First, we determined that by equally distributing resources, we could mail survey packets to 168 households
in each of the 178 districts. However, we decided it was better to oversample (i.e., mail more than an equal
number of survey packets) some districts that we expected had a high proportion of out-of-district shopping,
such as Cheraw District. More completed surveys from these districts (typically rural or small and adjacent to
more populated districts) would result in a lower margin of error, which would have a far greater positive impact
on the confidence of the shopping pattern results. On the other hand, we could reduce the number of packets
mailed to districts that we suspected had very little out-of-district shopping, such as Poudre District.
Second, we also consulted imputation percentage results from the Census Bureau’s American Community
Survey to flag school districts where we might expect lower response rates. We slightly increased the number
of survey packets mailed to these districts. Finally, in ten districts, we acquired fewer households mailing
addresses than called for in our sampling plan, in which case we mailed survey packets to all available addresses
in those districts. In total, we mailed survey packets to 30,295 addresses.
The survey was primarily executed via mail. Mailed packets included a cover letter, survey instrument, map
(where needed along the Front Range), and postage-paid return envelope. Shortly after the full survey packet
was mailed, a postcard reminder was sent to all survey recipients to further encourage response. An incentive
was also offered in the form of a prize drawing. Respondents could enter to win one of five $50 Visa gift cards.
As required by law, an alternate mode of entry was also provided.
As data collection proceeded, Corona monitored responses by district and deemed it necessary to boost
response. Due to the overall project timeline, this was done via two methods: (1) Corona sent 3,910 postcards
to a new sample of mailing addresses, encouraging residents in the household to respond to the survey online.
Postcards were primarily sent to small and rural districts. Each household received a unique one-use access
code. The same incentive described above was used. (2) Corona worked with an online sample provider
(Dynata) to collect additional responses online via their panel. Corona screened all participants to ensure
eligibility and a small incentive was provided via Dynata. For both online surveys, the survey was programmed
to mimic the paper survey to reduce any mode differences. A total of 3,368 (2,078 via the first mail survey,
1,275 via the panel survey, and 15 via the postcard survey).
Using the data gathered in this survey, the research team developed a family of predictive models to
estimate geographic shopping patterns in each district for each product category. As a first step, the team
reviewed all responses and assigned each home and purchase location given by respondents to an individual
school district. In cases where a city provided by a respondent included multiple school districts, the response
was assigned to one of the possible school districts in that city based on a randomization function weighted by
the number of businesses in the city that were located in each school district. Purchase locations outside of
Colorado were removed from the analysis.
After all responses were assigned to a district, the team further cleaned the data by individually inspecting
any purchase districts that were more than 100 driving miles from the respondent’s home location. If the
purchase location was in a major city or in a city in the same region of the state, it was deemed to be valid.
However, in cases where the purchase location was in a completely different region of the state, that data point
was removed from the analysis as an outlier.
Once the data were fully cleaned, the team developed predictive models to forecast the purchase district(s)
and proportions of purchases from each purchase district for residents of each district. For example, the team
developed data that show what proportion of haircut purchases by residents of District A were made in Districts
B, C, D, etc. The goal of this was to take into account the prices of goods not only in a district’s own business
community, but also in other nearby communities. Depending on the particular geography, shopping patterns
for any district might include only one district where shopping occurs or might include many districts. The final
shopping patterns matrices are presented in Appendix F.