friedman test

[Amr Ahmed]

Hello,
Can we use Friedman test to compare two dependent variables between two groups?
xpose is required before friedman test, is there a way to perform the test in the ordinary wide arrangement of the data?
Best regards,

[Marcos Almeida]

With regard to the first question: no. That wouldn’t be the Friedman test. Maybe you wish to take a look at the structural equation model (SEM) approach.

With regard to the second question: first, friedman is a user-written program whose author is Rich Goldstein, a very active member of this Forum. That said, there is the user-written skilmack program. I’ve never used it, since it is aimed at Friedman style test with missing data. The good news are: you won’t need to transpose in order to perform the test.

Hopefully that helps.

[Dave Airey]

Be careful about ties when using that user written command: https://www.statalist.org/forums/for...ings-mack-test.

[Rich Goldstein]

I wrote the Friedman test ado file many years ago because a client insisted on it; over the years a number of people have asked if I minded if they took it over and re-wrote to be more friendly and I always said, "fine - do it"; however, no one has so it remains as is; I have never used it since that client and don't expect I ever will so I will not be re-writing it

why do you want to use it? - maybe someone can suggest an alternative

[Joseph Coveney]

Although it's intended for more versatile use in computing Cochran-Mantel-Haenszel statistics, the user-written emh (for "extended Mantel-Haenszel") command from SSC will also do Friedman's test.

It gets the same results as shown in Dave's post for R's implementation of the Friedman test.

.ÿ
.ÿversionÿ16.1

.ÿ
.ÿclearÿ*

.ÿ
.ÿquietlyÿimportÿexcelÿ"GrangerÿData.xlsx",ÿsheet(Sheet1)ÿfirstrow

.ÿlist,ÿnoobsÿseparator(0)

ÿÿ+-----------+
ÿÿ|ÿAÿÿÿBÿÿÿCÿ|
ÿÿ|-----------|
ÿÿ|ÿ8ÿÿÿ8ÿÿÿ7ÿ|
ÿÿ|ÿ7ÿÿÿ6ÿÿÿ6ÿ|
ÿÿ|ÿ6ÿÿÿ8ÿÿÿ6ÿ|
ÿÿ|ÿ8ÿÿÿ9ÿÿÿ7ÿ|
ÿÿ|ÿ5ÿÿÿ8ÿÿÿ5ÿ|
ÿÿ|ÿ9ÿÿÿ7ÿÿÿ7ÿ|
ÿÿ|ÿ7ÿÿÿ7ÿÿÿ7ÿ|
ÿÿ|ÿ8ÿÿÿ7ÿÿÿ7ÿ|
ÿÿ|ÿ8ÿÿÿ6ÿÿÿ8ÿ|
ÿÿ|ÿ7ÿÿÿ6ÿÿÿ6ÿ|
ÿÿ|ÿ7ÿÿÿ8ÿÿÿ6ÿ|
ÿÿ|ÿ9ÿÿÿ9ÿÿÿ6ÿ|
ÿÿ+-----------+

.ÿ
.ÿgenerateÿbyteÿpidÿ=ÿ_n

.ÿ
.ÿrenameÿ(AÿBÿC)ÿsco=

.ÿquietlyÿreshapeÿlongÿsco,ÿi(pid)ÿj(Cdn)ÿstring

.ÿencodeÿCdnÿ,ÿgenerate(cdn)ÿlabel(Conditions)

.ÿ
.ÿemhÿscoÿcdn,ÿanovaÿstrata(pid)ÿtransformation(rank)

ExtendedÿMantel-Haenszelÿ(Cochran-Mantel-Haenszel)ÿStratifiedÿTestÿofÿAssociation

ANOVAÿ(RowÿMeanÿScores)ÿStatistic:
Qÿ(2)ÿ=ÿÿ7.6000,ÿPÿ=ÿ0.0224ÿ
Transformation:ÿRanks


.ÿ
.ÿexit

endÿofÿdo-file


.

[Amr Ahmed]

Thanks for these comprehensive replies
I would like to compare heart failure class (4 grades) before and after two treatments. Do you think EMH is suitable for this comparison?
I used Silkmack before to compare the change in Dyspnea grade (4 grades) before and after heart transplantation cause I couldn't perform Friedman in Stata, but I got questions from the reviewers why I used this test, as they believe it can be used only if there are missing data. How the results of Smack test in comparison to Friedman test?
Thanks again

[Joseph Coveney]

I would like to compare heart failure class (4 grades) before and after two treatments. Do you think EMH is suitable for this comparison?

If you're looking for assessing a treatment × time interaction term, then no.

[Amr Ahmed]

There is no time variable, we assessing the grade pre and post surgery in two groups

[Joseph Coveney]

There is no time variable, we assessing the grade pre and post surgery in two groups

Pre and post represent the time variable. You're looking for something akin to

manova pre_nyha post_nyha = i.trt
manovatest , test(-1 1 0) ytransform(1 -1)

which tests the treatment × time interaction, just the same as reshaping the dataset long forming an explicit time variable and fitting a regression model with the treatment-by-time interaction term made explicit. See the illustration below to see what I mean about their equivalence and about the presence of a time factor, implicit in the pre/post (a.k.a., mulitvariate, wide) arrangement such as you think of your dataset, and explicit in the long form.

.ÿ
.ÿversionÿ16.1

.ÿ
.ÿclearÿ*

.ÿ
.ÿsetÿseedÿ`=strreverse("1551702")'

.ÿ
.ÿquietlyÿdrawnormÿout0ÿout1,ÿdoubleÿcorr(1ÿ0.5ÿ\ÿ0.5ÿ1)ÿn(24)

.ÿgenerateÿbyteÿpidÿ=ÿ_n

.ÿgenerateÿbyteÿtrtÿ=ÿmod(_n,ÿ2)

.ÿ
.ÿ*
.ÿ*ÿBeginÿhere
.ÿ*
.ÿquietlyÿmanovaÿout0ÿout1ÿ=ÿi.trt

.ÿtempnameÿCÿT

.ÿmatrixÿinputÿ`C'ÿ=ÿ(-1ÿÿ1ÿ0)

.ÿmatrixÿinputÿ`T'ÿ=ÿ(ÿ1ÿ-1)

.ÿmanovatestÿ,ÿtest(`C')ÿytransform(`T')

ÿTransformationÿofÿtheÿdependentÿvariables
ÿ(1)ÿÿÿÿout0ÿ-ÿout1

ÿTestÿconstraint
ÿ(1)ÿÿÿÿ-ÿ0.trtÿ+ÿ1.trtÿ=ÿ0

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿWÿ=ÿWilks'ÿlambdaÿÿÿÿÿÿLÿ=ÿLawley-Hotellingÿtrace
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿPÿ=ÿPillai'sÿtraceÿÿÿÿÿRÿ=ÿRoy'sÿlargestÿroot

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿSourceÿ|ÿStatisticÿÿÿÿÿÿÿÿdfÿÿÿÿF(df1,ÿÿÿÿÿdf2)ÿ=ÿÿÿFÿÿÿProb>F
ÿÿÿÿÿÿÿÿÿÿ-----------+-------------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿmanovatestÿ|Wÿÿÿ0.9331ÿÿÿÿÿÿÿÿÿ1ÿÿÿÿÿÿ1.0ÿÿÿÿÿ22.0ÿÿÿÿÿ1.58ÿ0.2225ÿe
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|Pÿÿÿ0.0669ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ1.0ÿÿÿÿÿ22.0ÿÿÿÿÿ1.58ÿ0.2225ÿe
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|Lÿÿÿ0.0716ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ1.0ÿÿÿÿÿ22.0ÿÿÿÿÿ1.58ÿ0.2225ÿe
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|Rÿÿÿ0.0716ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ1.0ÿÿÿÿÿ22.0ÿÿÿÿÿ1.58ÿ0.2225ÿe
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|-------------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿResidualÿ|ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ22
ÿÿÿÿÿÿÿÿÿÿ-------------------------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿeÿ=ÿexact,ÿaÿ=ÿapproximate,ÿuÿ=ÿupperÿboundÿonÿF

.ÿ
.ÿquietlyÿreshapeÿlongÿout,ÿi(pid)ÿj(tim)

.ÿquietlyÿmixedÿoutÿi.trt##i.timÿ||ÿpid:ÿ,ÿremlÿdfmethod(satterthwaite)ÿ///
>ÿÿÿÿÿÿÿÿÿnoconstantÿresiduals(unstructured,ÿt(tim))ÿnolrtestÿnolog

.ÿcontrastÿR.trt#tim,ÿsmall

Contrastsÿofÿmarginalÿlinearÿpredictions

Marginsÿÿÿÿÿÿ:ÿasbalanced

-----------------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿ|ÿÿÿÿÿÿÿÿÿdfÿÿÿÿÿÿÿÿddfÿÿÿÿÿÿÿÿÿÿÿFÿÿÿÿÿÿÿÿP>F
-------------+---------------------------------------------
outÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿtrt#timÿ|ÿÿÿÿÿÿÿÿÿÿ1ÿÿÿÿÿÿ22.00ÿÿÿÿÿÿÿÿ1.58ÿÿÿÿÿ0.2225
-----------------------------------------------------------

.ÿ
.ÿexit

endÿofÿdo-file


.

So, what I said in #7 holds. You'd probably be better off with your NYHA class data in long format with patient ID to identify the pairs and using something like meologit in a hierarchical regression model that is directly analogous to the mixed example that I illustrate immediately above. Or perhaps using the preoperative NYHA class as a covariate and using something like ologit in a manner analogous to ANCOVA.

[Rich Goldstein]

following up on #4 above, Nick Cox has reminded me about the 2004 version (which I thank Nick for doing); so if you want to use the friedman command directly, ensure that you have the most recent version:

. which friedman
/Users/rich/Library/Application Support/Stata/ado/plus/f/friedman.ado
*! 2.0.1 NJC 27 Nov 2004
*! 1.01 Richard Goldstein  17 July 1991

[Dave Airey]

Looking at that ado code, I see the formulas used were from 2.11 and 2.8 and on page 458 of the Gibbons reference (4th edition, PDF found online, connect at your own risk, <http://erecursos.uacj.mx/bitstream/handle/20.500.11961/2064/Gibbons%2C%202003.pdf?sequence=14&isAllowed=y>) where it goes on to describe further corrections if ties are present. Ties were just something I stumbled on previously.