Cross-over RCT - multiple regression?

[Maggie ZX]

First time posting a question, apologizes if I am not doing this correctly.

I have a placebo-controlled, 4-way crossover trial: (1) placebo, (2) drug A, (3) drug B, and (4) combo of drug A and B. Each patient has a baseline pain score (continuous, 0 to 10) and 4 separate "post-treatment" pain scores, one for each of the 4 drug regiments (placebo/drugA/drugB/combo)

I also have adverse effects reported for each of the drug regimens (eg. fatigue yes/no).

So, the primary outcome (pain) and secondary outcome (fatigue)

I would be interested to see if certain baseline characteristics (I can use them as continuous variables or turn them into categorical variables by binning) are associated with 1) pain response and 2) adverse events, but I am uncertain how to handle the cross-over nature of the data.

I would need, would appreciate any insights. (I am reading about the mixed effect model but I can't seem to understand how this would apply to my situation).


I have thought about doing

Model (reduction in pain drug A relative to placebo) ~ (baseline 1 + baseline 2 + baseline 3)
but it seems like I would have to repeat this 2 more times with (drug B relative to placebo) and (combo relative to placebo) and i'm not sure if that is correct.


Thanks,

Maggie

[Joseph Coveney]

I would be interested to see if certain baseline characteristics (I can use them as continuous variables . . .) are associated with 1) pain response and 2) adverse events, but I am uncertain how to handle the cross-over nature of the data.

If you're interested in seeing—that is, visualizing—the association, then sum the pain scores across drug treatments, subtract the placebo score (or baseline scores, if there is a separate baseline measurement for each treatment condition), and plot the net pain sumscore against the baseline characteristic value. If the relationship is nonlinear, you should be able to pick that out, too.

Same with the fatigue state: sum the fatigue values (0 = No, 1 = Yes) across treatment conditions (I'll assume that this will be zero for baseline) and plot the sumscore against baseline characteristic value. If needed in order to aid visualization—for example, if you get a lot of close clusters of fatigue sumscores in a narrow range of baseline characteristic values—then jitter the symbols (it's an option of the graphing command) on the graph.

help egen
help graph twoway scatter

[Maggie ZX]

Thank you for your comment, Joseph! I should have been more clear in that I am having difficulties with choosing an appropriate statistical analysis for this 4 way cross-over. I have been reading about ANCOVA vs. -xtmixed- but please forgive me I'm still confused about which one to choose and how the code might look like.

My research question: Do baseline characteristics predict response to treatment for pain?

Before starting the 4-way cross-over RCT, each participant completes questionnaires to get baseline scores (for example baseline anxiety score [continuous], depression score[continuous]). [These would be my baseline predictors]

Then, each participant completes 4 of the following treatments: placebo, drugA, drugB, combo of drugsA+B with appropriate washout in between. After each of the 4 treatments, they report a pain response (continuous, pain intensity at the beginning of each treatment period MINUS pain near the end of the treatment period)

So that data would look something like this

ID	baseline_anxiety	baseline_depression	treatment	pain
1	15	12	placebo	16
1	15	12	drugA	15
1	15	12	drugB	12
1	15	12	drugAB	10
2	11	9	placebo	13
2	11	9	drugA	10
2	11	9	drugB	9
2	11	9	drugAB	2

so the baseline_anxiety and baseline_depression scores do not change for any individual patient.

Thank you!

[Joseph Coveney]

My research question: Do baseline characteristics predict response to treatment for pain?

From your snippet of a data listing, your crossover study design doesn't have enough of a handle on confounding to be able to answer your research question.

But you could create posttreatment sumscores for pain as I described above and graphically examine their relationship to pretreatment scores for anxiety and depression. Something like the following if you have no missing data.

bysort ID: generate double sco = sum(pain)
quietly by ID: keep if _n == _N

set more on
foreach var of varlist baseline_* {
    graph twoway scatter sco `var', mcolor(black) msize(small) ///
        ytitle(Pain Sumscore) ylabel( , angle(horizontal) nogrid) ///
        xtitle(strproper(subinstr("`var'", "_", " ", 1)) + " Score")
    more
}

If you have missing values, for example, as a result of dropout, then you could use the average pain score.