Maybe you have too many predictors and you are over fitting? How many cases and predictors do you have? Maybe you have to do feature subset selection as part of your modelling and evaluate it together with your model. See:

http://www.researchgate.net/publication/215739042_Fish_recruitment_prediction_using_robust_supervised_classification_methods

Article Fish recruitment prediction, using robust supervised classif...

Try looking up John Boscardin's paper on Harrell's Optimism for Logistic Regression. He and his group have developed a nice SAS macro that spits out a lot of nice statistics for model selection while bootstrapping. It can be a little tempermental, but it might be worth your while to figure out your way around it.

Which algorithm are you using to estimate the coefficients of your logistic regression mocel? Least squares? Mínimum Chi²? ...

Dear Wen Wang,

Kindly see the attachment PDF for your doubt.

Hi Jose,

Thank you for your response!

I had around 1000 cases and around 400 predictors orginally. Then I selected around 50 less redundant predictors.

One thing I didn't mention in my orgianl post is that these predictors were generated from raw data based on a certain association algorithm, so many individual predictors itself can predict outcomes pretty good - not in perfect manner. I want to see if I can get better prediction by combining these individual predictors.

After I posted this question I split my predictors to two groups based on their positive/negative  association to the outcomes and ran logistic regression on two sets of predictors separately. Then the perfect prediction problem is gone. 

But I don't understand if the perfect prediction is caused by any individual predictor, how come it's just gone after I split them to two groups?

Hi Nora and Senthilvel,

Thank you for the information! I read through all of them and they're very informative.

Hi Noel,

I used least squares. Thank you!

Hi wen wang, thank you for sharing your problem. I suggest you that you can ran logistic regression of all predictors with backward elimination. you will get a set of potential predictors. Then you can apply bootstrapping for the model with significantly associated predictors variables.

Please, I need help!

My regression analysis produced a perfect model after using Ten independent variables on a dependent variable. Could it be that the variables are too many or there is a specific technique that I lack to go around it?

Hi Arinze,

Did you use any cross-validation methods to evaluate your model?

Thanks,

Wen