Estimating Daytime Population Density

Check out the journal article about this project.

I was recently asked: “how might someone figure out the local daytime population density across the Bay Area from public data?” My answer, in short, was that you really couldn’t accurately. But you could at least produce a coarse, biased estimate. Here’s how.

I examined the Bay Area’s tract-level daytime population density using three input data products: the 2010 TIGER/Line census tracts shapefile with DP1 attributes, the 2010 California LEHD LODES data, and the census bureau’s 2010 US states shapefile. I preferred the 2010 census demographic data to (more recent) ACS data because the ACS tract-level variables are five-year rolling averages. Given this, I preferred not to compare 2014 LODES data to 2010-2014 ACS data as the Bay Area experienced substantial housing, economic, and demographic upheaval over this interval – patterns obscured in the ACS rolling average. To avoid inconsistent comparison, I opted for more stale – but more accurate and comparable – data.

I used Python and JavaScript to produce the analysis, the choropleth map above, and the interactive Leaflet web map below (code on GitHub). First, I loaded the LODES data. LODES is block-level and notoriously noisy (and synthetic!) so I aggregated and summed the O-D pairs to the tract-level, at which it converges reasonably well to the observed distribution. I then merged these data with tract-level geometries and populations within ten Bay Area counties (the standard nine-county ABAG/MTC region plus Santa Cruz county).

Next I estimated daytime population density (persons/km²) as tract total population minus outbound commuters plus inbound commuters, normalized by land area. I mapped it by trimming the geometries to California’s landmass. This is slightly coarse-grained, but it makes the bay much more legible than when the water is covered by tracts, and besides, we normalized by land area (this does however raise an interesting question about all those houseboats in Sausalito). Then I projected it and produced a choropleth with seven quantiles. Finally, I dumped the data to GeoJSON to make an interactive Leaflet web map of daytime population density:

So how good are these estimates: what about confidence intervals? Well, we cannot calculate them in a meaningful way from these data, as they are not sampled. The decennial census is a complete enumeration and the LODES data is an administrative payroll enumeration. Had I used ACS data, we could have looked at the population sample estimates and standard errors to determine confidence intervals, but this still would not take into account the LODES enumeration.

The latter points to another issue that makes any confidence in this analysis – and its interpretability – far more fraught. We systematically ignore or undercount the existence and flow of tourists, shoppers, students, the self-employed, various government workers, and populations that are less-legible to these data products, such as certain minority groups and the homeless. For instance, according to its post-enumeration survey, the 2010 census overcounted white Americans and undercounted black and Hispanic Americans. We also know such artifacts tend to exhibit a very lumpy geography. Alternative data, such as smartphone traces, could tell us other sides of this story but suffer the same fundamental problem: biased toward certain populations plus balkanized data ownership.

So let’s muse briefly on takeaways and limitations. If you inspect the map above, density patterns are straightforward and pass the smell test – urban cores around the Bay Area clearly stand out, including Palo Alto and downtown San Francisco. But there are anomalies. Due to its student and government worker populations (which LODES misses), UC Berkeley’s campus shows an absurdly low daytime density.

If we are interested in hypothesis-testing the received wisdom on Jane Jacobs’s “chance encounters,” vitality, and their relationship with human density, artifacts like this would confound our findings. And what of other places that would be prime locations for encounter, but whose daytime populations are drastically under-represented by residence and commute, e.g., Fisherman’s Wharf (tourists), regional parks (picnickers, joggers, hikers, etc.), and various large schools.

To further consider and expand the Fisherman’s Wharf point, not all urban spaces were created equal: the characteristics, culture, and type of density matter as well. Smoothed-out spaces of consumption and escape do not encourage the same types of human interaction as do spaces of challenge and difference. The built form and the transportation–land use connection similarly must be controlled in any such study. For example, an office skyscraper with underground parking might exhibit extremely high daytime population density. But these workers may be sealed in their cars to and from the office, take an elevator up to their desk, and then interact only with a handful of colleagues each day. Does this foster more chance encounters and “urban life” than Jane Jacobs’s lower-density sidewalk ballet? Finally, what about demographics? What is the value or impact of daytime density in wealthy tracts versus impoverished tracts? What about in a demographically-mixed tract with lower density but more opportunities for cultural exchange? What about homeless populations – how does their density contribute to urban life? And, in each of these questions, whose formulation of “good” urban life might we refer to?

Urban studies and planning have uncovered innumerable advantages to density (e.g., vitality, reduced VMT and energy consumption, increased pooling and matching agglomeration efficiencies, etc.). But fundamental to any questions of urbanization, planning, and policy are who we are analyzing, who we are missing, who we are serving, and what values we are embedding consciously or unconsciously in our models. These are critical open questions in scholarship and practice as big data and algorithms fundamentally transform the disciplines of urban planning, policy, and social science. How can we make visible the various essential characteristics, behaviors, and even people that are so often missed or oversimplified in traditional quantitative urban studies and data?