I was at the announcement of the trial data by the investigators this morning. The results were also discussed by a statistician not involved with the trial dir

I was at the announcement of the trial data by the investigators this morning. The results were also discussed by a statistician not involved with the trial directly. Not sure that hacker news needs an in depth analysis of the trial design and outcome, but anyway... In summary, the trial looked at patients who had incurable prostate cancer that had spread and stopped responding to traditional hormonal treatments. Patients had to have no symptoms to be enrolled on this trial. The trial compared abiraterone, a new type of hormonal treatment plus steroids to just being on steroids alone. The question that the trial was asking can be framed as "Does taking abiraterone prolong the period of time that a patient can be symptom free and their cancer is under control?" That was the primary endpoint. A secondary endpoint was survival, or how long they lived. Built in to the study design was a planned interim analysis of the data. Normally, the investigators are not allowed to just look at the data as it accrues continuously because it creates bias. It is a standard approach to allow one look at the data when say 40% of the expected events (in this case, development of progressive cancer) have occurred. When they did this, the difference between getting the drug and not getting the drug was considered big enough to unblind the trial. This means that everyone finds out if they were getting placebo or the real drug. There are multiple reasons why one may decide to do this. First and foremost is that if the drug really works, it would be unethical to leave people on a placebo for long periods, because it is clear they have an inferior outcome. Stopping early also speeds the time to drug approval and market entry for a given indication. There are somewhat esoteric statistical procedures that can determine at an interim analysis if the accrued data shows a statistically significant difference between groups, which would therefore justify stopping the trial, while adjusting for the impact of repeated testing on the type I error rate (ie the chance of finding an association that doesn't really exist). In this trial, there is no doubt that the primary endpoint was positive and clinically and statistically significant, even with only 40% of people on the trial having had a progression event. So they stopped. The data on whether abiraterone reduces the chance of death are not mature, because not enough of the people on the trial have died yet, so those results will be awaited. Stopping early means that these results may be biased however, because the trial is no longer blinded. This was a large and well done trial, that employed standard statistical practices.