Announcement

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

  • Difference in differences with PPML and Log-OLS reveal substantially different results

    Dear Statalisters,
    I have differences-in-differences setting where I have multiple treatments that occur at different times for different treated units. My main specification is the following one:
    Y_it= β_1 Post_it + δi + γt+ u_it where Post_it is the value of treatment for application i at week t, and δi and γt are application and time fixed effect parameters that are estimated.

    I use following commands to estimate the above specification:
    Code:
    xtpqml downloads post timefe*, fe i(app) cluster(app)
    and also:
    Code:
    xtreg ln(downloads + 1) post timefe*, fe vce(cluster app)
    However, although the number of observations are the same, results are very different: while I get a positive and significant coefficient (0.464) on the post variable in the first specification, I get a negative and significant coefficient in the second specification (-0,13)

    My dependent variable has the mean of 3.53, standard deviation of 16.7. Negative binomial (NB) fits better than Poisson specification, however given that no true fixed-effect estimator has yet been proposed in the NB, I use Poisson.

    Which specification I have to use?
    I can also attach the part of the data if it will make it easier
    Last edited by Erdem Yilmaz; 18 Mar 2019, 15:41.

  • #2
    Dear Erdem Yilmaz,

    The two dependent variables are very different and therefore it is not surprising that you get different results. If you care about the treatment effect on downloads, you need to stick to the first method. Note also that the fact that the marginal distribution of downloads is over-dispersed does not imply that NB is preferable to Poisson regression. Anyway, as you say, with FE you need to use Poisson.

    Best wishes,

    Joao

    Comment


    • #3
      Dear Professor Silva,
      Thank you very much for your reply, indeed I do cite some of your papers along with other references to justify the usage of ppml. Actually, one of the reviewers wrote me that: "Why don't you start by presenting the results with something simpler, like OLS with log transformed DV." . Therefore, I re-run some of the analysis with log-transformed DV. However, I am having hard time to explain why I observe what I observe. In most of the papers, just the magnitude of the coefficients change when people switch from log OLS specification to Poisson or the other way around. Do you have any suggestions for an explanation?
      Best regards,
      Erdem
      Last edited by Erdem Yilmaz; 19 Mar 2019, 13:14.

      Comment


      • #4
        Dear Erdem Yilmaz,

        That referee is a bit silly... Anyway, why don't you try using the logs without adding the 1? That will drop a lot of observations but that may be preferable to distorting the relationship as you are doing now. As for why the results are so different, that must depend no the nature of your data an your model.

        Best wishes,

        Joao

        Comment

        Working...
        X