How to do external validation of regression models Very basic question here, so bear with me... 
I have a data set with 241 patients with 16 variables plus diagnosis (malignant vs benign). There are 3 previously published logistic regression formulas in the field:


*

*modelS ~ A + B + C + D + E + F 

*modelH ~ modelS + G 

*modelG ~ A + B + C + H


I want to perform an external validation of each. Of course, I have plugged in the variables and have thus calculated the predicted probabilities of malignancy for each patient. 


*

*Is there a more formal way to compare the performance of these prediction formulas? Of course I have derived my own formula (following Harrell's methodology and using package rms). How can I properly compare my formula with the previously published ones? 

*Is it valid to take the parameters the previous formulas have found to be "significant" and derive the coefficients from my data set, thus deriving newModelS, newModelH and newModelG and then run stats (like C, likelihood chi2, etc) on those? I can also compare the newly replicated formulas by AICc,  but does that inform on the previously published formulas? 
The more general question is "How do you perform an external validation of a logistic regression formula"?
 A: Regarding part of #1, perhaps a better and more formal way to proceed is to put in a variable that is the logit of a published model, and add to it all of its component variables less one term.  Do a chunk likelihood ratio $\chi^2$ for the added value of all the components.  That is a test of lack of fit of the published model.
Regarding comparing performance on an external dataset, you can build a model with the logits of two models and see if each one adds predictive information to the other using Wald or likelihood ratio $\chi^2$ tests.  Then there are methods of Pencina and others (R Hmisc package improveProb function).
A: Regarding 1. There may be, but I would guess that a formal method isn't very useful (although some journal editor or pointy haired boss may want one). Rather ask "Do the models look the same? Would anyone care about the differences?"
Regarding 2. I don't understand the question.
Regarding 3. Well.... what's wrong with writing down your model, writing down theirs, and looking? It sounds like you may be re-asking Q1. If the models are substantively different, you can then start to ask why they are. E.g. Different data sets, different models, different methods (e.g. you used splines and they used polynomials) etc. 
