5
$\begingroup$

Suppose $X=\log(Y)$ can be modeled by a mixture of two normal distributions with proportion $p$ of $X_1$ and proportion $1-p$ of $X_2$, where $X_1\sim\mathcal N(U_1, \sigma^2_1)$ and $X_2\sim\mathcal N(U_2, \sigma^2_2)$.

How do you calculate $E(Y)$; i.e., $E(\exp(X))$ where $X$ is a mixture of two normals?

| cite | improve this question | |
$\endgroup$
  • 2
    $\begingroup$ Are you asking about a linear combination of normals or a mixture? The notation you've used is a little ambiguous in this regard and a correct answer will depend on which you are interested in. $\endgroup$ – cardinal Aug 16 '12 at 17:58
  • 2
    $\begingroup$ @Cardinal Given that the OP has twice used the word "mixture" and did not equate "$p * X_1 + (1-p) * X_2$" with a $\mathcal{N}(p U_1 + (1-p)U_2, p^2 \sigma_1^2 + (1-p)^2 \sigma_2^2)$ distribution, it would be erroneous to interpret this as a linear combination, don't you think? $\endgroup$ – whuber♦ Aug 16 '12 at 18:37
  • 1
    $\begingroup$ @whuber: I think that's very highly likely; on the other hand, both answers here have done their calculations assuming (intentionally or inadvertently) the latter rather than the former; and, there was a very recent question where this confusion popped up. So, my response was, shall we say, conservative rather than speculative. :-) $\endgroup$ – cardinal Aug 16 '12 at 18:41
  • 1
    $\begingroup$ @Cardinal Thanks. And, to be perfectly clear--because I think you're holding back a little out of politeness--the answer to the mixture interpretation differs substantially from the answer to the linear combination interpretation. $\endgroup$ – whuber♦ Aug 16 '12 at 18:44
  • 1
    $\begingroup$ I am sorry not getting back earlier. I mean mixture of normals, not linear combination of two normals. $\endgroup$ – qqq Aug 16 '12 at 21:34
7
$\begingroup$

I'll try to give an answer for the mixture case. Let's formalize the set-up. We consider a random variable $X$ and an indicator random variable $I$, with $P[I=1] = 1-P[I=2] = p$, independent of $X$. Furthermore, for the mixture we have that the law of $X$ given that $I=1$ is the law of $X_1$, which is Gaussian with mean $U_1$ and variance $\sigma^2_1$; and, if $I=2$, the law is that of $X_2$ with law $N(U_2,\sigma^2_2)$.

Then, for $Y=\exp(X)$ we can calculate the expectation as $$ \begin{align} E[Y] &= E[\exp(X)] = p E[\exp(X)|I=1] + (1-p) E[\exp(X)|I=2] \\ &= p E[\exp(X_1)] + (1-p) E[\exp(X_2)] \\ &= p \exp(U_1+ \sigma^2_1/2) + (1-p) \exp(U_2+ \sigma^2_2/2) \end{align} $$ by using the expectation of a log-normal.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ (+1) There are a couple (inconsequential) typos in the first paragraph. $\endgroup$ – cardinal Aug 16 '12 at 20:51
  • $\begingroup$ Follow up exercise (for the interested reader): Imitate this approach to (easily) generalize to arbitrary finite mixture distributions and arbitrary (Borel measurable) function $g(x)$ in place of $\exp(x)$. $\endgroup$ – cardinal Aug 16 '12 at 20:54
  • $\begingroup$ @cardinal I have corrected some ... don't see any anymore ... but it's 11pm here in Europe ... $\endgroup$ – Ric Aug 16 '12 at 20:58
  • $\begingroup$ I used this initially, but started to doubt whether it is correct. If X ~ mixed normal of X1 and X2 with mixing probability p, does that mean exp(X) ~ mixed lognormal with same mixing probability p? I started to doubt this because this estimator seems to be biased (different from sample mean) for my data. Any suggestions on how to reduce the bias? $\endgroup$ – qqq Aug 16 '12 at 21:42
  • 1
    $\begingroup$ @Richard: Sorry guys, my complete story: I need to estimate mean revenue (the population is very heavy tailed). Sample mean is a natural unbiased estimator but has a wide confidence interval. I hope to find an estimator with smaller MSE. I fitted normal mixture to log transformed data. The fit seems good visually. But the point estimator I got (as what Richard suggested before) seems to be different from sample mean. I wonder whether I calculated it correctly and therefore raised the original question. If right, how to reduce the bias? Thanks for your suggestions! $\endgroup$ – qqq Aug 17 '12 at 23:07
0
$\begingroup$

For instance, if you are dealing with a two component mixture, the first moment is calculated as follows: $\mu_{\rm mixture}=\pi_{1} \cdot \mu_{1} + \pi_{2} \cdot \mu_{2}$, where $\pi_{i}\ (i=1,2)$ stands for the components' weights and $\mu_{i}\ (i=1,2)$ for the means.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ This is true, as far as it goes, but it ignores the non-linear transform (exp) in the problem statement. $\endgroup$ – gung - Reinstate Monica♦ Jul 16 '14 at 18:15

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.