Consider two independent random samples of sizes $n_1$ and $n_2$ ($n_1\neq n_2$ may be the case) on independent random variables $x_1$ and $x_2$. That is, we have one iid sample of size $n_1$ from the distribution that $x_1$ follows, another iid sample of size $n_2$ from the distribution that $x_2$ follows, and these two samples were independently taken. Assume that $x_1$ and $x_2$ have finite fourth moments.

Let $\overline{x}_1=n_j^{-1}\sum_{i=1}^{n_j}x_j$ and $s^2_j=(n_j-1)^{-1}\sum_{i=1}^{n_j}(x_{ji}-\overline{x_j})^2$ for $j=1,2$. Define the test statistic $$t=\frac{\overline{x}_1-\overline{x_2}}{s}$$ where $s=\sqrt{\frac{s^2_1}{n_1}+\frac{s^2_2}{n_2}}$.

Under the null hypothesis $\mathbb{H}_0:\mathbb{E}(x_1)=\mathbb{E}(x_2)$, is $t$ asymptotically standard normal as $n_1$ and $n_2 $ diverges to infinity? If not, when is $t$ asymptotically standard normal? Is there another test statistic that has an asymptotic normal distribution under the null?

I note that $t$ is a function of $\{x_{1i}\}_{i=1}^{n_1}\cup\{x_{2i}\}_{i=1}^{n_2}$.

Compare with the article "Student's t-test" on Wikipedia.

  • 1
    $\begingroup$ This is essentially the Behrens-Fisher problem, en.wikipedia.org/wiki/Behrens-Fisher_problem $\endgroup$ – Alecos Papadopoulos Feb 24 at 20:41
  • $\begingroup$ @AlecosPapadopoulos (+1) But in the Behrens–Fisher problem we do not consider the case where $n_1\to\infty$ and $n_2\to\infty$. Right? Are you arguing that my problem reduces to the Behrens-Fisher problem? $\endgroup$ – Elias Feb 24 at 20:46
  • $\begingroup$ No we don't, but I wanted you to concentrate on that point exactly... what happens to the numerator and to the denominator of the $t$ statistic here as the sample sizes go to infinity under the null? $\endgroup$ – Alecos Papadopoulos Feb 25 at 3:10
  • $\begingroup$ @AlecosPapadopoulos They approach $\mathbb{E}(x_1)-\mathbb{E}(x_2)=0$ and 0. I don't think that is a particularly interesting question though. $t$ has an asymptotic normal distribution if $n_1=n_2$, for example, and so there is an answer to this question. What is your point exactly? $\endgroup$ – Elias Feb 25 at 7:30

Consider $$W \equiv \frac {\sqrt{n_2}}{n_1} \sum_{i=1}^{n_1}X_{1i} - \frac 1{\sqrt{n_2} }\sum_{i=1}^{n_2}X_{2i}$$

Let $v_1, v_2$ denote the true variances of the two distributions. Given the various assumptions and under the null we have

$$E(W) = 0, \text{Var}(W) = \frac{n_2}{n_1}v_1 + v_2$$

Assume in addition that $n_2/n_1 \to c < \infty,\;\;\; n_1, n_2 \to \infty$

Again given the assumptions, as $n_1, n_2 \to \infty$

$$\frac {W}{\sqrt{\frac{n_2}{n_1}v_1+v_2}} \to_p \frac {W}{\sqrt{cv_1+v_2}}\to_d N(0,1)$$

More over, by the assumptions and Slutsky's theorem

$$\frac {W}{\sqrt{\frac{n_2}{n_1}s_1+s_2}} - \frac {W}{\sqrt{cv_1+v_2}} \to_p 0$$

| cite | improve this answer | |
  • 1
    $\begingroup$ Since $E(W)=(\sqrt{n}_1-\sqrt{n}_2)E(x_1)$, this only works if $n_1=n_2$. One might restrict the sample to the first $\min(n_1,n_2)$ observations. I wonder how this relates to the asymptotic distribution of $t$. $\endgroup$ – Elias Feb 27 at 9:50

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.