1
$\begingroup$

For an exam preparation I'm trying to solve the following question, but I get stuck. The question is

One measure of the homogeneity of a multinomial population with $k$ cells and probabilities $p = (p_1,\cdots,p_k)$ is the sum of the squares of the probabilities $S(p) = \sum_{i=1}^k p_i^2$. Note that $\frac{1}{k}\leq S(p)\leq 1$, with higher values indicating greater heterogeneity. Given a sample of size $n$ from this population, we may estimate $S(p)$ by $S(\hat{p})$, where $\hat{p} =$ $(\hat{p}_1,\cdots,\hat{p}_k)$ and $\hat{p}_i$ is the proportion of the observations that fall in cell $i$. What is the asymptotic distribution of $S(\hat{p})$?

Following a proof for Neyman's $\chi^2$, I define random variables $X_{ij}$ such that $P(X_{ij} = 1) = p_j$ and $P(X_{ij} = 0) = 1 - p_j$. Then $EX_{ij} = p_j$ and $EX_i = p$ where $p = (p_1,\cdots,p_k)$. Furthermore the covariance matrix of $X_1,...,X_n$ is given by $\Sigma = \begin{bmatrix} p_1(1-p_1) & -p_1p_2 & \cdots & -p_1p_c\\ -p_1p_2 & p_2(1-p_2) & \cdots & -p_2p_c \\ \vdots & \vdots & & \vdots \\ -p_1p_c & -p_2p_c & \cdots & p_c(1-p_c)\end{bmatrix}$.

Then $S(\hat{p}) = \sum_{i=1}^k \begin{pmatrix} \frac{n_i}{n} \end{pmatrix}^2$ where $n_i$ denotes the number of trials that result in outcome $i$ for $i = 1,...,c$. So $S(\hat{p}) = \bar{X}_n^T\bar{X}_n$.

Starting here I get stuck trying to find an asymptotic distribution for $S(\hat{p})$. I know that $\sqrt{n}(\bar{X}_n - p) \xrightarrow{\mathcal{L}} \mathcal{N}(0, \Sigma)$, but I fail to see how I can derive the asymptotic distribution of $\bar{X}_n^T\bar{X}_n$ from here. I hope someone can help me out!

$\endgroup$
1
$\begingroup$

Use the multivariate form of the delta method for $g: \mathbb R^k \to \mathbb R$ defined by $x\mapsto x^\intercal x$. So, for $\sqrt n (\bar X_n - p)\stackrel{d}{\longrightarrow}N(0,\Sigma)$, we have $$\nabla g_{(p)}=2p\,\,\,\,\, \text{and} \,\,\,\,\, \sqrt n (\bar {X}_n^\intercal\bar X_n - p^\intercal p)\stackrel{d}{\longrightarrow}N(0,\nabla g_{(p)}^\intercal\Sigma\nabla g_{(p)})= N(0,4p^\intercal\Sigma p).$$

$\endgroup$
  • $\begingroup$ what if the vector p is uniform, you will get $p^T \Sigma p = 0$. $\endgroup$ – yuguaw Feb 4 at 7:33

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.