I was wondering if it is possible to make any economic implications by regressing the same set of independent variables with different outcome variables.

For instance, regressing [Industry, Years, Production Level] on [Revenue, Profits].

Given the above variables, if both models are significant, what would be the implications? Does it imply that [Industry, Years, Production Level] can be used to explained both the change of [Revenue, Profits].

If it is, what are the appropriate regression models? Are there any concerned issues regarding the results?

  • $\begingroup$ I think both models will be significant because given modest scalability Production Level will correlate with Revenue and given some nearly constant profit margin Profits will correlate Revenue. $\endgroup$
    – Giskard
    Jun 3 '15 at 6:03

Using matrix notation, denote profits by $\Pi$, revenues by $R$, and the explanatory variables by $\mathbf X$.

Assuming a linear setup, we look at

$$\mathbf \Pi = \mathbf X\mathbf a + \mathbf u \tag{1}$$

and $$\mathbf R = \mathbf X \mathbf b + \mathbf v \tag{2}$$

Are profits correlated with Revenues? Experience say they are. So there is also a relationship

$$\mathbf \Pi = \gamma \mathbf R + \mathbf \varepsilon \tag{3}$$

Inserting $(2)$ into $(3)$ we get

$$\mathbf \Pi = \mathbf X(\gamma\mathbf b) + \gamma \mathbf v + \mathbf \varepsilon \tag{4}$$

Looking at $(1)$ and $(4)$, we have the equivalence

$$\mathbf a = \gamma \mathbf b \tag{5}$$

Will we obtain it through estimation? Not that easily.

Least -squares estimation will give us, for $(1)$

$$ \mathbf {\hat a} = \left(\mathbf X'\mathbf X\right)^{-1}\mathbf X'\mathbf \Pi \tag{6}$$

and for $(2)$

$$\mathbf {\hat b} = \left(\mathbf X'\mathbf X\right)^{-1}\mathbf X'\mathbf R \tag{7}$$

Insert $(3)$ into $(6)$ to get

$$\mathbf {\hat a} = \left(\mathbf X'\mathbf X\right)^{-1}\mathbf X'\left(\gamma \mathbf R + \mathbf \varepsilon\right) $$

$$\implies \mathbf {\hat a} = \gamma \mathbf {\hat b} + \left(\mathbf X'\mathbf X\right)^{-1}\mathbf X'\mathbf \varepsilon \tag{8}$$

The interesting realization comes at this point: given the postulated relation $(1)$ it is unlikely that the variables in $\mathbf X$ will be uncorrelated with $\mathbf \varepsilon$, the error term in $(3)$.

This means that even with a large sample, calculating the ratios of the individual coefficients in the vectors $\mathbf {\hat a},\;\mathbf {\hat b}$, we will not recover $\gamma$, since $\left(\mathbf X'\mathbf X\right)^{-1}\mathbf X'\mathbf \varepsilon$ won't converge to zero.


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.