Re: st: IV and matching estimator

["Dimitriy V. Masterov" <dvmaster@gmail.com>]

Ayman,

I am not so sure. Lagged impressions for the same user are a function
of the unobserved error, so including them is problematic since they
are not exogenous (assuming the unobserved error is constant across
time).

The other IV you have in mind may not work either. A valid instrument
needs to wiggle X (impressions) for reasons that have nothing to do
with the error. Your clones see a lot of impressions for the same
reasons you do: they are enthusiastic consumers.

Think about this in the following way. We are interested in learning
the effect of the seeing the trailer on going to see the Hobbit movie.
I am very excited about this movie. I am going to see it anyway. I
watch the trailer several times. You look at my friends, who are also
nerds. They watch the trailer. You then predict that I also see the
trailer since all my friends did as well. You use that prediction in
the model. Does that allow you to identify the causal effect? Probably
not so much.

DVM
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/