1
$\begingroup$

Say I have a set $S = \{x_1, \dots, x_m\}$, where $1 \le x_i \le n$, all distinct, with median $M$. I take a sample $T$ of size $t$ from $S$, with replacement. I call $Y$ the median of $T$. What is the probability distribution of $Y$?

Specifically, I'm trying to argue that $t$ needs to be large (in particular, not in $o(n)$) for $P[M - \frac{n}{10} \le Y \le M + \frac{n}{10}] \ge \frac{2}{3}$ (i.e. for it to be a 10% approximation to the value of the median). This leads me to wonder about $F_Y(y) = P[Y \le y]$. Full disclosure: This last bit is my homework, so you don't need to tell me the answer to that :p

Intuitively, $Y$ will be with high probability something whose rank is near $\frac{m}{2}$, but its value can be arbitrarily far off from $M$, I'd think. Experimentally, with $m = n = t$, the probability above is smaller than $\frac{1}{2}$ for $n \in \{1, \dots, 8\}$. I tried to show this by running the above algorithm on a given set $S$, then modifying $S$ by a monotonic increasing function, arguing that the result should be the same if I run the algorithm again, and seeing that the actual value $Y$ can now be much more spread out than before (thus not meeting the bound), but handling $Y$ is complicated.

I found this paper which explicitly gives the CDF for $Y$ on page 23, but it seems awfully hairy. Could it be simplified for the case $n = m = t$, perhaps?

$\endgroup$

1 Answer 1

2
$\begingroup$

For simplicity let's suppose $n=m=t = 2k-1$ is odd, so if $x_1 < \ldots < x_m$, the median $M$ is $x_k$. Your sample median $Y \le x_j$ iff at least $k$ of your sample values are $\le x_j$. The number of values $\le x_j$ in your sample is a binomial random variable with parameters $n$ and $p = j/n$, i.e. $$ P\left[Y \le x_j\right] = \sum_{i=k}^n {n \choose i} p^i (1 - p)^{n-i} $$ Using the normal approximation to the binomial, $$ P\left[Y \le x_j \right] \approx 1 - \Phi \left(\dfrac{k - j}{\sqrt{j - j^2/n}}\right)$$ where $\Phi$ is the standard normal CDF.

$\endgroup$

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.