Announcement

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

  • Dynamic panel regression (xtdpd) with only time-series fixed effect without cross-sectional fixed effect

    Dear All,

    I am struggling to find out how I can run a dynamic panel regression (xtdpd) with only time-series fixed effect without cross-sectional fixed effect.

    Currently, I have a code as follow:

    xtset id year
    xtdpd f L.f ret year1986-year2013, dgmmiv(f).

    As far as I understand, this code estimates the time fixed effects of year1986-year2013 within each id by default. However, I was wondering if I could estimate the time fixed effects for all id's, i.e., with no cross-sectional fixed effect in the dynamic panel regression. It doesn't need to be xtdpd as long as I can consistently estimate the dynamic panel with only time-fixed effects included.

    Any comment at this moment will be greatly appreciated, and please fee free to let me know if anything is not clear in my question.

    In case helpful, I attach the result of the current code. (Note that too many instruments would be my next question in another posting.)

    Many thanks for your generous help and I look forward to talking with you soon.

    Kind regards,
    Minsoo
    Attached Files
    Last edited by Minsoo Kim; 01 Mar 2016, 15:13.

  • #2
    First of all, the question is why you do not want to have cross-sectional "fixed" effects and how your desired econometric model looks like. In particular, do you still assume that your errors are serially uncorrelated even without time-invariant cross-sectional effects? In that case, you can just use the plain regress (or ivregress) command because your lagged dependent variable would no longer be endogenous. However, this would be a very strong assumption on your error term. (You can, of course, test for serial correlation.)

    Otherwise, you need to make assumptions about the covariance structure of your error term. If your errors exhibit first-order serial correlation (due to whatever reason), the lagged dependent variable is no longer exogenous. In principal, you can use the xtdpd command to specify appropriate instruments (in particular lags of the dependent variable that reach sufficiently far into the past such that they are uncorrelated with the contemporaneous error term). Technically, the xtdpd command does not impose a fixed-effects structure of the error term (although it was primarily constructed for this purpose). All you need to do is to choose valid instruments for whatever structure of your errors you assume.

    Please have a closer look at the help file and the manual for xtdpd as well as the related literature if you want to proceed along these lines. You do not only have to specify instruments for the endogenous variables but for all right-hand side variables in your model. In particular, for your (exogenous) time dummies you would need to specify the option liv(year1986-year2013) which says that they are "instrumented" by themselves. Similarly, you need to specify appropriate instruments for your ret variable. Again, how exactly you specify those instruments, and whether you do it for the first-differenced equation or the level equation, depends first and foremost on your model assumptions.

    A final advise on the number of instruments. The dimension of your longest time series (27 periods) is relatively long. This explains the large number of instruments resulting from your dgmmiv() option. (Again, please make yourself familiar with the help file and the manual.) You might want to reduce the maximum number of lags used, as an arbitrary example: dgmmiv(f, lagrange(2 4)).
    https://twitter.com/Kripfganz

    Comment

    Working...
    X