Announcement

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

  • GEE and Fixed Effects Tobit

    Hello everyone,


    I have been doing extensive reading for some time to decide upon some appropriate models for analyzing my data, but I could use some input from others. I have a dataset that examines firms as the unit of analysis over the course of 8 years (t=8, n= ~23000). Each firm has many observations at each time point. The dependent variable is often zero (75-80% of the time), however, these are observed zeros and technically not quite censored values. Additionally, no firm has all zeros within panel and no firm has all zeros in a given time period. To be clear, the DV could be >0 for all firms in the sample for a given observation and each firm is in the business of potentially providing a non-zero value on the DV. The dependent variable is observed in dollars. I am inquiring about the following analytical techniques (GEE and Tobit) for my data. I understand they may not be the only models I can employ, but I am curious about their suitability specifically.

    For the primary analysis technique, I am considering either GEE or fixed effects Tobit, and I may use both with one or more being used for robustness.

    I think xtgee could be a good command being that I could xtset on firm and control for within firm dependence of observations. Also, from reading, GEE appears to stable with regard to distributional assumptions and I believe it does not assume normality among residuals. Would this be one appropriate way to approach my data?

    I am also considering FE Tobit. It seems like this could be a good way to go since Tobit can handle many zeros on the dependent variable. However, I understand there is a incidental parameters problem with FE Tobit, but it seems that there is not necessarily agreement on this point. Also, given that I have a large sample size and a reasonable T, it is my understanding that using FE Tobit should not be an issue. Can anyone shed some insight on this? I would love to hear an informed opinion on this point.

    For additional robustness, I am looking at running a fixed effects linear model, which I think should be fine. But I am curious to hear others thoughts on xtgee and FE Tobit for my data.

    Thank you all in advance!


    Last edited by andrew rich; 02 Dec 2018, 18:46.

  • #2
    Andrew, what is the dependent variable you are looking at and what is its expected range? (Knowing both those things might help people in making recommendations). For example, it might be helpful for others to know if you are predicting something like how much the firm spends on R&D or invests in lobbying (i.e. lots of real zeros but firms could also be millions of dollars.)

    My (very limited) understanding of Tobit models is that it assumes that the process that generates the zero observations is the same one that generates the positive values (just that no one observes the negative part of the distribution because it is truncated to zero), If that's not the case, then you might want to use some sort of 2-stage model, where the 1st stage predicts whether the firm makes R&D investments and the 2nd stage predicts, given that a firm makes an investment, how much it makes.

    Comment


    • #3
      Hello David,

      The DV is lobbying investments. All of the firms in the sample make lobbying investments and do so for different strategic reasons, so it's not a case where some firms are engaged in R&D expenditures and some firms are not, for example.

      Thank you.

      Comment


      • #4
        You'll increase your chances of a useful answer by following the FAQ on asking questions - provide Stata code in code delimiters, readable Stata output, and sample data using dataex.

        Your original posting said there were many zeros, and then you say all firms make such expenditures - does this mean there are no zeros? If all firms make such expenditures in all periods, then there is no need for a tobit. I generally find the xtreg, etc. routines easier to use than the GEE approach but this may be a matter of personal preferences. Some of the xt routines have gee equivalents that give the same output. While some recommend i.panelvariable in tobit as an approximation to a fixed effect tobit, as I understand it there is no statistically justified fixed effect tobit.

        Comment


        • #5
          Hello Phil,

          Thank you for your thoughtful reply. Both are true, there are many zeros on the DV and all the firms in the dataset make such expenditures each year. Expenditures are at the state level, however, and not all states receive the expenditures each year so they are a 0 on the DV. However, there is a theoretically valid reason for keeping these zeros in the dataset so I do not want to just dispense with them.

          Thanks

          Comment

          Working...
          X