How to set up a multivariate (=multi-response) multilevel model?

[Alex Gamma]

Hello Statasticians,

I have repeated measures of an outcome (response to a drug) that I want to predict by a time-invariant predictor (some EEG parameter). The repeated measures are indexed by the variable “repetition” and have been taken at roughly monthly intervals.

  * Example generated by -dataex-. To install: ssc install dataex
  clear
  input byte(id repetition medresponse) float eegpredictor
  1 1 . 13.438
  1 2 2 13.438
  1 3 1 13.438
  1 4 1 13.438
  1 5 1 13.438
  2 1 3 16.175
  2 2 . 16.175
  2 3 . 16.175
  2 4 3 16.175
  2 5 0 16.175
  2 6 3 16.175
  3 1 .   .612
  3 2 .   .612
  3 3 1   .612
  3 4 1   .612
  3 5 2   .612
  3 6 1   .612
  end

I want to include information about the variation in the outcome, that’s why I thought of using a multilevel model.

Questions:

A) What model would you recommend for my purposes?
B) Assuming multilevel is a good choice, how do I set up the data and what would be the command syntax to use?

Background: I’ve found some advice by Tabachnik & Fidell (section 15.5.5, p. 821) saying that I need to dummy-code the different repeats of the outcome and then use this as my level-1 predictor, but I’m not quite sure how to set up the data for it.

Thanks,
Alex

Reference: Tabachnik BG, Fidell LS. Using multivariate statistics. Pearson Education, Inc. 5th ed. Boston; 2007.

[Alex Gamma]

I think I was confused when I posted this. I don't have different outcome types, so I can just go with a regular (univariate) multilevel model. Deleting the post seems not possible, though...

[Joseph Coveney]

You might want to look into meologit. Keep in mind that the estimation command is large-sample (it expects many IDs). You've also got a lot of missing data.

If your scientific interest lies in assessing predictor strength (of the value of your electroencephalographic whatever for the monthly median drug response category), then, as a first pass, you might just want to take an average and standard deviation of your monthly drug-response medians by animal (or patient), and fit a variance-weighted least squares linear regression to see whether it's worthwhile to pursue anything more involved.

version 14.1

clear *
set more off

input byte(aid month response) float eeg
  1 1 . 13.438
  1 2 2 13.438
  1 3 1 13.438
  1 4 1 13.438
  1 5 1 13.438
  2 1 3 16.175
  2 2 . 16.175
  2 3 . 16.175
  2 4 3 16.175
  2 5 0 16.175
  2 6 3 16.175
  3 1 .   .612
  3 2 .   .612
  3 3 1   .612
  3 4 1   .612
  3 5 2   .612
  3 6 1   .612
end

collapse (mean) mean = response (sd) sd = response, by(eeg) // Assumes no two animals share an EEG thingy value to three decimal places

generate float iv = 1 / sd / sd
graph twoway scatter mean eeg [aweight=iv], mcolor(black) msymbol(circle_hollow) ///
    ytitle(Mean Monthly Median Drug Response Category Mouthful) ylabel(0(1)3, angle(horizontal) nogrid) ///
    xtitle(Pretreatment EEG Thingy)
// But see http://www.stata.com/statalist/archive/2008-07/msg01143.html

vwls mean c.eeg, sd(sd)

exit