What's the problem?
One of the primary issues with retrospective enrollment is the potential introduction of a form of selection bias into the experiment's results, as only those who actually survived the relevant intervention can be enrolled in this manner. If anyone died as a result of the intervention, their data would not be included in the final results of the experiment. This form of bias is more precisely called survivorship bias, as it colors the conclusion towards the survivors (a sub-group) rather than the full population.
This problem thus potentially arises when there is a significant risk that participants could die as a result of the intervention being studied. It is therefore particularly relevant for studies where death is one of the primary endpoints. This is often the case, for example, when dealing with artificial heart valves or similar implants.
Another potential issue with retrospective enrollment is so-called immortal time bias. This type of bias can occur when there's a period through the experiment where a relevant event cannot occur for certain participants (as they otherwise would not have been included or would have been placed in a different group in the study). Imagine, for example, wanting to study whether a certain treatment can prevent relapse for a specific group of patients after a procedure. The treatment is practically carried out within a certain timeframe after the procedure. One looks at a group that has received the preventative treatment and a group that has not. It turns out that those who received the treatment had fewer relapses.
However, this does not necessarily mean that the preventative treatment had a positive effect. None of those in the group who received preventative treatment could (by nature) have experienced a relapse before the time of the preventative treatment – if so, they would have been placed in the other group of the experiment. Therefore, one cannot rule out that any difference between the groups could be entirely or partially attributed to a skewing between participants, rather than a positive effect of the treatment. This problem thus potentially arises when the distribution of participants in the experiment's different groups is sensitive to the conditions being studied in the project.
How can it be solved?