Determining an effect of a seatbelt enforcement and a public outreach program in a town We conducted a seat belt use observation study in two similar sociodemographic towns with different media markets. We sampled two towns and measured outcomes in a pre and post format.  One town was a control site. The other was the treatment site. The treatment site police department conducted an enforcement and public outreach program. Over 3,000 seat belt observations were collected in each town pre and post. The treatment site went from 85.6 percent to 91.2 percent increase in belt use...the control site went from 87.7 to 85.2 percent. What statistical analysis can I use to determine the effect of the treatment program?
 A: A great review of statistical methodology for policy evaluation can be found in "Analysis of longitudinal data to evaluate a policy change." by French, Heagerty 2008. Although most of the methods extend to multiple sites (like 50 US states) or several time periods (like 20 quarters pre/post).
Your example can be seen as a special case of either scenario. You have a treatment group and a control group, and the basic question is: did the proportion experiencing an event at one site decrease relative to the proportion experiencing the event at the other site.
I assume this is a cluster longitudinal design without replicates. Replicates, in this case, would mean knowing whether the specific individual at each site was sampled for seatbelt use, which isn't the case here. 
With 2 sites, 2 time periods, and balanced design, these data are amenable to a difference-in-differences design using logistic regression. You would fit a binary effect for pre-post and a binary effect for site, as well as their product and test this against a simpler model which only controls for a pre-post indicator and a site indicator. The complex model is a saturated model, perfectly predicting the observed seatbelt proportion at both sites. The simpler model makes an assumption, under the null hypothesis, that the odds of seatbelt use over time are proportional for the two sites. This is the same as saying that the product term in the complex model is 1 exactly.
A likelihood ratio test of these two models directly tests the hypothesis that the pre-post proportion-change at site 1 was different than the pre-post proportion change at site 2. 
