Census Intractables

For a city whose inhabitants love to claim ownership of some kind, it’s amazing what we overlook. I grew up a few minutes outside of New York City (in New Jersey) and lived in Manhattan and Brooklyn for over ten years. I loved the familiarity and knowing my away around; on the flip side, I also loved stumbling upon street names and subway stations that I hadn't seen before.

I was caught off guard the first time I saw Governeur St. and Montgomery. They are over a hundred years old. First, they were home to tenement buildings, and later on, blocks and blocks of public housing.  I read about the history of the Vladeck houses, the first city-funded projects, founded in 1941, still down on nearby Water St. In 2005, the 24 six-story buildings finally had elevators installed that went past the fifth floor. For Jer Thorp's class in data representation, I looked at census data to learn more about this neglected area, as a way to better understand the neighborhood on the whole. I limited the scope of the assignment to sixteen census tracts that comprise the area below Bowery to the East River (focusing mostly on Community Board 3, which includes Chinatown).

What struck me about the census is the difficulty of categorization. We own our identities and what we call them. The census attempts to name us, and to some extent, it's very useful. As far as looking at tract data over time, these definitions are limited, occasionally absurd.

ACQUIRING DATA

Each resource had its potential shortcomings. The ACS (American Census Survey), for example, is distributed to 300,000 households, monthly, on a rolling basis, whereas the census is for one date. ACS releases five-year reports, with margins of error, but it only goes back to the 90s. While the aggregation tool, Social Explorer draws from census data from the 40s to present, it neither reports margin of error, nor normalizes geographies at the tract level. But it's such a pleasure to use. The Geolytics neighborhood database was my best bet, as it normalizes tracts from 1970 to 2000 and maintains roughly the same variables for each decade. Unfortunately, the database is only available on Internet Explorer and often crashes during the middle of a call.

I loaded the data into R and made heatmaps to get a sense of trends, as I discussed here, here, and here. My two main takeaways were: 1) there is a surprising amount of white poverty on the Lower East Side, through 2000 and 2) the tract zoning is only somewhat useful. I don't know how to account for specific streets, though it's clear enough that the poorest regions are where the projects are. One glaring error is that I didn't include Asian/Pacific Islander, an oversight. The central problem of my dataset is that it's impossible to compare one year to another. In 2000, respondents were able to check multiple boxes for race; categories expanded each year. In 2000, the census had included a wide range of Latin-American countries to check off, whereas 1980 only included Puerto Rico, Chile, Colombia and Mexico.

Heatmap Impressions

1. Grids and color palettes are useful in isolating patterns we wouldn’t see otherwise, but they can also quash reality. By dividing up the region into races and tracts, I obliterated patterns of interrelationships. How did class dynamics play out, for example? At the same time, my demographic research is based around race and class, so I need to isolate those variables.

2. The underlying theme here is systems of exclusion: socioeconomic inequality, census categories, arbitrary tract zones and heatmap colors. For example, the census couldn’t take into account squatters or community gardens, significant aspects of community-building in the 70s and 80s. “Asian” first became distinct from “Asian, Hawaiian and Pacific Islander” in the 90s census. The same year, “Homeless” was a new category. In 2000, respondents had a choice of selecting more than one box for race. The differences between the 1980 and 2000 censuses are pretty astounding.

2. I realized I lacked data. I had marriage/household, poverty, overall population info, mostly. In order to understand who these people are, I needed to get more specific; for my purposes, I am prioritizing education level, employment, age ranges, and languages spoken (Spanish only, English only, Spanish and English, etc.).

The difficulty in assessing demographic shifts between 1980 and 2000 is that, broadly, there are two possibilities: 1. demographics are changing 2. self-identification is changing. There is no way to know. By using coarse racial categories to keep my data uniform, I was messing with this ambiguity. I could recombine the split groups and ‘80 and '90, but that wouldn't change the fact a person can check all the boxes for race. 

3. Looking at percentages in each tract doesn’t tell us anything about specific streets. Early on in this project (see images at top), I juxtaposed a heatmap with a census tract map and a Google map. Each of these images are meaningful yet short-sighted in their own way. So, I'd have to refine my data, then come up with another approach way to highlight social phenomena. 

I scrapped the racial data, and returned to my dataset to see what other fields I could analyze. The census is crazy—it tracks whether water heating fuel is kerosene or gas and the number of people who get their water from an individually-dug vs. individually-drilled well. I acquired data for average commute time (under 20 minutes, between 20 and 44 minutes, over 45 minutes).

I also considered other ways to display the information. Spreadsheet and heatmap grids were useful to a point—it was hard to understand the differences between tracts without having a map out at the same time. What I couldn't ignore, though, was that the heatmaps were pretty. As diagrams, they'e very basic: they connect and relate concepts. Perhaps ironically, their simplicity canceled out my ability to draw any conclusions. I saw gradations of color, I could point out the spikes, but I couldn't extrapolate much further. There are multiple factors that could lead to a high concentration of single mothers in one tract; at such a granular level, I couldn't explain why that was.

At Jer's suggestion, I tried a new visualization approach, displaying the data in all its messy glory. I was particularly inspired by Lisa Jevbratt's piece, "Migration," which looks at IP addresses over time and the overlap of which years' IP's have websites. I also appreciated her description of two other, related works. I like the idea of data visualization as a kind of third logic, somewhere between image and text. Images that stand as their own entities, not as representations of, or supplements to another object or concept.

A new dataset: commute times

I began to look at racial population data in Processing and then moved on to commute times. In spirit of Jevbratt's abstractions, I represented each tract as a cluster of points, where the number of points and radius are relative to the size of the population. I placed each tract relative to its distance from the average latitude and longitude, which I set as the middle of the screen. Each year is spaced 100 pixels apart.

When I first created this piece in 2012, I wrote that: In the future, l'd like to 1) make the two-dimensional logic of the piece more understandable: place the tracts at their actual latitude and longitude and create a key. And 2) place the tracts relative to the percentage change in a variable from year to year. 

UPDATE 2016: The end result looks like an interesting constellation, though I don't think the technique of displaying the data in a less-obvious, artistic context really conveys the commute times dataset that I chose to work within the end. When working with geographic data and displaying how people's commute habits changed over time, the end result could have been as simple as a overlaying the data points in Mapbox.