[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Clarification: Instrumental variable design: no control group

From   John Antonakis <>
Subject   st: Clarification: Instrumental variable design: no control group
Date   Sun, 28 Sep 2008 22:55:26 +0200

Hi again:

Sorry, what I said below about the raters is unclear, as was my explanation on the design. I would just like to clarify two things. Raters were in groups, and the raters rated 4 participants (hence the rater fixed effects). The y variable is actually a speech; Time 1 and Time 2 speeches were randomly assigned to the groups of raters (but groups never saw the same participant twice). The DV was measured two different ways: an absolute measure (how good the participant was at giving the speech) and a relative measure (the raters rank ordered the 4 speeches from best to worst).


Here goes again.................

I have estimated the following model in STATA:

(where participant i is measured on two occassions)

xi: xtivreg y (mv = prepost) controls i.rater i.participant, re robust


prepost = a dummy variable (0=pretreatement, 1=posttreatment)

mv = the manipulated check (i.e., a manipulation check to see whether
the participants actually responded to the treatment)

controls = refers to participant individual difference controls (e.g.,
personality variables, age, set)

i.rater = rater fixed effects (given that groups of raters rated the
participant on the DV)

i.participant = participant fixed effects (given that participants were
rated twice)

Now this design does not have a control group. It would have been
stronger if it did, but I could not randomly assign individuals to a
treatment and control. Everyone had to receive the treatment. The way to
mitigate this limitation was to measure the manipulation and then
regress the DV on the manipulation (instrumented by the prepost dummy).
My thinking is that the manipulation check serves as the independent
endogenous variable and the intervention dummy variable as the exogenous
treatment. If the change in y is predicted by the manipulation check,
which in turn is predicted by the prepost treatment, then I can be more
certain about our causal deductions as compared a simple repeated
measures design in which the manipulation check does not serve as the
endogenous independent variable. Because the manipulation check is
modelled as a predictor, other putative causes can, for most intents and
purposes, be ruled out.

Has anyone seen a published example of such as design that they can
point me to? If so, is there a particular name given to this design or
is it simply a IV design? Also, any suggestions on how to improve causal



Prof. John Antonakis
Associate Dean
Faculty of Business and Economics
University of Lausanne
Internef #618
CH-1015 Lausanne-Dorigny

Tel ++41 (0)21 692-3438
Fax ++41 (0)21 692-3305

* For searches and help try:
* For searches and help try:

© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index