8
$\begingroup$

I want to simulate a sample from a mixture normal distribution such that

$$p\times\mathcal{N}(\mu_1,\sigma_1^2) + (1-p)\times\mathcal{N}(\mu_2,\sigma_2^2) $$

is restricted to the interval $[0,1]$ instead of $\mathbb{R}$. This means I want to simulate a truncated mixture of normal distributions.

I know that there are some algorithm to simulate a truncated normal (i.e from this question) and corresponding package in R to do this. But how can I simulate a truncated mixture normal? Is it the same if I simulate two truncated normal of $\mathcal{N}(\mu_1,\sigma_1^2)$ and $\mathcal{N}(\mu_2,\sigma_2^2$) to make a truncated mixture normal?

$\endgroup$
  • 3
    $\begingroup$ If it's on the unit interval, why not use betas instead of normals? For $\alpha=\beta>1$, the distribution is symmetric and unimodal and bounded on the unit interval. $\endgroup$ – Sycorax Mar 23 '15 at 14:45
  • 2
    $\begingroup$ If you don’t need your simulations to be very fast, you can do it using rejection sampling: (1) sample $x$ from the mixture of two normals, (2) if $x$ is not in $[0,1]$, go back to step 1, (3) output $x$. (but user777 is right, do you have a good reason to chose this distribution instead of a mixture of betas?) $\endgroup$ – Elvis Mar 23 '15 at 14:53
  • 1
    $\begingroup$ @user777 a Truncated Gaussian mixture has a different distribution from a Beta distribution and cannot be swapped out just because you can enforce symmetry and the same support. $\endgroup$ – mjnichol May 11 '15 at 21:15
12
$\begingroup$

Simulation from a truncated normal is easily done if you have access to a proper normal quantile function. For instance, in R, simulating $$ \mathcal{N}_a^b(\mu,\sigma^2)$$where $a$ and $b$ denote the lower and upper bounds can be done by inverting the cdf $$\dfrac{\Phi(\sigma^{-1}\{x-\mu\})-\Phi(\sigma^{-1}\{a-\mu\})}{\Phi(\sigma^{-1}\{b-\mu\})-\Phi(\sigma^{-1}\{a-\mu\})} $$ e.g., in R

x = qnorm(  pnorm(a,mu,sigma) + runif(1)*(pnorm(b,mu,sigma) - pnorm(a,mu,sigma))  )

Otherwise, I developed a truncated normal accept-reject algorithm twenty years ago.

If we consider the truncated mixture problem, with density $$ f(x;\theta) \propto \left\{p\varphi(x;\mu_1,\sigma_1)+(1-p)\varphi(x;\mu_2,\sigma_2)\right\}\mathbb{I}_{[a,b]}(x) $$ it is a mixture of truncated normal distributions but with different weights: $$ f(x;\theta) \propto p\left\{\Phi(\sigma_1^{-1}\{b-\mu_1\})-\Phi(\sigma_1^{-1}\{a-\mu_1\}) \right\}\dfrac{\sigma_1^{-1}\phi(\sigma_1^{-1}\{x-\mu_1\})}{\Phi(\sigma_1^{-1}\{b-\mu_1\})-\Phi(\sigma_1^{-1}\{a-\mu_1\})} \\[15pt] +(1-p)\left\{\Phi(\sigma_2^{-1}\{b-\mu_2\})-\Phi(\sigma_2^{-1}\{a-\mu_2\}) \right\}\dfrac{\sigma_2^{-1}\phi(\sigma_2^{-1}\{x-\mu_2\})}{\Phi(\sigma_2^{-1}\{b-\mu_2\})-\Phi(\sigma_1^{-1}\{a-\mu_2\})} $$ Therefore, to simulate from a truncated normal mixture, it is sufficient to take $$x=\begin{cases} x_1\sim\mathcal{N}_a^b(\mu_1,\sigma_1^2) &\text{with probability }\\ &\qquad p\left\{\Phi(\sigma_1^{-1}\{b-\mu_1\})-\Phi(\sigma_1^{-1}\{a-\mu_1\}) \right\}\big/\mathfrak{s}\\ x_2\sim\mathcal{N}_a^b(\mu_2,\sigma_2^2) &\text{with probability }\\ &\qquad(1-p)\left\{\Phi(\sigma_2^{-1}\{b-\mu_2\})-\Phi(\sigma_2^{-1}\{a-\mu_2\}) \right\}\big/\mathfrak{s} \end{cases} $$ where \begin{align} \mathfrak{s}=&p\left\{\Phi(\sigma_1^{-1}\{b-\mu_1\})-\Phi(\sigma_1^{-1}\{a-\mu_1\}) \right\}+ \\ &(1-p)\left\{\Phi(\sigma_2^{-1}\{b-\mu_2\})-\Phi(\sigma_2^{-1}\{a-\mu_2\}) \right\} \end{align}

$\endgroup$
  • $\begingroup$ Thanks you very much Xi'an.Your answer help me understand how to simulate a sample from a truncated mixture distribution, in general. $\endgroup$ – Alexy Mar 23 '15 at 16:11
  • $\begingroup$ Why can't we just draw the sample from the first normal with probability p and the second distribution with probability 1 - p? $\endgroup$ – mjnichol May 11 '15 at 21:25
  • 1
    $\begingroup$ Ah! I think I see the issue. It is because the entire distribution is being truncated, not each distribution separately. If each sub-distribution of the mixture were individually truncated before being added into the mixture then we would be able to simply sample from the distribution according to the relative weights of each sub-distribution, right? $\endgroup$ – mjnichol May 11 '15 at 22:03
  • 1
    $\begingroup$ @mjnichol It is a mixture but with different weights than $p$ and $1-p$. $\endgroup$ – Xi'an May 12 '15 at 5:21
  • 2
    $\begingroup$ @mjnichol: in that case, you would have $$p\mathcal{N}_a^b(\mu_1,\sigma_1^2)+(1-p)\mathcal{N}_a^b(\mu_2,\sigma_2^2)$$so yes indeed this would work. $\endgroup$ – Xi'an May 12 '15 at 17:50

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.