Mapping 5 Year Data

I’m wondering if there’s a best practice for reporting on/mapping the 5 year data with such wide margins of error, such as a 24.7% poverty rate w/ an MOE of +/-36.2, or even 0% poverty rate w/ an MOE of +/-28. For example, is there an MOE threshold that we could establish as too high to report on (leaving many counties in our map without data). This seems especially tricky when looking at differences between groups – such as difference between white and black poverty rates - when one race/ethnicity has a higher MOE than the other
  • Not sure if you're using ESRI products for mapping, but maybe their article on Margin of Error can help:

    Also, I attended an Association of American Geographers annual conference several years back where David Wong of George Mason University presented a ArcGIS extension he created that overlays a cross-hatching pattern on top of your data values, based upon the margin of error: The "documentation" PDF on their download page is worth reading.
  • This is challenge with the ACS data. We encountered the same issues working on a variety of reports.

    The ESRI report mentioned above is very helpful, we used those recommendations of categories in looking at data, along with a fair bit of research into acceptable data use. The Colorado State Department of Demography also has an informative website.

    During my research I did not find a clear set of thresholds created, which makes sense. I think this also really comes down to helping users learn to use the ACS.

    With counties the option of aggregating, as you would with tracts or bg is not an option. Some of those MOE are larger than the estimate, is there an alternate data source? I too found that looking at data based on race and ethnicity and poverty level created some high MOEs. In one of our reports, the Equity and Opportunity Assessment, we used census data, but in the charts we marked the data with the excessive MOE differently and noted it.

    [Updated on 1/11/2016 4:22 PM]
  • Just reiterating Patrick's earlier post ... a paper by David Wong and Min Sun (2013) titled "Handling Data Quality Information of Survey Data in GIS: A Case of Using the ACS data" is available at:

    [Updated on 1/11/2016 9:23 PM]
  • Thanks, all!

    I'm reviewing the materials you passed along.

    Sarah, I am interested in how you marked the data as excessive, and your examination of race/ethnicity in general. I was hoping to compare differences in poverty levels among race/ethnicity, but that's an even more difficult challenge, as MOEs vary dramtically among groups.
  • Hello Derek,
    The EOA is posted on this page. income and race and ethnicity of residents. We used an empty fill and noted the MOE in the narrative. For example look at chart 6.8 median income by race and ethnicity. If we had kept those values without noting excessive MOE it would tell a different story.

    [Updated on 1/12/2016 4:27 PM]
  • Hello, the way I have approached this problem on a census tract level in a county is to group geographies by ones that have no statistically significant differences within the group. I then end up with a number of groupings where I can say these census tracts (geographies) all have the highest (or lowest, etc) % of the population in poverty in the county. Unfortunately, it take a fair bit of calculation to get there, but I feel allows easier interpretation of the data.
  • Sorry for the late entry in this discussion. I would add a caution about the Esri approach mentioned above which is based on the Coefficient of Variation (CV). CV's make sense when working with metrics like median income where it is important to normalize the estimate. However, CV's can be misleading when the estimate is near zero. For example if the estimate of the percent below poverty is 0% with an MOE of +/- 5%, the CV is infinitely large. The Esri approach would classify this as "Low Reliability" even though we can be quite confident that the percent below poverty is less than 5%.

    In fact, I would not recommend using CVs with percentages at all. There is no need to normalize percentages - they are already normalized. I would just use the MOE directly to assess data reliability.
  • Hi all,

    Thanks again for feedback! Here's the final product. Some of the methodology can be found in the feedback. Happy to answer questions: