Dear Avi Z

It seems you should make your decision according to ethical consideration of your intervention. So that Interim analysis is suggested.

Ahalan Avi,

You have a lot of different things going on in your post, and it is making the question somewhat convoluted. Let's try and take things one at a time:

First, you wrote that your new intervention should perform no worse than your current treatment, but then you write that you expect it to perform better. These statements lead to two different approaches, so you need to be clear on what it is that you expect.

Second, you say that you expect that the two treatments to have somewhat different results at 1 year. So there is a problem here in that at the one year mark of the study you will expect to show no effect. Thus, it doesn't make sense to design a study that will require significant resources - but will show no effect.

Third, you write that you do expect to see a short term benefit to the new treatment over the current treatment. This is both good and bad. The good is that you expect to see an effect, the bad is that the effect will wear off somewhere between month 6 and 12.

So where does that leave us?

As you suggested yourself, you could implement the study with a 12 month follow-up period, and measure the monthly effect with an interaction effect (time X treatment). This will show that the new intervention had an effect earlier than the current treatment, but that the effect was short-lasting. This would be the preferred approach. 

You could also set the outcome to be a threshold of some sort, and then use a time-to-event analysis to demonstrate that the new treatment hit this threshold earlier than the current treatment. This approach is not recommended because (a) it collapses a continuous variable into to dichotomous variable and therefore results in a loss of information, and (b) the loss of effectiveness between 6 and 12 months will not be captured.

I hope this lengthy response is helpful for you. Be' hatzlakha!

Ariel

Thank you all for replying. Ariel, ahalan, let me be slightly more clear, I did not mean that there will be no effect. I'll give you an imaginary example from a different field to illustrate. Let's say a man is in pain, and currently you have a drug for pain, and it taken the drug 3 hours to work, but when it does, it handles the pain quite good. Then the current treatment, will treat the pain just as good (at least as good as the current drug), but will do it much faster, which is of course beneficial for the patients who don't need to suffer.

The time to event approach, as was also suggested by Alvaro, is very nice, however, my fear is that with be effort I will have up to 4 follow-up visits (and in the worst case only 3), and survival analysis might not be adequate.

So combining your advises, I tend to think that the best way is to declare a primary endpoint of non-inferiority on the long run (the last follow-up), to calculate the sample size accordingly, and a secondary endpoint of superiority at a smaller time point, leaving only one concern: for non-inferiority the sample sizes are usually smaller, will that leave me enough power to test the secondary endpoint (with an interaction like you said - a longitudinal model: GEE or mixed model).

If this will be an non-inferiority trial, your delta value should not have anything to do with time. If you feel that the effect will wax and wane against time, than an early interim analysis may be misguiding with narrow confidence interval range; thus showing an positive effect. I think for drugs that show temporary or short lasting benefit, adaptive clinical trial strategies do not work. But if you are doing a trial with a chemotherapeutic drug with sustained or long lasting effect; than an interim analysis may be feasible.