What is the best way to reduce false negative percent in the model? My logistic regression model seems to identify successes very well (about 85% - 94%), but fails to identify the failures (only identifying 18% - 32% correctly). 
I have thought of weighting the success and failures differently, but I am not sure of how to do this R. I have tried:
Fr <- ifelse (y==0, 2, 1)
model <- glm(y~factor(x1) + factor(x2), weights=Fr, 
                                        family = "binomial", 
                                        data = data)

but did not get any improvement this way.
 A: Logistic regression really predicts odds, and as such, probabilities. The default predicted class is just the one with the highest probability. There is nothing really to prevent you from moving the probability threshold around from 0.5 to, say, 0.7, or 0.3 to get a better balance between false positives and negatives.
By the way: how does this post relate to your tag 'cross-validation'?
A: To second Greg Snow's comment, where things went wrong was considering classification accuracy.  Probability models are designed to estimate probabilities.  Classification is arbitrary and artificial, and seldom necessary except in the special case where you know the utilities of false "positives" and false "negatives".  Useful summary indexes include c-index = ROC area (but not the ROC itself, which seldom leads to good decisions) or its highly related Somers' $D_{xy}$ rank correlation coefficient ($D_{xy} = 2(c-.5)$).  There are also good $R^2$ measures.  Defer classification until the decision point; it's not necessary during the analysis.
A: Missclassification rate is not the best measure of fit.  Consider if you had a coin that came up heads 60% of the time, the optimal prediction for minimizing missclassification is to predict heads 100% of the time, but that tells you nothing of the underlying science that you realy want to learn about.
