2
$\begingroup$

I have two random variables X and Y which came from different inverse gaussian (IG) distributions: $$ X \sim IG(\mu_1,\lambda_1)$$ $$ Y \sim IG(\mu_2,\lambda_2)$$

I need to find a probability of X>Y. It seems, that i need to find a pdf and cdf for variable Z=X-Y

$$ p(X-Y|\mu_1,\mu_2,\lambda_1,\lambda_2)$$

Is it possible to find and analytical solution for this problem? I will need these functions for further MLE estimation.

$\endgroup$
11
  • 1
    $\begingroup$ You're likely to make more progress by evaluating $\Pr(X/Y \gt t)$ for general $t\gt 0,$ because that immediately eliminates both the shape parameters. $\endgroup$
    – whuber
    May 15 at 15:55
  • $\begingroup$ @whuber I'm clearly missing something. I get $\frac{e^{\frac{\lambda_1}{\mu_1}+\frac{\lambda_2}{\mu_2}} \sqrt{\frac{\lambda_1 \lambda_2 \left(\lambda_2 \mu_1^2+\lambda_1 \mu_2^2 t\right)}{\lambda_1+\lambda_2 t}} K_1\left(\frac{\sqrt{\frac{(\lambda_1+t \lambda_2) \left(\lambda_2 \mu_1^2+t \lambda_1 \mu_2^2\right)}{t}}}{\mu_1 \mu_2}\right)}{\pi \mu_1 \mu_2 t}$ for the pdf of $T=X/Y$ so I don't see finding 1 minus the cdf for $t=1$ eliminates the both of the shape parameters. $\endgroup$
    – JimB
    May 15 at 20:00
  • $\begingroup$ And $K_1$ is the modified Bessel function of the second kind: $K_n(z)$ with $n=1$. $\endgroup$
    – JimB
    May 15 at 20:09
  • $\begingroup$ @whuber Thank you for a nice idea. Unfortunately, my knowledge of algebra is rather poor and I couldn't figure out how I could derive a pdf and cdf for this value. I tried to use these links for this: en.wikipedia.org/wiki/Ratio_distribution; mathworld.wolfram.com/RatioDistribution.html; math.stackexchange.com/questions/4292320/… $\endgroup$
    – zlon
    May 16 at 6:24
  • $\begingroup$ @JimB Generally, suppose $X/\lambda_x$ is assumed to be in one distributional family and $Y/\lambda_y$ in another family (which could be the same) and both families are supported on the positive reals. Because the event $X\gt Y$ is equivalent to $X/Y\gt \lambda_x/\lambda_y=t,$ the question of finding the chance of this event is reduced to that of finding the CDF of $X/Y.$ $\endgroup$
    – whuber
    May 16 at 12:57

1 Answer 1

2
$\begingroup$

As mentioned by @whuber, you're better off attempting to obtain the distribution of $X/Y$ than $X - Y$. Because both $X$ and $Y$ are supported on the positive reals one can write $Pr(X>Y)$ as $Pr(X/Y > 1)=$. So this will involve obtaining the cdf of $X/Y$.

Using Mathematica (and Maple and MATLAB and others can certainly do the same) I get the following for the pdf:

dist = TransformedDistribution[x/y, 
 {x \[Distributed] InverseGaussianDistribution[n1, s1], 
  y \[Distributed] InverseGaussianDistribution[m2, s2]}];
pdf = Simplify[PDF[dist, t], Assumptions -> t > 0]

pdf of X/Y

I don't believe that there's a nice closed form for the cdf so you'll need use numerical integration to get the cdf. Assuming you might be doing this in R:

# Define pdf
  pdf <- function(t, m1, s1, m2, s2) {
    (exp(s1/m1 + s2/m2)*
     sqrt((s1*s2*(m1^2*s2 + m2^2*s1*t))/(s1 + s2*t))*
     besselK(sqrt(((m1^2*s2 + m2^2*s1*t)*(s1 + s2*t))/
     t)/(m1*m2),1))/(m1*m2*pi*t)            
  }
  
# Define cdf
  cdf <- function(t0, m1, s1, m2, s2) {
    1 - integrate(pdf, 0, t0, m1=m1, s1=s1, m2=m2, s2=s2)$value  
  }
  
# Pr(X > Y)
  1 - cdf(1, 1, 2, 2, 3)
# [1] 0.7420883

A partial simplification

We have $X \sim IG(\mu_1, \lambda_1)$ and $Y\sim IG(\mu_2,\lambda_2)$ and want to determine $Pr(X>Y)$. Note that we can write $X=\lambda_1 X’$ where $X’\sim IG(\mu_1/\lambda_1, 1)$ and $Y=\lambda_2 Y’$ where $Y’\sim IG(\mu_2/\lambda_2,1)$. We have

$$Pr(X>Y)=Pr(\lambda_1 X’ > \lambda_2 Y’)=Pr\left(\frac{X’}{Y’}>\frac{\lambda_2}{\lambda_1}\right)$$

If we let $\rho_1=\frac{\mu_1}{\lambda_1}$ and $\rho_2=\frac{\mu_2}{\lambda_2}$, then the pdf for $X’/Y’$ is given by

$$\frac{e^{\frac{1}{\rho_1}+\frac{1}{\rho_2}} \sqrt{\frac{\rho_1^2+\rho_2^2 t}{t+1}} K_1\left(\frac{\sqrt{\frac{(t+1) \left(\rho_1^2+t \rho_2^2\right)}{t}}}{\rho_1 \rho_2}\right)}{\pi \rho_1 \rho_2 t}$$

There doesn’t appear to be a closed form for the cdf so numerical integration will need to be used. One would need to evaluate 1 minus the cdf evaluated at $\lambda_2/\lambda_1$ to find $Pr(X>Y)$. So with this simplification to achieve the objective (finding $Pr(X>Y)$), we only need 3 parameters rather than 4: $\mu_1/\lambda_1$, $\mu_2/\lambda_2$, and $\lambda_2/\lambda_1$. However, given the likely need to estimate all 4 parameters ($\mu_1$, $\lambda_1$, $\mu_2$, and $\lambda_2$) there isn’t much of a savings.

$\endgroup$
2
  • $\begingroup$ The whole point to the scale parameter observation is that it simplifies the problem, eliminating two of the four parameters. Unfortunately, the difficulty with integrating the pdf persists. $\endgroup$
    – whuber
    May 16 at 17:33
  • $\begingroup$ @JimB Thank you very much for your answer! Unfortunately, now I'm not so sure I understood and used it correctly. Could you look up the "successor" question to this one. stats.stackexchange.com/questions/575591/… $\endgroup$
    – zlon
    May 17 at 11:56

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.