1
$\begingroup$

Let’s say we would like to predict the sales of a company (Y), with using the size of the company and two month dummies. We only use one company over a period of 20 years.

${\rm sales}_t=\ \propto+\beta_1{\rm size}_t\ +\ \beta_2{\rm January}_t+\beta_3{\rm October}_t+\varepsilon_t$

1st question: Is it allowed to add an 2001 dummy, when there are already January and October dummies included in the regression equation?
If it is allowed and considering significance, how would the 2001 dummy be interpretated:

  1. For observations of the year 2001, the predicted sales are $\beta_4$ units higher/lower than in the other years, everything else constant.
  2. For observations of the year 2001 and not into January and October, the predicted sales are $\beta_4$ units higher/lower than in the other years, everything else constant.

2nd question: In case it is allowed to include the year dummy. Let’s say we want to know how the effect of size on sales is changing in the year 2001. To test this, an interaction term is created. ${\rm sales}_t=\ \propto+\beta_1{\rm size}_t\ +\ \beta_2{\rm January}_t+\beta_3{\rm October}_t{\ +\ \beta}_4{2001}_t{\ +\ \beta}_5{2001*size}_t+\varepsilon_t$

Which of the interpretations of the interaction term would be correct (if $\beta_1 {\ \beta}_5$ are significant):

  1. For every additional unit of size, the predicted sales are increasing/decreasing by $\beta_1$ units, but in 2001, the sales are increasing/decreasing by $\beta_1+{\ \beta}_5$ units, everything else constant.
  2. For every additional unit of size, the predicted sales are increasing/decreasing by $\beta_1$ units, but in 2001 and not in January or October, the sales are increasing/decreasing by $\beta_1+{\ \beta}_5$ units, everything else constant.

I am unsure if it is allowed to add the year dummy and if it is, then I am unsure if I need the "and not into January and October" when interpreting it. I couldn't find answers for that in books and in that forum.

Please let me know, when you have suggestions how I can improve my question.

Thank you!

$\endgroup$

1 Answer 1

0
$\begingroup$

1st question: Another graph is indicating, that size as well as sales are both highly varying in 2008 (financial crisis). Is it allowed to add an 2008 dummy, when there are already January and October dummies included in the regression equation?

Yes, no problem with that, since an observation that belongs to 2008 may or may not be in January or October. You are not "allowed" adding dummies whenever it creates a complete multicollinearity. E.g., adding dummy variable for each one of the $12$ months.

For the second question, you don't need the "not in January or October" part, since when you state that everything else hold fixed it includes both when the observation is, e.g., from January (i.e., January $= 1$), and when the observation is not from January (i.e., January $= 0$)$. In both cases the marginal effect of the other variable is the same.

$\endgroup$
4
  • $\begingroup$ thank you very much. So for both questions my interpretation "1" is correct? $\endgroup$ Aug 7, 2021 at 16:59
  • $\begingroup$ Yes, "1" is correct $\endgroup$
    – V. Vancak
    Aug 7, 2021 at 17:20
  • $\begingroup$ I was just wondering, if it necessary in question 2 to include the year 2008 outside of the interaction term. Would it be allowed if I remove the 2008 dummy and $\beta_4$ refers only to the interaction term with the 2008 dummy? Would this change the interpretation? $\endgroup$ Aug 9, 2021 at 12:15
  • $\begingroup$ Yes, it is allowed although not always suggested as the best practice. Generally, the $\beta_4$ stands for the so-called "main effect" of the year, while $\beta_5$ for the interaction. Hence, once you remove the year dummy, you will not be able to say something about the marginal effect of the year (itself), but only of the year-size variable. $\endgroup$
    – V. Vancak
    Aug 9, 2021 at 12:20

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .