I want to construct an NP-Test for simple Null- and Alternative-hypothesis. In particular the likelihood for the Nullhypothesis and Alternativehypothesis is given by $$h_0(x)=\chi_{(-2,2)}(x)\\h_1(x)=\frac{1}{\sqrt{2\pi}}\exp(-\frac{1}{2}x^2)$$ respectively. This means I want to test by considering a sample, if the underlying distribution is uniform on $(-2,2)$ or (standard)normally distributed.

First I construct the Likelihoodquotient $T(x)=\frac{h_1(x)}{h_0(x)}$ which yields $$T(x)= \begin{cases} \frac{1}{\sqrt{2\pi}}\exp(-\frac{1}{2}x^2),\ -2\le x\le 2\\ \infty, \quad\quad\quad\quad\quad \ \quad\quad\quad \text{else} \end{cases}$$

In the examples I have seen so far it is used that $T$ is monotonic, because then I can just refer the the distribution of the data under the Nullhypothesis. But in this case I have no idea on how to proceed.

I somehow have to compute a $\gamma\in[0,1]$ and $k\in[0,\infty)$ such that $$\alpha=P_0(T>k)+\gamma P_0(T=k)$$ where $P_0$ is the probability distribution corresponding to $h_0$ and $\alpha$ is the significance level.


1) The density of your uniform in (-2,2) is 1/4, not 1. This has an impact on your $T$ and $k$.

2) You can forget about $\gamma$, because both distributions are continuous, so $P_0(T=k)=0$.

3) This is somewhat nonstandard because $T$ indicates that $h_1$ is better outside $(-2,2)$, $h_0$ is better in a set of the shape $(-2,-a)\cup(a,2)$, and $h_1$ is better in $(-a,a)$, because of symmetry (you can make a drawing to see this). You need to find $a$ so that $P_0[(-2,-a)\cup(a,2)]=1-\alpha$, which is easy. If you have that, you don't even need $k$, or you could read it off easily.

  • $\begingroup$ I like this answer! Could you elaborate on what you mean by "better"? How would I reach your conclusion from merely considering $\alpha=P_0(T>k)$? $\endgroup$ – EpsilonDelta Jan 7 '20 at 20:06
  • $\begingroup$ "$h_1$ is better" means that $T=h_1/h_0$ is rather large (or, in other words, $h_1$ is relatively large compared to $h_0$, where the word "relatively" is connected to the fact that what values of $T=h_1/h_0$ are taken as "large" depends on $\alpha$ through $k$). If you start thinking from "$\alpha=P_0(T>k)$" you realise that you must choose $k$ in such a way that the set in which $T$ is large enough has probability $\alpha$, or equivalently probability $1-\alpha$ where $T$ is small. This is what I did, I chose a region in which $T$ is small so that its probability is $1-\alpha$. $\endgroup$ – Lewian Jan 8 '20 at 16:29
  • $\begingroup$ I then realised that this is easier and clearer in terms of $a$ than in terms of $k$. If you have $a$ it isn't difficult to find $k$, but in order to define and run the test, $k$ isn't even needed if you know $a$. $\endgroup$ – Lewian Jan 8 '20 at 16:31

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.