2x2 AB/BA crossover trial analysis resources/advice

[Dil Singh]

Hi all.

I would like to conduct the analysis of a trial comparing two drugs (A and B) on reducing pain associated with a procedure.

Subjects are randomly exposed to either drug A or B then undergo a procedure and asked to score the pain associated with that procedure 0-10. After a washout period, they are exposed to drug B or A (the one they were not initially exposed to) and then undergo the same procedure, again being asked to score their pain 0-10. Subjects are also asked to score their satisfaction with each drug, again on score 0-10. Subjects are also asked to state their overall preference (drug A or B).

Although I can find resources on crossover trials (such as the textbook by Senn 2002, as well as some online explanations of crossover trials), I am unable to find a decent resource with regard to the analysis of one in Stata.

Using "help pkcross" provides some information, but is somewhat brief.

Can anyone help? Indeed, can pkcross be used for experiments such as this or is it more designed for pharmacokinetic experiments as the name suggests?

Regards

[Joseph Coveney]

can pkcross be used for experiments such as this

Yes. Have you looked at the user's manual entry for -pkcross-? It gives three worked examples for a 2 × 2 crossover study.

On the popup window that displays after -help pkcross-, click on they hyperlink at the upper left corner that is labeled "(View complete PDF manual entry)".

I take it that yours is a "superiority study" and not one of therapeutic equivalence.

[Dil Singh]

Hi.

Thank you for your suggestion, which was good to reshape the data using "pkshape".

However, is there a means of doing a t-test (comparing mean differences by group) or chi-squared test based (based on a binary outcome of patient preference) on these crossover trial data, in keeping with the correct analyses, namely that they are not parallel trial data?

This seems to be possible in other statistical packages such as R or SAS looking online, but am unable to see anything for Stata.

And yes, it is a superiority study comparing two drugs used to reduce pain from a procedure.

Many thanks again in advance.

Regards, Dilraj Thind

[Joseph Coveney]

However, is there a means of doing a t-test (comparing mean differences by group) or chi-squared test based (based on a binary outcome of patient preference) on these crossover trial data, in keeping with the correct analyses, namely that they are not parallel trial data?

This seems to be possible in other statistical packages such as R or SAS looking online, but am unable to see anything for Stata.

It's right there in the ANOVA table, "Treatment effect", and it's under the header "Intrasubjects", which gives you your clue that these aren't parallel trial data:

.ÿquietlyÿwebuseÿchowliu

.ÿpkshapeÿidÿseqÿperiod1ÿperiod2,ÿorder(RTÿTR)

.ÿpkcrossÿoutcome
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿsequenceÿvariableÿ=ÿsequence
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿperiodÿvariableÿ=ÿperiod
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿtreatmentÿvariableÿ=ÿtreat
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿcarryoverÿvariableÿ=ÿcarry
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿidÿvariableÿ=ÿid

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿAnalysisÿofÿvarianceÿ(ANOVA)ÿforÿaÿ2x2ÿcrossoverÿstudy
ÿÿÿÿÿSourceÿofÿVariationÿÿ|ÿPartialÿSSÿÿÿdfÿÿÿÿÿÿÿÿMSÿÿÿÿÿÿÿÿÿÿFÿÿÿÿÿProbÿ>ÿFÿ
ÿÿÿÿ----------------------+--------------------------------------------------
ÿÿÿÿÿIntersubjectsÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿÿÿÿSequenceÿeffectÿ|ÿÿÿÿÿ276.00ÿÿÿÿ1ÿÿÿÿÿÿ276.00ÿÿÿÿÿÿ0.37ÿÿÿÿÿÿ0.5468
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿResidualsÿ|ÿÿÿ16211.49ÿÿÿ22ÿÿÿÿÿÿ736.89ÿÿÿÿÿÿ4.41ÿÿÿÿÿÿ0.0005
ÿÿÿÿ----------------------+--------------------------------------------------
ÿÿÿÿÿIntrasubjectsÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿÿÿTreatmentÿeffectÿ|ÿÿÿÿÿÿ62.79ÿÿÿÿ1ÿÿÿÿÿÿÿ62.79ÿÿÿÿÿÿ0.38ÿÿÿÿÿÿ0.5463
ÿÿÿÿÿÿÿÿÿÿÿÿPeriodÿeffectÿ|ÿÿÿÿÿÿ35.97ÿÿÿÿ1ÿÿÿÿÿÿÿ35.97ÿÿÿÿÿÿ0.22ÿÿÿÿÿÿ0.6474
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿResidualsÿ|ÿÿÿÿ3679.43ÿÿÿ22ÿÿÿÿÿÿ167.25
ÿÿÿÿ----------------------+--------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿTotalÿ|ÿÿÿ20265.68ÿÿÿ47

ÿÿÿÿOmnibusÿmeasureÿofÿseparabilityÿofÿtreatmentÿandÿcarryoverÿ=ÿÿÿ29.2893%

.

I suppose that if you wanted to see it as an actual t statistic, then you could:

.ÿmixedÿoutcomeÿi.sequenceÿi.treatÿi.periodÿ||ÿid:ÿ,ÿremlÿdfmethod(satterthwaite)ÿnolrtestÿnolog

Mixed-effectsÿREMLÿregressionÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿobsÿÿÿÿÿ=ÿÿÿÿÿÿÿÿÿ48
Groupÿvariable:ÿidÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿNumberÿofÿgroupsÿÿ=ÿÿÿÿÿÿÿÿÿ24

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿObsÿperÿgroup:
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿminÿ=ÿÿÿÿÿÿÿÿÿÿ2
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿavgÿ=ÿÿÿÿÿÿÿÿ2.0
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿmaxÿ=ÿÿÿÿÿÿÿÿÿÿ2
DFÿmethod:ÿSatterthwaiteÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿDF:ÿÿÿÿÿÿÿÿÿÿÿminÿ=ÿÿÿÿÿÿ22.00
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿavgÿ=ÿÿÿÿÿÿ24.37
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿmaxÿ=ÿÿÿÿÿÿ31.50

ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿF(3,ÿÿÿÿ22.00)ÿÿÿÿ=ÿÿÿÿÿÿÿ0.32
Logÿrestricted-likelihoodÿ=ÿ-197.03719ÿÿÿÿÿÿÿÿÿÿProbÿ>ÿFÿÿÿÿÿÿÿÿÿÿ=ÿÿÿÿÿ0.8096

------------------------------------------------------------------------------
ÿÿÿÿÿoutcomeÿ|ÿÿÿÿÿÿCoef.ÿÿÿStd.ÿErr.ÿÿÿÿÿÿtÿÿÿÿP>|t|ÿÿÿÿÿ[95%ÿConf.ÿInterval]
-------------+----------------------------------------------------------------
ÿÿÿÿsequenceÿ|
ÿÿÿÿÿÿÿÿÿTRÿÿ|ÿÿ-4.795834ÿÿÿ7.836273ÿÿÿÿ-0.61ÿÿÿ0.547ÿÿÿÿ-21.04727ÿÿÿÿÿ11.4556
ÿÿÿÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿtreatÿ|
ÿÿÿÿÿÿÿÿÿÿTÿÿ|ÿÿ-2.287501ÿÿÿÿ3.73326ÿÿÿÿ-0.61ÿÿÿ0.546ÿÿÿÿ-10.02981ÿÿÿÿ5.454806
ÿÿÿÿ2.periodÿ|ÿÿÿ-1.73125ÿÿÿÿ3.73326ÿÿÿÿ-0.46ÿÿÿ0.647ÿÿÿÿ-9.473557ÿÿÿÿ6.011057
ÿÿÿÿÿÿÿ_consÿ|ÿÿÿ85.82292ÿÿÿ6.137768ÿÿÿÿ13.98ÿÿÿ0.000ÿÿÿÿÿ73.31286ÿÿÿÿ98.33297
------------------------------------------------------------------------------

------------------------------------------------------------------------------
ÿÿRandom-effectsÿParametersÿÿ|ÿÿÿEstimateÿÿÿStd.ÿErr.ÿÿÿÿÿ[95%ÿConf.ÿInterval]
-----------------------------+------------------------------------------------
id:ÿIdentityÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿ|
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿvar(_cons)ÿ|ÿÿÿ284.8196ÿÿÿ113.9151ÿÿÿÿÿÿ130.0551ÿÿÿÿ623.7527
-----------------------------+------------------------------------------------
ÿÿÿÿÿÿÿÿÿÿÿÿÿÿÿvar(Residual)ÿ|ÿÿÿ167.2468ÿÿÿ50.42679ÿÿÿÿÿÿ92.62129ÿÿÿÿ301.9984
------------------------------------------------------------------------------

.

but you can see that it's the same.

I think that you'll be wasting your time with the preference outcome (recency phenomenon will give you a strong confounding period effect), but you would fit a logistic regression model using -xtlogit , fe- or -xtlogit , re- with the same predictors and examine the regression coefficient for treatment.

If you see something different in R or SAS that you cannot match up with the above, then point it out in another post, and I or someone else on the list might be able to help you with the cross-walk between the brands of software.