Thursday, March 21, 2013

Model vs. Reality

My twitter feed currently is ablaze with discussions about sexism in tech and science. As a member of the social group that goes through life with difficulty set to "easy", of course none of this has happened to me. I don't know what it feels like to go to a conference where people constantly comment on my looks or gender instead of my work. As I lack the experience, and I also lack solid data, I don't want to write about that in this post.

However, as a computational biologist, I do have some experience with model vs. reality clashes, and I believe that might be the reason why people on the internet are surprised about the existence of female scientists or engineers. People also tend to get upset when they realize their mental model doesn't match reality, which might explain some of the emotional upset males show in the discussions I mentioned in the first paragraph.

Now, instead of taking the easy way out and blaming this on internet stupidity, I want to put another theory out there. People often don't get statistics. In my day-to-day work, I frequently run into publications where there is at best a loose correlation between the data and the model supposed to explain said data. And I believe that's what is happening here as well.

So by the time-proven method of pulling data out of my hat, I propose that when you graph the "ability to do science" against the sex chromosome layout of individuals, you'll get something similar to this:

So far, so good. The problem is that despite "ability to do science" not clustering for any sane definition of a cluster, the mental model of many people seems to look like this:

If you have settled for a given model, there is quite some inertia to stay with your chosen model, even if the data doesn't back it up. If reality dares to come up with conflicting data, blame reality! The ripples of a lot of mental models running into reality hard are currently washing over my twitter feed. The inertia of sticking with your model makes it hard to realize it, but in the end when reality and your model disagree, it is easier to change your model. In my example graph, a lot of lines I could put in there would likely have a similar quadratic error. To me, this is a warning sign that my model probably is bad. In the example, the conclusion should be that not only "women are bad at science, men are great" (the red line) is wrong, but also every other attempt at constructing a linear correlation between the parameters. "Ability to do science" and "sex chromosome layout" are orthogonal characteristics*. Also, why is there an arrow on the x axis, when we're looking at discrete parameters?

* According to my hat, of course.

No comments: