0
$\begingroup$

When deriving that there is an optimal criterion in the Signal Detection Theory literature that maximizes Proportion Correct (PC), we usually arrive at the following expression by setting the first derivative of PC with respect to $\lambda$ to zero. In other words:

From :

$ PC = s Z(d'-\lambda) + (1-s) Z(\lambda)$

To :

$ \frac{\partial PC}{\partial \lambda} = -s z(d'-\lambda)+(1-s) z (\lambda) = 0 $

by setting $\frac{\partial PC}{\partial \lambda} = 0$ where $z$ is the pdf of a univariate zero mean gaussian with $(\mu,\sigma^2)=(0,1)$, and $Z$ is its respective cdf. $s$ is the proportion of target present trials, so $s$ and $1-s$ are both bounded $s\in[0,1]$.

The problem now is that I can't seem to analytically prove that a maximum exists by finding the second derivative of the above expression and showing that within a certain range it is negative (thus showing there is a maximum via concavity). I know that the optimal value is (but can't seem to prove that it is a maximum (vs a minimum):

$\lambda^* = \frac{d'}{2} - \frac{1}{d'}\log(\frac{s}{1-s})$

I computed the second derivative and am stuck at the following equation, without knowing how to show that it will be negative within a certain domain of $\lambda$:

$\frac{\partial^2 PC}{\partial \lambda^2} = -\frac{(d'-\lambda)^3}{2}\text{exp}(-(1/2)(d'-\lambda)^2)+\frac{\lambda^3}{2}\text{exp}(-(1/2)\lambda^2)$

Question: For what values of $\lambda$ is $\frac{\partial^2 PC}{\partial \lambda^2} < 0$ ?

$\endgroup$
2
  • 1
    $\begingroup$ There's a sign error in your calculation of the first derivative: if that's not just a typo, it will propagate into errors later--and appears to have done so. $\endgroup$
    – whuber
    Commented Aug 17, 2018 at 18:08
  • $\begingroup$ @whuber sign has been corrected, thanks for catching the mistake! $\endgroup$
    – Mecasickle
    Commented Aug 17, 2018 at 22:04

1 Answer 1

0
$\begingroup$

After correcting for a change of sign addressed in the comments, when we compute the second derivative we get:

$\frac{\partial^2 PC}{\partial \lambda^2} = (-s)z(d'-\lambda)(d'-\lambda)-(1-s)z(\lambda)\lambda$

From here it follows that $\frac{\partial^2 PC}{\partial \lambda^2}<0$ only if $d'>\lambda$ as all the terms in the expression above are positive, which is equivalent to the case in which the Hit Rate is greater than 0.5 or 50%. Recall that $d'=Z^{-1}(HR) + \lambda$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.