Hacker News Reader

Comment by contravariant | Hacker News Reader

Just playing devil's advocate here, but with the numbers given in the essay (10 out of 250 male entries tagged for deletion vs 2 out of 12 female entries) I got a p-value of approximately 10% [1]. By most criteria this wouldn't qualify as statistically significant. This doesn't immediately disqualify the entire essay, but the claim that women are 'over four times more likely to be deleted' is somewhat dubious.

[1]: http://www.wolframalpha.com/input/?i=1+-+(1+-+12%2F262)%5E12...

View on HN

There are many things that are not statistically significant, but still interesting when it seems likely* that a larger sample would give a similar result.

Only 17.47% (as of 26 February 2018) of the English Wikipedia's biographies are about women: https://en.wikipedia.org/wiki/Wikipedia:WikiProject_Women_in...

EDIT: I rechecked the quote from the article about the likelihood of deletion and found it to be "in this cohort of Erdős Ones, the pages for women were over four times more likely than men’s to be nominated for speedy deletion." (my italics). FYI.

> Only 17.47% (as of 26 February 2018) of the English Wikipedia's biographies are about women:

Wikipedia being a tertiary source may, through unbiased processes, simply be reflecting bias in the secondary sources. Biographies of pre-modern women are far less common than of their male contemporaries, IME; that's less true of my modern and especially extremely current figures, but Wikipedia isn't exclusively focused on current figures.

There is no such thing as a tertiary source. Either you have a primary source, or you have hear say.

You may not like secondary and tertiary sources, but they absolutely are well-defined categories, and Wikipedia's overtly declared mission is to be a tertiary source, which why original research and most direct citation of primary, rather than secondary, sources is restricted by policy.

That % isn't really surprising considering the patriarchal history of the world. You couldn't possibly get to 50% unless you over-represented women in the near present or adjusted notability standards.

However if it were 17.47% for current times, that would be a significant bias that needs to be addressed. It still might not hit 50/50 given reduced opportunities for women still today, but it should be moving ever closer.

The fraction of women with Erdos number one is already very small, no reason to assume that they are less important than corresponding man with the same Erdos number.

The conclusion is based on consideration of the entire population of biographical Wikipedia entries with an Erdos Number of One. It is not statistical except in some lay sense that any reference to percentages makes something statistical.

Talking about p-values in contexts where one has all the facts and the additional fact that one has all the facts suggests a gross misunderstanding of statistical concepts, their utilities, and the circumstances in which they apply.

I know I'm leaning more towards Bayesian than frequentist, but I vehemently disagree with the notion that applying statistics in this case is meaningless.

First of all I at least don't have all the information. If I had it would indeed be meaningless to infer the reason behind the deletion attempts, as I'd know already, but the fact remains that I don't. By this logic trying to see how likely this was to happen assuming articles were just randomly marked for deletion is a perfectly reasonable test.

Second of all, how exactly is this situation different from tossing a coin several times? If I toss a coin 50 times, record the results, and then melt down the coin, surely it'd still be valid to conclude that it was biased if it came up heads 49 times out of 50? Or conversely that it wasn't biased if it came up tails 26 times out of 50?

Arguing that statistics only applies to experiments that can be repeated suggests a gross misunderstanding of statistical concepts, their utility and the circumstances in which they apply.

By this logic trying to see how likely this was to happen assuming articles were just randomly marked for deletion is a perfectly reasonable test.

Maybe I'm misunderstanding Ben's point, but I don't think he's objecting to this as a frequentist. You have correctly calculated that given the small sample size of female candidates, there is about a 10% chance that random deletion of entries would produce a disparity this large.

I think Ben's issue is that since we already know that the deletion process is nonrandom, this tells us nothing we do not already know. The question we're actually interested in is what the factors are that cause deletion requests, not just the ones that correlate with them, and the numerical value you've calculated doesn't give us much insight.

So maybe you could talk more about what you think the calculated p-value actually means, and what we can do with it?

From a Bayesian point of view randomness is always merely a lack of knowledge. In that sense there's no problem with viewing the outcome of a deterministic process as something random.

Viewed from that perspective what this p-value means is that merely knowing the proportion of articles that gets marked for deletion is already enough to conclude that it is likely (or at least not impossible) for a group of 21 articles to have at least 2 articles marked for deletion. In that sense the knowledge that of the 21 articles written on women, 2 were marked for deletion doesn't really tell us something that we didn't already know.

Of course that doesn't mean we've exhausted all available knowledge, but you're going to have to look further than just how many articles get deleted.

The prior position staked on p-values makes the posterior claim of a Bayesian basis devilishly improbable.

I find it somehow fitting that the women in question would be very likely to tear apart this claim in the same way.

OT: How did you compute this p-value directly? I have never seen this approach and would be really grateful if you could point me towards a source for this approach.

I just used the binomial distribution with n=12, p=12/262 (p = number of articles with deletion attempts / total number of articles). Then I just calculated the probabilities of 0 and 1 deletions and subtracted those from 1 to get the probability of 2 or more deletion attempts.

great, thanks!

Although it will give you more or less reasonable results, I don't see a null hypothesis for which this approach does not ultimately become flawed: Assuming that you are testing the hypothesis that the two proportions agree, then, for instance, using a fixed binomial parameter of 12/262 becomes wrong for any number of deletion attempts different from 2; that is, part of the assumed model includes the outcomes you are testing: The case "3 deletions", which is part of the "2 or more extreme" probability you calculate, would have led to a different value of $p$.

Various statistical tests exist for taking this into account. With Fisher's exact test [0], we find a $p$-value close to yours:

    In [6]: from scipy.stats import fisher_exact, binom

    In [7]: binom(12, 12/262).sf(1)
    Out[7]: 0.10211478280975775

    In [8]: fisher_exact([[2, 10], [10, 240]])[1]
    Out[8]: 0.098322319292822702

However, had we not had such a large male sample size, the values might differ greatly; say, for instance, that we had had 3 deletion attempts out of 33 articles on males instead. Then we find

    In [12]: binom(12, 5/45).sf(1)
    Out[12]: 0.39171131326925679

    In [13]: fisher_exact([[2, 10], [3, 30]])[1]
    Out[13]: 0.59808767522891315

[0]: https://en.wikipedia.org/wiki/Fisher's_exact_test

That's fair, although I believe my method is justified precisely because there was such a large male sample size.

I suppose I could have made a more conservative estimate by not using the female samples in the calculation of 'p'.

I also don't quite like the way Fisher's test fixes the total number of deleted articles, but I guess a more sophisticated test assuming only the total number of female and male articles to be fixed could get a bit complicated.

Reply on news.ycombinator.com