We are currently working in our local hospital on a paper about the influence of the nephrologist on patients with kidney failure.

We are aiming to look at a couple of clinical (laboratory) outcomes such as blood pressure and eGFR (reflects the functioning of the kidney). All the outcomes are numeric (about 10).

We will gather data from patients from 1 year before treatment to 1.5 year after the start of the treatment.

However, the amount of available data varies across the patients. For some patients we only have a few data points, for other patients we have quite a lot of data points.

For this reason we are thinking to look/compare the average values at different moments.

Something like:

  • t1: 1 year till 0.5 year prior treatment
  • t2: 0.5 year prior treatment till start treatment
  • t3: start treatment till 0.5 after start treatment
  • t4: 0.5 year after start treatment till 1 year after start treatment
  • t5: 1 year after start treatment till 1.5 year after start treatment

However, I'm not quite sure if this is the right way to go. The goal is to look if the clinical values are stabilizing (or improving) after the start of the treatment. In other words if the kidney of the patients is detoriating at a slower speed.

My questions are:

  • What kind of model would be most suitable for this kind of data and question?
  • Would it be an option to look at the slope coefficients of the clinical values at the different moments?
  • Are there minimal requirements on the amount of data for each patient?

I hope my question is clear, if not please let me know, and I will try to explain it some more.

Any help/advice is appreciated! Thanks already in advance!


1 Answer 1


It sounds like this is a good candidate for growth modeling. Raudenbush and Bryk (2002) is the seminal source for this. See Chapter 6.

Their approach assumes reliable measurement, so if that is an issue, you might also try Latent Growth Modeling via any available SEM program.

Speaking only for the first case, the approach based on hierarchical linear modeling does not assume that all participants contribute data at all time-points, nor does it require equal spacing of time points. However, it is crucial for interpretation of the effect to choose some sensible time points, and it seems that you have (~ every 6 months). With 5 time points you can model a quartic growth function, which is probably not necessary, but suffice to say that 5 time points is good enough for linear or quadratic growth modeling!

I find the lme4 package in R to be the most useful for hierarchical linear modeling, as it enables all the data to live in one dataset, however, there is a program called HLM written by the authors of the book that people also use.

  • $\begingroup$ Hello Rick, thanks a lot for your suggestion. Will definitely look at the book you are referring to. Fortunately, I'm experienced with R so the lme4 package will be explored. $\endgroup$
    – Erik
    Commented Jan 27, 2021 at 20:18
  • $\begingroup$ Awesome! I think you'll find it very helpful! There are other more general resources for "mixed-effects" / "multilevel" models, but theirs is really the best for longitudinal studies IMO $\endgroup$
    – Rick Hass
    Commented Jan 28, 2021 at 15:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.