Do you have a bias bias?
I'm finishing up Enlightenment Now by Canadian, Harvard Psychology Professor Steven Pinker and I came across a term with which I was unfamiliar. Here is my attempt to understand it by explaining it. If anyone out there who knows this idea better than I do, please add/revise my explanation.
The benefits of simplicity are overlooked because of a bias bias. ie. Preference for more sophisticated models designed to avoid bias but may not lead to the most accurate prediction. When dealing with predictions, error occurs for two reasons, the first is due to bias, the inability of the model to represent certain patterns that are being observed. So if we're trying to predict how many men vs. women were going to apply for a job in engineering, our prediction model might be biased if we are only basing our prediction on past data, since today's women might be more likely to apply for engineering jobs than those of the past. The other type of error comes from variance, which is the model's sensitivity to different observations of the same problem, like a different sample from the same population. So maybe we are holding this study in Canada but our model predicts accurately in rural but not urban populations.
A 'bias bias' is when you prefer a model that limits bias but pays little attention to variance. You might go through rigorous efforts to limit your bias by considering a plethora of factors and viewpoints, making your decision making process in depth, dynamic, and complicated but in the end suffer the fate of a poor prediction all the same. Why? Because you put too much attention on avoiding bias and not enough on attention on what was probably the more simple decision making formula.
Let's see if I can take my own medicine and explain it more plainly (and I don't mean that in a condescending way, this is a test to see if I have the ability to explain this concept in terms most people can understand, because if I can't then it's evidence that I actually don't understand it myself at all.)
Scenario: A group deciding which movie to rent at the store (we're going 90's for our example today)
Complicated decision making model designed to reduce bias
We poll the group to gather data on move genre preferences, previously viewed titles, and willingness to watch a move they have previously viewed on a scale of 1-10. We analyze the data based on socioeconomic, racial, and religious backgrounds. The data discovers that the horror genre subgroup is voting as a block to sway the data away from rom-coms. To compensate we add a weighted point system so that each preferred genre has an equal voting share. When the final calculations are tallied, all genre preference biases are accounted for and corrected and the movie selected turns out to be "Encino Man" which was not on anyone's top five. Besides, they spent 2 hours in the store and no longer have time to watch the movie anyway.
Simple decision making model
The majority of the group wants to watch 'Reservoir Dogs', the three who preferred 'Arachnophobia' are asked if watching 'Reservoir Dogs' would be a dealbreaker. They reply, "no". The group gets 'Reservoir Dogs' and group enjoyment is maximized.
Occam's Razor has told us for centuries that often the simplest solution tends to be the right one. While I am a proponent of critical analysis, we often bog ourselves down with unnecessary variables and assumptions that complicate an otherwise fairly straight forward decision.
If you have a bias bias, then maybe you are making your decisions too complicated. Maybe you are so intent on eliminating the bias that might account for 10% of the potential error that you forego the common sense that accounts for the other 90%.
The benefits of simplicity are overlooked because of a bias bias. ie. Preference for more sophisticated models designed to avoid bias but may not lead to the most accurate prediction. When dealing with predictions, error occurs for two reasons, the first is due to bias, the inability of the model to represent certain patterns that are being observed. So if we're trying to predict how many men vs. women were going to apply for a job in engineering, our prediction model might be biased if we are only basing our prediction on past data, since today's women might be more likely to apply for engineering jobs than those of the past. The other type of error comes from variance, which is the model's sensitivity to different observations of the same problem, like a different sample from the same population. So maybe we are holding this study in Canada but our model predicts accurately in rural but not urban populations.
A 'bias bias' is when you prefer a model that limits bias but pays little attention to variance. You might go through rigorous efforts to limit your bias by considering a plethora of factors and viewpoints, making your decision making process in depth, dynamic, and complicated but in the end suffer the fate of a poor prediction all the same. Why? Because you put too much attention on avoiding bias and not enough on attention on what was probably the more simple decision making formula.
Let's see if I can take my own medicine and explain it more plainly (and I don't mean that in a condescending way, this is a test to see if I have the ability to explain this concept in terms most people can understand, because if I can't then it's evidence that I actually don't understand it myself at all.)
Scenario: A group deciding which movie to rent at the store (we're going 90's for our example today)
Complicated decision making model designed to reduce bias
We poll the group to gather data on move genre preferences, previously viewed titles, and willingness to watch a move they have previously viewed on a scale of 1-10. We analyze the data based on socioeconomic, racial, and religious backgrounds. The data discovers that the horror genre subgroup is voting as a block to sway the data away from rom-coms. To compensate we add a weighted point system so that each preferred genre has an equal voting share. When the final calculations are tallied, all genre preference biases are accounted for and corrected and the movie selected turns out to be "Encino Man" which was not on anyone's top five. Besides, they spent 2 hours in the store and no longer have time to watch the movie anyway.
Simple decision making model
The majority of the group wants to watch 'Reservoir Dogs', the three who preferred 'Arachnophobia' are asked if watching 'Reservoir Dogs' would be a dealbreaker. They reply, "no". The group gets 'Reservoir Dogs' and group enjoyment is maximized.
Occam's Razor has told us for centuries that often the simplest solution tends to be the right one. While I am a proponent of critical analysis, we often bog ourselves down with unnecessary variables and assumptions that complicate an otherwise fairly straight forward decision.
If you have a bias bias, then maybe you are making your decisions too complicated. Maybe you are so intent on eliminating the bias that might account for 10% of the potential error that you forego the common sense that accounts for the other 90%.
Comments
Post a Comment