2
2. The CISS Sample Design
CISS was designed independent of other NHTSA surveys. The target population for the CISS is
all police-reported motor vehicle crashes on a traffic way, each involving a passenger vehicle
1
and in which a passenger vehicle is towed from the scene for any reason. This definition is
slightly different from the CDS, which required that a vehicle be towed due to damage. This
change was made because sometimes it was difficult to determine why a vehicle was towed.
Because a direct collection of crashes in the nation is infeasible, the CISS crash sample is
selected in multiple stages to produce a nationally representative probability sample.
At the first stage, 3,117 counties in the United States were grouped into 1,784 Primary Sampling
Units. A PSU in the CISS is either a county or a group of counties. U.S. territories, some remote
counties in Alaska, and small islands of Hawaii were excluded. PSUs have been formed in such a
way that there is a 90 percent chance to have at least 5 fatal crashes every year inside each PSU
and the end-to-end distance of a PSU was 65 miles for an urban area and 130 miles for a rural
area.
The 1,784 PSUs were stratified into 24 strata by the four Census regions, urban/rural, total
highway/primary/secondary road miles, and total expected number of crashes. Each of the 1,784
PSUs in the frame was assigned a measure of size (MOS) equal to the combination of its
estimated seven types (defined by injury severity and vehicle model year) of crash counts.
From each of the 24 PSU strata, 2 PSUs were selected by a probability proportional-to-size
(PPS) sampling method. In addition, one large PSU was selected with certainty. This resulted in
a total of 49 PSUs. Then a sequence of PSU sub-samples was selected from the 49 PSUs with
decreasing sample sizes. In this process the PSU strata were collapsed when necessary. This
process produced a sequence of nested PSU samples. These nested PSU samples allow NHTSA
to change the PSU sample size without reselecting the sample. The final PSU sample is the result
of multiphase sampling and the PSU sample selected in such a way remained generally PPS.
For the 2017 CISS, for example, 24 PSUs were selected from 12 PSU strata (2 PSUs selected per
stratum) using the above process. Consequently, the PSU sampling rate was very low in each
stratum. Because of the reduction of PSU sample size (from 49 to 24), no PSU was selected with
certainty. All sampled PSUs cooperated with NHTSA’s data collection request.
The Secondary Sampling Units are police jurisdictions. Within each selected PSU, PJs were
stratified into three PJ strata by their estimated measure of size - a combination of crash counts in
six categories of interest. The Pareto sampling method (Rosén, 1997) was used to select PJ
samples from each PJ stratum. The Pareto sampling method produces overlapping samples when
a new sample is reselected. This reduces the changes to the existing PJ sample when a new PJ
sample needs to be selected because of PJ frame (the collection of all PJs in the selected PSU)
changes. The PJ inclusion probability under the Pareto sampling is approximately PPS (Rosén,
1997). In 2017 CISS, across the 24 sampled PSUs, a total of 182 PJs were selected and 168 PJs
1
CISS-applicable vehicles are the same as CDS-applicable vehicles: passenger cars, light trucks, vans, and sport
utility vehicles with gross vehicle weight rating (GVWR) less than 10,000 lbs.