Reducing complex phenomena to one factor so that they can be digested on a 30-second news segment or put into soundbites by any politician is an unwelcome consequence of modern media. I have no doubt many factors cause demographic changes such as the one discussed. Testing for the weight of each factor needs scientific work and controlled experiments which are pretty much impossible to carry out, unfortunately. Until then, things will be talked and written about, but remember, it's all fluff. Someday, some historian will even write the definitive history of women in early computer science, and I have no doubt his conclusions will be as disputed as the fall of the Roman empire is today.
But these discussions will still serve as vehicles for political agendas, most of which will be of the variety "more people should be like me".
Survival rate by stage:
Submitted resume gets reviewed:
1.05 : 1 in favor of women.
Reviewed resume gets a code review:
1.02 : 1 in favor of women.
Reviewed code gets a phone interview:
1 : 1.96 in favor of men.
Completed phone interview leads to an in-person interview:
1.31 : 1 in favor of women.
Completed interview leads to an offer:
2.92 : 1 in favor of women.
Candidate declines offer:
2 : 1 in favor of women.Not suggesting that's what fogcreek do, of course, but just to demonstrate that a coding test could easily have a gender bias - or a cultural one, for that matter.
Lots of times you need to develop stuff without domain knowledge. My startup targets women's fashion, which I know very little about. If I were to give up or fail because the problem involves plackets and silhouettes, I'd have already quit and gone back to NJ.
@diolpah it is a very small sample, it probably carries no scientific relevance. It is, however, the author's fault for using it, not yours. Your analysis looks correct as far as the tiny dataset presented is concerned.
EDIT: "Because we can’t ask applicants their gender, we guessed based on first names." Nevermind. The author is a jackass for even including these data in the first place, it's as good as meaningless, even within that small cohort. Alex, Ashley, Sam, Pat, Lindsey, Jessie, any hippy name (see: my name), Quinn, Casey, Rowan... crap pseudoscience is crap.
And, just to be clear, there is _plenty_ of discrimination and bias against women. But that shouldn't detract from the fact that you're pointing out a major outlier in their data.
Unfortunately, without knowing the test, it's hard to even guess what happened there. Is it focusing on syntactic vs. semantical correctness? Does it require deep specialist knowledge, or broad generalist knowledge? How is it graded?
Or is it simply a function of the tiny sample size?
Without pointing out what you conclude from that data, people will infer what you concluded. And since it's a loaded subject, they will often arrive at "what a misogynist jerk" without ever knowing what you were trying to say.
Corollary: When pointing out data on a loaded subject, it might be beneficial to point out what you conclude, and why. At least that way, only one of the two factions can flame you ;)
(Disclaimer: I am a woman, so I obviously disagree with the "women can't code theory". I'd still love to find out _why_ women are filtered out by that test so disproportionately, simply because it might hold clues as to why women are not interested in CS)
Also, the sample size is crap and the deviation is crap relative to the sample size, so you can't make very many truly meaningful observations from it.