Does high collinearity (>0.9) have any consequences on convex hull and Gower's distance calculations? I'm analyzing a dataset that includes time, time^2 and time^3 and this seems to be an issue. I'm not officially part of the class so feel free not to reply. Thanks a lot, Fabrizio --------------------------------------------------------- Fabrizio Gilardi Weatherhead Center for International Affairs Harvard University 1737 Cambridge Street Cambridge, MA 02138 (USA) Phone +1 617 495 1467 Fax +1 617 495 8292 fgilardi at wcfia.harvard.edu www.fabriziogilardi.org --------------------------------------------------------- _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

That's what I thought. If two variables are very similar then it shouldn't matter. But then why do I get very different results when I run these two models and counterfactuals? z.out1 <- zelig(y ~ x1 + x2 + x3 + time, model="logit", data=data) z.out2 <- zelig(y ~ x1 + x2 + x3 + time + I(time^2) + I(time^3), model="logit", data=data) range.x1 <- seq(min(data$x1), max(data$x1), length.out=100) x.out1 <- setx(z.out1, x1=range.x1) x.out2 <- setx(z.out2, x1=range.x1) cfact1 <- whatif(~ x1 + x2 + x3 + time, data=data, cfact=x.out1) cfact2 <- whatif(~ x1 + x2 + x3 + time + I(time^2) + I(time^3), data=data, cfact=x.out2) cfact3 <- whatif(~ x1 + x2 + x3 + time, data=data, cfact=x.out2) - In cfact1 about 80% of counterfactuals are in the convex hull - In cfact2 no counterfactual is in the convex hull - In cfact3 (which is based on the model with time^2 and time^3 but evaluates the counterfactual without time^2 and time^3) again about 80% of counterfactuals are in the convex hull - Adding / dropping the other variables makes no difference. It is really time^2 and time^3 that matter for the counterfactuals. What is going on here? Again, feel free to ignore this message. Thanks a lot, fabrizio On 28 Apr 2008, at 20:03, Gary King wrote: --------------------------------------------------------- Fabrizio Gilardi Weatherhead Center for International Affairs Harvard University 1737 Cambridge Street Cambridge, MA 02138 (USA) Phone +1 617 495 1467 Fax +1 617 495 8292 fgilardi at wcfia.harvard.edu www.fabriziogilardi.org ---------------------------------------------------------

...I wonder how terrible it would be, comparatively speaking, if one was to impute five times then match on each of the imputation data sets, then run the analysis on each of the 5 imputed and then matched data sets, and lastly average the results. Does anyone have any idea how well that would do? Johannes _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l

it wouldn't be terrible but it might be a pain since you'd likely have different numbers of observations for each imputed data set. why don't you instead temporarily create a data set that is the average of the 5 imputed data sets (so the observed data are the same as in each of the 5 and the missing data are the averages of the imputations). then match on that so you know what obervations to delete. then delete those observations from each of the original 5 data sets and use those. this is also a hack but would be less of a pain. Gary On Tue, 29 Apr 2008, Johannes Castner wrote: the _______________________________________________ gov2001-l mailing list gov2001-l at lists.fas.harvard.edu http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l