# Bootstrapping from Census data

I have a question about bootstrapping correlated values from grouped data. The context is using Census data grouped by region, $$R$$ (tract or block group). Each region has a list of estimated values $$V$$: median income, population in poverty, age 18 to 64, etc. Each estimate has a margin of error, $$MOE$$. This $$MOE$$ is easily translated to a std error for a 90% confidence interval.

When I use this data, I would like to sample from the regional values according to the distribution of the errors. If I create an error term $$e(v,r)$$ independently for each region $$r$$ and value $$v$$ then I am assuming independence of the $$V$$'s when they are very likely to be correlated.

How should I set up a sampling system to include the correlation patterns within the regions? Is there a common name for this process?

TIA.

• (1) What do you know about the correlations within the regions and how do you know it? (2) Which Census data are you using? In the US American Community Survey you can obtain detailed files constructed specifically to support doing this kind of calculation. They consist of 100 realizations created by the Census Bureau.
– whuber
Commented Apr 7, 2023 at 22:14
• For #1, housing prices and incomes are correlated, income and education levels are correlated. I'd rather assume some correlation and then use the data to determine it versus assume the correlations are zero. #2, yes using ACS 5-year estimate data. Are you referring to the Variance tables, census.gov/programs-surveys/acs/data/variance-tables.html ? Thanks I did not know about these. Commented Apr 10, 2023 at 15:37
• Yes, I was referring to the replicate estimate tables. They should be able to address any issues about correlations among variables. Your post is ambiguous in the sense that it can be understood (as I first did) as referring to spatial correlation. Whether correlation among variables is an issue depends on the analysis you are contemplating.
– whuber
Commented Apr 10, 2023 at 16:22