Announcement

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

  • Addressing endogeneity using difference in independent variable

    I’m working with panel data where the variables are group level indicators of performance. To put simply, the predictor is a group-level aggregated quantity (e.g., average reputation of members) which is time varying over several periods (the predicted variable being group performance). I have reason to believe that the predictor is not strictly exogenous since at times the group is constituted with an aim to make it perform well. However, a “part” of the predictor is exogeneous – it happens when a group member suddenly exits the group in one of the periods (death or some reason, which is strictly exogenous). So, for identification, I am thinking of creating two components of the predictor in my dataset: the first is the group level (reputation) measure assuming no exogenous shock – i.e., the group member has not left the group), and the second component would be the delta(predictor) ONLY there is an exogenous shock (death or some other reason) – this delta(predictor) would be a negative quantity if the exiting group member has an above-average reputation, and would be a positive quantity if the exiting group member has a below-average reputation. In any case, the second component would be the exogenous component of the predictor – and its coefficient should be ideally significant when testing for the proposed hypothesis. Now having said this, to slightly complicate the matters, I am using Cox regression (predicted is a duration variable) with time-varying covariates, BUT that is beside the point since the essential question I have from you all is whether my strategy makes sense.

  • #2
    Reza:
    welcome to this forum.
    My jerk-knee reaction to your post is: why not choosing (and testing) (at least) one instrument for this (potentially) endogenous predictor?
    What above implies to switch from -stcox- to a panel data command, though.
    Kind regards,
    Carlo
    (StataNow 18.5)

    Comment

    Working...
    X