2
$\begingroup$

In the Wikipedia page on the Neyman-Pearson Lemma, they discuss the example where there are $n$ iid Normal RVs with a known mean and a population variance that is the null hypothesis, and another that is the alternative hypothesis. They get the test statistic $\Sigma (x_i - \mu)^2$

They then jump to saying "we should reject $H_0$ if $\Sigma (x_i - \mu)^2$ is sufficiently large and that we can get the critical value by scaling the test-statistic to become a chi-square distribution, they never explain why.

Various sites online explain that basically converting the test statistic to become the sample variance over the population variance will turn the formula into a chi-square distribution. They do it as follows:

$\chi^2_{n-1} = \frac{(n-1)\frac{\Sigma (x_i - \mu)^2}{n-1}}{\sigma_0^2}$

$\chi^2_{n-1} = \frac{(n-1)s^2}{\sigma_0^2}$

But then I am confused as to how this would relate to a significance value $\alpha$. Like if it is $\alpha = 0.05$, how does this translate back into our original problem of "debunking/rejecting the null hypothesis"

Not looking for the answer, I just want some help explaining why the test statistic is the way it is.

| cite | improve this question | |
$\endgroup$
2
$\begingroup$

Neyman-Pearson says reject for $Q = \sum_i (X_i - \mu)^2 > c,$ with $c$ chosen so that $P(Q > c) = \alpha$ under $H_0.$ Then the problem is to find the numerical value of $c.$

Under $H_0,$ we have that $Z_i = (X_i - \mu)/\sigma_0$ is standard normal so that $Z_i^2\sim \mathsf{Chisq}(1).$ Thus $Q/\sigma_0^2 \sim\mathsf{Chisq}(n).$ We can use printed chi-squared tables or software to find the number $b$ with $P(Q/\sigma_0^2 > b) = \alpha.$ What is $c$ in terms of $b?$

Addendum: Below we simulate a million samples, each with $n = 5$ observations from $\mathsf{Norm}(100, 15).$

In the histogram at left we show that a histogram of simulated values of $(n-1)S^2/\sigma^2$ (called dist.1) fits $\mathsf{Chisq(4)}$ (solid curve), but not $\mathsf{Chisq(5)},$ as mentioned in a Comment. Informally, one says that one degree of freedom has been 'lost' estimating unknown $\mu.$

By contrast, the histogram at right illustrates that simulated values of $nQ/\sigma^2$ (called 'dist2`) do fit $\mathsf{Chisq}(5),$ as in the discussion above.

enter image description here

R code for the simulation is shown below (but not the code to make the histograms).

set.seed(409)  # for reproducibility
m = 10^6;  n = 5;  mu = 100;  sg = 15
x = rnorm(m *n, mu, sg);  MAT = matrix(x, nrow=m)
s.sq = apply(MAT, 1, var)
dist.1 = (n-1)*s.sq/sg^2 col="brown", lwd=2, lty="dotted")
q = rowSums((MAT - mu)^2)
dist.2 = q/sg^2
| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ If I understand you correctly, stating that $P(Q / \sigma_0^2 > b)$ would mean that $c=\frac{b}{\sigma_0^2}$, and that $Q/\sigma_0^2 = \chi_n^2$ but then why do the websites convert our Q into the sample standard deviation and have $\chi_{n-1}^2 = \frac{(n-1)\frac{Q}{(n-1)}}{\sigma_0^2} $ ? $\endgroup$ – QuantumHoneybees Apr 8 '19 at 21:20
  • 1
    $\begingroup$ Your 1st sentence fragment is OK. Addressing the rest: When $\mu$ is known [as I surmised from your $\sum_i (X_I - \mu)^2],$ you have $Q/\sigma_0^2 \sim \mathsf{Chisq}(n).$ When $\mu$ is unknown and estimated by $\bar X,$ then $(n-1)S^2/\sigma_0^2 \sim \mathsf{Chisq}(n-1),$ which is somewhat more difficult to prove. // Just as inference on $\mu$ uses either z or t test depending on whether $\sigma$ is known or not, respectively; inference on $\sigma$ is slightly different depending on whether $\mu$ is known or not. $\endgroup$ – BruceET Apr 8 '19 at 22:35
  • $\begingroup$ So $\mu$ is known, you are correct. Second, if it is not known, is it ok to use the estimator? I've studied estimators, and I know why the formula works, I guess my question is: "is it fine to use an estimator for hypothesis testing"? and why would that work, intuitively? $\endgroup$ – QuantumHoneybees Apr 8 '19 at 22:40
  • 1
    $\begingroup$ If $\mu$ known ok to use $Q$ as discussed in my Answ. If $\mu$ not known must use $S^2$ as discussed in my previous Comment. Two separate situations. Don't mix them. $\endgroup$ – BruceET Apr 8 '19 at 22:43
  • 1
    $\begingroup$ The dist'n of $(n-1)S^2/\sigma^2$ is not trivial to derive, but it is easy to simulate; see Addendum. $\endgroup$ – BruceET Apr 10 '19 at 1:50

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.