Suppose I have a classification problem about whether or not a person has blue eyes. There is a training set with 87,000 observations and a test set with 22,000 observations. Now suppose after fitting a model and predicting on the test set I get this confusion matrix:
So I get a false positive rate of 25% and a false negative of 2%.
Now I run the model on a much bigger data set, 5 million observations.
If my goal is to estimate the total number on persons with blue eyes in this population, can I correct this number by the false positive and false negative rates?
In this example that would mean reducing the total number I got by 2% and then adding 25%.
My question is really how much can I rely on the results from the test set and is it acceptable to correct the total number by these results.