Can one use different samples in pre and post tests? Generally I have seen that pre and post intervention tests are applied on the same sample from a population undergoing an intervention. e.g. same set of students may be tested in an educational intervention in both pre and post tests.
Would it be valid to test different random samples in both the pre and post test (using the same sampling methods)? 
If this is statistically valid, this would have an advantage that the agency responsible for intervention cannot pay special attention to the individuals who are part of the pre-test sample.
 A: If you take a different sample of participants at the pre- and post-test measurement points, you are in effect back to a traditional between-subject design. The main drawback is that the precision of effect estimates and the power of the corresponding hypothesis tests will be lower (possibly much lower) because inter-individual differences will end up in the error term.
In fact, measures taken before the intervention have a much more limited use in this scenario because they cannot be easily matched to the post-intervention data (in the traditional before/after design, each participant is traditionally said to “serve as his or her own control”). Unless you want to study some ancillary hypotheses, you could therefore just as well dispense with them and simply compare post-intervention scores to a properly selected control group measured at the same time. If people were randomly assigned to the intervention, this is as clean a measure of effectiveness as you are going to get.
If you don't have a control group, causal interpretations are in any case very difficult and you cannot simply assume that any measured change results from your intervention. If you do have a control group, you still need a way to make contact and assign participants to a condition before the intervention, identify them, track participation and ensure compliance/lack of contamination between the groups, which could negate some of the practical advantage of the suggested design.
Another common approach is to keep the groups separate by testing the intervention in different geographical areas/different institutions instead of randomly assigning individual people to each condition. In educational research, some interventions also cannot be assigned to individual students but only make sense at the school level. Here again inference is more difficult compared to a traditional before-after design and data from such studies should be analyzed with appropriate techniques (multilevel/hierarchical models, etc.) so the question becomes moot and you have many more design options to consider.
All this shows that this is not a black-and-white (valid/invalid) situation and explains why many studies are still planned following the traditional before/after design.
