View PDF Version - RePub - Erasmus Universiteit Rotterdam

Statistical analysis
Dynamic prediction of response to PEG-IFN 65
In a previous study baseline factors infl uencing the SR-rate were described in detail
and a prediction model (PEG-IFN HBV Treatment Index) was developed which provides
a subject specifi c prediction of SR. 9 Baseline factors associated with SR were: HBVgenotype,
age, gender, baseline HBV DNA (copies/ml, log10 ), ALT (loge ) and previous
treatment with IFN.
In the current study the PEG-IFN HBV Treatment Index was fi rst calculated for each
subject and was used throughout the analysis as an offset, i.e. a subject specifi c starting
value with the corresponding regression coeffi cient set to 1. The baseline prediction
of SR, Pbaseline , was hereafter updated with data on HBV DNA and ALT during therapy
applying logistic regression analysis techniques. For comparison purposes two model
approaches were considered: the fi rst was updating the model at each visit with the
new information, resulting in 8 models; one for each visit (week 4, 8, 12, 16, 20, 24, 28
and 32), the second an overall generalized estimating equations model11 using information
of all visits and adding the visit time as a continuous factor. The latter model
approach allows repeated measurement data and hence reduces the 8 models of the
fi rst approach into one overall model. Technically, each visit per patient is treated as
an observation and the model then corrects for the fact that each patient contributes
with 8 visits. This model is referred to as the dynamic logistic regression model and is
sometimes called a pooled logistic regression model. 12-13 The treatment duration (visit
time) was added as a linear variable to the model, a restricted cubic spline was used to
check the linearity assumption. The effect of the crude HBV-DNA (log10 ) as well as the
effect of HBV DNA log10-decline compared to baseline (= HBV DNAlog10 at time t – HBV
DNAlog10 at baseline) was studied. ALT was entered in the models as measured and also
transformed logarithmically. Interactions between HBV DNA and ALT with treatment
duration (visit time) and with HBV-genotype were considered. Changes in HBV DNA and
ALT measurements in prior visits were also studied. Discrimination between the different
models was quantifi ed by the c-statistics, which is the area under the receiver operating
curve. The best model-fi t was assessed comparing these (the higher the better) and the
Akaike’s Information criteria (AIC) or the quasi-likelihood information criteria (QIC) for
the generalized estimating equations method (the lower the better). Cross validation,
leaving one out, was performed to establish overall performances of the models.
Finally a cut-off value of HBV DNA during treatment was sought to fi nd a clinical useful
guiding rule for (dis)continuation of therapy. The optimal cut-off point was established in
a multivariable setting14 including the baseline PEG-IFN HBV Treatment Index: explanatory
plots were evaluated, followed by the maximum chi-square approach. Subsequently
the cut-off point search was repeated in 500 bootstrap samples for validation of the