Obviously it's not possible to pick a sample size large enough to eliminate all bias. However it's important to pick a sample that doesn't inherently bias the effect that is being measured. This has to apply to both the original and the replication.

The response to the paper points out few examples where the replication was not judicious in picking a sample that minimized its influence on the measured effect. Even besides that, the original study had sampling and scoring procedures which contradicted its original aims.

Lastly, sensationalist headlines of "more than half of all papers are not replicable" are blatantly wrong and not what the original paper even states.

As to your point, I agree that only being able to replicate in one sample type is grounds for skepticism, but more replications have to be done to ensure the replication sample is not the one that is flawed. One can't really draw conclusions unless the replication is repeated a large number of times across a wider sample space.

Link to response paper in case anyone's interested: http://projects.iq.harvard.edu/files/psychology-replications...