5 Limitations and Future Directions

The utility-based accessibility metrics we present and apply in this paper are evaluated from a discrete choice model estimated on simulated decision makers constructed from a third-party passive origin-destination matrix. This methodological choice has some strengths: Foremost among these is the ability to readily and affordably construct a large dataset on an infrequent trip purpose. Most destination choice and activity location models are estimated on small-sample household travel surveys. Securing sufficient responses to estimate a rich behavioral model on a trip purpose as infrequent as parks has proven prohibitively expensive outside of extensive research activities (e.g., Kaczynski et al. 2016). Using passive data sets to increase the effective sampling rate possible in a discrete choice model is a potentially powerful strategy, and its application here is an important contribution of our work.

At the same time, passive data sets available from commercial providers do not reveal any details about the specific trip makers beyond what can be learned from their residence block group. In this research we were able to determine whether a device resided in a block group with a high proportion of low-income households, but could not have confidence that a particular device belonged to a member of such a household. Similarly, there is no information on what kind of trip the device-holder actually accomplished at each park. These limitations combined mean that it would likely be infeasible to directly observe devices that traveled to the converted streets during the COVID-19 lockdowns. The ideal dataset for estimating individual park activity location choices generally and in special situations remains a high-quality, large-sample household survey of real individuals.

The individual-level demographic data would also be valuable in understanding more clearly the observed heterogeneity in response among different income or ethnic groups. The trends and correlations revealed in the presented models may reflect situational inequities rather than true preferences. For example, the distinct observed parameters on size and distance for block groups with high minority populations may indicate that areas with large minority populations tend to have smaller parks that are more geographically distributed relative to other areas of the region. This interpretation could also explain some of the non-intuitive response observed in our models, especially in regards to playgrounds.

We limited our analysis to home locations and parks in Alameda County, California. It is possible that some Alameda residents visit parks in neighboring counties, just as it is possible that parks in Alameda County attract trips from outside the county borders. This is most likely for block groups and parks on the north and south borders of the county. The scope of this analysis was determined by the passive data set available for the research, but the county boundaries are not a general requirement for all studies of this kind.

The distance to a park was represented in this study using a walk network retrieved from the OpenStreetMap project. Though perhaps superior to a Euclidean distance, this measure still has many limitations. First, we were unable to verify the integrity of the underlying network information; based on our prior experience, it is likely that some broken or improperly connected links artificially inflated the measured distance for an unknown number of park / block group pairs. A more serious limitation, however, is that experienced travel distances are a function of the transport mode employed by the traveler. Using bare distances does not provide any detail on how access to parks might be increased with improved transit service, for example. Using a mode-choice model logsum as a multi-modal impedance term in the activity location choice model would enable this kind of analysis.

The monetary benefits we present in this analysis are heavily dependent on two separate assumptions. First, reasonable researchers might have selected different values of time or cost coefficients. Second, the decision to assign one benefit to each household could also have been made differently. A change in either assumption would lead to a highly different total benefits estimate, but it would not change the distribution of the benefits, which is the objective of this study. At some level, converting the esoteric measure of choice model logsums into a unit that can be conveniently compared against other policies is desirable to help the public and policy makers evaluate such decisions. Further research should establish guidelines and practices for applying accessibility logsums in monetary cost-benefit analyses.

Of course, COVID-19 led to the closure of some park facilities — playgrounds, pavilions, and in some cases entire parks — that were not captured in this analysis. These closures would lead to a decrease in the consumer surplus for park access, which might overwhelm or at least change the distribution of positive benefits we measured here.