3 ms·
I think the author is missing some really key points here. Namely that GPA and test scores are highly correlated. I'd go so far as to say that there is a confou
by godelski 4y ago
I think the author is missing some really key points here. Namely that GPA and test scores are highly correlated. I'd go so far as to say that there is a confounding variable that predicts both. I'd even say there are several upstream predictors, but I think we all know that (or I thought we did). Every single graph they show demonstrates this yet they keep missing it. The top comment right now (by ajwinn[0]) notes that they used a heatmap like the one in the article for determining admission. This is great because it builds multiple pathways to gain admission. That's honestly how we should treat variables like these: rely on either.
As to explain the confusion in the models, well when you have confounding variables like this it is easy to reduce one variable's weight nearly to zero and place the rest in the other variable. So if I'm right (might not be) I think you could also show that the inverse is true, that grades aren't a good predictor and test scores are. The breakdown suggests that this is true too tbh.
But I also disagree with some of the analysis and techniques here. It's not always easy to use models that have continuous(ish) variables (GPA/ACT/income) and discrete ones (race/gender). I bring this up because there's some hints that they aren't a stats expert. For example there's the line
> The “likely-to-graduate” score becomes a probability after a sigmoid transformation.
Which just isn't true. A Softmax function[1] creates a probability distribution but a Sigmoid[2] doesn't. A Sigmoid bounds the output to [0, 1] (or [-1,1]) but being bound on that range isn't the only requirement to a probability distribution. A sigmoid does not guarantee that the sum of the elements is 1 (cumulative density), but a softmax does. In fact, it almost never will. (I mention this in my ML class a lot so make sure you all don't make this mistake too! I'll be really annoying if you do)
There's some more errors and novice approaches here and so I just want to note this because we shouldn't just take the results and conclusions as is when the modeling is bad. But I still think he makes some good (and bad) points and those are definitely up for discussion.
[0] https://news.ycombinator.com/item?id=33338809 https://news.ycombinator.com/item?id=33338809
[1] https://en.wikipedia.org/wiki/Softmax_function https://en.wikipedia.org/wiki/Softmax_function
[2] https://en.wikipedia.org/wiki/Sigmoid_function https://en.wikipedia.org/wiki/Sigmoid_function