I have 15 minutes, so here are some thoughts about ‘sibling-oocyte’ trials in IVF. Generally, these involve taking a woman’s oocytes(eggs) produced from ovarian stimulation, and randomly dividing in half. Each half gets a different treatment.

Sounds good? You get to compare eggs to others from the same person, thereby controlling for patient factors. And if you’re measuring outcomes at the level of the egg (fertilisation? Embryo quality?) then you’re good, provided you use stat analysis accounting for...

clustering within patients (eggs from same person have correlated outcomes). However, people often use these designs and try to measure patient-level outcomes, such as live birth. This introduces problems.

For example, people will select an embryo from whichever batch of eggs did best, will transfer to woman, and will then assign live birth outcome to whichever treatment arm that was. Then compare between arms.

Maybe not obvious, but this is an odd comparison. Comparing treatment A in a group whose eggs did best in treatment A to treatment B in a group whose eggs did best in treatment B. This isn’t a randomised comparison.

One solution people have used is to randomise at two levels. Randomly divide eggs and prospectively randomise women to have transfer in one or the other arms. Better. You can get a randomised comparison of live birth here. BUT...

You’ve still ended up with treatments that aren’t like reality, because you only have half the number of eggs to work with compared to real life. So you expect more cycle cancellation due to total fertilisation failure etc. If there are modest effects on fert rates...

This might translate into larger differences on cycle outcomes in split-oocyte designs compared to reality, where you have more eggs to work with. You have materially changed the treatment.

Failure to appreciate this point appears to have led to greater enthusiasm for ICSI in non-male factor infertility than was warranted. Sibling oocyte trials inevitably suggested greater effect on total fert failure.

Short version: split-oocyte designs are nice for looking at oocyte-level measures. For cycle and patient-level outcomes? Not so much.