I have been trying to determine the proper statistical test for comparing binned proportional data between groups. My data set is individual subjects (4-5 per group) with 76-165 cells per individual. Each cell has a variable X that can range from 0-1. I binned the data for each individual in 0.2 sized bins, converted each bin to a proportion of the total (for each individual), and averaged by group (WT/KO). (link to example data plotted in MATLAB, https://imgur.com/WyMGMyS).

It seems like a common method is to run a Wilcoxon test on each bin independently. My question would be, is this an appropriate method given that, as proportion data, the values in bins are not independent of each other? Alternatively, I have tried an ANOVA method in JMP (Analyze->Fit Model->Standard Least Squares->Y=variable, constructor model group*bin interaction), but this method seems very insensitive to effects in single bins if "bin" is a continuous variable, but listing bin as ordinal seems inappropriate (example JMP data entry for 1 individual, https://imgur.com/MANc58S). Has anyone encountered a similar issue? Also, if the Wilcoxon test is appropriate is there a way to run it in JMP on each bin independently without having to exclude all the other bins in the chart and run it 5 times per variable, X? Or would that be simpler to code in MATLAB using the ranksum function?

Thank you, and please let me know if this question is not appropriate/appropriate format.

*edit: link to data, https://github.com/kylemxm/ShepherdGit/blob/master/ODI%20example%20Data.mat

  • $\begingroup$ No need to bin at all. Push the original data through the command/function/routine in question. $\endgroup$ – Nick Cox Dec 10 '18 at 20:15
  • $\begingroup$ This appears to be a duplicate post from stackoverflow.com/q/53712090/8239061 $\endgroup$ – SecretAgentMan Dec 10 '18 at 20:22
  • $\begingroup$ Yes it is, I originally posted there but I was directed here as a more appropriate forum for this question. Deleting the Stackoverflow post now to avoid confusion. $\endgroup$ – Kyle Jenks Dec 10 '18 at 20:47
  • $\begingroup$ Its better to analyze the data as a single regression rather than binned observations. Can you link us to the data somewhere? $\endgroup$ – Demetri Pananos Dec 10 '18 at 21:03
  • $\begingroup$ Edited post to have link to data used to generate graph. First column of cell is WT, 2nd column is KO. Each matrix is an individual. I originally did the analysis as a Kolmogorov–Smirnov test, but that required combining all of the cells for all of the individuals. My thought for that is that it artificially increases my N to number of cells rather than number of individuals. Is there a way to do regression of continuous data, averaging across individuals in a group, without averaging into discrete bins? $\endgroup$ – Kyle Jenks Dec 10 '18 at 22:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.