Calculating Residential Segregation Indices

The purpose of this repository is to provide a simple and reproducible R pipeline to investigate residential segregation (RS) using US census data. The pipeline contains two components:

1. pulling decennial US census data of Year 2000, 2010, 2020 via R package tidycensus
2. calculating three residential segregation indices*, including dissimilarity, isolation and interaction indices, at the preferred geographical level, e.g. county or census tract level

*: The implementation of the three scores in this pipeline is neither endorsement nor an emphasize. Researchers should study all residential segregation indices to choose the ones that match with their study objectives.

We hope this work would relieve social scientists from repetitive data pulling, and inspire to adapt the reproducible pipeline.

Note: If you don’t have the time to go through the whole documentation, please finish reading the Remarks Section before modifying the code

Get Started

1. Download R and RStudio IDE
2. Install the necessary workflow packages targets and renv if you don’t already have
3. Open the R project in RStudio and call renv::restore() to install the required R packages. Please give permission to install the necessary packages. This will mirror the version of packages used in the creation of the work exactly.
4. Acquire your census api key string via https://api.census.gov/data/key_signup.html, and replace at the beginning of the _targets.R file
5. Modify the code to reflect your research needs. We highlight the places that requires customization with the tag TODO:, which can be enlisted via a global search, i.e. cmd/control + shift + f.
   • Change year, states, and geographic levels, where the upper level is your preferred level and the lower level is the level constituent the upper level. For example, in order to calculate county level indices (upper level), we need to have census tract level statistics (lower level).
   • Confirm if the variable codes in census databases match with your preferred variables. The variable code is year-specific, i.e. could be different depending on the year you use. For example, the same variable code P003003 means Total!!Population of one race!!White alone in 2000 data, and means Total!!Black or African American alone 2010 data
6. call targets::tar_make() in the console to run the pipeline.

Examples

In this section, we provide two examples for calculating RS indices of one state (stored in master branch) or multiple states (stored in meds_desert branch) respectively.

One State Example: 2010 Alabama Dissimilarity Index at County Level

In this example, we provide the pipeline to calculate the indices for a single state. As an bonus, a section of code that plots a index to the map are supplied, as shown in Figure 1*. Figure 1 includes two maps of the 2010 Alabama County Level dissimilarity index, White (majority) with respect to Black (minority), calculated using different definitions of lower level geographic unit, i.e. census tract level and block level.

*: The example pipelie only produces one of the plots, where caption had been manually modified.

Figure 1: 2010 Alabama Dissimilarity Index at county level calculated with census tract level statistics (a) and block level statistics (b)

Note: For those who are interested in 2020 AL Indices, please refer to test case in Issue 7 as an example of configuration

Multiple States Example: 2010 RS Indices of Medication Desert at Census Tract Level

The example demonstrate how to calculate residential segregation for multiple states collectively, either via an intput file or via an inline code. Please find the pipeline via meds_desert branch.

Remarks

In this section, we discuss our observations when creating the indices, which includes thoughts on numeric calculation with census data, interpretation, and practice of data sharing.

Numeric Calculation

1. During the calculation, we observe that depending on how the areal units are defined, it is possible to have areas with no majority or minority population at all, i.e. n_majority = 0 or n_minority = 0. This complicates the calculation of the scores, for example, introducing infinite or NaN as a score. Without finding any remedies in the literature yet, we defined these indices as missing values collectively.

2. It is very important to confirm if your variable codes match with your anticipated variable with the census data base. Even though we build error prevention mechanism in the code to numerically verifies, we are not certain it will catch the error 100% particularly with the flexibility that allow users’ customization. We provided how code changes in different years, see Get started.

3. In the calculation, we do not assume that the minority numerically complements the majority, i.e the numbers of minority and majority sums to the total. These indices would be different from the indices calculated with the complementing assumption.

4. The indices are claculated using their definition, which means they are possibly directionally different in their interpretation, e.g. interaction and isolation indices. The user will have to define their own reverse coding function to yield directionally consistent interpretation.

Interpretation

The interpretation of residential segregation indices gets complicated quickly depending on their areal definition. Hence, we don’t offer too much suggestions. We highly recommended the user to carefully go through Massey and Denton (1988) for more details when planning which indices and which areal unit to use in calculation. For people who seeks real-world example, we defer to Iceland and Weinberg (2002).

The followings are a few questions we had when calculating the residential segregation across the US, instead of specific metropolitan areas in previous studies.

1. The residential segregation indices are caluclated by agregating statistics of smaller areal units, where the lower arear units can be defined differently. For example, when calculating the residentital segregation indices at the county level, it is possible to define the smaller areal unit be census tract or block. Different definitions can yield inconsistent scores, both the magnitude of scores and the ranking of the scores. How to interpret the inconsistency due to the different definitions of lower area units remains as a question to us.

2. During calculating RS across multiple states, we observed there are few majority and minority with in an areal unit when we don’t assume majority and minority sum to total. For example, if we defined the White as majority and the Black as minority, there are few majority and minority in a areal unit that are in Indian reservation. Are the indices still well-defined in this case?

3. With improved efforts to collect more diverse racial information, it is possible to have individuals who have more than one racial background. For example, in 2010 census data, we have a variable code P003008 for Total!!Two or More Races. How to utilize this informaiton in calculating racial RS can lead to a fruitful discussion.

Data Sharing Practice

1. When sharing the indices with FIPS and GEOID as identifier, it is very important to treat FIPS and GEOID are characters/strings instead of numeric values because of the leading zeros. This practice will save more time and produce less error when sharing with others.

Questions/Contribution

We prefer questions or bug reports via Issues tab of the repository, such that the answer to your question can serve a broader audience. We are also open to questions via Email if you don’t feel comfortable with the aforementioned approach.

If you would like to contribute to this tutorial, we are welcome any contribution via pull requests so that you get proper credit.

References