I have two independent random normal variables $Y_1$ and $Y_2$. The sum is also normally distributed: $Y_1 + Y_2 \sim N(\mu_1+\mu_2,V_1+V_2)$

The normally-scaled inverse gamma (NSIG) distribution is a conjugate prior. I'd like to compute the posterior distribution for the sums $(\mu_1+\mu_2,V_1+V_2)$ which govern the distribution of $Y_1 + Y_2$, given a sample of $(Y_1,Y_2)$. What is this distribution? Is it still NSIG?


Given $y_1$ and $y_2$, the posteriors on $(\mu_i,V_i)$ are independent: $$ (\mu_i,V_i)|y_i \sim \mathcal{N}(\xi_i,V_i/\lambda^\mu_i)\times \mathcal{IG}(\lambda^{\sigma}_i,\alpha_i)\,, $$ where the normal distribution is the conditional distribution of $\mu_i$ given $y_i$ and $V_i$ [see formula (2.7) in Bayesian Core!] Then this means that, conditional on $(V_1,V_2)$ and $(y_1,y_2)$, $$ \mu_1+\mu_2\sim \mathcal{N}(\xi_1+\xi_2,V_1/\lambda^\mu_1+V_2/\lambda^\mu_2)\,. $$ Unfortunately, the marginal distribution of $V_1+V_2$ is not so straightforward since it is the sum of two inverse gamma random variables with different scale and shape parameters... Even the sum $V_1^{-1}+V_2^{-1}$ does not enjoy a simple expression when the scale parameters are not the same!

  • 1
    $\begingroup$ Note also that the posterior of $\mu_1+\mu_2$ does not depend on $V_1+V_2$... $\endgroup$ – Xi'an Feb 1 '12 at 21:33
  • $\begingroup$ @user388027: given that you know (in closed form) the posterior distribution of $(V_1,V_2)$, you can use simulation to learn about everything on the posterior distribution of $V_1+V_2$. $\endgroup$ – Xi'an Feb 1 '12 at 21:37

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.