Consider the model:


Where $D_1=\{0,1\}$ and $D_2=\{0,1\}$ are binary (dummy) variables, and $y_i$ is a continuos variable in levels.

In this model, for example $\beta_1$ is interpreted as the gap between the category when $D_1=1$ and $D_1=0$. Besides, this gap is in absolute terms, say in terms of the units of the $y_i$ variable.

Now, consider the similar model but dependent variable is in logs:

$$\ln (y_i)=\beta_0+\beta_1D_1+\beta_2D_2+u_i$$

I would thought that, for example, $\beta_1$ is the gap between the mentioned categories but in relative terms, say a proportion or percentage. Say in average category $D_1=1$ receives $\beta_1*100$ percent more (or less) of $y_1$ than $D_1=0$ in average.

But I found another approximation in my class for this gap, and it takes the gap starting with the fact that the relative gap:

$$Rel. gap\; between\; [D_1=1]\; and\; [D_1=0]=\frac{y_1-y_0}{y_0}*100\% = \left(\frac{y_1}{y_0}-1\right)*100\%$$

Where $y_1=y(D_1=1)$ and $y_0=y(D_1=0)$

Taking this into account we can get:

$$\ln (y_1)-\ln(y_0)=\beta_1 \implies \frac{y_1}{y_0}=e^{\beta_1}$$

Replacing this into the relative gap equation specified earlier we get:

$$Rel. gap\; between\; [D_1=1]\; and\; [D_1=0]=\left ( e^{\beta_1} -1 \right)*100\%$$

My question is: What is specifically the difference between those two approximations? And even if one of those perhaps is wrong, because the resulting gaps that you can obtain are way different. Moreover, which one should one use with academic purposes.


  • $\begingroup$ This would be easier to expalain and understand if you did not subtract $1$. If $e^{\beta_1}=2.5$ then having $D_1=1$ multiplies the estimate for $y$ by $2.5$, while if $e^{\beta_2}=4$ then having $D_2=1$ multiplies the estimate for $y$ by $4$, and having both multiplies the estimate for $y$ by $10$. So much simpler than saying that these increase the estimates for $y$ by $150\%$, $300\%$ and $900\%$ respectively $\endgroup$
    – Henry
    Commented Nov 21, 2020 at 22:39

1 Answer 1


As you can see from the plot, they are not the same. Infact, the 'rule of thumb', i.e. your first formula, only holds in approximation. The error gets larger for large coefficient values.



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.