Which post hoc test can I use after a friedman test?

[#1]

Hi there.

I'm using the Friedman test on my data and now I would like to know which of the non-parametric comparisons can I use as a post hoc test. The test I wanted to use was Bonferroni, for making the adjustments and error corrections. What should I do?

[#2]

I recently updated from version 15 to 17 and found that the non parametric post hoc analysis option was now greyed out when a block was included in fit Y by X using Friedman as a non parametric test. It should be possible to use a LSD test to show individual differences between the rank sums of each pair, this was possible in version 15. Why has this option be removed in version 17?

[#3]

Hello @CompositeRam426,

It would help having a sample of your anonymized data to be sure about your goal and check which test could be the most appropriate.

If you're doing multiple comparisons tests, you may choose Tukey HSD (in parametric test) or Steel-Dwass All Pairs (in non-parametric test, multiple comparisons) that both protect overall error test.

I hope I understood your question correctly and that this first answer may help you, 

[#4]

Hi Enclosed is a data file. Assessor is a block and Treatments a factor with four levels. Data is % of a VAS line (0-150 mm). The datafile includes a script for a Friedman analyses using Fit Y by X, non-parametric tests. But theres is no available post-hoc analyse for the individual comparisons (greyed out) . 

Typically an LSD test on the rank sum can be use as Post Hoc analyses as show in the enclosed formula.

Kind regards

Flemming

[#5]

Hi @flvs,

So if I understand your problem correctly, you would like to compare treatments based on the value of the "Data" column. You have 3 alternative treatments and one reference (standard) treatment.

I'm afraid in your configuration, the Friedman test will be the only option you have : "If you specify a Block variable in the launch window and there are equal counts for each combination of Block and X variable level, the Friedman Rank Test is the only Nonparametric option available. If you specify a Block variable in the launch window and there are unequal counts, none of the Nonparametric options are available." from JMP Help : The Oneway Platform Options (jmp.com)

• A) The other non parametric test I was mentioning (Steel-Dwass All Pairs, or Steel with Control if you are only interested in comparisons between alternatives and standard treatment (and not between alternatives)) is not available when there is a variable Block. You would have to remove your block "Assessor" to use it/them.
• B) One other option to have more information could be to use the platform Fit Model, to model the Data column depending on the Treatment with Assessor as a random effect. You can then visualize with the profiler which treatment is similar to reference treatment (alternative 3) and which ones are better compared to standard treatment (1 and 2).
• C) I may have missed something reading your file too quickly, but it seems that your data do respect the three assumptions behind parametric tests (data distribution is quasi-normal, homogeneity of variances and independance of observations), so perhaps you could still use Tukey HSD (for doing all comparisons while protecting overall error test) or Dunnett's test with Control (for doing comparison between alternatives and the reference). These parametric tests do support block variable.

I joined the datatable you provided with the added scripts for the three options.

Hope these suggestions will help you,

[#6]

Hi Victor

Thanks for your suggestions and you are correct that data does not violate the assumption of normal distribution of the error.

I am teaching undergraduates various statistical methods for sensory analyses using rank data, where Friedman is the tool used in for example international standards for sensory analyses (ISO 8587), why I am insist on using this method. As the ranks within each assessor is not independent it would not be correct to use Wilcoxon Rank test.  

As mentioned there previously was no parametric individual comparison available in JMP, but it was removed in the new version 17, which is actually my question.

PS. Using the fit model with random effect is by the way an interesting way of analysing such data (given normality).

Kind regards

Flemming

[#7]

Ok, thank you for the context about this analysis @flvs !

I'm sorry not to be able to help you further. Maybe @Mark_Bailey will have an answer about this feature change in JMP 17 ?

Kind regards,
 

[#8]

The Friedman test provided by the Oneway platform requires a complete block design. Use this example to test your version of JMP:

dt = Open( "$SAMPLE_DATA/Snapdragon.jmp" );
obj = dt << Oneway(
       Y( :Y ),
       X( :Soil ),
       Block( :Block )
);
obj << Friedman Rank Test( 1 );

The post hoc tests following the Friedman test are planned for the next version of JMP.

[#9]

It's unfortunate that there's no multiple comparison available for the Friedman test. Are there any suggestions on how to proceed, if one want to perfom them? Could a similar test be performed in the Fit Model platform, with multiple comparison? I.e. a non-parametric alternative to RM-ANOVA