How to get CV% for an unbalanced Precision study

[#1]

Hello JMP community,

I have some data for Precision study with an unbalanced design. As shown below, there are ten runs performed with reagent lot 2 and 3 on two instruments (A and B). The fully automated instrument has three docks (locations) for the 96-well plate. The same sample was tested with 3 replicates in each of the ten runs by one operator. The goal is to calculate CV% for intra-run, inter-run, inter-lot, inter-instrument, inter-dock (plate location), and total CV%.

Below is the screenshot of the data in JMP. (The file can be found in the attachment. ) 

Below is what I did for the analysis. I selected "Crossed" for the Variability Model.

My questions are:

1. Is this the right way to analyze the data?

2. How to get the CV% for each factor based on REML Variance Components Estimates results?

3. Only Run and Run*Replicate showed some values under Var Component. The others are all 0. I feel this is not quite right. Something wrong here?

4. Since the design is not balanced, what else needs to be considered for the analysis to account for that?

Thank you very much!

[#2]

I think your model is over-specified. You should omit Replicate. It will automatically be treated as the lowest level in your hierarchy of factors. Try repeating your analysis without explicitly including Replicate as a factor.

You will still have a problem. You have more than an imbalance. You have missing combinations that are required to estimate some terms.

[#3]

Thank you, Mark, for your comments. I most likely need to repeat this experiment. Could you share some insights regarding the major combinations that were missing so I can have a better design?

Also, I wonder how to get the CV% for various terms based on the variance components estimates below.

Thanks again.

[#4]

Hi @Jinweidu80 : You will not be able to estimate any of the interactions. To estimate an interaction between any two factors F1 and F2, each level of F1 must be done with each level of F2. For your three factors (reagent_lot, Instrument, and Dock), this requirement does not hold; so, as @Mark_Bailey said, you have missing combinations that are required.

IMHO...all you can really do is have a model with "Run" (random) as your only effect.  And even though you can do that, how will you interpret the results?

• The first six runs are either 3-A-1 or 3-B-1 (reagent lot-instrument-Dock).
• The next two runs are 2-A-1.
• Your last two runs are 3-A-2 and 3-A-3.

If there is/isn't "significant" variability between runs, how will that be interpreted wrt reagent lot, instrument, and Dock?