0
$\begingroup$

A quick question to check some reasoning: consider the chi-square test for a 2x2 table. For sake of argument, there's 500 patients in each group / category (male and female) and 10% of the total have some condition - if that consists of 60 men and 40 women with the illness, a chi-squared test gives a statistic of 4.444, significant at $\alpha < 0.05$ indicating a difference between men and women. Let's consider what happens if this disease were Poisson distributed, with expected value $\lambda = 50$ cases per group. Then we can simulate the probability of finding $k$ cases in a subset of people, and it looks like this:

enter image description here

The reason this is throwing me a little is that the areas under the tails are more than I'd expect at $\alpha = 0.05$; for example, a chi-square test suggests you'd find significance if men had $k\leq 40$ and $k\geq 59$ cases, but the sum of the area under the curve between $0-40$ and $59-100$ seems quite substantial (quick calculation suggests one has about ~18% chance of lying in this region). I'm wondering if the reason for this is that I can't just consider one group in isolation, and have to multiply probability - for example, if I see 60 cases in men, I must see 40 cases in women? Or is there some reason Poisson assumptions (or similar) don't hold here? Any insight welcome, just to make sure I properly understand what's going on....

$\endgroup$
1
$\begingroup$

Correct, you cannot just look at one group in isolation.

If the number of illnesses in men (and women) is Poisson with mean 50, then there is about 9% chance of observing 60 or more illnesses among men. There is also about a 9% chance that there will be 40 or fewer illnesses among women. It is very rare that both of these will happen (about 0.8%). There is also 0.8% chance of 60 or more women and 40 or fewer men. So, those add up to 1.6% so far. But, there are many other combinations that would result in a significant chi-square test; e.g. 70 vs. 48, etc. When you add up all those, you will get close to 5%. If you know the formula for the chi-square statistic, then for each x (illness in men) between 0 and 500 find all the corresponding values of y (illness in women) that would make the statistic significant. Then add the probabilities of all those events P[X=x and Y=y] from x=0,...,500.

$\endgroup$
1
  • $\begingroup$ Perfect, thanks! $\endgroup$ – DRG Feb 19 at 17:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.