It's a somewhat complicated situation and sorry about my phrasing, but it's my first time here. Suppose I have random normal variable $X$ ~ $N( \mu, \sigma^2)$, which represents some true effect(s). Then I try to estimate the mean of this effects using the estimates from a few (sometimes underpowered) trials -> Meta-Analysis, and do the following transformation:

$Y$ = $c/X^2$, where $c$ is a constant $> 0$ and $Y$ must be $> 0$ and therefore x>0. As a side note, $\sigma$ is for example 90% of $\mu$ and divided by the 97.5%-quantile of the standard normal distribution. What distribution has $Y$? But most importantly how can I determine/estimate the central tendency of a sample from this resulting distribution?

What I found out til now, just squaring that $X$ would result in a non-central chi-square distribution with a non centrality parameter and then just do the inverse (maybe?) and still not knowing the central tendency. Also with that 90%-$\sigma$, I have some x<0 and with very small x I get a super long tail (which doesn't reflect any reality anymore, it's just a theoretical number). Tested with $\mu=6$ and $\sigma=2.76$, c=11128.6 and 3 trials 99%,80% and 10% powered depending on the trial size in a simulation with n=10000.

I worked with the Median, but it doesn't seem to catch the center well.


1 Answer 1


The description of your experiment is somewhat unclear, so I will limit to answer the question about distributions.

With your notation, $X/\sigma \sim N(\mu/\sigma,1)$ so the square of that, $(X/\sigma)^2$ has a noncentral chisquare distribution with 1 df (degree of freedom) and noncentrality parameter $\lambda=(\mu/\sigma)^2$.

Then we can write $Y=c/X^2 = \frac{c/\sigma^2}{(X/\sigma)^2}$ so it is distributed like a constant ($c/\sigma^2$) times an inverse chi-squared rv (random variable) with 1 df and the noncentrality parameter $\lambda$.

You will find implementations of the noncentral chi-squared many places, so it will be useful to express the density (cdf, ...) for the inverse noncentral chi-square using that. Let $U$ be a nonnegative rv and $t>0$, the density of $U$ is $f$. Then we find the density of $1/U$ as $f_{1/U}(t)=f_U(1/t)/t^2$ so you can apply that.

In R there is an implementation (on CRAN) in package invgamma. But it is very easy to write yourself:

dinvchisq  <-  function(x,df,ncp=0,log=FALSE) {
    res <- dchisq(1/x,df,ncp,log=TRUE) - 2*log(x)
    if(log) res else exp(res)
  • $\begingroup$ Could you be more specific about which "inverse gamma distribution" this might be a special case of? Exactly how would one match the noncentrality parameter with the shape and scale parameters of the Inverse-Gamma? (It doesn't look possible.) $\endgroup$
    – whuber
    Commented Feb 2, 2019 at 18:51
  • 1
    $\begingroup$ You are right, I remove that reference. $\endgroup$ Commented Feb 11, 2019 at 13:59

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.