1) You run a trial
2) The treated cohort does not perform better than placebo
3) But you look at the data and find that a subgroup did better than placebo
4) You then deep dive into the characteristics of the subgroup and choose a specific characteristic that most of the patients in the group share. For example, being blond or being on memantine.
5) You run a new trial that will recruit only patients that share this characteristic.
In this way, you introduce selection bias in your drug testing process. Because of this bias, it is likely that the new trial will fail as well due to the high likelihood that the observed performance/effect in the subgroup in the original trial was by chance.
However, this is common in many trials. While there are ways to correct this bias in advance, many biotechs neglect it.