Consider three random variables $\epsilon_1, \epsilon_2, X$. Let $F_{\epsilon_i}(\cdot| x)$ denote the cumulative distribution function (cdf) of $\epsilon_i$ conditional on $X=x$ for any $i\in \{1,2\}$. Suppose $X$ has support $\mathcal{X}$.

In what follows I will impose restrictions on $F_{\epsilon_i}(\cdot| x)$ $\forall i \in \{1,2\}$ and I want to investigate the impact of these restrictions on the cdf of $\epsilon_1-\epsilon_2$ conditional on $X=x$, denoted by $F_{\epsilon_1-\epsilon_2}(\cdot| x)$.

Case 1 (zero median): Suppose $F_{\epsilon_1}(0|x)=F_{\epsilon_2}(0|x)=\frac{1}{2}$ $\forall x \in \mathcal{X}$. What can we say about $F_{\epsilon_1-\epsilon_2}(0|x)$ for any $x \in \mathcal{X}$? Can we claim that $F_{\epsilon_1-\epsilon_2}(0|x)=\frac{1}{2}$ $\forall x \in \mathcal{X}$?

Case 2 (symmetry): Suppose $F_{\epsilon_1}(u|x)=1-F_{\epsilon_1}(-u|x)$ and $F_{\epsilon_2}(u|x)=1-F_{\epsilon_2}(-u|x)$ $\forall u\in \mathbb{R}\cup\{-\infty, \infty\}$ $\forall x \in \mathcal{X}$. Can we claim that $F_{\epsilon_1-\epsilon_2}(u|x)=1-F_{\epsilon_1-\epsilon_2}(-u|x)$ $\forall u\in \mathbb{R}\cup\{-\infty, \infty\}$ $\forall x \in \mathcal{X}$?

Case 3 (independence): Suppose $F_{\epsilon_1}(\cdot|x)=F_{\epsilon_1}(\cdot|x')$ $\forall x,x'\in \mathcal{X}$ and $F_{\epsilon_2}(\cdot|x)=F_{\epsilon_2}(\cdot|x')$ $\forall x,x'\in \mathcal{X}$. Can we claim that $F_{\epsilon_1-\epsilon_2}(\cdot|x)=F_{\epsilon_1-\epsilon_2}(\cdot|x')$ $\forall x,x'\in \mathcal{X}$?


The answer to the first two "can we claim" questions is no.

Case 1: Consider the case where $\epsilon_1$ is $-1$ and $1$ with probability $1/2$, but that $\epsilon_2 = \epsilon_1$ with probability $1$. Then $F_{\epsilon_1 - \epsilon_2}(0 \mid x) = 1$ for any $x$, regardless of what the variable $X$ is or what its relationship is to $\epsilon_1, \epsilon_2$. On the other hand, if all the variables are independent of one another, then you can easily cook up a case where the claim would hold. Consequently, not much can be said about this case in general.

Case 2: Consider a case where the sample space for the variables $\epsilon_1, \epsilon_2$ is $\{a, b, c\}$, each with probability $1/3$. Let $\epsilon_1(a) = -1, \epsilon_1(b) = 0, \epsilon_1(c) = 1$, while $\epsilon_1(a) = 0, \epsilon_2(b) = 1, \epsilon_2(c) = -1$. Let $X$ be independent of $\epsilon_1$ and $\epsilon_2$. Then the variable $\epsilon_1 - \epsilon_2$ is $-1$ with probability $2/3$ and is $1$ with probability $1/3$, so it is not symmetric.

Case 3: The answer to this claim is yes. If $X$ is independent of $\epsilon_1$ and $\epsilon_2$, then it is independent of any function of those (such as $\epsilon_1 - \epsilon_2$) by a classical theorem.


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.