I am running an experiment looking at feed timing and harvest timing (harvest is defined as the end of the culture duration or end of the experiment) for different cell culture flasks. As I design my DoE I realize that the harvest timing allows me to "cheat" a little bit with my conditions. Say I am looking at the viability (cell health so to speak) if I harvest on day 14 vs day 15. When I run my experiment, i can actually "harvest" a small sample on D14 and thus analyze each condition for both time points (days 14 and 15). This will save me # of conditions but I'm aware I may be breaking certain assumptions when JMP analyzes each condition separately (not true independent events). When I try and make a DoE I know JMP is really suggesting having separate conditions for D14 vs D15 harvest as opposed to just analyzing the same condition at 2 time points.
What are the shortcomings of this method? I know there's something called a time series analysis but I've never used it and in this case we are only looking at 2 time points.
The issue is randomization, not timing. What you are doing is not wrong. What you are doing is different. JMP needs to know about that difference when you design the experiment and when you analyze the data.
The reason is that by default JMP assumes that you reset each factor between each run. You always start over or from scratch. In your case, the results at two time points had the same staring point. They are correlated. If that correlation is not accounted for, then you can have more of both type I errors and type II errors.
The answer is to change the definition of the factors to be "hard to change." Then you can define the number of "whole plots." In your case, you are randomizing half as often (i.e., you get a pair of time points from each condition), so enter n/2 for the number of whole plots where n is the number of runs.
JMP automatically saves information to the data table so that the analysis is correct.
In addition to Mark's suggestion (which is to handle this as a split-plot design a very useful application) you could also treat those two measures as repeats within treatment combinations. The repeats could be used to estimate an average and a variance for the treatment combinations. Each of those Y's could be analyzed independently. There is also nothing wrong with treating them as two Y's as long as you are careful in interpreting the quantitative analysis (NID(0,variance) assumptions). See how well they correlate (multivariate analysis).