8
$\begingroup$

Consider a typical 2x2 table of frequencies (shown in this image):
two by two table
Notation: The row variable is denoted R and takes on values 0 or 1; the column variable is denoted C and takes on values 0 or 1. The cells of the table indicate the frequency of each combination of R and C; for example, $b$ is the frequency of R=0 and C=1. For purposes of my question, assume that the cell counts are divided by the total, so that the cell values are the joint probabilities of the cells.

I want to express the cell probabilities in terms of the phi coefficient (which is a measure of correlation with formula provided below) and the marginal probabilities: $\mu_R\equiv p(R\!=\!1) = c+d$ and $\mu_C\equiv p(C\!=\!1) = b+d$. That is, I want to invert the following system of four equations: $$\begin{align} \phi &\equiv (ad-bc)/\sqrt{(a+b)(c+d)(a+c)(b+d)} \tag{by defn}\\ \mu_{R} &= c+d \tag{by defn}\\ \mu_{C} &= b+d \tag{by defn}\\ 1 &= a+b+c+d \tag{constraint} \end{align}$$ and, of course, $0 \le a,b,c,d \le 1$. In other words, I would like to solve for $a$, $b$, $c$, and $d$ in terms of $\phi$, $\mu_{R}$, and $\mu_{C}$.

This problem has probably been solved by somebody before, but my searches have not yielded a source, and my feeble attempts at algebra have not produced an answer, and I cannot find online system-of-(nonlinear)-equation inverters that handle this case.

$\endgroup$

1 Answer 1

4
+50
$\begingroup$

We easily recognize every factor in the denominator of $\phi$, because $a+b=1-\mu_R$ and $a+c=1-\mu_C$. Let's therefore start with a tiny simplification to avoid writing lots of square roots:

$$\Delta=ad - bc = \phi \sqrt{\mu_R(1-\mu_R)\mu_C(1-\mu_C)}.$$

Let's find $d$:

$$\eqalign{d &= (1)d = (a+b+c+d)d = ad +bd +cd + d^2 \\ &= ad + (-bc + bc) + bd + cd + d^2 \\ &= (ad - bc) + (c+d)(b+d) \\&= \Delta + \mu_R\mu_C.}$$

Finding $a$, $b$, and $c$ proceeds similarly due to the symmetries of the problem: interchanging the columns swaps $a$ and $b$, $c$ and $d$, while changing $\mu_C$ to $1-\mu_C$ and negating $\Delta$, whence $$c = -\Delta + \mu_R(1-\mu_C).$$

Interchanging the rows swaps $a$ and $c$, $b$ and $d$, while changing $\mu_R$ to $1-\mu_R$ and negating $\Delta$, whence

$$b = -\Delta + (1-\mu_R)\mu_C.$$

Swapping both rows and columns yields

$$a = \Delta + (1-\mu_R)(1-\mu_C).$$


Given these expressions for $a,b,c,d$, it is simple to check that $a+b+c+d=1, c+d=\mu_R,$ and $b+d=\mu_C$, and only a little bit harder to verify that $ad-bc=\Delta$.

$\endgroup$
2
  • $\begingroup$ One note to others who might use this (correct!) answer: It can yield values of a, b, c, or d that are negative. In other words, not all combinations of phi in [-1,1], mu_R in [0,1], and mu_C in [0,1] can be created by probability matrices. To whuber: Thank you! $\endgroup$ Jul 14, 2016 at 12:01
  • $\begingroup$ That's correct, John, but I made no mention of that fact because presumably $\mu_R$, $\mu_C$, and $\phi$ had been obtained from a valid table in the first place. Assuming $\mu_R$ and $\mu_C$ are valid frequencies (in the interval $[0,1]$), $\Delta$ will be real. It must lie in the interval $$[-\min(\mu_R\mu_C, (1-\mu_R)(1-\mu_C)), \ \min(\mu_R(1-\mu_C), (1-\mu_R)\mu_C)].$$ $\endgroup$
    – whuber
    Jul 14, 2016 at 13:16

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.