So far I have been under the impression that you can "linearize" multiplicative models of the form (1) $y=\alpha * \beta_1x_1 * \beta_2x_2 * \beta_3x_3 $ and exponential models of the form (2) $y=\alpha * x_1^{\beta_1} * x_2^{\beta_2} * x_3^{\beta_3} $ by taking the logarithm and then doing a regular OLS estimation.

For exponential models this makes sense to me as the logarithm gives us (3) $y=\alpha + \beta_1*log(x_1) + \beta_2*log(x_2) + \beta_3*log(x_3) $ but for the multiplicative model we receive (4) $y=\alpha + log(\beta_1*x_1) + log(\beta_2*x_2) + log(\beta_3*x_3) $.

Can we estimate a regression like (4) with standard OLS? How do we tell the statistics software the difference to (3)? How would a regression estimation like this look like in R using the lm command?

  • $\begingroup$ Model (1) isn’t identified. It’s indistinguishable from $y=\beta x_1 x_2 x_3$. $\endgroup$ – The Laconic Dec 20 '18 at 13:15

There is no need to "linearize" the "multiplicative" form. You have $\alpha\beta_1\beta_2\beta_3x_1x_2x_3$. So simply multiplying all three x variables, then the coefficient of their product is $\alpha\beta_1\beta_2\beta_3$. In this setup, you're assuming there is no intercept. For the exponential form, everything you have in (3) is correct other than the $\alpha$ in the log solution should be $\ln(\alpha)$.

In R, you'd have the first setup as:

lm(y ~ 0 + x1:x2:x3) # 0 + takes out the intercept

You get one coefficient, it is $\hat\alpha\hat\beta_1\hat\beta_2\hat\beta_3$

For the second setup:

lm(log(y) ~ log(x1) + log(x2) + log(x3))

You get four coefficients: $\{\hat\delta,\hat\beta_1,\hat\beta_2,\hat\beta_3\}$ where $\delta=\ln(\alpha)$.


You can rewrite the model

$$(1) \ \ y_i = \alpha\beta_1x_1\beta_2x_2\beta_3x_3u_i$$

by taking logs

$$\log y_i = [\log \alpha + \log\beta_1 + \log \beta_2 + \log\beta_3]+ \log x_1 +\log x_2+ \log x_3 + \log u_i$$

and in estimating the equation

$$\log y_i = \lambda_0 + \lambda_1 \log x_1 +\lambda_2 \log x_2+ \lambda_3 \log x_3 + e_i$$ you would get an estimate of $\lambda_0$ $$\lambda_0 = [\log \alpha + \log\beta_1 + \log \beta_2 + \log\beta_3]$$

as the Laconic says $(\alpha,\beta_1,\beta_2,\beta_3)$ are not seperately identifiable.

You would also get estimates of $\lambda_1,\lambda_2,\lambda_3$ and could test if they are all equal to 1 as they should be if $(1)$ where the true model.


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.