East: Example 4: A Case Study:

Design and Interim Monitoring of a Randomized Clinical Trial using EaSt
Protocol 8433 (conducted in the United States by the Cancer and Leukemia Oncology Group B--CALGB) was a phase III clinical trial of standard radiotherapy with or without a neoadjuvant combination chemotherapy in patients with regional stage III non-small-cell lung cancer. The study was activated in 1985, monitored at administratively convenient time-points, and terminated two-years earlier than an equivalent fixed-sample study would have been. This case-study demonstrates how one might design and monitor protocol 8433 using the EaSt software package.

1. Study Design

The median survival time of stage III non-small-cell lung cancer patients receiving standard radiotherapy is known to be 9 months. Based on past history it is expected that patients will accrue to this study at the rate of 5 per month. The goal is to design a study having 80\% power to detect an increase in median survival from 9 to 13.5 months, using a two-sided test at the 5% significance level. East has a convenient spread-sheet like user interface which makes it very easy to input these requirements, observe their ramifications interactively, and alter the input parameters if the ramifications are unsatisfactory. The final East output screen compares a fixed-sample plan with a group-sequential plan based on 3 interim analyses and early stopping contingent on crossing an O'Brien-Fleming stopping boundary.

Both plans achieve the desired 80% power with a sample size of 220 patients. However the fixed sample plan requires that the study remain open until 191 deaths are observed. This is expected to take 54.24 months under the null hypothesis and 59.89 months under the alternative hypothesis. That is, we are committed to a 5 year study if the alternative hypothesis is true.

The group-sequential plan gives us some flexibility to stop early by crossing the O'Brien-Fleming stopping boundary. Notice that under the alternative hypothesis we expect to close the study in 52.05 months and thereby trim 8 months from the fixed sample plan without compromising its power or significance level. We do pay a small price for this potential savings however. We are now committed to keeping the study open until 194 deaths are observed (instead of 191) unless we can exit earlier by crossing a stopping boundary. The maximum expected study duration is thus increased from 59.89 months under the fixed sample plan to 61.93 months under the group-sequential plan. 

2. Interim Monitoring

Protocol 8433 was activated in early 1985 as a group-sequential clinical trial. Using East we have demonstrated that a study design based on 3 interim looks at the data, spread out uniformly over a total of 194 deaths, with early stopping contingent on crossing an O'Brien-Fleming stopping boundary, would yield 80% power and limit the type-1 error to 5%. Notwithstanding these design constraints the Lan-Demets error spending function methodology adopted in East gives us considerable flexibility for interim monitoring. We are not obliged to monitor the data exactly 3 times, nor are we obliged to monitor the data after equal numbers of deaths. Under Lan-Demets we are free to monitor the data after equal numbers of deaths, after equal intervals of time, or even monitor sporadically. In all cases the 5% type-1 error rate will be preserved. This flexibility is extremely valuable as it is often administratively impossible to prescribe in advance when and how often the interim monitoring will be performed.

The first interim analysis of protocol 8433 was performed in the Fall of 1985, when only 10 deaths had been observed. East computed the stopping boundary needed to stop with only 10/194, or 5% of the total information (or process time) available. To stop the study at the first interim look the nominal p-value based on a logrank test would have to be no greater than $10^{-10}$. This small a p-value was not attained and so the study proceeded to a second interim analysis in the fall of 1986, at which time a total of 34 deaths (or 17% of the total information) had been observed. Now East calculated that the nominal logrank p-value needed for early stopping would have to be 0.0000014. Since the observed p-value was 0.0071, the study remained open and a third interim analysis was performed in the Winter of 1986 with 50 observed deaths (or 26% of the total information). Now the nominal p-value for early stopping was 0.00008. The observed p-value was 0.0015 and he study continued on to a fourth interim analysis in the Spring of 1987 by which time a 100 deaths (52% of the total information) had accumulated. This time the nominal p-value needed for early stopping was was 0.0042 while the observed p-value was 0.0015. Thus the study was terminated.

It is almost impossible to stop a study at the early stages of a clinical trial if the O'Brien-Fleming stopping boundary is adopted. For this reason the O'Brien-Fleming stopping boundary is very popular. It closely resembles the fixed-sample approach in the early stages of the study when the data are not very stable, but does provide the possibility of early stopping later on. Although initially designed for 3 interim looks, each after equal numbers of deaths, the study was actually run along different lines. There were 4 interim looks, and they were not spaced out equally over the deaths.

Nevertheless the Lan-Demets methodology ensured that the type-1 error of 5% was preserved. Protocol 8433 was a positive study and it did stop two years earlier than a corresponding fixed-sample design would have stopped. This resulted in substantial ethical and economic advantages to the CALGB.