I am using IPUMS ACS5 2024 microdata to caculate PUMA-level sums of single men and women in their 30s. (I’m defining ‘single’ here as having a MARST value other than 1, 2, or 9.) To check my numbers, I also calculated similar ratios using pre-calculated Census estimates.
I’m finding that the results of these two calculations diverge somewhat. The mean absolute value of the PUMA-level difference between a Census-estimate-based sum and a microdata-based sum is 175 (for women) and 193 (for men). This isn’t a huge difference relative to the mean number of singles per PUMA (4098 for women and 4654 for men), but it’s still noticeable. The overall correlations between the Census- and microdata-based sums are 0.991 (for men) and 0.992 (for women).
I understand that microdata calculations can diverge from Census-based estimates, but are such large differences to be expected? (I checked my code pretty carefully, but there’s always a possibility that I’m missing something.)
Here are visual comparisons of these differences:
The differences that you are finding between the published estimates and those in the Public Use Microdata Sample (PUMS) are within the expected range. As a reference, I pulled up Table B12002: “Sex by Marital Status for the Population 15 Years and Over” for 2024 (on census.data.gov) and selected estimates and margins of error for Alabama PUMAs. In the screenshot below the margins of error are well above the discrepancies that you are finding.
These differences are due to the published estimates being drawn from a larger internal file than the PUMS. It’s also worth noting that both figures are estimates that are affected by sampling variability and nonsampling error. The ACS technical documentation includes instructions and worked examples for approximating margins of error. The sources of uncertainty in ACS estimates are described in detail in the Understanding and Using ACS Data Handbook.
Our spatial geography project, IPUMS NHGIS, can also be used to download these published estimates with their corresponding margins of error.
Perfect–thanks so much for the detailed explanation, Ivan! It’s good to know that these differences are indeed reasonable.
I think part of my confusion was that, if I remember correctly, some 2020 Census estimates I created using the PPMF (privacy-protected microdata file) matched the Census’s published estimates exactly. So I was initially surprised to see some significant differences between my ACS microdata-based analyses and the Census’s own ACS estimates. But now I understand that such differences are to be expected. 
