Replicating 2018 ACS poverty estimates

I am trying to replicate the published estimate of the proportion of people living in poverty according to the one-year 2018 ACS. We have been able to replicate the 13.1 percent proportion of people in poverty that was published here: https://data.census.gov/cedsci/all?q=&g=&t=Poverty&table=S1701&tid=ACSST1Y2018.S1701&hidePreview=false&lastDisplayedRow=16

However, we are not able to replicate the denominator of people for whom poverty status is determined. According to my calculations (see Stata code below), the denominator is 319725862, but in the published table it is 319184033. This difference of 541,829 cases is far outside the margin of error of 23,581 people. I have tried excluding people living in group quarters, but that does not make up the difference.

I appreciate your help.

gen povlevel = .
replace povlevel = 1 if poverty < 100 & poverty >0 & poverty !=. //<100% FPL
else replace povlevel = 2 if poverty <200 & poverty >= 100 & poverty !=. //100-199% FPL
else replace povlevel = 3 if poverty >= 200 & poverty !=. //200+% FPL
svyset cluster [pweight=perwt], strata(strata)
svy: tab povlevel, count format(%12.0f)

This is likely due to the fact that the public use microdata files do not utilize the full ACS sample. The tables on data.census.gov use the full ACS sample, which is only accessible internally to Census Bureau researchers. The weights are not designed to sum to population totals for every possible subset of the data, so there are often small discrepancies like this between the official published estimates and the estimates from public use microdata. There is more on this directly from the Census Bureau here.

1 Like