The results become more meaningful when the predicted antibody level is plotted against time. This is done for Models 2 and 3 with the outlying observation removed (Figure 6.7). The quadratic model (Model 2) curve is also plotted to
PT_RESID
A AAA AAB
0.000 PT2_RESI
0.005
0.010
Figure 6.6 (continued).
MONTH
Figure 6.7 Predicted virus antibody level vs. time (Models 2 and 3). Curve:_, cubic;
MONTH
Figure 6.7 Predicted virus antibody level vs. time (Models 2 and 3). Curve:_, cubic;
Table 6.7 Results from Model 1 with time centred about its mean of 10 months.
Model
Fixed effects
G matrix and residual
1 - with altered time origin Intercept 3.34(0.2 7)
0.44 0.013
0.56
1.13 0.056
0.56
0.0042
0.0042
make the conclusions more robust since the cubic coefficient is of dubious significance.
The curves for the two models differ markedly for higher values of time. However, only a small proportion of observations were made after 24 months (5%) and the cubic coefficient will be based largely on these. The models are therefore only really plausible up to 24 months, for which they are quite similar. The virus antibody levels decrease most rapidly initially, then flatten. The divergent curves illustrate how over-interpretation might occur when part of them is only based on a small amount of data.
Was this article helpful?
Post a comment