Wilcoxon matched-pair sign-rank test -signrank- with a grouping variable ( by(groupvar) )

[Rasmus Green]

Hi!

I'm conducting a study in which I have 3 controls per case, matched on age, cancer stage and cancer grade. I have a few categorical and a few continuous variables i want to compare between the groups. Most continuous variables are not normally distributed. Thus, I've understood that I have non-independent data and should go for non-parametric descriptive tests, thus the Wilcoxon sign-rank test. You can see an example below, the indicator variable being cohort and the variable I want to compare is weight_merge. My patients are overweight, so this is not normally distributed.

* Example generated by -dataex-. To install: ssc install dataex
clear
input float(cohort weight_merge)
1  54
0  54
1 120
1  50
1 101
1 158
0  74
0  53
1  77
1  80
1  58
1  98
1 129
0 106
0 138
1  70
1  76
1  72
1  72
0  59
0  81
1  64
1  90
0  63
1  62
1  49
1  64
1  74
0  79
1  53
1  58
1  73
1  66
1 110
0 103
0  70
1  75
1  80
1  68
1  96
1 116
0  95
1 113
1  91
1  89
1  72
1  80
1 100
1  86
1  86
1  50
1  86
1  87
1  95
1  77
1 109
1  60
1  63
1  92
0  69
1  67
0  72
1  62
1  75
0  47
1  92
0  85
0 104
1  75
1  80
0  75
1  58
1  61
1  85
1  91
0  96
1 130
1 116
1  73
0  62
1 150
1  53
0  56
0  73
1  69
1 114
1  50
1  81
0  60
1  77
1 119
1  68
1  62
1 150
1  95
1  72
1  98
1  98
1  66
1  77
end
label values cohort cohort
label def cohort 0 "Study", modify
label def cohort 1 "Control", modify

Now, with other "comparison of means" test, Stata allows the option by(groupvar), but not for -signrank-. In either case, even with -ttest-, specifying by(groupvar) assumes unpaired data, according to the help section.
It seems that Stata needs paired data to be ordered as measurement_pre and measurement_post in order to do paired comparisons (ttest measurement_pre == measurement_post) for all these types of test, but in my case that is unattainable since I have 3 controls per case and I cannot tie a certain control to a certain case and I have about 5 variables that I want to compare.

Is there a user-written command for this? Or, how big of an error could I introduce by instead just running for unpaired comparisons with ranksum weight_merge, by(cohort) even though my data is dependent due to the matching? Before matching they were independent, controls were taken from a large, completely separate set of patients.

This question was also posted on ResearchGate, and I will make sure to post any good answer there as well.

https://www.researchgate.net/post/Ho...d23584d80a347d

Best regards!

//Rasmus W Green, PhD student, Karolinska Institute, Dept. of Women's and Children's Health, Stockholm, Sweden.

[Joseph Coveney]

I don't see the 3:1 matching. Where is the variable that matches the "3 controls per case . . . on age, cancer stage and cancer grade"?

With such a matching variable, you could use the user-written vanelteren or, better, the user-written emh, both available from SSC.

Otherwise, what you would seem to have is an independent-group (observational) study with eligibility criteria for age, and cancer stage and grade.

[Rasmus Green]

Hi and thanks for your response,

the indicator variable is cohort, but no, you don't see the 3:1 ratio because this is just the first 100 listings (I've got more than 500 observations).

I'll look into those commands, thank you! If they work I'll post an update, because that would be exactly what I'm looking for.