I need to compare multiple groups

[#1]

Hi - I have a table (n ~ 12,000 rows) with a column (GROUP_NUM) that identifies the group to which the row belongs (n = 54 possible groups).  I have another column (COMPOUND_X) that contains the contaminant concentration for that row.  I am trying to compare the contaminant concentrations associated with all the rows for each group. I.e. which groups are most alike?

Note:  I am assuming my data are not normally distributed.   I am using JMP version 12.

How can I do this in JMP?  Thanks in advance!

Here is what my table looks like:

[#2 · Accepted solution]

Assuming it is available in JMP 12 (I know it was available in at least JMP 14), you could use the Nonparametric Multiple Comparisons in the Fit Y by X platform. This will give you a table of similarities and doesn't assume normally distributed data. I've attached a subset of the sample file Probe.jmp with a script saved to the data table that does this.

[#3]

Thanks very much Jed!

[#4]

Hi @learning_JSL,

The platform mentioned by @Jed_Campbell is the right place to start.
However, due to the high number of groups (more than 50) involved in this multiple comparison, I would highly recommend to use the "Steel-Dwass" test (instead of Wilcoxon) if you are interested in all comparisons, or "Steel with Control" if you want to compare each group to a Control group (provided these tests are available in JMP 12).

These tests are available in the same menu, but protect against overall error rate :

• You have specified an error rate (alpha = 0,05 by default) for a two groups comparison, so your confidence level in this 2 groups comparison test will be 1-alpha = 0,95 (95% confidence).
• The "problem" here is that you have 54 possible groups, creating 1431 comparisons involving each possible pair of groups. You overall confidence level for all comparisons done will be (without correction or adjustment) :

  (1 - alpha) ^ (Total number of comparisons)

And with 1431 possible comparisons, your overall confidence level will be close to 0 (1,32.10^-32)...

So the risk of doing type I error (false positive, falsely rejecting the null hypothesis that there is no statistically significant mean difference for pairs of groups, aka detecting falsely a significant difference between means of groups pairs) is quite high if you don't use the right test or adjust your confidence level. Steel-Dwass is a non-parametric test controlling the overall error rate.

If you need more info about these tests, you can look at the help section : Nonparametric Multiple Comparisons Reports (jmp.com)

I hope this answer will help you choose the most appropriate statistical test,

[#5]

Excellent advice and most appreciated Victor.  Thanks for circling back on this!

[#6]

I suggest spending time visualizing the data before investing a lot of time in numerical analyses. Plotting your data will enhance your understanding, suggest patterns, and help identify anomalies. JMP makes such interactive exploration easy.

You might also use unsupervised learning techniques like clustering to discover similar and dissimilar groups.