Post-Hoc Connecting Letters report Displayed on Graphs

[scon]

A great function to add would be the ability to take a post-hoc analysis (Example Tukey's letter report showing treatment 1, 2,3,4 with corresponding letters A, AB, BC, A) and import those letters directly into a bar graph above the columns. 

Post Hoc on Graphs.png

 

[MTOF]

I believe you can already do that. At least if you use Fit Model: under Effect Details run 'LSMeans Tukey HSD' for the parameter in question, under the Tukey hotspot select Save Connecting Letters Table, Run the 'Bar Chart' script embedded in the new table. Example shown below.

MTOF_0-1605729226914.png

[bobmorrane]

@scon yes this wold be very useful to have! I'm surprised it hasnt been included a while ago actually. This is one of the basic features other softwares have and it's one of the most usual way to display statistical analysis.

@MTOF yes and no. The graph you show has several issues compared to what is shown above. First, it sorts the treatments in the alphabetical order of letters associated, which is not necesarily good to do. If your initial graph was sorted as treatment 1,2.... 10, you wouldn't want it to be sorted in a different way once you add the letters.

Another problematic thing is that each bar is separated, even though some may be connected by several letters. A more useful way to plot that would be to have overlapping bars, to clearly show who is different from who. 

Last but not least, the error bars from the initial graph are replaced by a confidence interval representative of the model used but not the initial data (If I'm not wrong). 

[MTOF]

@bobmorrane I agree, the original graph is not the same as the solution in JMP. I just wanted to point out that this graph exists (it's well hidden...) as I asked JMP to include something like this myself many years ago. Regarding your points: 1) the sorting, I agree, but you can just change the sorting yourself afterwards, or make your own graph from the saved data table. 2) overlapping bars, I think that would be to messy to be useful - at least for our data. 3) error bars, I disagree, the error bars on the graph should be confidence intervals from the model (or even better LSD bars). Confidence intervals based on raw data are not interesting, the individual observed variation may vary a lot by chance - the modeled variation is a better representation of the "estimated" uncertainty of each mean. 

[bobmorrane]

@MTOF  I guess your point on which error bars to use depends on who you ask and what type of data they work on. I work on biological data and I'm really not convinced that generating models on the small datasets I use would be really meaningful. Given the large variability of data and the many factors that can influence it, I'd rather just have the real data points and not estimates. 

Could anyone make a nice add-in for the graph builder that includes the letters on the graph? I'm sure someone skilled in scripting could do this in no time with graphlets. 

Is there a place with a wishlist for scripts and addins?