Poverty via PUMS and Poverty using table B17001

When I compute poverty status using the PUMS variable POVPIP <100 (percent) to create a variable "Below Poverty Level"  and "Above or equal to Poverty Level'  I get no missing data and a total population of 124,553 (OHIO CLEVELAND CENTRAL PUMA state ohio 39 puma fips 00906)  whereas ACS 2020 vintage 5 year table B17001 gives a total population of 120,938 with a universe of  "Population for whom poverty status is determined "  for the PUMA. So the PUMS data gives a difference of 3615.  Presumably the PUMS excess consists of "population for whom poverty status is NOT determined."  Does anyone know what PUMS variable needs to be used to set the 3615 individuals to "undetermined?"  Could these be group quarters people? https://www.census.gov/topics/income-poverty/poverty/guidance/poverty-measures.html  For example people in college dormitories don't have poverty status assigned.  ACS table B01001 (Age by Sex) gives a total population for the PUMA of 124,903. So the PUMS calculation for total population looks correct (I using the R survey package to handle weights)  I would expect that the PUMS and ACS detail tables would not give exactly the same total population because of disclosure avoidance Thanks in advance. Dave Dorer

  • Hi Dave -

    There are a couple of issues here. First the specific question. On the poverty variable you're using (I'm assuming you're using POVPIP) there are system-missing cases. Those indicate persons for whom poverty status is not determined.

    Second, PUMS estimates have always differed from published tabulations. "Because PUMS data consist of a subset of the full ACS sample, tabulations from the ACS PUMS will not match those from published tables of ACS data." (See page 16 here: https://www.census.gov/content/dam/Census/library/publications/2021/acs/acs_pums_handbook_2021_ch04.pdf

    For more information on the accuracy of PUMS, I'd suggests: https://www2.census.gov/programs-surveys/acs/tech_docs/pums/accuracy/2016_2020AccuracyPUMS.pdf 

    Hope this helps!

  • Dear Beth,

    Thanks for replying on the weekend.  You must be a night owl like me. I'm hoping that someone has reconciled the PUMS calculation to the ACS detailed tables numbers for poverty level. When I do this with other variables I usually come within what looks like the sampling error of the PUMS records (if I were being rigorous I would check the replicate weights but I haven't dug that deep yet.) What I am hoping for is being able to figure out how to use other PUMS variables to tease out the subgroup for whom "poverty status is not determined."  If someone on the Forum hasn't done it, I'll submit to census.data@census.gove and there should be an expert who knows how the calculation is made. Have a great W.E.

    I want to use the PUMS data to build a multivariate loglinear model with poverty status being one of the variables. Then I'm going to "rake" the loglinear fit to adjust to the B17001 table marginal for a census tract (along with some other ACS detail tables)   I need to get the loglinear model total N (from the PUMS data PUMA area to add up to (about) the B17001 marginal total for the PUMA so I know that the PUMS based calculation agrees with the calculation for B17001. I need to know that B17001 is defined the same way as the PUMS calculation.   I seem to need a 3rd category poverty below/poverty above/undetermined Then I can use B17001 for a census tract to "pull down" the PUMA level loglinear model to the census tract.  That's where I;m headed with this.

  •  I was able to answer my own question after doing some programming.  I have checked this for 1 PUMA and will do more in the coming days. Here is my solution as of now

    I was able to solve the problem by removing unrelated children (RELSHIPP) less than 15 years old (AGEP) and those in group quarters (TYPEHUGQ).

  •  I was able to answer my own question after doing some programming.  I have checked this for 1 PUMA and will do more in the coming days. Here is my solution as of now

    I was able to solve the problem by removing unrelated children (RELSHIPP) less than 15 years old (AGEP) and those in group quarters (TYPEHUGQ).

  • More details on a few adjustments and an email from the people at the Census:

     I've just been through a round with the people at the Census on reconciling "below poverty" from the PUMS with "below poverty" for table B17001.  Poverty is not defined for households only for families which includes only people in the household who are related to the "head of household."  For example a renter not related to the family is not included in the calculation and is not part of the "universe." For the calculation with partial details see: https://www.census.gov/topics/income-poverty/poverty/guidance/poverty-measures.html.   You add up the income for all family members and then apply the formula which is based on total family income, number of family members and number of children.  Here is the email that I just received from the people at the Census: (regarding matching numbers from B17001 and the calculation using the PUMS variables POVPIP RELSHIPP TYPEHUGQ AGEP.

    Hi David,
    According to the poverty subject matter experts, "It is possible that they [PUMS and B17001] are not matching because the poverty universe excludes a few more types of populations  1. children under the age of 15 who are not related to the householder 2. people living in institutional group quarters (nursing homes and correctional facilities) 3. people living in college dormitories 4. people living in military barracks. Perhaps by excluding all not in the universe he will be closer, that being said we have confidence in the estimates he mentions from 2020. The estimate 120,888 is within the moe of the PUMA that can be seen S1701 (120,938) on data.census.gov, so we have confidence in it."
    Let me know if this helps or not David.