4
$\begingroup$

I am trying to define a bivariate exponential distribution $(S, T)$ with marginals $S\sim\mathrm{Exp}(\lambda_S)$ and $T\sim\mathrm{Exp}(\lambda_T)$ for $\lambda_S > \lambda_T$. I would like the joint distribution of $S$ and $T$ to have two properties:

  1. $\rho(S, T)$ can be controlled by some parameter of the joint distribution
  2. $S\leq T$

The literature has many examples of distributions that fulfill one of these two properties. For instance, the BVE of Marshall and Olkin (1967) can be constructed from independent random variables $\tilde S, \tilde T$, and $C$ and selected constant $\lambda_C < \lambda_T$ as

\begin{align*} \tilde S &\sim \mathrm{Exp}(\lambda_S-\lambda_C) \\ \tilde T &\sim \mathrm{Exp}(\lambda_T-\lambda_C) \\ C &\sim \mathrm{Exp}(\lambda_C) \\ S &= \min(\tilde S, C) \\ T &= \min(\tilde T, C) \end{align*}

Here, $S$ and $T$ have the desired marginal distributions, and $\lambda_C$ controls the degree to which they are correlated (requirement 1). However, we cannot guarantee that $S\leq T$ (requirement 2).

Alternately, we could use independent random variables $\tilde S$ and $T$ to construct

\begin{align*} \tilde S &\sim \mathrm{Exp}(\lambda_S-\lambda_T) \\ T &\sim \mathrm{Exp}(\lambda_T) \\ S &= \min(\tilde S, T) \end{align*}

Again, $S$ and $T$ have the desired marginal distributions. This time $S\leq T$ (requirement 2), but we have no way to control the correlation (requirement 1).

Is there a bivariate exponential distribution that meets both of my requirements?

$\endgroup$
0

1 Answer 1

4
$\begingroup$

Consider $E \sim \mathcal{E}( \lambda_S - \lambda_C)$ and $C \sim \mathcal{E}(\lambda_C)$. Now define \begin{align*} S & = \min \{ E , C \} \\ T & = \min \{ a E , C\} \end{align*} where $$ a = \frac{ \lambda_S - \lambda_C}{\lambda_T - \lambda_C} > 1 . $$

You get $ S \sim \mathcal {E}(\lambda_S)$ and $ T \sim \mathcal{E} (\lambda_T)$, by construction $S \leq T$ and you have some ability to control correlation between two variables. Analytically, $$\rho(S,T) = \frac{\lambda_S+\lambda_C(1-\lambda_T/\lambda_S-\lambda_S/\lambda_T)}{\lambda_S-\lambda_C} . $$ If $ \lambda_C \approx 0$ you get $S \approx E$ and $S\approx a E$ hence $\rho \approx 1$, but if $\lambda_C \approx \lambda_T$ then $T \approx C$ and $S \approx \min\{E, C\}$, meaning $\rho\approx\lambda_T/\lambda_S$.

With the following simulation I obtained an estimation of $0.625$ for the correlation with $\lambda_S = 2$, $\lambda_T = 1.1$ and $\lambda_C = 1$ (analytically we know $\rho(S,T)=\frac{139}{220}\approx 0.632$).

set.seed(1234)
lambda_S = 2; lambda_T = 1.1; lambda_C = 1
a = (lambda_S -lambda_C)/(lambda_T -lambda_C)

E = rexp(10000,  lambda_S - lambda_C)
C = rexp(10000,  lambda_C)

S =  apply(rbind(E,C), 2, min)
TT = apply(rbind( a*E,C), 2, min)

cor(S, TT)
# 0.6254737

Changing $\lambda_S = 10$ gives $\rho(S,T)=\frac{1979}{9900}\approx 0.200$.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.