ArcGIS Living Atlas

How aggregation resolves reliability concerns for American Community Survey data

Many GIS analysts work with American Community Survey (ACS) data from the U.S. Census Bureau. This data is based on a sample of the population, therefore, like all survey data, it contains error due to sampling. Census publishes margins of error along with each and every estimate they publish, which is incredible. Viewing margins of error for a specific measure on a specific tract can erroneously lead users to dismiss the entire ACS. Worse, users may not know that the typical geographic analysis of ACS data actually increases the data reliability by a noticeable amount.

The way that Census is able to publish estimates down to the census tract and even block group level is by coarsening/aggregating across years. They pool together 5 years of data in order to produce estimates at such fine geographic levels. In addition to coarsening across time, we can also coarsen across space.

While this may seem like heresy to many geographers, there are benefits to coarsening, or aggregating up geographically. Many people need data disaggregated by race/ethnicity, gender, income, and more. Well, coarsening can make this data more reliable. Some cities have gone through robust validation processes to create their own geographies that are coarser than census tracts. For example, New York City has their own Neighborhood Tabulation Areas, and Houston has their own Super-Neighborhoods. We can tap into geography to test your own groups of tracts. You can call them anything you like. I’ve been calling them super-tracts.

Introducing: the ACS Summarization App

This lightweight app is designed to help you get a quick sense of how many tracts you’ll need to aggregate together to get an estimate that meets your desired reliability. The left panel lets you search and select layers and attributes. If you select a count, the tract centroids will be symbolized by size. If you select a percent, the tract centroids will be symbolized by color.

The right panel lets you see your summary statistics. The gauge shows the Coefficient of Variation (CV), which is calculated on-the-fly. The lower the coefficient, the higher the reliability. Conversely, the higher the coefficient, the lower the reliability.

Coefficient of Variation (CV) Reliability
 CV <= 12 high
12 < CV <= 40 medium
40 < CV low

Use the sketch tools on the top of the map to create your super-tract. Watch the gauge change as you modify your sketch!

As a circle is expanded to select several tracts, the reliability gauge on the top right goes down, and the summary statistics are presented on the bottom right.

Still not convinced? Watch this 5-minute video:

 

Tract summarization best practices

First, draw a shape around the neighborhood or district you know you want to analyze. Get the basic shape on the map to start.

Preserve patterns in the map

Try to follow the patterns in the map when combining. For example, combine high values with other high values and lower values with other low values. The symbology in this app can help, but your own local on-the-ground knowledge is invaluable here. I realize this is hard when aggregating just a few tracts, and also balancing other considerations. However, combining areas with wildly differing characteristics will smooth out the numbers and become less informative.

Be mindful of tracts with an estimate of zero

Tracts with an estimate of zero are symbolized by the transparent teal symbols (for counts). These are generally in places such as airports, cemeteries, open land, and so on, but they can be anywhere.

A very faint, small teal circle by the Long Beach airport is highlighted, along with a few other faint, small teal circles in the South Bay area of Los Angeles.

Even zero estimates have margins of error, which means there may be a few individuals in your population of interest in these tracts. They will not add anything to your estimate, but because they have a nonzero margin of error, they could add slightly to your error.

Use official estimates when available

Additionally, this aggregation method will only give you an approximation for both the estimate and the margin of error using the approximation formulas in the ACS Handbook. Census produces official estimates for school districts, incorporated places, county subdivisions, congressional districts, and much more. If you’re aggregating up to get values of a defined census geography, such as a city boundary, check data.census.gov for an official estimate you can use. Not only does the official estimate use the true boundary, but it will also have a lower MOE than you’d get from aggregating tracts, since it’s not being approximated.

Adjust your reliability comfort-level depending on the nature of the estimate

Ideally, we’d like all estimates to have high reliability, with a low coefficient of variation. However, this will be hard to achieve for very small populations, for example:

In cases like these, could you live with a medium level of reliability if that means finer geographic detail?

Also, neighboring tracts can have different reliability for the same attribute. Sometimes tracts do have reliable estimates, so there’s no need to aggregate unnecessarily. Why lose the geographic detail?

Use this app to iterate quickly and dynamically

Some last few details:

This app is designed to give you a jumping-off point for creating aggregations of tracts. It’s designed to help you iterate quickly and dynamically, to get a quick sense of how much you’ll have to aggregate up geographically in order to meet your reliability comfort-level. Geography can actually help us work with error, rather than being scared by it.

About the author

(she/her/hers) Diana loves working with data. She has over a decade of experience as a practitioner of demography, sociology, economics, policy analysis, and GIS. Diana holds a BA in quantitative economics and an MA in applied demography. She is a senior product engineer on ArcGIS Living Atlas of the World's Policy Maps team. Diana enjoys strong coffee and clean datasets, usually simultaneously.

0 Comments
Inline Feedbacks
View all comments

Next Article

Searching for Features in Maps and Apps

Read this article