06_conclusion.qmd

# Conclusions and Recommendations {#sec-conclude}

Access to nutrition is a critical topic that has been a frequent focus of
academic literature in public health, community planning, and engineering, but
the varying and incomplete quantitative definitions of access have perhaps
limited efforts at developing solutions.

In this research, the model we developed and informed with location-based
services data suggested that policies that increased the availability and
quality of grocery stores would do more to enhance access to nutrition than
strategies that increased the transit level of service and lowered active transport 
travel times in three
different Utah communities. These findings may not be general, but they largely
support the review of food desert literature in @beaulac2009 as well as 
critical evaluations by @shannon2014 and @wright2016.

## Limitations {#sec-limitations}

A number of assumptions made in @sec-methods lead to limitations that other
reasonable researchers might pursue differently, and thereby obtain marginally
different outcomes.

In selecting a survey instrument with which to collect the store attribute data,
we selected an existing and validated instrument from the nutrition environment
literature. The NEMS-S focus on low-calorie and low-fat alternatives may be
somewhat outdated in view of modern nutritional guidelines. For example, the
NEMS-S does not track the availability and price of poultry, as "lean" poultry
is not a goods category in the way that ground beef comes with multiple fat
contents. Thus a major source of protein in typical American diets [@fns2021]
was not traced across stores in each community. On a more basic level, the
NEMS-S attempts to measure a store's stocking of goods that researchers believe
are beneficial, and does not measure either what people wish to buy, or what
they are actually buying at a store. Future research might attempt to survey
shoppers on what they actually purchased at each store --- or collect receipts
of their purchases --- though this would substantially raise the difficulty of
collecting data.

Most households do not obtain all their groceries at a single store, though this
research of necessity assumed that a simulated person chose exactly one store
from stores available to them. Similarly, the location-based services data
provided by StreetLight and used to identify which stores people traveled to
only reveal whether a device was identified inside a geographic polygon, and not
what they were actually doing in that polygon. This research had no way of
distinguishing, for example, whether an individual device observed at a dollar
store or a super market (e.g. Wal-Mart) was there to purchase groceries or some
other household goods that might not be offered at more traditional food
markets.

A number of simplifying assumptions concern the socioeconomic and
spatio-temporal detail supplied to the choice models and accessibility
calculators. The StreetLight data do not contain any demographic information on
the individuals making trips, beyond the inferences made possible by the
residence block group. This makes it difficult to estimate whether lower income
households are more or less sensitive to travel distances or prices.
Additionally, the research assumed that every trip was from the
population-weighted block group centroid, which may vary substantially from the
actual distance traveled, especially in large block groups. The methodology also
used travel times calculated in the AM peak hour; though this time maximizes the
availability of transit options, it is not a typical peak time for grocery
shopping. The mode choice model was also selected for general convenience, and
not fully calibrated to grocery trips in the specific regions. All of these
limitations could potentially be relaxed by using a synthetic population with
detailed socioeconomic data and parcel-level location choices
determined by an activity-based model, as proposed by @dong2006. In this
exercise, which we leave to future research, the grocery maintenance trips could
be explicitly modeled, with synthetic individuals of unique characteristics
choosing destinations that are available on the course of their other daily
activities, using their chosen travel modes. This proposal would effectively 
extend the methods of  @widener2014 within the framework of explicit
activity-travel modeling.

Finally, the scenarios presented in @sec-scenarios are designed to illustrate
potential applications of this accessibility methodology, with a comparative
analysis of strategies to improve access to nutrition. Selecting different
sites, attribute levels, or transport policies might substantially change the
scale or rank-ordering of the estimated benefits. A comprehensive search for the
location that would maximize benefits would be an interesting exercise, which we
also leave to future research.