5. PROGRAM EVALUATION (continued)5.3.2 Analytic Approach
The objective of the speed analysis was to uncover any changes in the speeds of vehicles on the test segments between the first wave of measurement when no countermeasures had been mounted by this program and the subsequent data collection waves. Two decisions concerning the nature of the analyses conducted were essentially dictated by the data themselves and the particulars of the test. First, it was decided not to compare the results from Phoenix and Peoria in the same analysis. Combining the data from the two cities would have been of questionable validity because of the different underlying variances as a result of the use of the bin counters in Phoenix .
The second decision was to address the speed analysis as a set of 10 case studies based on each of the 10 test segments described earlier. As shown in Table 59, the widely varying countermeasures applied across the segments suggested that aggregate, citywide analyses would be largely uninformative with respect to the efficacy of individual specific interventions. Table 59 also references the table number in which the speed results for each segment can be found.
Since the primary objective of the statistical analysis was to assess the relationship between the various speed countermeasures and reductions in speed, ANOVA was the indicated primary statistical technique. The ANOVAs for the study were executed through the General Linear Model routine in SPSS® Release 11.5. This technique provided parameter estimates and hypotheses tests of the main and interactive effects of the various countermeasures or experimental factors. Statistically significant main effects and interactions were followed by inspection of the respective subgroup and cell means involved in those effects. All pairwise comparisons of the individual means were evaluated by the Sidak Test. This is a multiple comparison technique for adjusting t-tests for alpha inflation that results when a large number of contrasts are tested. All hypotheses were tested at p £ .05. That is, any difference that would be expected by chance less than 5% of the time was considered to be a true effect.
Chi-square tests were used to evaluate effects on certain binary and nominal transforms of the speed scale (e.g., proportion driving seven or more miles over the speed limit).
The complexity of the experimental plan and absence of countermeasures for some of the waves of measurement required several ANOVA models based on different subsets of the data. For example, in Phoenix waves two and three of measurement were only done on one road which was of particular interest to the participating city personnel. Therefore, these two waves were deleted for some of the ANOVAs so that there would be a full replication of measures at all test locations.
The final analytical strategy involved conducting separate wave x road segment ANOVAs for each city with speed as the dependent variable. The expectation was that road would show a significant main effect since the test segments within each city were quite different as described earlier. A significant main effect of wave would suggest that speeds had changed in the city over the various measurement periods. A significant wave x road segment interaction would suggest that any speed effect was differential across the test sites.
The ANOVAs for both Phoenix and Peoria yielded main effects of wave and road segment as well as a wave x road segment interaction that were significant with a P<.0001. Thus, it was meaningful to examine the pairwise comparisons in light of the known changes on each roadway. This was accomplished by combining the wave and road variables into a single variate and re-running the ANOVAs to produce a comparison of all wave and road pairs. The result was a mean speed differential between each pair of waves on each test segment along with a test of the significance of that difference.
Table 59. Summary of Countermeasures Implemented on Each Test Segment