1
$\begingroup$

I'm having some trouble fully understanding partial correlation and I was wondering if some of you can shred some light on my confusion.

Let's consider the following scenario: It is a known fact that heart disease is related to social and economic status. However, I want to understand if Anger is also a factor. So the obvious next step is to find the correlation between Anger and heart disease while controlling for social and economic status.

There are a couple ways I can do this. The one popular way that I found online was to use partial correlation (ppcor in R). However, when I looked into how they did partial correlation, it didn't make a lot of sense to me mathematically. The way they do it is: let's say they have 3 variables (X, Y , Z) and we want to correlate X and Y while taking into consideration Z, they take the residuals from correlating X and Y, then X and Z, then they correlated the two residuals to get the result.

This doesn't make a lot of sense to me, if residuals are variance that are not explained through correlation, then wouldn't it make more sense to only take the residuals from X and Z, then correlate that residual with Y, that way we can see if Y can explain the variance that is not explained by X and Z, therefore "controlling" for Z?

$\endgroup$
  • 2
    $\begingroup$ residuals from correlating X and Y Correlation cannot produce residuals. It is regression - directed correlation - that can. X is regressed by Z and residuals saved (i.e. Z is washed out from X). Y is regressed by Z and residuals saved (i.e. Z is washed out from Y). All three variables must be standardized initially. Then the two residuals correlate with each other. $\endgroup$ – ttnphns Jun 19 '16 at 20:22
  • 1
    $\begingroup$ Regarding the statement "So the obvious next step is to find the correlation between Anger and heart disease while controlling for social and economic status" : en.wikipedia.org/wiki/Correlation_does_not_imply_causation $\endgroup$ – shrey Jun 20 '16 at 10:22
0
$\begingroup$

You say that, with $X=\alpha_1+\beta_1Z+\epsilon_1$ and $X=\alpha_2+\beta_2Z+\epsilon_2$, you would prefer to consider $Corr(\epsilon_1,Y)$ rather than $Corr(\epsilon_1,\epsilon_2)$ . However, please notice that, given $\epsilon_1$ is independent from $Z$, you would not change the covariance ($Cov(\epsilon_1,Y)=Cov(\epsilon_1,\epsilon_2)$), so you would only replace the standard error of $\epsilon_2$ with the one of $Y$. You would thus lead your correlation toward zero for no reason. In your case (let's call Anger $A$, SocioEconomic Status $SES$ and Heart Disease $HD$), you would try to see how much "residual" anger (after controlling for $SES$) affects $HD$. The problem is that, after regressing $A$ on $SES$, what is left (the residuals) cannot explain the part of $HD$ that has already been explained by $SES$. So, by using $HD$ instead of its residuals, you would basically calculate a correlation that would be bounded (both from below and above) by the explanatory power of $SES$ on $HD$. In practice, after controlling for $SES$, in this way you'd have that the partial correlation between $HD$ and any other variable would not range between $-1$ and $+1$, but between $-k$ and $+k$ (with $k<1$, and the higher the explanatory power of $SES$ on $HD$, the lower $k$).

$\endgroup$

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.