1
$\begingroup$

Neyman-Pearson Lemma says that the most powerful test, $\phi(x)$, of size $\alpha$ (probability of false alarm), for testing $H_0:\theta=\theta_0$ versus $H_1:\theta=\theta_1$ is the likelihood ratio test of the form:

$$ \phi(x)=\begin{cases} 1 & l(x)>k \\ \color{#C00}{p} & l(x)=k \\ 0 & l(x)<k \\ \end{cases} $$

where, $p$ is the probability of accepting $H_1$ at $k.$

$k$ is threshold.

$l(x)$ is the likelihood ratio, $l(x)=\frac{f_{\theta_1}(x)}{f_{\theta_0}(x)}.$

Probability of false alarm, $P_f=\color{#C00}{p} \cdot P[l(x)=k \hspace{2 mm}| \hspace{2 mm} H_0 ]\hspace{2 mm}+\hspace{2 mm} P[l(x)>k \hspace{2 mm}| \hspace{2 mm} H_0 ].$

Probability of detection, $P_d=p \cdot P[l(x)=k \hspace{2 mm}| \hspace{2 mm} H_1 ]\hspace{2 mm}+\hspace{2 mm} P[l(x)>k \hspace{2 mm}| \hspace{2 mm} H_1 ].$

Question Why Probability of false alarm, $P_f$, is $\mathbf{not}$ defined as:

$P_f=\color{#C00}{(1-p)} \cdot P[l(x)=k \hspace{2 mm}| \hspace{2 mm} H_0 ]\hspace{2 mm}+\hspace{2 mm} P[l(x)>k \hspace{2 mm}| \hspace{2 mm} H_0 ].$

My understanding for saying above equation is that: if the probability of occurring $H_1$ is $p$, then the probability of occurring $H_0$ is $(1-p).$ And therefore, the probability of selecting $H_0$ at $k$ is $(1-p).P[l(x)=k \hspace{2 mm}| \hspace{2 mm} H_0].$

Where my understanding is wrong? Any comments please.

$\endgroup$
3
$\begingroup$

I am not sure what you mean by 'probability of occurring $H_0$', but that might be where you are confused. Note that $H_1$ and $H_0$ are not treated as random events and thus they have no probability. Instead, we assume either that $H_1$ is true or that $H_0$ is true, and compute probabilities of random events given $H_0$ or given $H_1$.

False alarm is defined as rejecting $H_0$ (accepting $H_1$) while $H_0$ is actually true. 'Probability of false alarm' in this context means the probability of false alarm given that $H_0$ is true. That is, \begin{equation} P[\textrm{We accept }H_1 \mid H_0]. \end{equation} We can divide this probability into 3 separate events depending on whether $l(x)<k,l(x)=k,l(x)>k$: \begin{equation} =P[\textrm{We accept }H_1 \cap l(x)<k \mid H_0] + P[\textrm{We accept }H_1 \cap l(x)=k \mid H_0] + P[\textrm{We accept }H_1 \cap l(x)>k \mid H_0] \end{equation} The first term is obviously $0$ as we never accept $H_1$ when $l(x)<k$. The third term is just $P[l(x)>k \mid H_0]$ as $H_0$ is always accepted when $l(x)>k$. For the second term, \begin{equation} P[\textrm{We accept }H_1 \cap l(x) = k \mid H_0] = P[\textrm{We accept }H_1 \mid l(x)=k, H_0]P[l(x)=k \mid H_0] = pP[l(x)=k \mid H_0] \end{equation} that is, for the event in question to occur, both $l(x)=k$ and acceptance of $H_1$ must occur. The probability of this is computed by multiplying the (marginal) probability of $l(x)=k$ by the conditional probability of accepting $H_1$ conditional on $l(x)=k$.

Combining the three terms, we get the probability of false alarm \begin{equation} = pP[l(x)=k \mid H_0] + P[l(x)>k \mid H_0]. \end{equation}

Note that if one treated $H_0$ as a random event, one could compute the marginal probability of false alarm, that is, to have false alarm, we need both $H_0$ to be true and $H_1$ to be accepted. This would result in multiplying the aforementioned probability by $P[H_0]$. But, outside Bayesian framework $H_0$ is not treated as a random event, and thus only probabilities of accepting $H_1$ that can be defined are the ones given $H_0$ or given $H_1$.

$\endgroup$

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.