How to apply tract crosswalk with SEs

NHGIS
1

I’m trying to apply the tract crosswalk (2010>2020) to ACS data. How do I apply the crosswalk to MOEs of count estimates? After transforming the MOE to the variance, can I apply the relevant weight and then aggregate by tract ID? Thank you

2

The ACS handbook, Understanding and Using American Community Survey Data, provides guidance on how to calculate measures of error for derived estimates (Ch. 8, p. 59). The guidance describes how to compute MOEs for sums, which entails taking the square root of the sum of squared MOEs.

That guidance is directly relevant in any case where the crosswalk associates multiple source zones (e.g., 2010 tracts) wholly to a single target zone (e.g., a 2020 tract). But in other cases, wherever the crosswalk weight is between 0 and 1, there is no well-established guidance that I know of. It remains an open research subject.

The situation is greatly complicated by the additional interpolation error introduced when using the crosswalks. The weights in the crosswalk are not exact specifications of the portion of source zone characteristics in each target zone. They are estimates, which introduces error beyond the survey error represented by the ACS MOEs.

We haven’t yet developed any protocols for how users could estimate interpolation error when using the crosswalks. We only provide guidance recommending highly that users start from the smallest possible units, which will minimize the frequency and size of interpolation errors. I.e., if the ACS data you’re crosswalking are available for block groups, you should get block-group-level data and use the crosswalks from 2010 block groups to 2020 tracts to obtain more accurate data for 2020 tracts, rather than using tract-to-tract crosswalks.

If you’d still like to get an estimate of survey error–setting aside the interpolation error–I would suggest using our crosswalks in this way:

  1. Join the crosswalk to the MOE data, as you would for ACS count estimates.
  2. Square each MOE
  3. For the MOEs only, change all nonzero weights in the crosswalk to equal 1. (This will increase the amount of MOE allocated from each source zone to each target zone, which is appropriate given the uncertainty in how much of each source zone’s sample is located in each target zone.)
  4. Then group by target zone (2020 tracts, in your case) and sum the squared MOEs, each multiplied by the adjusted weights
  5. Take the square root of the summed squared MOEs

I’d emphasize that this would only give an inexact estimate of the survey error, and, as I’ve said, doesn’t fully take into account the interpolation error.