Causal Inference in Statistics

Base: Correlation does not mean casualty.

  • If X and Y are statistically dependent, X does not necessarily cause Y (or Y cause X). 相关性不代表有因果性
  • If X causes Y, then X & Y are very likely to be statistically dependent (but not always, there is extreme condition). 但是因果性代表相关性

Study 1. V Structure:

  • Chain

$$ X\rightarrow Y\rightarrow Z $$

Z and X are likely dependent. However, Z and X are independent, conditional on Y.

$$ P(Z=z|X=x,Y=c)=P(Z=z|Y=c) $$i.e.

i.e.

\(f_x: X=u_x\)

\(f_y: Y=84-X+u_Y := c\)

\(f_z: Z=100\underbrace{Y}_{c}+u_z\)

Now, Z and X are independent.

Therefore, we know, in the Chain:

$$ X\equiv Z$$

$$ X\bot Z|Y $$

  • Folk

$$ Y\leftarrow X\rightarrow Z $$

Y and Z are likely dependent. However, Y and Z are independent conditional on X.

$$P(Z=z|Y=y, X=c)=P(Z=z|X=c)$$

While conditioning on intermediate node X, then Z and Y are independent.

$$Y\bot Z|X$$

  • Collider

$$ X\rightarrow Z\leftarrow Y $$

X and Y are independent. However, X and Y are dependent conditional on Z.

$$ P(X=x|Y=y, Z=c)\neq P(X=x|Z=c) $$

i.e.

If we know \(Z=X+Y+u_Z:=c\), then \( X=c-Y-u_Z\), and thus X and Y become dependent conditional on \(Z=c\). Otherwise, \(X=u_X\) and \(Y=u_Y\).

Once, conditioning on \(Z\), the way gets connected. Otherwise (unconditional), we get independent.

P.S. Descendent of Z:

$$ X (or\ Y)\rightarrow Z\rightarrow W $$

Similarly, we get in the Collider:

$$ X\bot Y $$

$$ X\equiv Y|Z $$

$$ X\equiv Y |W $$

  • See notes for further studies.

Reference

Pearl, J., Glymour, M. and Jewell, N.P., 2016. Causal inference in statistics: A primer. John Wiley & Sons.