Biostatistics 537: Longitudinal Data Analysis - Fall 2008
Problem Set 4 - Due: Thursday October 30, 2008
The data for this problem are from the Riesby et al., article that we have discussed in class.
This study examined the relationship in depressed inpatients between the drug plasma levels - the
antidepressant imipramine (IMI) and its metabolite desimipramine (DMI) - and clinical response
as measured by the Hamilton Depression Rating Scale (HDRS). In class, we noted that there was
a significant relationship across time between the drug plasma levels (specifically, desimipramine)
and depression. What I would like you to do for this assignment is examine the degree to which this
posited relationship is influenced by the variance-covariance structure (of the dependent measure
across time) that characterizes different statisticalmodels of the data. The dataset RIESBYT4.DAT
is available on the class website and contains the following variables:
field 1: Patient ID
field 2: HDRS change from baseline score
field 3: a field of ones (is “one” the loneliest variable?) - ignore this variable
field 4: Week - from 0 (week 2) to 3 (week 5)
field 5: sex (0 = male 1 = female) - ignore this variable
field 6: diagnostic group (0 = non-endogenous 1 = endogenous)
field 7: Imipramine (IMI) plasma levels (in ln units)
field 8: Desimipramine (DMI) plasma levels (in ln units)
For this problem (as in problem 3), I would like you to combine the drug plasma levels into one
variable - the natural log (ln) of the ratio of DMI to IMI (i.e., lnDMI - lnIMI). Let’s denote this
variable as LDIM. For this problem set do the following:
1. Consider a model with fixed effects of WEEK, WEEK2, LDIM, ENDOG, and the interaction
of ENDOG by LDIM. Decide on either ML or REML estimation and then perform a co-
variance structure selection using ideas discussed in class. What covariance structure do you
settle upon (note: you may want to consider a few models with random effects, or covariance
pattern models, or random effects plus autocorrelated errors of some sort)? What criteria do
you use to make this selection? What is your interpretation of the covariance structure and
the fixed effects in your model? Summarize your findings.
2. Suppose Researcher A says “covariance structure, my foot! If compound symmetry is good
enough for my hairstyle, it’s good enough for me!” and decides to do an analysis using the
same fixed effects as above, but only allowing for a CS structure on the dependent variable
across time. Is Research A likely to report any dubious findings with regards to the fixed
effects in the model?
3. Suppose Researcher B says “covariance structure, my eye! If unstructured is good enough
for my closet, it’s good enough for me!” and decides to do the same analysis, but using an
unstructured covariance structure for the dependent variable across time. Is Researcher B
likely to report any dubious findings with regards to the fixed effects in the model?
4. Summarize your feelings regarding covariance structure selection, and its place in statistical
modeling of longitudinal data.