Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Bertel Teilfeldt Hansen <statalistbertel@gmail.com> |
To | statalist@hsphsun2.harvard.edu |
Subject | st: Problems with plots from generalized sensitivity analysis (GSA) |
Date | Fri, 25 May 2012 13:18:38 +0200 |
Dear Statalist, I'm doing a paper about post-conflict elections wherein I use Masataka Harada's excellent ado-program gsa to examine possible bias from omitted variables. The program produces contour curves of generated variables similar to those of Imbens' sensitivity analysis (2003) and is described in detail in Harada (2012). The problem is that my contour curves turn out fine (and, as hoped, lie far away from other included variables), but the points representing gsa's saved variables are often extremely scattered. Some of them are very far from the observed variables, and some of them are very close. Also, gsa seems to save points which have the same correlation with the treatment, but very different correlations with the outcome, implying that they should not have an equally confounding influence on the treatment effect. This also happens the other way around (i.e. same correlation with the outcome, but different correlations with the treatment) and persists when the precision-option is set very low (meaning that the accepted span of influence for variables to be saved is narrowed). A sample of the commands I run are included below (I run the newest update of stata 12.1 on Windows XP and 7): Halving the coefficient on the treatment: - gsa outcome_binary treatment_continuous controlvar1 controlvar2 controlvar3 ... , vce(cluster unit_var) tau(0.08) nplots(10) observation(1000) precision(1) scatter - Making the treatment effect insignificant on the 0.1-level: - gsa outcome_binary treatment_continuous controlvar1 controlvar2 controlvar3 ... , vce(cluster unit_var ) tstat(1.645) nplots(10) observation(1000) precision(1) scatter - I'm thinking the problem might be caused by me having very few observations (I have between 60 and 70, depending on the specification), but other than that, I'm all out of ideas. I hope you can help! Thanks for your consideration! Best wishes - Bertel Ph.d.-student Bertel Teilfeldt Hansen University of Copenhagen Department of Political Science Email: bth@ifs.ku.dk Phone: +45 35324104 Litterature Imbens, Guido W. 2003. “Sensitivity to Exogeneity Assumptions in Program Evaluation.” The American Economic Review 93(2):126–132. Harada, Masataka. 2012 "Generalized Sensitivity Analysis." Working paper available here: http://home.uchicago.edu/~masa/docs/Harada-GeneralizedSensitivityAnalysis.pdf . * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/