been measured and presented – the item of interest? Or a proxy? Is the proxy a valid way to evaluate the original intent?

 

The last point one can think analytically about is the well-known yet easy-to-forget one: Does the data show that A seems to be connected to (correlated) B? Or does it prove that A causes B? We’d do well to remember that in the social sciences, causation is very hard to prove because the “everything else being the same” part of the experiment is so hard to achieve.

 

I felt these are good rules of thumb, since they are all generic warnings and questions that we can ask ourselves on most data we are presented with.