1
$\begingroup$

I am trying to find a way to calculate the mean of my normal distribution if I know a certain percentile. Sounds simple, but I'm really stuck.

The purpose is to calculate a limit of detection (LD) based on Lloyd Currie's work on the issue and his proposal for a IUPAC definition. First, I calculate a critical value (LC) above which the chance of finding noise is less than e.g. 5%, so that I can control my false positives. Now, in order to control my false negatives, I define the LOD as the point above which a true signal will have a chance of less than e.g. 5% that an estimator falls below the LC.

Critical value LC, limit of detection LD and percentages alpha/beta for false positives/negatives http://www.intechopen.com/source/html/46058/media/image9.png

So I have a value where a certain percentage has to be reached, and in order to achieve that, I need to set my mean (LD) accordingly. (The variance is fixed.)

My question is, can I do this mathematically somehow? In Excel and IDL, the only functions I've found were either for a standard normal distribution, or they needed the mean as an input and couldn't output it. And since the integral of the normal distribution is the error function, I can't simply solve the equation for µ.

Any ideas how to tackle this problem?

$\endgroup$
3
$\begingroup$

Let $\Phi^{-1}(p)$ denote the $p$-quantile of the standard normal distribution. The normal being a location scale family, its quantiles $F^{-1}(p)$ are related to that of the standard normal by $$ F^{-1}(p)=\mu+\sigma\Phi^{-1}(p) $$ By assumption, $F^{-1}(p)$ and $\sigma$ are known, and you may compute $\Phi^{-1}(p)$ corresponding to the known $p$-quantile. Then, solve for $\mu$.

$\endgroup$
3
$\begingroup$

Recall that the cumulative distribution function (CDF) $F$ for the normal distribution with mean $\mu$ and standard deviation $\sigma$ can be written in terms of the standard normal CDF $\Phi$ as follows:

$$ F(x; \mu, \sigma) = \Phi\left( \frac{x - \mu}{\sigma} \right) $$

So you know $\sigma$, $p = F(x; \mu, \sigma)$ and want to solve for $\mu$.

$$p = \Phi\left( \frac{x - \mu}{\sigma} \right) $$

Hence:

$$ \mu = x - \sigma \Phi^{-1}(p) $$

$\endgroup$
  • $\begingroup$ Thank you! Quick question, are you missing an 1/sigma in your first formula? The result seems to be correct, but the normalisation factor is missing in the first formula (maybe just a typo or I'm an idiot)? $\endgroup$ – PoorYorick Oct 21 '16 at 12:28
  • $\begingroup$ @Fearabbit What 'normalization' factor are you thinking should be there? $\endgroup$ – Glen_b -Reinstate Monica Oct 21 '16 at 16:16
  • $\begingroup$ Do you maybe confuse the distribution function and the density? $\endgroup$ – Christoph Hanck Oct 21 '16 at 17:07
  • $\begingroup$ Yeah, as I suspected, I was confused by the integral. You can transform the CDF into the standard CDF easily by substitution, but then you get a term for the derivative, and that cancels out the $1/\sigma$. $\endgroup$ – PoorYorick Oct 25 '16 at 9:15

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.