Let $X_1$ and $X_2$ be binomial random variables with respective parameters $n_1, p_1$ and $n_2, p_2$. Show that when $n_1$ and $n_2$ are large, an approximate level $\alpha$ test of $H_0 : p_1 = p_2$ versus $H_1 : p_1 \neq p_2$ is as follows, reject $H_0$ if $$ \frac{|X_1/n_1-X_2/n_2|}{\sqrt{\frac{X_1+X_2}{n_1+n_2} \left( 1 - \frac{X_1+X_2}{n_1+n_2}\right) \left(\frac{1}{n_1}+\frac{1}{n_2}\right)}} > z_{\alpha/2} $$ Hint below in screenshot.

My attempt

Given $H_0$ is true I can say that $p = p_1 = p_2$ and since $n_1$ and $n_2$ are large, I can use the normal approximation to the binomial to say that

$$ V = \frac{X_1 - n_1 p}{\sqrt{n_1 p q}} = \frac{ \frac{X_1}{n_1} - p}{\sqrt{\frac{p q}{n_1}}} \, \dot\sim \, N(0,1)$$

$$ W = \frac{X_2 - n_2 p}{\sqrt{n_2 p q}} = \frac{ \frac{X_2}{n_2} - p}{\sqrt{\frac{p q}{n_2}}} \, \dot\sim \, N(0,1)$$

Then we have $\frac{V-W}{\sqrt{2}} \dot\sim N(0,1)$ and we can build a two sided hypothesis using the fact that

$$P \left( -z_{\alpha/2} \le \frac{V-W}{\sqrt{2}} \le z_{\alpha/2} \right) = 1-\alpha $$

The book answer seems superior because you don't need to know $p_1$ or $p_2$. However I'm having difficulty getting rid of those two values. Thanks for your help and patience!

Book problem with hint

enter image description here


Define $\hat p_i=X_i/n_i$ as the observed binomial proportion, $i=1,2$.

Since $n_1,n_2$ are large, by CLT $$\frac{\sqrt{n_i}(\hat p_i-p_i)}{\sqrt{p_i(1-p_i)}}\stackrel{L}\longrightarrow N(0,1)\quad,\,i=1,2$$

For a formal derivation of the result, suppose $n=n_1+n_2$ and that $\min(n_1,n_2)\to\infty$ such that $n_1/n\to\lambda \in(0,1)$ (which implies $n_2/n\to1-\lambda$). Hence assuming $X_1$ and $X_2$ are independent,

$$\frac{\sqrt{n}\left((\hat p_1-\hat p_2)-(p_1-p_2)\right)}{\sqrt{\frac{p_1(1-p_1)}{\lambda}+\frac{p_2(1-p_2)}{1-\lambda}}}\stackrel{L}\longrightarrow N(0,1)$$

If $p$ is the common value of $p_1$ and $p_2$ under $H_0$, then

$$\frac{\sqrt{n}(\hat p_1-\hat p_2)}{\sqrt{p(1-p)\left(\frac{1}{\lambda}+\frac{1}{1-\lambda}\right)}}\stackrel{L}\longrightarrow N(0,1)\tag{1}$$

Let $\hat\lambda=n_1/n$ and define $$\hat p=\hat\lambda \hat p_1+(1-\hat\lambda)\hat p_2=\frac{1}{n}(X_1+X_2)$$

Now note that $$\hat p\stackrel{P}\longrightarrow p$$

Therefore, $$\frac{1}{\sqrt{\hat p(1-\hat p)}}\stackrel{P}\longrightarrow\frac{1}{\sqrt{p(1-p)}}\qquad,\,\hat p\ne 0,1$$

Or, $$\frac{\sqrt{p(1-p)}}{\sqrt{\hat p(1-\hat p)}}\stackrel{P}\longrightarrow 1\tag{2}$$

Applying Slutsky's theorem on $(1)$ and $(2)$, we get under $H_0$,

$$\frac{\sqrt{n}(\hat p_1-\hat p_2)}{\sqrt{p(1-p)\left(\frac{1}{\lambda}+\frac{1}{1-\lambda}\right)}}\times \frac{\sqrt{p(1-p)}}{\sqrt{\hat p(1-\hat p)}}\stackrel{L}\longrightarrow N(0,1)$$

That is, the test statistic under $H_0$ is given by $$\color{blue}{T=\frac{\sqrt{n}(\hat p_1-\hat p_2)}{\sqrt{\hat p(1-\hat p)\left(\frac{1}{\hat\lambda}+\frac{1}{1-\hat\lambda}\right)}}\stackrel{L}\longrightarrow N(0,1)}$$

(The expression above is the same as the one given in your question.)

We reject $H_0$ approximately at level $\alpha$ if $|\text{observed }T|>z_{\alpha/2}$.

  • $\begingroup$ Excellent answer, thank you. $\endgroup$ – HJ_beginner May 22 at 21:57
  • $\begingroup$ Quick notation question - what does the L over the arrow stand for? Does that mean in the limit as n goes to infinity the distribution approaches the same distribution as as unit normal? Something like that? Thanks. $\endgroup$ – HJ_beginner May 24 at 18:34
  • 1
    $\begingroup$ Yes, $L$ denotes convergence in law/convergence in distribution. $\endgroup$ – StubbornAtom May 24 at 18:37
  • $\begingroup$ Thanks for confirming. You rock! $\endgroup$ – HJ_beginner May 24 at 19:06

Interpret $X_i$ as a sum of $n_i$ Bernoulli rv's with parameter $p_i$ so the LLN shows that $\frac {X_i}{n_i} \approx p_i$ for large $n_i$. Now the CLT gives the hint and the denominator of the test statistic can be found by substituting $X_i/n_i=p_i=p$.

  • $\begingroup$ Ah, thank you that's a good point, another implication of $n_1$ and $n_2$ large is you can use the LLN for $p_i$ $\endgroup$ – HJ_beginner May 22 at 2:42

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.