1
$\begingroup$

In the context of a linear regression model where the independent variable ($X$) was log-transformed, like:

$Y = \alpha + \beta·ln(X)$

Is there a straightforward way to transform a regression coefficient ($\beta$) that was estimated using log-transformed independent variable $X$?

I know that in this model, an increase of 1% in $X$ would mean an increase of $\beta·ln(X)$ in the value of $Y$.

Is there a way to transform $\beta$ so I could say that an increase of 1% in $X$ means an increase of $\beta'·X $ in variable $Y$?

Improve this question
$\endgroup$
5
  • 1
    $\begingroup$ Possible duplicate of Back-transformation of regression coefficients $\endgroup$
    – Greenparker
    Commented Jul 5, 2016 at 13:05
  • $\begingroup$ Since $X$ was transformed, not $\beta$, a "back-transformation" of $\beta$ would seem to make no sense. Could you clarify what you are hoping to achieve? $\endgroup$
    – whuber ♦
    Commented Jul 5, 2016 at 13:15
  • $\begingroup$ $\beta$ is estimated here using $ln(X)$, so interpretation dependens on $ln(X)$ [as in the main text: an increase of 1% in $X$ results in an increase of $\beta·ln(X)$ for $Y$]. What I wonder to achieve (if possible) is an expression of the change in $Y$ due to $X$ in terms not of $ln(X)$ but $X$ itself, using $ln(X)$ in regression analysis. $\endgroup$
    – elcortegano
    Commented Jul 5, 2016 at 13:21
  • 1
    $\begingroup$ But of course: it's right in front of us. If $X$ changes to $\alpha X$ (for positive $\alpha$, of course) then $\log(X)$ becomes $\log(\alpha)+\log(X)$, whence $Y$ becomes $Y + \beta\log\alpha$. But there is no "back-transformation" of any sort needed in this analysis. $\endgroup$
    – whuber ♦
    Commented Jul 5, 2016 at 13:51
  • 1
    $\begingroup$ @whuber, why not develop that into an answer? I don't see an answer beyond that, and this isn't really a duplicate of the linked thread as it is about the transformation of Y whereas this is about the transformation of X. $\endgroup$
    – gung - Reinstate Monica
    Commented Jul 5, 2016 at 14:35

2 Answers 2

Reset to default
1
$\begingroup$

You can't transform in the sense you'd wish in the question, i.e. preserving the interpretation of the coefficient $\beta$. In the log transformed regression $\beta$ becomes the sensitivity of the dependent variable to the percentage change in the independent variable (IV), instead of the sensitivity to the simple change in IV. In other words, $\beta$ in the log transformed case isthe measure of the relative sensitivity rather than absolute one.

Improve this answer
$\endgroup$
0
1
$\begingroup$

You can get some intuition for small changes in $X$ by considering

$$ \begin{align} Y+\delta Y&=\alpha + \beta \ln(X + \delta X)\\ &=\underbrace{\alpha + \beta \ln X}_Y + \beta\ln\left(1+\frac{\delta X}{X}\right)\\ \therefore\delta Y&\approx \beta\frac{\delta X}{X}, \end{align} $$

where the approximation is obtained by a Taylor series expansion of the logarithm to first order in $\delta X$.

In other words, a 1% increase in $X$ gives you an increase of $0.01\beta$ in the dependent variable. But this only holds for small changes in $X$.

Improve this answer
$\endgroup$
1
  • 1
    $\begingroup$ When $\log(X)$ is used in a model, the natural way to represent changes to $X$ is multiplicative, not additive. The question is clear about that: it refers to a "1% change," which is multiplicative. $\endgroup$
    – whuber ♦
    Commented Jul 5, 2016 at 13:53

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.