I have data containing several variables. I ran a regression model. Prior to running the model I have normalized the dependent variable Y and the independent variables X1 and X2.

After receiving the output I want to interpret the results. For example, if the coefficient of X1 is 0.15, I know that it means that for addition of one standard deviation of X1, there is an increase of 0.15 standard deviations in Y, but this is not clear.

I want to go back to the original units of Y and X for interpretation. How do I do that ? Can I simply take the "normalized coefficients", multiply by the standard deviation of Y and add the mean of Y?

Something about it doesn't make me comfortable.

  • 1
    $\begingroup$ Would you please post a link to the raw data? I would like to surface fit "Y = f(X1, X2)"and see what I find with an equation search. $\endgroup$ – James Phillips Jan 23 '19 at 16:02

Firstly, why did you normalized Y? It will make your output harder to interpret, and it is often not necessary to standardize the dependent variable.

I presume you have centered and scaled you X's, you can backtransform them to interpret,

        # run your model with the X's standardized

        mean <- mean(x1)
        sd <- sd(x1)


#you can also plot

curve(b0 + b1*(x1-mean)/sd, add=TRUE) 

I also recommend to use Y at its original scale, since standardization does not change the distribution shape.

  • $\begingroup$ Can I say that the change in Y is actually the "standardized coefficient" multiplied by the standard deviation of Y ? I think it makes sense $\endgroup$ – user3275222 Jan 23 '19 at 16:08

Standardising both the dependent and independent variables can be useful for presentation and coefficient interpretation, normally in simple linear regression, whenever the Pearson correlation coefficient $r_{xy}$ is of interest. The OLS estimation $\hat{\beta} = r_{xy} \times \frac{s_y}{s_x} $. Clearly, if a variable $z$ has been standardised to ${z\prime}$ then it's standard deviation must equal unity, since $$ \sqrt{\mathbb{V}[{z\prime}]} = \sqrt{\mathbb{V}\left[\frac{z-\bar{z}}{\sqrt{\mathbb{V}[z]}} \right]} = \sqrt{\frac{1}{\mathbb{V}[z]}{\mathbb{V}[z-\bar{z}]}} = 1 $$ Therefore, when both $y$ and $x$ have been standardised $s_y=s_x=1$ and the OLS estimator for $\hat\beta$ is the correlation coefficient $r_{xy}$. Note also that the OLS estimator of $\alpha$ is zero for standardised variables since $\bar{x}=\bar{y}=0$.

In your multiple regression context, if your $X_i$ are independent then the $\beta_i$ will be equal to the correlation coefficients if you had performed separate simple regressions.

If the $X_i$ are not independent, each $\beta_k$, a partial correlation coefficient, shows the relationship between $Y$ and $X_k$ with the other $X_i$ fixed. This can have a different value to the simple correlation coefficient between $Y$ and $X_k$ in a simple linear regression ignoring the other $X_i$.


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.