pluviosilla
- 17
- 0
The slope of a fitted line = Cov(X,Y)/Var(X). I've seen the derivation of this, and it is pretty straightforward, but I am still having trouble getting an intuitive grasp. The formula is extremely suggestive and it is bothering me that I can't quite see its significance.
Perhaps, my mental block comes from thinking of X values as points on a line and I should instead be thinking of two parallel series, one of x values and another of y values that are parameterized using a third variable, t for time, for example. Thus values for X would not be sequential, like x values on the x-axis. They will fluctuate around a mean. Sometimes y values will "co-vary", i.e. fluctuate in *tandem* with x values but some of that apparent covariance is deceptive. It is really just noise, which is why the slope of the fitted line cannot be simply Cov(X,Y). We must divide by Var(X) in order to subtract out that *accidental* coincidence (or covariance) of X & Y.
Is it something like that?
Perhaps, my mental block comes from thinking of X values as points on a line and I should instead be thinking of two parallel series, one of x values and another of y values that are parameterized using a third variable, t for time, for example. Thus values for X would not be sequential, like x values on the x-axis. They will fluctuate around a mean. Sometimes y values will "co-vary", i.e. fluctuate in *tandem* with x values but some of that apparent covariance is deceptive. It is really just noise, which is why the slope of the fitted line cannot be simply Cov(X,Y). We must divide by Var(X) in order to subtract out that *accidental* coincidence (or covariance) of X & Y.
Is it something like that?