4
$\begingroup$

Suppose I have an urn with N marbles, with frequencies p and q for red and black marbles, and with p > 0,5. I take a sample of r marbles.

It sounds intuitive to say that deviations from the mean should be expected to occur in the direction of the more frequent marble (say for example, if I have 90 red and 10 black, then with r = 10, 10 red should be more likely than 8 red, 2 black). That is, that $Prob(\hat{p} > p) > Prob(\hat{p} < p)$. Or equivalently, $Prob(\hat{p} \ge p) > Prob(\hat{p} \le p)$.

But that is not always the case. Take p = 0,51 and r = 2, for example. What's going on here is that the expected value is not a possible sample value, but it is close to it, so $\hat{p}=0,5$ will count as a case where $\hat{p} < p$. So let's get rid of this sort of cases by taking as an additional premise that $p .r \in \mathbb{Z}$.

¿Will the result hold in this case?

What would be needed to answer positively is that

  • For sampling with replacement: $$\sum_{k=p.r}^{r} \binom{r}{k} p^k (1-p)^{r-k} > \sum_{k=0}^{p.r} \binom{r}{k} p^k (1-p)^{r-k}$$

  • For sampling without replacement: $$\frac{\binom{p.N}{p.r} \binom{N - p.r}{r - p.r}}{N \choose r} > \frac{\binom{(1-p).N}{(1-p).r} \binom{N - (1-p).r}{r - (1-p).r}}{N \choose r}$$ (could also use a cumulative hypergeometric distribution but this seems easier)

$\endgroup$

2 Answers 2

3
$\begingroup$

Here's a positive answer for the binomial distribution in the case $p=1-\frac1{n}$. (I suppose you could check some other cases like $p=1-\frac2n$ in a similar way.)


We need to prove $$\binom{n}{n}p^n > \sum_{k=0}^{n-2}\binom{n}k p^k(1-p)^{n-k},$$ $$\left(1-\frac1n\right)^n > \sum_{k=0}^{n-2}\binom{n}k \left(1-\frac1n\right)^k\left(\frac1n\right)^{n-k},$$ $$\left(n-1\right)^n > \sum_{k=0}^{n-2}\binom{n}k \left(n-1\right)^k.$$ The last right-hand side is actually, per Wolfram Alpha $$ \sum_{k=0}^{n-2} (n-1)^k \binom{n}{k} = n^n-n (n-1)^{n-1}-(n-1)^n$$ so we need $$n^n<2(n-1)^n+n(n-1)^{n-1}$$ $$\left(1+\frac1{n-1}\right)^n<2+\frac{n}{n-1} = 3 + \frac1{n-1}$$ or, with $m=n-1$, $$\left(1+\frac1m\right)^m < \frac{3+\frac1m}{1+\frac1m} = \frac{3m+1}{m+1} = 3-\frac2{m+1}=3-\frac2n$$ This is true: the left-hand side is bounded above by $e$, and the right-hand side surpasses $e$ by $m=7$, and it can also be checked for $m\in\{1,\dots,6\}$.

$\endgroup$
2
  • $\begingroup$ I decided it was better to remove my old comments and instead give a partial answer to your actual question :) $\endgroup$ Mar 30, 2015 at 1:24
  • 1
    $\begingroup$ Awesome! thanks a lot! Just a tiny correction though: in the last line, for m = 1 (n = 2) both sides are equal, as should be expected, since if n = 2 then p = 1 - 1/n = 1/2. Now I just have to see how to use this to get the general result. I'll report if I find anything :) $\endgroup$
    – Cartucho
    Mar 30, 2015 at 14:20
2
$\begingroup$

It's not true for sampling without replacement. Consider e.g. $N=9$, $p = 2/3$, $r = 6$. Note that the possible numbers of red balls in the sample are $3,4,5,6$, with $4$ corresponding to $\widehat{p} = p$. Since $P(3) = P(5) + 2 P(6)$, we have $P(\widehat{p} > p) = P(5) + P(6) < P(3)$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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