I'm working on a project involving regression models with multiple interaction terms. An example regression would be:

$Y$ ~ $A$ + $B$ + $C$ + $D$ + $AB$ + $CD$

where $AB$ = $A*B$ and $CD$ = $C*D$

All variables are continuous, normally distributed, and symmetric, with a mean of 0 and sd of 1.

I know the correlations between every term and $Y$, the correlations between $A$,$B$,$C$, and $D$, as well as the correlations between the two interaction terms ($AB$ and $CD$) and $A$,$B$,$C$, and $D$.

My intuition is that using this information I can figure out the correlation between $AB$ and $CD$ without first simulating them, but I'm not sure how to do it. Does anyone have any ideas?


1 Answer 1


I don't think you would need the conditions related to the response $Y$, because what you are interested is completely determined by the joint distribution of the $4$-dimensional random vector $(A, B, C, D)$. In fact, \begin{align} \operatorname{Corr}(AB, CD) = \frac{\operatorname{Cov}(AB, CD)}{\sqrt{\operatorname{Var}(AB)\operatorname{Var}(CD)}}, \end{align} where \begin{align} & \operatorname{Var}(AB) = E[A^2B^2] - (E[AB])^2, \\ & \operatorname{Var}(CD) = E[C^2D^2] - (E[CD])^2, \\ & \operatorname{Cov}(AB, CD) = E[ABCD] - E[AB]E[CD]. \end{align}

With the conditions you stated, $E[AB], E[CD]$ can be determined. However, higher-order moments $E[A^2B^2], E[C^2D^2], E[ABCD]$ cannot be determined.

For example, consider $(A_1, B_1, C_1, D_1) \sim N_4(0, I_{(4)})$ and $(A_2, B_2) \sim N_2(0, I_{(2)})$, $C_2 = (2\xi - 1)A_2, D_2 = (2\xi - 1)B_2$, where $\xi \sim \text{Bin}(1, 1/2)$ is independent of $(A_2, B_2)$. It is then easy to verify that

  1. $A_i \sim N(0, 1), B_i \sim N(0, 1), C_i \sim N(0, 1), D_i \sim (0, 1), i = 1, 2$.
  2. All pairwise correlations of $(A_i, B_i, C_i, D_i)$ are $0$, $i = 1, 2$.
  3. All pairwise correlations of $(A_iB_i, A_i, B_i, C_i, D_i)$ and $(C_iD_i, A_i, B_i, C_i, D_i)$ are $0$, $i = 1, 2$.

Therefore both $(A_1, B_1, C_1, D_1)$ and $(A_2, B_2, C_2, D_2)$ give exactly the same information in your question. However, \begin{align} & E[A_1B_1C_1D_1] = E[A_1]E[B_1]E[C_1]E[D_1] = 0, \\ & E[A_2B_2C_2D_2] = E[(2\xi - 1)^2A_2^2B_2^2] = 1. \end{align}

This shows $\operatorname{Corr}(A_1B_1, C_1D_1) = 0$, while $\operatorname{Corr}(A_2B_2, C_2D_2) = 1$, i.e., the correlation between interaction terms cannot be uniquely determined given the conditions presented.

  • $\begingroup$ Thanks for your answer! A couple follow-up questions. First, I'm not following your subscript notation, sorry! What is the difference between A1 and A2 here? Second, I have tried to do some reading on this - I found this paper, which states on pg. 315 that given my assumptions (I meant to include multivariate normality previously, sorry!) that E[ABCD], Var(AB), and Var(CD) are known (equations 7 & 11). Does the additional assumption help to resolve things? $\endgroup$
    – David B
    Jan 4, 2023 at 1:43
  • $\begingroup$ If you would like to assume multivariate normality, then $E[ABCD]$ (hence the correlation you are interested in) is of course uniquely determined. What I was trying to demonstrate is that, without the assumption of jointly normal, you can construct two random vectors $(A_1, B_1, C_1, D_1)$ and $(A_2, B_2, C_2, D_2)$ such that each of them satisfies the conditions you listed, while $\operatorname{Corr}(A_1B_1, C_1D_1) \neq \operatorname{Corr}(A_2B_2, C_2D_2)$. $\endgroup$
    – Zhanxiong
    Jan 4, 2023 at 1:52

Your Answer

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

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