Discontinuous growth curve analysis - NLIN? GLIMMIX? Spline? MM?

[plf515]

Using SAS 9.4 1M6 on Windows.

I have been hired to replace a statistician who left a project.

One of my tasks is to help my boss revise a paper that the earlier statistician worked on. I can't get into specifics (it hasn't been published yet) but, briefly, the dependent variable was measured at 5 time points - 3 at or before baseline and 2 after treatment. The previous statistician used what she called "discontinuous growth curve analysis" but cited no source and the results tables don't make it clear what exactly she did.  My boss is not at all statistically inclined and, therefore, can't really help with this. The reviewers did not object to the previous analysis.

I'm thinking that this sounds like a model where there were multiple slopes fit to the results at different time points and that it would be expected that the baseline values would (on average) show no change while the slope from the 3rd baseline value to the first followup and from the first followup to the second would be expected to be nonzero.

I've not done this before in SAS. I'm not sure how to model it. I am considering:

PROC NLIN (seems right, but I have not used NLIN before)

PROC GLIMMIX with a spline effect, restricting it to one knot.
PROC GLIMMIX with an MM effect (I just found out about these).

Any thoughts are welcome

Peter

[PaigeMiller]

Here's an example using PROC NLIN to fit a segmented model

https://documentation.sas.com/?cdcId=pgmsascdc&cdcVersion=9.4_3.4&docsetId=statug&docsetTarget=statu...

[plf515]

Thanks. That one uses a single "subject" and the only covariate is time (and it has a lot of times). I have about 70 subjects and only 5 times - so, I'm wondering if NLIN would be better than GLIMMIX

[PaigeMiller]

I am not aware of any way to force GLIMMIX to fit a discontinuous model. I guess you could manually split up the data into the discrete portions that you wish to fit (if that fits your problem), although that may have disadvantages in estimating errors across the entire data set.

[PGStats]

Create a variable "time since treatment", which takes value zero before treatment, to your data. Put both "time" (general trend) and "time since treatment" (treatment effect) in your model. Start with a linear model.

[Rick_SAS]

It is not clear to me what your model is, but you can use repeated interior knots in a spline effect to get a piecewise polynomial. See the discussion, examples, and references at "Nonsmooth models and spline effects."

[Ksharp]

Since your data is Repeated Measure Experiment , I think you should use PROC NLMIXED / MIXED / GLIMMIX .