Secondly, the data of the days with less than five
patient samples analyzed were also eliminated. Since data were
collected over a period of several months, analytical drifts or shifts
might have occurred and therefore we decided to normalize our
values by subtracting the daily median from the individual patient
results. Using these normalized COI values, we had to incorporate
the uncertainty of the calculated medians in our formulas as well:
X < COL − (p95 − p950)xp,i − (p97, 5 − p950)m
(for the non − competive assays)
X > COL − (p5 − p50)xp,i − (p2, 5 − p50)m
(for the competive assays)
with (p95–p50) xp,i and (p5–p50) xp,i being the spread between the
95th or 5th percentile and the median of our normalized patient
values; (p97,5–p50)m and (p2,5–p50)m being the proportional half
of the 95% confidence interval on our calculated daily median. In
order to quantify the uncertainty of the calculated medians, bootstrapping
statistics was performed on our normalised dataset [10].
1000 resamples of 5 individual patient results were taken. This generated
a dataset of 1000 medians, of which the distribution was
calculated. This distribution is in fact an overestimation of the real
variation because often more than 5 patient results per day are
available.