Data collection gone awry -- now what do I do?

#1
Trying to keep this as brief as possible ...

I handle the stats for a pilot project for my employer (a government agency). We're trying to determine whether a particular type of technical assistance improves regulatory compliance for a certain type of business. To do this, we designed a methodology incorporating (1) a round of baseline visits, (2) our intervention, and then (3) a follow-up round of visits.

The site visits are performed by local government employees who work for other agencies. Each local government jurisdiction was a separate stratum. Unfortunately, we've had a problem with one of our strata not following our instructions about how to select sites for a visit. (We had given them a randomized list and very clear instructions, but they decided to do something a little different and didn't tell us until it was too late.)

The employees conducting the visits for this stratum (a county) have asked us to split their stratum into multiple strata based on their respective areas of responsibility. Obviously, I don't want to do that, but for reasons I won't go into here, I think I have no choice at this point.

I'd love to have some input about:

1. They already messed up the data collection for the first round of site visits because they did not select facilities at random. I cannot re-do the data collection for this step. Besides explaining what happened and the increased likelihood of selection bias, is there anything else I can do? I don't believe any of my variables will allow me to control for self-selection bias, so I think I'm kinda stuck. Yes? Anything I'm overlooking?

2. If I break my problem stratum in two, I've increased the likelihood a facility will get selected for a visit. Is there a way I can control for that? Should I try to weight cases from this jurisdiction? Or should I just let it go?

3. If I break the problem stratum in two, which formula do I use to determine the number of site visits needed for a valid sample? Until now, I've used a formula to determine differences in proportions based on 2 samples. This is the second round of visits, but these will technically be new strata. Is this a 1-sample or a 2-sample test? I'm inclined to use the larger 2-sample number to ensure a valid sample, but would like some advice about the potential downside to this approach.

4. Am I overly concerned about splitting the stratum? They already messed up the data collection. Is splitting the stratum really going to make it any worse?

5. Since the beginning of the project, I have made it clear to management that the non-parametric nature of this study means they cannot use the results to generalize about the rest of the industry statewide. They say they understand that. The improper methodology in my problem stratum also means I can't generalize baseline results even for that county. Assuming the local folks follow instructions this time, anyone see any problem in generalizing follow-up results to the rest of that county? Does breaking the stratum in two affect the results?

Of course the problem strata accounts for almost one-third of my entire universe ... naturally. Anyone have any brilliant ideas or insights to help me figure out how to salvage this? I'm also interested in any problems I've overlooked. Thanks!