Running multivariate principle component analysis with missing values

[#1]

I'm using JMP 16.2.0 to run a principle component (multivariate) analysis through the Analyze menu option. My data contains 4 reps of 15 treatments each in two categories (i.e. 120 reps in total). The data has undergone a Grubbs test and the outliers have been removed from the reps resulting in gaps in my data table. When I run the PC analysis for all the variables as an overall analysis and 'by category' I have no problem. However, when I run it 'by treatment' (see attached image) I get the following error for each treatment with missing values: The Default estimation method was not used because the data contain missing values and the number of variables exceeds the number of observations. Estimation method switched to Pairwise.

Q1: Is there another way to run the PC analysis with the missing data? This will have to be without coding as I'm fairly new at this and don't know how to properly code in JMP/SAS.

Q2: If I run all of it pairwise, will this still give me reliable data?

[#2 · Accepted solution]

Before using a multivariate analysis, you used a univariate outlier test on a small sample (4 replicates). What does the multivariate analysis look like using all the data? The apparent univariate outliers might be correlated in higher dimensions.

What is Treatment? Are the other variables the responses to the Treatment changes?

Have you read the documentation for PCA? See the description at the end of this chapter to understand the different methods used to estimate the variance. You should not need scripting to obtain the correct analysis.

[#3]

Thanks for the reply. The multivariate PCA output using all the data shows higher correlation (default method) than without the outliers (pairwise method). There is no difference between the output when using the pairwise and default methods on data where no outliers were detected.

Yes, treatments with different fertilisers were measured to obtain nutrient values in response to the treatments (variables).

Thanks for the info - this helps a lot!