nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Does JMP automatically calculate ranks in Steel-Dwass

PeterSchmedes

The study: I am to analyze a non-parametric dataset where two types of twine were cultered in the same tank to test preference in settlement of algal spores. Three levels of agitation was tested. Using 3 replicates of the twine (2) and agitation (3) results in 6 combinations or datasets. They were cultered in 9 experimental units and 2*9 pieces of twine was used.

Can I use the Steel-Dwass to compare the median of the 6 datasets without doing a Friedman test first? Or should I save ranks using the Fit X by Y and then do a 2-factor ANOVA on ranks by blocking with one factor (equal to Friedman test) to get the effect of main factors? And then use Steel-Dwass in own-calculated ranks? Or does JMP automatically calculate and use ranked data when Steel-Dwass is performed?

Accepted answers

All comments

Mark_Bailey

First of all, what is your response? What is the result of each replicate test? Is it a measurement? A count?

I understand that you have 2 levels of twine and 3 levels of agitation for a full factorial design of 6 treatments. These were replicated twice 18 runs. What is the definition of the 9 experimental units?

The Steel-Dwass non-parametric test is for pair-wise comparisons, so it appears in the JMP Oneway platform. You could create a new data column, Treatment, that crossed the factor columns, Twine and Agitation, and perform this simple test but you might be missing the benefit of fitting a model to the data. You could use Analyze > Fit Model to define a linear predictor with Twine, Agitation, and Twine*Agitation. Your response, of course, is in the Y role. The method ('personality') that is appropriate depends on your answer to my first question above. I suggest considering this approach in addition to the simple group comparison.

Mark_Bailey

I should add that the choice of a non-parametric (distribution-free) method is a good opportunity to maintain the power of the test when the assumptions of the parametric methods are not met. On the other hand, the power of the parametric method is almost always greater than the non-parametric alternatives when the assumptions are met.

What led you to choose the non-parametric tests in this case?

PeterSchmedes

My two response variables are the number of settled spores and settled algae seedlings on the twines in each replicate. Actually, I started the experiment and cultured it for 7 weeks, and I counted both spores and seedlings every week (for 7 weeks) - thus, the experiement is done once and repeated measurements was sampled. As a start I want to assess any difference at the final date. To remind you, the two types of twine was cultured in pairs in the same tank (experimental unit) to which I applied three levels of agitation, in triplicates (= 9 tanks or exp units with totally 18 twines, nine pieces of each).

I think it is appropriate to choose a non-parametric analysis due to the fact that the twines were not independent from each other. I have read that by computing a 2factor ANOVA on the ranks by use of Fit X by Y and block by one factor, is effectively the same as a Friedman test (which JMP 13 does not have) and this test correspond to a non-parametric 2-factor ANOVA. The Steel-Dwass can only be used as a post hoc test?

Is it better to use the "Wilcoxon Test" in JMP 13?.

Mark_Bailey

Well, your experimental units define a 'split-plot' experiment. The Twine is the easy to change factor and the Agitation is the hard to change factor.

The correlation is not handled any better by a non-parametric test. The correlation is handled by a random effect for when Agitation is reset between runs.

The fact that you count spores means that you might be able to use ordinarly least squares regression with normal errors or you could use a genearlized linear model with Poisson errors.

PeterSchmedes

To do ordinarly least squares regression with normal errors or genearlized linear model with Poisson errors, would I have to arrange the datasets in 5 columns ?

Column1 (date), Column2 (FactorA, Twine with 2 levels), Column3 (FactorB; Agitation with 3 levels), Column4 (Response= spores), Column5 (Response2 =Seedlings)

For Split plot design I would construct the model effects by adding Twine (2levels) as main factor and add Agitation (3 levels) nested within Twine. Should I also add date (also as nominal variable) and cross it with Twine (I collected triplicate data on spore and seedling counts on 7 dates). Then I have read to correct the error term in the Model to Random Effect (in Attributes) and use EMS in Method.

Would it be a shame to treat it as a crossed design on only final counts?

As you mentioned, I could make a new factor column by combining the two main factors ("treatment") and assess the difference in final mean count.