4
$\begingroup$

I have a quick question regarding dummies for a regression that I am running.

I am running regressions of responses to a survey question about health on a vector of independent variables (e.g., unemployment, age, income). I am also considering including controls for the month in which individuals were interviewed. I expect that responses to the survey question about health might vary seasonally.

My "month of interview" variable is a continuous variable that runs from 1-12 (based on the actual month of interview). Would it make the most sense to include this variable in my regression as a continuous variable, or to include 11 "month" dummies? On the one hand, if I use the month variable as a continuous variable, I can still pick up long-range patterns ("beginning of year" vs. "end of year"). On the other hand, the month dummies might be more informative. However, they would make my regression output table long and unwieldy.

Alternatively, I could alter the month dummy into a four-category dummy with four seasons.

Any advice is greatly appreciated!

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ It depends a lot on your data and whether there really is a year-round trend. The simple answer is to try both and cross-validate to determine which generalizes better. $\endgroup$
    – Arya McCarthy
    Commented Apr 3, 2021 at 16:25
  • $\begingroup$ Thanks! Unfortunately, I don't have enough sophisticated statistical knowledge to know how to cross-validate or what this would involve. But I appreciate your advice! $\endgroup$
    – Annika Jason
    Commented Apr 3, 2021 at 16:34

1 Answer 1

Reset to default
3
$\begingroup$

Options:

  1. Continuous variable 1-12: This would assume a linear trend through the year. It is unlikely that the effect of this covariate in December will be 12x the effect of the covariate in January.
  2. Categorical variable (dummy encoding): This method affords a good deal of flexibility in how the seasonality is represented. I recommend this.
  3. Quarterly variable (1-4 or dummy encoding): I don't recommend this unless you know that your seasonality is really related to the "seasons" or to something like "quarterly earnings reports"
  4. Sinusoidal variable: I might use this if I had multiple years of data and thought there was an up-down pattern throughout the year that repeats.

In the end, I recommend the categorical variable approach unless there is a strong reason for one of the other options.

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thank you! That's a great clarification, I appreciate it! $\endgroup$
    – Annika Jason
    Commented Apr 3, 2021 at 18:22
  • $\begingroup$ 5. Combine option 1. with a spline $\endgroup$
    – kjetil b halvorsen ♦
    Commented Jul 5, 2021 at 18:24

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.