I am working on a project to predict when eggs will hatch based on the environment. To evaluate how accurate the model is, I plan to use the residuals from the predicted date of hatching and the actual date of hatching. For example, if the model predicts an egg will hatch on Sept. 10th and it actually hatches on Sept. 15th, then I can say this data point has a residual of 5. Analyzing these residuals will be extremely straight forward with many methods available.
The problem I have is that sometimes the eggs never hatch (i.e. the environment is ill-suited for the eggs so some of them die). I cannot simply throw these data points out since the best model should be able to 'predict' this. For example, if one model predicts an egg will hatch on Sept. 10th, a second predicts it will hatch on Sept. 20th, and the egg never hatches, then I can say that the second model preformed better than the first. This is because of complicated biology reasons that I do not think are particularly relevant, but I will go into if anyone asks.
I do not see how to quantify the residuals for these unhatched eggs. My best guess is that I am going about this whole process wrong. Maybe analyzing residuals is not the best way to determine how accurate my model is. If I treat hatching vs not hatching as a presence–absence data, then there are a few techniques available to me. However, doing this ignores the first case, where eggs do hatch and residuals are available. I really have no good ideas on how to go about this. What should I be doing?
