More Accurate Birth cohorts

I’ve seen the post “ACS: How do errors in birthyr average to zero?” and my question is in the same vein. Is there anyway to assign observations to more accurate birth cohorts? The idea that “the errors averaged across the entire sample will be close to zero” is very misleading if someone wants to use it for any sort of analysis, the best case scenario is significant attenuation bias concerns. Is it possible to construct a more accurate birth year through available data? Or any other way around this issue?

The response from staff indicated the code book for the quote above should be clarified and I concur. It is tacitly encouraging the variable to be used when it has some pretty big concerns (which the rest of the codebook tries to indicate).

EDIT: I realize some nature to how I’m using the data would be helpful. Basically I need, using ACS data from 2005-2017, to assign people to the same cohorts for schooling purposes, ie be able to refer to the 1992 class for Alabama in some capacity while keeping outcome variables such as income. Previously, I had been using the birth year + 6 for year started Kindergarten if born in first 3 quarters and + 7 if born in quarter 4 (I realize this isn’t perfect either but is small measurement error compared to the birth year issue).

Thank you for your time,
Jon M.

As noted in the forum post you reference, the error for BIRTHYR in ACS data is either 0 or 1. I will follow up with the IPUMS USA team to request that the documentation on the comparability tab clearly reflect this. For some research applications this single year error may be problematic. Unfortunately, there is no way to more accurately estimate year of birth given the detail available in public use ACS data. You could certainly use different data that allow you to get at birth year more accurately assuming you don’t require variables or geography that are specific to ACS data.

Thank you for your response!

1 Like