sequential randomization

[benkeser]

The trial is a sequentially randomized trial with possible re-randomization based on previous response. At baseline, participants are randomized 1:1:1 into three arms, texting intervention, mobile app intervention, and standard of care. At 3 month increments participants in the text and app arms who do not "improve" their risk behaviors are eligible for re-randomization into a third intervention, e-coaching.

Because it's a randomized trial, at each time I know the probability of receiving each treatment given the past. I don't necessarily have to input the known probabilities, but I can't figure out how to properly stratify the fitPropensity command and can't digest what the error messages are actually telling me.

Attached is an example data table.

• ID = partic. id
• covariate = a binary baseline covariate
• Y = the outcome at month 12
• txt = on the texting intervention
• app = on the app intervention
• t = time (0 = baseline, 3, 6, 9, 12 months)
• monitor = whether they are seen at t + 3
• risk = risk behavior score, if risk at t+3 is bigger than or equal to risk at t (and risk at t != 0) and -
• txt == 1 | app == 1, then they are eligible at t+3 for rerandomizaiton to e-coaching. Y is the sum of risk at 3, 6, 9, and 12.
• trt = multinomial treatment indicator 1 = txt, 2 = app
• Y_scale = Y scaled by maximum value of Y
• trt.setXX = treatment levels for rules I'm interested in (no int, all txt int, all app int, txt + e-coaching when risk doesn't improve, app + e-coaching when risk doesn't improve)
• eligible = whether or not someone is eligible for re-randomization at t
• ecoach = on the e-coach intervention
• coach.tminus1 = on e-coach intervention at previous time point
• cens = censoring = 0 for everyone, since it seemed like providing monitoring status made stremr happier than censoring + missing values

Here's what I've tried:

gform_MONITOR <- "monitor ~ risk + covariate"
# would perhaps prefer this to be intercept only, but ~1 syntax seems to piss everything off
gform_TRT <- "trt ~ risk"
stratify_TRT <- list(trt = c(
  "t == 0", # baseline randomization prob.
  "t == 3 & eligible == 1", # at time 3 prob. of getting e-coaching amongst those eligible (should be 1/2)
  "t == 6 & eligible == 1", # at time 6 prob. of getting e-coaching amongst those eligible (should be 1/2)
  "t == 9 & eligible == 1", # at time 9, etc... 
  "t == 3 & txt == 1 & eligible == 0", # at time 3, txt folks should stay on txt with prob. 1 if not eligible for e-coach
  "t == 6 & txt == 1 & eligible == 0", # at time 6, txt folks should stay on txt with prob. 1 if not eligible for e-coach
  "t == 9 & txt == 1 & eligible == 0", # etc...
  "t == 3 & app == 1 & eligible == 0", # ditto app people who are not eligible
  "t == 6 & app == 1 & eligible == 0", # etc 
  "t == 9 & app == 1 & eligible == 0", # etc
  "t == 3 & app == 0 & txt == 0", # ditto standard of care people, i.e. trt == 0
  "t == 6 & app == 0 & txt == 0", # etc
  "t == 9 & app == 0 & txt == 0", # etc
  "t == 6 & ecoach.tminus1 == 1",  # once you get to e-coaching, you stay on e-coaching with prob. 1
  "t == 9 & ecoach.tminus1 == 1"))

I'm just wondering how to properly specify these regressions and whether there is a way to make them intercept only or not.

testDT.RData.zip