STL_DEMOGRAPHY_NhoodPop

This repository contains code for creating a data set originally requested by Janelle O'Dae at the St. Louis Post-Dispatch, containing both total population and race by neighborhood in St. Louis between 1940 and 2017. These use the "modern" neighborhood boundaries, which have only been in use for the last few decades.

Background

Census tracts, which are commonly used units geographic entities in demography and other social sciences, are often used as proxies for "neighborhoods" despite the fact that they rarely map onto real-world neighborhood boundaries (either as they are subjectively understood or objectively delineated). They also change each decade when the U.S. Census Bureau administers the decennial census.

As part of a larger pilot project to understand the impact of racial residential segregation, I have obtained census data, with the help of my research assistant Carter Hanford, for the City of St. Louis. In order to clearly communicate how these changes have progressed over time in geographic units that make sense to most St. Louis residents, I have used these tract-level population counts as the basis for producing neighborhood population estimates. Each year's estimate was produced using R and a technique called areal weighted interpolation. This is typically used for population data (like what we have here) that are overlapping but whose boundaries are incongruent. This research is funded by the SLU Research Institute.

Sources

• Underlying census tract geometry and census data for 1940-1980 were obtained from the IPUMS National Historical Geographic Information System (NHGIS).
• Underlying decennial census data for 1990-2010 and ACS data 5-year estimates for periods ending in 2011 through 2017 were obtained from the Census Bureau via API (using the tidycensus package for R)
• Underlying census tract geometry for 1990-2010 were obtained from the Census Bureau via API (using the tidycensus package for R); all 2011 through 2017 ACS data use the 2010 tract boundaries
• Neighborhood geometry were obtained from the City of St. Louis

Methods

The original census tract geometry for 1940-2000 were cleaned prior to inclusion in this repository to remove tracts outside of the City of St. Louis (they were nationwide files). Data for 1940 through 2000 were imported one year at a time, cleaned, and interpolated to the neighborhood boundaries. These neighborhood estimates were then collapsed into a single data frame. Unit tests were used to verify identification variables uniquely identified each historical census tract (since identification numbers changed each decade). Unit tests were also used to verify that each individual counted in a given census year had been redistributed into a neighborhood - the sum of all census tracts for a given year should equal the sum of all neighborhood populations for that same year.

Data for 2010 through 2017 were downloaded and cleaned one at a time, and then combined into a single data frame. That single data frame was then interpolated to the neighborhood boundaries. As before, unit tests were also used to verify that each individual counted in a given census year had been redistributed into a neighborhood.

The same process was used to create counts for white and black residents by neighborhood. Note that, in 1940, the Census counted only white and non-white residents.

The two data frames, one for the historical data (1940-2000) and one for the modern data (2010-2017) were then combined and written. All R code is found in docs/buildPop.Rmd and docs/buildRace.Rmd. All original source data can be found in data/spatial/ and data/tabular.

Results

Data

This repository contains two clean neighborhood-level data sets - one for population and one for race. These can be found in data/clean/.

Data Visualizations

This repository also contains maps of both population density and percent African American by year for 1950-2010 as well as 2017. These are found in results/. This is an example map, showing the percentage of non-white residents by neighborhood in 1940:

Notes

Non-Integer Data

Users of the data should note that the estimates are non-integer values - the Franz Park population in 1940 has been estimated to be 3308.9920443033734. This, of course, is not possible in a "real world" sense - there are perhaps 3309 or 3307 individuals (though, for reasons described below, the actual population could also be different). This is a by-product of the areal weighted interpolation method, and values can be rounded to achieve an integer estimate.

Sources of Error - All Data

The process I use to create the neighborhood estimates, areal weighted interpolation, introduces some area because it assumes that individuals are evenly spread throughout the source features. Imagine a census tract that is evenly split between a residential area and a commercial area with no residential population. Areal weighted interpolation isn't capable of estimating this level of complexity - it assumes that half the people live in the the commercial area. So, it isn't perfect. It is the simplest approach and the most common for estimating data for overlapping but incongruent spatial features. This approach therefore also introduces noise into the estimates.

Once particular note is that the neighborhood data used for estimates have had six parks removed from them. This reduces the amount of error in the estimation by not allowing anyone to be counted as a resident of the areas covered Carondelet Park, Tower Grove Park, Forest Park, Fairground Park, Penrose Park, O'Fallon Park, Belfontaine/Calvary Cemetery, Missouri Botanical Garden, and Wilmore Park.

Sources of Error - Estimates Based on the ACS

1. The American Community Survey 5-year estimates are pooled over a period of 5 years, so they are the most reliable of the 5-year estimates but the least current. Some analysts use the midpoint year (so 2013-2017's 5 year estimate midpoint year would be 2014) as the "actual year" that these data represent for this reason.
2. The ACS are different than Decennial Census values - the ACS estimates come with a margin of error that represents the number of individuals (plus or minus) the population value may be off by. Each margin of error variable has a _m suffix.

So, bottom line, these data are not census data but rather population estimates, which have been used to produce neighborhood-level estimates. At each level of abstraction, then, there is going to be some error. They're therefore indicative of trends but fundamentally different than then decennial census data.