Survivorship Bias (and War Math) - What You Don't Know Can Hurt You
The Misconception: You should focus on the successful if you wish to become successful.
The Truth: When failure becomes invisible, the difference between failure and success may also become invisible.
Since well before Freakanomics, there is a lot of writing around that try to show how the "truth" is actually the reverse of conventional wisdom. Often this is because the people using the data don't understand what the data actually represents. To borrow Rumsfeld's words, there are "unknown unknowns".
This article talks about the problems with data, and the need to have common sense brought to it. It doesn't matter how good the modelling or analysis of the data is if the people doing it don't understand what the data is.
One early example is looking at WW2 bombers. If the bullet holes in the ones that returned occur in similar patterns, wouldn't that show the best place to put additional armour?
Actually no, the holes showed where a bomber could be shot and still survive the flight home. What they didn't, or couldn't, analyse were the bullet holes on the planes that didn't make it back, the analysis was in danger of suffering from surivorship bias.
It is easy to do. After any process that leaves behind survivors, the non-survivors are often destroyed or muted or removed from your view. If failures becomes invisible, then naturally you will pay more attention to successes. Not only do you fail to recognize that what is missing might have held important information, you fail to recognize that there is missing information at all.
And thinking about it in modern terms, it's very rare to see an unsuccessful case study. There are a few that have moments of crisis, but these are usually for dramatic effect. If something fails, there's not usually much publicity in analysing it.
The successes, on the other hand, are much more exciting, but
the problem here is that you rarely take away from these inspirational figures advice on what not to do, on what you should avoid, and that’s because they don’t know.
Where there's data, there's analysis, and there are some incredibly complex models that can show relationships between almost anything. But just because there is data, that doesn't mean it's the complete set of data, and as data goes into a model, it gets less and less transparent and harder to question, especially if the answers that are being generated seem rational.
Digital often suffers from this in the marketing department - we can't measure print impact, but we can measure clicks, so that's our proxy for success. It's a variant of Maslow's "unconscious incompetencies"; when a task looks easy because of the things you don't know you don't know.
That doesn't mean the data you have shouldn't be analysed (starting an analysis can often mean the start of a move from "unconscious incompetence" to "conscious incompetence"), but any analysis should at least try to look at the broader picture of what that analysis is trying to accomplish, and whether the data is appropriate for that.
After all, "conscious incompetence" is probably the minimum we should accept.
