No announcement yet.
  • Filter
  • Time
  • Show
Clear All
new posts

  • Ommited Industry Diversification - Fixed Effects Model

    Dear all,

    according to Allayannis and Weston (2001) I am trying to do a fixed effects regression with an industry diversification (idivers) variable. The variable is 1 if the firm operates in two or more industry segments and zero otherwise. For classifying in different industry segments I am using SICCODES.

    I am using the following setup:
    tsset id year
    xtreg y x1 x2 x3 idivers i.year, fe cluster(id)
    I want to control for time and industry fixed effects. Therefore I have some questions:

    1. As I learned so far I can't include i.industry in the model because the company effects, which I declare via "tsset id" already contain these industry effects, right?

    2. Stata always ommits my idivers variable because of collinearity. Is this industry diversification also contained in the company effects?

    3. How can authors like Allayannis and Weston estimate a fixed effects model containing a industry diversification variable and controlling for time and industry effects?

    Best regards,


  • #2
    1. & 2.: as per -fe-machinery, time-invariant predictors are wiped out (hence, if -i.industry- does not change within the same panel as time goes by, it is unavodably cancelled out);
    3. I do not know the paper you mention, so I cannot reply on it. If you can't track down any clues in the methods section of the article, ask Authors directly via email.
    Last edited by Carlo Lazzaro; 14 Jan 2019, 09:44.
    Kind regards,
    (Stata 15.1 SE)


    • #3
      Instead of saying "contain these effects," which is somewhat vague, think of it in these terms: if the variable you want to add is the same in every observation of a given firm, then it will be colinear with the firm fixed effects and it will be omitted. So you can only add industry diversification to your model if at least some of the firms experience a change in their industry diversification during the course of their observations in the data.

      As for question 3: first, Allayannis and Weston 2001 may be folklore in your shop, but this is an international, multi-disciplinary forum and many people here (most, I imagine) have no idea what this is. Please read the Forum FAQ before you next post here. Among the things you will learn there is that if you are going to refer to a publication you need to provide a complete reference. Next, understand that even with a complete reference, the number of people who can access the publication without paying for it and investing a big chunk of time in reading it may be limited, so you may not get an answer here. If you don't, your best bet may be to try contacting the authors.

      Added: Crossed with #2, which says more or less the same things, more succinctly.


      • #4
        1. You cannot include so called "nested" fixed effects. Imagine it this way. Your company fixed effects (id) demean all variables by company. This means that the average value for each variable for each company will be zero. Now take the average over industries. This is just an average of the firm averages. The average of a bunch of zeros is just zero. Therefore, including industry fixed effects doesn't make sense (as it's goal is to make the average zero, which it already is).

        2. This depends on how you calculated the idivers variable. Does it change over time? Does it change over industries? Does it change per industry per time period?
        3. This is the same question as 2.