Hacker News Reader

Comment by jim-greer | Hacker News Reader

This analysis is based on crowd-sourced data on shootings. I'm not saying that makes it invalid, but I'd need to know more about the dataset before trusting the results. It seems hard to have it be complete. The official FBI data has the same problem. We need to require local departments to report all shootings in a standardized way.

Edit: the author acknowledges the incompleteness of the data in the conclusions section. Oddly, he doesn't think that's likely to affect the mean "since the sample used herein is a large and random subset of the to-be-completed data set". That doesn't really make sense to me. How would random sampling of incomplete data improve the results?

http://regressing.deadspin.com/deadspin-police-shooting-data...

View on HN

"How would random sampling of incomplete data improve the results?"

It doesn't. The author means there is some actual population of shootings, and this dataset is a random sample of them. Since there are no systematic biases in the collection of the data, ie, it's not the case that the shooting of a black is more likely to be recorded than that of a white, the random sampling of the full set is sufficient. That's an assumption, of course, but given that assumption the rest works.

It doesn't improve the results, it just doesn't harm them.

I'm not sure that's what he's saying: "the sample used herein is a large and random subset of the to-be-completed data set". I read that as meaning database is to-be-completed, and he's sampling from that.

What seems to be missing to me is any support for the "random sampling" claim. My starting assumption is that reports on police shootings would be full of systematic biases running in all sorts of directions.

Without some actual argument that this data isn't horribly biased by exactly the issues it's trying to investigate, I'm not sure what worth to give it.

Specifically, random sampling assumes that data points are independent and identically distributed: https://en.wikipedia.org/wiki/Independent_and_identically_di...

As others have pointed out, that is not a great assumption in this case.

It isnt, and indeed it is rarely satisfied. The real question is how sensitive the resulting model may be to violations of this assumption.

Greg Ridgway looked at a natural experiment (proportion of drivers pulled over 1 hour before dawn, when the color of the driver is very hard to determine a priori, vs that proportion 1 hour after dawn, where it is easy) back when he was at RAND, and came to a similar conclusion. Ymmv.

Obviously, "pulled over" is not the same as "shot" but the objective data is easier to come by. Especially if you don't tell the LAPD what specifically you're studying beforehand.

Reply on news.ycombinator.com