JMP Genomics ANOVA and One-way ANOVA

[#1]

Hi, I am using ANOVA and One-way ANOVA in JMP Genomics for simple RNAseq analysis where we compare two groups. The p-values are quite different from each other in both analyses, and also we have not been able to replicate them in R packages. It would greatly help to know e.g. which ANOVA or One-way ANOVA approach is used (e.g. Wilcoxon).

1) Why would the results differ for the ANOVA and One-way ANOVA?

2) Where can I find more information on exactly which ANOVA method is used?

Thank you so much for your consideration. I hope my question makes sense.

[#2]

Hi @AS ,

Which version of JMP Genomics are you using by the way?

JMP Genomics user guide might assistance in answering these questions. Take a look here. It is a description of One-way ANOVA and how it compares to ANOVA. It includes which SAS PROCs are used.

The  p-values will differ from the many packages in R which implement one-way ANOVA and ANOVA in different ways. Take DESeq2 for example. To get a similar result to DESeq2, make sure that the option Shrink variances using Empirical Bayes it turned on within the Options tab of the One-way ANOVA dialog box.  This will be the main difference between different R packages and JMP Genomics. There will also be differences in application of Multiple Test Correction methods as well.

I hope that helps.

[#3]

Hi Chris, thanks for your fast feedback and consideration. I am using JMPGenomics 9, with JMP13.2.1.

A colleague of mine just used R base function aov to try to replicate the one-way anova results with the same expression values I had used and was not able to replicate them, but taking it from your advice I think there some settings which are different. I ran ANOVA (with diagnosis as fixed effect) and One-way ANOVA from within JMPGenomics and get different results for the p-values. Not sure what to use. Would you have advice how to chose the most suitable method? This might be hard to tell, without knowledge of all the settings.

[#4]

Hi @AS ,   Thanks for your question.  Both ANOVA and One Way ANOVA in JMP G use standard linear mixed model approaches based on normal theory to calculate the resulting t-and F-tests.   When differences occur with results from other packages, the typical place to look is in the degrees of freedom approximation used behind the t- and F-tests.   Different approximations for these are possible and there is not necessarily one best answer, especially if there are missing data.   If you want to break things down further, consider the numerator and denominator of a specific single-degree-of-freedom t-test for a few select genes to really get to the bottom of it.

Suggest transforming the p-values to -log10 scale and plotting them versus each other.   They should be highly correlated although not perfectly so, providing basically the same ordering of the genes.   If they are radically different, something may be off in the model setup itself in terms of the factors specified.   Tech support would need more specifics to help you decipher exactly what is going on in your situation if you want to contact them.   

One Way ANOVA in JMPG performs its calculations directly in a SAS data step for speed, although it is limited to one-way designs with one blocking factor.  Note though that multi-way designs can be converted to one-way by creating a single super factor that has all possible levels.    ANOVA in JMPG calls Proc Mixed with BY groups.    

Remember also to use the log2 fold change itself (numerator of t-statistic) to sort genes for reproducibility.   Those with larger fold changes are typically more reproducible.    In volcano plots, look for genes in the upper left and right corners.