This is arguably the strongest crash-based evaluations of Road Diet implementation.

Two likely reasons the results differ from the original Iowa results (below) is that the re-analysis involved a much larger reference group than was used in the original study, and the re-analysis provided  more weight to longer sites (while the original study weighted all treatment sites equally regardless of length).

Differences between the IA sites and those in CA/WA may be a function of traffic volumes and characteristics of the urban environments where the Road Diets were implemented. AADT for the IA sites ranged from 3,718 to 13,908 and were predominately on U.S. or State routes passing through  small towns; AADT for the sites in CA and WA ranged from 6,194 to 26,376 and were predominately on corridors in suburban environments that surrounded larger cities.

Sites with lower crash reduction  factors (CRFs) generally had higher traffic volumes, suggesting the possibility of diminishing safety benefits as traffic volumes increase.

The authors recommended that the choice of which CRF to use should be based on characteristics of the site being considered. If the proposed treatment site is more like the small-town Iowa sites, then the 47% reduction found in IA should be used. If the treatment site is part of a corridor in a suburban area of a larger city, then the 19% reduction should be used. If the proposed site matches neither of these site types, then the combined 29% reduction  is most appropriate.

The traditional before-and-after approach estimated a reduction  in total crashes between 42 and 43%.

A yoked/group comparison analysis found  a 37% reduction  in total crashes and a 46% reduction in PDO crashes (both statistically significant). The reductions in crash rates (per vehicle mile traveled) were 47% for total crashes and 45% for PDO crashes (both statistically significant).

The empirical Bayes (EB) approach estimated a 44% reduction  in total crashes.

This is a relatively  strong crash-based evaluation of Road Diet implementation.  The methodology is a refinement from the 2005 study by Li and Carriquiry.

Unlike the use of linear regression models to estimate expected crash frequencies, this study allowed for different slopes during the "before" and the "after" periods by including a change-point in the model and for the interaction of treatment and slope. As a result, the model allows for a slight increase in crash frequency during the months immediately preceding and following the conversion.

The number of comparison sites (15) is much smaller than the number  of reference sites (296) used in the EB
analysis performed by FHWA (2010).

In general, with elapsed time, the expected number of crashes per mile at each site in the treatment group continues to decrease faster than the number at the corresponding paired site in the control group.

For all treatment sites combined, the frequency of crashes per mile decreased an estimated 34.8%, from 23 pre-treatment to 15 post-treatment, whereas the crash frequency per mile for control sites decreased 6.2%, from 16 pre to 15 post. This would  suggest an estimated 29% net reduction  in the frequency of crashes per mile associated with the Road Diet treatments.

For all treatment sites combined, the annual crash rate per 100MVMT decreased an estimated 43.9%, from 792 pre-treatment to 442 post, whereas the crash rate for control sites decreased 25.5%, from 652 pre to 486 post. This would  suggest an estimated 18% net reduction in the crash rate per 100MVMT associated with the Road Diet treatments.

Stout, T.B. 2005. Before and After Study of Some Impacts of Four-lane to Three-lane Roadway Conversions. Unpublished paper: Iowa State University.
Stout, T.B; Pawlovich, M.; Souleyrette, R.R.; and Carriquiry, A.  2005. Safety Impacts of "Road Diets" in Iowa. Unpublished paper: Iowa State University.
Stout, T.B. Year unknown. Matched Pair Safety Analysis of Four-Lane to Three-Lane Roadway Conversions In Iowa. Unpublished paper: Iowa State University.

The studies reported a greater difference in crash reduction  between  the study segments and the yoked segments than was found between the study segments and the citywide data, which the author(s) attributed to greater variation in the changes in crashes in the yoked segments. The implied degree of effectiveness for the yoked comparison was larger than for the citywide comparisons, and according to the author, might be an artifact of the selection of the yoked segments.

The methodology did not account for possible regression-to-mean effects, and no tests of statistical significance were provided.