Rules of thumb for determining if tract-level data are too noisy/unreliable?


I’m using tract-level estimates from a 5-year ACS sample (number of renter-occupied households), and am trying to filter out unreliable data. It’s clear when estimates are extremely unreliable (e.g., the margin of error is larger than the estimate), but are there generally accepted thresholds in determining data validity? Margin of error no more than X, with a sample size of at least Y?

Looking at Census’ guidance on using margins of error (, I understand how to determine if two estimates are statistically different. But if I’m trying to draw conclusions about trends across all tracts, and so need a reliable dataset of tract-level data, how can I go about only using tract estimates that are satisfactorily un-noisy?


I think there are some suggestions for dealing with this issue toward the end of this paper.
Folch, David C. Daniel Arribas-Bel, Julia Koschinsky and Seth E. Spielman. (2016). Spatial Variation in the Quality of American Community Survey Estimates. Demography, 53:5 pp. 1535-1554.

1 Like

Thanks John! This paper was helpful in describing the non-random pattern that MOEs follow in the ACS data. Though I might hope for a straightforward set of rules, it looks like their recommendations would vary on a case-by-case basis.

I found a PDF of the working paper version here: