In the answer by fmark in this question it is stated that:

We prefer natural logs (that is, logarithms base $e$) because, as described above, coefficients on the natural-log scale are directly interpretable as approximate proportional differences: with a coefficient of 0.06, a difference of 1 in $x$ corresponds to an approximate 6% difference in $y$, and so forth.

Just wondering, if I apply log10 to the dependent and independent variable(s) would the above not apply anymore?

Looking at this one could argue that the above only applies for -5% ... 5% anyway and ln and log are very similar around x = 1 ...

  • $\begingroup$ $log_{10}$ and $ln$ only differ by a (constant) factor so if you apply it to the dependent and the independent variables then it doesn't make a difference except for the constant, see purplemath.com/modules/logrules5.htm. If you denote the factor by $f$, i.e. $f\cdot ln(x)=log_{10}(x)$ then if $ln(y)=a+b\cdot ln(x)$ is equivalent to $f\cdot ln(x)= f\cdot a + f\cdot b \cdot ln(x)$ is equivalent to $log_{10}(y)=f\cdot a + b\cdot log_{10}(x)$ $\endgroup$ – user83346 Jul 23 '17 at 18:17
  • $\begingroup$ so the fitted parameter of a log10'ed independent variable has the same as that of an ln'ed independent variable variable if the dependent variable was log10'ed and ln'ed respectively? $\endgroup$ – cs0815 Jul 23 '17 at 18:20
  • $\begingroup$ except for the constant, see my comment $\endgroup$ – user83346 Jul 23 '17 at 18:20
  • $\begingroup$ Thanks. I was aware of this transformation but I am not sure I understand correctly. I fit log10(y) = alpha * log10(x) or ln(y) = beta * ln(x) can alpha be interpreted in the same manner as beta for small changes in x? $\endgroup$ – cs0815 Jul 23 '17 at 18:25
  • 1
    $\begingroup$ in that case $\alpha$ and $\beta$ should be identical, you can easily try that in e.g. R $\endgroup$ – user83346 Jul 23 '17 at 18:29

As $ln(x)$ and $log_{10}(x)$ only differ by a factor see e.g. this link it holds that $f \cdot ln(x) = log_{10}(x)$, so then it holds that, if $ln(y)=a+b \cdot ln(x)$ then also $f \cdot ln(y)=f \cdot a+f \cdot b \cdot ln(x)$ or $log_{10}(y)= f \cdot a + b \log_{10}(x)$, so the coeffient $b$ is exactly the same.

Some R code to illustrate that:




  • $\begingroup$ Although the coefficient remains the same, its interpretation is subtly different: it is no longer an approximate percent change in response per unit change in the (log) regressor. A difference of $1$ in $\log_{10} x$ now corresponds to a difference of $b$ in $\log_{10} y$, which (for small $|b|$) is approximately a $100\times b\log(10)\% \approx 230\, b\%$ change in $y$. $\endgroup$ – whuber Jul 24 '17 at 15:39
  • $\begingroup$ @whuber: I think a 1% change in $x$ corresponds to a b% change in $y$ in both cases , or am I wrong ? $\endgroup$ – user83346 Jul 24 '17 at 16:10
  • 1
    $\begingroup$ Yes, that is correct--which is why I characterized the difference in interpretation as "subtle". Notice that a "difference of 1" in $\log_{10}(x)$is not a $100\%$ change in $x$. Somebody who incorrectly supposes, as the OP explicitly does, that "ln and log are very similar around x=1" might very well fall into the trap of not realizing this. $\endgroup$ – whuber Jul 24 '17 at 16:29

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.