Stata Technical Bulletin
STB-33
Figure 2
In this example, using ses as an instrument for sbp has a dramatic effect on the results. Instead of having a weakly
significant positive effect on cholesterol level, systolic blood pressure is estimated to have a strongly significant negative effect.
Also, having instrumented sbp, the age variable works much better, having a strongly significant positive effect on chi.
Whether this revised model makes sense is beyond my expertise as an economist to determine. My intuition is that ses
might affect chi directly as well as indirectly through sbp. If that is true, ses should be included as a right-hand-side variable in
the regression and becomes unavailable for use as an instrument. However, experiments with including ses in this way and then
instrumenting sbp with the other variables in Garrett’s dataset have the same dramatic sign-switching effect on the coefficient
of sbp. So unless Garrett’s dataset is just random numbers, there may be something going on here that deserves a second look
by medical researchers.
Joanne Garrett’s contributions to helping Stata users interpret estimated coefficients include three additional equally useful
programs: Iogpred, presented with regpred in STB-26, and adjmean and adjprop, presented in STB-24. It would be desirable
to extend all of these programs in the same directions as regpred2 extends the capabilities of regpred. Hopefully some other
Stata user will do so and contribute the result to the STB or the statalist list server.
References
Garrett, J. 1995a. sg33: Calculation of adjusted means and adjusted proportions. Stata Technical Bulletin 24: 22-25.
-----. 1995b. sg42: Plotting predicted values from linear and logistic regression models. Stata Technical Bulletin 26: 18-23.
sg49.1 An improved command for paired t tests: Correction
John R. Gleason, Syracuse University, [email protected]
I have detected an error in the rmttest program described in sg49 in STB-30. Under certain conditions, the program could
report incorrect results when one of the two elements in a sample pair was missing.
Reference
Gleason, J. R. 1996. sg49: An improved command for paired t tests. Stata Technical Bulletin 30: 6-9.