D9901 enrollment estimate: Ocyanblue
First, to correct a mistake, I meant the censoring shown for the progression data would give a hint to the enrollment pattern, not the survival data. Sorry about that mistake. Hoooow-ever, I may still have a point. What I did to get an impression of the D9901 enrollment pattern was the following:
For the analysis, I used the (time, % non-progression) data points for the 10 censored patients in the treatment arm.
If, say, a patient is censored after 20 weeks and if 25% of the patients would have progressed by this point in time, it is reasonable to assume that 1/25%=4 patients were enrolled 20 weeks before the trial was locked (three of the four patients would statistically have evented before 20 weeks, and the one who was censored would event some time after 20 weeks). Doing this for all the 10 patients gives a set of 10 new (time, number of patients) data points.
The time frame for the progression results is 75 weeks, and to transform the time values to time after the start of this 75 week period, a new set of data points is generated by subtracting the time of censoring from 75, i.e. (75 - time, number of patients). Plotting the accumulated number of patients as a function of the "new" time, gives (I hope...), an impression of the enrollment pattern.
I made the plotting, and linear regression gave the following equation:
<accumulated number of patients>=0.8217*<"new" time>+14.6. R2=0.928...
Test1: Inserting "new" time=75 gives an accumulated number of patients (in the treatment arm) of 76, which is not too far from the real life number of 82.
Test2: Inserting "new" time = -9 (75+9=84, keeping in mind that the enrollment of D9901 took 84 weeks) gives an accumulated number of patients of 7. Which is perhaps a little off the real life number of zero, however, I don't think it invalidates the approach - this is not supposed to be exact!
I have to admit that this is at the very best a very rough indication of the enrollment pattern, but I think (until Iwfal convinces me of the opposite) that this is a valid approach. Comments are utmost welcome!
Br.
Bengt
PS! The data set that I used is based on very quick and dirty readings from fig.2 in the 2006 JCO-paper by Dr. Small. Some uncertainty is introduced by this lack of precission.
AI
