“Unbiased” hypothesis test — what does it mean actually? [duplicate]

Consider a statistical hypothesis test of size level $0<\alpha<1$ for testing a null hypothesis $H_0:\theta \in \Theta_0 \subset \Theta$ against an alternative hypothesis $H_1:\theta \in \Theta_1=\Theta – \Theta_0$. I know that the test is considered “unbiased” if its power function $\beta$ satisfies the condition:

$$\beta(\theta)\leq\alpha \hspace{0.3cm}\text{if}\hspace{0.3cm}\theta \in \Theta_0,$$
$$\beta(\theta)\geq\alpha \hspace{0.3cm}\text{if}\hspace{0.3cm}\theta \in \Theta_1.$$

I understand the definition but I do not understand what does this mean for the test? Does it mean that the test has more power than it should? And how do you compare the power of two tests, one “biased” and the other “unbiased”?


It means that the probability that the test rejects (its power) is always higher when the alternative is true than when the null is true.

Suppose, for example, that you use a standard t-test for the null $\theta\leq0$ against the alternative $\theta>0$. The standard rejection rule at $\alpha=0.05$ would be to reject if $t>1.645$ (for either a sample from a normal distribution or asymptotically, when a central limit theorem applies).

Now, suppose you were to use that rule (reject if $t>1.645$) to test $\theta=0$ against $\theta\neq0$. The probability that the test will reject will decrease the more negative the true $\theta$, as we shall rarely observe large positive t-ratios in that case. In particular, this test is be biased, as $\beta(\theta)<\alpha$ when $\theta\in\Theta_1\cap(-\infty,0)$.

For concreteness, we may compute this probability explicitly in the normal case, $X_i\sim N(\theta,1)$, with $\sigma^2=1$ assumed known for simplicity. Then,
the t-statistic for $\theta=0$ simply is $t=\sqrt{n}\bar{X}$ and $$\sqrt{n}(\bar{X}-\theta)\sim N(0,1)$$
which tends to 0 as $\theta\to-\infty$.

enter image description here

theta.grid <- seq(-.8,.8,by=.01)
n <- seq(10,90,by=20)
power <- 1-pnorm(qnorm(.95)-outer(theta.grid,sqrt(n),"*"))
colors <- c("#DB2828", "#40AD64", "#E0B43A", "#2A49A1", "#7A7969")
matplot(theta.grid,power, type="l", lwd=2, lty=1, col=colors)
legend("topleft", legend=paste0("n=",n), col=colors, lty=1, lwd=2)

Source : Link , Question Author : Enthusiastic , Answer Author : Christoph Hanck

Leave a Comment