1

The longitudinal dataset has n=275 and 8 measurement points. There are 3 groups (3 different drugs) with roughly n=80 each.

The complications are:

(1) Substantial dropout: only n=136 have all observations. Dropout is not missing at random, of course: severity predicts dropout, so do a few other variables in the dataset.

(2) I have 17 dependent variables (DVs) of interest. Dimension reduction (e.g., PCA) is not possible because the covariance of variables changes dramatically over time (there are 4-5 dimensions/components at baseline and 1-2 dimensions/components at study exit)

(3) The 17 DVs are all ordered-categorical, and have different ranges (from 0 to 2 to 0 to 4).

My research question is whether the 3 drugs (groups) have differential impact on the trajectory of the 17 DVs. The best possibility seems to be to fit 17 longitudinal mixed models or 17 generalized estimation equations to the data, and correct for multiple testing somehow; there is probably too little power to fit multivariate models with even 2 DVs (let alone 17).

The package lme4 using the lmer() function cannot handle ordinal data, and I cannot model skewed ordinal variables with 3 categories as gaussian.

The package ordinal using the clmm2() function deletes missing values listwise, deleting too many observations.

So I seem to be trapped between a rock and a hard place and don't know how to fit the model even in the univariate case.

Torvon
  • 823
  • 4
  • 10
  • 21

2 Answers2

2

Sounds like a mess. I can suggest that for item #3, you look at the new R package brms, which uses Stan (Bayesian tool) under the hood. (If that doesn't seem quite right, also look at rstanarm, which is also using Stan under the hood, but takes a different approach.) These will not run nearly as fast as lmer since they're doing MCMC sampling, and are working through a second underlying tool. (brms in particular needs a compiler and compiles your model before executing it.) But it can handle ordinal independent variables, and has several options for this.

It sounds to me that 17 DVs with only 136 complete observations won't work and you might want to consider combining your DVs down to a single score if that's possible. (Other, more knowledgeable folks can chime in on this: whether it's a good idea or not and how to do it if it is a good idea. I have no experience in doing it.) But a Bayesian analysis should give you as much information as possible to decide how uncertain your results are.

Wayne
  • 19,981
  • 4
  • 50
  • 99
  • Thanks. 5 decades of analyzing the sum of these 17 DVs have shown that the drugs work similar on the sum, allowing the possibility that they exhibit differences on the item level (since the items are not even remotely unidimensional). So I need to model them individually and not as a sum. – Torvon Mar 03 '16 at 15:39
0

Something from the nlme package ought to be able to handle the ordinal DVs, though you may have to come up with a custom linking function. Here's an example from another post. Non-linear mixed effects regression in R What's more worrisome is the missing data. Not sure what to do about that since its NMAR. Maybe sensitivity analysis or a pattern-mixture mixed effects model? http://www.lexjansen.com/wuss/2012/32.pdf

Community
  • 1
SeanGL
  • 1