Given a dataset with n variables and a binary outcome:
Are there assumption violations/risks to drawing conclusions if the entire dataset is used to determine which variables are valid if the second part of the analysis will rely on generating a classification algorithm using test and training datasets?
In essence, can I determine odds ratios with the entire dataset and then go and develop a classification algorithm by sampling to create with test and training sets?
Or should I be using the training set to determine the odds ratios of a given variable?
Any feedback is appreciated, I'm hoping not to collect another tumbleweed badge.