4
$\begingroup$

I am running a mediation model including the variables $X$, $M$, and $Y$.

The Spearman correlation between $X$ and $Y$ is +0.4, while the direct path c’ between $X$ and $Y$ has a negative coefficient of -0.206.

Further notes: All variables, that is, $X$, $M$, and $Y$, are the same kind of measurement. Hence, they do not operate on different scales or intervals. Also, I did not demean the data before running the moderation model.

Question: I cannot make sense of this result because it seems like a contradiction to me. Also, why does the direct path c’ yield a negative coefficient of -0.206 when a linear regression between X and Y, as shown in the image below, is positive? Am I confusing the direct path c’ with the total path c here?

enter image description here

enter image description here

enter image description here

$\endgroup$
1

1 Answer 1

5
$\begingroup$

It's quite easy for something like this to occur. First, note that a correlation only captures the covariance between two variables. It does not partial out the effects present from other predictors. After conditioning on other variables, this can greatly influence your regression coefficients (see for example suppression effects, which pushes a coefficient to essentially zero even with correlations seemingly "strong" enough). Here you are fitting two regression paths which use $X$ as a predictor. Therefore the variance that is tied up in $X$ that would normally only go to $Y$ is now being shared with $M$. There is also a strong correlation between $M$ and $Y$, which means more of the variance in $Y$ is being soaked up by $M$. This could be contributing in some way. You may want to think on what is contributing to that from a theoretical lens.

However as a side note, I would ask yourself what is causing the two seemingly distinct blobs of data in both of your scatters. You may have a case of omitted variable bias going on, which I would investigate. It may be that you have some unknown grouping variable that is influencing your association here.

$\endgroup$
1
  • $\begingroup$ Thank you. It is much more clear to me now. I also have a theoretical reason why M way better explains the variance in Y than X. Explaining this goes way beyond my question here though. You are also right about pointing out the two blobs. I know the reason for that, but I cannot really disentangle that in the data analysis (again, explaining this goes beyond my topic here). So thank you, I will accept your answer! $\endgroup$
    – Philipp
    Commented Jun 20 at 8:55

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.